Skip to content

GitLab

Commit 48ec26d1 authored by Paper99's avatar Paper99
Browse files

Update PhotoMaker V2

parent 1e78aa65
Hide whitespace changes
Inline Side-by-side
Showing with 3831 additions and 115 deletions
README.md
View file @ 48ec26d1
...@@ -10,17 +10,17 @@ ...@@ -10,17 +10,17 @@
## PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding [![Paper page](https://huggingface.co/datasets/huggingface/badges/resolve/main/paper-page-md-dark.svg)](https://huggingface.co/papers/2312.04461) ## PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding [![Paper page](https://huggingface.co/datasets/huggingface/badges/resolve/main/paper-page-md-dark.svg)](https://huggingface.co/papers/2312.04461)
[[Paper](https://huggingface.co/papers/2312.04461)] &emsp; [[Project Page](https://photo-maker.github.io)] &emsp; [[Model Card](https://huggingface.co/TencentARC/PhotoMaker)] <br> [[Paper](https://huggingface.co/papers/2312.04461)] &emsp; [[Project Page](https://photo-maker.github.io)] &emsp; [[Model Card](https://huggingface.co/TencentARC/PhotoMaker)] <br>
[[🤗 Demo (Realistic)](https://huggingface.co/spaces/TencentARC/PhotoMaker)] &emsp; [[🤗 Demo (Stylization)](https://huggingface.co/spaces/TencentARC/PhotoMaker-Style)] <br> [[💥New 🤗 Demo (PhotoMaker V2)](https://huggingface.co/spaces/TencentARC/PhotoMaker-V2)] &emsp; [[🤗 Demo (Realistic)](https://huggingface.co/spaces/TencentARC/PhotoMaker)] &emsp; [[🤗 Demo (Stylization)](https://huggingface.co/spaces/TencentARC/PhotoMaker-Style)] <br>
[[Replicate Demo (Realistic)](https://replicate.com/jd7h/photomaker)] &emsp; [[Replicate Demo (Stylization)](https://replicate.com/yorickvp/photomaker-style)] <be> [[Replicate Demo (Realistic)](https://replicate.com/jd7h/photomaker)] &emsp; [[Replicate Demo (Stylization)](https://replicate.com/yorickvp/photomaker-style)] <be>
If the ID fidelity is not enough for you, please try our [stylization application](https://huggingface.co/spaces/TencentARC/PhotoMaker-Style), you may be pleasantly surprised. If the ID fidelity is not enough for you, please try our [PhotoMaker V2](https://huggingface.co/spaces/TencentARC/PhotoMaker-V2) or [stylization application](https://huggingface.co/spaces/TencentARC/PhotoMaker-Style), you may be pleasantly surprised.
🥳 We release PhotoMaker V2. Please refer to [comparisons](./README_pmv2.md) between PhotoMaker V1, PhotoMaker V2, IP-Adapter-FaceID-plus-V2, and InstantID. Please watch [this video](https://photo-maker.github.io/assets/demo_pm_v2_full.mp4) for how to use our demo.
</div> </div>
--- ---
Official implementation of **[PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding](https://huggingface.co/papers/2312.04461)**.
### 🌠 **Key Features:** ### 🌠 **Key Features:**
...@@ -41,8 +41,9 @@ Now we know the implementation of **Replicate**, **Windows**, **ComfyUI**, and * ...@@ -41,8 +41,9 @@ Now we know the implementation of **Replicate**, **Windows**, **ComfyUI**, and *
## 🚩 **New Features/Updates** ## 🚩 **New Features/Updates**
- ✅ Jan. 20, 2024. An **important** note: For those GPUs that do not support bfloat16, please change [this line](https://github.com/TencentARC/PhotoMaker/blob/6ec44fc13909d64a65c635b9e3b6f238eb1de9fe/gradio_demo/app.py#L39) to `torch_dtype = torch.float16`, the speed will be **greatly improved** (1min/img (before) vs. 14s/img (after) on V100). The minimum GPU memory requirement for PhotoMaker is **11G** (Please refer to [this link](https://github.com/TencentARC/PhotoMaker/discussions/114) for saving GPU memory). - ✅ July 20, 2024. 💥 We release PhotoMaker V2 with **improved ID fidelity**. At the same time, it still maintains the generation quality, editability, and compatibility with any plugins that PhotoMaker V1 offers. We have also provided scripts for integration with [ControlNet](./inference_scripts/inference_pmv2_contronet.py), [T2I-Adapter](./inference_scripts/inference_pmv2_t2i_adapter.py), and [IP-Adapter](./inference_scripts/inference_pmv2_ip_adapter.py) to offer excellent control capabilities. Users can further customize scripts for upgrades, such as combining with LCM for acceleration or integrating with IP-Adapter-FaceID or InstantID to further improve ID fidelity. We will release technical report of PhotoMaker V2 soon. Please refer to [this doc](./README_pmv2.md) for a quick preview.
- ✅ Jan. 15, 2024. We release PhotoMaker. - ✅ January 20, 2024. An **important** note: For those GPUs that do not support bfloat16, please change [this line](https://github.com/TencentARC/PhotoMaker/blob/6ec44fc13909d64a65c635b9e3b6f238eb1de9fe/gradio_demo/app.py#L39) to `torch_dtype = torch.float16`, the speed will be **greatly improved** (1min/img (before) vs. 14s/img (after) on V100). The minimum GPU memory requirement for PhotoMaker is **11G** (Please refer to [this link](https://github.com/TencentARC/PhotoMaker/discussions/114) for saving GPU memory).
- ✅ January 15, 2024. We release PhotoMaker.
--- ---
...@@ -240,6 +241,7 @@ Provided by [@Gradio](https://twitter.com/Gradio/status/1747683500495691942) ...@@ -240,6 +241,7 @@ Provided by [@Gradio](https://twitter.com/Gradio/status/1747683500495691942)
# 🤗 Acknowledgements # 🤗 Acknowledgements
- PhotoMaker is co-hosted by Tencent ARC Lab and Nankai University [MCG-NKU](https://mmcheng.net/cmm/). - PhotoMaker is co-hosted by Tencent ARC Lab and Nankai University [MCG-NKU](https://mmcheng.net/cmm/).
- Inspired from many excellent demos and repos, including [IP-Adapter](https://github.com/tencent-ailab/IP-Adapter), [multimodalart/Ip-Adapter-FaceID](https://huggingface.co/spaces/multimodalart/Ip-Adapter-FaceID), [FastComposer](https://github.com/mit-han-lab/fastcomposer), and [T2I-Adapter](https://github.com/TencentARC/T2I-Adapter). Thanks for their great work! - Inspired from many excellent demos and repos, including [IP-Adapter](https://github.com/tencent-ailab/IP-Adapter), [multimodalart/Ip-Adapter-FaceID](https://huggingface.co/spaces/multimodalart/Ip-Adapter-FaceID), [FastComposer](https://github.com/mit-han-lab/fastcomposer), and [T2I-Adapter](https://github.com/TencentARC/T2I-Adapter). Thanks for their great work!
- Thanks to the [HunyuanDiT](https://github.com/Tencent/HunyuanDiT) team for their generous support and suggestions!
- Thanks to the Venus team in Tencent PCG for their feedback and suggestions. - Thanks to the Venus team in Tencent PCG for their feedback and suggestions.
- Thanks to the HuggingFace team for their generous support! - Thanks to the HuggingFace team for their generous support!
......
README_pmv2.md 0 → 100755
View file @ 48ec26d1
<p align="center">
<img src="https://photo-maker.github.io/assets/logo.png" height=70>
</p>
<!-- ## <div align="center"><b>PhotoMaker</b></div> -->
<div align="center">
## PhotoMaker V2: Improved ID Fidelity and Better Controllability Compared to PhotoMaker V1
[[🤗 Demo](https://huggingface.co/spaces/TencentARC/PhotoMaker-V2)]
</div>
When training PhotoMaker V2, we focused on improving ID fidelity. Compared to PhotoMaker V1, we introduced 1️⃣ new training strategies, incorporated 2️⃣ more portrait datasets, and utilized 3️⃣ a more powerful ID extraction encoder. We will release a technical report soon. Thank you all for your attention.
### 🌠 **Key improvements in PhotoMaker V2:**
1. **ID fidelity** has been **further improved**, especially for single image input and Asian facial inputs. Of course, feeding more facial images can still yield better results.
2. By integrating [ControlNet](./inference_scripts/inference_pmv2_contronet.py), [T2I-Adapter](./inference_scripts/inference_pmv2_t2i_adapter.py), and [IP-Adapter](./inference_scripts/inference_pmv2_ip_adapter.py), the generation process becomes **more controllable**. We provide corresponding scripts for reference. Additionally, PhotoMaker V2 allows users to achieve better ID consistency by combining it with IP-Adapter-FaceID, InstantID, and [character LoRA](https://github.com/TencentARC/PhotoMaker/discussions/14).
3. PhotoMaker V2 **inherits the promising features of PhotoMaker V1**, such as high-quality and diverse generation capabilities, and powerful text control. Additionally, it can still integrate previous applications like bringing characters from old photos or paintings back to reality, identity mixing, and changing age or gender.
## Comparisons with PhotoMaker V1, IP-Adapter-FaceID and InstantID
We selected the three most prevalent methods in ID personalization generation, namely PhotoMaker V1, [IP-Adapter-FaceID-Plus-V2](https://huggingface.co/h94/IP-Adapter-FaceID) ([best of IP-Adapter-FaceID](https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/195)), and [InstantID](https://github.com/InstantID/InstantID).
To ensure a fair comparison, we used the same base model ([RealVisXL-V4.0](https://huggingface.co/SG161222/RealVisXL_V4.0)) and scheduler ([Euler](https://huggingface.co/docs/diffusers/api/schedulers/euler)), and selected the best out of four randomly generated images from each method for visualization. The prompts and negative prompts were consistent:
Prompt: `instagram photo, portrait photo of a woman img holding two cats, colorful, perfect face, natural skin, hard shadows, film grain`
Negative Prompt: `(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth`
We can see that our method has **advantages** in maintaining ID fidelity and in the quality of the generated images
![comp_pm_v2_reba](https://github.com/user-attachments/assets/b978ffa2-97c9-4910-ab23-a2b2edd3be1d)
![comp_pm_v2_musk](https://github.com/user-attachments/assets/6b96d65b-813a-45e0-8f7a-25041dc4dc10)
![comp_pm_v2_yanzu](https://github.com/user-attachments/assets/b788b2b0-9166-4c9d-aa46-24ef1fb4e5a9)
![comp_pm_v2_yifei](https://github.com/user-attachments/assets/66fa8a73-8973-4e40-a094-c4cb3eec8d8a)
## Cooperation with ControlNet / T2I-Adapter / IP-Adapter
PhotoMaker V2 can collaborate with [T2I-Adapter’s doodle mode](https://huggingface.co/TencentARC/t2i-adapter-sketch-sdxl-1.0),
allowing for controlled image generation based on user drawings and prompts.
This feature can be experienced in [[🤗 our official demo]](https://huggingface.co/spaces/TencentARC/PhotoMaker-V2).
The following video is an example of the experience process:
https://github.com/user-attachments/assets/1303d684-89e4-49d2-8e8c-4b659c8b48e7
Additionally, PhotoMaker V2 can work with [ControlNet](https://github.com/lllyasviel/ControlNet) and [T2I-Adapter](https://github.com/TencentARC/T2I-Adapter) for layout control, such as edge, pose, depth, and more.
We provide two example scripts:
1. [inference_pmv2_contronet.py](./inference_scripts/inference_pmv2_contronet.py)
2. [inference_pmv2_t2i_adapter.py](./inference_scripts/inference_pmv2_t2i_adapter.py)
The image below is an example of controlled generation using pose through ControlNet:
![pm_v2_controlnet](https://github.com/user-attachments/assets/57767447-192c-4606-af2a-4206b5dbccf9)
Our sample scripts can be referred to:
[inference_pmv2_ip_adapter.py](./inference_scripts/inference_pmv2_ip_adapter.py)
The image below is an example:
![pm_v2_ipadapter](https://github.com/user-attachments/assets/89f95604-6cfa-4dde-b563-2d052bac14cc)
PhotoMaker V2, as a plugin, can work well with other plugins, such as IP-Adapter-FaceID or InstantID, to further improve ID fidelity, or combining with LCM for acceleration. We look forward to your exploration of more features, and welcome you to **provide PRs** or **contribute to the open-source community**
🥳 If you have built or known repositories or applications around PhotoMaker V2, please leave us a message in the discussion. We will include them in our README.
## LICENSE
Since PhotoMaker V2 relies on [InsightFace](https://github.com/deepinsight/insightface), it also needs to comply with its [license](https://github.com/deepinsight/insightface?tab=readme-ov-file#license).
gradio_demo/app.py
View file @ 48ec26d1
...@@ -17,7 +17,7 @@ from style_template import styles ...@@ -17,7 +17,7 @@ from style_template import styles
from aspect_ratio_template import aspect_ratios from aspect_ratio_template import aspect_ratios
# global variable # global variable
base_model_path = 'SG161222/RealVisXL_V3.0' base_model_path = 'SG161222/RealVisXL_V4.0'
try: try:
if torch.cuda.is_available(): if torch.cuda.is_available():
device = "cuda" device = "cuda"
...@@ -37,10 +37,9 @@ DEFAULT_ASPECT_RATIO = ASPECT_RATIO_LABELS[0] ...@@ -37,10 +37,9 @@ DEFAULT_ASPECT_RATIO = ASPECT_RATIO_LABELS[0]
# download PhotoMaker checkpoint to cache # download PhotoMaker checkpoint to cache
photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker", filename="photomaker-v1.bin", repo_type="model") photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker", filename="photomaker-v1.bin", repo_type="model")
torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
if device == "mps": if device == "mps":
torch_dtype = torch.float16 torch_dtype = torch.float16
else:
torch_dtype = torch.bfloat16
pipe = PhotoMakerStableDiffusionXLPipeline.from_pretrained( pipe = PhotoMakerStableDiffusionXLPipeline.from_pretrained(
base_model_path, base_model_path,
...@@ -54,15 +53,17 @@ pipe.load_photomaker_adapter( ...@@ -54,15 +53,17 @@ pipe.load_photomaker_adapter(
os.path.dirname(photomaker_ckpt), os.path.dirname(photomaker_ckpt),
subfolder="", subfolder="",
weight_name=os.path.basename(photomaker_ckpt), weight_name=os.path.basename(photomaker_ckpt),
trigger_word="img" trigger_word="img",
pm_version="v1",
) )
pipe.id_encoder.to(device) pipe.id_encoder.to(device)
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config) pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
# pipe.set_adapters(["photomaker"], adapter_weights=[1.0]) # pipe.set_adapters(["photomaker"], adapter_weights=[1.0])
pipe.fuse_lora() pipe.fuse_lora()
pipe.to(device)
@spaces.GPU(enable_queue=True) @spaces.GPU
def generate_image(upload_images, prompt, negative_prompt, aspect_ratio_name, style_name, num_steps, style_strength_ratio, num_outputs, guidance_scale, seed, progress=gr.Progress(track_tqdm=True)): def generate_image(upload_images, prompt, negative_prompt, aspect_ratio_name, style_name, num_steps, style_strength_ratio, num_outputs, guidance_scale, seed, progress=gr.Progress(track_tqdm=True)):
# check the trigger word # check the trigger word
image_token_id = pipe.tokenizer.convert_tokens_to_ids(pipe.trigger_word) image_token_id = pipe.tokenizer.convert_tokens_to_ids(pipe.trigger_word)
...@@ -185,8 +186,8 @@ If our work is useful for your research, please consider citing: ...@@ -185,8 +186,8 @@ If our work is useful for your research, please consider citing:
@article{li2023photomaker, @article{li2023photomaker,
title={PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding}, title={PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding},
author={Li, Zhen and Cao, Mingdeng and Wang, Xintao and Qi, Zhongang and Cheng, Ming-Ming and Shan, Ying}, author={Li, Zhen and Cao, Mingdeng and Wang, Xintao and Qi, Zhongang and Cheng, Ming-Ming and Shan, Ying},
booktitle={arXiv preprint arxiv:2312.04461}, booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2023} year={2024}
} }
``` ```
📋 **License** 📋 **License**
...@@ -310,4 +311,4 @@ with gr.Blocks(css=css) as demo: ...@@ -310,4 +311,4 @@ with gr.Blocks(css=css) as demo:
gr.Markdown(article) gr.Markdown(article)
demo.launch(share=False) demo.launch()
gradio_demo/app_v2.py 0 → 100644
View file @ 48ec26d1
import torch
import torchvision.transforms.functional as TF
import numpy as np
import random
import os
import sys
from diffusers.utils import load_image
from diffusers import EulerDiscreteScheduler, T2IAdapter
from huggingface_hub import hf_hub_download
import spaces
import gradio as gr
from photomaker import PhotoMakerStableDiffusionXLAdapterPipeline
from photomaker import FaceAnalysis2, analyze_faces
from style_template import styles
from aspect_ratio_template import aspect_ratios
# global variable
base_model_path = 'SG161222/RealVisXL_V4.0'
face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
face_detector.prepare(ctx_id=0, det_size=(640, 640))
try:
if torch.cuda.is_available():
device = "cuda"
elif sys.platform == "darwin" and torch.backends.mps.is_available():
device = "mps"
else:
device = "cpu"
except:
device = "cpu"
MAX_SEED = np.iinfo(np.int32).max
STYLE_NAMES = list(styles.keys())
DEFAULT_STYLE_NAME = "Photographic (Default)"
ASPECT_RATIO_LABELS = list(aspect_ratios)
DEFAULT_ASPECT_RATIO = ASPECT_RATIO_LABELS[0]
enable_doodle_arg = False
photomaker_ckpt = hf_hub_download(repo_id="TencentARC/PhotoMaker", filename="photomaker-v2.bin", repo_type="model")
torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
if device == "mps":
torch_dtype = torch.float16
# load adapter
adapter = T2IAdapter.from_pretrained(
"TencentARC/t2i-adapter-sketch-sdxl-1.0", torch_dtype=torch_dtype, variant="fp16"
).to(device)
pipe = PhotoMakerStableDiffusionXLAdapterPipeline.from_pretrained(
base_model_path,
adapter=adapter,
torch_dtype=torch_dtype,
use_safetensors=True,
variant="fp16",
).to(device)
pipe.load_photomaker_adapter(
os.path.dirname(photomaker_ckpt),
subfolder="",
weight_name=os.path.basename(photomaker_ckpt),
trigger_word="img",
pm_version="v2",
)
pipe.id_encoder.to(device)
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
# pipe.set_adapters(["photomaker"], adapter_weights=[1.0])
pipe.fuse_lora()
pipe.to(device)
@spaces.GPU
def generate_image(
upload_images,
prompt,
negative_prompt,
aspect_ratio_name,
style_name,
num_steps,
style_strength_ratio,
num_outputs,
guidance_scale,
seed,
use_doodle,
sketch_image,
adapter_conditioning_scale,
adapter_conditioning_factor,
progress=gr.Progress(track_tqdm=True)
):
if use_doodle:
sketch_image = sketch_image["composite"]
r, g, b, a = sketch_image.split()
sketch_image = a.convert("RGB")
sketch_image = TF.to_tensor(sketch_image) > 0.5 # Inversion
sketch_image = TF.to_pil_image(sketch_image.to(torch.float32))
adapter_conditioning_scale = adapter_conditioning_scale
adapter_conditioning_factor = adapter_conditioning_factor
else:
adapter_conditioning_scale = 0.
adapter_conditioning_factor = 0.
sketch_image = None
# check the trigger word
image_token_id = pipe.tokenizer.convert_tokens_to_ids(pipe.trigger_word)
input_ids = pipe.tokenizer.encode(prompt)
if image_token_id not in input_ids:
raise gr.Error(f"Cannot find the trigger word '{pipe.trigger_word}' in text prompt! Please refer to step 2️⃣")
if input_ids.count(image_token_id) > 1:
raise gr.Error(f"Cannot use multiple trigger words '{pipe.trigger_word}' in text prompt!")
# determine output dimensions by the aspect ratio
output_w, output_h = aspect_ratios[aspect_ratio_name]
print(f"[Debug] Generate image using aspect ratio [{aspect_ratio_name}] => {output_w} x {output_h}")
# apply the style template
prompt, negative_prompt = apply_style(style_name, prompt, negative_prompt)
if upload_images is None:
raise gr.Error(f"Cannot find any input face image! Please refer to step 1️⃣")
input_id_images = []
for img in upload_images:
input_id_images.append(load_image(img))
id_embed_list = []
for img in input_id_images:
img = np.array(img)
img = img[:, :, ::-1]
faces = analyze_faces(face_detector, img)
if len(faces) > 0:
id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
if len(id_embed_list) == 0:
raise gr.Error(f"No face detected, please update the input face image(s)")
id_embeds = torch.stack(id_embed_list)
generator = torch.Generator(device=device).manual_seed(seed)
print("Start inference...")
print(f"[Debug] Seed: {seed}")
print(f"[Debug] Prompt: {prompt}, \n[Debug] Neg Prompt: {negative_prompt}")
start_merge_step = int(float(style_strength_ratio) / 100 * num_steps)
if start_merge_step > 30:
start_merge_step = 30
print(start_merge_step)
images = pipe(
prompt=prompt,
width=output_w,
height=output_h,
input_id_images=input_id_images,
negative_prompt=negative_prompt,
num_images_per_prompt=num_outputs,
num_inference_steps=num_steps,
start_merge_step=start_merge_step,
generator=generator,
guidance_scale=guidance_scale,
id_embeds=id_embeds,
image=sketch_image,
adapter_conditioning_scale=adapter_conditioning_scale,
adapter_conditioning_factor=adapter_conditioning_factor,
).images
return images, gr.update(visible=True)
def swap_to_gallery(images):
return gr.update(value=images, visible=True), gr.update(visible=True), gr.update(visible=False)
def upload_example_to_gallery(images, prompt, style, negative_prompt):
return gr.update(value=images, visible=True), gr.update(visible=True), gr.update(visible=False)
def remove_back_to_files():
return gr.update(visible=False), gr.update(visible=False), gr.update(visible=True)
def change_doodle_space(use_doodle):
if use_doodle:
return gr.update(visible=True)
else:
return gr.update(visible=False)
def remove_tips():
return gr.update(visible=False)
def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
if randomize_seed:
seed = random.randint(0, MAX_SEED)
return seed
def apply_style(style_name: str, positive: str, negative: str = "") -> tuple[str, str]:
p, n = styles.get(style_name, styles[DEFAULT_STYLE_NAME])
return p.replace("{prompt}", positive), n + ' ' + negative
def get_image_path_list(folder_name):
image_basename_list = os.listdir(folder_name)
image_path_list = sorted([os.path.join(folder_name, basename) for basename in image_basename_list])
return image_path_list
def get_example():
case = [
[
get_image_path_list('./examples/scarletthead_woman'),
"instagram photo, portrait photo of a woman img, colorful, perfect face, natural skin, hard shadows, film grain",
"(No style)",
"(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth",
],
[
get_image_path_list('./examples/newton_man'),
"sci-fi, closeup portrait photo of a man img wearing the sunglasses in Iron man suit, face, slim body, high quality, film grain",
"(No style)",
"(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth",
],
]
return case
### Description and style
logo = r"""
<center><img src='https://photo-maker.github.io/assets/logo.png' alt='PhotoMaker logo' style="width:80px; margin-bottom:10px"></center>
"""
title = r"""
<h1 align="center">PhotoMaker V2: Improved ID Fidelity and Better Controllability than PhotoMaker V1</h1>
"""
description = r"""
<b>Official 🤗 Gradio demo</b> for <a href='https://github.com/TencentARC/PhotoMaker' target='_blank'><b>PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding</b></a>.<br>
The details of PhotoMaker V2 can be found in
<br>
<br>
For previous version of PhotoMaker, you could use our original gradio demos [PhotoMaker](https://huggingface.co/spaces/TencentARC/PhotoMaker) and [PhotoMaker-Style](https://huggingface.co/spaces/TencentARC/PhotoMaker-Style).
<br>
❗️❗️❗️[<b>Important</b>] Personalization steps:<br>
1️⃣ Upload images of someone you want to customize. One image is ok, but more is better. Although we do not perform face detection, the face in the uploaded image should <b>occupy the majority of the image</b>.<br>
2️⃣ Enter a text prompt, making sure to <b>follow the class word</b> you want to customize with the <b>trigger word</b>: `img`, such as: `man img` or `woman img` or `girl img`.<br>
3️⃣ Choose your preferred style template.<br>
4️⃣ <b>(Optional: but new feature)</b> Select the ‘Enable Drawing Doodle...’ option and draw on the canvas<br>
5️⃣ Click the <b>Submit</b> button to start customizing.
"""
article = r"""
If PhotoMaker V2 is helpful, please help to ⭐ the <a href='https://github.com/TencentARC/PhotoMaker' target='_blank'>Github Repo</a>. Thanks!
[![GitHub Stars](https://img.shields.io/github/stars/TencentARC/PhotoMaker?style=social)](https://github.com/TencentARC/PhotoMaker)
---
📝 **Citation**
<br>
If our work is useful for your research, please consider citing:
```bibtex
@article{li2023photomaker,
title={PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding},
author={Li, Zhen and Cao, Mingdeng and Wang, Xintao and Qi, Zhongang and Cheng, Ming-Ming and Shan, Ying},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year={2024}
}
```
📋 **License**
<br>
Apache-2.0 LICENSE. Please refer to the [LICENSE file](https://huggingface.co/TencentARC/PhotoMaker/blob/main/LICENSE) for details.
📧 **Contact**
<br>
If you have any questions, please feel free to reach me out at <b>zhenli1031@gmail.com</b>.
"""
tips = r"""
### Usage tips of PhotoMaker
1. Upload **more photos**of the person to be customized to **improve ID fidelty**.
2. If you find that the image quality is poor when using doodle for control, you can reduce the conditioning scale and factor of the adapter.
Besides, you could enlarge the ratio for more consistency of your doodle. <br>
If you have any issues, leave the issue in the discussion page of the space. For a more stable (queue-free) experience, you can duplicate the space.
"""
# We have provided some generate examples and comparisons at: [this website]().
css = '''
.gradio-container {width: 85% !important}
'''
with gr.Blocks(css=css) as demo:
gr.Markdown(logo)
gr.Markdown(title)
gr.Markdown(description)
# gr.DuplicateButton(
# value="Duplicate Space for private use ",
# elem_id="duplicate-button",
# visible=os.getenv("SHOW_DUPLICATE_BUTTON") == "1",
# )
with gr.Row():
with gr.Column():
files = gr.Files(
label="Drag (Select) 1 or more photos of your face",
file_types=["image"]
)
uploaded_files = gr.Gallery(label="Your images", visible=False, columns=5, rows=1, height=200)
with gr.Column(visible=False) as clear_button:
remove_and_reupload = gr.ClearButton(value="Remove and upload new ones", components=files, size="sm")
prompt = gr.Textbox(label="Prompt",
info="Try something like 'a photo of a man/woman img', 'img' is the trigger word.",
placeholder="A photo of a [man/woman img]...")
style = gr.Dropdown(label="Style template", choices=STYLE_NAMES, value=DEFAULT_STYLE_NAME)
aspect_ratio = gr.Dropdown(label="Output aspect ratio", choices=ASPECT_RATIO_LABELS, value=DEFAULT_ASPECT_RATIO)
submit = gr.Button("Submit")
enable_doodle = gr.Checkbox(
label="Enable Drawing Doodle for Control", value=enable_doodle_arg,
info="After enabling this option, PhotoMaker will generate content based on your doodle on the canvas, driven by the T2I-Adapter (Quality may be decreased)",
)
with gr.Accordion("T2I-Adapter-Doodle (Optional)", visible=False) as doodle_space:
with gr.Row():
sketch_image = gr.Sketchpad(
label="Canvas",
type="pil",
crop_size=[1024,1024],
layers=False,
canvas_size=(350, 350),
brush=gr.Brush(default_size=5, colors=["#000000"], color_mode="fixed")
)
with gr.Group():
adapter_conditioning_scale = gr.Slider(
label="Adapter conditioning scale",
minimum=0.5,
maximum=1,
step=0.1,
value=0.7,
)
adapter_conditioning_factor = gr.Slider(
label="Adapter conditioning factor",
info="Fraction of timesteps for which adapter should be applied",
minimum=0.5,
maximum=1,
step=0.1,
value=0.8,
)
with gr.Accordion(open=False, label="Advanced Options"):
negative_prompt = gr.Textbox(
label="Negative Prompt",
placeholder="low quality",
value="nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry",
)
num_steps = gr.Slider(
label="Number of sample steps",
minimum=20,
maximum=100,
step=1,
value=50,
)
style_strength_ratio = gr.Slider(
label="Style strength (%)",
minimum=15,
maximum=50,
step=1,
value=20,
)
num_outputs = gr.Slider(
label="Number of output images",
minimum=1,
maximum=4,
step=1,
value=2,
)
guidance_scale = gr.Slider(
label="Guidance scale",
minimum=0.1,
maximum=10.0,
step=0.1,
value=5,
)
seed = gr.Slider(
label="Seed",
minimum=0,
maximum=MAX_SEED,
step=1,
value=0,
)
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
with gr.Column():
gallery = gr.Gallery(label="Generated Images")
usage_tips = gr.Markdown(label="Usage tips of PhotoMaker", value=tips ,visible=False)
files.upload(fn=swap_to_gallery, inputs=files, outputs=[uploaded_files, clear_button, files])
remove_and_reupload.click(fn=remove_back_to_files, outputs=[uploaded_files, clear_button, files])
enable_doodle.select(fn=change_doodle_space, inputs=enable_doodle, outputs=doodle_space)
input_list = [
files,
prompt,
negative_prompt,
aspect_ratio,
style,
num_steps,
style_strength_ratio,
num_outputs,
guidance_scale,
seed,
enable_doodle,
sketch_image,
adapter_conditioning_scale,
adapter_conditioning_factor
]
submit.click(
fn=remove_tips,
outputs=usage_tips,
).then(
fn=randomize_seed_fn,
inputs=[seed, randomize_seed],
outputs=seed,
queue=False,
api_name=False,
).then(
fn=generate_image,
inputs=input_list,
outputs=[gallery, usage_tips]
)
gr.Examples(
examples=get_example(),
inputs=[files, prompt, style, negative_prompt],
run_on_click=True,
fn=upload_example_to_gallery,
outputs=[uploaded_files, clear_button, files],
)
gr.Markdown(article)
demo.launch()
inference_scripts/inference_pmv2.py 0 → 100644
View file @ 48ec26d1
# !pip install opencv-python transformers accelerate
import os
import sys
import numpy as np
import torch
from diffusers.utils import load_image
from diffusers import EulerDiscreteScheduler
from huggingface_hub import hf_hub_download
from photomaker import PhotoMakerStableDiffusionXLPipeline
from photomaker import FaceAnalysis2, analyze_faces
face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
face_detector.prepare(ctx_id=0, det_size=(640, 640))
try:
if torch.cuda.is_available():
device = "cuda"
elif sys.platform == "darwin" and torch.backends.mps.is_available():
device = "mps"
else:
device = "cpu"
except:
device = "cpu"
torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
if device == "mps":
torch_dtype = torch.float16
output_dir = "./outputs"
os.makedirs(output_dir, exist_ok=True)
photomaker_path = hf_hub_download(repo_id="TencentARC/PhotoMaker", filename="photomaker-v2.bin", repo_type="model")
prompt = "instagram photo, portrait photo of a woman img, colorful, perfect face, natural skin, hard shadows, film grain, best quality"
negative_prompt = "(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth"
# initialize the models and pipeline
### Load base model
pipe = PhotoMakerStableDiffusionXLPipeline.from_pretrained(
"SG161222/RealVisXL_V4.0", torch_dtype=torch_dtype
).to("cuda")
### Load PhotoMaker checkpoint
pipe.load_photomaker_adapter(
os.path.dirname(photomaker_path),
subfolder="",
weight_name=os.path.basename(photomaker_path),
trigger_word="img" # define the trigger word
)
### Also can cooperate with other LoRA modules
# pipe.load_lora_weights(os.path.dirname(lora_path), weight_name=lora_model_name, adapter_name="lcm-lora")
# pipe.set_adapters(["photomaker", "lcm-lora"], adapter_weights=[1.0, 0.5])
pipe.fuse_lora()
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
pipe.enable_model_cpu_offload()
### define the input ID images
input_folder_name = './examples/scarletthead_woman'
image_basename_list = os.listdir(input_folder_name)
image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
input_id_images = []
for image_path in image_path_list:
input_id_images.append(load_image(image_path))
id_embed_list = []
for img in input_id_images:
img = np.array(img)
img = img[:, :, ::-1]
faces = analyze_faces(face_detector, img)
if len(faces) > 0:
id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
if len(id_embed_list) == 0:
raise ValueError(f"No face detected in input image pool")
id_embeds = torch.stack(id_embed_list)
# generate image
images = pipe(
prompt,
negative_prompt=negative_prompt,
input_id_images=input_id_images,
id_embeds=id_embeds,
num_images_per_prompt=2,
start_merge_step=10,
).images
for idx, img in enumerate(images):
img.save(os.path.join(output_dir, f"output_pmv2_{idx}.jpg"))
\ No newline at end of file
inference_scripts/inference_pmv2_contronet.py 0 → 100644
View file @ 48ec26d1
# !pip install opencv-python transformers accelerate
import os
import sys
import numpy as np
import torch
from diffusers.utils import load_image
from diffusers import EulerDiscreteScheduler, ControlNetModel
from huggingface_hub import hf_hub_download
from controlnet_aux import OpenposeDetector
from photomaker import PhotoMakerStableDiffusionXLControlNetPipeline
from photomaker import FaceAnalysis2, analyze_faces
face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
face_detector.prepare(ctx_id=0, det_size=(640, 640))
try:
if torch.cuda.is_available():
device = "cuda"
elif sys.platform == "darwin" and torch.backends.mps.is_available():
device = "mps"
else:
device = "cpu"
except:
device = "cpu"
torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
if device == "mps":
torch_dtype = torch.float16
output_dir = "./outputs"
os.makedirs(output_dir, exist_ok=True)
photomaker_path = hf_hub_download(repo_id="TencentARC/PhotoMaker", filename="photomaker-v2.bin", repo_type="model")
openpose = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
controlnet_pose_model = "thibaud/controlnet-openpose-sdxl-1.0"
controlnet_pose = ControlNetModel.from_pretrained(
controlnet_pose_model, torch_dtype=torch_dtype,
).to("cuda")
prompt = "instagram photo, a photo of a woman img, colorful, perfect face, natural skin, hard shadows, film grain, best quality"
negative_prompt = "(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth"
# download an image
pose_image = load_image(
"./examples/pos_ref.png"
)
pose_image = openpose(pose_image, detect_resolution=512, image_resolution=1024)
# initialize the models and pipeline
controlnet_conditioning_scale = 1.0 # recommended for good generalization
### Load base model
pipe = PhotoMakerStableDiffusionXLControlNetPipeline.from_pretrained(
"SG161222/RealVisXL_V4.0",
controlnet=controlnet_pose,
torch_dtype=torch_dtype,
).to("cuda")
### Load PhotoMaker checkpoint
pipe.load_photomaker_adapter(
os.path.dirname(photomaker_path),
subfolder="",
weight_name=os.path.basename(photomaker_path),
trigger_word="img" # define the trigger word
)
### Also can cooperate with other LoRA modules
# pipe.load_lora_weights(os.path.dirname(lora_path), weight_name=lora_model_name, adapter_name="lcm-lora")
# pipe.set_adapters(["photomaker", "lcm-lora"], adapter_weights=[1.0, 0.5])
pipe.fuse_lora()
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
pipe.enable_model_cpu_offload()
### define the input ID images
input_folder_name = './examples/scarletthead_woman'
image_basename_list = os.listdir(input_folder_name)
image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
input_id_images = []
for image_path in image_path_list:
input_id_images.append(load_image(image_path))
### extract insightface embedding
input_id_images = []
for image_path in image_path_list:
input_id_images.append(load_image(image_path))
id_embed_list = []
for img in input_id_images:
img = np.array(img)
img = img[:, :, ::-1]
faces = analyze_faces(face_detector, img)
if len(faces) > 0:
id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
if len(id_embed_list) == 0:
raise ValueError(f"No face detected in input image pool")
id_embeds = torch.stack(id_embed_list)
# generate image
images = pipe(
prompt,
negative_prompt=negative_prompt,
input_id_images=input_id_images,
id_embeds=id_embeds,
controlnet_conditioning_scale=controlnet_conditioning_scale,
image=pose_image,
num_images_per_prompt=2,
start_merge_step=10,
).images
for idx, img in enumerate(images):
img.save(os.path.join(output_dir, f"output_pmv2_cn_{idx}.jpg"))
\ No newline at end of file
inference_scripts/inference_pmv2_ip_adapter.py 0 → 100644
View file @ 48ec26d1
# !pip install opencv-python transformers accelerate
import os
import sys
import numpy as np
import torch
from diffusers.utils import load_image
from diffusers import EulerDiscreteScheduler
from huggingface_hub import hf_hub_download
from photomaker import PhotoMakerStableDiffusionXLPipeline
from photomaker import FaceAnalysis2, analyze_faces
face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
face_detector.prepare(ctx_id=0, det_size=(640, 640))
try:
if torch.cuda.is_available():
device = "cuda"
elif sys.platform == "darwin" and torch.backends.mps.is_available():
device = "mps"
else:
device = "cpu"
except:
device = "cpu"
torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
if device == "mps":
torch_dtype = torch.float16
output_dir = "./outputs"
os.makedirs(output_dir, exist_ok=True)
photomaker_path = hf_hub_download(repo_id="TencentARC/PhotoMaker", filename="photomaker-v2.bin", repo_type="model")
prompt = "portrait photo of a woman img, colorful, perfect face, best quality"
negative_prompt = "(asymmetry, worst quality, low quality, illustration), open mouth"
# # initialize the models and pipeline
### Load base model
pipe = PhotoMakerStableDiffusionXLPipeline.from_pretrained(
"SG161222/RealVisXL_V4.0", torch_dtype=torch_dtype,
).to("cuda")
pipe.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models", weight_name="ip-adapter_sdxl.bin")
pipe.set_ip_adapter_scale(0.7)
print("Loading images...")
style_images = [load_image(f"./examples/statue.png")]
### Load PhotoMaker checkpoint
pipe.load_photomaker_adapter(
os.path.dirname(photomaker_path),
subfolder="",
weight_name=os.path.basename(photomaker_path),
trigger_word="img" # define the trigger word
)
### Also can cooperate with other LoRA modules
# pipe.load_lora_weights(os.path.dirname(lora_path), weight_name=lora_model_name, adapter_name="lcm-lora")
# pipe.set_adapters(["photomaker", "lcm-lora"], adapter_weights=[1.0, 0.5])
pipe.fuse_lora()
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
pipe.enable_model_cpu_offload()
### define the input ID images
input_folder_name = './examples/scarletthead_woman'
image_basename_list = os.listdir(input_folder_name)
image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
input_id_images = []
for image_path in image_path_list:
input_id_images.append(load_image(image_path))
id_embed_list = []
for img in input_id_images:
img = np.array(img)
img = img[:, :, ::-1]
faces = analyze_faces(face_detector, img)
if len(faces) > 0:
id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
if len(id_embed_list) == 0:
raise ValueError(f"No face detected in input image pool")
id_embeds = torch.stack(id_embed_list)
# generate image
images = pipe(
prompt,
negative_prompt=negative_prompt,
input_id_images=input_id_images,
id_embeds=id_embeds,
ip_adapter_image=[style_images],
num_images_per_prompt=2,
start_merge_step=10,
).images
for idx, img in enumerate(images):
img.save(os.path.join(output_dir, f"output_pmv2_ipa_{idx}.jpg"))
\ No newline at end of file
inference_scripts/inference_pmv2_t2i_adapter.py 0 → 100644
View file @ 48ec26d1
# !pip install opencv-python transformers accelerate
import os
import sys
import numpy as np
import torch
from diffusers.utils import load_image
from diffusers import EulerDiscreteScheduler, T2IAdapter
from huggingface_hub import hf_hub_download
from controlnet_aux import OpenposeDetector
from photomaker import PhotoMakerStableDiffusionXLAdapterPipeline
from photomaker import FaceAnalysis2, analyze_faces
face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
face_detector.prepare(ctx_id=0, det_size=(640, 640))
try:
if torch.cuda.is_available():
device = "cuda"
elif sys.platform == "darwin" and torch.backends.mps.is_available():
device = "mps"
else:
device = "cpu"
except:
device = "cpu"
torch_dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
if device == "mps":
torch_dtype = torch.float16
output_dir = "./outputs"
os.makedirs(output_dir, exist_ok=True)
photomaker_path = hf_hub_download(repo_id="TencentARC/PhotoMaker", filename="photomaker-v2.bin", repo_type="model")
openpose = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
# load adapter
adapter = T2IAdapter.from_pretrained(
"TencentARC/t2i-adapter-openpose-sdxl-1.0", torch_dtype=torch_dtype,
).to("cuda")
prompt = "instagram photo, a photo of a woman img, colorful, perfect face, natural skin, hard shadows, film grain, best quality"
negative_prompt = "(asymmetry, worst quality, low quality, illustration, 3d, 2d, painting, cartoons, sketch), open mouth"
# download an image
pose_image = load_image(
"./examples/pos_ref.png"
)
pose_image = openpose(pose_image, detect_resolution=512, image_resolution=1024)
# initialize the models and pipeline
adapter_conditioning_scale = 0.8 # recommended for good generalization
adapter_conditioning_factor = 0.8
### Load base model
pipe = PhotoMakerStableDiffusionXLAdapterPipeline.from_pretrained(
"SG161222/RealVisXL_V4.0",
adapter=adapter,
torch_dtype=torch_dtype,
).to("cuda")
### Load PhotoMaker checkpoint
pipe.load_photomaker_adapter(
os.path.dirname(photomaker_path),
subfolder="",
weight_name=os.path.basename(photomaker_path),
trigger_word="img" # define the trigger word
)
### Also can cooperate with other LoRA modules
# pipe.load_lora_weights(os.path.dirname(lora_path), weight_name=lora_model_name, adapter_name="lcm-lora")
# pipe.set_adapters(["photomaker", "lcm-lora"], adapter_weights=[1.0, 0.5])
pipe.fuse_lora()
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
pipe.enable_model_cpu_offload()
### define the input ID images
input_folder_name = './examples/scarletthead_woman'
image_basename_list = os.listdir(input_folder_name)
image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
input_id_images = []
for image_path in image_path_list:
input_id_images.append(load_image(image_path))
id_embed_list = []
for img in input_id_images:
img = np.array(img)
img = img[:, :, ::-1]
faces = analyze_faces(face_detector, img)
if len(faces) > 0:
id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
if len(id_embed_list) == 0:
raise ValueError(f"No face detected in input image pool")
id_embeds = torch.stack(id_embed_list)
# generate image
images = pipe(
prompt,
negative_prompt=negative_prompt,
input_id_images=input_id_images,
id_embeds=id_embeds,
adapter_conditioning_scale=adapter_conditioning_scale,
image=pose_image,
num_images_per_prompt=2,
start_merge_step=10,
).images
for idx, img in enumerate(images):
img.save(os.path.join(output_dir, f"output_pmv2_t2ia_{idx}.jpg"))
\ No newline at end of file
photomaker/__init__.py
View file @ 48ec26d1
from .model import PhotoMakerIDEncoder from .model import PhotoMakerIDEncoder
from .model_v2 import PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken
from .resampler import FacePerceiverResampler
from .pipeline import PhotoMakerStableDiffusionXLPipeline from .pipeline import PhotoMakerStableDiffusionXLPipeline
from .pipeline import PhotoMakerStableDiffusionXLPipeline
from .pipeline_controlnet import PhotoMakerStableDiffusionXLControlNetPipeline
from .pipeline_t2i_adapter import PhotoMakerStableDiffusionXLAdapterPipeline
# InsightFace Package
from .insightface_package import FaceAnalysis2, analyze_faces
__all__ = [ __all__ = [
"FaceAnalysis2",
"analyze_faces",
"FacePerceiverResampler",
"PhotoMakerIDEncoder", "PhotoMakerIDEncoder",
"PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken",
"PhotoMakerStableDiffusionXLPipeline", "PhotoMakerStableDiffusionXLPipeline",
"PhotoMakerStableDiffusionXLControlNetPipeline",
"PhotoMakerStableDiffusionXLAdapterPipeline",
] ]
\ No newline at end of file
photomaker/insightface_package.py 0 → 100755
View file @ 48ec26d1
import numpy as np
# pip install insightface==0.7.3
from insightface.app import FaceAnalysis
from insightface.data import get_image as ins_get_image
###
# https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/165#issue-2055829543
###
class FaceAnalysis2(FaceAnalysis):
# NOTE: allows setting det_size for each detection call.
# the model allows it but the wrapping code from insightface
# doesn't show it, and people end up loading duplicate models
# for different sizes where there is absolutely no need to
def get(self, img, max_num=0, det_size=(640, 640)):
if det_size is not None:
self.det_model.input_size = det_size
return super().get(img, max_num)
def analyze_faces(face_analysis: FaceAnalysis, img_data: np.ndarray, det_size=(640, 640)):
# NOTE: try detect faces, if no faces detected, lower det_size until it does
detection_sizes = [None] + [(size, size) for size in range(640, 256, -64)] + [(256, 256)]
for size in detection_sizes:
faces = face_analysis.get(img_data, det_size=size)
if len(faces) > 0:
return faces
return []
photomaker/model_v2.py 0 → 100755
View file @ 48ec26d1
# Merge image encoder and fuse module to create an ID Encoder
# send multiple ID images, we can directly obtain the updated text encoder containing a stacked ID embedding
import torch
import torch.nn as nn
from transformers.models.clip.modeling_clip import CLIPVisionModelWithProjection
from transformers.models.clip.configuration_clip import CLIPVisionConfig
from . import FacePerceiverResampler
VISION_CONFIG_DICT = {
"hidden_size": 1024,
"intermediate_size": 4096,
"num_attention_heads": 16,
"num_hidden_layers": 24,
"patch_size": 14,
"projection_dim": 768
}
class MLP(nn.Module):
def __init__(self, in_dim, out_dim, hidden_dim, use_residual=True):
super().__init__()
if use_residual:
assert in_dim == out_dim
self.layernorm = nn.LayerNorm(in_dim)
self.fc1 = nn.Linear(in_dim, hidden_dim)
self.fc2 = nn.Linear(hidden_dim, out_dim)
self.use_residual = use_residual
self.act_fn = nn.GELU()
def forward(self, x):
residual = x
x = self.layernorm(x)
x = self.fc1(x)
x = self.act_fn(x)
x = self.fc2(x)
if self.use_residual:
x = x + residual
return x
class QFormerPerceiver(nn.Module):
def __init__(self, id_embeddings_dim, cross_attention_dim, num_tokens, embedding_dim=1024, use_residual=True, ratio=4):
super().__init__()
self.num_tokens = num_tokens
self.cross_attention_dim = cross_attention_dim
self.use_residual = use_residual
print(cross_attention_dim*num_tokens)
self.token_proj = nn.Sequential(
nn.Linear(id_embeddings_dim, id_embeddings_dim*ratio),
nn.GELU(),
nn.Linear(id_embeddings_dim*ratio, cross_attention_dim*num_tokens),
)
self.token_norm = nn.LayerNorm(cross_attention_dim)
self.perceiver_resampler = FacePerceiverResampler(
dim=cross_attention_dim,
depth=4,
dim_head=128,
heads=cross_attention_dim // 128,
embedding_dim=embedding_dim,
output_dim=cross_attention_dim,
ff_mult=4,
)
def forward(self, x, last_hidden_state):
x = self.token_proj(x)
x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
x = self.token_norm(x) # cls token
out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
if self.use_residual: # TODO: if use_residual is not true
out = x + 1.0 * out
return out
class FuseModule(nn.Module):
def __init__(self, embed_dim):
super().__init__()
self.mlp1 = MLP(embed_dim * 2, embed_dim, embed_dim, use_residual=False)
self.mlp2 = MLP(embed_dim, embed_dim, embed_dim, use_residual=True)
self.layer_norm = nn.LayerNorm(embed_dim)
def fuse_fn(self, prompt_embeds, id_embeds):
stacked_id_embeds = torch.cat([prompt_embeds, id_embeds], dim=-1)
stacked_id_embeds = self.mlp1(stacked_id_embeds) + prompt_embeds
stacked_id_embeds = self.mlp2(stacked_id_embeds)
stacked_id_embeds = self.layer_norm(stacked_id_embeds)
return stacked_id_embeds
def forward(
self,
prompt_embeds,
id_embeds,
class_tokens_mask,
) -> torch.Tensor:
# id_embeds shape: [b, max_num_inputs, 1, 2048]
id_embeds = id_embeds.to(prompt_embeds.dtype)
num_inputs = class_tokens_mask.sum().unsqueeze(0) # TODO: check for training case
batch_size, max_num_inputs = id_embeds.shape[:2]
# seq_length: 77
seq_length = prompt_embeds.shape[1]
# flat_id_embeds shape: [b*max_num_inputs, 1, 2048]
flat_id_embeds = id_embeds.view(
-1, id_embeds.shape[-2], id_embeds.shape[-1]
)
# valid_id_mask [b*max_num_inputs]
valid_id_mask = (
torch.arange(max_num_inputs, device=flat_id_embeds.device)[None, :]
< num_inputs[:, None]
)
valid_id_embeds = flat_id_embeds[valid_id_mask.flatten()]
prompt_embeds = prompt_embeds.view(-1, prompt_embeds.shape[-1])
class_tokens_mask = class_tokens_mask.view(-1)
valid_id_embeds = valid_id_embeds.view(-1, valid_id_embeds.shape[-1])
# slice out the image token embeddings
image_token_embeds = prompt_embeds[class_tokens_mask]
stacked_id_embeds = self.fuse_fn(image_token_embeds, valid_id_embeds)
assert class_tokens_mask.sum() == stacked_id_embeds.shape[0], f"{class_tokens_mask.sum()} != {stacked_id_embeds.shape[0]}"
prompt_embeds.masked_scatter_(class_tokens_mask[:, None], stacked_id_embeds.to(prompt_embeds.dtype))
updated_prompt_embeds = prompt_embeds.view(batch_size, seq_length, -1)
return updated_prompt_embeds
class PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken(CLIPVisionModelWithProjection):
def __init__(self, id_embeddings_dim=512):
super().__init__(CLIPVisionConfig(**VISION_CONFIG_DICT))
self.fuse_module = FuseModule(2048)
self.visual_projection_2 = nn.Linear(1024, 1280, bias=False)
cross_attention_dim = 2048
# projection
self.num_tokens = 2
self.cross_attention_dim = cross_attention_dim
self.qformer_perceiver = QFormerPerceiver(
id_embeddings_dim,
cross_attention_dim,
self.num_tokens,
)
def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds):
b, num_inputs, c, h, w = id_pixel_values.shape
id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w)
last_hidden_state = self.vision_model(id_pixel_values)[0]
id_embeds = id_embeds.view(b * num_inputs, -1)
id_embeds = self.qformer_perceiver(id_embeds, last_hidden_state)
id_embeds = id_embeds.view(b, num_inputs, self.num_tokens, -1)
updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask)
return updated_prompt_embeds
if __name__ == "__main__":
PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken()
\ No newline at end of file
photomaker/pipeline.py
View file @ 48ec26d1
from typing import Any, Callable, Dict, List, Optional, Union, Tuple #####
from collections import OrderedDict # Modified from https://github.com/huggingface/diffusers/blob/v0.29.1/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py
import os # PhotoMaker v2 @ TencentARC and MCG-NKU
# Author: Zhen Li
#####
# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import inspect
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import PIL import PIL
import numpy as np
import torch import torch
from torchvision import transforms as T from transformers import CLIPImageProcessor
from safetensors import safe_open from safetensors import safe_open
from huggingface_hub.utils import validate_hf_hub_args from huggingface_hub.utils import validate_hf_hub_args
from transformers import CLIPImageProcessor, CLIPTokenizer
from diffusers import StableDiffusionXLPipeline from diffusers import StableDiffusionXLPipeline
from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
from diffusers.loaders import (
StableDiffusionXLLoraLoaderMixin,
TextualInversionLoaderMixin,
)
from diffusers.callbacks import (
MultiPipelineCallbacks,
PipelineCallback,
)
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.utils import ( from diffusers.utils import (
_get_model_file, _get_model_file,
is_transformers_available, USE_PEFT_BACKEND,
logging, deprecate,
is_torch_xla_available,
scale_lora_layers,
unscale_lora_layers,
) )
from . import PhotoMakerIDEncoder if is_torch_xla_available():
import torch_xla.core.xla_model as xm
XLA_AVAILABLE = True
else:
XLA_AVAILABLE = False
from . import (
PhotoMakerIDEncoder, # PhotoMaker v1
PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken, # PhotoMaker v2
)
PipelineImageInput = Union[ PipelineImageInput = Union[
PIL.Image.Image, PIL.Image.Image,
...@@ -28,6 +68,81 @@ PipelineImageInput = Union[ ...@@ -28,6 +68,81 @@ PipelineImageInput = Union[
] ]
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
"""
Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
"""
std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
# rescale the results from guidance (fixes overexposure)
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
# mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
return noise_cfg
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
scheduler,
num_inference_steps: Optional[int] = None,
device: Optional[Union[str, torch.device]] = None,
timesteps: Optional[List[int]] = None,
sigmas: Optional[List[float]] = None,
**kwargs,
):
"""
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
Args:
scheduler (`SchedulerMixin`):
The scheduler to get timesteps from.
num_inference_steps (`int`):
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
must be `None`.
device (`str` or `torch.device`, *optional*):
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
timesteps (`List[int]`, *optional*):
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
`num_inference_steps` and `sigmas` must be `None`.
sigmas (`List[float]`, *optional*):
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
`num_inference_steps` and `timesteps` must be `None`.
Returns:
`Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
second element is the number of inference steps.
"""
if timesteps is not None and sigmas is not None:
raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
if timesteps is not None:
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
if not accepts_timesteps:
raise ValueError(
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
f" timestep schedules. Please check whether you are using the correct scheduler."
)
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
timesteps = scheduler.timesteps
num_inference_steps = len(timesteps)
elif sigmas is not None:
accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
if not accept_sigmas:
raise ValueError(
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
f" sigmas schedules. Please check whether you are using the correct scheduler."
)
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
timesteps = scheduler.timesteps
num_inference_steps = len(timesteps)
else:
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
timesteps = scheduler.timesteps
return timesteps, num_inference_steps
class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
@validate_hf_hub_args @validate_hf_hub_args
def load_photomaker_adapter( def load_photomaker_adapter(
...@@ -36,6 +151,7 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -36,6 +151,7 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
weight_name: str, weight_name: str,
subfolder: str = '', subfolder: str = '',
trigger_word: str = 'img', trigger_word: str = 'img',
pm_version: str = 'v2',
**kwargs, **kwargs,
): ):
""" """
...@@ -107,17 +223,25 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -107,17 +223,25 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
if keys != ["id_encoder", "lora_weights"]: if keys != ["id_encoder", "lora_weights"]:
raise ValueError("Required keys are (`id_encoder` and `lora_weights`) missing from the state dict.") raise ValueError("Required keys are (`id_encoder` and `lora_weights`) missing from the state dict.")
self.num_tokens =2
self.pm_version = pm_version
self.trigger_word = trigger_word self.trigger_word = trigger_word
# load finetuned CLIP image encoder and fuse module here if it has not been registered to the pipeline yet # load finetuned CLIP image encoder and fuse module here if it has not been registered to the pipeline yet
print(f"Loading PhotoMaker components [1] id_encoder from [{pretrained_model_name_or_path_or_dict}]...") print(f"Loading PhotoMaker {pm_version} components [1] id_encoder from [{pretrained_model_name_or_path_or_dict}]...")
id_encoder = PhotoMakerIDEncoder() self.id_image_processor = CLIPImageProcessor()
if pm_version == "v1": # PhotoMaker v1
id_encoder = PhotoMakerIDEncoder()
elif pm_version == "v2": # PhotoMaker v2
id_encoder = PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken()
else:
raise NotImplementedError(f"The PhotoMaker version [{pm_version}] does not support")
id_encoder.load_state_dict(state_dict["id_encoder"], strict=True) id_encoder.load_state_dict(state_dict["id_encoder"], strict=True)
id_encoder = id_encoder.to(self.device, dtype=self.unet.dtype) id_encoder = id_encoder.to(self.device, dtype=self.unet.dtype)
self.id_encoder = id_encoder self.id_encoder = id_encoder
self.id_image_processor = CLIPImageProcessor()
# load lora into models # load lora into models
print(f"Loading PhotoMaker components [2] lora_weights from [{pretrained_model_name_or_path_or_dict}]") print(f"Loading PhotoMaker {pm_version} components [2] lora_weights from [{pretrained_model_name_or_path_or_dict}]")
self.load_lora_weights(state_dict["lora_weights"], adapter_name="photomaker") self.load_lora_weights(state_dict["lora_weights"], adapter_name="photomaker")
# Add trigger word token # Add trigger word token
...@@ -131,17 +255,44 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -131,17 +255,44 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
self, self,
prompt: str, prompt: str,
prompt_2: Optional[str] = None, prompt_2: Optional[str] = None,
num_id_images: int = 1,
device: Optional[torch.device] = None, device: Optional[torch.device] = None,
prompt_embeds: Optional[torch.FloatTensor] = None, num_images_per_prompt: int = 1,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None, do_classifier_free_guidance: bool = True,
negative_prompt: Optional[str] = None,
negative_prompt_2: Optional[str] = None,
prompt_embeds: Optional[torch.Tensor] = None,
negative_prompt_embeds: Optional[torch.Tensor] = None,
pooled_prompt_embeds: Optional[torch.Tensor] = None,
negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
lora_scale: Optional[float] = None,
clip_skip: Optional[int] = None,
### Added args
num_id_images: int = 1,
class_tokens_mask: Optional[torch.LongTensor] = None, class_tokens_mask: Optional[torch.LongTensor] = None,
): ):
device = device or self._execution_device device = device or self._execution_device
if prompt is not None and isinstance(prompt, str): # set lora scale so that monkey patched LoRA
batch_size = 1 # function of text encoder can correctly access it
elif prompt is not None and isinstance(prompt, list): if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
self._lora_scale = lora_scale
# dynamically adjust the LoRA scale
if self.text_encoder is not None:
if not USE_PEFT_BACKEND:
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
else:
scale_lora_layers(self.text_encoder, lora_scale)
if self.text_encoder_2 is not None:
if not USE_PEFT_BACKEND:
adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
else:
scale_lora_layers(self.text_encoder_2, lora_scale)
prompt = [prompt] if isinstance(prompt, str) else prompt
if prompt is not None:
batch_size = len(prompt) batch_size = len(prompt)
else: else:
batch_size = prompt_embeds.shape[0] batch_size = prompt_embeds.shape[0]
...@@ -157,15 +308,40 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -157,15 +308,40 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
if prompt_embeds is None: if prompt_embeds is None:
prompt_2 = prompt_2 or prompt prompt_2 = prompt_2 or prompt
prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
# textual inversion: process multi-vector tokens if necessary
prompt_embeds_list = [] prompt_embeds_list = []
prompts = [prompt, prompt_2] prompts = [prompt, prompt_2]
for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
input_ids = tokenizer.encode(prompt) # TODO: batch encode if isinstance(self, TextualInversionLoaderMixin):
prompt = self.maybe_convert_prompt(prompt, tokenizer)
text_inputs = tokenizer(
prompt,
padding="max_length",
max_length=tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
text_input_ids, untruncated_ids
):
removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
print(
"The following part of your input was truncated because CLIP can only handle sequences up to"
f" {tokenizer.model_max_length} tokens: {removed_text}"
)
clean_index = 0 clean_index = 0
clean_input_ids = [] clean_input_ids = []
class_token_index = [] class_token_index = []
# Find out the corresponding class word token based on the newly added trigger word token # Find out the corresponding class word token based on the newly added trigger word token
for i, token_id in enumerate(input_ids): for i, token_id in enumerate(text_input_ids.tolist()[0]):
if token_id == image_token_id: if token_id == image_token_id:
class_token_index.append(clean_index - 1) class_token_index.append(clean_index - 1)
else: else:
...@@ -181,7 +357,7 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -181,7 +357,7 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
# Expand the class word token and corresponding mask # Expand the class word token and corresponding mask
class_token = clean_input_ids[class_token_index] class_token = clean_input_ids[class_token_index]
clean_input_ids = clean_input_ids[:class_token_index] + [class_token] * num_id_images + \ clean_input_ids = clean_input_ids[:class_token_index] + [class_token] * num_id_images * self.num_tokens + \
clean_input_ids[class_token_index+1:] clean_input_ids[class_token_index+1:]
# Truncation or padding # Truncation or padding
...@@ -193,32 +369,122 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -193,32 +369,122 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
max_len - len(clean_input_ids) max_len - len(clean_input_ids)
) )
class_tokens_mask = [True if class_token_index <= i < class_token_index+num_id_images else False \ class_tokens_mask = [True if class_token_index <= i < class_token_index+(num_id_images * self.num_tokens) else False \
for i in range(len(clean_input_ids))] for i in range(len(clean_input_ids))]
clean_input_ids = torch.tensor(clean_input_ids, dtype=torch.long).unsqueeze(0) clean_input_ids = torch.tensor(clean_input_ids, dtype=torch.long).unsqueeze(0)
class_tokens_mask = torch.tensor(class_tokens_mask, dtype=torch.bool).unsqueeze(0) class_tokens_mask = torch.tensor(class_tokens_mask, dtype=torch.bool).unsqueeze(0)
prompt_embeds = text_encoder( prompt_embeds = text_encoder(clean_input_ids.to(device), output_hidden_states=True)
clean_input_ids.to(device),
output_hidden_states=True,
)
# We are only ALWAYS interested in the pooled output of the final text encoder # We are only ALWAYS interested in the pooled output of the final text encoder
pooled_prompt_embeds = prompt_embeds[0] pooled_prompt_embeds = prompt_embeds[0]
prompt_embeds = prompt_embeds.hidden_states[-2] if clip_skip is None:
prompt_embeds = prompt_embeds.hidden_states[-2]
else:
# "2" because SDXL always indexes from the penultimate layer.
prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
prompt_embeds_list.append(prompt_embeds) prompt_embeds_list.append(prompt_embeds)
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
class_tokens_mask = class_tokens_mask.to(device=device) # TODO: ignoring two-prompt case class_tokens_mask = class_tokens_mask.to(device=device) # TODO: ignoring two-prompt case
# get unconditional embeddings for classifier free guidance
zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
elif do_classifier_free_guidance and negative_prompt_embeds is None:
negative_prompt = negative_prompt or ""
negative_prompt_2 = negative_prompt_2 or negative_prompt
# normalize str to list
negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
negative_prompt_2 = (
batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
)
uncond_tokens: List[str]
if prompt is not None and type(prompt) is not type(negative_prompt):
raise TypeError(
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
f" {type(prompt)}."
)
elif batch_size != len(negative_prompt):
raise ValueError(
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
" the batch size of `prompt`."
)
else:
uncond_tokens = [negative_prompt, negative_prompt_2]
negative_prompt_embeds_list = []
for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
if isinstance(self, TextualInversionLoaderMixin):
negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
max_length = prompt_embeds.shape[1]
uncond_input = tokenizer(
negative_prompt,
padding="max_length",
max_length=max_length,
truncation=True,
return_tensors="pt",
)
negative_prompt_embeds = text_encoder(
uncond_input.input_ids.to(device),
output_hidden_states=True,
)
# We are only ALWAYS interested in the pooled output of the final text encoder
negative_pooled_prompt_embeds = negative_prompt_embeds[0]
negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
negative_prompt_embeds_list.append(negative_prompt_embeds)
negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
return prompt_embeds, pooled_prompt_embeds, class_tokens_mask if self.text_encoder_2 is not None:
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
else:
prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
bs_embed, seq_len, _ = prompt_embeds.shape
@property if do_classifier_free_guidance:
def interrupt(self): # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
return self._interrupt seq_len = negative_prompt_embeds.shape[1]
if self.text_encoder_2 is not None:
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
else:
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
if do_classifier_free_guidance:
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
if self.text_encoder is not None:
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
# Retrieve the original scale by scaling back the LoRA layers
unscale_lora_layers(self.text_encoder, lora_scale)
if self.text_encoder_2 is not None:
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
# Retrieve the original scale by scaling back the LoRA layers
unscale_lora_layers(self.text_encoder_2, lora_scale)
return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds, class_tokens_mask
@torch.no_grad() @torch.no_grad()
def __call__( def __call__(
...@@ -228,6 +494,8 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -228,6 +494,8 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
height: Optional[int] = None, height: Optional[int] = None,
width: Optional[int] = None, width: Optional[int] = None,
num_inference_steps: int = 50, num_inference_steps: int = 50,
timesteps: List[int] = None,
sigmas: List[float] = None,
denoising_end: Optional[float] = None, denoising_end: Optional[float] = None,
guidance_scale: float = 5.0, guidance_scale: float = 5.0,
negative_prompt: Optional[Union[str, List[str]]] = None, negative_prompt: Optional[Union[str, List[str]]] = None,
...@@ -235,11 +503,13 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -235,11 +503,13 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
num_images_per_prompt: Optional[int] = 1, num_images_per_prompt: Optional[int] = 1,
eta: float = 0.0, eta: float = 0.0,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
latents: Optional[torch.FloatTensor] = None, latents: Optional[torch.Tensor] = None,
prompt_embeds: Optional[torch.FloatTensor] = None, prompt_embeds: Optional[torch.Tensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None, negative_prompt_embeds: Optional[torch.Tensor] = None,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None, pooled_prompt_embeds: Optional[torch.Tensor] = None,
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
ip_adapter_image: Optional[PipelineImageInput] = None,
ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
output_type: Optional[str] = "pil", output_type: Optional[str] = "pil",
return_dict: bool = True, return_dict: bool = True,
cross_attention_kwargs: Optional[Dict[str, Any]] = None, cross_attention_kwargs: Optional[Dict[str, Any]] = None,
...@@ -247,16 +517,22 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -247,16 +517,22 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
original_size: Optional[Tuple[int, int]] = None, original_size: Optional[Tuple[int, int]] = None,
crops_coords_top_left: Tuple[int, int] = (0, 0), crops_coords_top_left: Tuple[int, int] = (0, 0),
target_size: Optional[Tuple[int, int]] = None, target_size: Optional[Tuple[int, int]] = None,
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, negative_original_size: Optional[Tuple[int, int]] = None,
callback_steps: int = 1, negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, negative_target_size: Optional[Tuple[int, int]] = None,
clip_skip: Optional[int] = None,
callback_on_step_end: Optional[
Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
] = None,
callback_on_step_end_tensor_inputs: List[str] = ["latents"], callback_on_step_end_tensor_inputs: List[str] = ["latents"],
# Added parameters (for PhotoMaker) # Added parameters (for PhotoMaker)
input_id_images: PipelineImageInput = None, input_id_images: PipelineImageInput = None,
start_merge_step: int = 0, # TODO: change to `style_strength_ratio` in the future start_merge_step: int = 10, # TODO: change to `style_strength_ratio` in the future
class_tokens_mask: Optional[torch.LongTensor] = None, class_tokens_mask: Optional[torch.LongTensor] = None,
id_embeds: Optional[torch.FloatTensor] = None,
prompt_embeds_text_only: Optional[torch.FloatTensor] = None, prompt_embeds_text_only: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds_text_only: Optional[torch.FloatTensor] = None, pooled_prompt_embeds_text_only: Optional[torch.FloatTensor] = None,
**kwargs,
): ):
r""" r"""
Function invoked when calling the pipeline for generation. Function invoked when calling the pipeline for generation.
...@@ -280,9 +556,29 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -280,9 +556,29 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
[`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
`tuple`. When returning a tuple, the first element is a list with the generated images. `tuple`. When returning a tuple, the first element is a list with the generated images.
""" """
callback = kwargs.pop("callback", None)
callback_steps = kwargs.pop("callback_steps", None)
if callback is not None:
deprecate(
"callback",
"1.0.0",
"Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
)
if callback_steps is not None:
deprecate(
"callback_steps",
"1.0.0",
"Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
)
if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
# 0. Default height and width to unet # 0. Default height and width to unet
height = height or self.unet.config.sample_size * self.vae_scale_factor height = height or self.default_sample_size * self.vae_scale_factor
width = width or self.unet.config.sample_size * self.vae_scale_factor width = width or self.default_sample_size * self.vae_scale_factor
original_size = original_size or (height, width) original_size = original_size or (height, width)
target_size = target_size or (height, width) target_size = target_size or (height, width)
...@@ -300,10 +596,17 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -300,10 +596,17 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
negative_prompt_embeds, negative_prompt_embeds,
pooled_prompt_embeds, pooled_prompt_embeds,
negative_pooled_prompt_embeds, negative_pooled_prompt_embeds,
ip_adapter_image,
ip_adapter_image_embeds,
callback_on_step_end_tensor_inputs, callback_on_step_end_tensor_inputs,
) )
self._interrupt = False self._guidance_scale = guidance_scale
self._guidance_rescale = guidance_rescale
self._clip_skip = clip_skip
self._cross_attention_kwargs = cross_attention_kwargs
self._denoising_end = denoising_end
self._interrupt = False
# #
if prompt_embeds is not None and class_tokens_mask is None: if prompt_embeds is not None and class_tokens_mask is None:
...@@ -328,30 +631,36 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -328,30 +631,36 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
device = self._execution_device device = self._execution_device
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
# corresponds to doing no classifier free guidance.
do_classifier_free_guidance = guidance_scale >= 1.0
assert do_classifier_free_guidance
# 3. Encode input prompt # 3. Encode input prompt
num_id_images = len(input_id_images) lora_scale = (
self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
)
num_id_images = len(input_id_images)
( (
prompt_embeds, prompt_embeds,
_,
pooled_prompt_embeds, pooled_prompt_embeds,
_,
class_tokens_mask, class_tokens_mask,
) = self.encode_prompt_with_trigger_word( ) = self.encode_prompt_with_trigger_word(
prompt=prompt, prompt=prompt,
prompt_2=prompt_2, prompt_2=prompt_2,
device=device, device=device,
num_id_images=num_id_images, num_id_images=num_id_images,
class_tokens_mask=class_tokens_mask,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=self.do_classifier_free_guidance,
negative_prompt=negative_prompt,
negative_prompt_2=negative_prompt_2,
prompt_embeds=prompt_embeds, prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds,
class_tokens_mask=class_tokens_mask, negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
lora_scale=lora_scale,
clip_skip=self.clip_skip,
) )
# 4. Encode input prompt without the trigger word for delayed conditioning # 4. Encode input prompt without the trigger word for delayed conditioning
# encode, remove trigger word token, then decode # encode, remove trigger word token, then decode
tokens_text_only = self.tokenizer.encode(prompt, add_special_tokens=False) tokens_text_only = self.tokenizer.encode(prompt, add_special_tokens=False)
...@@ -368,36 +677,40 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -368,36 +677,40 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
prompt_2=prompt_2, prompt_2=prompt_2,
device=device, device=device,
num_images_per_prompt=num_images_per_prompt, num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=do_classifier_free_guidance, do_classifier_free_guidance=self.do_classifier_free_guidance,
negative_prompt=negative_prompt, negative_prompt=negative_prompt,
negative_prompt_2=negative_prompt_2, negative_prompt_2=negative_prompt_2,
prompt_embeds=prompt_embeds_text_only, prompt_embeds=prompt_embeds_text_only,
negative_prompt_embeds=negative_prompt_embeds, negative_prompt_embeds=negative_prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds_text_only, pooled_prompt_embeds=pooled_prompt_embeds_text_only,
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
lora_scale=lora_scale,
clip_skip=self.clip_skip,
) )
# 5. Prepare the input ID images # 5. Prepare timesteps
timesteps, num_inference_steps = retrieve_timesteps(
self.scheduler, num_inference_steps, device, timesteps, sigmas
)
# 6. Prepare the input ID images
dtype = next(self.id_encoder.parameters()).dtype dtype = next(self.id_encoder.parameters()).dtype
if not isinstance(input_id_images[0], torch.Tensor): if not isinstance(input_id_images[0], torch.Tensor):
id_pixel_values = self.id_image_processor(input_id_images, return_tensors="pt").pixel_values id_pixel_values = self.id_image_processor(input_id_images, return_tensors="pt").pixel_values
id_pixel_values = id_pixel_values.unsqueeze(0).to(device=device, dtype=dtype) # TODO: multiple prompts id_pixel_values = id_pixel_values.unsqueeze(0).to(device=device, dtype=dtype) # TODO: multiple prompts
# 6. Get the update text embedding with the stacked ID embedding # 7. Get the update text embedding with the stacked ID embedding
prompt_embeds = self.id_encoder(id_pixel_values, prompt_embeds, class_tokens_mask) if id_embeds is not None:
id_embeds = id_embeds.unsqueeze(0).to(device=device, dtype=dtype)
prompt_embeds = self.id_encoder(id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds)
else:
prompt_embeds = self.id_encoder(id_pixel_values, prompt_embeds, class_tokens_mask)
bs_embed, seq_len, _ = prompt_embeds.shape bs_embed, seq_len, _ = prompt_embeds.shape
# duplicate text embeddings for each generation per prompt, using mps friendly method # duplicate text embeddings for each generation per prompt, using mps friendly method
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
# 7. Prepare timesteps
self.scheduler.set_timesteps(num_inference_steps, device=device)
timesteps = self.scheduler.timesteps
# 8. Prepare latent variables # 8. Prepare latent variables
num_channels_latents = self.unet.config.in_channels num_channels_latents = self.unet.config.in_channels
...@@ -416,6 +729,7 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -416,6 +729,7 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
# 10. Prepare added time ids & embeddings # 10. Prepare added time ids & embeddings
add_text_embeds = pooled_prompt_embeds
if self.text_encoder_2 is None: if self.text_encoder_2 is None:
text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
else: else:
...@@ -428,53 +742,117 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -428,53 +742,117 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
dtype=prompt_embeds.dtype, dtype=prompt_embeds.dtype,
text_encoder_projection_dim=text_encoder_projection_dim, text_encoder_projection_dim=text_encoder_projection_dim,
) )
add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0) if negative_original_size is not None and negative_target_size is not None:
negative_add_time_ids = self._get_add_time_ids(
negative_original_size,
negative_crops_coords_top_left,
negative_target_size,
dtype=prompt_embeds.dtype,
text_encoder_projection_dim=text_encoder_projection_dim,
)
else:
negative_add_time_ids = add_time_ids
if self.do_classifier_free_guidance:
add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
prompt_embeds = prompt_embeds.to(device)
add_text_embeds = add_text_embeds.to(device)
add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
image_embeds = self.prepare_ip_adapter_image_embeds(
ip_adapter_image,
ip_adapter_image_embeds,
device,
batch_size * num_images_per_prompt,
self.do_classifier_free_guidance,
)
# 11. Denoising loop # 11. Denoising loop
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
# 11.1 Apply denoising_end
if (
self.denoising_end is not None
and isinstance(self.denoising_end, float)
and self.denoising_end > 0
and self.denoising_end < 1
):
discrete_timestep_cutoff = int(
round(
self.scheduler.config.num_train_timesteps
- (self.denoising_end * self.scheduler.config.num_train_timesteps)
)
)
num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
timesteps = timesteps[:num_inference_steps]
# 12. Optionally get Guidance Scale Embedding
timestep_cond = None
if self.unet.config.time_cond_proj_dim is not None:
guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
timestep_cond = self.get_guidance_scale_embedding(
guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
).to(device=device, dtype=latents.dtype)
self._num_timesteps = len(timesteps)
with self.progress_bar(total=num_inference_steps) as progress_bar: with self.progress_bar(total=num_inference_steps) as progress_bar:
for i, t in enumerate(timesteps): for i, t in enumerate(timesteps):
if self.interrupt: if self.interrupt:
continue continue
latent_model_input = ( # expand the latents if we are doing classifier free guidance
torch.cat([latents] * 2) if do_classifier_free_guidance else latents latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
)
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
if i <= start_merge_step: if i <= start_merge_step:
current_prompt_embeds = torch.cat( current_prompt_embeds = torch.cat(
[negative_prompt_embeds, prompt_embeds_text_only], dim=0 [negative_prompt_embeds, prompt_embeds_text_only], dim=0
) ) if self.do_classifier_free_guidance else prompt_embeds_text_only
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds_text_only], dim=0) add_text_embeds = torch.cat(
[negative_pooled_prompt_embeds, pooled_prompt_embeds_text_only], dim=0
) if self.do_classifier_free_guidance else pooled_prompt_embeds_text_only
else: else:
current_prompt_embeds = torch.cat( current_prompt_embeds = torch.cat(
[negative_prompt_embeds, prompt_embeds], dim=0 [negative_prompt_embeds, prompt_embeds], dim=0
) ) if self.do_classifier_free_guidance else prompt_embeds
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) add_text_embeds = torch.cat(
# predict the noise residual [negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0
) if self.do_classifier_free_guidance else pooled_prompt_embeds
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
added_cond_kwargs["image_embeds"] = image_embeds
# predict the noise residual
noise_pred = self.unet( noise_pred = self.unet(
latent_model_input, latent_model_input,
t, t,
encoder_hidden_states=current_prompt_embeds, encoder_hidden_states=current_prompt_embeds,
cross_attention_kwargs=cross_attention_kwargs, timestep_cond=timestep_cond,
cross_attention_kwargs=self.cross_attention_kwargs,
added_cond_kwargs=added_cond_kwargs, added_cond_kwargs=added_cond_kwargs,
return_dict=False, return_dict=False,
)[0] )[0]
# perform guidance # perform guidance
if do_classifier_free_guidance: if self.do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
if do_classifier_free_guidance and guidance_rescale > 0.0: if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
# compute the previous noisy sample x_t -> x_t-1 # compute the previous noisy sample x_t -> x_t-1
latents_dtype = latents.dtype
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
if latents.dtype != latents_dtype:
if torch.backends.mps.is_available():
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
latents = latents.to(latents_dtype)
if callback_on_step_end is not None: if callback_on_step_end is not None:
callback_kwargs = {} callback_kwargs = {}
...@@ -486,11 +864,11 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -486,11 +864,11 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
# negative_pooled_prompt_embeds = callback_outputs.pop( negative_pooled_prompt_embeds = callback_outputs.pop(
# "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
# ) )
# add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
# negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids) negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
# call the callback, if provided # call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
...@@ -499,13 +877,41 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -499,13 +877,41 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
step_idx = i // getattr(self.scheduler, "order", 1) step_idx = i // getattr(self.scheduler, "order", 1)
callback(step_idx, t, latents) callback(step_idx, t, latents)
# make sure the VAE is in float32 mode, as it overflows in float16 if XLA_AVAILABLE:
if self.vae.dtype == torch.float16 and self.vae.config.force_upcast: xm.mark_step()
self.upcast_vae()
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
if not output_type == "latent": if not output_type == "latent":
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] # make sure the VAE is in float32 mode, as it overflows in float16
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
elif latents.dtype != self.vae.dtype:
if torch.backends.mps.is_available():
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
self.vae = self.vae.to(latents.dtype)
# unscale/denormalize the latents
# denormalize with the mean and std if available and not None
has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
if has_latents_mean and has_latents_std:
latents_mean = (
torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
)
latents_std = (
torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
)
latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
else:
latents = latents / self.vae.config.scaling_factor
image = self.vae.decode(latents, return_dict=False)[0]
# cast back to fp16 if needed
if needs_upcasting:
self.vae.to(dtype=torch.float16)
else: else:
image = latents image = latents
return StableDiffusionXLPipelineOutput(images=image) return StableDiffusionXLPipelineOutput(images=image)
...@@ -522,4 +928,4 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): ...@@ -522,4 +928,4 @@ class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline):
if not return_dict: if not return_dict:
return (image,) return (image,)
return StableDiffusionXLPipelineOutput(images=image) return StableDiffusionXLPipelineOutput(images=image)
\ No newline at end of file
photomaker/pipeline_controlnet.py 0 → 100755
View file @ 48ec26d1
#####
# Modified from https://github.com/huggingface/diffusers/blob/v0.29.1/src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py
# PhotoMaker v2 @ TencentARC and MCG-NKU
# Author: Zhen Li
#####
# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import inspect
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import numpy as np
import PIL.Image
import torch
import torch.nn.functional as F
from safetensors import safe_open
from huggingface_hub.utils import validate_hf_hub_args
from transformers import (
CLIPImageProcessor,
CLIPTextModel,
CLIPTextModelWithProjection,
CLIPTokenizer,
CLIPVisionModelWithProjection,
)
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import (
FromSingleFileMixin,
IPAdapterMixin,
StableDiffusionXLLoraLoaderMixin,
TextualInversionLoaderMixin,
)
from diffusers.models import AutoencoderKL, ControlNetModel, ImageProjection, UNet2DConditionModel
from diffusers.models.attention_processor import (
AttnProcessor2_0,
LoRAAttnProcessor2_0,
LoRAXFormersAttnProcessor,
XFormersAttnProcessor,
)
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
USE_PEFT_BACKEND,
deprecate,
logging,
replace_example_docstring,
scale_lora_layers,
unscale_lora_layers,
)
from diffusers.utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
from diffusers import StableDiffusionXLControlNetPipeline, ControlNetModel, AutoencoderKL
from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
from diffusers.utils import _get_model_file
from . import (
PhotoMakerIDEncoder, # PhotoMaker v1
PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken, # PhotoMaker v2
)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
"""
Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
"""
std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
# rescale the results from guidance (fixes overexposure)
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
# mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
return noise_cfg
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
scheduler,
num_inference_steps: Optional[int] = None,
device: Optional[Union[str, torch.device]] = None,
timesteps: Optional[List[int]] = None,
sigmas: Optional[List[float]] = None,
**kwargs,
):
"""
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
Args:
scheduler (`SchedulerMixin`):
The scheduler to get timesteps from.
num_inference_steps (`int`):
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
must be `None`.
device (`str` or `torch.device`, *optional*):
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
timesteps (`List[int]`, *optional*):
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
`num_inference_steps` and `sigmas` must be `None`.
sigmas (`List[float]`, *optional*):
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
`num_inference_steps` and `timesteps` must be `None`.
Returns:
`Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
second element is the number of inference steps.
"""
if timesteps is not None and sigmas is not None:
raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
if timesteps is not None:
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
if not accepts_timesteps:
raise ValueError(
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
f" timestep schedules. Please check whether you are using the correct scheduler."
)
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
timesteps = scheduler.timesteps
num_inference_steps = len(timesteps)
elif sigmas is not None:
accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
if not accept_sigmas:
raise ValueError(
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
f" sigmas schedules. Please check whether you are using the correct scheduler."
)
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
timesteps = scheduler.timesteps
num_inference_steps = len(timesteps)
else:
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
timesteps = scheduler.timesteps
return timesteps, num_inference_steps
class PhotoMakerStableDiffusionXLControlNetPipeline(StableDiffusionXLControlNetPipeline):
@validate_hf_hub_args
def load_photomaker_adapter(
self,
pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
weight_name: str,
subfolder: str = '',
trigger_word: str = 'img',
pm_version: str = 'v2',
**kwargs,
):
"""
Parameters:
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
Can be either:
- A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
the Hub.
- A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
with [`ModelMixin.save_pretrained`].
- A [torch state
dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
weight_name (`str`):
The weight name NOT the path to the weight.
subfolder (`str`, defaults to `""`):
The subfolder location of a model file within a larger model repository on the Hub or locally.
trigger_word (`str`, *optional*, defaults to `"img"`):
The trigger word is used to identify the position of class word in the text prompt,
and it is recommended not to set it as a common word.
This trigger word must be placed after the class word when used, otherwise, it will affect the performance of the personalized generation.
"""
# Load the main state dict first.
cache_dir = kwargs.pop("cache_dir", None)
force_download = kwargs.pop("force_download", False)
resume_download = kwargs.pop("resume_download", False)
proxies = kwargs.pop("proxies", None)
local_files_only = kwargs.pop("local_files_only", None)
token = kwargs.pop("token", None)
revision = kwargs.pop("revision", None)
user_agent = {
"file_type": "attn_procs_weights",
"framework": "pytorch",
}
if not isinstance(pretrained_model_name_or_path_or_dict, dict):
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
if weight_name.endswith(".safetensors"):
state_dict = {"id_encoder": {}, "lora_weights": {}}
with safe_open(model_file, framework="pt", device="cpu") as f:
for key in f.keys():
if key.startswith("id_encoder."):
state_dict["id_encoder"][key.replace("id_encoder.", "")] = f.get_tensor(key)
elif key.startswith("lora_weights."):
state_dict["lora_weights"][key.replace("lora_weights.", "")] = f.get_tensor(key)
else:
state_dict = torch.load(model_file, map_location="cpu")
else:
state_dict = pretrained_model_name_or_path_or_dict
keys = list(state_dict.keys())
if keys != ["id_encoder", "lora_weights"]:
raise ValueError("Required keys are (`id_encoder` and `lora_weights`) missing from the state dict.")
self.num_tokens = 2
self.trigger_word = trigger_word
# load finetuned CLIP image encoder and fuse module here if it has not been registered to the pipeline yet
print(f"Loading PhotoMaker {pm_version} components [1] id_encoder from [{pretrained_model_name_or_path_or_dict}]...")
self.id_image_processor = CLIPImageProcessor()
if pm_version == "v1": # PhotoMaker v1
id_encoder = PhotoMakerIDEncoder()
elif pm_version == "v2": # PhotoMaker v2
id_encoder = PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken()
else:
raise NotImplementedError(f"The PhotoMaker version [{pm_version}] does not support")
id_encoder.load_state_dict(state_dict["id_encoder"], strict=True)
id_encoder = id_encoder.to(self.device, dtype=self.unet.dtype)
self.id_encoder = id_encoder
# load lora into models
print(f"Loading PhotoMaker {pm_version} components [2] lora_weights from [{pretrained_model_name_or_path_or_dict}]")
self.load_lora_weights(state_dict["lora_weights"], adapter_name="photomaker")
# Add trigger word token
if self.tokenizer is not None:
self.tokenizer.add_tokens([self.trigger_word], special_tokens=True)
self.tokenizer_2.add_tokens([self.trigger_word], special_tokens=True)
def encode_prompt_with_trigger_word(
self,
prompt: str,
prompt_2: Optional[str] = None,
device: Optional[torch.device] = None,
num_images_per_prompt: int = 1,
do_classifier_free_guidance: bool = True,
negative_prompt: Optional[str] = None,
negative_prompt_2: Optional[str] = None,
prompt_embeds: Optional[torch.Tensor] = None,
negative_prompt_embeds: Optional[torch.Tensor] = None,
pooled_prompt_embeds: Optional[torch.Tensor] = None,
negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
lora_scale: Optional[float] = None,
clip_skip: Optional[int] = None,
### Added args
num_id_images: int = 1,
class_tokens_mask: Optional[torch.LongTensor] = None,
):
device = device or self._execution_device
# set lora scale so that monkey patched LoRA
# function of text encoder can correctly access it
if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
self._lora_scale = lora_scale
# dynamically adjust the LoRA scale
if self.text_encoder is not None:
if not USE_PEFT_BACKEND:
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
else:
scale_lora_layers(self.text_encoder, lora_scale)
if self.text_encoder_2 is not None:
if not USE_PEFT_BACKEND:
adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
else:
scale_lora_layers(self.text_encoder_2, lora_scale)
prompt = [prompt] if isinstance(prompt, str) else prompt
if prompt is not None:
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
# Find the token id of the trigger word
image_token_id = self.tokenizer_2.convert_tokens_to_ids(self.trigger_word)
# Define tokenizers and text encoders
tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
text_encoders = (
[self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
)
if prompt_embeds is None:
prompt_2 = prompt_2 or prompt
prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
# textual inversion: process multi-vector tokens if necessary
prompt_embeds_list = []
prompts = [prompt, prompt_2]
for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
if isinstance(self, TextualInversionLoaderMixin):
prompt = self.maybe_convert_prompt(prompt, tokenizer)
text_inputs = tokenizer(
prompt,
padding="max_length",
max_length=tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
text_input_ids, untruncated_ids
):
removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
print(
"The following part of your input was truncated because CLIP can only handle sequences up to"
f" {tokenizer.model_max_length} tokens: {removed_text}"
)
clean_index = 0
clean_input_ids = []
class_token_index = []
# Find out the corresponding class word token based on the newly added trigger word token
for i, token_id in enumerate(text_input_ids.tolist()[0]):
if token_id == image_token_id:
class_token_index.append(clean_index - 1)
else:
clean_input_ids.append(token_id)
clean_index += 1
if len(class_token_index) != 1:
raise ValueError(
f"PhotoMaker currently does not support multiple trigger words in a single prompt.\
Trigger word: {self.trigger_word}, Prompt: {prompt}."
)
class_token_index = class_token_index[0]
# Expand the class word token and corresponding mask
class_token = clean_input_ids[class_token_index]
clean_input_ids = clean_input_ids[:class_token_index] + [class_token] * num_id_images * self.num_tokens + \
clean_input_ids[class_token_index+1:]
# Truncation or padding
max_len = tokenizer.model_max_length
if len(clean_input_ids) > max_len:
clean_input_ids = clean_input_ids[:max_len]
else:
clean_input_ids = clean_input_ids + [tokenizer.pad_token_id] * (
max_len - len(clean_input_ids)
)
class_tokens_mask = [True if class_token_index <= i < class_token_index+(num_id_images * self.num_tokens) else False \
for i in range(len(clean_input_ids))]
clean_input_ids = torch.tensor(clean_input_ids, dtype=torch.long).unsqueeze(0)
class_tokens_mask = torch.tensor(class_tokens_mask, dtype=torch.bool).unsqueeze(0)
prompt_embeds = text_encoder(clean_input_ids.to(device), output_hidden_states=True)
# We are only ALWAYS interested in the pooled output of the final text encoder
pooled_prompt_embeds = prompt_embeds[0]
if clip_skip is None:
prompt_embeds = prompt_embeds.hidden_states[-2]
else:
# "2" because SDXL always indexes from the penultimate layer.
prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
prompt_embeds_list.append(prompt_embeds)
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
class_tokens_mask = class_tokens_mask.to(device=device) # TODO: ignoring two-prompt case
# get unconditional embeddings for classifier free guidance
zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
elif do_classifier_free_guidance and negative_prompt_embeds is None:
negative_prompt = negative_prompt or ""
negative_prompt_2 = negative_prompt_2 or negative_prompt
# normalize str to list
negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
negative_prompt_2 = (
batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
)
uncond_tokens: List[str]
if prompt is not None and type(prompt) is not type(negative_prompt):
raise TypeError(
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
f" {type(prompt)}."
)
elif batch_size != len(negative_prompt):
raise ValueError(
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
" the batch size of `prompt`."
)
else:
uncond_tokens = [negative_prompt, negative_prompt_2]
negative_prompt_embeds_list = []
for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
if isinstance(self, TextualInversionLoaderMixin):
negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
max_length = prompt_embeds.shape[1]
uncond_input = tokenizer(
negative_prompt,
padding="max_length",
max_length=max_length,
truncation=True,
return_tensors="pt",
)
negative_prompt_embeds = text_encoder(
uncond_input.input_ids.to(device),
output_hidden_states=True,
)
# We are only ALWAYS interested in the pooled output of the final text encoder
negative_pooled_prompt_embeds = negative_prompt_embeds[0]
negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
negative_prompt_embeds_list.append(negative_prompt_embeds)
negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
if self.text_encoder_2 is not None:
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
else:
prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
bs_embed, seq_len, _ = prompt_embeds.shape
if do_classifier_free_guidance:
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
seq_len = negative_prompt_embeds.shape[1]
if self.text_encoder_2 is not None:
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
else:
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
if do_classifier_free_guidance:
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
if self.text_encoder is not None:
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
# Retrieve the original scale by scaling back the LoRA layers
unscale_lora_layers(self.text_encoder, lora_scale)
if self.text_encoder_2 is not None:
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
# Retrieve the original scale by scaling back the LoRA layers
unscale_lora_layers(self.text_encoder_2, lora_scale)
return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds, class_tokens_mask
@torch.no_grad()
def __call__(
self,
prompt: Union[str, List[str]] = None,
prompt_2: Optional[Union[str, List[str]]] = None,
image: PipelineImageInput = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 50,
timesteps: List[int] = None,
sigmas: List[float] = None,
denoising_end: Optional[float] = None,
guidance_scale: float = 5.0,
negative_prompt: Optional[Union[str, List[str]]] = None,
negative_prompt_2: Optional[Union[str, List[str]]] = None,
num_images_per_prompt: Optional[int] = 1,
eta: float = 0.0,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
latents: Optional[torch.Tensor] = None,
prompt_embeds: Optional[torch.Tensor] = None,
negative_prompt_embeds: Optional[torch.Tensor] = None,
pooled_prompt_embeds: Optional[torch.Tensor] = None,
negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
ip_adapter_image: Optional[PipelineImageInput] = None,
ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
guess_mode: bool = False,
control_guidance_start: Union[float, List[float]] = 0.0,
control_guidance_end: Union[float, List[float]] = 1.0,
original_size: Tuple[int, int] = None,
crops_coords_top_left: Tuple[int, int] = (0, 0),
target_size: Tuple[int, int] = None,
negative_original_size: Optional[Tuple[int, int]] = None,
negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
negative_target_size: Optional[Tuple[int, int]] = None,
clip_skip: Optional[int] = None,
callback_on_step_end: Optional[
Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
] = None,
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
# Added parameters (for PhotoMaker)
input_id_images: PipelineImageInput = None,
start_merge_step: int = 10, # TODO: change to `style_strength_ratio` in the future
class_tokens_mask: Optional[torch.LongTensor] = None,
id_embeds: Optional[torch.FloatTensor] = None,
prompt_embeds_text_only: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds_text_only: Optional[torch.FloatTensor] = None,
**kwargs,
):
r"""
Function invoked when calling the pipeline for generation.
Only the parameters introduced by PhotoMaker are discussed here.
For explanations of the previous parameters in StableDiffusionXLControlNetPipeline, please refer to https://github.com/huggingface/diffusers/blob/v0.25.0/src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py
Args:
input_id_images (`PipelineImageInput`, *optional*):
Input ID Image to work with PhotoMaker.
class_tokens_mask (`torch.LongTensor`, *optional*):
Pre-generated class token. When the `prompt_embeds` parameter is provided in advance, it is necessary to prepare the `class_tokens_mask` beforehand for marking out the position of class word.
prompt_embeds_text_only (`torch.FloatTensor`, *optional*):
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
provided, text embeddings will be generated from `prompt` input argument.
pooled_prompt_embeds_text_only (`torch.FloatTensor`, *optional*):
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
If not provided, pooled text embeddings will be generated from `prompt` input argument.
Returns:
[`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
[`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
`tuple`. When returning a tuple, the first element is a list with the generated images.
"""
callback = kwargs.pop("callback", None)
callback_steps = kwargs.pop("callback_steps", None)
if callback is not None:
deprecate(
"callback",
"1.0.0",
"Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
)
if callback_steps is not None:
deprecate(
"callback_steps",
"1.0.0",
"Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
)
if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
# align format for control guidance
if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
control_guidance_start = len(control_guidance_end) * [control_guidance_start]
elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
control_guidance_end = len(control_guidance_start) * [control_guidance_end]
elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
control_guidance_start, control_guidance_end = (
mult * [control_guidance_start],
mult * [control_guidance_end],
)
# 1. Check inputs. Raise error if not correct
self.check_inputs(
prompt,
prompt_2,
image,
callback_steps,
negative_prompt,
negative_prompt_2,
prompt_embeds,
negative_prompt_embeds,
pooled_prompt_embeds,
ip_adapter_image,
ip_adapter_image_embeds,
negative_pooled_prompt_embeds,
controlnet_conditioning_scale,
control_guidance_start,
control_guidance_end,
callback_on_step_end_tensor_inputs,
)
self._guidance_scale = guidance_scale
self._clip_skip = clip_skip
self._cross_attention_kwargs = cross_attention_kwargs
self._denoising_end = denoising_end
#
if prompt_embeds is not None and class_tokens_mask is None:
raise ValueError(
"If `prompt_embeds` are provided, `class_tokens_mask` also have to be passed. Make sure to generate `class_tokens_mask` from the same tokenizer that was used to generate `prompt_embeds`."
)
# check the input id images
if input_id_images is None:
raise ValueError(
"Provide `input_id_images`. Cannot leave `input_id_images` undefined for PhotoMaker pipeline."
)
if not isinstance(input_id_images, list):
input_id_images = [input_id_images]
# 2. Define call parameters
if prompt is not None and isinstance(prompt, str):
batch_size = 1
elif prompt is not None and isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
device = self._execution_device
if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
global_pool_conditions = (
controlnet.config.global_pool_conditions
if isinstance(controlnet, ControlNetModel)
else controlnet.nets[0].config.global_pool_conditions
)
guess_mode = guess_mode or global_pool_conditions
# 3. Encode input prompt
text_encoder_lora_scale = (
self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
)
num_id_images = len(input_id_images)
(
prompt_embeds,
_,
pooled_prompt_embeds,
_,
class_tokens_mask,
) = self.encode_prompt_with_trigger_word(
prompt=prompt,
prompt_2=prompt_2,
device=device,
num_id_images=num_id_images,
class_tokens_mask=class_tokens_mask,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=self.do_classifier_free_guidance,
negative_prompt=negative_prompt,
negative_prompt_2=negative_prompt_2,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds,
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
lora_scale=text_encoder_lora_scale,
clip_skip=self.clip_skip,
)
# print(negative_pooled_prompt_embeds.shape, pooled_prompt_embeds_text_only.shape)
# 4. Encode input prompt without the trigger word for delayed conditioning
# encode, remove trigger word token, then decode
tokens_text_only = self.tokenizer.encode(prompt, add_special_tokens=False)
trigger_word_token = self.tokenizer.convert_tokens_to_ids(self.trigger_word)
tokens_text_only.remove(trigger_word_token)
prompt_text_only = self.tokenizer.decode(tokens_text_only, add_special_tokens=False)
(
prompt_embeds_text_only,
negative_prompt_embeds,
pooled_prompt_embeds_text_only, # TODO: replace the pooled_prompt_embeds with text only prompt
negative_pooled_prompt_embeds,
) = self.encode_prompt(
prompt=prompt_text_only,
prompt_2=prompt_2,
device=device,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=self.do_classifier_free_guidance,
negative_prompt=negative_prompt,
negative_prompt_2=negative_prompt_2,
prompt_embeds=prompt_embeds_text_only,
negative_prompt_embeds=negative_prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds_text_only,
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
lora_scale=text_encoder_lora_scale,
clip_skip=self.clip_skip,
)
# 4.2 Encode ip_adapter_image
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
image_embeds = self.prepare_ip_adapter_image_embeds(
ip_adapter_image,
ip_adapter_image_embeds,
device,
batch_size * num_images_per_prompt,
self.do_classifier_free_guidance,
)
# 5. Prepare the input ID images
dtype = next(self.id_encoder.parameters()).dtype
if not isinstance(input_id_images[0], torch.Tensor):
id_pixel_values = self.id_image_processor(input_id_images, return_tensors="pt").pixel_values
id_pixel_values = id_pixel_values.unsqueeze(0).to(device=device, dtype=dtype) # TODO: multiple prompts
# 6. Get the update text embedding with the stacked ID embedding
if id_embeds is not None:
id_embeds = id_embeds.unsqueeze(0).to(device=device, dtype=dtype)
prompt_embeds = self.id_encoder(id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds)
else:
prompt_embeds = self.id_encoder(id_pixel_values, prompt_embeds, class_tokens_mask)
bs_embed, seq_len, _ = prompt_embeds.shape
# duplicate text embeddings for each generation per prompt, using mps friendly method
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
# 7. Prepare ControlNet image
if isinstance(controlnet, ControlNetModel):
image = self.prepare_image(
image=image,
width=width,
height=height,
batch_size=batch_size * num_images_per_prompt,
num_images_per_prompt=num_images_per_prompt,
device=device,
dtype=controlnet.dtype,
do_classifier_free_guidance=self.do_classifier_free_guidance,
guess_mode=guess_mode,
)
height, width = image.shape[-2:]
elif isinstance(controlnet, MultiControlNetModel):
images = []
for image_ in image:
image_ = self.prepare_image(
image=image_,
width=width,
height=height,
batch_size=batch_size * num_images_per_prompt,
num_images_per_prompt=num_images_per_prompt,
device=device,
dtype=controlnet.dtype,
do_classifier_free_guidance=self.do_classifier_free_guidance,
guess_mode=guess_mode,
)
images.append(image_)
image = images
height, width = image[0].shape[-2:]
else:
assert False
# 7. Prepare timesteps
timesteps, num_inference_steps = retrieve_timesteps(
self.scheduler, num_inference_steps, device, timesteps, sigmas
)
self._num_timesteps = len(timesteps)
# 8. Prepare latent variables
num_channels_latents = self.unet.config.in_channels
latents = self.prepare_latents(
batch_size * num_images_per_prompt,
num_channels_latents,
height,
width,
prompt_embeds.dtype,
device,
generator,
latents,
)
# 8.5 Optionally get Guidance Scale Embedding
timestep_cond = None
if self.unet.config.time_cond_proj_dim is not None:
guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
timestep_cond = self.get_guidance_scale_embedding(
guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
).to(device=device, dtype=latents.dtype)
# 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
# 9.1 Create tensor stating which controlnets to keep
controlnet_keep = []
for i in range(len(timesteps)):
keeps = [
1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
for s, e in zip(control_guidance_start, control_guidance_end)
]
controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
# 9.2 Prepare added time ids & embeddings
if isinstance(image, list):
original_size = original_size or image[0].shape[-2:]
else:
original_size = original_size or image.shape[-2:]
target_size = target_size or (height, width)
add_text_embeds = pooled_prompt_embeds
if self.text_encoder_2 is None:
text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
else:
text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
add_time_ids = self._get_add_time_ids(
original_size,
crops_coords_top_left,
target_size,
dtype=prompt_embeds.dtype,
text_encoder_projection_dim=text_encoder_projection_dim,
)
if negative_original_size is not None and negative_target_size is not None:
negative_add_time_ids = self._get_add_time_ids(
negative_original_size,
negative_crops_coords_top_left,
negative_target_size,
dtype=prompt_embeds.dtype,
text_encoder_projection_dim=text_encoder_projection_dim,
)
else:
negative_add_time_ids = add_time_ids
if self.do_classifier_free_guidance:
add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
prompt_embeds = prompt_embeds.to(device)
add_text_embeds = add_text_embeds.to(device)
add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
# 11. Denoising loop
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
if (
self.denoising_end is not None
and isinstance(self.denoising_end, float)
and self.denoising_end > 0
and self.denoising_end < 1
):
discrete_timestep_cutoff = int(
round(
self.scheduler.config.num_train_timesteps
- (self.denoising_end * self.scheduler.config.num_train_timesteps)
)
)
num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
timesteps = timesteps[:num_inference_steps]
is_unet_compiled = is_compiled_module(self.unet)
is_controlnet_compiled = is_compiled_module(self.controlnet)
is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
with self.progress_bar(total=num_inference_steps) as progress_bar:
for i, t in enumerate(timesteps):
# Relevant thread:
# https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1:
torch._inductor.cudagraph_mark_step_begin()
# expand the latents if we are doing classifier free guidance
latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
if i <= start_merge_step:
current_prompt_embeds = torch.cat(
[negative_prompt_embeds, prompt_embeds_text_only], dim=0
) if self.do_classifier_free_guidance else prompt_embeds_text_only
add_text_embeds = torch.cat(
[negative_pooled_prompt_embeds, pooled_prompt_embeds_text_only], dim=0
) if self.do_classifier_free_guidance else pooled_prompt_embeds_text_only
else:
current_prompt_embeds = torch.cat(
[negative_prompt_embeds, prompt_embeds], dim=0
) if self.do_classifier_free_guidance else prompt_embeds
add_text_embeds = torch.cat(
[negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0
) if self.do_classifier_free_guidance else pooled_prompt_embeds
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
# controlnet(s) inference
controlnet_prompt_embeds_text_only = torch.cat(
[negative_prompt_embeds, prompt_embeds_text_only], dim=0
)
if guess_mode and self.do_classifier_free_guidance:
# Infer ControlNet only for the conditional batch.
control_model_input = latents
control_model_input = self.scheduler.scale_model_input(control_model_input, t)
controlnet_prompt_embeds = controlnet_prompt_embeds_text_only.chunk(2)[1]
controlnet_added_cond_kwargs = {
"text_embeds": add_text_embeds.chunk(2)[1],
"time_ids": add_time_ids.chunk(2)[1],
}
else:
control_model_input = latent_model_input
controlnet_prompt_embeds = controlnet_prompt_embeds_text_only
controlnet_added_cond_kwargs = added_cond_kwargs
if isinstance(controlnet_keep[i], list):
cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
else:
controlnet_cond_scale = controlnet_conditioning_scale
if isinstance(controlnet_cond_scale, list):
controlnet_cond_scale = controlnet_cond_scale[0]
cond_scale = controlnet_cond_scale * controlnet_keep[i]
down_block_res_samples, mid_block_res_sample = self.controlnet(
control_model_input,
t,
encoder_hidden_states=controlnet_prompt_embeds,
controlnet_cond=image,
conditioning_scale=cond_scale,
guess_mode=guess_mode,
added_cond_kwargs=controlnet_added_cond_kwargs,
return_dict=False,
)
if guess_mode and self.do_classifier_free_guidance:
# Infered ControlNet only for the conditional batch.
# To apply the output of ControlNet to both the unconditional and conditional batches,
# add 0 to the unconditional batch to keep it unchanged.
down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
added_cond_kwargs["image_embeds"] = image_embeds
# predict the noise residual
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=current_prompt_embeds,
timestep_cond=timestep_cond,
cross_attention_kwargs=self.cross_attention_kwargs,
down_block_additional_residuals=down_block_res_samples,
mid_block_additional_residual=mid_block_res_sample,
added_cond_kwargs=added_cond_kwargs,
return_dict=False,
)[0]
# perform guidance
if self.do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
# compute the previous noisy sample x_t -> x_t-1
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
if callback_on_step_end is not None:
callback_kwargs = {}
for k in callback_on_step_end_tensor_inputs:
callback_kwargs[k] = locals()[k]
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
latents = callback_outputs.pop("latents", latents)
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
negative_pooled_prompt_embeds = callback_outputs.pop(
"negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
)
add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
# call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
progress_bar.update()
if callback is not None and i % callback_steps == 0:
step_idx = i // getattr(self.scheduler, "order", 1)
callback(step_idx, t, latents)
if not output_type == "latent":
# make sure the VAE is in float32 mode, as it overflows in float16
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
# unscale/denormalize the latents
# denormalize with the mean and std if available and not None
has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
if has_latents_mean and has_latents_std:
latents_mean = (
torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
)
latents_std = (
torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
)
latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
else:
latents = latents / self.vae.config.scaling_factor
image = self.vae.decode(latents, return_dict=False)[0]
# cast back to fp16 if needed
if needs_upcasting:
self.vae.to(dtype=torch.float16)
else:
image = latents
if not output_type == "latent":
# apply watermark if available
if self.watermark is not None:
image = self.watermark.apply_watermark(image)
image = self.image_processor.postprocess(image, output_type=output_type)
# Offload all models
self.maybe_free_model_hooks()
if not return_dict:
return (image,)
return StableDiffusionXLPipelineOutput(images=image)
\ No newline at end of file
photomaker/pipeline_t2i_adapter.py 0 → 100755
View file @ 48ec26d1
#####
# Modified from https://github.com/huggingface/diffusers/blob/v0.29.1/src/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py
# PhotoMaker v2 @ TencentARC and MCG-NKU
# Author: Zhen Li
#####
# Copyright 2024 TencentARC and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import inspect
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import numpy as np
import PIL.Image
import torch
from transformers import (
CLIPImageProcessor,
CLIPTextModel,
CLIPTextModelWithProjection,
CLIPTokenizer,
CLIPVisionModelWithProjection,
)
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import (
FromSingleFileMixin,
IPAdapterMixin,
StableDiffusionXLLoraLoaderMixin,
TextualInversionLoaderMixin,
)
from diffusers.models import AutoencoderKL, ImageProjection, MultiAdapter, T2IAdapter, UNet2DConditionModel
from diffusers.models.attention_processor import (
AttnProcessor2_0,
LoRAAttnProcessor2_0,
LoRAXFormersAttnProcessor,
XFormersAttnProcessor,
)
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
PIL_INTERPOLATION,
USE_PEFT_BACKEND,
logging,
replace_example_docstring,
scale_lora_layers,
unscale_lora_layers,
)
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
from diffusers.pipelines import StableDiffusionXLAdapterPipeline
from diffusers.utils import _get_model_file
from safetensors import safe_open
from huggingface_hub.utils import validate_hf_hub_args
from . import (
PhotoMakerIDEncoder, # PhotoMaker v1
PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken, # PhotoMaker v2
)
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
"""
Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
"""
std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
# rescale the results from guidance (fixes overexposure)
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
# mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
return noise_cfg
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
def retrieve_timesteps(
scheduler,
num_inference_steps: Optional[int] = None,
device: Optional[Union[str, torch.device]] = None,
timesteps: Optional[List[int]] = None,
sigmas: Optional[List[float]] = None,
**kwargs,
):
"""
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
Args:
scheduler (`SchedulerMixin`):
The scheduler to get timesteps from.
num_inference_steps (`int`):
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
must be `None`.
device (`str` or `torch.device`, *optional*):
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
timesteps (`List[int]`, *optional*):
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
`num_inference_steps` and `sigmas` must be `None`.
sigmas (`List[float]`, *optional*):
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
`num_inference_steps` and `timesteps` must be `None`.
Returns:
`Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
second element is the number of inference steps.
"""
if timesteps is not None and sigmas is not None:
raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
if timesteps is not None:
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
if not accepts_timesteps:
raise ValueError(
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
f" timestep schedules. Please check whether you are using the correct scheduler."
)
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
timesteps = scheduler.timesteps
num_inference_steps = len(timesteps)
elif sigmas is not None:
accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
if not accept_sigmas:
raise ValueError(
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
f" sigmas schedules. Please check whether you are using the correct scheduler."
)
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
timesteps = scheduler.timesteps
num_inference_steps = len(timesteps)
else:
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
timesteps = scheduler.timesteps
return timesteps, num_inference_steps
def _preprocess_adapter_image(image, height, width):
if isinstance(image, torch.Tensor):
return image
elif isinstance(image, PIL.Image.Image):
image = [image]
if isinstance(image[0], PIL.Image.Image):
image = [np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])) for i in image]
image = [
i[None, ..., None] if i.ndim == 2 else i[None, ...] for i in image
] # expand [h, w] or [h, w, c] to [b, h, w, c]
image = np.concatenate(image, axis=0)
image = np.array(image).astype(np.float32) / 255.0
image = image.transpose(0, 3, 1, 2)
image = torch.from_numpy(image)
elif isinstance(image[0], torch.Tensor):
if image[0].ndim == 3:
image = torch.stack(image, dim=0)
elif image[0].ndim == 4:
image = torch.cat(image, dim=0)
else:
raise ValueError(
f"Invalid image tensor! Expecting image tensor with 3 or 4 dimension, but recive: {image[0].ndim}"
)
return image
class PhotoMakerStableDiffusionXLAdapterPipeline(StableDiffusionXLAdapterPipeline):
@validate_hf_hub_args
def load_photomaker_adapter(
self,
pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
weight_name: str,
subfolder: str = '',
trigger_word: str = 'img',
pm_version: str = 'v2',
**kwargs,
):
"""
Parameters:
pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
Can be either:
- A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
the Hub.
- A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
with [`ModelMixin.save_pretrained`].
- A [torch state
dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
weight_name (`str`):
The weight name NOT the path to the weight.
subfolder (`str`, defaults to `""`):
The subfolder location of a model file within a larger model repository on the Hub or locally.
trigger_word (`str`, *optional*, defaults to `"img"`):
The trigger word is used to identify the position of class word in the text prompt,
and it is recommended not to set it as a common word.
This trigger word must be placed after the class word when used, otherwise, it will affect the performance of the personalized generation.
"""
# Load the main state dict first.
cache_dir = kwargs.pop("cache_dir", None)
force_download = kwargs.pop("force_download", False)
resume_download = kwargs.pop("resume_download", False)
proxies = kwargs.pop("proxies", None)
local_files_only = kwargs.pop("local_files_only", None)
token = kwargs.pop("token", None)
revision = kwargs.pop("revision", None)
user_agent = {
"file_type": "attn_procs_weights",
"framework": "pytorch",
}
if not isinstance(pretrained_model_name_or_path_or_dict, dict):
model_file = _get_model_file(
pretrained_model_name_or_path_or_dict,
weights_name=weight_name,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
revision=revision,
subfolder=subfolder,
user_agent=user_agent,
)
if weight_name.endswith(".safetensors"):
state_dict = {"id_encoder": {}, "lora_weights": {}}
with safe_open(model_file, framework="pt", device="cpu") as f:
for key in f.keys():
if key.startswith("id_encoder."):
state_dict["id_encoder"][key.replace("id_encoder.", "")] = f.get_tensor(key)
elif key.startswith("lora_weights."):
state_dict["lora_weights"][key.replace("lora_weights.", "")] = f.get_tensor(key)
else:
state_dict = torch.load(model_file, map_location="cpu")
else:
state_dict = pretrained_model_name_or_path_or_dict
keys = list(state_dict.keys())
if keys != ["id_encoder", "lora_weights"]:
raise ValueError("Required keys are (`id_encoder` and `lora_weights`) missing from the state dict.")
self.num_tokens =2
self.trigger_word = trigger_word
# load finetuned CLIP image encoder and fuse module here if it has not been registered to the pipeline yet
print(f"Loading PhotoMaker {pm_version} components [1] id_encoder from [{pretrained_model_name_or_path_or_dict}]...")
self.id_image_processor = CLIPImageProcessor()
if pm_version == "v1": # PhotoMaker v1
id_encoder = PhotoMakerIDEncoder()
elif pm_version == "v2": # PhotoMaker v2
id_encoder = PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken()
else:
raise NotImplementedError(f"The PhotoMaker version [{pm_version}] does not support")
id_encoder.load_state_dict(state_dict["id_encoder"], strict=True)
id_encoder = id_encoder.to(self.device, dtype=self.unet.dtype)
self.id_encoder = id_encoder
# load lora into models
print(f"Loading PhotoMaker {pm_version} components [2] lora_weights from [{pretrained_model_name_or_path_or_dict}]")
self.load_lora_weights(state_dict["lora_weights"], adapter_name="photomaker")
# Add trigger word token
if self.tokenizer is not None:
self.tokenizer.add_tokens([self.trigger_word], special_tokens=True)
self.tokenizer_2.add_tokens([self.trigger_word], special_tokens=True)
def encode_prompt_with_trigger_word(
self,
prompt: str,
prompt_2: Optional[str] = None,
device: Optional[torch.device] = None,
num_images_per_prompt: int = 1,
do_classifier_free_guidance: bool = True,
negative_prompt: Optional[str] = None,
negative_prompt_2: Optional[str] = None,
prompt_embeds: Optional[torch.Tensor] = None,
negative_prompt_embeds: Optional[torch.Tensor] = None,
pooled_prompt_embeds: Optional[torch.Tensor] = None,
negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
lora_scale: Optional[float] = None,
clip_skip: Optional[int] = None,
### Added args
num_id_images: int = 1,
class_tokens_mask: Optional[torch.LongTensor] = None,
):
device = device or self._execution_device
# set lora scale so that monkey patched LoRA
# function of text encoder can correctly access it
if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
self._lora_scale = lora_scale
# dynamically adjust the LoRA scale
if self.text_encoder is not None:
if not USE_PEFT_BACKEND:
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
else:
scale_lora_layers(self.text_encoder, lora_scale)
if self.text_encoder_2 is not None:
if not USE_PEFT_BACKEND:
adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
else:
scale_lora_layers(self.text_encoder_2, lora_scale)
prompt = [prompt] if isinstance(prompt, str) else prompt
if prompt is not None:
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
# Find the token id of the trigger word
image_token_id = self.tokenizer_2.convert_tokens_to_ids(self.trigger_word)
# Define tokenizers and text encoders
tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
text_encoders = (
[self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
)
if prompt_embeds is None:
prompt_2 = prompt_2 or prompt
prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
# textual inversion: process multi-vector tokens if necessary
prompt_embeds_list = []
prompts = [prompt, prompt_2]
for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
if isinstance(self, TextualInversionLoaderMixin):
prompt = self.maybe_convert_prompt(prompt, tokenizer)
text_inputs = tokenizer(
prompt,
padding="max_length",
max_length=tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
text_input_ids, untruncated_ids
):
removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
print(
"The following part of your input was truncated because CLIP can only handle sequences up to"
f" {tokenizer.model_max_length} tokens: {removed_text}"
)
clean_index = 0
clean_input_ids = []
class_token_index = []
# Find out the corresponding class word token based on the newly added trigger word token
for i, token_id in enumerate(text_input_ids.tolist()[0]):
if token_id == image_token_id:
class_token_index.append(clean_index - 1)
else:
clean_input_ids.append(token_id)
clean_index += 1
if len(class_token_index) != 1:
raise ValueError(
f"PhotoMaker currently does not support multiple trigger words in a single prompt.\
Trigger word: {self.trigger_word}, Prompt: {prompt}."
)
class_token_index = class_token_index[0]
# Expand the class word token and corresponding mask
class_token = clean_input_ids[class_token_index]
clean_input_ids = clean_input_ids[:class_token_index] + [class_token] * num_id_images * self.num_tokens + \
clean_input_ids[class_token_index+1:]
# Truncation or padding
max_len = tokenizer.model_max_length
if len(clean_input_ids) > max_len:
clean_input_ids = clean_input_ids[:max_len]
else:
clean_input_ids = clean_input_ids + [tokenizer.pad_token_id] * (
max_len - len(clean_input_ids)
)
class_tokens_mask = [True if class_token_index <= i < class_token_index+(num_id_images * self.num_tokens) else False \
for i in range(len(clean_input_ids))]
clean_input_ids = torch.tensor(clean_input_ids, dtype=torch.long).unsqueeze(0)
class_tokens_mask = torch.tensor(class_tokens_mask, dtype=torch.bool).unsqueeze(0)
prompt_embeds = text_encoder(clean_input_ids.to(device), output_hidden_states=True)
# We are only ALWAYS interested in the pooled output of the final text encoder
pooled_prompt_embeds = prompt_embeds[0]
if clip_skip is None:
prompt_embeds = prompt_embeds.hidden_states[-2]
else:
# "2" because SDXL always indexes from the penultimate layer.
prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
prompt_embeds_list.append(prompt_embeds)
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
class_tokens_mask = class_tokens_mask.to(device=device) # TODO: ignoring two-prompt case
# get unconditional embeddings for classifier free guidance
zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
elif do_classifier_free_guidance and negative_prompt_embeds is None:
negative_prompt = negative_prompt or ""
negative_prompt_2 = negative_prompt_2 or negative_prompt
# normalize str to list
negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
negative_prompt_2 = (
batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
)
uncond_tokens: List[str]
if prompt is not None and type(prompt) is not type(negative_prompt):
raise TypeError(
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
f" {type(prompt)}."
)
elif batch_size != len(negative_prompt):
raise ValueError(
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
" the batch size of `prompt`."
)
else:
uncond_tokens = [negative_prompt, negative_prompt_2]
negative_prompt_embeds_list = []
for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
if isinstance(self, TextualInversionLoaderMixin):
negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
max_length = prompt_embeds.shape[1]
uncond_input = tokenizer(
negative_prompt,
padding="max_length",
max_length=max_length,
truncation=True,
return_tensors="pt",
)
negative_prompt_embeds = text_encoder(
uncond_input.input_ids.to(device),
output_hidden_states=True,
)
# We are only ALWAYS interested in the pooled output of the final text encoder
negative_pooled_prompt_embeds = negative_prompt_embeds[0]
negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
negative_prompt_embeds_list.append(negative_prompt_embeds)
negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
if self.text_encoder_2 is not None:
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
else:
prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
bs_embed, seq_len, _ = prompt_embeds.shape
if do_classifier_free_guidance:
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
seq_len = negative_prompt_embeds.shape[1]
if self.text_encoder_2 is not None:
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
else:
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
if do_classifier_free_guidance:
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
bs_embed * num_images_per_prompt, -1
)
if self.text_encoder is not None:
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
# Retrieve the original scale by scaling back the LoRA layers
unscale_lora_layers(self.text_encoder, lora_scale)
if self.text_encoder_2 is not None:
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
# Retrieve the original scale by scaling back the LoRA layers
unscale_lora_layers(self.text_encoder_2, lora_scale)
return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds, class_tokens_mask
@property
def interrupt(self):
return self._interrupt
@torch.no_grad()
def __call__(
self,
prompt: Union[str, List[str]] = None,
prompt_2: Optional[Union[str, List[str]]] = None,
image: PipelineImageInput = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 50,
timesteps: List[int] = None,
sigmas: List[float] = None,
denoising_end: Optional[float] = None,
guidance_scale: float = 5.0,
negative_prompt: Optional[Union[str, List[str]]] = None,
negative_prompt_2: Optional[Union[str, List[str]]] = None,
num_images_per_prompt: Optional[int] = 1,
eta: float = 0.0,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
latents: Optional[torch.Tensor] = None,
prompt_embeds: Optional[torch.Tensor] = None,
negative_prompt_embeds: Optional[torch.Tensor] = None,
pooled_prompt_embeds: Optional[torch.Tensor] = None,
negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
ip_adapter_image: Optional[PipelineImageInput] = None,
ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
callback: Optional[Callable[[int, int, torch.Tensor], None]] = None,
callback_steps: int = 1,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
guidance_rescale: float = 0.0,
original_size: Optional[Tuple[int, int]] = None,
crops_coords_top_left: Tuple[int, int] = (0, 0),
target_size: Optional[Tuple[int, int]] = None,
negative_original_size: Optional[Tuple[int, int]] = None,
negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
negative_target_size: Optional[Tuple[int, int]] = None,
adapter_conditioning_scale: Union[float, List[float]] = 1.0,
adapter_conditioning_factor: float = 1.0,
clip_skip: Optional[int] = None,
# Added parameters (for PhotoMaker)
input_id_images: PipelineImageInput = None,
start_merge_step: int = 10, # TODO: change to `style_strength_ratio` in the future
class_tokens_mask: Optional[torch.LongTensor] = None,
id_embeds: Optional[torch.FloatTensor] = None,
prompt_embeds_text_only: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds_text_only: Optional[torch.FloatTensor] = None,
**kwargs,
):
r"""
Function invoked when calling the pipeline for generation.
Only the parameters introduced by PhotoMaker are discussed here.
For explanations of the previous parameters in StableDiffusionXLControlNetPipeline, please refer to https://github.com/huggingface/diffusers/blob/v0.25.0/src/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py
Args:
input_id_images (`PipelineImageInput`, *optional*):
Input ID Image to work with PhotoMaker.
class_tokens_mask (`torch.LongTensor`, *optional*):
Pre-generated class token. When the `prompt_embeds` parameter is provided in advance, it is necessary to prepare the `class_tokens_mask` beforehand for marking out the position of class word.
prompt_embeds_text_only (`torch.FloatTensor`, *optional*):
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
provided, text embeddings will be generated from `prompt` input argument.
pooled_prompt_embeds_text_only (`torch.FloatTensor`, *optional*):
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
If not provided, pooled text embeddings will be generated from `prompt` input argument.
Returns:
[`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
[`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
`tuple`. When returning a tuple, the first element is a list with the generated images.
"""
height, width = self._default_height_width(height, width, image)
device = self._execution_device
use_adapter = True if image is not None else False
print(f"Use adapter: {use_adapter} ｜ output size: {(height, width)}")
if use_adapter:
if isinstance(self.adapter, MultiAdapter):
adapter_input = []
for one_image in image:
one_image = _preprocess_adapter_image(one_image, height, width)
one_image = one_image.to(device=device, dtype=self.adapter.dtype)
adapter_input.append(one_image)
else:
adapter_input = _preprocess_adapter_image(image, height, width)
adapter_input = adapter_input.to(device=device, dtype=self.adapter.dtype)
original_size = original_size or (height, width)
target_size = target_size or (height, width)
# 1. Check inputs. Raise error if not correct
self.check_inputs(
prompt,
prompt_2,
height,
width,
callback_steps,
negative_prompt,
negative_prompt_2,
prompt_embeds,
negative_prompt_embeds,
pooled_prompt_embeds,
negative_pooled_prompt_embeds,
ip_adapter_image,
ip_adapter_image_embeds,
)
self._guidance_scale = guidance_scale
self._clip_skip = clip_skip
#
if prompt_embeds is not None and class_tokens_mask is None:
raise ValueError(
"If `prompt_embeds` are provided, `class_tokens_mask` also have to be passed. Make sure to generate `class_tokens_mask` from the same tokenizer that was used to generate `prompt_embeds`."
)
# check the input id images
if input_id_images is None:
raise ValueError(
"Provide `input_id_images`. Cannot leave `input_id_images` undefined for PhotoMaker pipeline."
)
if not isinstance(input_id_images, list):
input_id_images = [input_id_images]
# 2. Define call parameters
if prompt is not None and isinstance(prompt, str):
batch_size = 1
elif prompt is not None and isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
device = self._execution_device
# 3. Encode input prompt
lora_scale = (
cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
)
num_id_images = len(input_id_images)
(
prompt_embeds,
_,
pooled_prompt_embeds,
_,
class_tokens_mask,
) = self.encode_prompt_with_trigger_word(
prompt=prompt,
prompt_2=prompt_2,
device=device,
num_id_images=num_id_images,
class_tokens_mask=class_tokens_mask,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=self.do_classifier_free_guidance,
negative_prompt=negative_prompt,
negative_prompt_2=negative_prompt_2,
prompt_embeds=prompt_embeds,
negative_prompt_embeds=negative_prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds,
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
lora_scale=lora_scale,
clip_skip=self._clip_skip,
)
# 4. Encode input prompt without the trigger word for delayed conditioning
# encode, remove trigger word token, then decode
tokens_text_only = self.tokenizer.encode(prompt, add_special_tokens=False)
trigger_word_token = self.tokenizer.convert_tokens_to_ids(self.trigger_word)
tokens_text_only.remove(trigger_word_token)
prompt_text_only = self.tokenizer.decode(tokens_text_only, add_special_tokens=False)
(
prompt_embeds_text_only,
negative_prompt_embeds,
pooled_prompt_embeds_text_only, # TODO: replace the pooled_prompt_embeds with text only prompt
negative_pooled_prompt_embeds,
) = self.encode_prompt(
prompt=prompt_text_only,
prompt_2=prompt_2,
device=device,
num_images_per_prompt=num_images_per_prompt,
do_classifier_free_guidance=self.do_classifier_free_guidance,
negative_prompt=negative_prompt,
negative_prompt_2=negative_prompt_2,
prompt_embeds=prompt_embeds_text_only,
negative_prompt_embeds=negative_prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds_text_only,
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
lora_scale=lora_scale,
clip_skip=self._clip_skip,
)
# 5. Prepare the input ID images
dtype = next(self.id_encoder.parameters()).dtype
if not isinstance(input_id_images[0], torch.Tensor):
id_pixel_values = self.id_image_processor(input_id_images, return_tensors="pt").pixel_values
id_pixel_values = id_pixel_values.unsqueeze(0).to(device=device, dtype=dtype) # TODO: multiple prompts
# 6. Get the update text embedding with the stacked ID embedding
if id_embeds is not None:
id_embeds = id_embeds.unsqueeze(0).to(device=device, dtype=dtype)
prompt_embeds = self.id_encoder(id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds)
else:
prompt_embeds = self.id_encoder(id_pixel_values, prompt_embeds, class_tokens_mask)
bs_embed, seq_len, _ = prompt_embeds.shape
# duplicate text embeddings for each generation per prompt, using mps friendly method
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
# 6.1 Get the ip adapter embedding
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
image_embeds = self.prepare_ip_adapter_image_embeds(
ip_adapter_image,
ip_adapter_image_embeds,
device,
batch_size * num_images_per_prompt,
self.do_classifier_free_guidance,
)
# 7. Prepare timesteps
timesteps, num_inference_steps = retrieve_timesteps(
self.scheduler, num_inference_steps, device, timesteps, sigmas
)
# 8. Prepare latent variables
num_channels_latents = self.unet.config.in_channels
latents = self.prepare_latents(
batch_size * num_images_per_prompt,
num_channels_latents,
height,
width,
prompt_embeds.dtype,
device,
generator,
latents,
)
# 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
# 8.5 Optionally get Guidance Scale Embedding
timestep_cond = None
if self.unet.config.time_cond_proj_dim is not None:
guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
timestep_cond = self.get_guidance_scale_embedding(
guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
).to(device=device, dtype=latents.dtype)
# 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
if use_adapter:
if isinstance(self.adapter, MultiAdapter):
adapter_state = self.adapter(adapter_input, adapter_conditioning_scale)
for k, v in enumerate(adapter_state):
adapter_state[k] = v
else:
adapter_state = self.adapter(adapter_input)
for k, v in enumerate(adapter_state):
adapter_state[k] = v * adapter_conditioning_scale
if num_images_per_prompt > 1:
for k, v in enumerate(adapter_state):
adapter_state[k] = v.repeat(num_images_per_prompt, 1, 1, 1)
if self.do_classifier_free_guidance:
for k, v in enumerate(adapter_state):
adapter_state[k] = torch.cat([v] * 2, dim=0)
add_text_embeds = pooled_prompt_embeds
if self.text_encoder_2 is None:
text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
else:
text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
add_time_ids = self._get_add_time_ids(
original_size,
crops_coords_top_left,
target_size,
dtype=prompt_embeds.dtype,
text_encoder_projection_dim=text_encoder_projection_dim,
)
if negative_original_size is not None and negative_target_size is not None:
negative_add_time_ids = self._get_add_time_ids(
negative_original_size,
negative_crops_coords_top_left,
negative_target_size,
dtype=prompt_embeds.dtype,
text_encoder_projection_dim=text_encoder_projection_dim,
)
else:
negative_add_time_ids = add_time_ids
if self.do_classifier_free_guidance:
add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
prompt_embeds = prompt_embeds.to(device)
add_text_embeds = add_text_embeds.to(device)
add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
# 11. Denoising loop
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
# Apply denoising_end
if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
discrete_timestep_cutoff = int(
round(
self.scheduler.config.num_train_timesteps
- (denoising_end * self.scheduler.config.num_train_timesteps)
)
)
num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
timesteps = timesteps[:num_inference_steps]
with self.progress_bar(total=num_inference_steps) as progress_bar:
for i, t in enumerate(timesteps):
# expand the latents if we are doing classifier free guidance
latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
if i <= start_merge_step:
current_prompt_embeds = torch.cat(
[negative_prompt_embeds, prompt_embeds_text_only], dim=0
) if self.do_classifier_free_guidance else prompt_embeds_text_only
add_text_embeds = torch.cat(
[negative_pooled_prompt_embeds, pooled_prompt_embeds_text_only], dim=0
) if self.do_classifier_free_guidance else pooled_prompt_embeds_text_only
else:
current_prompt_embeds = torch.cat(
[negative_prompt_embeds, prompt_embeds], dim=0
) if self.do_classifier_free_guidance else prompt_embeds
add_text_embeds = torch.cat(
[negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0
) if self.do_classifier_free_guidance else pooled_prompt_embeds
if i < int(num_inference_steps * adapter_conditioning_factor) and (use_adapter):
down_intrablock_additional_residuals = [state.clone() for state in adapter_state]
else:
down_intrablock_additional_residuals = None
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
added_cond_kwargs["image_embeds"] = image_embeds
# predict the noise residual
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=current_prompt_embeds,
timestep_cond=timestep_cond,
cross_attention_kwargs=cross_attention_kwargs,
down_intrablock_additional_residuals=down_intrablock_additional_residuals,
added_cond_kwargs=added_cond_kwargs,
return_dict=False,
)[0]
# perform guidance
if self.do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
if self.do_classifier_free_guidance and guidance_rescale > 0.0:
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
# compute the previous noisy sample x_t -> x_t-1
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
# call the callback, if provided
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
progress_bar.update()
if callback is not None and i % callback_steps == 0:
step_idx = i // getattr(self.scheduler, "order", 1)
callback(step_idx, t, latents)
if not output_type == "latent":
# make sure the VAE is in float32 mode, as it overflows in float16
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
if needs_upcasting:
self.upcast_vae()
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
# cast back to fp16 if needed
if needs_upcasting:
self.vae.to(dtype=torch.float16)
else:
image = latents
return StableDiffusionXLPipelineOutput(images=image)
image = self.image_processor.postprocess(image, output_type=output_type)
# Offload all models
self.maybe_free_model_hooks()
if not return_dict:
return (image,)
return StableDiffusionXLPipelineOutput(images=image)
\ No newline at end of file
photomaker/resampler.py 0 → 100644
View file @ 48ec26d1
#### Borrowed from https://github.com/tencent-ailab/IP-Adapter/blob/main/ip_adapter/resampler.py
# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py
import math
import torch
import torch.nn as nn
from einops import rearrange
from einops.layers.torch import Rearrange
class FacePerceiverResampler(torch.nn.Module):
def __init__(
self,
*,
dim=768,
depth=4,
dim_head=64,
heads=16,
embedding_dim=1280,
output_dim=768,
ff_mult=4,
):
super().__init__()
self.proj_in = torch.nn.Linear(embedding_dim, dim)
self.proj_out = torch.nn.Linear(dim, output_dim)
self.norm_out = torch.nn.LayerNorm(output_dim)
self.layers = torch.nn.ModuleList([])
for _ in range(depth):
self.layers.append(
torch.nn.ModuleList(
[
PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
FeedForward(dim=dim, mult=ff_mult),
]
)
)
def forward(self, latents, x):
x = self.proj_in(x)
for attn, ff in self.layers:
latents = attn(x, latents) + latents
latents = ff(latents) + latents
latents = self.proj_out(latents)
return self.norm_out(latents)
# FFN
def FeedForward(dim, mult=4):
inner_dim = int(dim * mult)
return nn.Sequential(
nn.LayerNorm(dim),
nn.Linear(dim, inner_dim, bias=False),
nn.GELU(),
nn.Linear(inner_dim, dim, bias=False),
)
def reshape_tensor(x, heads):
bs, length, width = x.shape
# (bs, length, width) --> (bs, length, n_heads, dim_per_head)
x = x.view(bs, length, heads, -1)
# (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
x = x.transpose(1, 2)
# (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
x = x.reshape(bs, heads, length, -1)
return x
class PerceiverAttention(nn.Module):
def __init__(self, *, dim, dim_head=64, heads=8):
super().__init__()
self.scale = dim_head**-0.5
self.dim_head = dim_head
self.heads = heads
inner_dim = dim_head * heads
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
self.to_q = nn.Linear(dim, inner_dim, bias=False)
self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
self.to_out = nn.Linear(inner_dim, dim, bias=False)
def forward(self, x, latents):
"""
Args:
x (torch.Tensor): image features
shape (b, n1, D)
latent (torch.Tensor): latent features
shape (b, n2, D)
"""
x = self.norm1(x)
latents = self.norm2(latents)
b, l, _ = latents.shape
q = self.to_q(latents)
kv_input = torch.cat((x, latents), dim=-2)
k, v = self.to_kv(kv_input).chunk(2, dim=-1)
q = reshape_tensor(q, self.heads)
k = reshape_tensor(k, self.heads)
v = reshape_tensor(v, self.heads)
# attention
scale = 1 / math.sqrt(math.sqrt(self.dim_head))
weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards
weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
out = weight @ v
out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
return self.to_out(out)
class Resampler(nn.Module):
def __init__(
self,
dim=1024,
depth=8,
dim_head=64,
heads=16,
num_queries=8,
embedding_dim=768,
output_dim=1024,
ff_mult=4,
max_seq_len: int = 257, # CLIP tokens + CLS token
apply_pos_emb: bool = False,
num_latents_mean_pooled: int = 0, # number of latents derived from mean pooled representation of the sequence
):
super().__init__()
self.pos_emb = nn.Embedding(max_seq_len, embedding_dim) if apply_pos_emb else None
self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
self.proj_in = nn.Linear(embedding_dim, dim)
self.proj_out = nn.Linear(dim, output_dim)
self.norm_out = nn.LayerNorm(output_dim)
self.to_latents_from_mean_pooled_seq = (
nn.Sequential(
nn.LayerNorm(dim),
nn.Linear(dim, dim * num_latents_mean_pooled),
Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled),
)
if num_latents_mean_pooled > 0
else None
)
self.layers = nn.ModuleList([])
for _ in range(depth):
self.layers.append(
nn.ModuleList(
[
PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
FeedForward(dim=dim, mult=ff_mult),
]
)
)
def forward(self, x):
if self.pos_emb is not None:
n, device = x.shape[1], x.device
pos_emb = self.pos_emb(torch.arange(n, device=device))
x = x + pos_emb
latents = self.latents.repeat(x.size(0), 1, 1)
x = self.proj_in(x)
if self.to_latents_from_mean_pooled_seq:
meanpooled_seq = masked_mean(x, dim=1, mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool))
meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq)
latents = torch.cat((meanpooled_latents, latents), dim=-2)
for attn, ff in self.layers:
latents = attn(x, latents) + latents
latents = ff(latents) + latents
latents = self.proj_out(latents)
return self.norm_out(latents)
def masked_mean(t, *, dim, mask=None):
if mask is None:
return t.mean(dim=dim)
denom = mask.sum(dim=dim, keepdim=True)
mask = rearrange(mask, "b n -> b n 1")
masked_t = t.masked_fill(~mask, 0.0)
return masked_t.sum(dim=dim) / denom.clamp(min=1e-5)
pyproject.toml
View file @ 48ec26d1
[tool.poetry] [tool.poetry]
name = "photomaker" name = "photomaker"
version = "0.1.0" version = "0.2.0"
description = "PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding" description = "PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding"
authors = ["Li, Zhen", "Cao, Mingdeng", "Wang, Xintao", "Qi, Zhongang", "Cheng, Ming-Ming", "Shan, Ying"] authors = ["Li, Zhen", "Cao, Mingdeng", "Wang, Xintao", "Qi, Zhongang", "Cheng, Ming-Ming", "Shan, Ying"]
license = "Apache-2.0" license = "Apache-2.0"
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment