Skip to content

GitLab

"backends/v3/src/client/grpc_client.rs" did not exist on "f9910d13e296989f41e714c43eb60ce051359db3"
Unverified Commit 73ba8109 authored by Álvaro Somoza's avatar Álvaro Somoza Committed by GitHub
Browse files

[Community pipeline] SDXL Differential Diffusion Img2Img Pipeline (#7550) 

* initial-commit pipeline created

* updated README.md
parent 7956c36a
Show whitespace changes
Inline Side-by-side
Showing with 1703 additions and 66 deletions
examples/community/README.md
View file @ 73ba8109
......@@ -10,11 +10,12 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
| Example | Description | Code Example | Colab | Author |
|:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
|Differential Diffusion|[Differential Diffusion](https://github.com/exx8/differential-diffusion) modifies an image according to a text prompt, and according to a map that specifies the amount of change in each region.|[Differential Diffusion](#differential-diffusion)|[![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/exx8/differential-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/exx8/differential-diffusion/blob/main/examples/SD2.ipynb)|[Eran Levin](https://github.com/exx8) and [Ohad Fried](https://www.ohadf.com/)|
| HD-Painter | [HD-Painter](https://github.com/Picsart-AI-Research/HD-Painter) enables prompt-faithfull and high resolution (up to 2k) image inpainting upon any diffusion-based image inpainting method. | [HD-Painter](#hd-painter) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/PAIR/HD-Painter) | [Manukyan Hayk](https://github.com/haikmanukyan) and [Sargsyan Andranik](https://github.com/AndranikSargsyan) |
| Marigold Monocular Depth Estimation | A universal monocular depth estimator, utilizing Stable Diffusion, delivering sharp predictions in the wild. (See the [project page](https://marigoldmonodepth.github.io) and [full codebase](https://github.com/prs-eth/marigold) for more details.) | [Marigold Depth Estimation](#marigold-depth-estimation) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/toshas/marigold) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/12G8reD13DdpMie5ZQlaFNo2WCGeNUH-u?usp=sharing) | [Bingxin Ke](https://github.com/markkua) and [Anton Obukhov](https://github.com/toshas) |
| LLM-grounded Diffusion (LMD+) | LMD greatly improves the prompt following ability of text-to-image generation models by introducing an LLM as a front-end prompt parser and layout planner. [Project page.](https://llm-grounded-diffusion.github.io/) [See our full codebase (also with diffusers).](https://github.com/TonyLianLong/LLM-groundedDiffusion) | [LLM-grounded Diffusion (LMD+)](#llm-grounded-diffusion) | [Huggingface Demo](https://huggingface.co/spaces/longlian/llm-grounded-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1SXzMSeAB-LJYISb2yrUOdypLz4OYWUKj) | [Long (Tony) Lian](https://tonylian.com/) |
| CLIP Guided Stable Diffusion | Doing CLIP guidance for text to image generation with Stable Diffusion | [CLIP Guided Stable Diffusion](#clip-guided-stable-diffusion) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/CLIP_Guided_Stable_diffusion_with_diffusers.ipynb) | [Suraj Patil](https://github.com/patil-suraj/) |
| One Step U-Net (Dummy) | Example showcasing of how to use Community Pipelines (see https://github.com/huggingface/diffusers/issues/841) | [One Step U-Net](#one-step-unet) | - | [Patrick von Platen](https://github.com/patrickvonplaten/) |
| One Step U-Net (Dummy) | Example showcasing of how to use Community Pipelines (see <https://github.com/huggingface/diffusers/issues/841>) | [One Step U-Net](#one-step-unet) | - | [Patrick von Platen](https://github.com/patrickvonplaten/) |
| Stable Diffusion Interpolation | Interpolate the latent space of Stable Diffusion between different prompts/seeds | [Stable Diffusion Interpolation](#stable-diffusion-interpolation) | - | [Nate Raw](https://github.com/nateraw/) |
| Stable Diffusion Mega | **One** Stable Diffusion Pipeline with all functionalities of [Text2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py), [Image2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py) and [Inpainting](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) | [Stable Diffusion Mega](#stable-diffusion-mega) | - | [Patrick von Platen](https://github.com/patrickvonplaten/) |
| Long Prompt Weighting Stable Diffusion | **One** Stable Diffusion Pipeline without tokens length limit, and support parsing weighting in prompt. | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion) | - | [SkyTNT](https://github.com/SkyTNT) |
......@@ -45,7 +46,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
| CLIP Guided Images Mixing Stable Diffusion Pipeline | Сombine images using usual diffusion models. | [CLIP Guided Images Mixing Using Stable Diffusion](#clip-guided-images-mixing-with-stable-diffusion) | - | [Karachev Denis](https://github.com/TheDenk) |
| TensorRT Stable Diffusion Inpainting Pipeline | Accelerates the Stable Diffusion Inpainting Pipeline using TensorRT | [TensorRT Stable Diffusion Inpainting Pipeline](#tensorrt-inpainting-stable-diffusion-pipeline) | - | [Asfiya Baig](https://github.com/asfiyab-nvidia) |
| IADB Pipeline | Implementation of [Iterative α-(de)Blending: a Minimalist Deterministic Diffusion Model](https://arxiv.org/abs/2305.03486) | [IADB Pipeline](#iadb-pipeline) | - | [Thomas Chambon](https://github.com/tchambon)
| Zero1to3 Pipeline | Implementation of [Zero-1-to-3: Zero-shot One Image to 3D Object](https://arxiv.org/abs/2303.11328) | [Zero1to3 Pipeline](#Zero1to3-pipeline) | - | [Xin Kong](https://github.com/kxhit) |
| Zero1to3 Pipeline | Implementation of [Zero-1-to-3: Zero-shot One Image to 3D Object](https://arxiv.org/abs/2303.11328) | [Zero1to3 Pipeline](#zero1to3-pipeline) | - | [Xin Kong](https://github.com/kxhit) |
| Stable Diffusion XL Long Weighted Prompt Pipeline | A pipeline support unlimited length of prompt and negative prompt, use A1111 style of prompt weighting | [Stable Diffusion XL Long Weighted Prompt Pipeline](#stable-diffusion-xl-long-weighted-prompt-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1LsqilswLR40XLLcp6XFOl5nKb_wOe26W?usp=sharing) | [Andrew Zhu](https://xhinker.medium.com/) |
| FABRIC - Stable Diffusion with feedback Pipeline | pipeline supports feedback from liked and disliked images | [Stable Diffusion Fabric Pipeline](#stable-diffusion-fabric-pipeline) | - | [Shauray Singh](https://shauray8.github.io/about_shauray/) |
| sketch inpaint - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion Pipeline](#stable-diffusion-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
......@@ -57,10 +58,10 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
| Regional Prompting Pipeline | Assign multiple prompts for different regions | [Regional Prompting Pipeline](#regional-prompting-pipeline) | - | [hako-mikan](https://github.com/hako-mikan) |
| LDM3D-sr (LDM3D upscaler) | Upscale low resolution RGB and depth inputs to high resolution | [StableDiffusionUpscaleLDM3D Pipeline](https://github.com/estelleafl/diffusers/tree/ldm3d_upscaler_community/examples/community#stablediffusionupscaleldm3d-pipeline) | - | [Estelle Aflalo](https://github.com/estelleafl) |
| AnimateDiff ControlNet Pipeline | Combines AnimateDiff with precise motion control using ControlNets | [AnimateDiff ControlNet Pipeline](#animatediff-controlnet-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1SKboYeGjEQmQPWoFC0aLYpBlYdHXkvAu?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) and [Edoardo Botta](https://github.com/EdoardoBotta) |
| DemoFusion Pipeline | Implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973) | [DemoFusion Pipeline](#DemoFusion) | - | [Ruoyi Du](https://github.com/RuoyiDu) |
| DemoFusion Pipeline | Implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973) | [DemoFusion Pipeline](#demofusion) | - | [Ruoyi Du](https://github.com/RuoyiDu) |
| Instaflow Pipeline | Implementation of [InstaFlow! One-Step Stable Diffusion with Rectified Flow](https://arxiv.org/abs/2309.06380) | [Instaflow Pipeline](#instaflow-pipeline) | - | [Ayush Mangal](https://github.com/ayushtues) |
| Null-Text Inversion Pipeline | Implement [Null-text Inversion for Editing Real Images using Guided Diffusion Models](https://arxiv.org/abs/2211.09794) as a pipeline. | [Null-Text Inversion](https://github.com/google/prompt-to-prompt/) | - | [Junsheng Luan](https://github.com/Junsheng121) |
| Rerender A Video Pipeline | Implementation of [[SIGGRAPH Asia 2023] Rerender A Video: Zero-Shot Text-Guided Video-to-Video Translation](https://arxiv.org/abs/2306.07954) | [Rerender A Video Pipeline](#Rerender-A-Video) | - | [Yifan Zhou](https://github.com/SingleZombie) |
| Rerender A Video Pipeline | Implementation of [[SIGGRAPH Asia 2023] Rerender A Video: Zero-Shot Text-Guided Video-to-Video Translation](https://arxiv.org/abs/2306.07954) | [Rerender A Video Pipeline](#rerender-a-video) | - | [Yifan Zhou](https://github.com/SingleZombie) |
| StyleAligned Pipeline | Implementation of [Style Aligned Image Generation via Shared Attention](https://arxiv.org/abs/2312.02133) | [StyleAligned Pipeline](#stylealigned-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://drive.google.com/file/d/15X2E0jFPTajUIjS0FzX50OaHsCbP2lQ0/view?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
| AnimateDiff Image-To-Video Pipeline | Experimental Image-To-Video support for AnimateDiff (open to improvements) | [AnimateDiff Image To Video Pipeline](#animatediff-image-to-video-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://drive.google.com/file/d/1TvzCDPHhfFtdcJZe4RLloAwyoLKuttWK/view?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
| IP Adapter FaceID Stable Diffusion | Stable Diffusion Pipeline that supports IP Adapter Face ID | [IP Adapter Face ID](#ip-adapter-face-id) | - | [Fabio Rigano](https://github.com/fabiorigano) |
......@@ -76,6 +77,83 @@ pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custo
## Example usages
### Differential Diffusion
**Eran Levin, Ohad Fried**
**Tel Aviv University, Reichman University**
Diffusion models have revolutionized image generation and editing, producing state-of-the-art results in conditioned and unconditioned image synthesis. While current techniques enable user control over the degree of change in an image edit, the controllability is limited to global changes over an entire edited region. This paper introduces a novel framework that enables customization of the amount of change per pixel or per image region. Our framework can be integrated into any existing diffusion model, enhancing it with this capability. Such granular control on the quantity of change opens up a diverse array of new editing capabilities, such as control of the extent to which individual objects are modified, or the ability to introduce gradual spatial changes. Furthermore, we showcase the framework's effectiveness in soft-inpainting---the completion of portions of an image while subtly adjusting the surrounding areas to ensure seamless integration. Additionally, we introduce a new tool for exploring the effects of different change quantities. Our framework operates solely during inference, requiring no model training or fine-tuning. We demonstrate our method with the current open state-of-the-art models, and validate it via both quantitative and qualitative comparisons, and a user study.
![teaser-img](https://github.com/exx8/differential-diffusion/raw/main/assets/teaser.png)
You can find additional information about Differential Diffusion in the [paper](https://differential-diffusion.github.io/paper.pdf) or in the [project website](https://differential-diffusion.github.io/).
#### Usage example
```python
import torch
from torchvision import transforms
from diffusers import DPMSolverMultistepScheduler
from diffusers.utils import load_image
from examples.community.pipeline_stable_diffusion_xl_differential_img2img import (
StableDiffusionXLDifferentialImg2ImgPipeline,
)
pipeline = StableDiffusionXLDifferentialImg2ImgPipeline.from_pretrained(
"SG161222/RealVisXL_V4.0", torch_dtype=torch.float16, variant="fp16"
).to("cuda")
pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, use_karras_sigmas=True)
def preprocess_image(image):
image = image.convert("RGB")
image = transforms.CenterCrop((image.size[1] // 64 * 64, image.size[0] // 64 * 64))(image)
image = transforms.ToTensor()(image)
image = image * 2 - 1
image = image.unsqueeze(0).to("cuda")
return image
def preprocess_map(map):
map = map.convert("L")
map = transforms.CenterCrop((map.size[1] // 64 * 64, map.size[0] // 64 * 64))(map)
map = transforms.ToTensor()(map)
map = map.to("cuda")
return map
image = preprocess_image(
load_image(
"https://huggingface.co/datasets/OzzyGT/testing-resources/resolve/main/differential/20240329211129_4024911930.png?download=true"
)
)
mask = preprocess_map(
load_image(
"https://huggingface.co/datasets/OzzyGT/testing-resources/resolve/main/differential/gradient_mask.png?download=true"
)
)
prompt = "a green pear"
negative_prompt = "blurry"
image = pipeline(
prompt=prompt,
negative_prompt=negative_prompt,
guidance_scale=7.5,
num_inference_steps=25,
original_image=image,
image=image,
strength=1.0,
map=mask,
).images[0]
image.save("result.png")
```
### HD-Painter
Implementation of [HD-Painter: High-Resolution and Prompt-Faithful Text-Guided Image Inpainting with Diffusion Models](https://arxiv.org/abs/2312.14091).
......@@ -197,6 +275,7 @@ This pipeline can be used with an LLM or on its own. We provide a parser that pa
The following code has been tested on 1x RTX 4090, but it should also support GPUs with lower GPU memory.
#### Use this pipeline with an LLM
```python
import torch
from diffusers import DiffusionPipeline
......@@ -232,6 +311,7 @@ images[0].save("./lmd_plus_generation.jpg")
```
#### Use this pipeline on its own for layout generation
```python
import torch
from diffusers import DiffusionPipeline
......@@ -327,7 +407,7 @@ pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeli
pipe()
```
**Note**: This community pipeline is not useful as a feature, but rather just serves as an example of how community pipelines can be added (see https://github.com/huggingface/diffusers/issues/841).
**Note**: This community pipeline is not useful as a feature, but rather just serves as an example of how community pipelines can be added (see <https://github.com/huggingface/diffusers/issues/841>).
### Stable Diffusion Interpolation
......@@ -361,7 +441,7 @@ frame_filepaths = pipe.walk(
The output of the `walk(...)` function returns a list of images saved under the folder as defined in `output_dir`. You can use these images to create videos of stable diffusion.
> **Please have a look at https://github.com/nateraw/stable-diffusion-videos for more in-detail information on how to create videos using stable diffusion as well as more feature-complete functionality.**
> **Please have a look at <https://github.com/nateraw/stable-diffusion-videos> for more in-detail information on how to create videos using stable diffusion as well as more feature-complete functionality.**
### Stable Diffusion Mega
......@@ -411,7 +491,9 @@ images = pipe.inpaint(prompt=prompt, image=init_image, mask_image=mask_image, st
As shown above this one pipeline can run all both "text-to-image", "image-to-image", and "inpainting" in one pipeline.
### Long Prompt Weighting Stable Diffusion
Features of this custom pipeline:
- Input a prompt without the 77 token length limit.
- Includes tx2img, img2img. and inpainting pipelines.
- Emphasize/weigh part of your prompt with parentheses as so: `a baby deer with (big eyes)`
......@@ -419,6 +501,7 @@ Features of this custom pipeline:
- Precisely weigh part of your prompt as so: `a baby deer with (big eyes:1.3)`
Prompt weighting equivalents:
- `a baby deer with` == `(a baby deer with:1.0)`
- `(big eyes)` == `(big eyes:1.1)`
- `((big eyes))` == `(big eyes:1.21)`
......@@ -512,12 +595,14 @@ diffuser_pipeline = diffuser_pipeline.to(device)
output = diffuser_pipeline(speech_data)
plt.imshow(output.images[0])
```
This example produces the following image:
![image](https://user-images.githubusercontent.com/45072645/196901736-77d9c6fc-63ee-4072-90b0-dc8b903d63e3.png)
### Wildcard Stable Diffusion
Following the great examples from https://github.com/jtkelm2/stable-diffusion-webui-1/blob/master/scripts/wildcards.py and https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#wildcards, here's a minimal implementation that allows for users to add "wildcards", denoted by `__wildcard__` to prompts that are used as placeholders for randomly sampled values given by either a dictionary or a `.txt` file. For example:
Following the great examples from <https://github.com/jtkelm2/stable-diffusion-webui-1/blob/master/scripts/wildcards.py> and <https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#wildcards>, here's a minimal implementation that allows for users to add "wildcards", denoted by `__wildcard__` to prompts that are used as placeholders for randomly sampled values given by either a dictionary or a `.txt` file. For example:
Say we have a prompt:
......@@ -624,6 +709,7 @@ tvu.save_image(grid, f'{prompt}_{args.weights}' + '.png')
```
### Imagic Stable Diffusion
Allows you to edit an image using stable diffusion.
```python
......@@ -665,6 +751,7 @@ image.save('./imagic/imagic_image_alpha_2.png')
```
### Seed Resizing
Test seed resizing. Originally generate an image in 512 by 512, then generate image with same seed at 512 by 592 using seed resizing. Finally, generate 512 by 592 using original stable diffusion pipeline.
```python
......@@ -815,6 +902,7 @@ This example produces the following images:
![image](https://user-images.githubusercontent.com/4313860/198328706-295824a4-9856-4ce5-8e66-278ceb42fd29.png)
### GlueGen Stable Diffusion Pipeline
GlueGen is a minimal adapter that allow alignment between any encoder (Text Encoder of different language, Multilingual Roberta, AudioClip) and CLIP text encoder used in standard Stable Diffusion model. This method allows easy language adaptation to available english Stable Diffusion checkpoints without the need of an image captioning dataset as well as long training hours.
Make sure you downloaded `gluenet_French_clip_overnorm_over3_noln.ckpt` for French (there are also pre-trained weights for Chinese, Italian, Japanese, Spanish or train your own) at [GlueGen's official repo](https://github.com/salesforce/GlueGen/tree/main)
......@@ -853,6 +941,7 @@ if __name__ == "__main__":
image = pipeline(prompt, generator=generator).images[0]
image.save("gluegen_output_fr.png")
```
Which will produce:
![output_image](https://github.com/rootonchair/diffusers/assets/23548268/db43ffb6-8667-47c1-8872-26f85dc0a57f)
......@@ -927,7 +1016,8 @@ image = pipe(image=image, text=text, prompt=prompt).images[0]
```
### Bit Diffusion
Based https://arxiv.org/abs/2208.04202, this is used for diffusion on discrete data - eg, discreate image data, DNA sequence data. An unconditional discreate image can be generated like this:
Based <https://arxiv.org/abs/2208.04202>, this is used for diffusion on discrete data - eg, discreate image data, DNA sequence data. An unconditional discreate image can be generated like this:
```python
from diffusers import DiffusionPipeline
......@@ -938,7 +1028,7 @@ image = pipe().images[0]
### Stable Diffusion with K Diffusion
Make sure you have @crowsonkb's https://github.com/crowsonkb/k-diffusion installed:
Make sure you have @crowsonkb's <https://github.com/crowsonkb/k-diffusion> installed:
```
pip install k-diffusion
......@@ -963,6 +1053,7 @@ image.save("./astronaut_heun_k_diffusion.png")
To make sure that K Diffusion and `diffusers` yield the same results:
**Diffusers**:
```python
from diffusers import DiffusionPipeline, EulerDiscreteScheduler
......@@ -979,6 +1070,7 @@ image = pipe(prompt, generator=generator, num_inference_steps=50).images[0]
![diffusers_euler](https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/k_diffusion/astronaut_euler.png)
**K Diffusion**:
```python
from diffusers import DiffusionPipeline, EulerDiscreteScheduler
......@@ -996,12 +1088,14 @@ image = pipe(prompt, generator=generator, num_inference_steps=50).images[0]
![diffusers_euler](https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/k_diffusion/astronaut_euler_k_diffusion.png)
### Checkpoint Merger Pipeline
Based on the AUTOMATIC1111/webui for checkpoint merging. This is a custom pipeline that merges upto 3 pretrained model checkpoints as long as they are in the HuggingFace model_index.json format.
The checkpoint merging is currently memory intensive as it modifies the weights of a DiffusionPipeline object in place. Expect at least 13GB RAM Usage on Kaggle GPU kernels and
on colab you might run out of the 12GB memory even while merging two checkpoints.
Usage:-
```python
from diffusers import DiffusionPipeline
......@@ -1027,6 +1121,7 @@ prompt = "An astronaut riding a horse on Mars"
image = merged_pipe(prompt).images[0]
```
Some examples along with the merge details:
1. "CompVis/stable-diffusion-v1-4" + "hakurei/waifu-diffusion" ; Sigmoid interpolation; alpha = 0.8
......@@ -1037,15 +1132,14 @@ Some examples along with the merge details:
![Stable plus Waifu Sigmoid 0.8](https://huggingface.co/datasets/NagaSaiAbhinay/CheckpointMergerSamples/resolve/main/waifu_openjourney_inv_sig_0.8.png)
3. "CompVis/stable-diffusion-v1-4" + "hakurei/waifu-diffusion" + "prompthero/openjourney"; Add Difference interpolation; alpha = 0.5
![Stable plus Waifu plus openjourney add_diff 0.5](https://huggingface.co/datasets/NagaSaiAbhinay/CheckpointMergerSamples/resolve/main/stable_waifu_openjourney_add_diff_0.5.png)
### Stable Diffusion Comparisons
This Community Pipeline enables the comparison between the 4 checkpoints that exist for Stable Diffusion. They can be found through the following links:
1. [Stable Diffusion v1.1](https://huggingface.co/CompVis/stable-diffusion-v1-1)
2. [Stable Diffusion v1.2](https://huggingface.co/CompVis/stable-diffusion-v1-2)
3. [Stable Diffusion v1.3](https://huggingface.co/CompVis/stable-diffusion-v1-3)
......@@ -1088,6 +1182,7 @@ As a result, you can look at a grid of all 4 generated images being shown togeth
Implementation of the [MagicMix: Semantic Mixing with Diffusion Models](https://arxiv.org/abs/2210.16056) paper. This is a Diffusion Pipeline for semantic mixing of an image and a text prompt to create a new concept while preserving the spatial layout and geometry of the subject in the image. The pipeline takes an image that provides the layout semantics and a prompt that provides the content semantics for the mixing process.
There are 3 parameters for the method-
- `mix_factor`: It is the interpolation constant used in the layout generation phase. The greater the value of `mix_factor`, the greater the influence of the prompt on the layout generation process.
- `kmax` and `kmin`: These determine the range for the layout and content generation process. A higher value of kmax results in loss of more information about the layout of the original image and a higher value of kmin results in more steps for content generation process.
......@@ -1113,6 +1208,7 @@ mix_img = pipe(
)
mix_img.save('phone_bed_mix.jpg')
```
The `mix_img` is a PIL image that can be saved locally or displayed directly in a google colab. Generated image is a mix of the layout semantics of the given image and the content semantics of the prompt.
E.g. the above script generates the following image:
......@@ -1127,7 +1223,6 @@ E.g. the above script generates the following image:
For more example generations check out this [demo notebook](https://github.com/daspartho/MagicMix/blob/main/demo.ipynb).
### Stable UnCLIP
UnCLIPPipeline("kakaobrain/karlo-v1-alpha") provide a prior model that can generate clip image embedding from text.
......@@ -1209,10 +1304,8 @@ print(pipeline.prior_scheduler)
# }
```
`shiba-inu.jpg`
![shiba-inu](https://user-images.githubusercontent.com/16448529/209185639-6e5ec794-ce9d-4883-aa29-bd6852a2abad.jpg)
### UnCLIP Text Interpolation Pipeline
......@@ -1280,6 +1373,7 @@ output = pipe(image = images ,steps = 6, generator = generator)
for i,image in enumerate(output.images):
image.save('starry_to_flowers_%s.jpg' % i)
```
The original images:-
![starry](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/starry_night.jpg)
......@@ -1295,7 +1389,9 @@ The resulting images in order:-
![result5](https://huggingface.co/datasets/NagaSaiAbhinay/UnCLIPImageInterpolationSamples/resolve/main/starry_to_flowers_5.png)
### DDIM Noise Comparative Analysis Pipeline
#### **Research question: What visual concepts do the diffusion models learn from each noise level during training?**
The [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227) paper proposed an approach to answer the above question, which is their second contribution.
The approach consists of the following steps:
......@@ -1425,6 +1521,7 @@ image.save('tensorrt_mt_fuji.png')
### EDICT Image Editing Pipeline
This pipeline implements the text-guided image editing approach from the paper [EDICT: Exact Diffusion Inversion via Coupled Transformations](https://arxiv.org/abs/2211.12446). You have to pass:
- (`PIL`) `image` you want to edit.
- `base_prompt`: the text prompt describing the current image (before editing).
- `target_prompt`: the text prompt describing with the edits.
......@@ -1584,6 +1681,7 @@ image.save('tensorrt_img2img_new_zealand_hills.png')
This pipeline uses the Reference Control. Refer to the [sd-webui-controlnet discussion: Reference-only Control](https://github.com/Mikubill/sd-webui-controlnet/discussions/1236)[sd-webui-controlnet discussion: Reference-adain Control](https://github.com/Mikubill/sd-webui-controlnet/discussions/1280).
Based on [this issue](https://github.com/huggingface/diffusers/issues/3566),
- `EulerAncestralDiscreteScheduler` got poor results.
```py
......@@ -1629,6 +1727,7 @@ Output Image of `reference_attn=True` and `reference_adain=True`
This pipeline uses the Reference Control with ControlNet. Refer to the [sd-webui-controlnet discussion: Reference-only Control](https://github.com/Mikubill/sd-webui-controlnet/discussions/1236)[sd-webui-controlnet discussion: Reference-adain Control](https://github.com/Mikubill/sd-webui-controlnet/discussions/1280).
Based on [this issue](https://github.com/huggingface/diffusers/issues/3566),
- `EulerAncestralDiscreteScheduler` got poor results.
- `guess_mode=True` works well for ControlNet v1.1
......@@ -1674,12 +1773,12 @@ Output Image
![output_image](https://github.com/huggingface/diffusers/assets/24734142/7b9a5830-f173-4b92-b0cf-73d0e9c01d60)
### Stable Diffusion on IPEX
This diffusion pipeline aims to accelarate the inference of Stable-Diffusion on Intel Xeon CPUs with BF16/FP32 precision using [IPEX](https://github.com/intel/intel-extension-for-pytorch).
To use this pipeline, you need to:
1. Install [IPEX](https://github.com/intel/intel-extension-for-pytorch)
**Note:** For each PyTorch release, there is a corresponding release of the IPEX. Here is the mapping relationship. It is recommended to install Pytorch/IPEX2.0 to get the best performance.
......@@ -1690,10 +1789,13 @@ To use this pipeline, you need to:
|[v1.13.\*](https://github.com/pytorch/pytorch/tree/v1.13.0 "v1.13.0")|[v1.13.\*](https://github.com/intel/intel-extension-for-pytorch/tree/v1.13.100+cpu)|
You can simply use pip to install IPEX with the latest version.
```python
python -m pip install intel_extension_for_pytorch
```
**Note:** To install a specific version, run with the following command:
```
python -m pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
```
......@@ -1701,6 +1803,7 @@ python -m pip install intel_extension_for_pytorch==<version_name> -f https://dev
2. After pipeline initialization, `prepare_for_ipex()` should be called to enable IPEX accelaration. Supported inference datatypes are Float32 and BFloat16.
**Note:** The setting of generated image height/width for `prepare_for_ipex()` should be same as the setting of pipeline inference.
```python
pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="stable_diffusion_ipex")
# For Float32
......@@ -1710,6 +1813,7 @@ pipe.prepare_for_ipex(prompt, dtype=torch.bfloat16, height=512, width=512) #valu
```
Then you can use the ipex pipeline in a similar way to the default stable diffusion pipeline.
```python
# For Float32
image = pipe(prompt, num_inference_steps=20, height=512, width=512).images[0] #value of image height/width should be consistent with 'prepare_for_ipex()'
......@@ -1778,6 +1882,7 @@ print("Latency of StableDiffusionPipeline--fp32",latency)
This diffusion pipeline aims to accelarate the inference of Stable-Diffusion XL on Intel Xeon CPUs with BF16/FP32 precision using [IPEX](https://github.com/intel/intel-extension-for-pytorch).
To use this pipeline, you need to:
1. Install [IPEX](https://github.com/intel/intel-extension-for-pytorch)
**Note:** For each PyTorch release, there is a corresponding release of IPEX. Here is the mapping relationship. It is recommended to install Pytorch/IPEX2.0 to get the best performance.
......@@ -1788,10 +1893,13 @@ To use this pipeline, you need to:
|[v1.13.\*](https://github.com/pytorch/pytorch/tree/v1.13.0 "v1.13.0")|[v1.13.\*](https://github.com/intel/intel-extension-for-pytorch/tree/v1.13.100+cpu)|
You can simply use pip to install IPEX with the latest version.
```python
python -m pip install intel_extension_for_pytorch
```
**Note:** To install a specific version, run with the following command:
```
python -m pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
```
......@@ -1810,6 +1918,7 @@ pipe.prepare_for_ipex(torch.bfloat16, prompt, height=512, width=512)
```
Then you can use the ipex pipeline in a similar way to the default stable diffusion xl pipeline.
```python
# value of image height/width should be consistent with 'prepare_for_ipex()'
# For Float32
......@@ -1886,7 +1995,6 @@ CLIP guided stable diffusion images mixing pipeline allows to combine two images
This approach is using (optional) CoCa model to avoid writing image description.
[More code examples](https://github.com/TheDenk/images_mixing)
### Stable Diffusion XL Long Weighted Prompt Pipeline
This SDXL pipeline support unlimited length prompt and negative prompt, compatible with A1111 prompt weighted style.
......@@ -1953,6 +2061,7 @@ In the above code, the `prompt2` is appended to the `prompt`, which is more than
For more results, checkout [PR #6114](https://github.com/huggingface/diffusers/pull/6114).
### Example Images Mixing (with CoCa)
```python
import requests
from io import BytesIO
......@@ -2055,6 +2164,7 @@ image = pipeline(
num_inference_steps=50,
)["images"][0]
```
![mixture_tiling_results](https://huggingface.co/datasets/kadirnar/diffusers_readme_images/resolve/main/mixture_tiling.png)
### TensorRT Inpainting Stable Diffusion Pipeline
......@@ -2131,10 +2241,10 @@ output = pipeline(
seed=5525475061,
)["images"][0]
```
![Input_Image](https://huggingface.co/datasets/kadirnar/diffusers_readme_images/resolve/main/input_image.png)
![mixture_canvas_results](https://huggingface.co/datasets/kadirnar/diffusers_readme_images/resolve/main/canvas.png)
### IADB pipeline
This pipeline is the implementation of the [α-(de)Blending: a Minimalist Deterministic Diffusion Model](https://arxiv.org/abs/2305.03486) paper.
......@@ -2277,7 +2387,6 @@ for obj in range(bs):
This pipeline uses the Reference . Refer to the [stable_diffusion_reference](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#stable-diffusion-reference).
```py
import torch
from PIL import Image
......@@ -2322,7 +2431,6 @@ Output Image
Reference Image
![reference_image](https://github.com/huggingface/diffusers/assets/34944964/449bdab6-e744-4fb2-9620-d4068d9a741b)
Output Image
`prompt: A dog`
......@@ -2346,7 +2454,6 @@ FABRIC approach applicable to a wide range of popular diffusion models, which ex
the self-attention layer present in the most widely used architectures to condition
the diffusion process on a set of feedback images.
```python
import requests
import torch
......@@ -2400,13 +2507,12 @@ image.save("black_to_blue.png")
The original codebase can be found at [sd-fabric/fabric](https://github.com/sd-fabric/fabric), and available checkpoints are [dreamlike-art/dreamlike-photoreal-2.0](https://huggingface.co/dreamlike-art/dreamlike-photoreal-2.0), [runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5), and [stabilityai/stable-diffusion-2-1](https://huggingface.co/stabilityai/stable-diffusion-2-1) (may give unexpected results).
Let's have a look at the images (*512X512*)
Let's have a look at the images (_512X512_)
| Without Feedback | With Feedback (1st image) |
|---------------------|---------------------|
| ![Image 1](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/fabric_wo_feedback.jpg) | ![Feedback Image 1](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/fabric_w_feedback.png) |
### Masked Im2Im Stable Diffusion Pipeline
This pipeline reimplements sketch inpaint feature from A1111 for non-inpaint models. The following code reads two images, original and one with mask painted over it. It computes mask as a difference of two images and does the inpainting in the area defined by the mask.
......@@ -2432,20 +2538,20 @@ result.images[0].save("result.png")
original image mech.png
<img src=https://github.com/noskill/diffusers/assets/733626/10ad972d-d655-43cb-8de1-039e3d79e849 width="25%" >
<img src=<https://github.com/noskill/diffusers/assets/733626/10ad972d-d655-43cb-8de1-039e3d79e849> width="25%" >
image with mask mech_painted.png
<img src=https://github.com/noskill/diffusers/assets/733626/c334466a-67fe-4377-9ff7-f46021b9c224 width="25%" >
<img src=<https://github.com/noskill/diffusers/assets/733626/c334466a-67fe-4377-9ff7-f46021b9c224> width="25%" >
result:
<img src=https://github.com/noskill/diffusers/assets/733626/23a0a71d-51db-471e-926a-107ac62512a8 width="25%" >
<img src=<https://github.com/noskill/diffusers/assets/733626/23a0a71d-51db-471e-926a-107ac62512a8> width="25%" >
### Prompt2Prompt Pipeline
Prompt2Prompt allows the following edits:
- ReplaceEdit (change words in prompt)
- ReplaceEdit with local blend (change words in prompt, keep image part unrelated to changes constant)
- RefineEdit (add words to prompt)
......@@ -2477,6 +2583,7 @@ outputs = pipe(prompt=prompts, height=512, width=512, num_inference_steps=50, cr
And abbreviated examples for the other edits:
`ReplaceEdit with local blend`
```python
prompts = ["A turtle playing with a ball",
"A monkey playing with a ball"]
......@@ -2490,6 +2597,7 @@ cross_attention_kwargs = {
```
`RefineEdit`
```python
prompts = ["A turtle",
"A turtle in a forest"]
......@@ -2502,6 +2610,7 @@ cross_attention_kwargs = {
```
`RefineEdit with local blend`
```python
prompts = ["A turtle",
"A turtle in a forest"]
......@@ -2515,6 +2624,7 @@ cross_attention_kwargs = {
```
`ReweightEdit`
```python
prompts = ["A smiling turtle"] * 2
......@@ -2531,7 +2641,7 @@ Side note: See [this GitHub gist](https://gist.github.com/UmerHA/b65bb5fb9626c9c
### Latent Consistency Pipeline
Latent Consistency Models was proposed in [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://arxiv.org/abs/2310.04378) by *Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, Hang Zhao* from Tsinghua University.
Latent Consistency Models was proposed in [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://arxiv.org/abs/2310.04378) by _Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, Hang Zhao_ from Tsinghua University.
The abstract of the paper reads as follows:
......@@ -2539,7 +2649,7 @@ The abstract of the paper reads as follows:
The model can be used with `diffusers` as follows:
- *1. Load the model from the community pipeline.*
- *1. Load the model from the community pipeline.*
```py
from diffusers import DiffusionPipeline
......@@ -2566,8 +2676,6 @@ For any questions or feedback, feel free to reach out to [Simian Luo](https://gi
You can also try this pipeline directly in the [🚀 official spaces](https://huggingface.co/spaces/SimianLuo/Latent_Consistency_Model).
### Latent Consistency Img2img Pipeline
This pipeline extends the Latent Consistency Pipeline to allow it to take an input image.
......@@ -2598,8 +2706,6 @@ num_inference_steps = 4
images = pipe(prompt=prompt, image=input_image, strength=strength, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images
```
### Latent Consistency Interpolation Pipeline
This pipeline extends the Latent Consistency Pipeline to allow for interpolation of the latent space between multiple prompts. It is similar to the [Stable Diffusion Interpolate](https://github.com/huggingface/diffusers/blob/main/examples/community/interpolate_stable_diffusion.py) and [unCLIP Interpolate](https://github.com/huggingface/diffusers/blob/main/examples/community/unclip_text_interpolation.py) community pipelines.
......@@ -2646,12 +2752,14 @@ assert len(images) == (len(prompts) - 1) * num_interpolation_steps
```
### StableDiffusionUpscaleLDM3D Pipeline
[LDM3D-VR](https://arxiv.org/pdf/2311.03226.pdf) is an extended version of LDM3D.
The abstract from the paper is:
*Latent diffusion models have proven to be state-of-the-art in the creation and manipulation of visual outputs. However, as far as we know, the generation of depth maps jointly with RGB is still limited. We introduce LDM3D-VR, a suite of diffusion models targeting virtual reality development that includes LDM3D-pano and LDM3D-SR. These models enable the generation of panoramic RGBD based on textual prompts and the upscaling of low-resolution inputs to high-resolution RGBD, respectively. Our models are fine-tuned from existing pretrained models on datasets containing panoramic/high-resolution RGB images, depth maps and captions. Both models are evaluated in comparison to existing related methods*
Two checkpoints are available for use:
- [ldm3d-pano](https://huggingface.co/Intel/ldm3d-pano). This checkpoint enables the generation of panoramic images and requires the StableDiffusionLDM3DPipeline pipeline to be used.
- [ldm3d-sr](https://huggingface.co/Intel/ldm3d-sr). This checkpoint enables the upscaling of RGB and depth images. Can be used in cascade after the original LDM3D pipeline using the StableDiffusionUpscaleLDM3DPipeline pipeline.
......@@ -2661,7 +2769,8 @@ import os
import torch
from diffusers import StableDiffusionLDM3DPipeline, DiffusionPipeline
#Generate a rgb/depth output from LDM3D
# Generate a rgb/depth output from LDM3D
pipe_ldm3d = StableDiffusionLDM3DPipeline.from_pretrained("Intel/ldm3d-4c")
pipe_ldm3d.to("cuda")
......@@ -2671,8 +2780,8 @@ rgb_image, depth_image = output.rgb, output.depth
rgb_image[0].save(f"lemons_ldm3d_rgb.jpg")
depth_image[0].save(f"lemons_ldm3d_depth.png")
# Upscale the previous output to a resolution of (1024, 1024)
#Upscale the previous output to a resolution of (1024, 1024)
pipe_ldm3d_upscale = DiffusionPipeline.from_pretrained("Intel/ldm3d-sr", custom_pipeline="pipeline_stable_diffusion_upscale_ldm3d")
pipe_ldm3d_upscale.to("cuda")
......@@ -2687,6 +2796,7 @@ upscaled_depth.save(f"upscaled_lemons_depth.png")
'''
### ControlNet + T2I Adapter Pipeline
This pipelines combines both ControlNet and T2IAdapter into a single pipeline, where the forward pass is executed once.
It receives `control_image` and `adapter_image`, as well as `controlnet_conditioning_scale` and `adapter_conditioning_scale`, for the ControlNet and Adapter modules, respectively. Whenever `adapter_conditioning_scale = 0` or `controlnet_conditioning_scale = 0`, it will act as a full ControlNet module or as a full T2IAdapter module, respectively.
......@@ -2755,6 +2865,7 @@ images[0].save("controlnet_and_adapter.png")
```
### ControlNet + T2I Adapter + Inpainting Pipeline
```py
import cv2
import numpy as np
......@@ -2825,13 +2936,16 @@ images[0].save("controlnet_and_adapter_inpaint.png")
```
### Regional Prompting Pipeline
This pipeline is a port of the [Regional Prompter extension](https://github.com/hako-mikan/sd-webui-regional-prompter) for [Stable Diffusion web UI](https://github.com/AUTOMATIC1111/stable-diffusion-webui) to diffusers.
This code implements a pipeline for the Stable Diffusion model, enabling the division of the canvas into multiple regions, with different prompts applicable to each region. Users can specify regions in two ways: using `Cols` and `Rows` modes for grid-like divisions, or the `Prompt` mode for regions calculated based on prompts.
![sample](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/imgs/rp_pipeline1.png)
### Usage
### Sample Code
```
from from examples.community.regional_prompting_stable_diffusion import RegionalPromptingStableDiffusionPipeline
pipe = RegionalPromptingStableDiffusionPipeline.from_single_file(model_path, vae=vae)
......@@ -2865,11 +2979,14 @@ for image in images:
fileName = f'img-{time}-{i+1}.png'
image.save(fileName)
```
### Cols, Rows mode
In the Cols, Rows mode, you can split the screen vertically and horizontally and assign prompts to each region. The split ratio can be specified by 'div', and you can set the division ratio like '3;3;2' or '0.1;0.5'. Furthermore, as will be described later, you can also subdivide the split Cols, Rows to specify more complex regions.
In this image, the image is divided into three parts, and a separate prompt is applied to each. The prompts are divided by 'BREAK', and each is applied to the respective region.
![sample](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/imgs/rp_pipeline2.png)
```
green hair twintail BREAK
red blouse BREAK
......@@ -2877,7 +2994,9 @@ blue skirt
```
### 2-Dimentional division
The prompt consists of instructions separated by the term `BREAK` and is assigned to different regions of a two-dimensional space. The image is initially split in the main splitting direction, which in this case is rows, due to the presence of a single semicolon`;`, dividing the space into an upper and a lower section. Additional sub-splitting is then applied, indicated by commas. The upper row is split into ratios of `2:1:1`, while the lower row is split into a ratio of `4:6`. Rows themselves are split in a `1:2` ratio. According to the reference image, the blue sky is designated as the first region, green hair as the second, the bookshelf as the third, and so on, in a sequence based on their position from the top left. The terrarium is placed on the desk in the fourth region, and the orange dress and sofa are in the fifth region, conforming to their respective splits.
```
rp_args = {
"mode":"rows",
......@@ -2892,12 +3011,16 @@ terrarium on desk BREAK
orange dress and sofa
"""
```
![sample](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/imgs/rp_pipeline4.png)
### Prompt Mode
There are limitations to methods of specifying regions in advance. This is because specifying regions can be a hindrance when designating complex shapes or dynamic compositions. In the region specified by the prompt, the regions is determined after the image generation has begun. This allows us to accommodate compositions and complex regions.
For further infomagen, see [here](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/main/prompt_en.md).
### syntax
```
baseprompt target1 target2 BREAK
effect1, target1 BREAK
......@@ -2911,10 +3034,13 @@ target2 baseprompt target1 BREAK
effect1, target1 BREAK
effect2 ,target2
```
is also effective.
### Sample
In this example, masks are calculated for shirt, tie, skirt, and color prompts are specified only for those regions.
```
rp_args = {
"mode":"prompt-ex",
......@@ -2929,8 +3055,11 @@ green, tie BREAK
blue , skirt
"""
```
![sample](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/imgs/rp_pipeline3.png)
### threshold
The threshold used to determine the mask created by the prompt. This can be set as many times as there are masks, as the range varies widely depending on the target prompt. If multiple regions are used, enter them separated by commas. For example, hair tends to be ambiguous and requires a small value, while face tends to be large and requires a small value. These should be ordered by BREAK.
```
......@@ -2938,44 +3067,56 @@ a lady ,hair, face BREAK
red, hair BREAK
tanned ,face
```
`threshold : 0.4,0.6`
If only one input is given for multiple regions, they are all assumed to be the same value.
### Prompt and Prompt-EX
The difference is that in Prompt, duplicate regions are added, whereas in Prompt-EX, duplicate regions are overwritten sequentially. Since they are processed in order, setting a TARGET with a large regions first makes it easier for the effect of small regions to remain unmuffled.
### Accuracy
In the case of a 512 x 512 image, Attention mode reduces the size of the region to about 8 x 8 pixels deep in the U-Net, so that small regions get mixed up; Latent mode calculates 64*64, so that the region is exact.
```
girl hair twintail frills,ribbons, dress, face BREAK
girl, ,face
```
### Mask
When an image is generated, the generated mask is displayed. It is generated at the same size as the image, but is actually used at a much smaller size.
When an image is generated, the generated mask is displayed. It is generated at the same size as the image, but is actually used at a much smaller size.
### Use common prompt
You can attach the prompt up to ADDCOMM to all prompts by separating it first with ADDCOMM. This is useful when you want to include elements common to all regions. For example, when generating pictures of three people with different appearances, it's necessary to include the instruction of 'three people' in all regions. It's also useful when inserting quality tags and other things."For example, if you write as follows:
```
best quality, 3persons in garden, ADDCOMM
a girl white dress BREAK
a boy blue shirt BREAK
an old man red suit
```
If common is enabled, this prompt is converted to the following:
```
best quality, 3persons in garden, a girl white dress BREAK
best quality, 3persons in garden, a boy blue shirt BREAK
best quality, 3persons in garden, an old man red suit
```
### Negative prompt
Negative prompts are equally effective across all regions, but it is possible to set region-specific prompts for negative prompts as well. The number of BREAKs must be the same as the number of prompts. If the number of prompts does not match, the negative prompts will be used without being divided into regions.
### Parameters
To activate Regional Prompter, it is necessary to enter settings in rp_args. The items that can be set are as follows. rp_args is a dictionary type.
### Input Parameters
Parameters are specified through the `rp_arg`(dictionary type).
```
......@@ -2987,20 +3128,22 @@ rp_args = {
pipe(prompt =prompt, rp_args = rp_args)
```
### Required Parameters
- `mode`: Specifies the method for defining regions. Choose from `Cols`, `Rows`, `Prompt` or `Prompt-Ex`. This parameter is case-insensitive.
- `divide`: Used in `Cols` and `Rows` modes. Details on how to specify this are provided under the respective `Cols` and `Rows` sections.
- `th`: Used in `Prompt` mode. The method of specification is detailed under the `Prompt` section.
### Optional Parameters
- `save_mask`: In `Prompt` mode, choose whether to output the generated mask along with the image. The default is `False`.
The Pipeline supports `compel` syntax. Input prompts using the `compel` structure will be automatically applied and processed.
### Diffusion Posterior Sampling Pipeline
* Reference paper
- Reference paper
```
@article{chung2022diffusion,
title={Diffusion posterior sampling for general noisy inverse problems},
......@@ -3009,9 +3152,12 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
year={2022}
}
```
* This pipeline allows zero-shot conditional sampling from the posterior distribution $p(x|y)$, given observation on $y$, unconditional generative model $p(x)$ and differentiable operator $y=f(x)$.
* For example, $f(.)$ can be downsample operator, then $y$ is a downsampled image, and the pipeline becomes a super-resolution pipeline.
* To use this pipeline, you need to know your operator $f(.)$ and corrupted image $y$, and pass them during the call. For example, as in the main function of dps_pipeline.py, you need to first define the Gaussian blurring operator $f(.)$. The operator should be a callable nn.Module, with all the parameter gradient disabled:
- This pipeline allows zero-shot conditional sampling from the posterior distribution $p(x|y)$, given observation on $y$, unconditional generative model $p(x)$ and differentiable operator $y=f(x)$.
- For example, $f(.)$ can be downsample operator, then $y$ is a downsampled image, and the pipeline becomes a super-resolution pipeline.
- To use this pipeline, you need to know your operator $f(.)$ and corrupted image $y$, and pass them during the call. For example, as in the main function of dps_pipeline.py, you need to first define the Gaussian blurring operator $f(.)$. The operator should be a callable nn.Module, with all the parameter gradient disabled:
```python
import torch.nn.functional as F
import scipy
......@@ -3081,7 +3227,9 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
def get_kernel(self):
return self.kernel.view(1, 1, self.kernel_size, self.kernel_size)
```
* Next, you should obtain the corrupted image $y$ by the operator. In this example, we generate $y$ from the source image $x$. However in practice, having the operator $f(.)$ and corrupted image $y$ is enough:
- Next, you should obtain the corrupted image $y$ by the operator. In this example, we generate $y$ from the source image $x$. However in practice, having the operator $f(.)$ and corrupted image $y$ is enough:
```python
# set up source image
src = Image.open('sample.png')
......@@ -3099,18 +3247,23 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
save_image((src+1.0)/2.0, "dps_src.png")
save_image((measurement+1.0)/2.0, "dps_mea.png")
```
* We provide an example pair of saved source and corrupted images, using the Gaussian blur operator above
* Source image:
* ![sample](https://github.com/tongdaxu/Images/assets/22267548/4d2a1216-08d1-4aeb-9ce3-7a2d87561d65)
* Gaussian blurred image:
* ![ddpm_generated_image](https://github.com/tongdaxu/Images/assets/22267548/65076258-344b-4ed8-b704-a04edaade8ae)
* You can download those image to run the example on your own.
* Next, we need to define a loss function used for diffusion posterior sample. For most of the cases, the RMSE is fine:
- We provide an example pair of saved source and corrupted images, using the Gaussian blur operator above
- Source image:
- ![sample](https://github.com/tongdaxu/Images/assets/22267548/4d2a1216-08d1-4aeb-9ce3-7a2d87561d65)
- Gaussian blurred image:
- ![ddpm_generated_image](https://github.com/tongdaxu/Images/assets/22267548/65076258-344b-4ed8-b704-a04edaade8ae)
- You can download those image to run the example on your own.
- Next, we need to define a loss function used for diffusion posterior sample. For most of the cases, the RMSE is fine:
```python
def RMSELoss(yhat, y):
return torch.sqrt(torch.sum((yhat-y)**2))
```
* And next, as any other diffusion models, we need the score estimator and scheduler. As we are working with $256x256$ face images, we use ddmp-celebahq-256:
- And next, as any other diffusion models, we need the score estimator and scheduler. As we are working with $256x256$ face images, we use ddmp-celebahq-256:
```python
# set up scheduler
scheduler = DDPMScheduler.from_pretrained("google/ddpm-celebahq-256")
......@@ -3119,7 +3272,9 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
# set up model
model = UNet2DModel.from_pretrained("google/ddpm-celebahq-256").to("cuda")
```
* And finally, run the pipeline:
- And finally, run the pipeline:
```python
# finally, the pipeline
dpspipe = DPSPipeline(model, scheduler)
......@@ -3131,15 +3286,17 @@ The Pipeline supports `compel` syntax. Input prompts using the `compel` structur
).images[0]
image.save("dps_generated_image.png")
```
* The zeta is a hyperparameter that is in range of $[0,1]$. It need to be tuned for best effect. By setting zeta=1, you should be able to have the reconstructed result:
* Reconstructed image:
* ![sample](https://github.com/tongdaxu/Images/assets/22267548/0ceb5575-d42e-4f0b-99c0-50e69c982209)
* The reconstruction is perceptually similar to the source image, but different in details.
* In dps_pipeline.py, we also provide a super-resolution example, which should produce:
* Downsampled image:
* ![dps_mea](https://github.com/tongdaxu/Images/assets/22267548/ff6a33d6-26f0-42aa-88ce-f8a76ba45a13)
* Reconstructed image:
* ![dps_generated_image](https://github.com/tongdaxu/Images/assets/22267548/b74f084d-93f4-4845-83d8-44c0fa758a5f)
- The zeta is a hyperparameter that is in range of $[0,1]$. It need to be tuned for best effect. By setting zeta=1, you should be able to have the reconstructed result:
- Reconstructed image:
- ![sample](https://github.com/tongdaxu/Images/assets/22267548/0ceb5575-d42e-4f0b-99c0-50e69c982209)
- The reconstruction is perceptually similar to the source image, but different in details.
- In dps_pipeline.py, we also provide a super-resolution example, which should produce:
- Downsampled image:
- ![dps_mea](https://github.com/tongdaxu/Images/assets/22267548/ff6a33d6-26f0-42aa-88ce-f8a76ba45a13)
- Reconstructed image:
- ![dps_generated_image](https://github.com/tongdaxu/Images/assets/22267548/b74f084d-93f4-4845-83d8-44c0fa758a5f)
### AnimateDiff ControlNet Pipeline
......@@ -3283,6 +3440,7 @@ export_to_gif(result.frames[0], "result.gif")
This pipeline is the official implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973).
The original repo can be found at [repo](https://github.com/PRIS-CV/DemoFusion).
- `view_batch_size` (`int`, defaults to 16):
The batch size for multiple denoising paths. Typically, a larger batch size can result in higher efficiency but comes with increased GPU memory requirements.
......@@ -3306,6 +3464,7 @@ The original repo can be found at [repo](https://github.com/PRIS-CV/DemoFusion).
- `show_image` (`bool`, defaults to False):
Determine whether to show intermediate results during generation.
```py
from diffusers import DiffusionPipeline
......@@ -3337,7 +3496,9 @@ images = pipe(
show_image=True
)
```
You can display and save the generated images as:
```py
def image_grid(imgs, save_path=None):
......@@ -3361,6 +3522,7 @@ def image_grid(imgs, save_path=None):
image_grid(images, save_path="./outputs/")
```
![output_example](https://github.com/PRIS-CV/DemoFusion/blob/main/output_example.png)
### SDE Drag pipeline
......@@ -3403,6 +3565,7 @@ output_image.save("./output.png")
```
### Instaflow Pipeline
InstaFlow is an ultra-fast, one-step image generator that achieves image quality close to Stable Diffusion, significantly reducing the demand of computational resources. This efficiency is made possible through a recent [Rectified Flow](https://github.com/gnobitab/RectifiedFlow) technique, which trains probability flows with straight trajectories, hence inherently requiring only a single step for fast inference.
```python
......@@ -3419,9 +3582,10 @@ images = pipe(prompt=prompt,
guidance_scale=0.0).images
images[0].save("./image.png")
```
![image1](https://huggingface.co/datasets/ayushtues/instaflow_images/resolve/main/instaflow_cat.png)
You can also combine it with LORA out of the box, like https://huggingface.co/artificialguybr/logo-redmond-1-5v-logo-lora-for-liberteredmond-sd-1-5, to unlock cool use cases in single step!
You can also combine it with LORA out of the box, like <https://huggingface.co/artificialguybr/logo-redmond-1-5v-logo-lora-for-liberteredmond-sd-1-5>, to unlock cool use cases in single step!
```python
from diffusers import DiffusionPipeline
......@@ -3437,12 +3601,15 @@ images = pipe(prompt=prompt,
guidance_scale=0.0).images
images[0].save("./image.png")
```
![image0](https://huggingface.co/datasets/ayushtues/instaflow_images/resolve/main/instaflow_logo.png)
### Null-Text Inversion pipeline
This pipeline provides null-text inversion for editing real images. It enables null-text optimization, and DDIM reconstruction via w, w/o null-text optimization. No prompt-to-prompt code is implemented as there is a Prompt2PromptPipeline.
* Reference paper
- Reference paper
```@article{hertz2022prompt,
title={Prompt-to-prompt image editing with cross attention control},
author={Hertz, Amir and Mokady, Ron and Tenenbaum, Jay and Aberman, Kfir and Pritch, Yael and Cohen-Or, Daniel},
......
examples/community/pipeline_stable_diffusion_xl_differential_img2img.py 0 → 100644
View file @ 73ba8109
# 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 torchvision
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, 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.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
USE_PEFT_BACKEND,
deprecate,
is_invisible_watermark_available,
is_torch_xla_available,
logging,
replace_example_docstring,
scale_lora_layers,
unscale_lora_layers,
)
from diffusers.utils.torch_utils import randn_tensor
if is_invisible_watermark_available():
from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
if is_torch_xla_available():
import torch_xla.core.xla_model as xm
XLA_AVAILABLE = True
else:
XLA_AVAILABLE = False
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
EXAMPLE_DOC_STRING = """
Examples:
```py
>>> import torch
>>> from diffusers import StableDiffusionXLImg2ImgPipeline
>>> from diffusers.utils import load_image
>>> pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
... "stabilityai/stable-diffusion-xl-refiner-1.0", torch_dtype=torch.float16
... )
>>> pipe = pipe.to("cuda")
>>> url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png"
>>> init_image = load_image(url).convert("RGB")
>>> prompt = "a photo of an astronaut riding a horse on mars"
>>> image = pipe(prompt, image=init_image).images[0]
```
"""
# 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_img2img.retrieve_latents
def retrieve_latents(
encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
):
if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
return encoder_output.latent_dist.sample(generator)
elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
return encoder_output.latent_dist.mode()
elif hasattr(encoder_output, "latents"):
return encoder_output.latents
else:
raise AttributeError("Could not access latents of provided encoder_output")
# 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,
**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 support arbitrary spacing between timesteps. If `None`, then the default
timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
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:
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)
else:
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
timesteps = scheduler.timesteps
return timesteps, num_inference_steps
class StableDiffusionXLDifferentialImg2ImgPipeline(
DiffusionPipeline,
StableDiffusionMixin,
TextualInversionLoaderMixin,
FromSingleFileMixin,
StableDiffusionXLLoraLoaderMixin,
IPAdapterMixin,
):
r"""
Pipeline for text-to-image generation using Stable Diffusion XL.
This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
In addition the pipeline inherits the following loading methods:
- *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
- *LoRA*: [`loaders.LoraLoaderMixin.load_lora_weights`]
- *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
as well as the following saving methods:
- *LoRA*: [`loaders.LoraLoaderMixin.save_lora_weights`]
Args:
vae ([`AutoencoderKL`]):
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
text_encoder ([`CLIPTextModel`]):
Frozen text-encoder. Stable Diffusion XL uses the text portion of
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
text_encoder_2 ([` CLIPTextModelWithProjection`]):
Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
[CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
specifically the
[laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
variant.
tokenizer (`CLIPTokenizer`):
Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
tokenizer_2 (`CLIPTokenizer`):
Second Tokenizer of class
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
scheduler ([`SchedulerMixin`]):
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
"""
model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
_optional_components = [
"tokenizer",
"tokenizer_2",
"text_encoder",
"text_encoder_2",
"image_encoder",
"feature_extractor",
]
_callback_tensor_inputs = [
"latents",
"prompt_embeds",
"negative_prompt_embeds",
"add_text_embeds",
"add_time_ids",
"negative_pooled_prompt_embeds",
"add_neg_time_ids",
]
def __init__(
self,
vae: AutoencoderKL,
text_encoder: CLIPTextModel,
text_encoder_2: CLIPTextModelWithProjection,
tokenizer: CLIPTokenizer,
tokenizer_2: CLIPTokenizer,
unet: UNet2DConditionModel,
scheduler: KarrasDiffusionSchedulers,
image_encoder: CLIPVisionModelWithProjection = None,
feature_extractor: CLIPImageProcessor = None,
requires_aesthetics_score: bool = False,
force_zeros_for_empty_prompt: bool = True,
add_watermarker: Optional[bool] = None,
):
super().__init__()
self.register_modules(
vae=vae,
text_encoder=text_encoder,
text_encoder_2=text_encoder_2,
tokenizer=tokenizer,
tokenizer_2=tokenizer_2,
unet=unet,
image_encoder=image_encoder,
feature_extractor=feature_extractor,
scheduler=scheduler,
)
self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
self.register_to_config(requires_aesthetics_score=requires_aesthetics_score)
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
if add_watermarker:
self.watermark = StableDiffusionXLWatermarker()
else:
self.watermark = None
# Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(
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.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
lora_scale: Optional[float] = None,
clip_skip: Optional[int] = None,
):
r"""
Encodes the prompt into text encoder hidden states.
Args:
prompt (`str` or `List[str]`, *optional*):
prompt to be encoded
prompt_2 (`str` or `List[str]`, *optional*):
The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
used in both text-encoders
device: (`torch.device`):
torch device
num_images_per_prompt (`int`):
number of images that should be generated per prompt
do_classifier_free_guidance (`bool`):
whether to use classifier free guidance or not
negative_prompt (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation. If not defined, one has to pass
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
less than `1`).
negative_prompt_2 (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
`text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
prompt_embeds (`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.
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
argument.
pooled_prompt_embeds (`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.
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
input argument.
lora_scale (`float`, *optional*):
A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
clip_skip (`int`, *optional*):
Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
the output of the pre-final layer will be used for computing the prompt embeddings.
"""
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]
# 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])
logger.warning(
"The following part of your input was truncated because CLIP can only handle sequences up to"
f" {tokenizer.model_max_length} tokens: {removed_text}"
)
prompt_embeds = text_encoder(text_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)
# 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
# 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)
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
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
def prepare_extra_step_kwargs(self, generator, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
extra_step_kwargs["generator"] = generator
return extra_step_kwargs
def check_inputs(
self,
prompt,
prompt_2,
strength,
num_inference_steps,
callback_steps,
negative_prompt=None,
negative_prompt_2=None,
prompt_embeds=None,
negative_prompt_embeds=None,
ip_adapter_image=None,
ip_adapter_image_embeds=None,
callback_on_step_end_tensor_inputs=None,
):
if strength < 0 or strength > 1:
raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
if num_inference_steps is None:
raise ValueError("`num_inference_steps` cannot be None.")
elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0:
raise ValueError(
f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type"
f" {type(num_inference_steps)}."
)
if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
if callback_on_step_end_tensor_inputs is not None and not all(
k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
):
raise ValueError(
f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
)
if prompt is not None and prompt_embeds is not None:
raise ValueError(
f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
" only forward one of the two."
)
elif prompt_2 is not None and prompt_embeds is not None:
raise ValueError(
f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
" only forward one of the two."
)
elif prompt is None and prompt_embeds is None:
raise ValueError(
"Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
)
elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
if negative_prompt is not None and negative_prompt_embeds is not None:
raise ValueError(
f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
)
elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
raise ValueError(
f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
)
if prompt_embeds is not None and negative_prompt_embeds is not None:
if prompt_embeds.shape != negative_prompt_embeds.shape:
raise ValueError(
"`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
f" {negative_prompt_embeds.shape}."
)
if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
raise ValueError(
"Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
)
if ip_adapter_image_embeds is not None:
if not isinstance(ip_adapter_image_embeds, list):
raise ValueError(
f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
)
elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
raise ValueError(
f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
)
def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None):
# get the original timestep using init_timestep
if denoising_start is None:
init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
t_start = max(num_inference_steps - init_timestep, 0)
else:
t_start = 0
timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
# Strength is irrelevant if we directly request a timestep to start at;
# that is, strength is determined by the denoising_start instead.
if denoising_start is not None:
discrete_timestep_cutoff = int(
round(
self.scheduler.config.num_train_timesteps
- (denoising_start * self.scheduler.config.num_train_timesteps)
)
)
num_inference_steps = (timesteps < discrete_timestep_cutoff).sum().item()
if self.scheduler.order == 2 and num_inference_steps % 2 == 0:
# if the scheduler is a 2nd order scheduler we might have to do +1
# because `num_inference_steps` might be even given that every timestep
# (except the highest one) is duplicated. If `num_inference_steps` is even it would
# mean that we cut the timesteps in the middle of the denoising step
# (between 1st and 2nd derivative) which leads to incorrect results. By adding 1
# we ensure that the denoising process always ends after the 2nd derivate step of the scheduler
num_inference_steps = num_inference_steps + 1
# because t_n+1 >= t_n, we slice the timesteps starting from the end
timesteps = timesteps[-num_inference_steps:]
return timesteps, num_inference_steps
return timesteps, num_inference_steps - t_start
def prepare_latents(
self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None, add_noise=True
):
if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
raise ValueError(
f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
)
# Offload text encoder if `enable_model_cpu_offload` was enabled
if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
self.text_encoder_2.to("cpu")
torch.cuda.empty_cache()
image = image.to(device=device, dtype=dtype)
batch_size = batch_size * num_images_per_prompt
if image.shape[1] == 4:
init_latents = image
else:
# make sure the VAE is in float32 mode, as it overflows in float16
if self.vae.config.force_upcast:
image = image.float()
self.vae.to(dtype=torch.float32)
if isinstance(generator, list) and len(generator) != batch_size:
raise ValueError(
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
)
elif isinstance(generator, list):
init_latents = [
retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
for i in range(batch_size)
]
init_latents = torch.cat(init_latents, dim=0)
else:
init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
if self.vae.config.force_upcast:
self.vae.to(dtype)
init_latents = init_latents.to(dtype)
init_latents = self.vae.config.scaling_factor * init_latents
if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
# expand init_latents for batch_size
additional_image_per_prompt = batch_size // init_latents.shape[0]
init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
raise ValueError(
f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
)
else:
init_latents = torch.cat([init_latents], dim=0)
if add_noise:
shape = init_latents.shape
noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
# get latents
init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
latents = init_latents
return latents
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
dtype = next(self.image_encoder.parameters()).dtype
if not isinstance(image, torch.Tensor):
image = self.feature_extractor(image, return_tensors="pt").pixel_values
image = image.to(device=device, dtype=dtype)
if output_hidden_states:
image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
uncond_image_enc_hidden_states = self.image_encoder(
torch.zeros_like(image), output_hidden_states=True
).hidden_states[-2]
uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
num_images_per_prompt, dim=0
)
return image_enc_hidden_states, uncond_image_enc_hidden_states
else:
image_embeds = self.image_encoder(image).image_embeds
image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
uncond_image_embeds = torch.zeros_like(image_embeds)
return image_embeds, uncond_image_embeds
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
def prepare_ip_adapter_image_embeds(
self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
):
if ip_adapter_image_embeds is None:
if not isinstance(ip_adapter_image, list):
ip_adapter_image = [ip_adapter_image]
if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
raise ValueError(
f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
)
image_embeds = []
for single_ip_adapter_image, image_proj_layer in zip(
ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
):
output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
single_image_embeds, single_negative_image_embeds = self.encode_image(
single_ip_adapter_image, device, 1, output_hidden_state
)
single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
single_negative_image_embeds = torch.stack(
[single_negative_image_embeds] * num_images_per_prompt, dim=0
)
if do_classifier_free_guidance:
single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
single_image_embeds = single_image_embeds.to(device)
image_embeds.append(single_image_embeds)
else:
repeat_dims = [1]
image_embeds = []
for single_image_embeds in ip_adapter_image_embeds:
if do_classifier_free_guidance:
single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
single_image_embeds = single_image_embeds.repeat(
num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
)
single_negative_image_embeds = single_negative_image_embeds.repeat(
num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
)
single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
else:
single_image_embeds = single_image_embeds.repeat(
num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
)
image_embeds.append(single_image_embeds)
return image_embeds
def _get_add_time_ids(
self,
original_size,
crops_coords_top_left,
target_size,
aesthetic_score,
negative_aesthetic_score,
negative_original_size,
negative_crops_coords_top_left,
negative_target_size,
dtype,
text_encoder_projection_dim=None,
):
if self.config.requires_aesthetics_score:
add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,))
add_neg_time_ids = list(
negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,)
)
else:
add_time_ids = list(original_size + crops_coords_top_left + target_size)
add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size)
passed_add_embed_dim = (
self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
)
expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
if (
expected_add_embed_dim > passed_add_embed_dim
and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim
):
raise ValueError(
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model."
)
elif (
expected_add_embed_dim < passed_add_embed_dim
and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim
):
raise ValueError(
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model."
)
elif expected_add_embed_dim != passed_add_embed_dim:
raise ValueError(
f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
)
add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype)
return add_time_ids, add_neg_time_ids
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
def upcast_vae(self):
dtype = self.vae.dtype
self.vae.to(dtype=torch.float32)
use_torch_2_0_or_xformers = isinstance(
self.vae.decoder.mid_block.attentions[0].processor,
(
AttnProcessor2_0,
XFormersAttnProcessor,
LoRAXFormersAttnProcessor,
LoRAAttnProcessor2_0,
),
)
# if xformers or torch_2_0 is used attention block does not need
# to be in float32 which can save lots of memory
if use_torch_2_0_or_xformers:
self.vae.post_quant_conv.to(dtype)
self.vae.decoder.conv_in.to(dtype)
self.vae.decoder.mid_block.to(dtype)
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(
self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32
) -> torch.FloatTensor:
"""
See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
Args:
w (`torch.Tensor`):
Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings.
embedding_dim (`int`, *optional*, defaults to 512):
Dimension of the embeddings to generate.
dtype (`torch.dtype`, *optional*, defaults to `torch.float32`):
Data type of the generated embeddings.
Returns:
`torch.FloatTensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
"""
assert len(w.shape) == 1
w = w * 1000.0
half_dim = embedding_dim // 2
emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
emb = w.to(dtype)[:, None] * emb[None, :]
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
if embedding_dim % 2 == 1: # zero pad
emb = torch.nn.functional.pad(emb, (0, 1))
assert emb.shape == (w.shape[0], embedding_dim)
return emb
@property
def guidance_scale(self):
return self._guidance_scale
@property
def guidance_rescale(self):
return self._guidance_rescale
@property
def clip_skip(self):
return self._clip_skip
# 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.
@property
def do_classifier_free_guidance(self):
return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
@property
def cross_attention_kwargs(self):
return self._cross_attention_kwargs
@property
def denoising_end(self):
return self._denoising_end
@property
def denoising_start(self):
return self._denoising_start
@property
def num_timesteps(self):
return self._num_timesteps
@property
def interrupt(self):
return self._interrupt
@torch.no_grad()
@replace_example_docstring(EXAMPLE_DOC_STRING)
def __call__(
self,
prompt: Union[str, List[str]] = None,
prompt_2: Optional[Union[str, List[str]]] = None,
image: Union[
torch.FloatTensor,
PIL.Image.Image,
np.ndarray,
List[torch.FloatTensor],
List[PIL.Image.Image],
List[np.ndarray],
] = None,
strength: float = 0.3,
num_inference_steps: int = 50,
timesteps: List[int] = None,
denoising_start: Optional[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.FloatTensor] = None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
ip_adapter_image: Optional[PipelineImageInput] = None,
ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
guidance_rescale: float = 0.0,
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,
aesthetic_score: float = 6.0,
negative_aesthetic_score: float = 2.5,
clip_skip: Optional[int] = None,
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
map: torch.FloatTensor = None,
original_image: Union[
torch.FloatTensor,
PIL.Image.Image,
np.ndarray,
List[torch.FloatTensor],
List[PIL.Image.Image],
List[np.ndarray],
] = None,
**kwargs,
):
r"""
Function invoked when calling the pipeline for generation.
Args:
prompt (`str` or `List[str]`, *optional*):
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
instead.
prompt_2 (`str` or `List[str]`, *optional*):
The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
used in both text-encoders
image (`torch.FloatTensor` or `PIL.Image.Image` or `np.ndarray` or `List[torch.FloatTensor]` or `List[PIL.Image.Image]` or `List[np.ndarray]`):
The image(s) to modify with the pipeline.
strength (`float`, *optional*, defaults to 0.3):
Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
will be used as a starting point, adding more noise to it the larger the `strength`. The number of
denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
be maximum and the denoising process will run for the full number of iterations specified in
`num_inference_steps`. A value of 1, therefore, essentially ignores `image`. Note that in the case of
`denoising_start` being declared as an integer, the value of `strength` will be ignored.
num_inference_steps (`int`, *optional*, defaults to 50):
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
expense of slower inference.
denoising_start (`float`, *optional*):
When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be
bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and
it is assumed that the passed `image` is a partly denoised image. Note that when this is specified,
strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline
is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refining the Image
Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
denoising_end (`float`, *optional*):
When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
completed before it is intentionally prematurely terminated. As a result, the returned sample will
still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be
denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the
final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline
forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
guidance_scale (`float`, *optional*, defaults to 7.5):
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
`guidance_scale` is defined as `w` of equation 2. of [Imagen
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
usually at the expense of lower image quality.
negative_prompt (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation. If not defined, one has to pass
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
less than `1`).
negative_prompt_2 (`str` or `List[str]`, *optional*):
The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
`text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
num_images_per_prompt (`int`, *optional*, defaults to 1):
The number of images to generate per prompt.
eta (`float`, *optional*, defaults to 0.0):
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
[`schedulers.DDIMScheduler`], will be ignored for others.
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
to make generation deterministic.
latents (`torch.FloatTensor`, *optional*):
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
tensor will ge generated by sampling using the supplied random `generator`.
prompt_embeds (`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.
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
argument.
pooled_prompt_embeds (`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.
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
input argument.
ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
if `do_classifier_free_guidance` is set to `True`.
If not provided, embeddings are computed from the `ip_adapter_image` input argument.
output_type (`str`, *optional*, defaults to `"pil"`):
The output format of the generate image. Choose between
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] instead of a
plain tuple.
callback (`Callable`, *optional*):
A function that will be called every `callback_steps` steps during inference. The function will be
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
callback_steps (`int`, *optional*, defaults to 1):
The frequency at which the `callback` function will be called. If not specified, the callback will be
called at every step.
cross_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
`self.processor` in
[diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
guidance_rescale (`float`, *optional*, defaults to 0.7):
Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
[Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
Guidance rescale factor should fix overexposure when using zero terminal SNR.
original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
`original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
`crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
`crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
`crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
For most cases, `target_size` should be set to the desired height and width of the generated image. If
not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
aesthetic_score (`float`, *optional*, defaults to 6.0):
Used to simulate an aesthetic score of the generated image by influencing the positive text condition.
Part of SDXL's micro-conditioning as explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
Part of SDXL's micro-conditioning as explained in section 2.2 of
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to
simulate an aesthetic score of the generated image by influencing the negative text condition.
Examples:
Returns:
[`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`:
[`~pipelines.stable_diffusion.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 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`",
)
# 1. Check inputs. Raise error if not correct
self.check_inputs(
prompt,
prompt_2,
strength,
num_inference_steps,
callback_steps,
negative_prompt,
negative_prompt_2,
prompt_embeds,
negative_prompt_embeds,
ip_adapter_image,
ip_adapter_image_embeds,
callback_on_step_end_tensor_inputs,
)
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._denoising_start = denoising_start
self._interrupt = False
# 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
text_encoder_lora_scale = (
cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
)
(
prompt_embeds,
negative_prompt_embeds,
pooled_prompt_embeds,
negative_pooled_prompt_embeds,
) = self.encode_prompt(
prompt=prompt,
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,
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,
)
# 4. Preprocess image
# image = self.image_processor.preprocess(image) #ideally we would have preprocess the image with diffusers, but for this POC we won't --- it throws a deprecated warning
map = torchvision.transforms.Resize(
tuple(s // self.vae_scale_factor for s in original_image.shape[2:]), antialias=None
)(map)
# 5. Prepare timesteps
def denoising_value_valid(dnv):
return isinstance(dnv, float) and 0 < dnv < 1
timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
# begin diff diff change
total_time_steps = num_inference_steps
# end diff diff change
timesteps, num_inference_steps = self.get_timesteps(
num_inference_steps,
strength,
device,
denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None,
)
latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
add_noise = True if denoising_start is None else False
# 6. Prepare latent variables
latents = self.prepare_latents(
image,
latent_timestep,
batch_size,
num_images_per_prompt,
prompt_embeds.dtype,
device,
generator,
add_noise,
)
# 7. Prepare extra step kwargs.
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
height, width = latents.shape[-2:]
height = height * self.vae_scale_factor
width = width * self.vae_scale_factor
original_size = original_size or (height, width)
target_size = target_size or (height, width)
# 8. Prepare added time ids & embeddings
if negative_original_size is None:
negative_original_size = original_size
if negative_target_size is None:
negative_target_size = target_size
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, add_neg_time_ids = self._get_add_time_ids(
original_size,
crops_coords_top_left,
target_size,
aesthetic_score,
negative_aesthetic_score,
negative_original_size,
negative_crops_coords_top_left,
negative_target_size,
dtype=prompt_embeds.dtype,
text_encoder_projection_dim=text_encoder_projection_dim,
)
add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1)
if self.do_classifier_free_guidance:
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1)
add_time_ids = torch.cat([add_neg_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)
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,
)
# 9. Denoising loop
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
# 9.1 Apply denoising_end
if (
denoising_end is not None
and denoising_start is not None
and denoising_value_valid(denoising_end)
and denoising_value_valid(denoising_start)
and denoising_start >= denoising_end
):
raise ValueError(
f"`denoising_start`: {denoising_start} cannot be larger than or equal to `denoising_end`: "
+ f" {denoising_end} when using type float."
)
elif denoising_end is not None and denoising_value_valid(denoising_end):
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]
# preparations for diff diff
original_with_noise = self.prepare_latents(
original_image, timesteps, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator
)
thresholds = torch.arange(total_time_steps, dtype=map.dtype) / total_time_steps
thresholds = thresholds.unsqueeze(1).unsqueeze(1).to(device)
masks = map > (thresholds + (denoising_start or 0))
# end diff diff preparations
# 9.2 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:
for i, t in enumerate(timesteps):
if self.interrupt:
continue
# diff diff
if i == 0 and denoising_start is None:
latents = original_with_noise[:1]
else:
mask = masks[i].unsqueeze(0)
# cast mask to the same type as latents etc
mask = mask.to(latents.dtype)
mask = mask.unsqueeze(1) # fit shape
latents = original_with_noise[i] * mask + latents * (1 - mask)
# end diff diff
# 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)
# predict the noise residual
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
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=prompt_embeds,
timestep_cond=timestep_cond,
cross_attention_kwargs=cross_attention_kwargs,
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_dtype = latents.dtype
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)
else:
raise ValueError(
"For the given accelerator, there seems to be an unexpected problem in type-casting. Please file an issue on the PyTorch GitHub repository. See also: https://github.com/huggingface/diffusers/pull/7446/."
)
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)
add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_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 XLA_AVAILABLE:
xm.mark_step()
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)
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)
else:
raise ValueError(
"For the given accelerator, there seems to be an unexpected problem in type-casting. Please file an issue on the PyTorch GitHub repository. See also: https://github.com/huggingface/diffusers/pull/7446/."
)
# 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
# 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)
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