add MultiDiffusionPanorama pipeline (#2393)

* add MultiDiffusionPanorama pipeline

* fix docs naming

* update pipeline name, remove redundant tests

* apply styling.

* debugging information.

* fix: assertion values.

* fix-copies.

* update docs

* update docs

---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

docs/source/en/_toctree.yml

@@ -161,6 +161,8 @@
         title: Pix2Pix Zero
       - local: api/pipelines/stable_diffusion/self_attention_guidance
         title: Self-Attention Guidance
+      - local: api/pipelines/stable_diffusion/panorama
+        title: MultiDiffusion Panorama
       title: Stable Diffusion
     - local: api/pipelines/stable_diffusion_2
       title: Stable Diffusion 2

docs/source/en/api/pipelines/stable_diffusion/panorama.mdx

+<!--Copyright 2023 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.
+-->
+
+# MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation
+
+## Overview
+
+[MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation](https://arxiv.org/abs/2302.08113) by Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel.
+
+The abstract of the paper is the following:
+
+*Recent advances in text-to-image generation with diffusion models present transformative capabilities in image quality. However, user controllability of the generated image, and fast adaptation to new tasks still remains an open challenge, currently mostly addressed by costly and long re-training and fine-tuning or ad-hoc adaptations to specific image generation tasks. In this work, we present MultiDiffusion, a unified framework that enables versatile and controllable image generation, using a pre-trained text-to-image diffusion model, without any further training or finetuning. At the center of our approach is a new generation process, based on an optimization task that binds together multiple diffusion generation processes with a shared set of parameters or constraints. We show that MultiDiffusion can be readily applied to generate high quality and diverse images that adhere to user-provided controls, such as desired aspect ratio (e.g., panorama), and spatial guiding signals, ranging from tight segmentation masks to bounding boxes.
+
+Resources:
+
+* [Project Page](https://multidiffusion.github.io/).
+* [Paper](https://arxiv.org/abs/2302.08113).
+* [Original Code](https://github.com/omerbt/MultiDiffusion).
+
+## Available Pipelines:
+
+| Pipeline | Tasks
+|---|---|
+| [StableDiffusionPanoramaPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_panorama.py) | *Text-Guided Panorama View Generation* |
+
+<!-- TODO: add Colab -->
+
+## Usage example
+
+```python
+import torch
+from diffusers import StableDiffusionPanoramaPipeline, DDIMScheduler
+
+model_ckpt = "stabilityai/stable-diffusion-2-base"
+scheduler = DDIMScheduler.from_pretrained(model_ckpt, subfolder="scheduler")
+pipe = StableDiffusionPanoramaPipeline.from_pretrained(model_ckpt, scheduler=scheduler, torch_dtype=torch.float16)
+
+pipe = pipe.to("cuda")
+
+prompt = "a photo of the dolomites"
+image = pipe(prompt).images[0]
+image.save("dolomites.png")
+```
+
+## StableDiffusionPanoramaPipeline
+[[autodoc]] StableDiffusionPanoramaPipeline
+	- __call__
+	- all

docs/source/en/index.mdx

@@ -49,6 +49,7 @@ available a colab notebook to directly try them out.
 | [score_sde_vp](./api/pipelines/score_sde_vp) | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | Unconditional Image Generation | 
 | [semantic_stable_diffusion](./api/pipelines/semantic_stable_diffusion) | [**Semantic Guidance**](https://arxiv.org/abs/2301.12247) | Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ml-research/semantic-image-editing/blob/main/examples/SemanticGuidance.ipynb)
 | [stable_diffusion](./api/pipelines/stable_diffusion/text2img) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-to-Image Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb)
+| [stable_diffusion](./api/pipelines/stable_diffusion/panorama) | [**MultiDiffusion**](https://multidiffusion.github.io/) | Text-to-Panorama Generation |
 | [stable_diffusion](./api/pipelines/stable_diffusion/img2img) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Image-to-Image Text-Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb)
 | [stable_diffusion](./api/pipelines/stable_diffusion/inpaint) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | Text-Guided Image Inpainting | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb)
 | [stable_diffusion_2](./api/pipelines/stable_diffusion_2) | [**Stable Diffusion 2**](https://stability.ai/blog/stable-diffusion-v2-release) | Text-to-Image Generation |

src/diffusers/__init__.py

@@ -120,6 +120,7 @@ else:
         StableDiffusionInpaintPipelineLegacy,
         StableDiffusionInstructPix2PixPipeline,
         StableDiffusionLatentUpscalePipeline,
+        StableDiffusionPanoramaPipeline,
         StableDiffusionPipeline,
         StableDiffusionPipelineSafe,
         StableDiffusionPix2PixZeroPipeline,

src/diffusers/pipelines/__init__.py

@@ -55,6 +55,7 @@ else:
         StableDiffusionInpaintPipelineLegacy,
         StableDiffusionInstructPix2PixPipeline,
         StableDiffusionLatentUpscalePipeline,
+        StableDiffusionPanoramaPipeline,
         StableDiffusionPipeline,
         StableDiffusionPix2PixZeroPipeline,
         StableDiffusionSAGPipeline,

src/diffusers/pipelines/stable_diffusion/__init__.py

@@ -50,6 +50,7 @@ else:
     from .pipeline_stable_diffusion_inpaint_legacy import StableDiffusionInpaintPipelineLegacy
     from .pipeline_stable_diffusion_instruct_pix2pix import StableDiffusionInstructPix2PixPipeline
     from .pipeline_stable_diffusion_latent_upscale import StableDiffusionLatentUpscalePipeline
+    from .pipeline_stable_diffusion_panorama import StableDiffusionPanoramaPipeline
     from .pipeline_stable_diffusion_sag import StableDiffusionSAGPipeline
     from .pipeline_stable_diffusion_upscale import StableDiffusionUpscalePipeline
     from .pipeline_stable_unclip import StableUnCLIPPipeline

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -212,6 +212,21 @@ class StableDiffusionLatentUpscalePipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class StableDiffusionPanoramaPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class StableDiffusionPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/stable_diffusion/test_stable_diffusion_panorama.py

+# coding=utf-8
+# Copyright 2023 HuggingFace Inc.
+#
+# 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 gc
+import unittest
+
+import numpy as np
+import torch
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
+
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    EulerAncestralDiscreteScheduler,
+    LMSDiscreteScheduler,
+    PNDMScheduler,
+    StableDiffusionPanoramaPipeline,
+    UNet2DConditionModel,
+)
+from diffusers.utils import slow, torch_device
+from diffusers.utils.testing_utils import require_torch_gpu
+
+from ...test_pipelines_common import PipelineTesterMixin
+
+
+torch.backends.cuda.matmul.allow_tf32 = False
+
+
+class StableDiffusionPanoramaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = StableDiffusionPanoramaPipeline
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        unet = UNet2DConditionModel(
+            block_out_channels=(32, 64),
+            layers_per_block=2,
+            sample_size=32,
+            in_channels=4,
+            out_channels=4,
+            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
+            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+            cross_attention_dim=32,
+        )
+        scheduler = DDIMScheduler()
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            block_out_channels=[32, 64],
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+            latent_channels=4,
+        )
+        torch.manual_seed(0)
+        text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+        )
+        text_encoder = CLIPTextModel(text_encoder_config)
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        components = {
+            "unet": unet,
+            "scheduler": scheduler,
+            "vae": vae,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "safety_checker": None,
+            "feature_extractor": None,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        generator = torch.manual_seed(seed)
+        inputs = {
+            "prompt": "a photo of the dolomites",
+            "generator": generator,
+            # Setting height and width to None to prevent OOMs on CPU.
+            "height": None,
+            "width": None,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "output_type": "numpy",
+        }
+        return inputs
+
+    def test_stable_diffusion_panorama_default_case(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionPanoramaPipeline(**components)
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = sd_pipe(**inputs).images
+        image_slice = image[0, -3:, -3:, -1]
+        assert image.shape == (1, 64, 64, 3)
+
+        expected_slice = np.array([0.5101, 0.5006, 0.4962, 0.3995, 0.3501, 0.4632, 0.5339, 0.525, 0.4878])
+
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+
+    def test_stable_diffusion_panorama_negative_prompt(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionPanoramaPipeline(**components)
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        negative_prompt = "french fries"
+        output = sd_pipe(**inputs, negative_prompt=negative_prompt)
+        image = output.images
+        image_slice = image[0, -3:, -3:, -1]
+
+        assert image.shape == (1, 64, 64, 3)
+
+        expected_slice = np.array([0.5326, 0.5009, 0.5074, 0.4133, 0.371, 0.464, 0.5432, 0.5429, 0.4896])
+
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+
+    def test_stable_diffusion_panorama_euler(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        components["scheduler"] = EulerAncestralDiscreteScheduler(
+            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear"
+        )
+        sd_pipe = StableDiffusionPanoramaPipeline(**components)
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = sd_pipe(**inputs).images
+        image_slice = image[0, -3:, -3:, -1]
+
+        assert image.shape == (1, 64, 64, 3)
+
+        expected_slice = np.array(
+            [0.48235387, 0.5423796, 0.46016198, 0.5377287, 0.5803722, 0.4876525, 0.5515428, 0.5045897, 0.50709957]
+        )
+
+        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
+
+    def test_stable_diffusion_panorama_pndm(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        components["scheduler"] = PNDMScheduler()
+        sd_pipe = StableDiffusionPanoramaPipeline(**components)
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        # the pipeline does not expect pndm so test if it raises error.
+        with self.assertRaises(ValueError):
+            _ = sd_pipe(**inputs).images
+
+    def test_stable_diffusion_panorama_num_images_per_prompt(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+        sd_pipe = StableDiffusionPanoramaPipeline(**components)
+        sd_pipe = sd_pipe.to(device)
+        sd_pipe.set_progress_bar_config(disable=None)
+
+        # test num_images_per_prompt=1 (default)
+        inputs = self.get_dummy_inputs(device)
+        images = sd_pipe(**inputs).images
+
+        assert images.shape == (1, 64, 64, 3)
+
+        # test num_images_per_prompt=1 (default) for batch of prompts
+        batch_size = 2
+        inputs = self.get_dummy_inputs(device)
+        inputs["prompt"] = [inputs["prompt"]] * batch_size
+        images = sd_pipe(**inputs).images
+
+        assert images.shape == (batch_size, 64, 64, 3)
+
+        # test num_images_per_prompt for single prompt
+        num_images_per_prompt = 2
+        inputs = self.get_dummy_inputs(device)
+        images = sd_pipe(**inputs, num_images_per_prompt=num_images_per_prompt).images
+
+        assert images.shape == (num_images_per_prompt, 64, 64, 3)
+
+        # test num_images_per_prompt for batch of prompts
+        batch_size = 2
+        inputs = self.get_dummy_inputs(device)
+        inputs["prompt"] = [inputs["prompt"]] * batch_size
+        images = sd_pipe(**inputs, num_images_per_prompt=num_images_per_prompt).images
+
+        assert images.shape == (batch_size * num_images_per_prompt, 64, 64, 3)
+
+
+@slow
+@require_torch_gpu
+class StableDiffusionPanoramaSlowTests(unittest.TestCase):
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def get_inputs(self, seed=0):
+        generator = torch.manual_seed(seed)
+        inputs = {
+            "prompt": "a photo of the dolomites",
+            "generator": generator,
+            "num_inference_steps": 3,
+            "guidance_scale": 7.5,
+            "output_type": "numpy",
+        }
+        return inputs
+
+    def test_stable_diffusion_panorama_default(self):
+        model_ckpt = "stabilityai/stable-diffusion-2-base"
+        scheduler = DDIMScheduler.from_pretrained(model_ckpt, subfolder="scheduler")
+        pipe = StableDiffusionPanoramaPipeline.from_pretrained(model_ckpt, scheduler=scheduler, safety_checker=None)
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing()
+
+        inputs = self.get_inputs()
+        image = pipe(**inputs).images
+        image_slice = image[0, -3:, -3:, -1].flatten()
+
+        assert image.shape == (1, 512, 2048, 3)
+
+        expected_slice = np.array(
+            [
+                0.36968392,
+                0.27025372,
+                0.32446766,
+                0.28379387,
+                0.36363274,
+                0.30733347,
+                0.27100027,
+                0.27054125,
+                0.25536096,
+            ]
+        )
+
+        assert np.abs(expected_slice - image_slice).max() < 1e-2
+
+    def test_stable_diffusion_panorama_k_lms(self):
+        pipe = StableDiffusionPanoramaPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-2-base", safety_checker=None
+        )
+        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing()
+
+        inputs = self.get_inputs()
+        image = pipe(**inputs).images
+        image_slice = image[0, -3:, -3:, -1].flatten()
+
+        assert image.shape == (1, 512, 2048, 3)
+
+        expected_slice = np.array(
+            [
+                [
+                    0.0,
+                    0.0,
+                    0.0,
+                    0.0,
+                    0.0,
+                    0.0,
+                    0.0,
+                    0.0,
+                    0.0,
+                ]
+            ]
+        )
+
+        assert np.abs(expected_slice - image_slice).max() < 1e-3
+
+    def test_stable_diffusion_panorama_intermediate_state(self):
+        number_of_steps = 0
+
+        def callback_fn(step: int, timestep: int, latents: torch.FloatTensor) -> None:
+            callback_fn.has_been_called = True
+            nonlocal number_of_steps
+            number_of_steps += 1
+            if step == 1:
+                latents = latents.detach().cpu().numpy()
+                assert latents.shape == (1, 4, 64, 256)
+                latents_slice = latents[0, -3:, -3:, -1]
+
+                expected_slice = np.array(
+                    [
+                        0.18681869,
+                        0.33907816,
+                        0.5361276,
+                        0.14432865,
+                        -0.02856611,
+                        -0.73941123,
+                        0.23397987,
+                        0.47322682,
+                        -0.37823164,
+                    ]
+                )
+                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2
+            elif step == 2:
+                latents = latents.detach().cpu().numpy()
+                assert latents.shape == (1, 4, 64, 256)
+                latents_slice = latents[0, -3:, -3:, -1]
+
+                expected_slice = np.array(
+                    [
+                        0.18539645,
+                        0.33987248,
+                        0.5378559,
+                        0.14437142,
+                        -0.02455261,
+                        -0.7338317,
+                        0.23990755,
+                        0.47356272,
+                        -0.3786505,
+                    ]
+                )
+
+                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2
+
+        callback_fn.has_been_called = False
+
+        model_ckpt = "stabilityai/stable-diffusion-2-base"
+        scheduler = DDIMScheduler.from_pretrained(model_ckpt, subfolder="scheduler")
+        pipe = StableDiffusionPanoramaPipeline.from_pretrained(model_ckpt, scheduler=scheduler, safety_checker=None)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing()
+
+        inputs = self.get_inputs()
+        pipe(**inputs, callback=callback_fn, callback_steps=1)
+        assert callback_fn.has_been_called
+        assert number_of_steps == 3
+
+    def test_stable_diffusion_panorama_pipeline_with_sequential_cpu_offloading(self):
+        torch.cuda.empty_cache()
+        torch.cuda.reset_max_memory_allocated()
+        torch.cuda.reset_peak_memory_stats()
+
+        model_ckpt = "stabilityai/stable-diffusion-2-base"
+        scheduler = DDIMScheduler.from_pretrained(model_ckpt, subfolder="scheduler")
+        pipe = StableDiffusionPanoramaPipeline.from_pretrained(model_ckpt, scheduler=scheduler, safety_checker=None)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        pipe.enable_attention_slicing(1)
+        pipe.enable_sequential_cpu_offload()
+
+        inputs = self.get_inputs()
+        _ = pipe(**inputs)
+
+        mem_bytes = torch.cuda.max_memory_allocated()
+        # make sure that less than 5.2 GB is allocated
+        assert mem_bytes < 5.2 * 10**9