[ldm3d] Ldm3d upscaler to community pipeline (#5870)

---------
Co-authored-by: Aflalo <estellea@isl-gpu27.rr.intel.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>

docs/source/en/_toctree.yml

@@ -334,7 +334,7 @@
       - local: api/pipelines/stable_diffusion/upscale
         title: Super-resolution
       - local: api/pipelines/stable_diffusion/ldm3d_diffusion
-        title: LDM3D Text-to-(RGB, Depth)
+        title: LDM3D Text-to-(RGB, Depth), Text-to-(RGB-pano, Depth-pano), LDM3D Upscaler
       - local: api/pipelines/stable_diffusion/adapter
         title: Stable Diffusion T2I-Adapter
       - local: api/pipelines/stable_diffusion/gligen

docs/source/en/api/pipelines/overview.md

@@ -53,7 +53,7 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 | [Kandinsky 2.2](kandinsky_v22) | text2image, image2image, inpainting |
 | [Latent Consistency Models](latent_consistency_models) | text2image |
 | [Latent Diffusion](latent_diffusion) | text2image, super-resolution |
-| [LDM3D](stable_diffusion/ldm3d_diffusion) | text2image, text-to-3D |
+| [LDM3D](stable_diffusion/ldm3d_diffusion) | text2image, text-to-3D, text-to-pano, upscaling |
 | [MultiDiffusion](panorama) | text2image |
 | [MusicLDM](musicldm) | text2audio |
 | [Paint by Example](paint_by_example) | inpainting |

docs/source/en/api/pipelines/stable_diffusion/ldm3d_diffusion.md

@@ -14,6 +14,11 @@ specific language governing permissions and limitations under the License.
 
 LDM3D was proposed in [LDM3D: Latent Diffusion Model for 3D](https://huggingface.co/papers/2305.10853) by Gabriela Ben Melech Stan, Diana Wofk, Scottie Fox, Alex Redden, Will Saxton, Jean Yu, Estelle Aflalo, Shao-Yen Tseng, Fabio Nonato, Matthias Muller, and Vasudev Lal. LDM3D generates an image and a depth map from a given text prompt unlike the existing text-to-image diffusion models such as [Stable Diffusion](./overview) which only generates an image. With almost the same number of parameters, LDM3D achieves to create a latent space that can compress both the RGB images and the depth maps. 
 
+Two checkpoints are available for use:
+- [ldm3d-original](https://huggingface.co/Intel/ldm3d). The original checkpoint used in the [paper](https://arxiv.org/pdf/2305.10853.pdf)
+- [ldm3d-4c](https://huggingface.co/Intel/ldm3d-4c). The new version of LDM3D using 4 channels inputs instead of 6-channels inputs and finetuned on higher resolution images. 
+
+
 The abstract from the paper is:
 
 *This research paper proposes a Latent Diffusion Model for 3D (LDM3D) that generates both image and depth map data from a given text prompt, allowing users to generate RGBD images from text prompts. The LDM3D model is fine-tuned on a dataset of tuples containing an RGB image, depth map and caption, and validated through extensive experiments. We also develop an application called DepthFusion, which uses the generated RGB images and depth maps to create immersive and interactive 360-degree-view experiences using TouchDesigner. This technology has the potential to transform a wide range of industries, from entertainment and gaming to architecture and design. Overall, this paper presents a significant contribution to the field of generative AI and computer vision, and showcases the potential of LDM3D and DepthFusion to revolutionize content creation and digital experiences. A short video summarizing the approach can be found at [this url](https://t.ly/tdi2).*
@@ -26,12 +31,25 @@ Make sure to check out the Stable Diffusion [Tips](overview#tips) section to lea
 
 ## StableDiffusionLDM3DPipeline
 
-[[autodoc]] StableDiffusionLDM3DPipeline
+[[autodoc]] pipelines.stable_diffusion.pipeline_stable_diffusion_ldm3d.StableDiffusionLDM3DPipeline
 	- all
 	- __call__
 
+
 ## LDM3DPipelineOutput
 
 [[autodoc]] pipelines.stable_diffusion.pipeline_stable_diffusion_ldm3d.LDM3DPipelineOutput
 	- all
 	- __call__
+
+# Upscaler
+
+[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 from communauty pipeline.
+

docs/source/en/api/pipelines/stable_diffusion/overview.md

@@ -121,10 +121,16 @@ The table below summarizes the available Stable Diffusion pipelines, their suppo
             <td class="px-4 py-2 text-gray-700">
             <a href="./ldm3d_diffusion">StableDiffusionLDM3D</a>
             </td>
-            <td class="px-4 py-2 text-gray-700">text-to-rgb, text-to-depth</td>
+            <td class="px-4 py-2 text-gray-700">text-to-rgb, text-to-depth, text-to-pano</td>
             <td class="px-4 py-2"><a href="https://huggingface.co/spaces/r23/ldm3d-space"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue"/></a>
             </td>
         </tr>
+        <tr>
+            <td class="px-4 py-2 text-gray-700">
+            <a href="./ldm3d_diffusion">StableDiffusionUpscaleLDM3D</a>
+            </td>
+            <td class="px-4 py-2 text-gray-700">ldm3d super-resolution</td>
+        </tr>
         </tbody>
     </table>
     </div>

docs/source/ja/index.md

@@ -96,3 +96,4 @@ specific language governing permissions and limitations under the License.
 | [versatile_diffusion](./api/pipelines/versatile_diffusion) | [Versatile Diffusion: Text, Images and Variations All in One Diffusion Model](https://arxiv.org/abs/2211.08332) | Dual Image and Text Guided Generation |
 | [vq_diffusion](./api/pipelines/vq_diffusion) | [Vector Quantized Diffusion Model for Text-to-Image Synthesis](https://arxiv.org/abs/2111.14822) | Text-to-Image Generation |
 | [stable_diffusion_ldm3d](./api/pipelines/stable_diffusion/ldm3d_diffusion) | [LDM3D: Latent Diffusion Model for 3D](https://arxiv.org/abs/2305.10853) | Text to Image and Depth Generation |
+| [stable_diffusion_upscaler_ldm3d](./api/pipelines/stable_diffusion/ldm3d_diffusion) | [LDM3D-VR: Latent Diffusion Model for 3D VR](https://arxiv.org/pdf/2311.03226) | Image and Depth Upscaling |

examples/community/README.md

@@ -48,7 +48,8 @@ prompt-to-prompt | change parts of a prompt and retain image structure (see [pap
 |   Latent Consistency Pipeline                                                                                                    | Implementation of [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://arxiv.org/abs/2310.04378)                                                                                                                                                                                                                                                                                                                                                                                                                                      | [Latent Consistency Pipeline](#latent-consistency-pipeline)      | - |              [Simian Luo](https://github.com/luosiallen) |
 |   Latent Consistency Img2img Pipeline                                                                                                    | Img2img pipeline for Latent Consistency Models                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Img2Img Pipeline](#latent-consistency-img2img-pipeline)      | - |              [Logan Zoellner](https://github.com/nagolinc) |
 |   Latent Consistency Interpolation Pipeline                                                                                                    | Interpolate the latent space of Latent Consistency Models with multiple prompts                                                                                                                                                                                                                                                                                                                                                                                                                                    | [Latent Consistency Interpolation Pipeline](#latent-consistency-interpolation-pipeline) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1pK3NrLWJSiJsBynLns1K1-IDTW9zbPvl?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
-
+| 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) |
+|
 
 To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.
 ```py
@@ -2344,6 +2345,47 @@ images = pipe(
 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.
+
+'''py
+from PIL import Image
+import os
+import torch
+from diffusers import StableDiffusionLDM3DPipeline, DiffusionPipeline
+
+#Generate a rgb/depth output from LDM3D
+pipe_ldm3d = StableDiffusionLDM3DPipeline.from_pretrained("Intel/ldm3d-4c")
+pipe_ldm3d.to("cuda")
+
+prompt =f"A picture of some lemons on a table"
+output = pipe_ldm3d(prompt)
+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)
+pipe_ldm3d_upscale = DiffusionPipeline.from_pretrained("Intel/ldm3d-sr", custom_pipeline="pipeline_stable_diffusion_upscale_ldm3d")
+
+pipe_ldm3d_upscale.to("cuda")
+
+low_res_img = Image.open(f"lemons_ldm3d_rgb.jpg").convert("RGB")
+low_res_depth = Image.open(f"lemons_ldm3d_depth.png").convert("L")
+outputs = pipe_ldm3d_upscale(prompt="high quality high resolution uhd 4k image", rgb=low_res_img, depth=low_res_depth, num_inference_steps=50, target_res=[1024, 1024])
+
+upscaled_rgb, upscaled_depth =outputs.rgb[0], outputs.depth[0]
+upscaled_rgb.save(f"upscaled_lemons_rgb.png")
+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. 

src/diffusers/image_processor.py

@@ -33,6 +33,15 @@ PipelineImageInput = Union[
     List[torch.FloatTensor],
 ]
 
+PipelineDepthInput = Union[
+    PIL.Image.Image,
+    np.ndarray,
+    torch.FloatTensor,
+    List[PIL.Image.Image],
+    List[np.ndarray],
+    List[torch.FloatTensor],
+]
+
 
 class VaeImageProcessor(ConfigMixin):
     """
@@ -441,6 +450,18 @@ class VaeImageProcessorLDM3D(VaeImageProcessor):
 
         return pil_images
 
+    @staticmethod
+    def depth_pil_to_numpy(images: Union[List[PIL.Image.Image], PIL.Image.Image]) -> np.ndarray:
+        """
+        Convert a PIL image or a list of PIL images to NumPy arrays.
+        """
+        if not isinstance(images, list):
+            images = [images]
+
+        images = [np.array(image).astype(np.float32) / (2**16 - 1) for image in images]
+        images = np.stack(images, axis=0)
+        return images
+
     @staticmethod
     def rgblike_to_depthmap(image: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]:
         """
@@ -526,3 +547,102 @@ class VaeImageProcessorLDM3D(VaeImageProcessor):
             return self.numpy_to_pil(image), self.numpy_to_depth(image)
         else:
             raise Exception(f"This type {output_type} is not supported")
+
+    def preprocess(
+        self,
+        rgb: Union[torch.FloatTensor, PIL.Image.Image, np.ndarray],
+        depth: Union[torch.FloatTensor, PIL.Image.Image, np.ndarray],
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        target_res: Optional[int] = None,
+    ) -> torch.Tensor:
+        """
+        Preprocess the image input. Accepted formats are PIL images, NumPy arrays or PyTorch tensors.
+        """
+        supported_formats = (PIL.Image.Image, np.ndarray, torch.Tensor)
+
+        # Expand the missing dimension for 3-dimensional pytorch tensor or numpy array that represents grayscale image
+        if self.config.do_convert_grayscale and isinstance(rgb, (torch.Tensor, np.ndarray)) and rgb.ndim == 3:
+            raise Exception("This is not yet supported")
+
+        if isinstance(rgb, supported_formats):
+            rgb = [rgb]
+            depth = [depth]
+        elif not (isinstance(rgb, list) and all(isinstance(i, supported_formats) for i in rgb)):
+            raise ValueError(
+                f"Input is in incorrect format: {[type(i) for i in rgb]}. Currently, we only support {', '.join(supported_formats)}"
+            )
+
+        if isinstance(rgb[0], PIL.Image.Image):
+            if self.config.do_convert_rgb:
+                raise Exception("This is not yet supported")
+                # rgb = [self.convert_to_rgb(i) for i in rgb]
+                # depth = [self.convert_to_depth(i) for i in depth]  #TODO define convert_to_depth
+            if self.config.do_resize or target_res:
+                height, width = self.get_default_height_width(rgb[0], height, width) if not target_res else target_res
+                rgb = [self.resize(i, height, width) for i in rgb]
+                depth = [self.resize(i, height, width) for i in depth]
+            rgb = self.pil_to_numpy(rgb)  # to np
+            rgb = self.numpy_to_pt(rgb)  # to pt
+
+            depth = self.depth_pil_to_numpy(depth)  # to np
+            depth = self.numpy_to_pt(depth)  # to pt
+
+        elif isinstance(rgb[0], np.ndarray):
+            rgb = np.concatenate(rgb, axis=0) if rgb[0].ndim == 4 else np.stack(rgb, axis=0)
+            rgb = self.numpy_to_pt(rgb)
+            height, width = self.get_default_height_width(rgb, height, width)
+            if self.config.do_resize:
+                rgb = self.resize(rgb, height, width)
+
+            depth = np.concatenate(depth, axis=0) if rgb[0].ndim == 4 else np.stack(depth, axis=0)
+            depth = self.numpy_to_pt(depth)
+            height, width = self.get_default_height_width(depth, height, width)
+            if self.config.do_resize:
+                depth = self.resize(depth, height, width)
+
+        elif isinstance(rgb[0], torch.Tensor):
+            raise Exception("This is not yet supported")
+            # rgb = torch.cat(rgb, axis=0) if rgb[0].ndim == 4 else torch.stack(rgb, axis=0)
+
+            # if self.config.do_convert_grayscale and rgb.ndim == 3:
+            #     rgb = rgb.unsqueeze(1)
+
+            # channel = rgb.shape[1]
+
+            # height, width = self.get_default_height_width(rgb, height, width)
+            # if self.config.do_resize:
+            #     rgb = self.resize(rgb, height, width)
+
+            # depth = torch.cat(depth, axis=0) if depth[0].ndim == 4 else torch.stack(depth, axis=0)
+
+            # if self.config.do_convert_grayscale and depth.ndim == 3:
+            #     depth = depth.unsqueeze(1)
+
+            # channel = depth.shape[1]
+            # # don't need any preprocess if the image is latents
+            # if depth == 4:
+            #     return rgb, depth
+
+            # height, width = self.get_default_height_width(depth, height, width)
+            # if self.config.do_resize:
+            #     depth = self.resize(depth, height, width)
+        # expected range [0,1], normalize to [-1,1]
+        do_normalize = self.config.do_normalize
+        if rgb.min() < 0 and do_normalize:
+            warnings.warn(
+                "Passing `image` as torch tensor with value range in [-1,1] is deprecated. The expected value range for image tensor is [0,1] "
+                f"when passing as pytorch tensor or numpy Array. You passed `image` with value range [{rgb.min()},{rgb.max()}]",
+                FutureWarning,
+            )
+            do_normalize = False
+
+        if do_normalize:
+            rgb = self.normalize(rgb)
+            depth = self.normalize(depth)
+
+        if self.config.do_binarize:
+            rgb = self.binarize(rgb)
+            depth = self.binarize(depth)
+
+        return rgb, depth