Upload MatForgerPipeline

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+img/MatForger_gen-text.png filter=lfs diff=lfs merge=lfs -text
+img/MatForger_gen-img.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,151 @@
+---
+library_name: diffusers
+datasets:
+- gvecchio/MatSynth
+language:
+- en
+tags:
+- pbr
+- materials
+- svbrdf
+- 3d
+- textures
+license: openrail
+---
+
+<!-- # ⚒️ MatForger -->
+![alt text](./img/MatForger.png)
+
+
+> Three Textures for the Designers under the sky, \
+> Seven for the Artists in their studios of light, \
+> Nine for the Architects doomed to try, \
+> One for the Developers on their screens so bright \
+> In the Land of Graphics where the Pixels lie. \
+> One Forge to craft them all, One Code to find them, \
+> One Model to bring them all and to the mesh bind them \
+> In the Land of Graphics where the Pixels lie.
+
+<sup><sub>Our deep apologies to J. R. R. Tolkien</sub></sup>
+
+
+## 🤖 Model Details
+
+### Overview
+
+**MatForger** is a generative diffusion model designed specifically for generating Physically Based Rendering (PBR) materials. Inspired by the [MatFuse](https://arxiv.org/abs/2308.11408) model and trained on the comprehensive [MatSynth](https://huggingface.co/datasets/gvecchio/MatSynth) dataset, MatForger pushes the boundaries of material synthesis. 
+It employs the noise rolling technique, derived from [ControlMat](https://arxiv.org/abs/2309.01700), to produce tileable maps. The model generates multiple maps, including basecolor, normal, height, roughness, and metallic, catering to a wide range of material design needs.
+
+### Features
+- **High-Quality PBR Material Generation:** Produces detailed and realistic materials suited for various applications.
+- **Tileable Textures:** Utilizes a noise rolling approach to ensure textures are tileable, enhancing their usability in larger scenes.
+- **Versatile Outputs:** Generates multiple texture maps (basecolor, normal, height, roughness, metallic) to meet the requirements of complex material designs.
+- **Text and Image Conditioning:** Can be conditioned with either images or text inputs to guide material generation, offering flexibility in creative workflows.
+
+### Model Description
+MatForger architecture is based on **MatFuse**. It differs from it in using a continuous VAE instead of a vector quantized autoencoder (VQ-VAE). Additionally we distilled the multiencoder VAE into a single-encoder model, thus reducing the model complexity but retaining the disentangled latent representation of MatFuse.
+
+## ⚒️ MatForger at work
+
+MatForger can be conditioned via text prompts or images to generate high-quality materials. Following some examples of materials generated using MatForge. For each sample we report the prompt, the generated maps (basecolor, normal, height, roughness, metallic) and the resulting rendering.
+
+<details>
+    <summary>Text2Mat samples</summary>
+    <img src="./img/MatForger_gen-text.png" alt="Text2Mat generation samples">
+</details>
+
+<details>
+    <summary>Image2Mat samples</summary>
+    <img src="./img/MatForger_gen-img.png" alt="Image2Mat generation samples">
+</details>
+
+## 🧑‍💻 How to use
+
+MatForger requires a custom pipeline due to the data type.
+
+You can use it in [🧨 diffusers](https://github.com/huggingface/diffusers):
+
+```python
+import torch
+
+from PIL import Image
+
+from diffusers import DiffusionPipeline
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+pipe = DiffusionPipeline.from_pretrained(
+    "gvecchio/MatForger",
+    trust_remote_code=True,
+)
+
+pipe.enable_vae_tiling()
+
+pipe.enable_freeu(s1=0.9, s2=0.2, b1=1.1, b2=1.2)
+pipe.to(device)
+
+# model prompting with image
+prompt = Image.open("bricks.png")
+image = pipe(
+    prompt,
+    guidance_scale=6.0,
+    height=512,
+    width=512,
+    num_inference_steps=25,
+).images[0]
+
+# model promptiong with text
+prompt = "terracotta brick wall"
+image = pipe(
+    prompt,
+    guidance_scale=6.0,
+    height=512,
+    width=512,
+    num_inference_steps=25,
+).images[0]
+
+# get maps from prediction
+basecolor = image.basecolor
+normal = image.normal
+height = image.height
+roughness = image.roughness
+metallic = image.metallic
+
+```
+
+## 📉 Bias and Limitations
+
+The model was trained on a variety of synthetic and real data from the [MatSynth](https://huggingface.co/datasets/gvecchio/MatSynth) dataset. 
+However, it might fail to generate complex materials or patterns that differ significantly from the training data distribution.
+Additionally, the model can be conditioned using either images or text, however it might give unexpected results when promped with text as it was mainly trained to do img2material.
+
+**Note:** MatForge is a home-trained model, with limited resources. We will try to keepe it regularly updated and improve its performances. \
+We welcome contributions, feedback, and suggestions to enhance its capabilities and address its limitations. Please be patient as we work towards making MatForger an even more powerful tool for the creative community.
+
+## 💡 Upcoming features ideas
+
+As MatForger continues to evolve, we're working on several features aimed at enhancing its utility and effectiveness. As we continue to refine and expand its capabilities, here are some of the possible upcoming enhancements:
+
+- **Opacity**: Generate opacity map for materials requiring transparency.
+
+- **Material Inpainting**: A feature designed to allow users to modify and enhance materials by filling in gaps or correcting imperfections directly within the generated textures.
+
+- **Sketch-Based Material Generation**: We're exploring ways to convert simple sketches into detailed materials. This aims to simplify the material creation process, making it more accessible to users without in-depth technical expertise.
+
+- **Color Palette Conditioning**: Future updates will offer improved control over the color palette of generated materials, enabling users to achieve more precise color matching for their projects.
+
+- **Material Estimation from Photographs**: We aim to refine the model's ability to interpret and recreate the material properties observed in photographs, facilitating the creation of materials that closely mimic real-world textures.
+
+### 🎯 Ongoing Development and Openness to Feedback
+MatForger is a RESEARCH TOOL, thus its development is an ongoing process and highly subsceptible to our research agenda. 
+However, we are committed to improving MatForger's capabilities and addressing any limitations and implementing suggestions we receive from our users.
+
+### 🤝 How to Contribute
+**Feature Suggestions**: If you have ideas for new features or improvements, we're eager to hear them. Reach out to us! Your suggestions play a crucial role in guiding the direction of MatForger's development.
+
+**Dataset Contributions**: Enhancing the diversity of our training data can significantly improve the model's performance. If you have access to textures, materials, or data that could benefit MatForger, consider contributing.
+
+**Feedback**: User feedback is invaluable for identifying areas for improvement. Whether it's through reporting issues or sharing your experiences, your insights help us make MatForger better.
+
+## Terms of Use:
+We hope that the release of this model will make community-based research efforts more accessible. This model is governed by a Openrail License and is intended for research purposes.

model_index.json

@@ -0,0 +1,20 @@
+{
+  "_class_name": ["pipeline", "MatForgerPipeline"],
+  "_diffusers_version": "0.26.3",
+  "prompt_encoder": [
+    "encoder",
+    "MaterialPromptEncoder"
+  ],
+  "scheduler": [
+    "diffusers",
+    "DDIMScheduler"
+  ],
+  "unet": [
+    "diffusers",
+    "UNet2DConditionModel"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderKL"
+  ]
+}

pipeline.py

@@ -0,0 +1,877 @@
+import inspect
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision.transforms.functional as TF
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin
+from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import (
+    EXAMPLE_DOC_STRING,
+    rescale_noise_cfg,
+    retrieve_timesteps,
+)
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    BaseOutput,
+    deprecate,
+    logging,
+    replace_example_docstring,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from PIL import Image
+
+from diffusers import AutoencoderKL, DiffusionPipeline, UNet2DConditionModel
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+from dataclasses import dataclass
+
+
+def postprocess(
+    image: torch.FloatTensor,
+    output_type: str = "pil",
+):
+    """
+    Postprocess the image output from tensor to `output_type`.
+
+    Args:
+        image (`torch.FloatTensor`):
+            The image input, should be a pytorch tensor with shape `B x C x H x W`.
+        output_type (`str`, *optional*, defaults to `pil`):
+            The output type of the image, can be one of `pil`, `np`, `pt`, `latent`.
+
+    Returns:
+        `PIL.Image.Image`, `np.ndarray` or `torch.FloatTensor`:
+            The postprocessed image.
+    """
+    if not isinstance(image, torch.Tensor):
+        raise ValueError(
+            f"Input for postprocessing is in incorrect format: {type(image)}. We only support pytorch tensor"
+        )
+    if output_type not in ["latent", "pt", "np", "pil"]:
+        deprecation_message = (
+            f"the output_type {output_type} is outdated and has been set to `np`. Please make sure to set it to one of these instead: "
+            "`pil`, `np`, `pt`, `latent`"
+        )
+        deprecate(
+            "Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False
+        )
+        output_type = "np"
+
+    image = image.detach().cpu()
+
+    if output_type == "latent":
+        return image
+
+    # denormalize the image
+    image = image.clamp(-1, 1) * 0.5 + 0.5
+
+    materials = []
+    for i in range(image.shape[0]):
+
+        material = MatForgerMaterial()
+        material.init_from_tensor(image[i])
+
+        if output_type == "pt":
+            material.to_pt()
+
+        if output_type == "np":
+            material.to_np()
+
+        if output_type == "pil":
+            material.to_pil()
+
+        materials.append(material)
+
+    return materials
+
+
+@dataclass
+class MatForgerMaterial:
+    def __init__(
+        self,
+        basecolor: Optional[Union[Image.Image, np.ndarray, torch.FloatTensor]] = None,
+        normal: Optional[Union[Image.Image, np.ndarray, torch.FloatTensor]] = None,
+        height: Optional[Union[Image.Image, np.ndarray, torch.FloatTensor]] = None,
+        roughness: Optional[Union[Image.Image, np.ndarray, torch.FloatTensor]] = None,
+        metallic: Optional[Union[Image.Image, np.ndarray, torch.FloatTensor]] = None,
+    ):
+        self.basecolor = basecolor
+        self.normal = normal
+        self.height = height
+        self.roughness = roughness
+        self.metallic = metallic
+
+    def _to_numpy(self, image):
+        if image is None:
+            return None
+
+        if isinstance(image, Image.Image):
+            image = np.array(image)
+        elif isinstance(image, torch.FloatTensor):
+            image = image.cpu().numpy()
+        return image
+
+    def _to_pil(self, image):
+        if image is None:
+            return None
+
+        if isinstance(image, np.ndarray):
+            image = Image.fromarray(image)
+        elif isinstance(image, torch.FloatTensor):
+            image = TF.to_pil_image(image)
+        return image
+
+    def _to_pt(self, image):
+        if image is None:
+            return None
+
+        if isinstance(image, np.ndarray):
+            image = torch.from_numpy(image)
+        elif isinstance(image, Image.Image):
+            image = TF.to_tensor(image)
+        return image
+
+    def compute_normal_map_z_component(self, normal: torch.FloatTensor):
+        """
+        Compute the z-component of the normal map for a tensor of shape (2, H, W).
+
+        Parameters:
+        - normal_map (torch.Tensor): A tensor of shape (2, H, W) containing the x and y components of the normal map.
+
+        Returns:
+        - A tensor of shape (1, H, W) containing the z-component of the normal map.
+        """
+        # Normalize the normal map to the range [-1, 1]
+        normal = normal * 2 - 1
+
+        # Square the x and y components
+        squared = normal**2
+
+        # Sum along the first dimension (x^2 + y^2)
+        sum_squared = squared.sum(dim=0, keepdim=True)
+
+        # Compute z-component: sqrt(1 - (x^2 + y^2))
+        z_component = torch.sqrt(1 - sum_squared).clamp(
+            min=0
+        )  # Clamp to avoid negative values under sqrt
+
+        normal = torch.cat([normal, z_component], dim=0)
+        normal = normal * 0.5 + 0.5  # Denormalize to [0, 1]
+        return normal
+
+    def init_from_tensor(self, image: torch.FloatTensor):
+        assert image.shape[0] >= 8, "Input tensor should have at least 8 channels"
+        self.basecolor = image[:3]
+        self.normal = self.compute_normal_map_z_component(image[3:5])
+        self.height = image[5:6]
+        self.roughness = image[6:7]
+        self.metallic = image[7:8]
+
+    def to_pt(self):
+        # convert to pytorch tensor
+        self.basecolor = self._to_pt(self.basecolor)
+        self.normal = self._to_pt(self.normal)
+        self.height = self._to_pt(self.height)
+        self.roughness = self._to_pt(self.roughness)
+        self.metallic = self._to_pt(self.metallic)
+
+    def to_np(self):
+        # convert to numpy
+        self.basecolor = self._to_numpy(self.basecolor)
+        self.normal = self._to_numpy(self.normal)
+        self.height = self._to_numpy(self.height)
+        self.roughness = self._to_numpy(self.roughness)
+        self.metallic = self._to_numpy(self.metallic)
+
+    def to_pil(self):
+        # convert to PIL image
+        self.basecolor = self._to_pil(self.basecolor)
+        self.normal = self._to_pil(self.normal)
+        self.height = self._to_pil(self.height)
+        self.roughness = self._to_pil(self.roughness)
+        self.metallic = self._to_pil(self.metallic)
+
+    def as_dict(self):
+        return {
+            "basecolor": self.basecolor,
+            "normal": self.normal,
+            "height": self.height,
+            "roughness": self.roughness,
+            "metallic": self.metallic,
+        }
+
+
+@dataclass
+class MatForgerPipelineOutput(BaseOutput):
+    """
+    Output class for Stable Diffusion pipelines.
+
+    Args:
+        images (`List[PIL.Image.Image]` or `np.ndarray`)
+            List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+            num_channels)`.
+    """
+
+    images: List[MatForgerMaterial]
+
+
+class MatForgerPipeline(DiffusionPipeline, FromSingleFileMixin):
+
+    model_cpu_offload_seq = "prompt_encoder->unet->vae"
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        unet: UNet2DConditionModel,
+        prompt_encoder: nn.Module,
+        scheduler: KarrasDiffusionSchedulers,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            unet=unet,
+            prompt_encoder=prompt_encoder,
+            scheduler=scheduler,
+        )
+
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def encode_prompt(
+        self,
+        prompt,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            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`).
+            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.
+        """
+        if (
+            prompt is not None
+            and isinstance(prompt, str)
+            or isinstance(prompt, Image.Image)
+        ):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self.prompt_encoder.encode_prompt(prompt)
+
+        if self.prompt_encoder is not None:
+            prompt_embeds_dtype = self.prompt_encoder.dtype
+        elif self.unet is not None:
+            prompt_embeds_dtype = self.unet.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_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 and negative_prompt_embeds is None:
+            negative_prompt_embeds = self.prompt_encoder.encode_prompt(
+                [""] * batch_size  # TODO: Make this customizable
+            )
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(
+                dtype=prompt_embeds_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
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def decode_latents(self, latents):
+        deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+        deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+        latents = 1 / self.vae.config.scaling_factor * latents
+        image = self.vae.decode(latents, return_dict=False)[0]
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    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,
+        height,
+        width,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(
+                f"`height` and `width` have to be divisible by 8 but are {height} and {width}."
+            )
+
+        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 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)}"
+            )
+
+        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."
+            )
+
+        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}."
+                )
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            height // self.vae_scale_factor,
+            width // self.vae_scale_factor,
+        )
+        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."
+            )
+
+        if latents is None:
+            latents = randn_tensor(
+                shape, generator=generator, device=device, dtype=dtype
+            )
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stages where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+        that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        if not hasattr(self, "unet"):
+            raise ValueError("The pipeline must have `unet` for using FreeU.")
+        self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism if enabled."""
+        self.unet.disable_freeu()
+
+    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+    def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            timesteps (`torch.Tensor`):
+                generate embedding vectors at these timesteps
+            embedding_dim (`int`, *optional*, defaults to 512):
+                dimension of the embeddings to generate
+            dtype:
+                data type of the generated embeddings
+
+        Returns:
+            `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), 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
+
+    # def patch image
+    def patch_image(
+        self,
+        image: torch.FloatTensor,
+        patch_size: int,
+        overlap: float = 0.5,
+    ) -> torch.FloatTensor:
+        r"""
+        Patch the input image into smaller patches.
+
+        Args:
+            image (`torch.Tensor`):
+                The input image tensor to be patched. The tensor should have shape `(B, C, H, W)`.
+            patch_size (`int`):
+                The size of the patch.
+            overlap (`float`, *optional*, defaults to `0.25`):
+                The overlap between patches.
+
+        Returns:
+            `torch.Tensor`:
+                The patched image tensor.
+        """
+        # Get the number of channels
+        B, C, H, W = image.shape
+
+        # Calculate the stride for unfolding
+        stride = int(patch_size * (1 - overlap))
+
+        # Calculate required padding for height and width
+        pad_height = (H - patch_size) % stride
+        pad_width = (W - patch_size) % stride
+
+        # Adjust padding to fully cover the image dimensions
+        if pad_height > 0:
+            pad_height = stride - pad_height
+        if pad_width > 0:
+            pad_width = stride - pad_width
+
+        # Apply padding symmetrically to the bottom and right sides
+        image = F.pad(image, (0, pad_width, 0, pad_height), mode="circular", value=0)
+        H_padded, W_padded = image.shape[-2:]
+
+        # Unfold the padded image tensor into patches
+        image = image.unfold(2, patch_size, stride).unfold(3, patch_size, stride)
+
+        image = image.permute(0, 2, 3, 1, 4, 5)
+        image = image.reshape(-1, C, patch_size, patch_size)
+        return image, (H_padded, W_padded)
+
+    # def unpatch image with overlap
+    def unpatch_image(
+        self,
+        patches: torch.FloatTensor,
+        batch_size: int,
+        output_size: Tuple[int, int],
+        patch_size: int,
+        crop_size: Optional[Tuple[int, int]] = None,
+        overlap: float = 0.25,
+    ) -> torch.FloatTensor:
+        """
+        Reconstruct the original image from its patches using fold, averaging the overlaps.
+
+        Args:
+            patches (torch.Tensor): The patches of the image with shape `(B, C, H, W)`,
+                                where `B` is the effective batch size (number of patches),
+                                `C` is the channel depth, and `H`, `W` are the patch height and width.
+            batch_size (int): The effective batch size (number of patches).
+            output_size (tuple): The height and width of the original image before patching.
+            patch_size (int): The height and width of each patch (assuming square patches).
+            crop_size (tuple, *optional*): The height and width of the cropped image.
+            overlap (`float`, *optional*, defaults to `0.25`):
+                The overlap between patches.
+
+        Returns:
+            torch.Tensor: The reconstructed images of shape `(B, C, H, W)`.
+        """
+        # Set crop size if not provided
+        if crop_size is None:
+            crop_size = output_size
+
+        # Calculate the stride for folding
+        stride = int(patch_size * (1 - overlap))
+
+        # Calculate the number of patches per image
+        num_patches_per_image = patches.shape[0] // batch_size
+
+        patches = patches.view(
+            batch_size, num_patches_per_image, patches.shape[1], patch_size, patch_size
+        )
+        patches = patches.permute(0, 2, 3, 4, 1).contiguous()
+        patches = patches.view(
+            batch_size, patches.shape[1] * patch_size * patch_size, -1
+        )
+
+        # Use fold to reconstruct the images
+        reconstructed = F.fold(
+            patches, output_size=output_size, kernel_size=patch_size, stride=stride
+        )
+
+        # For averaging the overlaps, create a tensor of ones and fold it
+        mask = torch.ones_like(patches)
+        mask = F.fold(
+            mask, output_size=output_size, kernel_size=patch_size, stride=stride
+        )
+
+        # Average the accumulated values in the overlaps
+        reconstructed /= mask
+
+        # Crop the reconstructed image to the desired size
+        reconstructed = reconstructed[..., : crop_size[0], : crop_size[1]]
+
+        return reconstructed
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def guidance_rescale(self):
+        return self._guidance_rescale
+
+    # 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 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], PipelineImageInput, List[PipelineImageInput]
+        ] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        tileable: bool = True,
+        patched: bool = True,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 7.5,
+        negative_prompt: 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,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        **kwargs,
+    ):
+
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and (
+            isinstance(prompt, str) or isinstance(prompt, Image.Image)
+        ):
+            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
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            self.do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+        )
+
+        # For classifier free guidance, we need to do two forward passes.
+        # Here we concatenate the unconditional and text embeddings into a single batch
+        # to avoid doing two forward passes
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler, num_inference_steps, device, timesteps
+        )
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs.
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 6.2 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)
+
+        # 7. Denoising loop
+        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
+
+                # If patched diffusion
+                if patched:
+                    B = latents.shape[0]
+                    # patch the latents
+                    latents, size_padded = self.patch_image(
+                        latents, patch_size=32, overlap=0.0
+                    )
+                    # TODO: Improve prompt repeat when patching
+                    Bp = latents.shape[0]
+                    if prompt_embeds.shape[0] != Bp * 2:
+                        prompt_embeds = prompt_embeds.repeat_interleave(Bp // B, dim=0)
+
+                # 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
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_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 + self.guidance_scale * (
+                        noise_pred_text - noise_pred_uncond
+                    )
+
+                if self.do_classifier_free_guidance and self.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=self.guidance_rescale,
+                    )
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs, return_dict=False
+                )[0]
+
+                if patched:
+                    # unpatch the latents
+                    latents = self.unpatch_image(
+                        latents, B, size_padded, patch_size=32, overlap=0.0
+                    )
+
+                # noise rolling, baby!
+                # Based on 5.1. in https://arxiv.org/pdf/2309.01700.pdf
+                if tileable:
+                    roll_h = torch.randint(0, height, (1,)).item()
+                    roll_w = torch.randint(0, width, (1,)).item()
+                    latents = torch.roll(latents, shifts=(roll_h, roll_w), dims=(2, 3))
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or (i + 1) % self.scheduler.order == 0:
+                    progress_bar.update()
+
+        if not output_type == "latent":
+            if tileable:
+                # decode padded latent to preserve tileability
+                l_height = height // self.vae_scale_factor
+                l_width = width // self.vae_scale_factor
+                latents = TF.center_crop(
+                    latents.repeat(1, 1, 3, 3), (l_height + 4, l_width + 4)
+                )
+
+            # decode the latents
+            image = self.vae.decode(
+                latents / self.vae.config.scaling_factor,
+                return_dict=False,
+                generator=generator,
+            )[0]
+
+            # crop to original size
+            image = TF.center_crop(image, (height, width))
+        else:
+            image = latents
+
+        image = postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return image
+
+        return MatForgerPipelineOutput(images=image)

prompt_encoder/config.json

@@ -0,0 +1,4 @@
+{
+  "_class_name": "MaterialPromptEncoder",
+  "_diffusers_version": "0.26.3"
+}

prompt_encoder/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46470236d5adee6faf6bf0011dc6f67a0ece61041ed342b57301db02d5c58ff7
+size 1710544320

prompt_encoder/encoder.py

@@ -0,0 +1,80 @@
+from typing import List, Optional
+
+from diffusers.configuration_utils import ConfigMixin
+from diffusers.models.modeling_utils import ModelMixin
+from PIL import Image
+from transformers import (
+    AutoProcessor,
+    AutoTokenizer,
+    CLIPTextModelWithProjection,
+    CLIPVisionModelWithProjection,
+)
+
+
+class BasePromptEncoder(ModelMixin, ConfigMixin):
+    def __init__(self):
+        super().__init__()
+
+    def encode_text(self, text):
+        raise NotImplementedError
+
+    def encode_image(self, image):
+        raise NotImplementedError
+
+    def forward(
+        self,
+        prompt,
+        negative_prompt=None,
+    ):
+        raise NotImplementedError
+
+
+class MaterialPromptEncoder(BasePromptEncoder):
+    def __init__(self):
+        super(MaterialPromptEncoder, self).__init__()
+
+        self.processor = AutoProcessor.from_pretrained("openai/clip-vit-large-patch14")
+        self.tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-large-patch14")
+        self.clip_vision = CLIPVisionModelWithProjection.from_pretrained(
+            "openai/clip-vit-large-patch14"
+        )
+        self.clip_text = CLIPTextModelWithProjection.from_pretrained(
+            "openai/clip-vit-large-patch14"
+        )
+
+    def encode_text(self, text):
+        inputs = self.tokenizer(text, padding=True, return_tensors="pt")
+        inputs["input_ids"] = inputs["input_ids"].to(self.device)
+        inputs["attention_mask"] = inputs["attention_mask"].to(self.device)
+        outputs = self.clip_text(**inputs)
+        return outputs.text_embeds.unsqueeze(1)
+
+    def encode_image(self, image):
+        inputs = self.processor(images=image, return_tensors="pt")
+        inputs["pixel_values"] = inputs["pixel_values"].to(self.device)
+        outputs = self.clip_vision(**inputs)
+        return outputs.image_embeds.unsqueeze(1)
+
+    def encode_prompt(
+        self,
+        prompt,
+    ):
+        dtype = type(prompt)
+        if dtype == list:
+            dtype = type(prompt[0])
+
+        if dtype == str:
+            return self.encode_text(prompt)
+        elif dtype == Image.Image:
+            return self.encode_image(prompt)
+        else:
+            raise NotImplementedError
+
+    def forward(
+        self,
+        prompt,
+        negative_prompt=None,
+    ):
+        prompt = self.encode_prompt(prompt)
+        negative_prompt = self.encode_prompt(negative_prompt)
+        return prompt, negative_prompt

scheduler/scheduler_config.json

@@ -0,0 +1,21 @@
+{
+  "_class_name": "DDIMScheduler",
+  "_diffusers_version": "0.26.3",
+  "beta_end": 0.012,
+  "beta_schedule": "scaled_linear",
+  "beta_start": 0.00085,
+  "clip_sample": false,
+  "clip_sample_range": 1.0,
+  "dynamic_thresholding_ratio": 0.995,
+  "interpolation_type": "linear",
+  "num_train_timesteps": 1000,
+  "prediction_type": "epsilon",
+  "rescale_betas_zero_snr": false,
+  "sample_max_value": 1.0,
+  "set_alpha_to_one": false,
+  "skip_prk_steps": true,
+  "steps_offset": 1,
+  "thresholding": false,
+  "timestep_spacing": "leading",
+  "trained_betas": null
+}

unet/config.json

@@ -0,0 +1,67 @@
+{
+  "_class_name": "UNet2DConditionModel",
+  "_diffusers_version": "0.26.3",
+  "act_fn": "silu",
+  "addition_embed_type": null,
+  "addition_embed_type_num_heads": 64,
+  "addition_time_embed_dim": null,
+  "attention_head_dim": 8,
+  "attention_type": "default",
+  "block_out_channels": [
+    256,
+    512,
+    1024,
+    1024
+  ],
+  "center_input_sample": false,
+  "class_embed_type": null,
+  "class_embeddings_concat": false,
+  "conv_in_kernel": 3,
+  "conv_out_kernel": 3,
+  "cross_attention_dim": 768,
+  "cross_attention_norm": null,
+  "down_block_types": [
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D",
+    "DownBlock2D"
+  ],
+  "downsample_padding": 1,
+  "dropout": 0.0,
+  "dual_cross_attention": false,
+  "encoder_hid_dim": null,
+  "encoder_hid_dim_type": null,
+  "flip_sin_to_cos": true,
+  "freq_shift": 0,
+  "in_channels": 18,
+  "layers_per_block": 2,
+  "mid_block_only_cross_attention": null,
+  "mid_block_scale_factor": 1,
+  "mid_block_type": "UNetMidBlock2DCrossAttn",
+  "norm_eps": 1e-05,
+  "norm_num_groups": 32,
+  "num_attention_heads": null,
+  "num_class_embeds": null,
+  "only_cross_attention": false,
+  "out_channels": 18,
+  "projection_class_embeddings_input_dim": null,
+  "resnet_out_scale_factor": 1.0,
+  "resnet_skip_time_act": false,
+  "resnet_time_scale_shift": "default",
+  "reverse_transformer_layers_per_block": null,
+  "sample_size": 64,
+  "time_cond_proj_dim": null,
+  "time_embedding_act_fn": null,
+  "time_embedding_dim": null,
+  "time_embedding_type": "positional",
+  "timestep_post_act": null,
+  "transformer_layers_per_block": 1,
+  "up_block_types": [
+    "UpBlock2D",
+    "CrossAttnUpBlock2D",
+    "CrossAttnUpBlock2D",
+    "CrossAttnUpBlock2D"
+  ],
+  "upcast_attention": false,
+  "use_linear_projection": false
+}

unet/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67f2d6cf5583e3fc2559fbc587a9f01d09db8765a5b1ef38dc5d34e436b82143
+size 2213282432

vae/config.json

@@ -0,0 +1,31 @@
+{
+  "_class_name": "AutoencoderKL",
+  "_diffusers_version": "0.26.3",
+  "act_fn": "silu",
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "force_upcast": true,
+  "in_channels": 9,
+  "latent_channels": 18,
+  "layers_per_block": 2,
+  "norm_num_groups": 32,
+  "out_channels": 9,
+  "sample_size": 512,
+  "scaling_factor": 0.18215,
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ]
+}

vae/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79ef4da42cd429353b034bf07cfc7cff98c9214421e9c2291c0a945a17afc290
+size 335479204