Stablematerials

🚀 StableMaterials

StableMaterials is a diffusion-based model crafted for generating photorealistic physical-based rendering (PBR) materials. It combines semi-supervised learning with Latent Diffusion Models (LDMs) to create high-resolution, tileable material maps from text or image prompts. This model can infer both diffuse (Basecolor) and specular (Roughness, Metallic) properties, along with the material mesostructure (Height, Normal). 🌟

For more details, visit the project page or read the full paper on arXiv.

⚠️ Important Note

This repo contains the weight and the pipeline code for the base model in both the LDM and LCM verisons. The refiner model, along with its pipeline and the inpainting pipeline, will be released shortly.

✨ Features

🧩 Base Model

The base model generates low-resolution (512x512) material maps using a compression VAE (Variational Autoencoder) followed by a latent diffusion process. The architecture is based on the MatFuse adaptation of the LDM paradigm, optimized for material map generation with a focus on diversity and high visual fidelity. 🖼️

🔑 Key Features

• Semi-Supervised Learning: The model is trained using both annotated and unannotated data, leveraging adversarial training to distill knowledge from large-scale pretrained image generation models. 📚
• Knowledge Distillation: Incorporates unannotated texture samples generated using the SDXL model into the training process, bridging the gap between different data distributions. 🌐
• Latent Consistency: Employs a latent consistency model to facilitate fast generation, reducing the inference steps required to produce high-quality outputs. ⚡
• Feature Rolling: Introduces a novel tileability technique by rolling feature maps for each convolutional and attention layer in the U-Net architecture. 🎢

🚀 Quick Start

StableMaterials is designed for generating high-quality, realistic PBR materials for applications in computer graphics, such as video game development, architectural visualization, and digital content creation. The model supports both text and image-based prompting, allowing for versatile and intuitive material generation. 🕹️🏛️📸

💻 Usage Examples

Basic Usage

from diffusers import DiffusionPipeline
from diffusers.utils import load_image

# Load pipeline enabling the execution of custom code
pipe = DiffusionPipeline.from_pretrained(
    "gvecchio/StableMaterials", 
    trust_remote_code=True, 
    torch_dtype=torch.float16
)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
pipe = pipe.to(device)

# Text prompt example
material = pipe(
  prompt="Old rusty metal bars with peeling paint",
  guidance_scale=10.0,
  tileable=True,
  num_images_per_prompt=1,
  num_inference_steps=50,
).images[0]

# Image prompt example
material = pipe(
  prompt=load_image("path/to/input_image.jpg"),
  guidance_scale=10.0,
  tileable=True,
  num_images_per_prompt=1,
  num_inference_steps=50,
).images[0]

# The output will include basecolor, normal, height, roughness, and metallic maps
basecolor = material.basecolor
normal = material.normal
height = material.height
roughness = material.roughness
metallic = material.metallic

Advanced Usage

from diffusers import DiffusionPipeline, LCMScheduler, UNet2DConditionModel
from diffusers.utils import load_image

# Load LCM distilled unet
unet = UNet2DConditionModel.from_pretrained(
    "gvecchio/StableMaterials",
    subfolder="unet_lcm",
    torch_dtype=torch.float16,
)

# Load pipeline enabling the execution of custom code
pipe = DiffusionPipeline.from_pretrained(
    "gvecchio/StableMaterials", 
    trust_remote_code=True, 
    unet=unet,
    torch_dtype=torch.float16
)

# Replace scheduler with LCM scheduler
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
pipe = pipe.to(device)

# Text prompt example
material = pipe(
  prompt="Old rusty metal bars with peeling paint",
  guidance_scale=10.0,
  tileable=True,
  num_images_per_prompt=1,
  num_inference_steps=4, # LCM enables fast generation in as few as 4 steps
).images[0]

# Image prompt example
material = pipe(
  prompt=load_image("path/to/input_image.jpg"),
  guidance_scale=10.0,
  tileable=True,
  num_images_per_prompt=1,
  num_inference_steps=4,
).images[0]

# The output will include basecolor, normal, height, roughness, and metallic maps
basecolor = material.basecolor
normal = material.normal
height = material.height
roughness = material.roughness
metallic = material.metallic

📚 Documentation

🗂️ Training Data

The model is trained on a combined dataset from MatSynth and Deschaintre et al., including 6,198 unique PBR materials. It also incorporates 4,000 texture-text pairs generated from the SDXL model using various prompts. 🔍

🔧 Limitations

While StableMaterials shows robust performance, it has some limitations:

• It may struggle with complex prompts describing intricate spatial relationships. 🧩
• It may not accurately represent highly detailed patterns or figures. 🎨
• It occasionally generates incorrect reflectance properties for certain material types. ✨

Future updates aim to address these limitations by incorporating more diverse training prompts and improving the model's handling of complex textures.

📖 Citation

If you use this model in your research, please cite the following paper:

@article{vecchio2024stablematerials,
  title={StableMaterials: Enhancing Diversity in Material Generation via Semi-Supervised Learning},
  author={Vecchio, Giuseppe},
  journal={arXiv preprint arXiv:2406.09293},
  year={2024}
}

Model Architecture

🧩 Base Model

The base model generates low-resolution (512x512) material maps using a compression VAE (Variational Autoencoder) followed by a latent diffusion process. The architecture is based on the MatFuse adaptation of the LDM paradigm, optimized for material map generation with a focus on diversity and high visual fidelity. 🖼️

🔑 Key Features

• Semi-Supervised Learning: The model is trained using both annotated and unannotated data, leveraging adversarial training to distill knowledge from large-scale pretrained image generation models. 📚
• Knowledge Distillation: Incorporates unannotated texture samples generated using the SDXL model into the training process, bridging the gap between different data distributions. 🌐
• Latent Consistency: Employs a latent consistency model to facilitate fast generation, reducing the inference steps required to produce high-quality outputs. ⚡
• Feature Rolling: Introduces a novel tileability technique by rolling feature maps for each convolutional and attention layer in the U-Net architecture. 🎢

Information Table