StableNormal addresses the challenge of generating stable and sharp normal maps from single images using diffusion models. It is designed for researchers and developers in computer vision and graphics who require high-quality surface normal estimation, offering improved stability and accuracy over existing diffusion-based methods.

How It Works

StableNormal tailors diffusion priors for monocular normal estimation by reducing the inherent stochasticity of diffusion models like Stable Diffusion. This approach aims to enhance estimation stability, leading to "Stable-and-Sharp" normal maps. The method leverages a modified diffusion process to achieve more consistent and precise results compared to standard diffusion-based techniques.

Quick Start & Requirements

• Install via pip: pip install git+https://github.com/Stable-X/StableNormal.git
• Requires PyTorch.
• Pretrained weights are available via torch.hub or can be downloaded locally.
• Official website: https://stable-x.github.io/StableNormal
• Paper: https://arxiv.org/abs/2406.16864
• Hugging Face Model: https://huggingface.co/Stable-X/stable-normal-v0-1

Highlighted Details

• Accepted to SIGGRAPH Asia 2024 (Journal Track).
• Offers a "StableNormal-turbo" variant for 10x faster inference.
• Benchmarks show superior performance on datasets like DIODE, IBims-1, Scannet, and NYUv2 compared to GeoWizard, Marigold Normal, GenPercept, and DSINE.
• Includes scripts for computing evaluation metrics on the DIODE dataset.

Maintenance & Community

• Project is associated with SIGGRAPH Asia 2024.
• Releases include StableNormal and StableNormal-turbo.
• ModelScope and Hugging Face Spaces are available for exploration.

Licensing & Compatibility

• Licensed under Apache-2.0.
• Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The provided benchmarks focus on specific datasets and metrics; performance on diverse, real-world unseen data may vary. The "turbo" version might involve trade-offs in accuracy for speed.