Stable Fast 3D Documentation Summary

Introduction

Stable Fast 3D (SF3D) is a large-scale image-to-3D reconstruction model developed by Stability AI. It utilizes the TripoSR framework to convert a single image into a textured UV-unwrapped 3D mesh asset. This model is particularly efficient, generating a 3D model in under one second, and includes features for enhanced material prediction and asset application in downstream tasks.

Architecture

SF3D is built on a transformer-based architecture designed for image-to-3D conversion. It accepts a 512x512 pixel input and outputs a low-polygon, textured 3D model. The model incorporates a delighting process and predicts material parameters, such as roughness and metallic properties, to improve the realism of rendered objects.

Training

The model's training utilized the Objaverse dataset, filtered and curated for licensing compliance and suitability. Enhanced rendering techniques were employed to better simulate real-world image distributions, allowing the model to generalize effectively.

Guide: Running Locally

1. Clone the Repository:
   Visit the GitHub repository and clone it to your local environment.

2. Install Dependencies:
   Ensure all necessary libraries and dependencies are installed as per the repository's instructions.

3. Prepare Input Images:
   Input images should be formatted to 512x512 pixels.

4. Run the Model:
   Execute the model locally using the provided scripts in the repository.

5. Cloud GPUs:
   For an enhanced performance experience, consider using cloud GPU services such as AWS, Google Cloud, or Azure.

License

SF3D is available under the Stability AI Community License, which permits free use for research, non-commercial, and commercial purposes for entities with annual revenues under $1,000,000. Organizations exceeding this revenue must obtain an Enterprise License from Stability AI. For full license details, refer to Stability AI's license page.