sv3d

Introduction

Stable Video 3D (SV3D) is a generative model developed by Stability AI, designed to create orbital videos from a single still image of an object. It uses Stable Video Diffusion to generate videos by conditioning on a single frame.

Architecture

SV3D generates videos by producing 21 frames at a resolution of 576x576 pixels, using a provided image as a context frame. It offers two variants:

1. SV3D_u: Produces orbital videos without camera conditioning.
2. SV3D_p: Supports both single images and orbital views, allowing for 3D video creation along specified camera paths.

Training

The model was trained using a subset of the Objaverse dataset, selected to mimic real-world image distributions. This training approach enhances the model's generalization capabilities. The dataset is available under the CC-BY license.

Guide: Running Locally

1. Clone the Repository:

   git clone https://github.com/Stability-AI/generative-models
   cd generative-models
   

2. Install Dependencies: Ensure all required libraries and dependencies are installed.
3. Run the Model: Follow the usage instructions provided in the repository to execute the model.
4. Hardware Suggestion: For optimal performance, consider using cloud GPUs such as those provided by AWS, Google Cloud, or Azure.

License

SV3D is released under the Stability AI Community License. For commercial applications, refer to Stability AI's commercial license. The dataset used is licensed under CC-BY.