Skip to content
Get packs

Generative Models

Research code and model weights for Stability AI's video and 3D-view generation models.

GitHub Download 0.1.0
Stars
27.3k
Forks
3.1k
Licence
MIT
Language
Python
Last update
11 months ago
Type
AI model / agent

What it is

This repository holds the code for Stability AI's generative models, mainly video models that create new views of an object or turn a single image into an orbiting video. The README covers SV3D, SV4D and SV4D 2.0, which take a video or image and generate novel-view video from it. The models are released for research purposes, and the README shows command-line scripts for running them.

  • Generate a novel-view video from a short input video of a moving object
  • Create an orbital video from a single white-background image using SV3D
  • Run the SV4D 2.0 sampling script on a gif or mp4 file to produce new camera-view frames

Great for

  • Researchers studying video-to-4D or novel-view video synthesis
  • Developers who want to run SV3D or SV4D locally from the command line
  • People with a single image of an object who want an orbiting video of it, provided they have a suitable GPU

Not ideal for

  • People who want a ready-made consumer app: the README describes command-line scripts and a community gradio demo, not a polished product
  • Commercial users, since the models are described as released for research purposes and the README does not give commercial terms
  • Users without a capable GPU, since the README mentions low-VRAM workarounds, implying significant VRAM needs
Step-by-step guide: how to use Generative ModelsInstall, first run, examples, tips and FAQs

How to get it

  1. 1

    Clone and build from source

    You'll need: Python 3.10 · PyTorch with CUDA (version depends on your GPU) · Hugging Face model weights (sv4d2.safetensors and others)

    Create a virtual environment

    python3.10 -m venv .generativemodels

    Activate the environment

    source .generativemodels/bin/activate

    Install dependencies from requirements file

    pip3 install -r requirements/pt2.txt

    Install the package

    pip3 install .
  2. 2

    Run it

    • Clone the repository and set up the Python 3.10 virtual environment
    • Install PyTorch matching your CUDA version and the requirements file
    • Download sv4d2.safetensors from Hugging Face into checkpoints/
    • Run scripts/sampling/simple_video_sample_4d2.py with an input video path and an output folder
  3. 3

    Get the most out of it

    • Set --remove_bg=True for input videos with a plain background, using rembg
    • Use --encoding_t=1 and --decoding_t=1, or a smaller --img_size=512, if you have low VRAM
    • Lower num_steps below the default 50 to shorten sampling time at some quality cost
    • Use the 8-view model sv4d2_8views.safetensors if you want the alternative 5x8 setup

Downloads

Source code (.zip)main branch · straight from GitHubRelease 0.1.0Release notes and source

Strengths

  • Covers several related models (SV3D, SV4D, SV4D 2.0) in one repository
  • Includes concrete command-line quickstart examples in the README
  • Code is released under the MIT license

Trade-offs

  • The README is truncated and some model details and requirements are unclear from it
  • Models are described as for research purposes, which may limit real-world use
  • Setup needs Python, PyTorch with a matching CUDA version, and model weights downloaded separately from Hugging Face

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.