RepoPicks Radar

Updated every day around 23:30 UTC

Guides

Tune-A-Video

showlab/Tune-A-VideoReviewed by AI

Checked by RepoPicks Radar on

What is Tune-A-Video?

Tune-A-Video is the code behind a 2023 research paper (ICCV). You give it one short video and a sentence describing it, and it fine-tunes a Stable Diffusion image model for about 10 to 15 minutes on a strong NVIDIA card. After that you can ask for the same motion with a different subject or style, such as turning "a man is skiing" into "Spider Man is skiing on the beach". Everything happens in a terminal and in Python; there is no interface, and the project has not been touched since October 2023.

Tune-A-Video is free and open source (Apache-2.0) and can replace Runway and Google Veo 3.

How easy is it to install Tune-A-Video?

D · terminal, several tools or a specific graphics card · Linux

Linux is the only setup described: install the Python packages, download Stable Diffusion weights from Hugging Face yourself, edit a YAML config and run `accelerate launch train_tuneavideo.py` on a card like an A100. Getting a video out afterwards means writing your own Python script from the snippet in the README; a Colab notebook is linked for trying it without installing anything.

Coming soon: install in 1 click with the RepoPicks app.

Tune-A-Video on GitHub

Stars
4,359on Nov 10, 2026
Downloads
—all releases
Growth
—no gain in the last reading

Stars over time

4,3594,3604,361 Nov 5Nov 6Nov 7Nov 8Nov 9Nov 10 4,3594,3604,361 Nov 5Nov 8

Facts

Runs on
Linux
You need to install
Python, NVIDIA graphics card
To build from source
Python, Graphics card
Disk space
950 MB–4.3 GB · estimate
Accounts or keys
None
Cost
Free · Apache-2.0 code. The Stable Diffusion models it builds on have their own licences, and the training needs a rented or owned GPU.
License
Apache-2.0 (Permissive)
Can replace
Runway, Veo 3
Last change on GitHub
Oct 25, 2023

Tune-A-Video risk signals (6)

  • AttentionDependencies with known vulnerabilities

    Pinned dependencies with known vulnerabilities (OSV): torch: GHSA-3749-ghw9-m3mg, GHSA-47fc-vmwq-366v, GHSA-53q9-r3pm-6pq6; diffusers: GHSA-7wx4-6vff-v64p, GHSA-98h9-4798-4q5v, PYSEC-2026-2446

    automatic check

  • Attentionpinned versions from 2022

    requirements.txt pins torch==1.12.1, torchvision==0.13.1, diffusers[torch]==0.11.1 and bitsandbytes==0.35.4, all from 2022. Those versions are hard to install on a current Python and current graphics drivers, and they are the same pins the checks flagged for known vulnerabilities.

    AI review

  • Attentionnot maintained

    The last commit is from 2023-10-25, 1112 days before this check, there are no releases, 37 issues are open and the newest item in the README's News list is from February 2023.

    AI review

  • Infono interface and no finished program

    The repository has train_tuneavideo.py, configs and a Python module; the "Inference" section is a code snippet you are expected to paste into your own script. The only ready-made interface is a Gradio demo on Hugging Face built by someone else.

    AI review

  • Infoexample prompts use protected characters

    The README's result tables are built on prompts such as "Spider Man is skiing on the beach", "Wonder Woman, wearing a cowboy hat", "James Bond is dribbling a basketball", "Mr Potato Head" and "modern disney style", and it points to fine-tuned models like Modern Disney and Anything V4.0. Videos like these use characters and styles someone else owns.

    AI review

  • Infoedits real footage of people

    The method takes one real video and regenerates it with a different subject, as in the "A man is skiing" examples; the same steps work on footage of an identifiable person who did not agree to it.

    AI review

Evidence is shown in its original language.

What Radar could not check about Tune-A-Video

  • whether the 2022 package versions still install and run on a current machine and driver
  • how much graphics memory is really needed, since the only figure given is 10 to 15 minutes on one A100
  • how much disk the Stable Diffusion weights and the per-video fine-tuned copies take once you keep several of them
  • whether the linked Colab notebook and Hugging Face demo still work three years after their last update
  • what the Stable Diffusion and DreamBooth model licences allow you to do with the videos you make

Similar projects

The RepoPicks app

Coming soon: install in 1 click with the RepoPicks app.

Before you download

See the release page on GitHub