Stable Video Infinity
vita-epfl/Stable-Video-InfinityReviewed by AI
Checked by RepoPicks Radar on
What is Stable Video Infinity?
Stable Video Infinity, from the VITA lab at EPFL, is research code that extends an existing video model (Wan 2.1) so it can keep generating minutes-long clips from one picture plus a stream of text prompts, instead of a few seconds. Variants do cartoons, dancing from a skeleton and talking heads driven by audio. It is run from the terminal on Linux with a strong NVIDIA card (the team tested an A100 80 GB) after downloading tens of gigabytes of model files, or through community ComfyUI workflows. Training code and datasets are included.
Stable Video Infinity is free and open source (MIT) and can replace Runway, Kling, and Pika.
How easy is it to install Stable Video Infinity?
D · terminal, several tools or a specific graphics card · Linux
Linux only in practice, and a D: you create a conda environment, run pip install -e ., install flash_attn (the README links an issue for when that fails), download the Wan 2.1 14B base model plus the SVI adapters with huggingface-cli, and launch bash scripts or the Gradio demo. The environment the authors tested is an A100 with 80 GB of video memory.
Coming soon: install in 1 click with the RepoPicks app.
Stable Video Infinity on GitHub
- Stars
- 2,620on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +1stars in 6 days
Stars over time
Facts
- Runs on
- Linux
- You need to install
- Python, NVIDIA graphics card, FFmpeg
- To build from source
- Python, Graphics card, FFmpeg
- Disk space
- 1.1 GB–4.8 GB · estimate
- Accounts or keys
- None
- Cost
- Free · MIT code and freely downloadable model files; the base Wan 2.1 weights and the datasets come with their own terms. The README also says SVI-2.0 Pro is available on the Poe platform, through its chat interface or its API, hosted with support from @empiriolabsai.
- License
- MIT (Permissive)
- Last change on GitHub
- Sep 17, 2026
Stable Video Infinity risk signals (7)
- AttentionDependencies with known vulnerabilities
Pinned dependencies with known vulnerabilities (OSV): torch: GHSA-3749-ghw9-m3mg, GHSA-53q9-r3pm-6pq6, GHSA-63cw-57p8-fm3p; transformers: GHSA-29pf-2h5f-8g72, GHSA-37mw-44qp-f5jm, GHSA-4w7r-h757-3r74
automatic check
- Attentiontalking head from photo and audio
The SVI-Talk model takes "Image + Audio" and produces a "Talking video" (topic: audio-driven-talking-face), which makes videos of a person saying something they never said; the README has no notice about consent or misuse.
AI review
- Attentioncartoon variant of known characters
One of the variants is listed as "SVI-Tom&Jerry — Cartoon animation", and the README links an "8-minute crazy Tom & Jerry video made with SVI-Tom". The file svi-tom.safetensors is not in the repository: the README downloads it from Hugging Face (vita-video-gen/svi-model) with huggingface-cli. Cartoon characters are usually protected by copyright, and the rights holder's terms are outside the project's MIT license.
AI review
- Infoheavy hardware and downloads
README.md: "We have tested the environment with A100 80G, cuda 12.0" and the weights/ tree lists the Wan 2.1 I2V 14B base model in seven safetensors shards plus extra encoders for the talk and dance variants.
AI review
- Infothird party workflows and hosting
Much of the practical use goes through other people's work: community ComfyUI workflows (with warnings such as "Avoid using the wrong SVI 1.0 workflow in this repo" and quantized test versions), and the README says "SVI-2.0 Pro is now available on the Poe platform! You can access it through the Poe chat interface or integrate it via their API", thanking @empiriolabsai for the support.
AI review
- Infotraining data from third parties
The listed training datasets include MixKit, UltraVideo, a 5k subset of Hallo 3 and a TikTok dancing dataset from Kaggle, each with its own terms.
AI review
- Infoauthor claims not measured
Claims such as "This will never drift or forget in our 20 min test" and "infinite length" come from the authors' own tests.
AI review
Evidence is shown in its original language.
What Radar could not check about Stable Video Infinity
- How much video memory you really need on a consumer graphics card, and whether it runs at all below the tested A100 80 GB.
- Real disk space: the base Wan 2.1 14B model and the extra encoders are far larger than the estimate in the checks.
- How long one minute of video takes to generate.
- Whether it works on Windows or macOS; the instructions are Linux shell commands with flash_attn and CUDA 12.
- Whether the quality in the community videos is reproducible with your own images and prompts.
- What material each variant was fine-tuned on: the README lists the training datasets for the family, but not for svi-tom.