Cosmos3-Super-Image2Video: Video Generator — What It Does and What You Need to Run It

Cosmos3-Super-Image2Video is a video generator from NVIDIA that turns a still image into a short video clip. The tool belongs to the Cosmos family and handles the image-to-video task: an image goes in, an animated scene comes out. The code is open source, and the model weights are published separately.
What it does
The model performs one main task — generating video from an image. The authors describe Cosmos as an open platform of world models, datasets and tools for building physical AI. The description mentions robots, autonomous vehicles and smart infrastructure, but the developers do not detail the specific use cases of the image-to-video model itself.
The latest release is called Cosmos3. The repository is written mostly in Jupyter Notebook, which indicates that the model is used through interactive notebooks.
What you need to run it
It requires a CUDA platform — this is stated as described by the author and is not the result of independent testing. The developers do not name the exact CUDA version, but the documentation has a separate section with a question about which version to use.
The largest weights file is 4.6 GB, and the full set of all files is 121.6 GB. This is worth keeping in mind when planning disk space. The code license is non-standard and must be checked manually before use. The weights license is labeled "other," which also requires a separate review of the terms.
The code is available in the repository github.com/nvidia/cosmos, the weights at huggingface.co/nvidia/Cosmos3-Super-Image2Video. The official project website: nvidia.com/en-us/ai/cosmos.
Who it suits
The tool is aimed at developers and researchers who work with generative video models and have access to CUDA-capable hardware. The open source code makes it possible to study the implementation and adapt it to your own tasks.
For those looking for a ready-made application with a graphical interface, this option is unlikely to fit: work is done through notebooks and code. It is also worth estimating the volume of weights to download in advance and checking the license terms — both require manual review.
Cosmos3-Super-Image2Video is an open source research tool aimed at specialists who need image-to-video generation and are prepared to deal with non-standard terms of use.
Fact sheet
Repository · Model on HuggingFace · Developer’s site
| Task | image-to-video source |
|---|---|
| Code license | non-standard (verify manually) source |
| Weights license | other source |
| Platform | CUDA per the author’s description source |
| Largest weights file | 4.6 GB source |
| All weights files | 121.6 GB source |
| Library | cosmos source |
| Language | Jupyter Notebook source |
| Last code change | 2026-09-10 source |
| Repository created | 2024-12-30 source |
| Latest release | Cosmos3 source |
|---|---|
| Release date | 2026-06-01 source |
| GitHub stars | 11788 source |
| Forks | 866 source |
| Downloads per month | 59799 source |
| Model updated | 2026-08-26 source |
Values are collected automatically from official sources and were checked on 2026-09-10. Each one links to its source, and values taken from the developer’s pages also carry a verbatim quote — hover over the note. Pricing and versions are shown as of the check date and change most often; verify on the vendor’s site before buying.



