TTT Video DiT
test-time-training/ttt-video-ditReviewed by AI
Checked by RepoPicks Radar on
What is TTT Video DiT?
This is the official code from the paper "One-Minute Video Generation with Test-Time Training". It adds new layers to an existing text-to-video model (CogVideoX 5B) and trains it in stages so the videos get longer, up to 63 seconds, keeping the same style from start to end. It is a research project for people with machine-learning experience: you build a CUDA kernel from source, fetch the base model yourself and train on H100 cards. There is no app, no interface and no finished model to download.
TTT Video DiT is free and open source (MIT) and can replace Google Veo 3.
How easy is it to install TTT Video DiT?
E · developers only · Linux
Developers only, on Linux: you create a conda environment, pull a git submodule and compile the TTT-MLP kernel, which needs the CUDA toolkit 12.3+ and gcc 11+, and the authors say training is only supported on H100 cards. The base model weights have to be collected by hand from two other projects.
Coming soon: install in 1 click with the RepoPicks app.
TTT Video DiT on GitHub
- Stars
- 2,453on Nov 10, 2026
- Downloads
- —all releases
- Growth
- —no gain in the last reading
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
- 1.0 GB–4.6 GB · estimate
- Accounts or keys
- None
- Cost
- Free
- License
- MIT (Permissive)
- Can replace
- Veo 3
- Last change on GitHub
- Feb 25, 2026
TTT Video DiT risk signals (3)
- Infofinal model weights not released
README note under the download instructions: "These are the pretrained model weights, not the final model weights used to generate our demonstration videos." Only the CogVideoX-5b base weights are linked.
AI review
- Infoneeds datacenter gpu and a compiled kernel
README: "You must have cuda toolkit (12.3+) and gcc11+ installed to build the TTT-MLP kernel. We only support training on H100s for TTT-MLP."
AI review
- Infobuilds on a model with its own terms
The code is MIT, but it runs on THUDM's CogVideoX-5b weights fetched from Hugging Face, which carry their own licence that this repository does not restate.
AI review
Evidence is shown in its original language.
What Radar could not check about TTT Video DiT
- Whether anyone can reproduce the one-minute demo videos, since the finished weights are not published
- How much video memory and disk the training and sampling steps really need
- What renting H100 time would cost to train the staged schedule described in the README
- Whether the kernel still builds with current CUDA and PyTorch versions, as the last commit is from February 2026