ttt-video-dit

Why we included this project

Text-to-video models typically generate clips of only a few seconds, and this repo is the official code behind a research effort that stretches that to around a minute of coherent footage. It builds on the CogVideoX 5B diffusion transformer, adding Test-Time Training layers to track long-range dependencies across the whole video while keeping the original attention for fine detail inside each three-second segment. The repository walks you through a staged training schedule from 3 to 63 seconds, including the custom kernels you need to run it, so researchers and engineers can reproduce the results or build their own long-context experiments. Just be aware it expects H100 GPUs for training, so it's a research reference for teams with serious compute, not something you'd drop into a production pipeline. The README is thorough, with clear training and inference documentation, which makes it a good entry point for anyone wanting to work on long-form video generation.