Stable Video Infinity
Research code and weights for generating very long videos without drift
What it is
A research project from EPFL's VITA lab that extends video diffusion transformers built on Wan 2.1 and 2.2 so they can produce minutes-long videos without the color drift and quality loss that usually build up across clips. It trains the model to correct its own accumulated errors, and ships LoRA-style weights, a benchmark dataset, and variants for talking faces, dance, and cartoon footage. Researchers and ComfyUI users run it to chain prompts into long single-shot or multi-shot sequences.