StepVideo-T2V

StepVideo-T2V is StepFun's 30-billion-parameter text-to-video model, released with open weights on February 17, 2025. It uses a diffusion transformer with 3D full attention trained via flow matching, plus a deep-compression video VAE and a final video-based direct preference optimization stage to reduce artifacts.

It generates videos up to 204 frames from bilingual English and Chinese prompts and was presented as state-of-the-art among open video models at release on the Step-Video-T2V-Eval benchmark. A distilled Turbo variant and an image-to-video follow-up were released shortly after. Check the repository's license file for usage terms before commercial deployment.