Step-Video-T2V

Among the largest open video models released to date, Step-Video-T2V from StepFun packs 30 billion parameters into a diffusion transformer with 48 layers, 48 attention heads, and 3D full attention. A deep-compression Video-VAE achieves 16x16 spatial and 8x temporal compression, letting the model generate up to 204 frames at resolutions such as 544x992 or 768x768, and bilingual text encoders take prompts in English and Chinese. Training used flow matching with a Video-DPO preference optimization stage to improve output quality. The hardware bill is serious: peak memory of 78.55 GB means 80 GB GPUs are recommended, with 50-step generation taking about 860 seconds using flash attention, and multi-GPU parallel deployment is supported through a bundled script. A distilled Step-Video-T2V-Turbo variant cuts sampling to 10 to 15 steps. Code and weights are MIT licensed, the technical report is arXiv 2502.10248, and the repository has about 3,200 stars. It functions mostly as a research artifact and capability benchmark rather than a consumer tool, given the datacenter-class requirements.