StepVideo
30B open text-to-video model — 204-frame clips, bilingual prompts, SOTA claims on its own benchmark.
About
Step-Video-T2V is StepFun's 30-billion-parameter open text-to-video model, released February 2025 with weights, inference code, and a technical report. It generates up to 204 frames per clip using a high-compression Video-VAE (16x16 spatial, 8x temporal), bilingual text encoders (English + Chinese), and a 3D full-attention DiT trained with flow matching plus video-based DPO for artifact reduction. A Turbo distilled variant and a TI2V (text-driven image-to-video, 30B, 102 frames) sibling followed, with ComfyUI nodes available.
The honest picture: the full BF16 model wants ~80GB VRAM (auto-offload drops it to 24GB at 2x inference time; FP8 needs ~40GB) — this is datacenter or high-end workstation territory, not casual local use. Its 'SOTA' claims rest largely on StepFun's own Step-Video-T2V-Eval benchmark. But as an open, research-friendly video foundation with unusually long clip lengths and strong Chinese-language support, it fills a real gap.
What it can do
- Text-to-video up to 204 frames
- 30B DiT with flow matching + video DPO
- Bilingual English/Chinese prompt encoding
- Turbo distilled variant for faster inference
- TI2V sibling for text-driven image-to-video
Setup
Download weights from Hugging Face (stepfun-ai/stepvideo-t2v) or ModelScope, clone github.com/stepfun-ai/Step-Video-T2V, and run the provided inference scripts; ComfyUI-StepVideo nodes exist for visual workflows.
Prerequisites
80GB VRAM for BF16 (or 24GB with offload / 40GB FP8); CUDA; Python environment.
Use cases
Works with
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 54; Trending score 72/100 from researcher assessment.
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.