tools
Step-Video-T2V
Step-Video-T2V is an open-source text-to-video model that generates videos from English or Chinese prompts.
Step-Video-T2V turns text prompts into video clips using a 30-billion-parameter model. It supports bilingual prompts, video generation up to 204 frames, and local inference through supplied scripts.
The model is available through Yuewen Video online and as open-source code and weights. Local use requires Linux, CUDA, NVIDIA GPUs, and substantial GPU memory. No official Step-Video-T2V-specific pricing page was found.
Features
- Generates videos from text prompts
- Supports English and Chinese text encoders
- Generates videos up to 204 frames
- Uses a compressed Video-VAE for video representation
- Uses DiT with 3D full attention
- Includes Video-DPO preference optimization
- Provides multi-GPU parallel inference scripts
- Released under the MIT license
Use cases
- Generate short videos from written scene descriptions
- Prototype cinematic scenes and visual concepts
- Create Chinese- or English-prompted video demos
- Run local video generation on NVIDIA GPU servers
- Evaluate text-to-video systems with the Step-Video-T2V benchmark