stepfun-ai

Step-Video-T2V: Generate Videos from Text Prompts

Step-Video-T2V is a 30-billion-parameter text-to-video model that generates clips up to 204 frames from English or Chinese prompts. It offers downloadable weights and inference code, but practical use requires substantial GPU memory and multi-GPU setup.

View on GitHub Compare
Stars
3.2k
Forks
339
Stars, last 30 days
+2

Our take

Health score

24/100

How it is scored
Activity
19
Community
36
Issues
21
Pull requests
20

+2 stars in the last 30 days

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

3,189

Forks

335

Change since Nov 14, 2026: +1 stars·0 forks

Overview

Step-Video-T2V is a text-to-video generation model from StepFun. It turns English or Chinese text prompts into videos, targeting users who need high-fidelity, longer clips and can run large generative models on their own hardware.

The project combines a video compression autoencoder with a diffusion transformer and applies preference optimization to generated video. The repository provides inference code and model weights, as well as a distilled Turbo variant. It is best suited to research, experimentation, or production workflows with access to powerful NVIDIA GPUs, rather than lightweight local use.

Key Features

  • Generates videos from prompts in English and Chinese.
  • Supports video generation up to 204 frames.
  • Uses a video VAE with 16×16 spatial and 8× temporal compression.
  • Builds on a 30-billion-parameter DiT with 3D full attention.
  • Includes a Turbo model variant with inference-step distillation.
  • Provides multi-GPU inference code and links to model weights on Hugging Face and ModelScope.
  • Includes Step-Video-T2V-Eval, a benchmark based on 128 Chinese prompts across 11 categories.

Use Cases

  • Video generation researchers can study a large text-to-video model and its reported architecture and training approach.
  • Teams with multi-GPU servers can generate longer prompt-driven clips for creative prototyping or evaluation.
  • Developers comparing video-generation systems can use the included evaluation benchmark as a starting point.
  • Users who want to experiment with reduced inference steps can try the Turbo weights, subject to the same hardware constraints.

Getting Started

The repository recommends Python 3.10 or later, PyTorch 2.3 with CUDA 12.1, and a CUDA-capable NVIDIA GPU. Clone and install the project:

git clone https://github.com/stepfun-ai/Step-Video-T2V.git
cd Step-Video-T2V
pip install -e .

See the README for model downloads, GPU setup, and inference commands.

Alternatives

  • Wan2.2: Wan2.2 also offers local video generation, with text, image, and audio-driven workflows beyond Step-Video-T2V's text-prompt focus.
ProjectLanguageLicenseStarsStatus
Step-Video-T2VPythonMIT3.2kNot checked yet
Wan2.2PythonApache-2.017.7kMaintained

Considerations

  • This is a large model: the README reports peak memory requirements around 72–79 GB for its listed generation configurations. It recommends GPUs with 80 GB of memory and describes testing on four GPUs.
  • The documented primary inference path uses multi-GPU parallel deployment and separate services for text encoding and VAE decoding. Single-GPU inference and quantization are referenced through the external DiffSynth-Studio project.
  • The text encoder's self-attention supports CUDA compute capabilities sm_80, sm_86, and sm_90 only, according to the README.
  • The repository was last pushed on 2025-03-17. Its model and hardware requirements make it a specialized option, not a lightweight general-purpose video tool.

Found this useful?

Share it with someone who would like Step-Video-T2V.

Comparisons