WorldFoundry
English中文

Step-Video-T2V

Method overview: Architecture overview of Step-Video-T2V. Videos are represented by a high-compression Video-VAE, achieving 16x16 spatial and 8x temporal compression ratios. User prompts are encoded using two bilingual pre-trained.

Step-Video-T2V from StepFun: text-to-video and video-generation; Verified.

Edit this pageGitHubMarkdown

Step-Video-T2V is StepFun's large open text-to-video model (30B parameters upstream, MIT weights at stepfun-ai/stepvideo-t2v). In WorldFoundry the public pipeline is assembled from native role-based components with the shared diffusion runner — DiT, Step-LLM text encoder, CLIP, scheduler, latent initialization, and VAE decoding all execute in-process.

Method overview: Architecture overview of Step-Video-T2V. Videos are represented by a high-compression Video-VAE, achieving 16x16 spatial and 8x temporal compression ratios. User prompts are encoded using two bilingual pre-trained.
Method overview: Architecture overview of Step-Video-T2V. Videos are represented by a high-compression Video-VAE, achieving 16x16 spatial and 8x temporal compression ratios. User prompts are encoded using two bilingual pre-trained.

The recorded paper is arXiv 2502.10248.

Code: stepfun-ai/Step-Video-T2V.

Build your run

Pipeline
StepVideoT2VPipelineworldfoundry.pipelines.step_video.pipeline_step_video_t2v
TaskInteractive Video interactive-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  step-video-t2v \
  --pipeline.task-profile interactive-video \
  --json

What it is for

Pick step-video-t2v when you want this family's text-to-video line in WorldFoundry: a prompt in, a generated clip out. Use it as the catalog T2V start rather than an editor or a world-model stepper.

Upstream: a 30B DiT with a deeply compressed Video-VAE (16x16 spatial, 8x temporal) and bilingual text encoders. This repository re-assembles the native pipeline in-process; the former xFuser and caption/VAE HTTP-service runtime is not a backend. A checkpoint-backed 204-frame GPU artifact is verified for one prompt.

Catalog tasks include text-to-video and video-generation. Catalog aliases include step-video, stepvideo-t2v, and step_video_t2v. Documented modalities: text in, video out.

Recorded evaluation entry points include VBench (Standard open T2V quality suite).

Runtime contract

Optional prompt, negative_prompt, frames, fps, and height; artifacts generated_video and step_video_t2v.mp4.

CLI task profiles: interactive-video. WorldFoundry binds StepVideoT2VPipeline.

Family status: Verified.

Recorded interactive-video defaults: frames=204, fps=25, height=544, width=992.

A checkpoint-backed 204-frame GPU artifact at 992x544 and 50 steps is verified for one prompt. Use id step-video-t2v. At 30B scale plan for serious VRAM or multi-GPU setups.

Running in WorldFoundry

The catalog launch is worldfoundry-eval run step-video-t2v --pipeline.task-profile interactive-video. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8).

Recorded weight repositories include stepfun-ai/stepvideo-t2v. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh step-video-t2v. Add --download to fetch missing public checkpoints.

Recorded licenses include mit; status: Verified.