This documentation is organized around the public inference workflow. Start with Quick Start, then open the detailed page for the step you need.
| Guide | Purpose |
|---|---|
| Quick Start | Install dependencies, set the model root, and run one end-to-end generation. |
| Prompt Preparation | Rewrite plain prompts into structured JSON captions and optionally create per-sample negative prompts. |
| Diffusers Inference | Run diffusers inference for T2I, T2V, TI2V, and Refinement. |
| SGLang Diffusion Inference | Run SGLang Diffusion for parallel accelerated inference. |
| Few-Step (DMD Student) Sampling | Sample DMD-distilled student checkpoints in 8 steps with the trained DDIM→Euler geometry (t2v and ti2v). |
| Inference Performance Benchmark | Compare Dense/MoE Single GPU, CP8, DiT FSDP, VLM FSDP, and FP8 speed-memory trade-offs. |
- Convert plain prompts into structured prompts with the prompt rewriter.
- Optionally run auto negative to create a per-sample negative prompt, or use the default negative prompt.
- Run DiT inference with
scripts/inference.py --prompt_json. - Select
--backend diffusersor--backend sglangfor inference.