AGENT SKILL

video-inference

Build and validate one-shot TurboDiffusion Wan T2V/I2V inference commands without running models.

License 许可
Apache-2.0
最近更新
2026-10-08
Stars
★ 11
Forks
3

01 前言 Front Matter

name
video-inference
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Build and validate one-shot TurboDiffusion Wan T2V/I2V inference commands without running models.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

TurboDiffusion video inference

Use this sub-skill when the task is to plan or review a one-shot TurboDiffusion video generation command for:

  • Wan2.1 text-to-video (T2V) using a single DiT checkpoint.
  • Wan2.2 image-to-video (I2V) using separate high-noise and low-noise DiT checkpoints.
  • Prompt, image, model/checkpoint, quantization, attention, output-path, and no-download preflight decisions.

Do not use this sub-skill for interactive multi-turn generation; route that to interactive-serving. Route checkpoint conversion, quantization, rCM/SLA training, and checkpoint merging to training-and-checkpoints. Route CUDA extension, Sparse/SageSLA dependency, custom-op, and low-level acceleration failures to acceleration-backends.

Read first

Bundled helpers

These helpers only render shell commands. They do not import TurboDiffusion, download assets, run models, train, convert checkpoints, or access credentials.

Source-layout command quirk

TurboDiffusion's public one-shot scripts are usually launched from a source layout with helper packages such as imaginaire, rcm, serve, SLA, and ops importable from the inner turbodiffusion source directory. If a command fails with an import such as No module named imaginaire, add a public source-layout PYTHONPATH pointing at that inner source directory, for example PYTHONPATH=turbodiffusion, or run from an environment/layout where those helper packages are already importable.

04 同仓库更多 Skills