---

name: sana description: "Use and maintain Sana image, video, training, evaluation, and deployment workflows." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

Sana

Use this repo skill when the task names Sana, SANA-1.5, SANA-Sprint, SANA-Video, LongSANA, SANA-WM, SANA-Streaming, or Sol-RL, or when the request spans more than one Sana workflow family. This root skill is a router for the sub-skills below, not a long manual.

This skill is self-contained guidance distilled from the Sana repository. It does not require the original checkout to remain available once this tree has been generated.

Route By Task

User task	Read
Still-image generation, prompt-file image batches, ControlNet HED, 4-bit/8-bit image inference, 2K/4K tiling, Gradio image demos	image-generation
SANA-Video, LongSANA, SANA-WM, SANA-WM streaming, or SANA-Streaming V2V	video-world-streaming
Image/video training, data layout validation, FSDP/DDP launch planning, LoRA/DreamBooth, Sprint, Sol-RL, or Cosmos-RL boundaries	training-data-configs
Metrics, checkpoint conversion/export, sana-run, sana-upload, SGLang, ComfyUI, or deployment planning	evaluation-conversion-deployment

Common Entry Points

• python scripts/check_sana_install.py performs a safe import/CLI smoke check.
• python scripts/check_sana_install.py --skip-cli keeps the check import-only.
• python -m pip install -e . is the expected editable install path after the required Python 3.11/CUDA stack is ready.
• sana-run --help and sana-upload --help are the public entry points for the two console scripts.

Shared References

• references/troubleshooting.md: cross-cutting install, import, backend, and CLI triage.
• references/repo-provenance.md: source snapshot and staleness cues for this generated skill.
• references/repo-routing-metadata.json: structured managed-router metadata for this skill.

Start Here

1. Pick the route that matches the user request.
2. Read the sub-skill before drafting commands or explanations.
3. Use the bundled helper in that sub-skill if the workflow has one.
4. Keep final guidance self-contained; do not point runtime instructions back at the original checkout.

Boundaries

• Do not turn this root into a model runner, trainer, benchmark harness, or download assistant.
• Do not duplicate the full workflow manuals here; the sub-skills own those details.
• Do not assume CPU-only smoke checks prove a CUDA generation or distributed-training workflow works.