DEPLOY.md

# Omni-Video-Factory — public-models build (no secrets, no filter, no tracking)

A fully self-contained copy of the Omni-Video-Factory Space that works on a
fresh duplicate with NO environment variables or secrets:

- Model manifest is BUILT IN (DEFAULT_OMNI_VIDEOS in src/app_lib.py) and points
  at four PUBLIC Hugging Face files:
    * Comfy-Org/Wan_2.1_ComfyUI_repackaged - WAN VAE + UMT5-XXL text encoder
    * jorgmikel76/wan22EnhancedNSFWCameraPrompt_nsfwFASTMOVEFP8High - the
      WAN 2.2 14B "EnhancedNSFWCameraPrompt" high/low-noise FP8 model pair
      (the same model family the original space loads).
- The ComfyUI runtime is cloned from the PUBLIC repo selfitcamera/ComfyUIVideo
  (commit 17b9fb3) - no GitHub token required (the token gate was removed).
- The workflows load the models via ComfyUI's standard UNETLoader (GGUF
  loaders converted) - no model files need to be shipped in the repo.

## Removed (privacy/filter build)
- NSFW filter: src/config.pyc has NSFW_ENABLE=false and the code's
  _nsfw_enabled()/_nsfw_policy_applies() hard-return False, so the filter can
  never be re-enabled by config or env.
- Redis usage tracking disabled (redis = None): no per-IP counters, no country
  stats, no NSFW hit counting, no IP sets.
- Country/IP geo disabled (_request_country returns ""): no geo header reads,
  no external lookups to ipwho.is / ipapi.co. IPs never leave the server.
- Console snapshot logging disabled (_log_usage_snapshot no-op): no
  "用量统计 ... IP=..." lines in the Space logs.
- External LLM disabled (_llm_generate_scene_prompt_text): auto-scene splits
  prompts locally; no prompt text is POSTed to omnifilm.net or anywhere else.
- Gradio analytics remains (anonymous pings to gradio.app; no content).

## Deploy
1. Create a new Space: https://huggingface.co/new-space
   - SDK: Gradio, Python 3.12 (README frontmatter already sets this).
2. Upload the contents of this zip as the repo root (web UI: Add file ->
   Upload files, or git push).
3. Settings -> Hardware -> ZeroGPU (Nvidia L4S).
4. Settings -> Storage -> increase persistent storage to >= 50 GB (models are
   ~35 GB; HF grants this to ZeroGPU spaces on request, usually free).
   REQUIRED - the default 20 GB is not enough for the two 14.3 GB models.
5. Wait for first boot: it downloads the 4 models (~35 GB) and clones the
   Comfy runtime; first generation takes a while after the Space appears
   "Running". Watch the Space logs for "using workflow dir" / "runtime ready"
   / "all required models".

No env vars are needed. Optional overrides if you ever want them:
- OMNI_VIDEOS - replace the built-in model manifest (same 4-part #-format).
- HF_TOKEN - only if you point OMNI_VIDEOS at gated repos.
- GITHUB_TOKEN - only if you change OMNI_RUNTIME_GIT_REPO to a private repo.
- Do NOT set OMNI_PRIVATE_CONFIG_B64 / OMNI_PRIVATE_CONFIG_PY: they
  regenerate src/config.pyc at boot and could re-enable the filter.

## Pointing the Perchance page at it
Set CFG.api in the Perchance generator's index.html to:
  https://<your-username>-<space-name>.hf.space/gradio_api