audio/README.md

SFX training extension

Open-Sora-v2 itself does not provide a native SFX generator checkpoint. Keep SFX as a separate model/adapter until a licensed audio dataset and real audio generator are selected.

The intended v2.1 artifact layout is:

/audio/
  sfx_model.safetensors
  config.json

Training examples should contain licensed audio plus event metadata such as:

{
  "start": 1.25,
  "duration": 0.35,
  "event": "footstep",
  "intensity": 0.72
}

The video generation Job can then create a timeline and mux generated SFX into the final MP4. Do not claim SFX is trained until an actual audio model has been trained and evaluated.