---

name: inference description: "Operate MMAudio inference for text-to-audio, video-to-audio, experimental image-to-audio, the CLI demo, the Gradio UI, and programmatic generation." disable-model-invocation: true metadata: disco-role: operating license: MIT

MMAudio Inference

Use this sub-skill when the task is to run, package, or diagnose MMAudio inference only. It covers pretrained model selection, CLI demo usage, Gradio launch behavior, and programmatic generation wiring.

What this sub-skill covers

• demo.py for command-line text-to-audio and video-to-audio generation.
• gradio_demo.py for the browser UI, including the experimental image-to-audio tab.
• mmaudio.eval_utils.generate, load_video, load_image, and make_video for custom callers.
• mmaudio.model.networks.get_my_mmaudio and MMAudio.update_seq_lengths for duration-aware inference.
• Safe command construction for demo.py without executing the model.

Route elsewhere

• Training, checkpointing, and DDP setup -> training sub-skill.
• Dataset or feature preparation -> data-preparation sub-skill.
• Batch evaluation or onset metrics -> evaluation sub-skill.

Operating rules

1. Default to large_44k_v2 unless a smaller model or 16 kHz output is explicitly needed.
2. For any duration change, update both seq_cfg.duration and net.update_seq_lengths(...) before generation.
3. For video jobs, trust the duration returned by load_video(...); it truncates to the usable clip and sync coverage.
4. Use --skip_video_composite when the user only needs audio or when video reconstruction is unnecessary.
5. High-resolution inputs slow decode/encode work without improving quality; trim or re-encode instead of upscaling.
6. The Gradio script downloads and loads the model at import time, then exposes only its output directory for browser access.

Start here

• references/cli-reference.md
• references/api-reference.md
• references/gradio-interface.md
• references/troubleshooting.md
• scripts/build_demo_command.py

Notes

• This subtree is self-contained runtime guidance. Do not send future agents back to the original repository docs.
• Evidence labels used to build this sub-skill: README.md, docs/MODELS.md, demo.py, gradio_demo.py, mmaudio/eval_utils.py, mmaudio/model/networks.py, mmaudio/model/sequence_config.py, mmaudio/utils/download_utils.py.