MiniMax H3 Audio

MiniMax H3 Audio

Text-to-music / audio-only custom nodes for MiniMax H3 in ComfyUI. Shrinks the dummy video canvas (preserving the native 7:4 aspect ratio of 1344×768) so the shared DiT allocates most tokens to the audio stream, then decodes audio with core VAEDecodeAudio.

Nodes

| Node | Role | |------|------| | MiniMaxH3AudioPrompt | Compose T2VA fields (integrated_multimodal_description / overall_soundscape / non_diegetic_music) | | MiniMaxH3TextToAudio | clip + prompt + seconds → conditioning + AV latent (no video VAE needed) |

Typical Workflow

1. UNETLoader(minimax_h3_fl2va_bf16) → MiniMaxH3SigmaShift(12/3) → BasicGuider
2. CLIPLoader(qwen3vl_32b_minimax_h3_int8_convrot, type=minimax) → MiniMaxH3TextToAudio
3. MiniMaxH3AudioPrompt → MiniMaxH3TextToAudio.prompt
4. SamplerCustomAdvanced (res_multistep / simple / 20 steps) → VAEDecodeAudio(minimax_h3_audio_vae_fp32) → SaveAudioMP3

An example workflow JSON is available at example_workflows/minimax_h3_text_to_music.json.

Configuration & Defaults

• video_canvas: 224x128 (significantly faster generation than full 1344×768 resolution while maintaining audio quality)
• seconds: 15.0 (trained range is ~5–15s; longer durations are experimental)
• scene: General background scene text (H3 correlates music generation with visual descriptions)

Usage Notes & Limitations

• BasicGuider has no CFG support; negative prompts are ignored. Put exclusions in the positive music prompt text instead.
• Batch size is fixed at 1 (MiniMax H3 pack requirement).
• Recommended sigma shift settings: 12 / 3.

Running Tests

pytest tests/