MiniMax H3 Text to Audio

MiniMax H3 is a video model that generates sound and picture together - which is exactly the problem if all you want is a music track. Run it normally and you're paying (in VRAM and seconds) for a 1344×768 picture you're going to delete. MiniMaxH3TextToAudio is the workaround: it feeds H3 a dummy video canvas shrunk to 224×128 while keeping the audio latent at full quality, so the shared DiT spends its tokens on the sound and the video becomes cheap padding. The result is text-to-music at a practical cost, decoded by core ComfyUI's VAEDecodeAudio - no separate video VAE in the loop.

What it actually does

Take a clip, a prompt, and a length in seconds, and you get two outputs:

• positive - a CONDITIONING tensor, wired into a BasicGuider
• latent - an audio+video latent (a ComfyUI NestedTensor), wired into SamplerCustomAdvanced.latent_image

The prompt is tokenized and encoded on the spot (clip.tokenize + encode_from_tokens_scheduled), so you don't need a separate text-encode step.

The mechanism, ground truth

The source builds an empty NestedTensor with two branches: a video branch of zeros sized [1, 24, latent_t, h//16, w//16] at your canvas resolution, and a stereo audio branch [1, 32, 2, audio_t] at 40 fps. Your requested seconds get snapped up to the 17k+5 frame grid at 24 fps - the minimum real length is 124 frames, about 5.17 seconds, which is why seconds has a 5.16 floor. Two quirks baked into the code: batch is fixed at 1 (H3's NestedTensor path rejects bigger batches), and there's a warning logged once you cross 15.1s, because the ~5–15s window is the trained range and anything longer is experimental.

The inputs that matter

• clip - from CLIPLoader, the qwen3vl_32b_minimax_h3_int8_convrot clip loaded with type minimax. This is the tokenizer/encoder that actually understands the prompt format.
• prompt - wire in MiniMaxH3AudioPrompt's output. H3 wants its structured T2VA format (integrated_multimodal_description / overall_soundscape / non_diegetic_music), not a loose sentence; feeding a raw sentence is how you get mush.
• seconds - default 15. The trained range is ~5–15s; past 15.1s the node warns you're outside it. This is the field that actually sets the clip length (unlike the same-named field on the prompt node).
• video_canvas - the dummy video resolution, a dropdown of 7:4 options from 112x64 up to native 1344x768. Leave it at 224x128: it's the default for a reason, and the point of this node is not to generate video. Bigger canvases only slow you down for audio-only work.

Install and the model files

cd ComfyUI/custom_nodes
git clone https://github.com/t22m003/ComfyUI-MiniMaxH3Audio

Then restart. (Or use ComfyUI Manager and search "ComfyUI-MiniMaxH3Audio".) There are no Python dependencies in this pack - but you need the three H3 checkpoint files sitting in your models folders: minimax_h3_fl2va_bf16 (UNet), qwen3vl_32b_minimax_h3_int8_convrot (CLIP), and minimax_h3_audio_vae_fp32 (VAE). These come from ComfyUI's official MiniMax H3 integration, and they're big.

The trap to remember

BasicGuider has no CFG, so a negative prompt does nothing. Put exclusions in the positive music text instead - "no vocals", "no drums" - that's the author's explicit recommendation. And before you commit to H3 locally, check the MiniMax H3 Community License: it excludes the US, EU, UK, and South Korea, outputs included. If you're in one of those regions, this whole local pipeline isn't licensed for you, whatever the hardware can handle.