MiniMax H3 Multi-Rate Sampler (EXP/T8)

Most H3 sampling treats video and audio as one lump: N steps, done. MiniMaxH3MultiRateSamplerEXPT8 is the experimental sampler that splits them - the video gets fewer Euler macro steps while the audio gets more micro steps. The pitch is appealing: don't waste expensive model calls refining frames that are already fine, spend the extra budget on the audio that needs it. And MiniMax H3 is genuinely built for this kind of thinking, because it's a joint AV transformer where video and audio share compute.

Here's the caveat the tooltip smacks you with, and it's the most important sentence in this whole article: audio_steps is the actual number of full joint H3 DiT calls. Video and audio are generated together - the transformer runs as one model. So "4 video steps, 10 audio steps" does not cost 4 model evaluations, it costs about 10. Every audio micro step is a full joint forward pass. If you were hoping this was a free lunch, it isn't; it's a budget-allocation choice, not a speed hack.

Inputs

• model and av_latent - your H3 MODEL and the joint audio/video latent, same as any H3 sampler.
• video_steps - committed video Euler macro updates, default 4.
• audio_steps - audio micro updates, default 8, and it must be at least video_steps. The author's suggested starting point is 4/8.
• shift_video (12) and shift_audio (3) - the same sigma-shift values the pack uses across its dual-clock samplers; the defaults are the ones the rest of the ecosystem expects.

Outputs

model, sampler, and sigmas - the standard sampler trio, wired into a KSampler-style node. Note there's no report_json here; the node keeps its outputs minimal.

Install and where it fits

Pack install, once: ComfyUI Manager → search MiniMax H3 Audio T8, or cd ComfyUI/custom_nodes && git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8, then restart. No extra pip deps, no model downloads - the H3 DiT, CLIP and VAEs come from your existing setup.

Use it when you're in the regime where video looks done at 4 steps but the audio track still feels undercooked - exactly the scenario the README's 02-audio-control workflows explore. Where people get burned is thinking 4/8 is cheaper than 8/8. It isn't, really; you're just spending the 8 forward passes on audio refinement rather than spreading them evenly. If you're on a 16GB card, remember the pack's standing advice: keep canvas and frame count modest, because every step is a full joint AV forward. And it's Experimental, so A/B it against the stock dual-clock sampler before you build a production queue on it - the author's own README leans hard on "this is a fixed-single-material conclusion, not a general quality guarantee."