MiniMax H3 Audio Refine Compatibility Plan / 音频精修兼容计划 (T8 Advanced EXP)

The MiniMax H3 Audio Refine compatibility chain is three nodes that answer one question: can I tack a short audio refinement onto a first pass I already ran, without pretending it's the same thing as training on more steps? This Plan node is the middle of that chain. The Route node signs off on what your first pass actually was, this node builds the fixed four-NFE audio tail, and the Setup node hands it to the sampler. On its own it does nothing; in the chain it's the part that decides how much audio re-denoising you get.

Why you'd bother: H3 generates video and audio jointly in one omni-modal transformer. That's the model's whole party trick - native stereo audio that stays in sync because it was never a separate pass. But when you run a distilled low-step recipe like Turbo4 or PDD 8-step, the video can look great while the audio comes out muffled, metallic, or short on transient detail. The fix this pack ships is a partial-tail audio refinement: after the first pass finishes, run a few extra steps that re-denoise the audio slice of the joint AV latent. Only it's not actually audio-only - every step is still a full H3 transformer forward. The README says this in the small print: an uncached refine step costs roughly as much as a first-pass step. Don't expect a free lunch.

What the Plan node actually does

It takes the signed route from MiniMaxH3AudioRefineCompatibilityRouteT8Advanced and turns it into a concrete plan. Two things matter here:

• refine_steps is locked to 4 - the schema allows exactly 4 and nothing else. This is the "four-NFE final audio tail" and it's fixed by design.
• audio_denoise runs from 0.35 to 0.5, defaulting to 0.50. 0.35 is the conservative listen-and-compare point. Note what's not offered: 1.0, full audio regeneration, is deliberately left out of this node. If you want to re-roll the audio from scratch, this isn't the tool.
• refine_seed - set it once and keep it if you want A/B comparisons to mean anything.

Outputs are plan (feeds the Setup node), plus decision and report_json strings you can drop into a text preview. The author's own description is worth quoting because it's the pack's personality in one sentence: total NFE is reported as cost only, "never as a training-distribution equivalence claim." This pack will happily tell you the first pass ran 8 NFE and the refine adds 4 more, but it refuses to pretend 4+4 equals the quality of a native 8. That honesty is rare and it's why you should trust the report_json over vibes.

How it fits and how to install

The whole chain assumes you've already generated an H3 AV latent - Turbo4, Turbo8, learned two-pass, PDD, EAV, Prompt Relay, or long-video segments all feed the Route node first. So this is an advanced workflow, not a starting point. Use the pack's own 18-audio-refine example workflows rather than hand-wiring it.

Install is the standard ComfyUI Manager route: search "MiniMax H3 Audio T8" and hit install, then restart. Or clone manually:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

One thing this pack insists on, and it's real: update ComfyUI itself first. It uses the new comfy_api.latest, comfy.weight_adapter and native MiniMax H3 support, so if every T8 node goes red at once it's your ComfyUI version, not the workflow. The requirements.txt being empty is intentional - torch, torchaudio, numpy and Pillow come from ComfyUI. Don't go installing extras over your Torch/CUDA stack to fix it.

Where people get burned: chaining two sampler-owning nodes at once, or treating a refine candidate as automatically better. The pack's own blind test ended with "refined slightly wins on one clip, original sounded louder on another." Listen to both before you accept anything.