MiniMax H3 Audio Refine Phase 2 Plan / 音频精修二阶段计划 (T8 Advanced EXP)

The single-model Audio Refine plan gives you a flexible tail - anywhere from 1 to 8 steps, any denoise from 0.01 to 1.0. This node is the strict sibling for the dual-model branch: once the Model Route has proven that your two models actually pair, this node builds a deliberately narrow Phase 2 plan - exactly four steps, and a denoise locked to one of just two pre-registered values: 0.35 or 0.50. That's not a limitation so much as a commitment: with two models in the mix, the pack only trusts the settings it has actually validated.

The philosophy here is the same one running through the whole EXP side of this pack - when in doubt, narrow the space of things that can go wrong. A dual-model refine is more exotic than a single-model one, so it gets less freedom, not more. The plan it produces is signed, deterministic, and scoped to the audio tail only, ready for the Dual-Model Setup node to turn into sampler parts.

Inputs

• route (H3_T8_AUDIO_REFINE_MODEL_ROUTE) - from the Model Route node; the proven pairing this plan is built on
• refine_steps - fixed at 4; it's in the schema as an input, but 4 is the only value, so think of it as documentation
• audio_denoise - 0.35 to 0.50, default 0.5. 0.50 is the full validated point; 0.35 is the gentler one for when you want a lighter touch
• refine_seed - the seed for the tail

Outputs: plan (H3_T8_AUDIO_REFINE_PHASE2_PLAN), a decision, and report_json. The plan feeds Dual-Model Setup.

Installing it

Part of the T8mars/comfyui-minimax-h3-audio-T8 pack - search "MiniMax H3 Audio T8" in ComfyUI Manager, install, restart, or:

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

Update ComfyUI itself first (recent core: comfy_api.latest, comfy.weight_adapter, comfy.patcher_extension, comfy.ldm.minimax); requirements.txt is intentionally empty.

The honest take

There's nothing to configure here beyond seed and one of two denoise values, which is the point - but it's worth reading the restriction as the pack's measured claim about dual-model refine. If your route got proven and the tail still sounds wrong, the answer is almost never "the plan is too limited"; it's that your model pairing is subtly off, or the first pass wasn't what you declared. Don't fight the 0.35/0.50 fence by trying to get more denoise through some other node - you'd be hand-building an unvalidated tail, which is exactly the thing this whole chain exists to prevent. For most people, single-model Audio Refine is the better starting point; the dual-model route earns its complexity only when you genuinely need different model stacks for the two passes.