MiniMax H3 Audio Refine Plan / 音频精修尾段计划 (T8 Advanced EXP)

The second node in T8's Audio Refine chain, and the one that decides how the refine will run. The audit blessed the run; this node turns that blessing into a concrete, signed plan: how many steps, how much denoise, what seed, and which strategy for the model. What it deliberately does not do is sample - it produces a plan object that the next node (Dual-Clock Setup) turns into actual sampler parts. Think of it as writing the recipe and handing it to the kitchen.

Everything in the plan is fixed and deterministic, which is the whole point: a refine tail that isn't reproducible is a refine tail you can't debug. The built-in defaults are the pack's validated settings for this job - CFG 1, video shift 12 / audio shift 3 (the standard Turbo dual-clock shifts), a video mask of 0 and audio mask of 1 (only audio gets re-sampled, the picture stays pinned), the dual_clock_euler sampler and native_flow scheduler. A "KSampler-equivalent partial-tail" plan just means it matches what a KSampler would compute for that tail, but only for the audio portion.

Inputs that matter

• audit - from the Audio Refine Audit node; the plan revalidates it before committing
• refine_steps - 1–8, default 4. Low-step is the point of this feature: a handful of steps re-shape the audio without re-rolling the video
• audio_denoise - 0.01–1.0, default 0.5. How much of the audio's original noise schedule gets re-denoised; lower keeps closer to the original, higher re-imagines more
• refine_seed - the seed for the tail, so a good result is reproducible
• model_strategy - combo, default connected_model_explicit; the strategy for relating the refine model to the audited model

Outputs are the plan object (H3_T8_AUDIO_REFINE_PLAN), a decision, and report_json. Wire the plan into Dual-Clock 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

This node is where the "low-step" philosophy of the whole feature lives. The default 4 steps at 0.5 denoise is a reasonable starting point, but the two numbers you'll actually tune are audio_denoise and refine_seed. A common mistake is cranking audio_denoise toward 1.0 expecting a bigger change - it does re-imagine more, but it also drifts further from the source material the video is locked to, and the whole feature works because the video half never moves. If your refine sounds "off," it's usually the denoise, not the steps. And because the plan is deterministic, a good seed is a save point: find one that works and it'll reproduce every time you re-run.