MiniMax H3 SPEED Spectrum Dataset Finalize / 频谱数据集定稿 (Advanced)

This is the node where a pile of measurements becomes a claim. MiniMaxH3SPEEDSpectrumDatasetFinalizeT8Advanced takes the accumulated statistics you've built clip-by-clip and fits them into one H3 SPEED profile - the H3_T8_SPEED_PROFILE that can authorize the delta-optimal transition mode in the Plan node. And it refuses to do that on weak evidence. The description sets the bar in plain terms: fewer than 100 unique clips, or a weak fit, stays a research probe and cannot authorize validated delta-optimal mode. This is the gate that keeps the whole pipeline honest.

How it works

Feed it the accumulated spectrum_dataset from Accumulate/File and name the result (profile_name, default h3_dataset_spectrum_v1). Two thresholds decide whether the result is a usable profile or a probe:

• minimum_independent_clips (default 100, floor 100) - the size gate. Note the honest framing baked into the whole chain: "independent" means genuinely distinct source clips, tracked by Accumulate's batch-ID deduplication and provenance inputs - not 100 rows you re-accumulated from the same latent.
• minimum_r_squared (default 0.8) - the quality gate. R² measures how well the power-law model actually explains the data. A noisy or multi-modal sample set won't hit 0.8, and the node won't pretend otherwise.

The fit is per task/model/VAE binding - the profile inherits the task_family and the checkpoint/VAE fingerprints the dataset was accumulated under. A T2VA profile on a different checkpoint is not the same profile, and downstream, the Plan node's require_validated_profile policy will reject cross-binding use.

Outputs: spectrum_profile (into the Plan node's optional spectrum_profile input) and report_json with the fit statistics - R², clip count, the fitted parameters, and explicitly whether this cleared the bar for validated mode or remains a research probe.

Why the gate exists

Delta-optimal SPEED transitions compute where to switch resolutions from the spectrum math. If you feed that math a garbage fit, you get garbage transition points - and unlike a manual sigma you picked yourself, a garbage automated transition is harder to spot because it looks principled. The author's response is the 100-clip / 0.8-R² bar: a real dataset is expensive to assemble, which is exactly why it's trustworthy, and the cost of assembling one filters out the impulse to fake it. If your fit comes back as a probe, that's not a failure - that's the node correctly telling you your evidence isn't there yet.

Installing

Standard pack install: ComfyUI Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into custom_nodes, restart. No pip deps; no GPU - it's statistics over already-accumulated data. Lifecycle: Harvester (probe) → Accumulate (×many) → File (save/load) → this node → Plan (delta_optimal). Reach this node and it accepted your profile, you've done something most people won't bother doing - and you can trust your SPEED transitions accordingly.