MiniMax H3 SPEED Spectrum Harvester / 空间频谱标定 (Advanced)

SPEED's fancier transition math works by modelling the target model's latent spectrum - the frequency-domain power distribution of what the model actually generates. For WAN and Flux those spectra are known quantities people have fit and shared. For H3, they aren't - and the fastest way to get them wrong is to borrow a neighbour model's constants and pretend they transfer. MiniMaxH3SPEEDSpectrumHarvesterT8Advanced is the node that measures H3's own: it fits the power-law model P(omega)=A|omega|^-beta from real H3 video latent samples. The description's single most important sentence: "It never substitutes WAN/FLUX constants."

How it works

You feed it a video_latent - and note the tooltip's warning, which is load-bearing: it must be the separated H3 video LATENT [B,24,T,H,W], not the joint AV latent. The pack has nodes for pulling the video stream out of H3's nested AV latent; wire that, not the combined thing. The node samples up to max_temporal_samples (32) frames from the batch, computes the spatial power spectrum, and fits the A/β power law. Outputs: a spectrum_profile (the H3_T8_SPEED_PROFILE object) and a report_json with the fitted parameters and their quality.

You also declare the identity of the measurement:

• profile_name - labels this profile (h3_local_spectrum_probe by default).
• task_family - T2VA, I2VA, FL2VA, L2VA, Ref2VA, Hybrid. A spectrum measured on T2VA output is not automatically valid for Ref2VA.
• checkpoint_fingerprint and vae_fingerprint - where the Model+VAE Fingerprint node's hashes plug in. Default unrecorded is allowed, but the tooltip is blunt: "filenames alone are not proof."
• independent_clip_count (default 1) and minimum_r_squared (0.8). Here's the honesty built into the schema: typing 100 for one clip does not promote it. The tooltip spells it out - dataset status needs 100 actual batch entries, and independence is a provenance assertion, not a number you type. A single clip's fit is a research probe, not a dataset.

Why the probe/dataset distinction matters

Downstream, the Plan node's profile_policy is require_validated_profile by default. A probe profile doesn't pass that bar - it's information, not authorization. The author's stance is consistent across the whole SPEED family: you may build whatever you like, but the pipeline refuses to pretend a small sample is a calibrated fact. That's rare in this ecosystem and worth respecting.

Installing and pairing

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; you need H3 weights and a sampled latent. Practical wiring: a generated av_latent → video-latent separation → this node → (optional) the Accumulate node for building a real dataset. One more thing the report tells you: the fitted β and R² tell you whether your sample even fits the power-law shape - if R² is ugly, the "spectrum" story is weak no matter how many clips you have.