MiniMax H3 Spectrum (Experimental)

UC_MiniMaxH3Spectrum is the experimental sibling of the pack's H3 cache node, and it's weirder and more ambitious. Instead of reusing a cached result when consecutive steps are similar, it forecasts the complete post-transformer features with a Chebyshev polynomial model, runs the real transformer on a schedule of actual evaluations, and blends the two - all while keeping native output reconstruction intact. In plain terms: it watches how H3's features evolve across steps, fits a cheap model to that history, and predicts steps the 33B transformer never has to compute. The label says Experimental for a reason.

It's part of silveroxides/ComfyUI-UtilsCollection. If you're new to this, be honest with yourself about where you are: this is a researcher's tuning node with fifteen inputs, not a "click for free speed" button. The right way to approach it is to set everything to defaults and only touch one or two knobs, or to use it in forced_actual mode to measure overhead first.

The knobs that matter

• execution_mode - spectrum (default) forecasts on eligible steps; forced_actual runs everything for real, which is your benchmark mode for measuring patch overhead with zero forecasts.
• degree (default 1) - Chebyshev polynomial degree. Degree 1 enables the first-forecast bootstrap; higher degrees fit more expressive curves but disable it. Start at 1.
• ridge_lambda (default 0.1) - regularization that stabilizes the history fit; raise it if the forecast goes unstable.
• window_size (default 2) - the initial interval between actual transformer evaluations; flex_window (0.75) grows that interval after each refresh, so the model gets bolder over time.
• warmup_steps / tail_actual_steps (defaults 1/1) - the first and last solver steps always run the real transformer. Leave them.
• video_blend_weight (default 0.5) - how much of the output is pure spectral prediction vs local linear prediction. 1 = all spectral, 0 = none.
• audio_blend_weight (default 0) - spectral share for audio rows, kept at 0 by default because direct spectral mixing of audio can hurt continuity.
• offline_smoothing_replay (default on) - a two-pass smoothing replay that protects H3 audio continuity. It's there for a reason; don't turn it off on a whim.
• history_storage / offline_archive_storage - system_ram or vram for the feature history and replay anchors. system_ram is the default and the safe choice on tight VRAM.
• debug - logs forecast decisions, timing, retained memory, and a run summary. Turn it on for exactly one run to see what's happening.

Outputs

• model - the patched clone, fed to your sampler. Nothing else.

Installing it

Via the pack:

• ComfyUI Manager: search "ComfyUI-UtilsCollection".
• Or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart ComfyUI. No extra downloads beyond H3 itself (license-gated; US/EU/UK/Korea excluded).

Gotchas

This is a speed experiment, and speed experiments on a 33B model have consequences: turn on debug, run one spectrum pass and one forced_actual pass on the same seed, and compare before you trust it. If the forecast destabilizes, raise ridge_lambda or reduce video_blend_weight rather than throwing more degree at it. And if you just want a safer speedup, the plain UC_MiniMaxH3Cache node is the conservative option - Spectrum is for people who want to squeeze further and are willing to validate the output. The audio defaults are deliberately conservative because H3 audio is easy to break; leave audio_blend_weight at 0 unless you've confirmed the audio survives.