Spectrum Apply MiniMax H3

MiniMax H3 is a 33B omni-modal model that renders video and native stereo audio in one pass - which is exactly why it's slow. Every sampling step is a full transformer forward over packed text, audio, and video rows. Spectrum's pitch is short: run about half of those passes, predict the rest, and end up with output that's usually close enough to native that you can't tell in a fast cut. It's training-free - no distillation, no LoRA, no API key. The one thing it isn't is lossless.

What it actually does

On an actual step, Spectrum lets native H3 run normally and captures the packed target hidden state - the [audio | video] tail rows the model produces after its final transformer block. On a forecast step, it predicts that hidden state from the previous actual anchors (a small ridge-regularized polynomial fit) and skips the expensive H3 transformer blocks for that step entirely, resuming at the native output and sampler path. The text rows and keyframe/reference rows don't participate in the forecast; only the generated target does.

A typical 20-step Euler or ER-SDE run resolves to 11 actual evaluations and 9 forecasts - roughly half the transformer cost, without touching attention backends or the VAE. But it is an approximation: forecasted steps change the denoising trajectory, so even with the same seed and prompt your output won't be byte-identical to native H3. The maintainer is upfront about that, and the README tells you to A/B it against bypassed for quality-critical work.

Where it sits

MiniMax H3 model loader
-> LoRA / model patches
-> MiniMax H3 Sigma Shift
-> Spectrum Apply MiniMax H3   <- here
-> guider / sampler

You drop it on the model chain right before the sampler, and it returns a cloned, patched model (MODEL in, MODEL out). It only works on ComfyUI's native comfy.ldm.minimax.model.MiniMaxH3Model - point it at anything else and it raises a clear TypeError rather than silently doing nothing. It's fail-closed: samplers it hasn't reviewed, or custom noise samplers it can't prove safe, fall back to native execution instead of guessing. Reviewed samplers include Euler, native and RefDelta ER-SDE, H3 Turbo, and the RES multistep family.

The inputs that matter

Most widgets are calibration knobs you'll never touch. The ones a beginner actually sets:

• enabled - flip it off to bypass Spectrum entirely (your A/B control).
• blend_weight - video spectral share, default 0.5. This is your "how aggressive" dial.
• warmup_steps and tail_actual_steps - how many native steps run at the start and end. The warmup and tail are always actual, which is a big part of why forecasts stay sane.
• max_history and history_storage - how many actual anchors are kept and where (system_ram vs vram). Large H3 hidden histories can eat multiple GiB, so keep system_ram unless you're sure.

The default offline_smoothing_replay=true uses a two-pass capture-then-replay design that was the historically validated fix for audio stutter, at the cost of a second sampler pass. model_aware_mode stays off by default; set it to full with the default generic-correction settings for the current preferred quality path on native ER-SDE. That's an advanced lane - you can ship the defaults happily.

Installing it

ComfyUI Manager / the Comfy Registry: search ComfyUI-Spectrum-MiniMax-H3 and restart. Or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3.git
# restart ComfyUI

Updating is git pull --ff-only inside the folder, then restart. It has zero Python dependencies - no requirements to pip install. What you do need is a recent-enough ComfyUI (the reviewed contract starts at commit e377e263) and, separately, the H3 weights, VAE, and audio decoder loaded through ComfyUI's native nodes - the pack ships no models. And the weights themselves sit behind the MiniMax H3 Community License, which geofences out the US, EU, UK, and Korea. Running Spectrum doesn't change that; you still need to be licensed to run H3 at all.

Where people get burned

• EasyCache or LazyCache on the same model branch. They can bypass the native H3 observation Spectrum needs, so Spectrum detects them and just stays inactive. Don't stack them.
• Frozen frames with CK attention. Issue #41 is open for a second-generation freeze on some CK + Spectrum systems; update ComfyUI past the upstream peak-VRAM fix before blaming Spectrum.
• Saved workflows keep old values. ComfyUI serializes widget values, so updating the pack doesn't rewrite them - a workflow saved before the replay default changed still runs the old offline_smoothing_replay setting.
• Expecting identical output. Same seed, same workflow, Spectrum on vs off is not a bit-for-bit match. Treat it as a speed/quality tradeoff and measure it with the pack's own benchmark nodes when it matters.

It's the fastest path to "H3, but fewer expensive steps" in ComfyUI today, and it's honest about being an accelerator rather than a free upgrade. Set debug=true and watch the decision=actual|forecast log before you tune anything else.