Spectrum Apply MiniMax H3 for ComfyUI

If you only need the current setup facts, start here. The values below reflect the current xmarre Spectrum implementation checked on August 26, 2026, rather than older workflows that may still contain legacy settings.

MiniMax H3 normally performs an expensive transformer evaluation repeatedly throughout diffusion sampling. Spectrum reduces some of those evaluations by learning from hidden features captured during real H3 steps and forecasting selected future hidden states.

On an actual step, native MiniMax H3 runs normally and Spectrum records the target hidden feature after the transformer. On a forecast step, Spectrum predicts that feature from previous actual anchors, skips the H3 transformer blocks for that step, and continues through MiniMax H3's native output and sampler path.

This is why Spectrum is better described as an approximate inference accelerator rather than a cache that reproduces native H3 exactly.

With the current default degree-1 scheduling, a typical 20-step Euler or ER-SDE single-pass run can resolve to approximately:

A F A F A F A F A F A F A F A F A F A A

That corresponds to 11 actual H3 transformer evaluations and 9 forecasts.

Option 1 — ComfyUI Manager

1. 1Open ComfyUI Manager.
2. 2Search for ComfyUI-Spectrum-MiniMax-H3.
3. 3Install the custom node.
4. 4Restart ComfyUI.
5. 5Search the canvas for Spectrum Apply MiniMax H3.

Option 2 — Git

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

To update an existing Git installation:

cd ComfyUI/custom_nodes/ComfyUI-Spectrum-MiniMax-H3
git pull --ff-only

Restart ComfyUI after installing or updating the node.

Place Spectrum after the MiniMax H3 model loader, LoRAs or compatible model patches, and MiniMax H3 Sigma Shift — but before the guider and sampler.

External H3 patches that explicitly publish Spectrum compatibility metadata should also be applied before Spectrum.

The current upstream configuration is substantially more aggressive than some early Spectrum MiniMax H3 guides. For most users, the maintainer's recommendation is simple: start with the current defaults and change settings only when you have a specific quality, sampler or memory reason.

The current default configuration uses degree=1, warmup_steps=1, blend_weight=0.50, audio_blend_weight=0.00 and system-RAM history storage.

Spectrum uses an allowlisted, fail-closed compatibility model. Reviewed sampler paths can forecast; unknown or changed sampler contracts fall back to native H3 instead of guessing.

Yes. Turbo or LightX2V-style acceleration LoRAs can be combined with Spectrum, and the combination can reduce sampling time further.

The trade-off is quality. Maintainer testing found that stacking few-step acceleration with Spectrum can produce larger changes in composition, motion and fine detail than the normal 20-step Spectrum path. Treat the combination as a separate speed-versus-quality configuration, not as free acceleration.

Compare the few-step options in the MiniMax H3 Turbo Guide.

This does not mean one method is universally faster or better. Benchmark Spectrum and a cache method independently with the same workflow if you want to compare them.

Spectrum stores feature history for forecasting and, when offline smoothing replay is enabled, stores replay anchors as well. The current defaults keep both history_storage and offline_archive_storage in system RAM.

For most users, system RAM is the safer default. Large H3 hidden histories can consume multiple GiB at higher resolution or duration, while moving the history to VRAM can increase GPU memory pressure.

For the broader GPU and VRAM decision, see MiniMax H3 Local.

There is no single trustworthy percentage that applies to every MiniMax H3 workflow. Spectrum reduces transformer evaluations, but wall-clock speed depends on the sampler, resolution, duration, precision, references, model offloading, attention backend, CPU/GPU transfer cost and hardware.

Maintainer launch benchmark Historical / reported

This early maintainer benchmark used roughly 0.5 MP output, an 8-second clip, 20 steps, an RTX PRO 6000 under WSL, W8A8 loading and SageAttention. It is useful evidence that Spectrum can reduce sampling time, but it is not a universal benchmark for current releases or other GPUs.

It can. Spectrum is approximate rather than output-identical to native MiniMax H3. Motion, timing, composition and fine details can change because forecasted steps alter the denoising trajectory.

The trade-off tends to matter most when a shot depends on fast motion, brief fine details, hands, gaze, timing or a tightly constrained action trajectory. For simple or slower shots, the difference may be less important.

This does not mean Spectrum always produces a visibly worse video. It means the output should be evaluated as a different accelerated trajectory rather than assumed to be a lossless copy of native H3.

Spectrum is an approximate accelerator, so the same MiniMax H3 prompt and seed can produce a different output when Spectrum is enabled.

The current default enables offline_smoothing_replay=true. Spectrum first performs a causal capture pass to collect the required anchors, then performs a transformer-free replay pass using the accepted smoothing trajectory.

The second pass does not double the number of MiniMax H3 transformer evaluations because it reuses the captured archive instead of running the H3 transformer again.

Spectrum originates from the CVPR 2026 paper “Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration.” The research proposes a training-free method for predicting diffusion-model features across future sampling steps.

Instead of relying only on local feature reuse, Spectrum treats denoiser features as functions over diffusion time, approximates their evolution with Chebyshev polynomials, and fits the forecasting coefficients with ridge regression.

The original research evaluated Spectrum across several diffusion image and video models. The Spectrum Apply MiniMax H3 node is a later ComfyUI integration specifically adapted to native MiniMax H3 sampling; it should not be presented as if the CVPR paper itself were a MiniMax H3 paper.

Last verified: August 26, 2026

Primary sources

Community benchmark evidence

Community benchmarks are useful for understanding real-world performance, but they are not treated as universal specifications. Hardware, node version, sampler, model precision and workflow configuration should always be shown alongside timing claims.