H3RelayInternalSpectrum

Open the H3 Relay example workflow, expand the hidden nodes, and you'll find H3RelayInternalSpectrum - a wall of twenty-ish parameters that looks like an invitation to ruin your settings. It isn't. It's a vendored copy of the Comfy-Org Spectrum MiniMax H3 runtime (by xmarre), and it's the sampling layer that sits between your model and the sampler. The good news: the defaults are the validated configuration, and Sequence Start's spectrum_enabled toggle switches the whole thing on and off for you. Leave it on.

So what does Spectrum actually do? It's a step-forecasting and spectral-blending layer for H3. Instead of paying full inference cost on every diffusion step, it keeps a bounded history of actual solver features, fits a polynomial to them, and forecasts near-future steps. Then it blends the forecast against the real trajectory with a weight you control. On top of that, its default offline_smoothing_replay path addresses the classic H3 complaint: audio stutter and temporal inconsistency. It captures a local trajectory, then replays it offline applying blends using past and future anchors - at the cost of a second sampler pass, which is why the tooltip calls it the "historically validated H3 audio/stutter path."

The inputs that matter

• model (MODEL) - the model to wrap. Chain this node's model output into your sampler.
• enabled (default true) - Spectrum's on/off, independent of anything upstream.
• blend_weight (default 0.5) - the direct video spectral share: how much forecast vs. actual.
• audio_blend_weight (default 0) - the audio share. Default zero is deliberate: it prevents spectral mixing of the audio rows, which is what keeps H3's generated sound intact.
• degree (1) - polynomial degree. Anything above 1 disables bootstrap_first_forecast.
• warmup_steps (1) - how many initial native solver steps run before forecasting kicks in.
• max_history (8) - how much bounded causal history the forecaster keeps.
• tail_actual_steps (1) - requested final native tail; the sampler's own rules may override it.

The experimental pile - do not touch

model_aware_mode (default off), anchor_residual_feedback, selective_rollback_correction, generic_correction_*, model_aware_* - these are research switches, and the tooltips say so. The pattern to remember: the experimental modes that apply corrections explicitly tell you to disable offline_smoothing_replay first, and the supported defaults for all of them are the conservative values. If you're not reproducing a specific experiment, changing them makes results worse.

One automatic behavior worth knowing: bootstrap_first_forecast defaults to true but quietly disables itself (with a console warning) if your degree or warmup_steps settings make it invalid. It fails soft, not hard.

Why it's here

Spectrum is why the H3 Relay default is "Euler + beta57 + Spectrum": the combination gives you H3's intended manual sampling curve (16 steps, alpha 0.5 / beta 0.7) plus this smoothing layer, and Sequence Start wires spectrum_enabled straight into it. If you switch to a sampler Spectrum doesn't support, the pack simply runs without it - Spectrum is bypassed, not broken.

Install is the shared pack path: ComfyUI Manager → H3 Relay, or git clone https://github.com/akatz-ai/h3-relay into custom_nodes, restart. Needs the MiniMax H3 models (separately licensed, territorial) and FFmpeg.