H3RelayInternalSpectrum

A ComfyUI node in sampling/spectrum with 26 inputs and 1 output.

By akatz-ai·Created 8 days ago·Updated 4 days ago· 9

model

model

◄enabledtrue►

◄blend_weight0.50►

◄degree1►

◄ridge_lambda0.10►

◄window_size2.00►

◄flex_window0.75►

◄warmup_steps1►

◄tail_actual_steps1►

◄max_history8►

◄debugfalse►

◄history_storagesystem_ram►

◄bootstrap_first_forecasttrue►

◄anchor_residual_feedbackfalse►

◄selective_rollback_correctionfalse►

◄offline_smoothing_replaytrue►

◄audio_blend_weight0.00►

◄offline_archive_storagesystem_ram►

◄model_aware_modeoff►

◄model_aware_risk_threshold0.65►

◄model_aware_trust_shrinkagefalse►

◄model_aware_replay_generic_correctionfalse►

◄generic_correction_modecoordinate_rls►

◄generic_correction_limiterhard_clip►

◄generic_correction_limit0.40►

◄generic_correction_attenuationno_attenuation►

Categorysampling/spectrum

Inputs (26)

Name	Type	Default	Description
model	MODEL		—
enabled	BOOLEAN	true	—
blend_weight	FLOAT	0.500–1	Direct video spectral share. Audio uses the separate audio_blend_weight setting. In ordinary single-pass H3, video forecasts can still affect later audio through joint transformer calls.
degree	INT	11–16	Polynomial degree. Values other than 1 disable bootstrap_first_forecast.
ridge_lambda	FLOAT	0.100–10	—
window_size	FLOAT	2.001–16	—
flex_window	FLOAT	0.750–8	—
warmup_steps	INT	10–64	Initial native solver steps. Values above 1 disable bootstrap_first_forecast.
tail_actual_steps	INT	10–64	Requested final native tail. RES enforces its three-step solver tail. ER-SDE offline replay promotes only a penultimate step that the normal schedule would forecast, preserving a future exact terminal anchor without a blanket two-step tail.
max_history	INT	82–64	—
debug	BOOLEAN	false	—
history_storageopt	COMBO	system_ram	Storage for the bounded causal history, capped by max_history. Offline replay uses the separate offline_archive_storage setting.
bootstrap_first_forecastopt	BOOLEAN	true	Forecast solver step 1 from the actual step-0 feature. Requires degree=1 and warmup_steps<=1; incompatible settings disable it with a console warning.
anchor_residual_feedbackopt	BOOLEAN	false	Experimental video-scored actual-refresh guard; never injects a hidden residual. Disable offline_smoothing_replay before enabling this mode.
selective_rollback_correctionopt	BOOLEAN	false	Experimental thresholded, budgeted rollback for the reviewed deterministic Euler sampler only. Disable offline_smoothing_replay before enabling this mode.
offline_smoothing_replayopt	BOOLEAN	true	Compatibility-safe default and historically validated H3 audio/stutter path: capture a local-only trajectory, then apply configured blends using past and future anchors without causal video-to-audio feedback. It uses a second sampler pass and retains every actual anchor. Current controlled native ER-SDE testing favored full single-pass for one recurring temporal facial artifact; replay remains supported.
audio_blend_weightopt	FLOAT	0.000–1	Direct audio spectral share. The default 0 prevents spectral mixing of audio rows. The default offline replay path also isolates capture from video-to-audio trajectory coupling.
offline_archive_storageopt	COMBO	system_ram	Storage for every actual anchor retained until offline replay completes. This archive is not capped by max_history. Keep system_ram for constrained GPUs; vram is an explicit speed/memory tradeoff.
model_aware_modeopt	COMBO	off	Experimental model/patch-aware scheduling and confidence. 'schedule' may replace risky forecasts with actual evaluations. 'schedule_confidence' also adapts ridge regularization, usable degree, and spectral share without applying a correction. 'full' additionally applies the bounded generic latest-delta residual correction. Current controlled native ER-SDE testing prefers full single-pass among the compared model-aware quality modes. No equivalent conclusion is established for other samplers. The correction itself adds no denoiser forward.
model_aware_risk_thresholdopt	FLOAT	0.650–1	Advanced threshold for converting a prospective forecast into an actual model evaluation. Lower values are more conservative and may spend more NFEs.
model_aware_trust_shrinkageopt	BOOLEAN	false	Experimental research/reproduction switch for model_aware_mode='full'. The supported default is false and current native ER-SDE perceptual A/B testing does not recommend promotion. Offline replay keeps the rejected causal-kappa transfer disabled and uses shadow-only diagnostics. No transformer evaluation is added.
model_aware_replay_generic_correctionopt	BOOLEAN	false	Offline replay-only legacy/ablation switch for model_aware_mode='full'. False is the supported default: do not transplant the causal PR #39 latest-delta scalar onto the different future-bracket replay direction. True explicitly restores that old replay transfer for regression/scientific reproduction. The causal PR #39 correction, validation attenuation, local/spectral blending, scheduling, and transformer NFE are unchanged.
generic_correction_modeopt	COMBO	coordinate_rls	Full-mode correction controller. coordinate_rls is the validated default from three-run hidden-space and decoded-media ER-SDE evidence. Legacy retains the exact previous EWMA path for reproduction. Regional mode remains experimental. No mode adds a transformer evaluation.
generic_correction_limiteropt	COMBO	hard_clip	Generic-correction gain limiter. hard_clip with limit 0.40 is the validated full-mode default. rational with 0.25 remains the exact legacy reproduction setting.
generic_correction_limitopt	FLOAT	0.400.01–1	Symmetric gain-limit scale. 0.40 is validated with coordinate_rls/hard_clip; use 0.25 with legacy/rational/mode_default for the previous exact baseline.
generic_correction_attenuationopt	COMBO	no_attenuation	Full-mode attenuation policy. no_attenuation is the validated coordinate_rls default. mode_default with legacy exactly preserves the previous correction path. This setting adds no transformer evaluation.

Outputs (1)

Name	Type	Description
model	MODEL	—