SpectrumSEAKSamplerModGuidance

The old SEA sampler, now an alias

By sorryhyun·Created 4 months ago·Updated 22 days ago· 25

model
positive
negative
latent_image
clip

LATENT

◄seed0►

◄steps28►

◄cfg4.0►

◄sampler_name▾►

◄scheduler▾►

◄denoise1.00►

◄quality_tagshighres, best quality, score_7►

◄quality_negscore_1, score_2, score_3, worst quality, lowres, old, bad hands, bad anatomy►

◄mod_w_profilestep_i8_skip27►

◄refresh_ratio0.00►

◄adaptive_smc_alpha0.10►

◄xattn_boost1.00►

◄fsgfalse►

Another retired class name, same story as its sibling SpectrumKSamplerModGuidance. This one used to be "KSampler (Spectrum SEA + Mod Guidance)" - the version of the sampler built specifically around SEA content-aware step scheduling. When the pack unified its sampler nodes, this one folded into the same place: KSampler (Spectrum). The class key is kept as a hidden alias purely so that saved workflows referencing it keep loading - it no longer shows up if you search for it in the Add Node menu.

What SEA was, and why it's not a separate node anymore

SEA (Spectral-Evolution-Aware scheduling) is the smarter alternative to Spectrum's plain fixed-cadence caching window. Instead of always doing an actual forward every N cached steps regardless of what's happening in the image, SEA tracks the accumulated latent distance and only forces a real forward when that distance crosses a calibrated threshold - so compute gets spent on the steps that are actually moving the image, not on a metronome. It used to be its own node; now it's just the refresh_ratio dial on the unified sampler, which is exactly the schema this alias still carries:

refresh_ratio - 0 (default) is SEA auto: content-aware placement, calibrated so the total speed matches the plain window schedule at your step count. -1 turns SEA off entirely, giving you the plain growing-window schedule with no calibration delay. Anything above 0 is an explicit refresh ratio you set by hand - lower is faster and less faithful. The first run at a given resolution/steps/cfg/refresh_ratio combination pays a one-time calibration cost and caches the result to your ComfyUI user directory, so expect the very first render at a new config to be a touch slower.
The rest of the schema matches the unified sampler exactly: standard KSampler inputs, optional clip + quality_tags/quality_neg/mod_w_profile for modulation guidance, adaptive_smc_alpha for detail recovery, xattn_boost for text adherence, and fsg for the heavier Foresight Guidance path. Output is LATENT.

What to actually do

Old workflow using this node? Leave it - it loads and behaves exactly as before. New workflow? Add KSampler (Spectrum) instead; refresh_ratio is right there on it, same as it always was on this one.

Installing it

Same pack as every node here - ComfyUI Manager: search SpectrumKSampler, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/sorryhyun/ComfyUI-Spectrum-KSampler

Restart. Modulation guidance's ~12MB adapter auto-downloads to ComfyUI/models/anima_mod_guidance/ the first time you actually use it with a CLIP wired in.

Common issues & troubleshooting

Can't find it in the node search. Intentional - it's a hidden legacy alias. Use KSampler (Spectrum) going forward; the SEA dial is refresh_ratio on that node.

Speed varies between the first render and later ones at the same settings. That's SEA's calibration pass, not a bug - it only happens once per unique (resolution, steps, cfg, refresh_ratio) combination, then reuses the cached threshold.

Want the fastest possible first run with zero calibration wait? Set refresh_ratio = -1 to skip SEA and use the plain window schedule from frame one.

Categorysampling

Inputs (18)

Name	Type	Default	Description
model	MODEL		The model used for denoising the input latent.
seed	INT	00–18446744073709550000	—
steps	INT	281–10000	—
cfg	FLOAT	4.00–100	—
sampler_name	COMBO		45 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +39
scheduler	COMBO		9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3
positive	CONDITIONING		—
negative	CONDITIONING		—
latent_image	LATENT		—
denoise	FLOAT	1.000–1	—
quality_tags	STRING	highres, best quality, score_7	Quality tags to steer generation toward via modulation.
quality_neg	STRING	score_1, score_2, score_3, worst quality, lowres, old, bad hands, bad anatomy	Quality-negative baseline for the mod-guidance steering axis (delta = proj(quality_tags) − proj(quality_neg)). Leave EMPTY to reuse the CFG negative (legacy behavior). Set a clean counter-pole (e.g. 'worst quality, score_1') to decouple the quality axis from the broad CFG negative, which is anti-correlated with the intended quality direction. Does NOT change the CFG negative itself.
mod_w_profile	COMBO	step_i8_skip27	Per-block guidance schedule preset. 'off' disables modulation guidance entirely (no adapter download, no extra hook). 'step_i8_skip27' (default) protects early tonal-DC blocks 0–7 and the final compensation block 27, applying w=3 to blocks 8–26 — best overall quality but can occasionally show minor anatomy drift on drift-prone LoRAs. 'step_i14' is the SAFE option: steers only from block 14 onward, reliably stays inside the trained manifold at the cost of a slightly less expressive result. 'uniform_w3' recovers pre-0413 behavior (not recommended — prone to pink-collapse).
refresh_ratio	FLOAT	0.00-1–1	SEA scheduling dial. -1 = SEA off (plain growing-window schedule; accelerates from the first run, no calibration). 0 = SEA auto (match the window schedule's refresh fraction at this step count — same speed, smarter step placement). >0 = explicit refresh ratio (lower = faster, less faithful). The first run at each (resolution / steps / cfg / refresh_ratio) does a one-time full-compute calibration pass, then caches δ to the ComfyUI user dir for later runs.
adaptive_smc_alpha	FLOAT	0.100–1	α-adaptive Sliding-Mode Control CFG gain. 0 disables (vanilla CFG combine). 0.2 = production default — k_t := α·mean(\|v_cond − v_uncond\|) per step keeps the bang-bang correction in-band across CFG/σ/sample (paper's fixed k=0.1 was ~14× off on Anima at CFG=4). Recovers detail (fingers, eyes, text); outputs run slightly darker. Auto-disabled when CFG=1.
xattn_boost	FLOAT	1.001–3	Front-loaded cross-attn boost λ. 1.0 = off (exact identity). Scales every block's cross-attn residual on the conditional forward at high σ (the plan-writing window where text drive lives), strengthening weak-tag adherence and relation/attribute bindings without changing the render style. Norm-matched by default (renorm 'img' ρ0.5 — tunable on the Advanced node) so the boosted state stays on the trained norm shell instead of burning saturation. ~1.5 recommended; up to 3.0 for stubborn tags (higher λ can amplify unwanted caption tags like framing/crop priors). Boosts only actual forwards; forecast steps extrapolate from the boosted features. Composes with SMC-CFG / CFG++ / FSG / mod-guidance.
fsg	BOOLEAN	false	Foresight Guidance toward the golden path (one switch = the validated production stack: CFG++ λ=1.5 substrate + FSG band [0.59,0.75], K=3, on the 1024 tier @ ~28 steps). Needs CFG != 1. Because CFG++ replaces the cond/uncond combine, turning this ON disables SMC-CFG (they are mutually exclusive). Adds ~3·K forwards per in-band step. Band/K/Δσ/γ and the CFG++ λ are tunable on the Advanced node; re-tune if you change steps/resolution (the band moves).
clipopt	CLIP		CLIP encoder for encoding positive quality tags.

Outputs (1)

Name	Type	Description
LATENT	LATENT	—