KSampler (Spectrum + Mod Guidance Advanced)

This is the node you reach for once the plain KSampler (Spectrum) is working and you want to actually tune it - or once you need the one feature the simple node doesn't have at all: DCW. Same idea underneath (Chebyshev-forecasted block caching to skip most of Anima's transformer compute), but every preset is unwound into raw parameters, and the full correction stack - DCW, SMC-CFG, CFG++, Foresight Guidance, cross-attn boost - lives here in its tunable form instead of behind a single switch.

What's different from the simple node

The core caching mechanism is the same training-free trick this whole pack is built on - the same family as TeaCache-style accelerators, just tuned specifically for Anima's DiT. Where it diverges:

• No SEA / refresh_ratio. This node uses the plain growing-window schedule directly: window_size (starting cache window), flex_window (how fast the window grows after each real forward - raise this for more speed), warmup_steps (steps forced to full compute up front). blend_w, cheby_degree, and ridge_lambda are the forecaster's own math knobs - most people never touch these three.
• Modulation guidance is exposed raw. Instead of a mod_w_profile preset, you get mod_w (peak steering strength), mod_start_layer / mod_end_layer (which transformer blocks receive the delta - default 8 to 27, protecting Anima's early tonal-DC blocks and its final compensation block), mod_taper / mod_taper_scale (soften the last few steered blocks), and mod_final_w (steering at the output head, 0 by default - leave it there unless a specific LoRA needs it). There's also an adapter dropdown for swapping in a custom pooled_text_proj beyond the auto-downloaded default.
• DCW lives here, full stop. dcw_mode (off / manual / auto) is a post-step correction for flow-matching's SNR-t bias - it nudges each step's output toward (or away from) the post-CFG prediction, restricted by default to the Haar low-frequency subband (dcw_band_mask = LL), which the pack's own testing found strictly better than correcting the full spectrum. manual mode uses dcw_lambda directly (+0.01 is tuned for CFG ≥ ~2 at non-square aspects; flip to roughly -0.015 at CFG 1 / 1024², since the bias direction reverses there). auto predicts a per-prompt λ via a small auto-downloaded fusion head and ignores dcw_lambda - and conveniently it reuses this node's own clip and positive inputs, so there's nothing extra to wire.
• CFG++ and Foresight Guidance are separate controls here. cfgpp_lambda (0 = off, 1.5 is the validated point) replaces the plain CFG combine with a σ-scheduled one. fsg then runs pre-step latent calibration on top of it via forward-backward fixed-point iterations (fsg_k, default 3, each iteration costs roughly three extra DiT forwards) inside a σ band (fsg_band_lo/fsg_band_hi, tuned for the 1024px / 28-step case - re-tune if you change either).

The trap worth knowing before you flip switches

cfgpp_lambda and adaptive_smc_alpha (SMC-CFG) are mutually exclusive - both replace the same CFG-combine hook, and SMC wins if you leave both nonzero, so your CFG++ setting silently does nothing. If you want the Foresight Guidance path, set adaptive_smc_alpha = 0 and cfgpp_lambda = 1.5 together - that's the pack's own validated pairing. And fsg needs cfg != 1; on an Anima Turbo-style low-CFG setup it has nothing to calibrate against. One more from the tooltip worth respecting: keep fsg_gamma near your CFG scale (Anima's default is 4) - pushing it up toward CFG++'s effective in-band weight (~11) makes the calibration diverge instead of converge.

What a beginner should actually set

Start with window_size/flex_window at defaults, dcw_mode = off, adaptive_smc_alpha at its default 0.1 for a mild detail bump, and everything FSG-related left alone. Get a baseline you like, then start turning on DCW or FSG one at a time - this node has enough interacting switches that changing three things at once makes it hard to tell what helped.

Installing it

ComfyUI Manager - search SpectrumKSampler, install, restart. Manual:

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

Restart. Mod guidance's ~12MB adapter and DCW's auto-calibrator both auto-download on first use (to models/anima_mod_guidance/ and models/anima_dcw_calibrator/ respectively) - expect a short pause the first time you flip either on.

Common issues & troubleshooting

CFG++ isn't doing anything. Check adaptive_smc_alpha - if it's nonzero, SMC-CFG has already claimed the CFG-combine hook and your cfgpp_lambda is being ignored.

FSG produces no visible change. It's inert at cfg = 1, and fsg_k = 0 disables it outright even if the boolean is on.

Output looks noticeably darker with SMC-CFG on. That's expected - the α-adaptive combine clamps a small consistently-signed brightness component in Anima's velocity prediction along with the noise. Lower adaptive_smc_alpha or smc_cfg_lambda if the darkening bothers you more than the detail recovery is worth.

This is Anima-tuned, not universal. Every correction here - mod guidance, DCW, SMC-CFG's tuned constants - was derived from Anima's own training pipeline. It may run on other DiT checkpoints with compatible sampler internals, but the pack doesn't promise quality there.