DCW(+a)

If you've ever cranked CFG up to force the prompt through and instead got burnt, oversaturated garbage, that's not you doing it wrong - that's the tool being blunt. Standard CFG multiplies the whole cond − uncond error by a single scalar, all frequencies, all steps, at once. DCWModelPatch takes that one knob and cracks it open into three training-free corrections, each with its own toggle:

• DCW (Differential Correction in Wavelet domain) fixes the SNR-t bias - the tendency of a model's predicted clean image (x0_pred) to drift as noise accumulates.
• CWM (CFG Wavelet Mixing) applies different guidance strength to low vs high frequencies, and to early vs late denoising steps.
• SMC (Sliding Mode Control CFG) stabilizes the guidance error itself so high-CFG sampling stops oscillating.

No training, no model downloads, no API, no external dependencies. It's a single model_patches node you drop between your checkpoint (or LoRA chain) and KSampler.

How it actually works

Think of it as three hooks at different spots in the sampler:

• DCW lives in sampler_post_cfg_function - it runs after CFG has combined the conditioned and unconditioned predictions. It decomposes x0_pred into wavelet bands and nudges each band toward or away from the current noisy latent, with timing: low-frequency (composition/color) corrections peak in early steps, high-frequency (edge/texture) corrections peak late. Because it runs after CFG, it's independent of your CFG scale and sampler choice.
• CWM replaces the scalar CFG formula inside sampler_cfg_function. It splits the guidance error e = cond − uncond into wavelet bands and scales each one separately - "follow the prompt's composition hard early, leave the texture alone late," impossible with one CFG value.
• SMC runs inside the same hook, before CWM. It's straight control theory: a sliding surface built from the current and previous guidance error, corrected by a normalized switching term so the correction energy stays bounded. Net effect: the guidance trajectory stops wobbling at high CFG. Set smc_preset to Auto and it reads the model class name and applies paper-tested (λ, k) values for your model family.

Honest caveat: the author cites the DCW SNR-t bias paper and the CFG-Ctrl (SMC) CVPR work, but this is a new, niche pack with a thin community footprint. The README's numbers are the author's own tuning tables, not crowdsourced gospel - treat them as starting points and A/B test (right-click → Bypass does that for free).

The inputs that matter

Nine knobs sounds like a lot, but defaults are sane and you'll realistically touch three:

• lambda_l / lambda_h - DCW's low- and high-frequency correction strength. Start 0.05 / 0.010 for SD/SDXL-class models, roughly double for flow models. Set both to 0 to disable DCW.
• alpha_l / alpha_h - CWM's per-band CFG boost, expressed relative to your base CFG scale. 0 means identical to standard CFG, which is a nice safety property. Start 0.1–0.3 low-freq and 0.1–0.2 high-freq.
• smc_preset - Off (default), Auto (recommended), named presets per model family (SDXL, Flux, SD3/SD3.5, Qwen-Image, Cosmos/Wan…), or Custom, which unlocks the smc_lambda and smc_k sliders. Leave those alone until Auto disappoints you.

dcw_enabled and cwm_enabled are your A/B toggles. The single output is model (MODEL), which wires straight into KSampler's model port.

Installing it

Through ComfyUI Manager, search "ComfyUI-DCW". Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/namemechan/ComfyUI-DCW

Then restart ComfyUI. That's the whole install - no requirements.txt, no model files, pure PyTorch. Zero-dependency installs are rare in this ecosystem; enjoy it.

Where people get burned

The hook conflict, first and foremost. CWM and SMC take over sampler_cfg_function. If another node upstream has already registered that hook - some CFG-refiner and guidance-modifier nodes do - CWM and SMC are silently skipped with only a console warning. DCW uses a different hook (sampler_post_cfg_function) and always chains safely. If CWM/SMC appear to do nothing, check the console, then move this node earlier or disable the conflicting hook.

A few built-in guards worth knowing: odd-resolution latents are reflect-padded, corrected, and cropped back; fp8 tensors get upcast to bf16 for the math and restored; SMC runs in fp32 specifically to dodge the NaN/black-image failure you can hit when combining fp16 patches like SageAttention with this kind of math.

Flow models are the special case. Flux, Cosmos/Anima, and Wan cap their σ scale around 0.5, so corrections land at half strength - double your lambda/alpha values there. Also know the effect is strongest at low step counts (10–20 steps); at 30+ it gets subtle. And one quiet detail: with CWM or SMC active, the node forces ComfyUI to compute the unconditional pass even at CFG 1 - it needs that prediction to steer with, so your "CFG 1 = free speed" trick won't apply while it's on. If you're already happy with your sampler and CFG, this is strictly optional; if you're fighting oversaturation, blur, or prompt drift, it's worth an afternoon of A/B before you go shopping for a new sampler.