Apply DCW (Wavelet Patch)

One node, and it's a model patch

Apply DCW (Wavelet Patch) doesn't generate anything by itself. It's a model_function_wrapper - you drop it into the purple MODEL slot between your checkpoint loader and your KSampler, and it quietly rewrites how the model denoises on every step. Think of it like a LoRA that isn't trained: it changes behavior with zero retraining, and it changes nothing until you tell it to.

DCW stands for Dynamic Consistency Weighting, a technique from NVIDIA's AMAP research group (AMAP-ML/DCW, paper "Elucidating the SNR-t Bias of Diffusion Probabilistic Models"). Its whole pitch is few-step generation that doesn't fall apart. Anyone who's run LCM or Turbo at 4–8 steps knows the failure mode: the image comes out fast but colors go muddy, composition wobbles, details smear. That's not just "low quality" - it's low-frequency error piling up step over step. DCW attacks exactly that.

How it actually works

Here's the part that's genuinely clever. On each denoising step, the node:

1. Runs your model normally to get its prediction, then reconstructs the "clean image estimate" (x0) from the noisy latent and the current timestep.
2. Splits both the noisy input and that estimate into frequency bands with a single-level wavelet transform - low frequency (color, composition, big shapes) and high frequency (texture, edges).
3. Rebuilds the estimate, but only the low-frequency band gets corrected, blended toward the input by an amount that scales with how far along the denoise you are: low = estimate + strength × (input_low − estimate_low) × t/1000. Early steps lean hard on the input's structure; late steps trust the model.
4. Converts the fixed estimate back into a prediction ((input − x0_corrected) / timestep) so the sampler's math still holds, and hands that to the next step.

High-frequency detail is preserved untouched, which is the whole trick - you fix the color/composition drift without sanding off the texture. It also handles 5D video latents (it reshapes to 4D, processes, reshapes back) and pads odd-sized latents so the wavelet transform doesn't choke.

The inputs that matter

There are only three, and one of them you'll leave alone:

• model (required) - your MODEL. Feed it the output of your checkpoint/UNET loader.
• strength (FLOAT, 0–4, default 1.0) - the correction intensity. The README's guidance: 0.3 for ultra-fast 8-step generation, 1.0 for medium, 2.0+ is heavy and can artifact at very low steps. 0.0 is a clean passthrough, which makes A/B testing trivial.
• wavelet (haar / db2 / db3, default haar) - the wavelet basis. haar is fastest and a good default; db2 preserves a bit more detail; db3 is maximum detail but slower. In practice: start with haar, switch if you're chasing fine texture.

The single output is a patched MODEL - wire it straight into your sampler.

Installing it

ComfyUI Manager can find it if you search for "DCW", or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/DawnW0lf/ComfyUI-DCW-Diffusion-Color-Wavelets-Node.git
cd ComfyUI-DCW-Diffusion-Color-Wavelets-Node
pip install -r requirements.txt

Then restart ComfyUI. One README trap: its install section says cd comfyui-dcw, but no such directory exists - the folder is the repo name above. Don't be alarmed when cd fails.

Gotchas worth knowing

• The dependency that bites: pytorch_wavelets. It's a compiled package built against your installed PyTorch, and it's the most likely thing to fail during install (PyWavelets, the other dep, is a boring wheel). If pip's build errors out, the usual fix is building it from source against your torch version.
• This is brand-new, barely-adopted code. As of writing there's essentially no community chatter about it - v1.0.0, one node, one author. Treat it as research code: A/B it against your normal workflow before you trust it.
• Silent failure mode is silent. If something goes wrong mid-generation, the wrapper catches the exception, prints [DCW FATAL ERROR] to your ComfyUI console, and returns the unpatched output. So if your results look identical and you're not watching the console, it might be failing every step without telling you.
• The license is non-commercial. CC-BY-NC-SA 4.0, inherited from the NVIDIA research code. Fine for personal use and experiments; a hard no for selling what you make with it unless you get permission.

Where does it shine? Paired with a distilled/fast model at low step counts - that's the regime where low-frequency drift is worst and DCW has the most to fix. At 20+ steps on a full model you probably won't notice it, because you didn't have the problem in the first place.