WAS Latent Detail Boost

Sharpening has a fixed recipe. Blur a copy, subtract it to isolate the high-frequency detail, add that detail back scaled by an amount. That's unsharp masking, it's the only thing "sharpen" has ever meant, and the reason it looks bad when you crank it is that the same amount gets added everywhere - including on both sides of every edge, where it produces the bright rim and the dark outline people call haloing.

Do that in latent space and it's worse, because the latent's "high frequency" isn't texture, it's encoding noise. Sharpen a latent directly and you get dark outlines and an embossed, metallic look that no prompt asked for.

This node is a band-pass with three guards bolted on, and each one is aimed at a specific failure of plain sharpening.

What it does

It isolates one band of detail rather than "all the sharp stuff": sigma_small is the fine end (0.6 in latent blocks by default) and sigma_large the coarse end (1.4), and the gap between them is what gets boosted. A narrow band like 0.6 against 1.4 reads as texture; widen it and you start treating larger features. Set sigma_small to 0.0 and you reach the very finest detail - including noise, which is why 0.6 is a better default than 0.0. The two are swapped if you enter them the wrong way round, so don't worry about the order.

gain is how much of that isolated detail is added back. 0.35 is a gentle lift; above about 1.0 the texture starts to dominate the picture.

Then the guards.

The contrast limit is the interesting one. The detail is levelled against the local amount of contrast before it's added, measured over rms_sigma (1.2 by default) - so a smooth sky gains as much texture as a busy tree, instead of being left behind by a process calibrated for the busy areas. rms_floor puts a floor under that measurement so genuinely flat regions aren't amplified into grain; raise it toward 0.2 if skies come out sandy, lower it toward 0.0 to treat flat areas as hard as everything else. limit (1.25) squashes the detail before it lands, which is what stops a strong edge ringing - 4.0 flattens the strongest detail to a uniform level, 0.1 barely limits at all. Setting rms_sigma to 0.0 measures over the whole frame instead and gives you plain, uniform sharpening - the behaviour you came here to avoid, kept around as a comparison.

Edge protection holds the effect back on strong boundaries. edge_protect 0.45 halves it there, 1.0 leaves boundaries untouched, 0.0 disables the pass (and skips finding edges entirely). edge_sigma blurs before edges are detected, so fine texture isn't mistaken for a boundary and protected from the very treatment it wants; edge_threshold is how strong a boundary has to be, measured against the strongest in the picture; edge_softness is how gradually protection fades in - 0.1, because a hard-edged protected strip is visible as a strip.

Video, and seeing the masks

Video latents with a time axis are handled frame by frame, so a sequence gets the same treatment frame to frame instead of flickering.

Outputs: latent, plus mask and mask_preview so you can see which map the node is working from rather than guessing, and preview_mask_scale (8 by default, matching what most VAEs decode to) sets how many preview pixels each latent block becomes. preview_mode picks which of the two maps the node shows - edge_mask or detail_mask. The masks are the underrated part: when the boost isn't doing anything, the mask tells you whether it's the threshold or your image.

Where it fits

Between a latent upscale and the second sampler in a hires pass. The latent upscale interpolates and softens; this puts some detail back before the sampler re-resolves, which lets you sit at a slightly lower denoise and keep more of the composition you liked. Also at the tail of a chained sequence run, where contrast and detail have accumulated - though there, the more correct fix is turning the drift down rather than sharpening on top of it.

Installing it

It ships in WAS Node Suite v3 - MIT, by WASasquatch, 468 nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui.git

ComfyUI Manager, search WAS Node Suite v3, is the recommended route. ComfyUI 0.14.0+ and Python 3.10+; no pip packages, no downloads, no build. Config and folders land in <ComfyUI user dir>/was-node-suite/ on first start, which is why that first boot is a second slower and an update repeats it. Latent upscalers and this node are among the 27 in the extras feature group, on by default - check features.extras in config.yaml if it's missing from your menu. And if an older guide tells you to install scikit-image first, close the tab; that was v2.