Latent Channel Nonlinear Transform

Latent Channel Nonlinear Transform is where the Skoogeer-Noise pack gets weird on purpose. Where its sibling linear transform rotates and permutes channels smoothly, this one is a bank of eight destructive-ish operations - gating, quantization, hard and soft clipping, and three flavors of dropout - applied to your latent's channels. The results range from "posterized art print" to "the VAE ate a bad mushroom." If you're building glitch aesthetics, datamosh-adjacent looks, or just experimenting with what latent space can be pushed into, this is the node.

The honest framing: these are not fixes. They're disturbances. Each operation mangles channels in a specific way, and the fun (and the usefulness) is in the controlled mangle.

How it works

You pick an operation and it does its thing on selected channels:

• gate_multiply / gate_add - build a per-pixel gate from a random projection and multiply (or add) it into channels. This creates "alive" glitches that track existing image structure instead of blanket noise. gate_strength sets the gate's steepness, beta its strength, blur_radius smooths the gate map.
• quantize - rounds values to a fixed quantize_step, giving posterized, flat, vector-art texture.
• clip_hard / clip_soft - clamp values at clip_threshold. Hard = abrupt plastic flattening; soft (tanh-based) = gentler, dreamier compression.
• dropout_zero / dropout_noise / dropout_swap - zero out selected channels, replace them with matched noise, or swap them with other random channels. Zeroing is where you get "missing feature" hallucinations and sudden structure gaps.

Because these hit the latent, not the pixels, the artifacts track what the model is actually representing - one channel might carry a face's structure, and knocking it out makes the face dissolve while the rest of the image stands. That's the "latent" magic and the reason the results look unlike any pixel-space filter.

The controls that keep it sane

• selection_mode - all, random, top_variance, top_roughness (pick the channels carrying the most texture/detail), or explicit indices via selection_indices (comma-separated). This is your surgical control: glitch the texture channels but leave the structural ones alone.
• mix - blend between original (0) and transformed (1). Your emergency brake. Almost every "why is this ruined" is fixed by dropping mix to 0.3.
• match_stats - re-normalize each channel's mean/std after the transform. It stabilizes the latent so the sampler doesn't cascade the damage into something unrelated.

Everything else (seed, selection_fraction/count/order) follows the pack's common conventions.

Installing it

It's part of Skoogeer-Noise. ComfyUI Manager: search "Skoogeer-Noise", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/ttulttul/Skoogeer-Noise
# restart ComfyUI

Standard deps, no downloads.

Gotchas

• Start with quantize or clip_soft. They're the least likely to destroy the image and easiest to reason about. Save dropout_zero for when you're chasing pure chaos.
• Dropout ops can nuke a whole image fast. Even at selection_fraction 0.1, zeroing the wrong channel (the structural one) collapses the composition. Use top_roughness + lowest order to target fine detail instead of structure.
• Turn on match_stats before you panic. A lot of "it's ruined" is actually the channel distribution drifting out of the range the sampler expects; matching stats fixes that without changing the glitch character.
• Pair it with LatentChannelStatsPreview from the same pack - you want to see which channels are about to be mutated before you drop them.