Noise Expose

The whole problem with diffusion noise residue is that you can't see it. It's a 1–3-level-per-channel micro-noise coat that survives VAE decode, sits invisible under the final image, and is exactly what AI-detection tools latch onto. Noise Expose is the X-ray: it remaps your image so that invisible noise becomes loud, obvious color. Feed it any image, look at the output, and you'll know in one glance whether your workflow is shipping residue.

What it's for

This is a diagnostic node, not a fixer. Its two real jobs:

1. Before cleaning - see whether your image carries detectable residue in the first place.
2. After cleaning - verify the clean-up pass actually removed it.

It's the companion piece to the Latent Residue Cleaner in the same pack, and the cleaner's own debug mode uses this exact effect internally. You can also use it to compare two samplers or checkpoints and see which one leaves a cleaner image behind - the noise pattern shows up as color, so differences you'd never spot by eye become obvious side by side.

How it works

It's all color-space trickery, implemented in pure torch with no OpenCV. The image goes RGB → HSV, and each mode pushes noise into a visible channel:

• Dual Expose (default) - runs the transform on the image and on its inverse, then combines them. Catches noise in both bright and dark pixels. This is the one you'll live in.
• Value → Saturation (Original) - copies brightness into saturation. The baseline, but dark pixels stay hidden, so dark-region noise slips through.
• Triple Expose - adds a mid-shifted pass on top, so gray-pixel noise shows too. Use it when you're suspicious and Dual Expose comes back clean.
• Edge Noise - a 3×3 high-pass filter (original minus local mean, amplified ×5). Shows fine-grain noise structure regardless of brightness. Less "pretty", more "lab instrument".
• Hue Noise - hue channel at full saturation. It's the view that exposes the garbage-hue problem on near-white/near-black pixels that the cleaner's Hue Snap stage fixes.

The whole thing is a remap - a couple of tensor ops per pixel, no filtering that takes minutes. It runs on CUDA if you have it, CPU otherwise, and it's cheap either way.

Inputs and output

Two inputs total: image (any IMAGE tensor - a VAE decode output, a Load Image result, whatever you're auditing) and mode (the five options above). One output, image, which you can chain, save, or compare against another run. It's not an output node, so it won't pop into your save list on its own - put a Save Image after it if you want the exposed version on disk. Zoom in on the saved version; that's where the detail actually reads.

Installing it

Noise Expose ships in the Latent Residue Cleaner pack, so installing that gives you both nodes. In ComfyUI Manager, search Latent Residue Cleaner, or:

cd ComfyUI/custom_nodes
git clone https://github.com/L33chKing/ComfyUI_LatentResidueCleaner

Restart ComfyUI after. There's no requirements.txt, no models to download - the whole pack is two Python files and torch.

Troubleshooting

• The output looks loud and psychedelic. That's the point. You're looking at a visualization of noise, not a render. If it looks calm, your image is clean.
• Dual Expose comes back almost black or flat. Great news: you have very little residue and probably don't need the cleaner at all.
• Comparing before/after. Run Noise Expose on the same image before and after scrubbing, save both, and flip between them. The noise color should visibly drain away after the clean - if it doesn't, your cleaner settings aren't doing anything.

One honest note: this pack is new enough that there's no community folklore around Noise Expose yet - no shared "use Triple Expose for skin" wisdom to lean on. You'll be your own lab tech for a bit. That's fine; the modes are self-explanatory enough that five minutes of poking tells you everything.