Same Color Noise 🍌

Every seed you roll starts from independent random noise, which is exactly why a batch of otherwise-identical prompts can come back with wildly different overall color and tone - one warm, one cool, one blown-out, one flat. Same Color Noise is a NOISE generator built to fight that: it borrows color/tone information from a reference_latents input and biases your starting noise toward matching it, so a batch shares more of a family resemblance instead of each seed rolling completely independently.

Why there are sixteen channel toggles

ch_00 through ch_15 - sixteen boolean switches, all on by default - are the node's most distinctive feature, and they map onto the latent channel count of newer architectures. SD 1.5 and SDXL use a 4-channel latent space; the newer generation of models (Flux, SD3, and other current architectures) moved to 16 channels. This node gives you one on/off switch per channel, so you can decide which of the sixteen actually participate in the color-matching effect rather than applying it uniformly across all of them.

Practically: on a 16-channel-latent model, all sixteen toggles are meaningful. On an SD 1.5/SDXL workflow, only ch_00–ch_03 correspond to anything that exists in that model's latent space - the rest presumably do nothing there, since those channels aren't part of a 4-channel latent to begin with. That's a reasonable inference from the schema rather than something documented outright, so if you're on an older architecture, start by only touching the first four toggles and ignore the rest.

seed sets the base noise draw, and strength (-1 to 1) dials how hard the color-matching pulls toward the reference - negative values presumably push away from it rather than toward, which is worth testing deliberately if you want the opposite effect (more color divergence, not less) across a batch.

Using it

Output is NOISE - wire it into SamplerCustomAdvanced or this pack's SamplerCustomAdvancedPreview/SamplerCustomXY. It needs reference_latents connected to something - an encoded reference image, or a latent from an earlier generation you liked the tone of.

Installing it

Via ComfyUI Manager: search cgem156-ComfyUI, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/laksjdjf/cgem156-ComfyUI

Restart afterward. No model downloads, no extra dependencies - a pure noise-generation node.

Common issues & troubleshooting

Output looks discolored or broken rather than just tone-matched. Don't assume strength is the only knob - try turning channels off one at a time instead of adjusting strength first. Some channels carrying color-matched noise may interact badly with your specific model.

No visible effect at all. Confirm reference_latents is actually connected - this node needs a real reference, it doesn't generate one. If you're on an SD 1.5/SDXL workflow, also confirm you haven't accidentally left only the higher channel indices (which likely don't map to anything in a 4-channel latent) toggled on.

You're not sure this is the right tool for what you want. If your actual goal is character or subject consistency rather than overall color/tone consistency, this node is the wrong layer - look at this pack's ReferenceApply/MultipleReferenceApply nodes (self-attention-based reference injection) instead, or a proper adapter/edit-model workflow for anything identity-related.

No community track record. This is a genuinely niche corner of the pack with no README section and no discussion found anywhere - treat the channel-toggle behavior above as an informed reading of the schema, confirm it by testing on your own architecture before trusting it blindly.