SonarChannelNoise

A latent isn't one image - it's a stack of channels, and the model treats them as separate feature planes. SD 1.5 and SDXL latents have 4 channels; Flux and SD3 have 16. Almost every noise tool in the world generates the same noise for all of them. SonarChannelNoise is the exception: it lets you assign a different noise generator to each channel, so channel 0 might get brownian while channel 3 gets gaussian. Since each channel feeds a different part of what the model "sees," this is a surprisingly effective way to change a generation's character without touching the prompt.

The catch is in the node's wiring, and it trips everyone the first time. The sonar_custom_noise input is treated as a list, not a single item, and the list is built by chaining noise nodes through the usual optional inputs - with the item furthest from this node counting as channel 0. So the visual chain order is backwards from what intuition suggests: the noise node nearest to SonarChannelNoise maps to the last channel, the one furthest away maps to channel 0. Take a breath, read the tooltip, wire it once, and it'll click.

The other inputs:

• insufficient_channels_mode - what happens when you connect fewer noise items than the latent has channels. wrap (default) cycles back to the first item, repeat uses the last item for everything after, and zero fills remaining channels with zeros (the tooltip says generally not recommended - a zero channel means the model gets nothing on that plane).
• mix_count - how many of the connected items get selected each time noise is generated. At 1 (default) each channel is deterministic from its assigned item; higher values pick a subset per call, which adds variation.
• normalize and factor - the standard strength controls.

Output is SONAR_CUSTOM_NOISE, consumed wherever custom noise is accepted. Install is the pack's usual one-liner, no model files:

cd ComfyUI/custom_nodes
git clone https://github.com/blepping/ComfyUI-sonar

Where this shines and where it bites: the power is that you can give each channel its own texture; the frustration is that you can't see the channels, so tuning is empirical - change one item, rerun, compare. Start with 4-channel SDXL, two noise types, wrap mode, and see whether a specific channel assignment does something you like before adding complexity. And because channel assignments interact with the sampler, this is definitely a node for ancestral/SDE samplers or the sonar momentum ones - converging samplers won't inject the per-channel noise at all. It's fiddly, it's invisible-to-the-eye work, and for that subset of generations it's the only node in the pack that does what it does.