NoisyLatentLike

ComfyUI-sonar started life as a momentum-sampler port and quietly turned into a noise toybox. If you want to see why, this is the node to start with. NoisyLatentLike generates noise shaped like a reference latent and optionally adds it to that latent - it's the generic "give me a latently-sized noise field of a specific character" node, and every custom noise chain in the pack eventually feeds into it.

For a beginner the mechanism is simple: you hand it a latent, it makes noise that matches that latent's size and channel layout, and if add_to_latent is on, it adds the noise to the latent instead of to an empty one. That second mode is the img2img trick - you're rebuilding the noisy starting state that img2img needs, but with whatever exotic noise you want instead of the usual gaussian.

The inputs that actually matter

There are eight required inputs but you'll set maybe four:

• noise_type (default gaussian) - the star of the show. Thirty-eight options: gaussian, brownian, collatz, green_test, grey, the highres_pyramid family, laplacian, the onef_* (1/f) variants, and more. Don't overthink the first time; grab brownian or green_test and see what it does to a fixed seed.
• seed - ordinary noise seed.
• latent - the reference latent everything keys off.
• add_to_latent - off by default. The tooltip's blunt advice: "Generally should be enabled for img2img workflows." In that mode you also want to connect mul_by_sigmas_opt (a SIGMAS socket) and model_opt, which scales the generated noise by the first sigma so the noise level actually matches your denoise strength.

The rest are tune-ables: multiplier scales noise strength (applied after sigma scaling), repeat_batch repeats generation for larger batches, cpu_noise (default true) decides CPU vs GPU - only matters for types that support GPU generation, basically Brownian - and normalize (default true) pins noise to 1.0 strength before scaling. Leave normalize on unless you know why you're turning it off.

There's also the custom_noise_opt socket: connect a custom noise chain here and noise_type is ignored entirely. That's how the SonarAdvanced*, SonarBlended*, and SonarChannelNoise nodes hand their output onward.

Output and wiring

Output is a LATENT. In a text-to-image workflow you can feed it into a SamplerCustom or a KSampler as the initial latent - the Sonar sampler nodes take the same noise and it behaves equivalently. In img2img you chain it with add_to_latent enabled, model and sigmas connected, and let your sampler consume the result.

Installing

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

or "ComfyUI-sonar" via ComfyUI Manager, then restart. No models, no extra requirements.

The honest caveats

The author is upfront that the noise types are best-effort - "pink" noise here may not be what a signal-processing textbook calls pink noise. It doesn't matter; the point is the character, not the taxonomy. Second, seeds aren't portable across GPU types because generation hops between CPU and GPU. And since the toybox is the whole point, spend a minute with the README's docs/base_noise_types.md - it's the only real documentation the noise types have, and it's worth more than an hour of random knob-turning.