SonarAdvancedDistroNoise

Gaussian noise is the boring default because the math says it's right - but "right" and "interesting" are different goals, and SonarAdvancedDistroNoise exists for the second one. It replaces the sampler's noise with samples drawn from any of 26 probability distributions that PyTorch's torch.distributions provides: uniform, beta, cauchy, exponential, gamma, gumbel, laplacian, poisson, geometric, studentt, weibull, vonmises, and more. Heavy-tailed distributions like Cauchy or Laplacian throw in occasional extreme values that give generations a different texture than smooth gaussian; discrete ones like Poisson or geometric produce grainy, quantized patterns. If you've ever wondered what your latent would look like born from a Gumbel distribution - well, now you can find out.

It's a pick-one node, which is the good news. The distribution dropdown selects your flavor, and the rest of the inputs are organized so you only touch the ones relevant to your choice:

• factor and rescale - the pack-wide strength controls. rescale non-zero makes this item plus its chained items scale to add up to that value.
• quantile_norm - default 0.85, and you should probably leave it near there. Distributions with heavy tails will otherwise blow out the noise range, and this clips outliers to the given quantile. Set to 1.0 or 0.0 to disable. (Experimental bonus: negative values treat values closest to zero as the outliers instead.)
• quantile_norm_mode - what dimensions the normalization considers; batch (the default) flattens first, the nonflat_* modes can inject strong row/column influence.
• result_index - some distributions return a batch of samples; this picks which one to keep. Negative indexes count from the end.
• Then the long tail: normal_mean/normal_std for normal, cauchy_median/cauchy_sigma, gamma_concentration/gamma_rate, poisson_rate, uniform_low/uniform_high, beta_concentration0/1, and on down the list. Only the params for your selected distribution actually matter; the rest sit idle.

One heads-up: lots of these parameter fields are strings rather than floats (e.g. "0.5 0.5" for multi-valued ones like dirichlet_concentration), so when you paste a value from the docs, keep the quotes.

Output is SONAR_CUSTOM_NOISE, chained through the optional sonar_custom_noise_opt input, and consumed wherever a custom noise input is accepted - a Sonar sampler, NoisyLatentLike, or SamplerConfigOverride. Install is the standard one-liner with no model files:

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

The trap is the same one the whole pack sets: ancestral/SDE samplers are the ones that inject noise, so that's where your exotic distribution actually appears - pair this with Euler a or DPM++ SDE. And combine it with the quantile_norm default before judging a distribution, because a raw Cauchy sample without clipping will just nuke the latent. Start with laplacian or cauchy, keep the quantile on, and see what texture does to your output before you go full vonmises.