SonarLatentOperationQuantileFilter

SonarLatentOperationQuantileFilter applies quantile normalization to a latent mid-sampling, clipping the extreme outlier values that can blow out a generation. If your images sometimes come out with harsh, blown-out highlights or crunchy artifacts that look like a few values ran away with the whole latent, this is the targeted fix: it doesn't average, it clips to the distribution.

What quantile normalization means

Normal "normalize to 1.0" squashes everything proportionally, so one wild outlier drags the whole distribution down. Quantile normalization works differently: it finds the value at a given quantile (say the 85th percentile) and treats everything beyond it as an outlier, clipping or reshaping it. Result: the bulk of your latent keeps its relative character, and the insane tail gets tamed. The tooltip example is precise - a quantile of 0.75 means outliers above the 75th percentile get handled.

The inputs that matter

• quantile - default 0.85. 1.0 or 0.0 disables the normalization entirely. Values like 0.75–0.85 are the sensible range; the right number depends on how extreme your inputs are. The experimental twist: a negative quantile treats values closest to zero as the "extreme" ones instead.
• strategy - how outliers are handled, and this is a deep one: 43 choices. The accessible headliners are clamp (pin to the boundary), scale_down, tanh, and the various sin/sigmoid shapes. zero returns zero outside the quantile range and reverse_zero keeps only the outliers - the tooltip notes those two only make sense if you're adding the result to other noise. Most of the 43 are experiments; start with clamp or tanh.
• dim - which dimensions to normalize over: global or an explicit dim. Image latents are batch/channel/row/column; video latents add a frame dim (so 2 is the frame axis on 5D latents, but the row on 4D ones - mind the shape you're actually feeding).
• flatten - default true, flattens before normalizing. Try disabling if you want per-row/column influence, though the tooltip warns that produces strong row/column bias.
• norm_factor and norm_power - the multiplier just before clipping (leave at 1) and the exponent applied after (0.5 default; below -0.3 gets strange).

Output is a LATENT_OPERATION, meant for a consumer like the pack's SonarApplyLatentOperationCFG - the node description notes this one is the operation-based sibling of the pack's standalone quantile-normalization noise node, and it applies to whatever you attach it to (denoised, uncond, and so on).

Where you'd use it

Anything where outliers are the failure mode: high-CFG generations that blow out, latent edits that spike values, or after aggressive negative-space operations. Because it runs mid-sampling via an operation consumer, you can also schedule it to only tame the early or late steps.

Install

ComfyUI Manager → search "ComfyUI-sonar" → Install → restart, or:

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

No models, no extra deps. The author's disclaimer that a lot of this pack is "janky" and best-effort is especially true for the 43 strategy variants - treat the exotic ones as experiments, not gospel.