SonarApplyLatentOperationCFG

Normally a latent operation happens before or after sampling - you pass a latent through a filter node, then hand it to the sampler. SonarApplyLatentOperationCFG does the weirder and more interesting thing: it wraps your MODEL so that a LATENT_OPERATION gets applied during the sampling loop, on the model's own predictions. Model in, model out, and everything downstream - KSampler, SamplerCustom, whatever - just works, because the patch rides along inside the model object.

A LATENT_OPERATION is ComfyUI's pluggable "transform this latent" interface. ComfyUI ships a few built-in ones, and this pack adds its own, SonarLatentOperationQuantileFilter, which is presumably what most people wire up here. The point of applying one mid-CFG is to reshape the denoised image at the point where it actually matters - think of it as an in-loop filter that can suppress or sculpt features while the sampler is still working, rather than after the fact.

The mode dropdown is the heart of it. The default cond_sub_uncond is what ComfyUI's own latent operations use - it operates on the classifier-free-guidance result (cond minus uncond), which is the thing CFG is actually steering. The other modes (denoised, cond, uncond, denoised_sub_uncond, etc.) operate on different internal predictions, and the tooltip flags that the non-sub_uncond ones won't work with pred_flip_mode. In short: stay on cond_sub_uncond until you have a reason not to.

The inputs a beginner actually sets:

• start_sigma / end_sigma - the sigma range where the effect is active. start_sigma at -1 means "the model's max sigma" (i.e. from the very beginning); end_sigma at 0 means "until the end." Narrow this to apply the operation only in the middle of sampling.
• blend_strength - how hard the operation's output mixes with the original. 1.0 = full replacement, 0.0 = nothing.
• blend_scale_mode - scales the blend over time. The recommended reverse_sampling is strong at the start of sampling and fades to 0 at the end - the tooltip's reasoning: these operations usually work better early, when structure is being established.
• pred_flip_mode - applies the operation to the noise prediction instead of the image prediction. The tooltip is disarmingly honest: no real reason it should be better, just something to try.
• operation_1 through operation_5 - the actual operations, applied in sequence. immediate_blend forces blending after each one rather than only at the end.

Output is a MODEL, so this slots between your checkpoint loader and your sampler node. The CFG math here is real - this is the same family of model-patching that ComfyUI's own _cfg_pp samplers use, where the sampler steers using the unconditional prediction. That's also why require_uncond exists: at CFG 1.0 ComfyUI skips the uncond pass entirely for speed, and an operation that needs uncond would silently do nothing unless you opt back in.

Install is the pack standard, no model downloads:

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

It's a power-user node - if you're not sure what operation you're applying or why mid-sampling beats post-processing, start with SonarLatentOperationQuantileFilter at a modest blend_strength and reverse_sampling and see whether the in-loop version does something you can't get by filtering the latent afterward. Sometimes it does, and that's the whole appeal.