SonarPerDimNoise

SonarPerDimNoise calls your noise generator multiple times along a chosen dimension instead of once for the whole latent. That sounds academic until you hit the actual use case: video. If you want each frame to get its own independent noise draw - rather than one noise tensor shared across frames - this is the node that gives it to you.

What it's actually for

The pack's other noise nodes generate one noise tensor sized to the latent. For a 5D video latent, that means the frames share noise structure unless something splits them up. SonarPerDimNoise walks along a dimension (default 0, which is the batch/frame axis) and calls the sampler per chunk, so each frame's noise is a fresh draw. The README's own example - "moving slices of 3D Perlin noise into the batch dimension" - is the same idea applied to animation: slice a volumetric noise field and let each frame use its own slice.

The inputs

• custom_noise - the noise chain being called per slice.
• dim - which dimension to split on. 0 is batch/frames and the safe default. The tooltip warns plainly: be careful with dims above 1, because those are spatial, and you might end up calling a slow noise sampler hundreds of times.
• chunk_size - how many items per call. With dim=0, chunk_size=2, and a batch of 4, the sampler gets called twice - the first call covers the first two items, the second call the last two. This is your control over the tradeoff between "one call, shared noise" (large chunk) and "fully independent" (chunk 1).
• shrink_dim - when enabled, the reference latent is chunk-sized in that dimension, so each call generates only the slice it needs. Faster and more memory-efficient, but some noise types don't work well when the target is smaller. When disabled, noise is generated at full size and sliced after, which is safer but slower.
• normalize_noise (source normalized initially) and normalize (default/forced/disabled) round it out, plus factor and rescale.

Output is SONAR_CUSTOM_NOISE like everything else in the pack.

The trap

dim above 1 is where people get burned. Set dim=2 (a spatial axis) with a slow noise type and you've just created a loop that runs for every row - that's the "hundreds of calls" the tooltip warns about, and it can turn a 5-second gen into a minutes-long one. The tooltip's advice is the right instinct: if you're not deliberately doing something spatial, keep dim at 0. Also check shrink_dim when you do venture into spatial dims - it's the difference between a fast experiment and a long one.

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. This node's whole reason to exist is video latents, so if you're using it, you're already in the pack's more advanced territory - and you'll want SonarCustomNoiseParameters nearby with frames_to_channels for noise types that can't handle 5D latents directly.