SonarResizedNoise

Noise doesn't have to be generated at the exact size of your latent. SonarResizedNoise lets you take one noise item and resize it to a different width and height before it hits the sampler - which matters more than it sounds, because the size of your noise changes its character. Big noise is coarse and blobby; small noise is fine and grainy. This node is how you dial that in independently of your image resolution.

The author is refreshingly upfront about the use case: "Scaling noise is something you'd pretty much only use to create weird effects. For normal workflows, leave this on crop." So treat it as a deliberate tool, not a default.

The inputs that matter

• custom_noise - the noise chain to resize.
• width / height (default 1152) - the target size in latent pixels. The tooltip carries an important hint: "This should almost always be set to a higher value than the image you're actually sampling." Oversized noise that gets cropped down keeps more of its structure; undersized noise upscaled to fill the latent smears and looks mushy. Note one latent pixel ≈ eight image pixels for most models.
• downscale_strategy (crop / scale) - what happens when the source is bigger than the target. crop is the normal default.
• crop_mode - nine positions (center, corners, edges) plus crop_offset_horizontal / crop_offset_vertical to nudge the window around. Offsets are clamped into the available space, so a top_right crop with a positive horizontal offset does nothing - it's already as far right as it can go.
• upscale_mode / downscale_mode - which resampling method to use (bilinear, nearest, bicubic, area, etc.). nearest-exact is the default; area and bislerp are worth a try when you want softer results.
• initial_reference - when the sampler's reference latent isn't the same size as your requested width/height, whether to crop or scale to match. Default prefer_crop is fine; the note says cropping usually isn't possible anyway.
• factor and normalize - strength and the standard rebalance control.

Output is a SONAR_CUSTOM_NOISE, feeding a Sonar sampler, SamplerConfigOverride, or NoisyLatentLike like any other chain item.

A practical pattern

Resized noise shines in the "weird effects" department the author mentions: generate a small, low-resolution noise (say 256x256 latent), upscale it with area or bislerp to your full latent size, and the sampling starts from soft, structured noise instead of white static - a low-frequency bias that can push compositions toward smoother, more abstract shapes. Combined with the pack's scheduling node, you can use it for the first N percent of sampling and fall back to Gaussian after, which is exactly how the pack's own examples mix extreme noise types.

Installing

Part of ComfyUI-sonar - install once via ComfyUI Manager (search "ComfyUI-sonar") or:

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

Restart ComfyUI, find it under advanced/noise. No extra dependencies or model downloads. And remember the pack-level caveat: inputs and internals have shifted between versions, so if a saved workflow breaks after an update, it's usually the pack changing, not you.