SUNoiseLatent

Most of this pack is about the noise that ancestral samplers inject during denoising. This node is about the noise your image starts from. SUNoiseLatent generates the initial latent noise itself - scaled uniform noise instead of the standard Gaussian - and hands it to the sampler as a NOISE object. If you've ever seen the "RandomNoise" node in ComfyUI and wanted to build your own noise instead of relying on the built-in distribution, this is that, with an on/off switch for structure.

What it actually does

It creates a noise tensor shaped like the latent image - same batch, channels, height and width - drawn from a uniform distribution scaled to [-scale, +scale], and returns it as a NOISE object. That object plugs into the noise input on KSampler (the extra port on the right side, not the seed widget). Wire it there and your run starts from this noise instead of the default randn.

The two noise_type options are the same split as the sampler nodes:

• standard - pure uniform noise per pixel. The argument for it is the SUNoise one: Gaussian's unbounded tails occasionally dump extreme values into the latent, and a bounded distribution avoids that entirely.
• multires - uniform noise generated at full and half resolution, the half-res layer upscaled and summed with the full-res one, then rescaled back to [-scale, +scale]. The result has spatial structure - low-frequency blobs rather than pure static. This is the option that actually changes composition-level behavior, because structured initial noise biases what the model does with large-scale shapes.

The scale slider (default 1.0, negative values allowed) controls the amplitude: how strong the initial perturbation is relative to the latent. For img2img-style starts you'll want it small; for full text-to-image runs, 1.0 is the sane starting point.

The seed (noise_seed, default 0) is where reproducibility comes from. The noise generator is deterministic - same seed, same noise - and unlike the in-loop sampler noise, it's derived directly from this widget rather than borrowed from your KSampler seed.

Wiring it up

KSampler
  model        → CheckpointLoader
  positive     → CLIPTextEncode
  negative     → CLIPTextEncode
  latent_image → EmptyLatentImage (or your img2img encode)
  noise        → SUNoiseLatent (NOISE out)

That noise port only appears in recent-enough ComfyUI builds, so update if you don't see it. Everything else about the KSampler behaves normally.

Install and setup

Same pack, same story - ComfyUI Manager, search "SUNoise", or:

cd ComfyUI/custom_nodes
git clone https://github.com/bvhari/ComfyUI_SUNoise

Restart. No dependencies to install, no models to download. GPL-3.0.

The honest take

This is the most underrated node in the pack, and also the most situational. multires noise is a genuinely useful tool for img2img and inpainting-heavy workflows, because structured initial noise survives the first denoising steps and steers composition - people pay good money (well, node complexity) for Perlin-style alternatives that do roughly this. But for plain text-to-image, standard SUNoise at the start is a marginal tweak; the effect mostly washes out by step 20.

Where people get burned: mixing SUNoiseLatent with a sampler node from the same pack and then wondering why the seed no longer "works." Both nodes derive their noise from seeds, but they're separate mechanisms - change one and the other's noise stays put. If you want everything seeded from one place, feed the same seed into both and treat that as your project seed.

Also worth noting: the pack has basically no community footprint - no tutorials, no "this fixed my workflow" posts. You're on your own for dialing scale and noise_type per model, which is fine because it's one slider and one dropdown. Fix the seed, try both noise types, keep what looks better.