NoiseImage_AS

NoiseImage_AS generates a pure random-noise IMAGE at whatever size you ask for. Every pixel is a random float between 0 and 1, which makes a classic static-gray field. It's a dead-simple utility with two honest use cases: testing image pipelines end to end, and feeding noise into img2img-style workflows as a starting texture.

It comes from flyingshutter's As_ComfyUI_CustomNodes, a dependency-free playground pack. No models, no installs, no fuss.

How it works

The node calls torch.rand to create a (1, height, width, 3) tensor - one image, three RGB channels, every value uniform random in [0, 1). No distribution tuning, no seeded reproducibility option, just noise. Width and height step in increments of 64, defaulting to 512.

One honest detail worth knowing: the node has an idx input that looks like a seed, but the source never actually uses it. torch.rand draws from PyTorch's global RNG, and this node doesn't set or consume any seed. So idx does nothing - you'll get different noise on every run regardless of what you put in it. If you wanted reproducible noise, this isn't the node for that (and it's a good reminder this pack is a self-described playground).

The inputs

• width - image width, in steps of 64 (default 512).
• height - image height, in steps of 64 (default 512).
• idx - present, but inert in the current code.

Output: a single IMAGE of random noise.

Where you'd use it

• Pipeline smoke-testing. Stuck with a workflow that fails halfway? Feed it a noise image instead of a real load to isolate whether the problem is in the image-loading stage or the processing stages. Deterministic-ish debugging, minus the determinism.
• Texture start for img2img. Noise fed into an img2img sampler at high denoise acts as a near-total restart of an image, and some creative workflows deliberately layer noise before denoising for texture effects.
• Test data. Any node that needs an image-shaped tensor to exercise logic without waiting for a generation.

For anything fancier - gaussian noise, perlin/texture noise, seeded noise - other packs do it properly. This one is uniform static, plain and simple.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/flyingshutter/As_ComfyUI_CustomNodes

Restart ComfyUI and it's under ASNodes (or ComfyUI Manager → "As_ComfyUI_CustomNodes").

Common issues

• idx doesn't control anything. If you crank it expecting reproducible noise, you'll be confused - it's a no-op in the shipped code. Different noise every run is the expected behavior.
• Stepped dimensions. Width/height snap to multiples of 64. Asking for 500 gives you a rounding to the step grid, not 500 exactly.
• Uniform vs. gaussian. This is uniform noise, which has a different character than the gaussian noise diffusion models work with internally. Don't expect it to behave like a noise scheduler's latent noise.

It's a two-line node that does exactly one thing. For a quick noise field it's fine; for reproducibility, look elsewhere.