NoiseVariations

A surprisingly practical node hiding in a very experimental pack. NoiseVariations takes a LATENT, adds gaussian noise scaled by a scale value, and returns the latent - that's the entire job. It's the simple, blunt sibling of the fancier bending nodes: where they intercept model layers and rewrite activations, this one just scatters noise into the latent directly.

Why would you want that? Because it's the fastest way to get a "same seed, different output" effect in ComfyUI. Change the seed, the whole sampling trajectory changes; add a touch of this node's noise instead and you nudge the image sideways while keeping the overall composition and prompt intent roughly intact. It's a variation dial rather than a different take.

How it works

Under the hood it's two lines:

input["samples"] = input["samples"] + torch.randn_like(input["samples"]) * scale

Random normal noise, scaled by your scale, added to every latent sample. The noise is fresh each run, so you don't control it with a seed input on this node - each execution rolls new random numbers. Output is the LATENT, ready to continue down the graph.

Inputs and outputs

• input (LATENT) - from an Empty Latent Image, a VAEEncode, or another latent operation.
• scale (FLOAT, default 0, range −100 to 100) - the amount of noise. At 0 it's a pure pass-through; negative values subtract the same noise, which is not the same as silence - it still injects random signal.

Output: LATENT.

Use it between the empty-latent generator and the sampler, or between two sampling stages. Because latent noise at low scale reads as fine-grained "texture variation," it's most useful right before a final pass: nudge the latent, let the sampler polish it, and you get a series of related images you can batch for comparison.

Installing

Ships in the ComfyUI-Model-Bending pack:

cd ComfyUI/custom_nodes
git clone https://github.com/abuzreq/ComfyUI-Model-Bending

Restart ComfyUI and refresh the tab, or search ComfyUI-Model-Bending in ComfyUI Manager. Deps: kornia + scikit-learn, no downloads. Keep the clone folder as ComfyUI-Model-Bending (the README's ComfyUI-Web-Bend-Demo name is stale).

Where people get tripped up

Don't crank the scale and expect "more variation" to mean "better." In latent space, large noise isn't texture - it's a new image. The range goes to ±100, but genuinely useful values live in the tiny band near zero, like 0.05–0.3; past 1 you've basically resampled from scratch and you'd be better off just rolling a new seed. Also worth knowing: this node adds fresh noise every run with no seed control, so if you're trying to reproduce a specific variation deterministically, this isn't the tool - a seed-driven approach is. For quick exploratory batches though, it's the cheapest variation generator in the pack, and it doesn't require you to learn any bending paths to use it.