GodnessMerger_NoiseInjection

This is the odd one out in the Godness Merger family: everything else in that group is about combining two checkpoints, but GodnessMerger_NoiseInjection takes a single model and hands you back a perturbed version of the same one. It's the "add controlled randomness directly to the weights" tool, not a merge tool - the kind of thing people reach for when a checkpoint (often itself the product of a merge) has converged a little too hard on one look and you want to knock it slightly off that local optimum without retraining anything.

The mechanism

You pick an operation - gaussian or random - plus a mean, a std, a ratio, and a seed. Read literally: ratio controls what fraction of the model's weights get touched at all, and mean/std shape the noise added to the ones that are. gaussian almost certainly means the added noise is drawn from a normal distribution centered at mean with spread std - that's the standard meaning of those two parameters together, and it's the more controllable of the two modes since Gaussian noise concentrates near the mean instead of spreading unpredictably. What exactly random mode does differently isn't spelled out anywhere in the README (it's all screenshots, no parameter docs), so take that distinction as "two flavors of noise, pick the one that gives you results you like" rather than something with a precisely documented difference. seed pins which weights get hit and with what values, so the result is reproducible.

The knobs that matter

• operation - gaussian (default) or random. Start with gaussian; it's the more predictable of the two.
• ratio (default 0.98) - how much of the model gets perturbed. The high default suggests this is meant to touch nearly everything by a small amount, not a small slice by a lot - pair it with a small std rather than cranking mean/std up.
• mean (default 0.05) / std (default 0.01) - small by design. These aren't meant to visibly wreck the model; the defaults suggest a subtle nudge, not a scramble.
• seed - reproducibility, as always with this pack's random-anything nodes.

Output is a plain MODEL - wire it straight into whatever normally consumes a checkpoint (a KSampler, or another merge node if you're chaining this into a bigger recipe).

Installing it

Manager: search "SP-Nodes," install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/bananasss00/ComfyUI-SP-Nodes

Restart ComfyUI afterward. No extra dependency or model download for this node specifically - it operates on a model you've already loaded.

Where this goes wrong

Push mean, std, or ratio too far from the small defaults and you will hear it - degraded output, artifacts, or a model that's visibly noisier than the source, because you're literally adding noise to every weight the ratio selects. If that happens, that's the node working as described, not a bug; back the values off. Start from the defaults and nudge one parameter at a time rather than changing all four at once, or you won't know which knob caused the difference.

The other thing to keep in mind: like the rest of this pack's random-anything nodes, "random" and "gaussian" are only reproducible if you keep the seed fixed. If you land on a noise pattern that improves a checkpoint in some subtle way, write the seed down - there's no way to recover it after the fact just by looking at the output model.