Conditioning Variation

You found a composition you love and you want to explore variations of the same prompt without re-rolling the image. The blunt tool is the sampler seed - but that changes everything, including the noise layout you just fell for. The smarter tool is to keep the sampler seed locked and vary the conditioning instead. That's exactly what Conditioning Variation is for: it nudges your prompt's embedding with seeded noise, giving you coherent sibling images from one seed.

How the nudge works

The mechanism is where it earns its keep. If you just added random noise to the embedding, the nudge would scale with how loud the conditioning happens to be - a loud prompt would get thrown around, a quiet one barely tickled. This node scales the noise to each token's own magnitude, so strength means the same thing regardless of what you're conditioning. At the default 0.1 you get close variations; crank it higher and you drift further from the prompt.

The preserve_magnitude toggle (on by default) is the quality knob. When it's on, each token gets rescaled back to its original magnitude after the nudge, so the variation changes the direction of the embedding - its content - without changing its activation energy. The README's framing is the right one: on, you get coherent variety rather than louder or washed-out conditioning; off, you get a rawer perturbation. Leave it on and save "off" for when you're chasing chaos.

The four knobs

The inputs are exactly four: conditioning, strength (0–2, default 0.1), seed (the variation seed - change it to get a different variation), and preserve_magnitude. One CONDITIONING output goes right back onto the wire. To A/B it, lock the sampler seed, set this to 0.1, and roll just the variation seed between runs - the sampler seed stays fixed, so you're seeing pure conditioning variation, not noise re-rolls.

A practical pairing: this is the conditioning-side cousin of the sampler's control_after_generate. Where the sampler seed gives you wild swings, this gives you measured drift - the difference is night and day when you're iterating on a character's outfit or a scene's mood rather than fishing for a new image.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-NynxzExperimental

or search "Nynxz" in ComfyUI Manager, then restart. No dependencies or downloads; it's a pure torch transform on the conditioning tensor. And like everything in this pack it's explicitly moving, so pin the version if a workflow depends on it.