Flux Mix Noise

FluxNoiseMixer is a latent-blending utility. It takes two LATENTs - your image latent and a noise latent - and mixes them together, with options for how much, which blend mode, and whether to inject fresh random noise on top. It's not glamorous and it's not required by any of the pack's headline workflows, but it's the kind of node that makes image-editing experiments easier: interpolate between an inversion's latent and pure noise, nudge how much structure survives, or add a controlled amount of randomness to an otherwise deterministic edit.

How it works

The node copies your latent, then applies a couple of independent operations:

• random_noise (FLOAT, default 0.0) - how much fresh gaussian noise to mix into the noise input first, using random_mix_type (mix or add). A mix is a normalized blend; an add is noise + random_noise_latent * amount. This lets you take a stable inversion latent and de-stabilize it a little.
• mix_percent (FLOAT, default 0.5) - how much of the (possibly noise-added) noise latent to blend into your image latent, using mix_type (mix or add). The mix mode normalizes the result so the total energy stays roughly constant - (1-p)*latent + p*noise over a normalizing factor - which is the right way to interpolate two latents. add mode is cruder: latent + noise * mix_percent.
• take_diff (BOOLEAN) - if on, subtracts latent * mix_percent from the result. It's a "difference" mode for when you want the delta between what you started with and what you mixed, rather than the blend itself.

Output:

• LATENT - the blended latent, ready to feed a decoder, another sampler, or back into an inversion pipeline.

When you'd actually use it

Mostly in editing/inversion experiments: after RF-Inversion produces a semantic noise latent, you can use this node to mix in a different seed's noise and get a controlled "how much of the original survives" sweep without re-running the inversion. The mix mode's normalization is what makes that sweep meaningful - plain addition drifts off-energy and gives you mush.

Installation

Standard for the pack: ComfyUI Manager → search "ComfyUI-Fluxtapoz" → install → restart, or git clone https://github.com/logtd/ComfyUI-Fluxtapoz into custom_nodes. Only einshape, no models to download.

Common issues

• Result looks washed out or too dim - you're probably in add mode where energy isn't preserved. Switch to mix for interpolation.
• "I set random_noise and nothing changed." If random_mix_type is mix, a random_noise of 0 means no change by design; and in add mode the amount is multiplied by mix_percent's noise, not your latent - read the labels carefully, the two mix types are applied to different operands.
• It's a utility, not a sampler - output is a LATENT, so it feeds anything that takes a latent, but don't expect it to generate anything on its own.

Once you're doing serious inversion work, this is the node you'll reach for when you want to poke the latent between two runs without rebuilding the graph.