OCSNoise ImmiscibleReference

Normal noise is lazy: its random values don't care about your image at all. Immiscible noise is the opposite - it's generated so that it's as different as possible from a reference, like oil and water refusing to mix. That's the idea behind the paper this node implements (arxiv 2406.12303), and OCSNoise ImmiscibleReference is the standalone version of it: generate noise, then reshuffle it so each piece sits far from (or close to) a reference latent. The README is careful to note the technique was designed for training, not inference - so think of this as an experimental flavor that sometimes gives sampling a useful kick, not a guaranteed upgrade.

What it is

A custom noise node (OCS_NOISE output) that takes a reference - either a LATENT or another noise generator - and produces immiscible noise from it. When a sampler requests noise for ancestral or SDE sampling, the node generates a batch of candidate noise and then optimizes the assignment so the noise is maximally (or minimally) distant from the reference. It's the "repelling" ingredient in the OCS noise kitchen.

The inputs that matter

• reference (optional) - a LATENT to repel from. Attach this or custom_noise_ref, not both.
• custom_noise_ref (optional) - an alternative reference source: use noise from another generator as the thing to repel from.
• custom_noise (required) - the base noise being immiscibilized.
• size (default 64) - number of batch repeats to generate for the matching. Setting it to 0 disables immiscible processing entirely, which makes this node a pass-through - handy for A/B comparisons.
• batching (default channel) - what dimension gets matched: channel, batch, row, column, and the cycle_* variants. Column mode reshapes the input and can eat a lot of VRAM; row is slow but less brutal. Stick with channel to start.
• maximize (default false) - false pushes noise away from the reference (the classic immiscible direction); true pulls it closer.
• distance_scale (default 0.1) - the v1/v2 switch: 0 gives you the original (v1) immiscible behavior, non-zero enables the v2 mode that also scales the input noise. 0.1 is the README's suggested starting point.
• blend (default 1.0) - how much of the output is immiscible noise vs. normal noise. custom_noise_blend lets you pick a different noise type for the normal side.

Output is OCS_NOISE, ready for a sampler's custom_noise input.

Why you'd bother

Immiscible noise in sampling is a "shuffle the noise to be maximally spread out relative to the current image" trick. Its biggest practical draw is in Restart sampling - the pack supports immiscible noise for restarts too, and some users report it changing how re-noised steps behave. It's genuinely experimental, so treat it as a sampler-flavor experiment: set it up, A/B against plain noise, and keep it only if it helps.

Installing and gotchas

Part of the pack - Manager search "Overly Complicated Sampling" or git clone https://github.com/blepping/comfyui_overly_complicated_sampling into ComfyUI/custom_nodes, restart. No models or mandatory deps. The blend_mode note: only slerp works well, and that requires installing ComfyUI-bleh. Set blend: 0 or size: 0 to quickly disable and compare.