Nodes/ComfyUI-ZeroCLIP-nodes/ZeroClip-C Coherent Entropy
ComfyUI Node

ZeroClip-C Coherent Entropy

ZeroClip-C Coherent Entropy

By MushroomFleet·Created 5 months ago·Updated 5 months ago· 0
ZeroClip-C Coherent Entropy
  • projection
  • conditioning
seed_int1
region_seed42
blend0.50
mode
dimension768

The problem with pure ZeroClip-C is that every seed is a fresh roll of the dice - nothing relates to anything else. ZeroClip-C Coherent Entropy is the fix: it blends two entropy samples together so you can generate a family of images that share a visual "region" while each stays individually varied. Same region_seed, different seed_int → related images, like characters who belong to the same world.

It's the C-variant answer to "how do I get consistency out of randomness?" You pick a region seed to define the family's shared base, then vary the local seed per image. The blend control decides how strongly each image sticks to the family base versus going off on its own.

The inputs

  • seed_int (default 1) - the local seed, for per-entity variation. Change this per image.
  • region_seed (default 42) - the shared seed, kept constant across the whole family.
  • blend (FLOAT, default 0.5, 0–1) - where between the two samples the output lands. The tooltip is the clearest part of this node: 0 = pure local, 1 = pure region, 0.3 = mild coherence, 0.7 = strong. So 0.7 keeps a family tightly together; 0.3 lets members wander.
  • mode - pure or guided, same as the base C node (guided needs projection).
  • dimension (default 768) - embedding width, matching your checkpoint family.
  • projection (optional) - for guided mode.

Output: conditioning for the KSampler positive input.

How it works

Mechanically it's a slerp - spherical linear interpolation - between the local sample (from seed_int) and the region sample (from region_seed), at the blend fraction. Slerp, rather than plain linear interpolation, keeps the result on the unit sphere the whole way, which is important because both endpoints live on the sphere and a straight-line average would drift off it. The blend control is essentially the dial for "how much family resemblance."

Install

Part of ComfyUI-ZeroCLIP-nodes. ComfyUI Manager (search "ZeroCLIP"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/MushroomFleet/ComfyUI-ZeroCLIP-nodes

Restart ComfyUI; it's under ZeroClip/C - Pure Entropy. No pip install, no artifacts in pure mode.

Gotchas

  • blend semantics run the opposite of what you might guess. At 1.0 you get pure region - every image collapses to the same conditioning, so a whole family of identical images. You want something like 0.3–0.7 for "related but distinct." Don't leave it at the default and assume it's safe.
  • Pure mode family resemblance is still weak. Region coherence works best with the guided mode + projection, because the shared base actually sits near meaningful CLIP territory. In pure mode the "family" is a family of random vectors that happened to be slerped - expect looser visual ties.
  • No memory between runs. Same region_seed and same local seed always produce the same blend - deterministic, yes, but only if you actually keep the seeds the same.
CategoryZeroClip/C - Pure Entropy

Inputs (6)

NameTypeDefaultDescription
seed_intINT10–18446744073709550000Local seed for per-entity variation
region_seedINT420–18446744073709550000Region seed (shared across region)
blendFLOAT0.500–10=pure local, 1=pure region. 0.3=mild coherence, 0.7=strong
modeCOMBO2 options: pure, guided
dimensionINT76864–4096
projectionoptZEROCLIP_C_PROJECTION

Outputs (1)

NameTypeDescription
conditioningCONDITIONING