ZeroClip-C Conditioning

ZeroClip-C Conditioning is the "try before you buy" node of this pack: it produces seed-driven CONDITIONING with zero setup - no anchor libraries, no trained models, no downloads. In pure mode it just samples a random point on the unit hypersphere, deterministically from your seed. The name of the category ("Pure Entropy") is not an accident. This is the closest the pack gets to "free exploration."

If you want to understand the whole ZeroCLIP idea in five minutes, wire this node straight into a KSampler with an Empty Conditioning negative and start rolling seeds. No artifacts, no loaders, instant text-free generation. It's also the most honest demonstration of what these nodes are doing: you're replacing the text encoder with a deterministic number → vector function, and the model is left to make sense of an embedding that no human ever wrote.

The inputs

• seed_int (default 42) - a single u64 entropy seed. Note it's a plain integer, not the four-part ZEROCLIP_SEED that A/B/D use - C works differently.
• mode - pure or guided. Pure needs nothing; guided needs the optional projection input connected.
• dimension (default 768) - the embedding width; the tooltip marks 768/1024/2048 as the meaningful values. Required for pure mode, because a pure sample has to be drawn at the right dimensionality.
• projection (optional) - the ZEROCLIP_C_PROJECTION from a Load Projection node. Required for guided mode, ignored for pure.

Output: conditioning, standard CONDITIONING for the KSampler positive input.

How it works

Pure mode: the seed is FNV-1a hashed, then for each of the D dimensions it derives two child hashes and turns them into a standard normal via Box-Muller - generating a random Gaussian vector that's then normalized to the unit sphere. Deterministic, but effectively uniform over the sphere. Guided mode: the same sample gets projected onto the principal components of real CLIP embeddings (plus the mean), pulling it near the CLIP manifold so the output reads more "language-like" than "white noise."

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 model files in pure mode. Guided mode only needs projection.npz in models/zeroclip/.

The honest take

Pure C is zero-effort and genuinely fun, but "unpredictable, pure prior resonance" is doing a lot of work in that sentence - pure mode has no semantic structure at all. You're not browsing concepts; you're rolling dice in a 768-dimensional space and hoping the model's prior finds something resonant. That's why the guided mode and the A/B/D variants exist. Use C pure to test the pack, use guided when you want C but with a hint of language structure, and use A when you actually want related outputs from nearby seeds.