ZeroClip Seed From Random

ZeroCLIP conditioning wants a four-part seed tuple (concept, style, mood, world), but you almost never want to type four numbers every run. ZeroClip Seed From Random lets you feed it one master integer - which means you can wire ComfyUI's standard randomize-seed widget straight into a text-free workflow and explore the space by clicking "Queue" until you hit something good.

It's the sibling of ZeroClip Seed Pack: that node is for deliberate, nudgeable exploration where you sweep concept_id by hand. This one is for the other 90% of the time, when you just want to roll dice. You pick which depending on whether you're dialing in or hunting.

What it does

One input: seed (INT, default 0) - the master seed, and the tooltip says it plainly: use ComfyUI's randomize widget. One output: a ZEROCLIP_SEED that feeds the same A/B/D conditioner nodes as Seed Pack.

The mechanism is a simple deterministic derivation: it runs the master seed through FNV-1a, then takes four "child hashes" (salted sub-hashes) and splits them into concept_id, style_id, mood_salt, and world_seed. Same master seed, same four values, every time. It's not random in the "unpredictable" sense - it's a stable expansion of one number into four.

Why you'd reach for it

The ZeroCLIP philosophy is "same seed, same image, forever" - conditioning is a pure function of the numbers. Seed From Random is how you get that reproducibility while still getting the familiar ComfyUI loop of randomizing until something lands, then locking it. Find a great image, note the master seed, and you can regenerate the exact conditioning on any machine.

Install

Part of the ComfyUI-ZeroCLIP-nodes pack. Install via ComfyUI Manager (search "ZeroCLIP"), or:

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

Restart ComfyUI; the node appears under ZeroClip/. No pip install - just torch and numpy, which ComfyUI already ships.

The trap every ComfyUI user hits once

If you wire a KSampler-style seed widget here expecting it to stay put, remember the control_after_generate setting: by default it's set to randomize, and it fires after the run - so the seed shown in the box is the one used next, not the one that just made the image you're looking at. If you find a great result and want to keep it, flip control_after_generate to fixed on that input. The same gotcha applies to any seed widget, but it stings more here because the whole point is reproducibility. The community workaround (set widget control mode to "Before" once per install) fixes it globally if you keep losing seeds.