提示词丨随机组合

The old wildcard trick - __hair__, __outfit__, let the graph roll new values every run - is still the fastest way to explore a prompt space, and it works without any __ syntax here. QING_PromptRandomizer is the QING pack's take: three text pools (one candidate per line), you say how many to pick from each, and it assembles a prompt. Fixed seed means reproducible; -1 means a roll of the dice.

What it does

Inputs: base_prompt (the fixed part), then three pools - pool_a, pool_b, pool_c - each a multiline box where every line is one candidate. pick_a / pick_b / pick_c (default 0 each) say how many to draw from each pool. seed (default -1, meaning "new random each time"; set any non-negative value to make a roll reproducible). unique_pick (default true) draws without replacement, and joiner (default ", ") is the separator between the base prompt and the picked items.

Outputs: prompt (the assembled string), picked_items (what was actually drawn, tagged A:/B:/C: so you can see which pool each came from), and a debug/info string with seed, pool sizes, and pick counts.

Mechanically it's Python's random.Random(seed) doing sample/choice - nothing exotic, which means the semantics are the familiar ones: a fixed seed always produces the same roll for the same pools.

Where you'd reach for it

Batch variation is the killer use: pool A = hair colors, pool B = outfits, pool C = settings, base prompt fixed. Run 30 images and every one differs; find a good combo, copy the picked_items line or lock the seed, and you can reproduce the exact prompt. It's the same idea as the wildcard lists from the prompt-engineering doc, just without the syntax layer - and it composes cleanly with the pack's QING_PromptTemplate if you want the roll inserted into a structured template.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/GAO-SHIQING/ComfyUI-QING
cd ComfyUI-QING
python install_dependencies.py

Restart ComfyUI after. Pure stdlib random, no dependencies. Manager: search "ComfyUI-QING." (README's clone URL has a GAOSHI-QING typo - repo is GAO-SHIQING/ComfyUI-QING.)

Things to know

Two gotchas, one in the pools and one in ComfyUI itself. Pool math: with unique_pick on (the default), asking to pick 3 from a pool that only has 2 entries raises an error - the node refuses rather than repeats. Keep pools at least as big as your pick counts. The caching one is subtler: the node doesn't force an always-rerun (no IS_CHANGED override), and ComfyUI aggressively caches node results. With seed at -1 and unchanged inputs, a cached run can serve you the previous roll instead of a fresh one - which is the exact opposite of what you wanted from a randomizer. If your prompts aren't changing between runs, nudge an input, toggle a widget, or re-queue rather than assuming the -1 did its job.