GALIAIS-Nodes Prompt Random Pool

Want to batch a bunch of looks without hand-typing thirty variants? The Prompt Random Pool is the wildcard picker: you give it a pool of candidate tags, it draws a handful, and hands you a ready-to-wire prompt string. The trick that makes it worth using over the dozens of randomizers out there is the seed - a fixed 随机种子 gives you the exact same picks every run, so a lucky draw is reproducible instead of being a one-time accident.

What you set

• 候选词条 - your pool. One tag per line (or comma-separated); the pack splits on commas, newlines, and semicolons, so pasting a comma list works fine.
• 选择数量 - how many to draw. Default 1, up to 100.
• 随机种子 - the reproducibility knob. Same seed, same selection, always. Set it to 0 to roll the dice fresh each run.
• 允许重复 - off by default, which means no repeats until you run out of pool (it samples without replacement). Flip it on and you get independent picks, duplicates allowed.
• 权重 - applied to each selected tag. Leave at 1 unless you're on a model that understands (tag:1.3) syntax (see the caveat below).

The 随机提示词 output is the joined string; the 元信息JSON output records exactly which tags were selected and the seed you used - which is how you debug "why did batch 14 look different" three days later.

The honest caveat

The weight feature wraps every selected tag in (tag:weight) CLIP-style syntax. That syntax is alive and well on SDXL-lineage anime models (Illustrious, NoobAI, Pony), but it's silently discarded by LLM-encoded models like Anima and Flux - and this pack's other nodes are explicitly tuned for Anima-style output ordering. So if your target model is Anima, keep 权重 at 1 and let the raw tags do the work. The random picking itself is model-agnostic and safe everywhere.

How it fits a workflow

This is the node you reach for when you want prompt variation as a controlled variable. Fix everything else, tie 随机种子 to your KSampler seed via a shared seed node, and the selected prompt tags become part of the batch's identity. Crank the seed between runs to explore, then pin the good seed and lock in that exact composition forever. It's also handy inside the pack's own flow - a random pool feeding 主体 or 场景 slots in the Typed Composer V2 gives you cheap batch diversity without touching the rest of the prompt structure.

Installing

Standard for this pack:

cd ComfyUI/custom_nodes
git clone https://github.com/GALIAIS/GALIAIS-Nodes.git

Restart ComfyUI. No extra dependencies, no model downloads - this node is pure string shuffling. The pack itself needs no Danbooru DB for this one; only the tag-resolution nodes do. Nothing to trip on here, which is honestly the nicest thing about it.