Lumi Shuffle Prompt

Same prompt, same seed, same image - that's the deal, until you shuffle the order of the prompt itself. Lumi Shuffle Prompt exists to break that monotony: it takes your prompt text, scrambles the token order, and hands back a rearranged version. Wire it between your prompt source and your text encoder, and the same content can generate genuinely different compositions without you changing a word.

The mechanism is exactly what the README says: it strips newlines and commas, splits the remaining text on spaces, shuffles with a seeded RNG, and rejoins with single spaces. That means a red car, city street, sunset becomes something like street sunset red a city car - order scrambled, words preserved.

Whether that's a feature depends on what model you're feeding. This is where the model-family split in modern prompting really bites. On the SDXL/booru-tag lineage (Illustrious, NoobAI, Pony), tag order has only mild effect and shuffling is a cheap way to get variants from a fixed seed. On LLM-encoded models where "your prompt is an instruction, not a token bag," shuffling changes the meaning - "a red car" and "car a red" are different instructions, and you'll often get visibly worse or broken results. The node doesn't know which world you're in; you do.

The inputs that matter

• text - the prompt to shuffle (multiline).
• seed - the shuffle seed. Same text + same seed = same shuffled order, every time. Change it to roll a different arrangement.

Output: shuffled text, a single STRING you can feed into your encoder, a wildcard processor, or straight to the LLM prompt processor for enhancement.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/illuminatianon/comfyui-lumi-tools
cd comfyui-lumi-tools
uv sync

Restart ComfyUI, or grab "Lumi Tools" from ComfyUI Manager.

The gotchas

Three practical notes. First, the shuffling is word-level, not tag-level - commas get removed, so multi-word tags like cyberpunk city travel as a unit, but a comma-separated list loses its structure. If you rely on comma boundaries, shuffle the result downstream of your structure, not before. Second, keep the seed wired to your workflow-wide seed (Lumi Seed works great here) so the whole graph advances coherently. Third, and this is the one that gets people: if you're on an LLM-encoded model, test the shuffled output on a fixed seed before you trust it - a scrambled instruction can quietly produce garbage that looks fine in the prompt box.