Multi Vocab String List (Nukun)

Prompt mixing usually means a wildcard text file, a randomize seed, and a lot of squinting at your output trying to remember which combo made the good one. Multi Vocab String List (Nukun) attacks the other side of that problem: instead of one random grab-bag, it gives you four vocabulary slots, each with its own deterministic word cursor. You don't shuffle blindly - you walk through combinations like flipping cards, one at a time, and the good one is reproducible.

How it works

Each of the four slots has three widgets: a vocab_file picker, an amount (how many words to draw), and a word_index. The vocab file is a comma-separated word list - bundled resources like adjectives.csv, animals_mythical_creatures.csv, camera_composition.csv, visual_art_styles.csv, plus Little Doom LoRA keyword sets - or anything you drop into the pack's resources/ folder (.csv, .txt, or .json all work). It also reads ComfyUI/user/vocab.json.

The clever part is word_index. It's a block cursor, not a per-word counter: incrementing it walks forward through non-overlapping shuffled blocks of words before reshuffling, so consecutive runs never overlap. With ComfyUI's control-after-generate dropdown you can keep, increment, decrement, or randomize each slot's cursor independently - one slot advances while another stays fixed, and every combination is replayable on the same seed.

The inputs that matter

• vocab_file_1..4 - which deck each slot draws from.
• amount_1..4 - words per slot; 0 disables the slot, which is the intended way to use fewer than four.
• word_index_1..4 - the block cursors. This is where the control-after-generate menu earns its keep.
• chain (optional) - an incoming string; when connected, the combined output is appended after it.

Outputs are combined (all active slots joined) plus slot_1 through slot_4 as separate strings, so you can wire individual slots to different encoders or keep one as a filename component.

What it's good for

This pack pairs it naturally with Ollama Prompt Refiner (Nukun) - feed the combined output in as a random "word salad" and let a local LLM turn it into a coherent prompt. But it's just as useful standing alone: put characters in slot 1, settings in slot 2, style in slot 3, camera in slot 4, and sweep the cursors. Amounts larger than the vocabulary clamp to the full list, and words are deduplicated before sampling, so you won't get "red red sky."

Installing it

It's in the Nukun_ComfyUI_Nodes pack - ComfyUI Manager (search "Nukun") or:

cd ComfyUI/custom_nodes
git clone https://github.com/OnekoSL/Nukun_ComfyUI_Nodes

Restart ComfyUI. No model downloads, just numpy/Pillow/scipy/PyWavelets. One gotcha: a new vocab file you add to resources/ only appears in the dropdown after a restart - the file list is built at load time. And if you want to freeze a good combo, note the word index and the seed, not just the words - that's what makes it reproducible.