Prompt Tag Cleanup

Prompts arrive messy. A concatenated pile of tags has duplicates, stray commas from unconnected inputs, and line breaks that become part of a tag if you're not careful. Prompt Tag Cleanup splits that mess into tags, drops what you don't want, optionally sorts and caps the list, and joins it back up - while correctly treating emphasis like (tag:1.4) as the same tag as the plain spelling. That last bit is the part that separates it from a naive string dedupe.

It's the polite-bouncer node: most of what it does is invisible until you wire it into the path and realize your prompt stopped growing a copy of every tag on every concat. If you build prompts from multiple Text Concatenate branches - a subject block plus a style block plus a quality block - you've almost certainly shipped duplicates without noticing, because the eye skips repeated words. This node notices.

The switches that matter

• delimiter - what separates tags in the incoming text. The default , treats comma-separated tags correctly. Leave it empty and every word is its own tag, which is rarely what you want.
• dedupe / ignore_case / ignore_emphasis - the deduping trio. ignore_case on means Neon Glow and neon glow are one tag; ignore_emphasis on means (neon glow:1.4) and neon glow are one. The tooltip's suggested combo - emphasis ignored, keep set to last - keeps the weighted spelling when a duplicate exists, so your (tag:1.4) survives the cleanup.
• keep - first or last: which spelling of a duplicate survives.
• remove_empty - drops tags holding nothing. This is what clears the run of bare commas an unconnected input leaves behind.
• collapse_whitespace - runs of spaces, tabs and line breaks inside a tag become one space. This is what removes the accidental line breaks from a pasted multi-line prompt.
• sort - none keeps written order (which preserves word order, the thing models actually read), or sort alphabetically/frequency if that's your workflow.
• limit - keep at most this many tags, counted after everything else runs. 0 keeps all. Useful for trimming a prompt to fit a token budget.

Outputs

text is the tidied prompt, joined back up with join_with (the default , is ordinary prompt spelling; \n gets one tag per line). tags is the surviving set as one LIST for the list nodes, tag_strings is the same tags as a STRING list - a node reading it runs once per tag, which is how you'd render once per tag if that's the experiment. count and removed report what survived and what went, and removed of 0 means your prompt was already clean.

The honest limits

Deduping is exact-match based, not semantic - it won't notice cat and feline overlap, and nothing should expect it to. Sorting by anything other than none reorders your tags, and for tag-driven models word position matters, so only sort when you mean it. The node also operates on the string level: it doesn't know or care which checkpoint you're prompting, so the quality-tag ordering advice from your model's docs is still your job.

Installing

Ships in WAS Node Suite v3 - ComfyUI Manager (search WAS Node Suite v3) or:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui

Restart after cloning; needs ComfyUI 0.14.0+ and Python 3.10+. Like every node in the v3 pack it installs nothing extra - no pip packages, no downloaded weights - which is a deliberate change from the old v2 suite's heavy dependency load.

Natural neighbors: put it after Prompt Parse when wildcard draws produce multi-tag lines you want deduped, and before a CLIP Text Encode when you want a clean, deterministic prompt string you can also save.