Prompt Cleaner

Here is a prompt you will actually end up with, because Danbooru gives you tags in its own format and you paste them:

drawing_bow, sunna_(zenless_zone_zero), (masterpiece:1.3), best_quality, tagme

Here is what Anima actually wants:

sunna \(zenless zone zero\), drawing bow, (masterpiece:1.3), best quality

Three things changed: underscores became spaces, the series brackets got escaped so they survive, and tagme - a pure metadata tag that means "someone should tag this" - vanished. Prompt Cleaner does all three in one node, at paste time.

Why the underscores matter, and why the brackets matter more

Anima is a 2B DiT on Cosmos-Predict2 with a Qwen3-0.6B text encoder, and it takes Danbooru tags markedly better than paraphrases of them. But it does not take them in Danbooru's storage format. The community rule from the Anima prompting thread is blunt: replace underscores with spaces, score_7 being the one exception. That is the whole market for this node - it is the difference between drawing_bow (a token sequence the model has never seen) and drawing bow (a concept it definitely has).

The bracket part is the subtler trap. In ComfyUI, (sunna (zenless zone zero)) is not a character name with a series in it - token_weights parses round brackets as attention syntax, so the series gets eaten as if it were ((...)) emphasis. The node escapes them to \(zenless zone zero\) because escape_important in comfy/sd1_clip.py is the one escape sequence that code path understands, and Anima's encoder routes through it. That's why the fix holds for Anima and not just for SDXL checkpoints.

What you set, and what you get

The node has a permanent text_in socket, an autogrowing bank of extra sockets, and the usual multiline text box. This is the part people get wrong: they are not either/or. Everything connected plus whatever is typed gets concatenated in the order text_in → extra sockets → text box, with newlines between, and then cleaned as one prompt. So dedupe and the blacklist work across sources - if your character node and your style node both emit masterpiece, you get it once. Autogrow gives you up to ten extra sockets and mints a fresh one whenever you connect the last free one; that's normal behaviour, remove it if you don't want it.

After that come the cleaning switches, all of which now live behind one ⚙ settings dialog on the node (they are still real, declared inputs - same names in the workflow JSON, same values in an API-format run, they are just drawn at zero height). For a beginner, four matter:

• bracket_policy - leave it on escaped parentheses only. The "escape everything" option writes \[ and \{, which ComfyUI never un-escapes because square and curly brackets are not syntax there, so you inject literal backslashes into your prompt for nothing.
• underscore_to_space - on by default, and it will turn score_7 into score 7. If you use score tags (the model card recommends score_7), turn this off.
• strip_blacklist plus the blacklist box - the built-in list kills tagme, watermark, signature, artist request and year 20xx. Plain lines are whole-tag matches, so signature won't eat signature move; re: lines are regexes, and a broken one prints a console warning instead of failing your run.
• dedupe - case-insensitive, keeps the first occurrence.

The single output is cleaned_text, a one-line string you wire into any CLIP Text Encode (or into the pack's own encode node). You don't need a preview node to see it - the result is painted onto the node, and this node is registered as an output node, so it runs and shows the text even with nothing plugged into it. That is the fastest way to check what your pipeline is actually sending.

One opinionated note on keep_weights: it protects (tag:1.2) from being escaped, which is the right call, since mangling a weight into \(tag:1.2\) would be worse. But on Anima prompt weighting is discarded by the LLM encoder path, so those brackets are literal punctuation to the tokenizer. Booru block, comma+space, no weights is still what gets you better images.

Install

No dependencies at all - requirements.txt is comments only and pyproject.toml declares an empty dependency list, so nothing gets pip-installed behind your back. The one requirement is a recent ComfyUI: this pack uses the V3 node API (comfy_api.latest) and declares requires-comfyui >= 0.3.60.

cd ComfyUI/custom_nodes
git clone https://github.com/L134283/comfyui-promote-cleaner
# restart ComfyUI

Or search Prompt Cleaner in ComfyUI Manager, or comfy node install comfyui-promote-cleaner. You will find it under the Prompt Cleaner category in the right-click menu and in Essentials → Basics. To update: git pull in that folder, then hard-refresh the browser.

Where people get burned

The buttons aren't there. + text input and ⚙ settings are added by a frontend extension (web/js/prompt_cleaner.js). If you don't see them, check the browser console and hit Ctrl+F5 - custom node JS is cached aggressively. The backend node still works without them; you just lose the dialog.

No nodes at all after restart. Almost always a ComfyUI older than 0.3.60. Update ComfyUI, not the pack.

The two dropdowns are in Chinese. Deliberate: they are parameter values written into the workflow JSON, not translatable UI strings, so the author kept them verbose rather than showing two identical-looking commas. The tooltips explain them.

Fair warning on expectations: this is a 2026 pack from a single author with essentially no community footprint yet, so you're early. But every failure mode above is a frontend or version issue - the cleaning engine is plain Python with no third-party imports, and that's the part that usually rots.