Prompt Cleaner (CLIP Text Encode)

Every Anima workflow has a CLIP Text Encode in it. This node is that node with the tag cleaning baked in - you paste raw Danbooru output, it fixes the underscores and the brackets, and it hands the sampler a CONDITIONING you can use immediately. One fewer wire, one fewer place to forget that sunna_(zenless_zone_zero) is not what the model wants to see.

What it is

It is the pack's cleaning engine followed by a normal encode, in a single node: same rules, same switches, same settings dialog as the standalone Prompt Cleaner, plus a clip input on the front and a conditioning output on the back. Anima's Qwen3-0.6B encoder reads real Danbooru tags far better than paraphrases, but it wants them with spaces instead of underscores and with the series brackets escaped so ComfyUI's token_weights doesn't eat (zenless zone zero) as attention syntax. That is the whole reason to install the pack.

Mechanically it does the obvious thing, which is worth saying because it means there are no surprises: it concatenates your text sources, cleans the result, calls clip.tokenize() on the cleaned string, and returns clip.encode_from_tokens_scheduled(tokens) as the conditioning. The escaping holds because the cleaned string is the thing that gets tokenized - escape_important in comfy/sd1_clip.py recognizes \( and \), and Anima's text encoder goes through that path, so a protected series name arrives at the model intact instead of being consumed as emphasis.

Inputs and outputs that matter

clip is the only input that makes this node different from its sibling, and it's the one to get right. Point it at the encoder output of whatever loaded your model - Anima's is the qwen_3_06b_base text encoder the tooltip names. If it arrives as None you wired a model instead of an encoder, and you'll get the familiar "your checkpoint does not contain a valid clip or text encoder model" error. Wiring mistake, not a bug in the node.

Then the text side, identical to the Prompt Cleaner: a permanent text_in socket, up to ten autogrowing extra sockets, and the multiline box - merged in the order text_in → extra sockets → box and cleaned as one prompt, so a masterpiece coming from both your character node and your style node collapses to one. That merge is the real workflow win: wire a tags node and a style node into one encode and stop hand-maintaining a mega-string.

The cleaning switches live behind the node's ⚙ settings dialog: bracket_policy (keep it on escaped parentheses only - "escape everything" writes \[/\{, which ComfyUI never un-escapes), keep_weights, underscore_to_space, dedupe, strip_blacklist, lowercase, separator, normalize_fullwidth, and the blacklist editor. They're hidden widgets, not deleted ones: the same names carry the same values into an API-format run.

Two outputs, and the second one is the reason to prefer this over a stock encode:

• conditioning → straight into your sampler's positive (or negative) input.
• cleaned_text → the cleaned single-line prompt, for reuse or for double-checking.

That second output is worth wiring somewhere: feed it into a stock CLIP Text Encode for your negative, into a filename node, or into a text preview so you can see what actually got sent. The node paints the cleaned text on its own canvas too.

One behavioural difference from the text-only node, and it's deliberate: this one is not an output node, so it only executes when something downstream consumes one of its outputs. That's correct - you don't want a text encoding burned on every queue just to update a preview - but it means that if you queue and the preview text on the node looks stale, the node simply wasn't reached. Check the sampler is actually wired to conditioning.

Install

Nothing to download, nothing to pip install - no models, no wheels, standard library only. It ships in the same pack as the cleaning node, and it needs ComfyUI >= 0.3.60 because the pack is written against the V3 node API.

cd ComfyUI/custom_nodes
git clone https://github.com/L134283/comfyui-promote-cleaner
# restart ComfyUI, then hard-refresh the browser (Ctrl+F5)

ComfyUI Manager users can search Prompt Cleaner, or run comfy node install comfyui-promote-cleaner. The nodes land in the Prompt Cleaner category and in Essentials → Basics; the encode variant is the one with the clip socket.

Where people get burned

The ⚙ and + text input buttons don't show up. They're injected by the pack's frontend JS. Hard-refresh before you file anything - that extension file is cached by the browser. The node still cleans and encodes without them; you just can't open the settings dialog.

score_7 is now score 7. underscore_to_space is on by default, and score tags are the documented exception to the "replace underscores" rule. Turn that switch off if you use them.

Weights look preserved but do nothing. keep_weights protects (tag:1.2) from being escaped, which is the safe behaviour. On Anima the weighting itself is discarded by the LLM encoder path, so it's inert punctuation rather than emphasis - the pack keeps your syntax valid instead of pretending to make it work. The Anima convention is comma+space tags with no weights at all.

Nothing runs. Not a bug - see above. This node is demand-driven, unlike the text-only one.

An old ComfyUI. If the node doesn't appear at all, check your version before anything else.