Text Match Pixaroma

Here's a workflow everyone eventually wants: put a vision model in front of a folder of images, ask it "Is the face sharp? Answer yes or no", and keep only the ones it said yes to. The model cooperates. The graph doesn't, because a text answer is a dead end - nothing downstream can branch on it.

Text Match Pixaroma is the adapter. Text in, a yes/no out, plus two gates that let a picture (or a prompt, or a latent) through only on a match or only without one.

How it works

Matching is literal, and that's a deliberate choice. Every word you type is escaped before use, so your list can never be read as a regular expression - no surprise metacharacters, no runaway patterns. Four modes:

• contains any word (the default) - at least one of your words is in the text
• contains all words - every one of them is
• is exactly one of them - the whole text is one of your words, with surrounding spaces, full stops and quotes ignored, so Yes. and "Yes" and **Yes** all count. This is the one you want after an LLM, because an LLM answering "yes" almost never types just yes.
• starts with one of them

You get ignore_case (on by default, so Yes and YES and yes are the same) and whole_words (also on by default). Whole-word matching is subtler than it sounds, and in a good way: it's implemented as "not glued to another letter", not as the usual word-boundary marker. That distinction is what lets cat match the filename 01_Cat - a normal word boundary fires between an underscore and a letter, the pack's rule doesn't. Since filenames are a main use case here, that's the difference between working and not.

The gates are the real feature. Behind the scenes they use ComfyUI's own execution blocker: a closed gate means everything downstream of it simply doesn't run. No error, no warning, no red box. That's the mechanism that turns a per-image yes/no into a whole-folder filter in a single Run.

Inputs, outputs, and the recipe

text is the thing being checked - an AI answer, a prompt, a filename (it's an input socket, so you wire it; you can't type into it). words is your list: one per line, or comma separated, and the first one is number 1. mode, ignore_case and whole_words are as above.

The optional value input is where the trick lives. Wire a picture into it and the two gates pass that picture; leave it unwired and the gates pass the text itself.

[Vision model: "Is the face sharp? Answer yes or no."] -> text
                                        words: yes
                                        value: <the IMAGE>
                                            |
                            if_match ---------+--> Save Image Pixaroma

Five outputs: match (a boolean), index (which word matched, 1 for the first in your list, 0 for none), word (the word itself), and the two gates if_match / if_no_match. Wire the gates to whatever comes next; match and index wire nicely into a Save node's filename or a Show Text so you can see what the model actually said.

Because this is a list node when fed a list, it runs once per item - so each picture is gated on its own answer, not on some aggregate. That's what makes the folder filter work.

Traps

"Nothing downstream ran" is the feature working. If a gate is closed, the run continues silently and your save node just doesn't fire. Beginners read that as a broken workflow. Check match on the node before you go hunting for a bug.

Whole words is on. You look for blurry, the model writes "the image is blur", no match. Turn whole_words off if you want partial hits.

Don't put an Image Grid after a gate. A grid needs the complete list to build, so if any single picture was held back, the whole grid is skipped. Wire the grid before the gate.

On a very old ComfyUI, one without the execution blocker class, the gates can't close - the node passes the value both ways rather than failing your Run, and match still tells the truth. Rare now, but that's why a filter might look like it isn't filtering.

Installing

ComfyUI Manager, search Pixaroma, Install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/pixaroma/ComfyUI-Pixaroma

No models, and nothing beyond the pack's one declared dependency (imageio-ffmpeg, for its video nodes) - this node is pure Python logic and costs nothing to run.

The pack is a solo effort from the designer behind Pixaroma (@pixaroma on YouTube, support over the Pixaroma Discord - GitHub issues and PRs are closed on purpose), and it's bundled in ComfyUI Easy Install on Windows. As with every node in it: if things look broken right after an update, Ctrl+Shift+R the browser tab before you debug anything.