TagRandomCategory

TagRandomCategory is the smarter cousin of TagRandom: instead of shuffling a list you wrote, it keeps a whole dictionary of categorized Danbooru tags in the pack and draws random ones from whatever category you name. Feel like a random prop in the shot? Type props. Need a random location? Type location. It turns "I don't know what to add" into a single text field, which is exactly the niche the author built it for.

What it actually does

The pack ships a category map of tens of thousands of Danbooru-style tags (the current code reads tag_category_v3.json, roughly 67k tags, each mapped to a list of categories like expression, color, hair_style). This node builds a pool of every tag whose category list contains what you asked for, then samples from it. The README's own framing is the best description: "I want some small prop" → props; "I want some kind of location" → location.

Two details make it more useful than a plain random pick:

• negative_category - draw from clothing but exclude anything that's also tagged color, so you get clothes without fighting the palette.
• Seeding - same seed, same selection, so you can lock in a roll you like.

Mechanically it samples count tags per category listed (with replacement), then de-duplicates and joins them with commas. Note the dedup: list three categories with count 2 and you can get up to six tags, but a tag that lives in two of your categories is emitted once, not twice.

The inputs that matter

• category - what to draw from, comma-separated if you want several. Matching is forgiving: lowercase, spaces become underscores, so hair style finds hair_style.
• negative_category - tags belonging to these categories get excluded.
• count - how many per category, 1–100.
• seed - the reproducibility knob.

Output is a single tags STRING, ready for a text concat or CLIPTextEncode.

Where people get burned

The author says it straight: if you feed every element of the workflow from this node, you get images that make no sense - "which is, of course, its own kind of interesting." That's the real failure mode. The randomness is per-category and uncoordinated, so a location draw and a props draw have no idea what each other picked. Fine for brainstorming, chaos for a character you need to keep consistent.

The other practical gotcha is the dedup shortfall. Because selection uses choices (with replacement) and then dedups, you can ask for count 3 and get back 2 if the random draws collided. Not a bug - just don't build a workflow that assumes an exact count.

Install

Same pack as the rest of comfyui_tag_filter - ComfyUI Manager, search "comfyui_tag_filter". Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/sugarkwork/comfyui_tag_fillter

Restart after. The repo has no dependencies and no models to download; the entire engine is those category JSON files, so this installs in seconds.