Random Character Picker (Image Saver)

This is the odd one out in the Image Saver pack - nothing to do with saving files. It reads a text file of character tags and randomly picks some to drop into your prompt. If you generate anime/booru-style images and want variety without hand-picking a character every run, this rolls the dice for you: point it at a tag list, tell it how many to grab, and it outputs a formatted tag string ready to feed a text encoder.

Think of it as a lightweight wildcard specifically shaped for danbooru-style character tags, with the formatting knobs those tags actually need.

How it works

You give it a file_path to a tag file and it selects count entries at random (the seed makes that repeatable). Then it formats them the way prompt tags want: replace_underscore turns character_name into character name, escape_parens backslash-escapes parentheses so they don't get read as prompt-weighting syntax, and delimiter sets what goes between tags. weight_by_count biases the random pick toward more common tags rather than treating every entry equally, and exclude / filter let you blocklist or narrow the pool. The include_core_tags and include_copyright toggles pull in the character's associated defining tags and source series when your file carries them.

The inputs and outputs that matter

• file_path - the tag file to draw from. This is the one you have to get right; no file, nothing to pick.
• count - how many characters to grab (1–1000).
• seed - fix it for reproducible picks, vary it for fresh ones.
• replace_underscore / escape_parens - leave both on for normal prompt use; they make the tags actually parse correctly.
• weight_by_count / filter / exclude - shape the pool and the odds.

The single output is tags (a STRING) - wire it into a CLIP Text Encode, or concatenate it with your base prompt.

Installing it

ComfyUI Manager: search ComfyUI Image Saver, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/alexopus/ComfyUI-Image-Saver
cd ComfyUI-Image-Saver
pip install -r requirements.txt

Restart, hard-refresh. The pack ships no tag file for you - you supply your own.

Where people get tripped up

The whole thing hinges on the file. Point file_path at a list that doesn't exist, or one in a format the node doesn't expect, and you get nothing useful - this is where most confusion starts, so confirm the path and that the file is a plain tag list. The include_core_tags / include_copyright toggles also only do something if your file actually carries that structured data; on a flat list of names they're inert.

And match your formatting to your model. replace_underscore and escape_parens are on by default because booru-trained SD 1.5 / SDXL checkpoints expect spaced, escaped tags - but if you're feeding a natural-language model, a pile of comma-separated danbooru tags isn't how it wants to be prompted. This is a tool for tag-based checkpoints; on newer LLM-encoder models it's the wrong shape of input.