TagForge-0.3B

Version 0.62

Decoder-only transformer that generates structured Danbooru tag sets for single-character images. It adheres mostly to Danbooru styling tags, so it will generate outfits based on actual, working outfit mixes.

Part of the TagForgeSuite (read below) ComfyUI node pack, but can be used in any LLM inferencing tool.

Version 0.7 is currently being trained. It will be larger, as I found out that 294M parameters are not enough to hold all the reasoning/relational data to keep all the styles in mind.

Architecture

Parameter	Value
Vocab size	153,728
Embedding dim	768
Transformer layers	8
Attention heads	12
FFN dim	3,072
Parameters	~294M
Dtype	bfloat16

Usage

Load via the TagForge Loader node in TagForgeSuite (not yet available). The node uses the original-format tagforge_tokenizer.json (included). For raw transformers use, tokenizer.json is the HuggingFace WordLevel tokenizer.

To generate an output, use the command GEN, FILL, or EDIT. To classify tags into their assigned body or scene areas, use CLASSIFY.

For example, to generate a full output: [GEN] 1girl, tomboy, petite, overalls, small_breasts, red_hair, pixie_cut, blue_eyes, t-shirt

This input might output the following:

OUTPUT: RATING : s | SUBJECT : 1girl , petite , tomboy | BODY : breasts , small_breasts | META : simple_background , white_background | HAIR : pixie_cut , red_hair | FACE : blue_eyes | UPPER_BODY : ATTIRE : overalls , shirt , t-shirt | LOWER_BODY : ATTIRE : shorts | EXPRESSION : blush , looking_at_viewer , open_mouth , smile | ACTION : STATE : standing

You would then use the nodes in TagForgeSuite to direct the tags to the TagForgeSuite Sampler, or convert them to a classic prompt.

If you add a RATING to your input (g, s, q or e just like in Danbooru), the model will strictly adhere to your requirements, but it will not leave out the tags you have given it.

Ratings are g (general/safe for work), s (sensitive), q (questionable) and e (explicit) - and e is very explicit.

License

GPL-3.0