tipo-kgen

KGen - A System for Prompt Generation to Improve Text-to-Image Performance

KGen is a project that utilizes Large Language Models (LLMs) to generate prompts for Text-to-Image (T2I) models.

The goal is to enable T2I models to use more complicated and detailed captions during training while maintaining good usability. Which avoid the requirements of multi-level caption or caption dropout for "better flexibility".

This project contains the implementation of some prompt gen model, framework and sampling method.

The implemented framework or sampling method can also be used in other domain of field, but this project will only focus in prompt generation at this time.

Usage

Installation:

pip install tipo-kgen

Use in code: Read the Example code or TIPO-test script for more informations.

TGTS

TGTS: Token Group Tree Sampling for Efficient Multi-Variants Sampling in Language Model

Implementation can be found in sampling module.
Experiment scripts can be found in scrtips

TIPO

TIPO: Text to Image with text Presampling for Optimal prompting

Arxiv Paper: https://arxiv.org/abs/2411.08127

TIPO is a LLM model system designed for generating detailed prompt from input tags or caption. Unlike DTG, TIPO can handle both tags and Natural language. In theory, you can also design your own tag in linguistic way. (For example, long blue hair is acceptable tag in TIPO and will not break the model). The main difference between TIPO and DTG is:

1. TIPO is trained with both Natural Language captions and Danbooru tags, the "nl+tags" data are also not only from danbooru but also some general text-image dataset like Coyo-11M
2. TIPO is trained with better format which achieve some ability on "generate meta infos" such as artists/characters. (or, you can say TIPO have ability to "choose" which artist tag is suitable with current content)
3. TIPO is trained on 30M entries dataset, with more than 25M entries have NL caption, more than 18M entries have tags

Model card

	TIPO-200M	TIPO-200M-ft	TIPO-500M
Arch	LLaMA	LLaMA	LLaMA
Max ctx length	1024	1024	1024
Batch Size	2048	2048	3584
Training dataset	Danbooru, GBC10M, 5epoch Danbooru, GBC10M, Coyo11M, 3epoch	Danbooru(pixtral), Coyo11M, 2epoch	Danbooru, GBC10M, Coyo11M, 5epoch
Real Token Seen*	40B token	50B (10B more from TIPO-200M)	30B token
Training Hardware	RTX 3090 x 4	RTX 3090 x 4	H100 x 8
Training Time	420 hour`	120 hour`	100 hour`
Huggingface	KBlueLeaf/TIPO-200M · Hugging Face	KBlueLeaf/TIPO-200M-ft · Hugging Face	KBlueLeaf/TIPO-500M · Hugging Face

*: We only count "non-padding token" in the token seen, since all the training data have very large length range.<br/> `: Since the training data is pretty short, it cost more time to reach same token seen than general LLM pretraining.
```` As reference, with 4096 as max ctx length and almost all the data have reach that length, you may only need 2days to reach 10B token seen on RTX 3090 x 4 with 200M model.

Usage

A Simple DEMO for TIPO (with t2i functionality included): https://huggingface.co/spaces/KBlueLeaf/TIPO-DEMO

TIPO-extension: https://github.com/KohakuBlueleaf/z-tipo-extension

Cite

@misc{yeh2024tipotextimagetext,
      title={TIPO: Text to Image with Text Presampling for Prompt Optimization}, 
      author={Shih-Ying Yeh and Sang-Hyun Park and Giyeong Oh and Min Song and Youngjae Yu},
      year={2024},
      eprint={2411.08127},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2411.08127}, 
}

DanTagGen

DanTagGen is an early project under KGen, trained on the Danbooru tag system. Danbooru tags often have "overlaps" or "duplicates", such as:

• "long hair" and "very long hair"
• "thighhighs", "black thighhighs", and "black legwears"

Although users naturally avoid these duplications, the model may benefit from having the complete set of tags for better results, as that is how they were trained.

In addition to overlapping tags, "character tags" also need to be mentioned. For simplicity, a "DreamBooth style" prompt can be used to illustrate this:

• Original: a dog
• DreamBooth: a [V] dog
• What User wants: [V]

As shown above, users tend to ignore all the "descriptions" that directly point to the character. In this situation, utilizing LLMs to "connect" the "trigger word" and "related description" is a promising approach. DanTagGen is an experimental project to prove this concept.

Architecture

DanTagGen uses the LLaMA architecture with 400M parameters.

Training

DanTagGen is trained on posts with the top 75% favorite count in Danbooru, which amounts to 5 million entries.

More details about the architecture and training can be found on the Hugging Face page: KBlueLeaf/DanTagGen-beta · Hugging Face

Usage

• Hugging Face Space: DTG Demo - a Hugging Face Space by KBlueLeaf
• SD-WebUI Extension: KohakuBlueleaf/z-a1111-sd-webui-dtg: A sd-webui extension for utilizing DanTagGen to "upsample prompts". (github.com)
• ComfyUI Node: toyxyz/a1111-sd-webui-dtg_comfyui: A sd-webui extension for utilizing DanTagGen to "upsample prompts". (github.com)