TIPO (Text to Image with text presampling for Prompt Optimization) is a technique that improves the quality and usability of text-2-image models. 

TIPO uses LLM to preprocess text queries, making them more accurate and informative. It accepts both natural language Prompts and Danbooru tag format.

The basic idea behind the method is that more detailed and specific queries lead to more accurate image generation, while unspecific queries lead to a wider range but less accurate results.

TIPO generates multiple detailed variants of a query from one simple query, thereby expanding the space of possible results and increasing the probability of obtaining the desired image. 

We present 2 models of TIPO, both built on LLaMA 400M, trained on Danbooru2023, GBC10M and Coyo-HD-11M sets with a total of 30 mlrn tokens.

TIPO (Text to Image with text presampling for Prompt Optimization) is a technique that improves the quality and usability of text-2-image models. 

TIPO uses LLM to preprocess text queries, making them more accurate and informative. It accepts both natural language Prompts and Danbooru tag format.

The basic idea behind the method is that more detailed and specific queries lead to more accurate image generation, while unspecific queries lead to a wider range but less accurate results.

TIPO generates multiple detailed variants of a query from one simple query, thereby expanding the space of possible results and increasing the probability of obtaining the desired image. 

We present 2 models of TIPO, both built on LLaMA 400M, trained on Danbooru2023, GBC10M and Coyo-HD-11M sets with a total of 30 mlrn tokens. 

https://github.com/KohakuBlueleaf/KGen