This model is trained on flux.1-dev2pro. Caption with dan tags and NL(pixtral-12b). Dataset from danbooru2024 use cheesechaser selected some artist for total 1w+ high quality illustrations(thanks for deepghs). Similar to the format of TIPO(both, tag first), recommend to use it to detail prompt or Gacha. This model have no artist name tags.

How to prompt

Use dan tags to describe outfit and what objects you wanna add.

And use NL to describe the pose in detail for the Character.

And recommend to use TIPO-200M-ft-F16 get detail prompt.

Model Type

Q4_1 model is for llama.cpp Q4 inference.

fp16 model(too large to upload) is for regular inference.

dev2pro_pretrain model(too large to upload) is base model for train on it, the model above is continue training with particular dataset on dev2pro_pretrain model.

描述:

训练词语: