This model cannot generate NSFW content.

This model is only trained on female cosplayers.

Be sure to use sgm_uniform and deis/dpmpp_2m for inference with around 50 steps.

This model is a training test conducted on Flux. Due to limited data and the way it is labeled, this model cannot replicate the cosplay of a specific character. For example, it cannot reproduce the cosplay of "Konan" from Naruto. What is highly consistent in the dataset is the beauty of the coser, so this model is less of a "cosplay model" and more of a model for beautiful Chinese cosers.

The dataset labeling combines natural language and tags, with high labeling quality. It is hoped that under low-weight usage, the model's understanding of the text will not be overly distorted.

The trigger phrase is "Cosplay photography." Always place this phrase at the beginning of your prompt.

An example of a prompt could be:

Cosplay photography, Portrait of goth girl, long black wavy hair, huge breasts,lace choker, depth of field,heavy bokeh, cinematic lighting,cinematic color,dramatic lighting,ultra detailed,looking away,

Since fine-tuning T5-XXL is not meaningful, this training did not involve fine-tuning T5-XXL. Therefore, you can certainly write your prompts using pure natural language. It is more recommended to use some Vision-Language Models (VLMs), such as GPT-4o, to infer images and generate prompt suggestions.