Style Transfer with Auto-Tagger

Using a Ghibli Style LoRA

Workflow Description:

Flux context and GPT-image-1 are here and they are quite powerful tools. One of the main uses I see is style transfer. What I propose here is a workflow that uses Flux (or any other Img2Img2 model) to perform style transfer using both a LoRA I trained (yes it is oversized) and a custom tagger to close the gap with multimodal models like GPT-4. The tagger of choice is Miaoshouai, a Florence 2 based tagger, which in my opinion does a good job at tagging while also allowing to replace tags in the generated caption. The workflow is designed to take an original image and apply a Ghibli-style transformation, while maintaining the essence of the original content.