The method is as below.
1.
Copy source model to be a target model
2.
Render each image in both models with same view point
3.
Transform “image rendered by target model” using InstructPix2Pix guided by” image rendered by source model” and text instruction for stylization
4.
Use SDS to update target model
They insisted that they targeted the inconsistency problem with img2img diffusion format, but in my opinion this doesn’t guarantees high structure preservation.