Anime content training

[vicentecarro]

Hi,
Regarding training for Anime content, which differs significantly from real world content in terms of surface normals (due to stylization, shading conventions, and art stuff), from your experience, would it make more sense to (re)train the model from scratch, or to fine-tune the existing one, despite the domain gap?
If the latter, do you have any recommendations? Actually, in both cases, do you have any recommendations in general?

[hugoycj]

I recently found that Flux Kontext Normal Lora is very promsing for image different from training domain, maybe you could have a try https://www.reddit.com/r/StableDiffusion/comments/1meu533/normal_map_lora_for_flux_kontext_yes_is_works_and/

[vicentecarro]

Thank you, I'll check it.