Plans for Better East Asian Identity Consistency in v1.3?
Plans for Better East Asian Identity Consistency in v1.3?
Hello, and thank you for creating this LoRA. I have been using it frequently and really enjoying it.
Version 1.2 feels noticeably better than v1.1. The improvement in identity preservation is substantial, especially when restaging a person in a new scene.
However, I still find identity consistency slightly less reliable for some ethnicities, particularly East Asian subjects. It can produce a very convincing likeness, but the results are somewhat inconsistent.
My main question is whether you are planning to expand the training for a possible v1.3 with:
- More poses and camera angles
- Greater facial and appearance diversity
- Broader representation across different ethnicities
- More general human identity data
I would be especially interested in improved consistency for East Asian subjects, including Chinese, Japanese, and Korean faces. More diverse East Asian training data would likely help with this.
I also saw the character reference-sheet capability mentioned on the model card, and I read elsewhere that you may be doing additional character-sheet training. I am very excited about that direction, but I am also curious whether you plan to continue expanding the general-person training data.
To be clear, v1.2 already works very well for me. When I generate five images, one or two often preserve the person’s identity extremely closely, which I am very happy with.
The remaining weakness I notice is mostly with side profiles and three-quarter views. In those angles, the face can occasionally look slightly less like the reference, particularly in facial proportions and structure.
Honestly, I feel that even another 20% improvement in identity consistency would make this an exceptionally complete and definitive identity-editing model.
Are you planning to train future versions with more varied human subjects, facial structures, ethnicities, poses, and viewing angles?
Thank you again for your work. I am already very satisfied with v1.2 and look forward to seeing how the next version develops.