I'm happy to present you the new version which now produces much better results Yet, it's far from perfection. The main obstacle for me is the lack of anime/cartoon stylized images. To improve the results I had to add some photos to my dataset, which affected the style but not significantly After I have made 8 attempts of training this v2 model I can conclude that it's very hard for lora to understand the 3d structure of a bicycle, I suspect having 500+ images might help but currently it's not possible
I'm happy to present you the new version which now produces much better results
Yet, it's far from perfection. The main obstacle for me is the lack of anime/cartoon stylized images.
To improve the results I had to add some photos to my dataset, which affected the style but not significantly
After I have made 8 attempts of training this v2 model I can conclude that it's very hard for lora to understand the 3d structure of a bicycle, I suspect having 500+ images might help but currently it's not possible