v3: It might be better than v1 and v2.

v1: Trained on Hassaku XL 1.3 with 357 images from the game cg.

This model doesn't work very well, and I'll train another one