Some feedback based on v0.3. I think the final output depends largely on the quality of the character model used. For example, the アロナ model trained by Khanon is excellent - it retains almost all of the character's features. It performs well in v0.3 and requires very little parameter tuning or prompting. On the other hand, the quality of the XC2 model is average, which means I had to generate thousands of images and pick out the few that accurately represented the character's face.

I will try v1.0 soon.