I just tested the new Gemini 3.1 Flash image preview model. The model inherited the 3.0 Pro problem, which sometimes includes hallucinations even when a reference image is provided. It also tries to change the user-submitted photo to achieve better editing results.
This is in contrast to the Flash 2.5 model, which is good at face editing using a reference image.
I will try the new model with my manga generation pipeline. I hope it can solve the inconsistency issues, as what I am doing right now is using 2.5 Flash for quick character creation using a base image, and then using 3.0 Pro for text and caption generation.