In my limited testing, at least in terms of maintaining consistency between input and output for Asian faces, it has even regressed.

Actually, Gemini 3 is about the same, and doesn't feel as good as Claude 4.5. I have a feeling it's been fine-tuned for a cool front-end marketing effect.

Furthermore, I really don't understand why AI Studio, now requiring me to use its own API for payment, still adds a watermark.