How does Nano Banana Pro achieve such high accuracy in CJK text rendering, especially Korean (Hangul) text rendering?

Gemini API
,
1

Hi Google DeepMind team,

I’m impressed by Nano Banana Pro’s Korean text rendering accuracy. The official blog mentions “enhanced multilingual reasoning” but I’d love to understand the technical approach:

  1. What architecture changes enable accurate Hangul rendering? (Hangul has 11,172 possible syllable combinations)

  2. Is there a specialized Korean glyph encoder or does it use the Gemini 3 Pro text encoder’s multilingual capabilities?

How was the Korean text training data constructed?

I’m researching ways to teach multiple languages ​​in a limited environment. Any information would be greatly appreciated.

Thank you!