Screenshot translation: text inside images, too

Recognition and translation happen in one pass. Splitting them (read first, translate second) looks tidier, but a misread character gets passed straight into translation and compounds the error. A multimodal model sees the glyphs and the layout at once, so blurry text, vertical text and meme fonts all come out better.

You get the source text as well as the translation: the source lets you check what was read, and it is also what gets metered — counted against your monthly quota by recognised character count, the same unit as text translation.