Same artwork. Different language.

Panels and text regions are found separately, then classified as dialogue, thought, narration, SFX, signs, or background text. The page is never treated as one block of OCR.

Before anything is translated, Overset assembles the chapter, the scene, the conversation, the speaker, and your glossary into one structured context.

The model returns a literal reading, a recommended translation, alternatives, a confidence score, and a note when the line is genuinely ambiguous.

Only the source text is removed. Simple bubbles get their background restored; text over artwork is reconstructed in the smallest region that will do.

Translated text is fitted into the original bubble by adjusting line breaks first, spacing second, and font size last.

Everything the AI produced is editable — the OCR, the speaker, the region boundaries, the reading order, the wording, and the typesetting.