Comic translation is far more demanding than ordinary text translation. It blends visual art, varied speech bubble geometry, vertical layouts, sound effects, and distinct character voices. Generic screen translators fail because they translate each bubble in isolation—losing pronouns, context, and recurring names.
1. Why the Model Needs Context
Even the strongest model, looking at a single page, can't know what happened three pages ago or how two characters address each other. Fumitoon sends every page to a vision-language model together with this context:
- Story summary: A short summary of the earlier pages goes with each new page, so split sentences and references are understood.
- Series glossary: Character names, their gender, how others address them and recurring terms are kept per series, the most frequent first.
- Your corrections: A glossary entry you edit is locked and the model won't change it. While reading you can correct a bubble yourself or change a name everywhere in a chapter at once.
2. Finding Bubbles and Removing the Lettering
Before the translation is set, the speech bubbles and narration boxes are located. Only the pixels of the letters are erased, not the whole bubble, and the gap is filled from the surrounding background, so bubble outlines and nearby art survive. The translation is set in a comic font close to the measured size and original colour of the lettering.
3. Preserving Drawn Sound Effects (SFX)
Hand-drawn sound effects are part of the illustration itself. Overwriting dynamic SFX with flat text damages the artist's composition, so Fumitoon leaves drawn sound effects untouched and translates dialogue and narration.
