Skip to main content
Inpainting reconstructs the artwork beneath source text. A precise removal area usually matters more than a larger model.

Review automatic cleanup

Detection gives each text block a mask over its source glyphs, reaching a few pixels past them but never into another text’s box. Inpainting fills each mask and stores the result as the block’s clean-up, shown under its translation. Rerunning inpainting replaces the clean-up. Most lettering sits on plain paper or a gradient, so Koharu fills it without the model when the surrounding pixels allow: The model redraws each text with 256 pixels of context around it, at the page’s own scale, and never across two panels. Only the masked pixels come back. Inspect the result at normal reading size and high zoom. Look for glyph fragments, missed punctuation or furigana, erased bubble borders, repeated textures, and smeared screentone.

Repair a missed region

1

Paint the area

Choose Remove, set the brush size, and cover only the source marks that need reconstruction. Use controlled strokes and avoid neighboring line art.
2

Let inpainting run

Release the stroke. Koharu inpaints that area at once with the selected inpainting model and adds the result as a new raster layer, limited to the painted area by its mask.
3

Touch up and inspect

Use Color picker, Brush, and Eraser for small deterministic corrections. Check the result again at normal zoom.
To start a text over, delete its text block: its clean-up goes with it and the original lettering shows again.

Choose a model

Choose the inpainting model in Settings → Pipeline. LaMa is the default and takes no prompt. RORem, FLUX.3 Klein, MOSS-Image, and Qwen-Image 2.1 are generative alternatives that follow the Prompt there and need more memory and time. See vision and inpainting models.
Brush strokes land on the selected raster layer, a Remove result included. Clear the selection before painting to keep manual touch-ups on a new raster layer.