Skip to main content
Open Settings → Pipeline and choose a model for each stage. Every model runs on your machine; the workflow stays the same whichever you choose.

Detection

Koharu Layout HighRes, an RF-DETR segmenter, finds panels, speech bubbles, text, and sound effects, and gives each text block its mask and reading order. Text, SFX, bubble, and panel thresholds set how confident each class must be; empty fields use the model’s own. Lower thresholds keep more uncertain regions and may add false positives. Compare several representative pages before changing them, then run detection again on those pages.

OCR

Manga OCR AR is licensed CC BY-NC-SA 4.0, so keeping the default holds you to non-commercial use. Check each model card for its license.
OCR reads detected text blocks. Check detection first when text is missing: changing the recognizer cannot recover a block that was never detected.

Inpainting

LaMa, fine-tuned on manga, is the default. Start here for routine lettering removal.
Flat and smoothly shaded backgrounds are filled without any model, so the choice matters most for lettering over line art and screentone. See cleanup.

First runs

Weights download from Hugging Face, pinned to a commit, the first time a model loads, and the model stays loaded until the settings switch away from it. A first run includes download and loading costs, so compare warmed models on identical pages before choosing a project default.