> ## Documentation Index
> Fetch the complete documentation index at: https://koharu.rs/llms.txt
> Use this file to discover all available pages before exploring further.

# Vision and inpainting models

> Choose processors for detection, OCR, and artwork reconstruction.

Open **Settings → Pipeline** and choose a model for each stage. Models own their preprocessing and options; the [workflow](/en/guides/processing) stays the same.

## Detection

**Koharu Layout RF-DETR Seg 2XL** detects text, speech bubbles, and panels, and produces segmentation data for later stages.

Text, bubble, and panel thresholds control how much evidence each class needs. Lower thresholds retain more uncertain regions and may increase false positives. Compare several representative pages before changing thresholds.

## OCR

| Processor        | Use                                               |
| ---------------- | ------------------------------------------------- |
| PaddleOCR-VL 1.6 | Default general vision-language OCR               |
| Manga OCR        | Japanese manga text                               |
| Baberu OCR       | Alternative manga-oriented recognition            |
| Hayai OCR        | Chinese, Korean, Japanese, and English manga text |

<Note>
  OCR reads detected text regions. Check detection first when text is missing: changing the
  recognizer cannot recover a region that was never passed to it.
</Note>

## Inpainting

<Tabs>
  <Tab title="Direct">
    **LaMa** is the default. **AOT Inpainting** is another direct reconstruction model. Start here for routine lettering removal.
  </Tab>

  <Tab title="Generative">
    **FLUX.2 Klein** accepts a reconstruction prompt. **RORem Mixed** is manga-focused and exposes positive and negative prompts.

    These choices generally need more runtime data, memory, and time. Describe surrounding artwork and the removal of letters; typeset the translation with text layers.
  </Tab>
</Tabs>

## Profiles and first runs

Supported processors keep independent settings profiles. Returning to a generative model restores its prompt fields instead of inheriting another model's options.

Models load on demand. A first run includes download and loading costs, so compare warmed models on identical pages before selecting a project default.