---
license: apache-2.0
base_model: jzhang533/PaddleOCR-VL-For-Manga
tags:
  - gguf
  - ocr
  - manga
  - paddleocr
library_name: gguf
---

# PaddleOCR-VL-For-Manga, GGUF

GGUF conversion of [jzhang533/PaddleOCR-VL-For-Manga](https://huggingface.co/jzhang533/PaddleOCR-VL-For-Manga),
a PaddleOCR-VL fine-tune for Japanese manga text. Converted for use in the browser
via [wllama](https://github.com/ngxson/wllama), which needs GGUF rather than
safetensors.

| file | size | what it is |
| --- | --- | --- |
| `PaddleOCR-VL-manga-Q4_K_M.gguf` | 286 MB | decoder |
| `mmproj-Q8_0.gguf` | 570 MB | vision projector |

Both files are required.

## Usage

```sh
llama-mtmd-cli -m PaddleOCR-VL-manga-Q4_K_M.gguf --mmproj mmproj-Q8_0.gguf \
  --image crop.png -p "OCR:" --jinja --temp 0
```

Use the prompt `OCR:` — these weights are trained on it. The model expects a
crop of a single text region, not a whole page.

## Quantization

Converted with `convert_hf_to_gguf.py` (llama.cpp b10150) to F16, then quantized
with `llama-quantize` to Q4_K_M. The vision projector is Q8_0.

The upstream vision config declares `SiglipVisionModel`; the converter's mmproj
path expects `PaddleOCRVisionModel`, so that field was renamed before converting.
No weights were altered.

On a 7-crop Japanese manga page, Q4_K_M output was character-identical to F16
except for one ambiguous handwritten kanji, so the smaller file costs nothing in
practice.

## Licence

Apache-2.0, inherited from PaddleOCR-VL. Credit for the fine-tune belongs to
[jzhang533](https://huggingface.co/jzhang533); this repository only changes the
file format.