---
tags:
- ml-intern
---
# Comic Character Extractor

## Overview

This space automatically extracts **all characters** from a comic page image and re-renders each with a **neutral reference prompt** (plain background, no emotion, no pose, no action).

Built by cloning the [prithivMLmods/Qwen-Image-Edit-2511-LoRAs-Fast](https://huggingface.co/spaces/prithivMLmods/Qwen-Image-Edit-2511-LoRAs-Fast) studio UI and adding an automatic multi-character workflow.

## How It Works

1. **Character Detection** — Grounding DINO (tiny) detects all people/characters in the comic page.
2. **Segmentation** — Segment Anything Model (SAM, vit-base) segments each detected character.
3. **Extraction** — Each character is cropped onto a plain white background.
4. **Neutral Re-render** — Qwen Image Edit 2511 renders each character with a neutral prompt:
   > `white_background clean flat neutral character portrait, front_view face forward standing_pose neutral_expression, no_emotion no_action no_gesture`

## Features

- **Auto-detects all characters** — No need to specify how many characters are on the page.
- **Neutral reference output** — Plain background, no emotions, no poses, no actions.
- **Studio UI** — Dark-themed interface with canvas view, filmstrip, and inspector panel.
- **Keyboard shortcuts** — `Ctrl+Enter` to auto-extract, `Escape` to close zoom.

## Usage

1. Upload a comic page image (PNG/JPG).
2. Click **"Auto-Extract Characters"** or press `Ctrl+Enter`.
3. Each character appears as a separate image in the filmstrip.
4. Click any character in the filmstrip to view it on the canvas.
5. Use the Download button to save individual character images.

## Models Used

| Component | Model |
|-----------|-------|
| Character Detector | `facebook/grounding-dino-tiny` |
| Segmenter | `facebook/sam-vit-base` |
| Image Editor | `Qwen/Qwen-Image-Edit-2511` |

## Citation

This project uses:
- [Grounding DINO](https://arxiv.org/abs/2301.12092)
- [Segment Anything Model (SAM)](https://arxiv.org/abs/2304.02675)
- [Qwen-Image-Edit](https://huggingface.co/Qwen)

## License

MIT