---
title: Medical Deepfake Detector
emoji: 🩻
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.2.0
python_version: "3.10.13"
app_file: app.py
pinned: false
---

# Medical Deepfake Detector

Hugging Face ZeroGPU Space for single-image medical deepfake detection.

## What It Does

- Upload one image or select an example from `imgs/`
- Resize to `1024x1024` before inference
- Run the Hugging Face model with the same `transformers` inference path as `infer_single_image_transformers.py`
- Parse the `deepfake` bounding box from model output
- Map the box back to the original image and render the result

## Required Configuration

Set these in Space `Settings -> Variables and secrets`.

- `MODEL_ID`: Hugging Face model repo id used by `from_pretrained` if you want to override the default
- `HF_TOKEN`: required only when the model repo is private

`MODEL_ID` is the single source of truth for model selection. If it is not set, the app uses `RichardChenZH/MedForge-Reasoner`.

## Example Gallery

- Example images are loaded from the repo `imgs/` directory
- Supported suffixes: `.png`, `.jpg`, `.jpeg`, `.webp`, `.bmp`
- Invalid files such as `.DS_Store` are ignored
- Add, replace, or delete examples by editing `imgs/` only

## Runtime Behavior

- `@spaces.GPU(...)` wraps the model loading and generation path used by the Space
- Image resize, bbox parsing, reverse scaling, and drawing follow `infer_single_image_transformers.py`
- ZeroGPU-compatible runtime pins are declared in `requirements.txt`
- Missing bbox or invalid bbox math fails fast instead of silently returning a fallback result

## Local Run

```bash
source /opt/miniconda3/etc/profile.d/conda.sh
conda activate base
pip install -r requirements.txt
python app.py
```

If you want to override the default model repo, set `MODEL_ID` before launching.