metadata
library_name: transformers
pipeline_tag: image-text-to-text
tags:
- image-forensics
- image-manipulation-detection
- visual-tool-use
- qwen3-vl
ForgeryVCR
ForgeryVCR is a visual-centric image-forensics agent built on Qwen3-VL-4B-Instruct. It actively calls ELA, FFT, NPP, and Zoom-In tools, classifies an image as authentic or tampered, and localizes manipulated regions with bounding boxes.
The released merged GRPO checkpoint is stored under GRPO/.
Download
hf download youqiwong/ForgeryVCR \
--repo-type model \
--include "GRPO/*" \
--local-dir weights/ForgeryVCR
Use the checkpoint from the ForgeryVCR repository:
INFERENCE_MODEL_PATH=weights/ForgeryVCR/GRPO \
INFERENCE_DATASETS="all" \
bash scripts/run_inference.sh
The full tool definitions, prompts, preprocessing, SAM2 mask generation, and evaluation code are maintained in the project repository.