---
license: apache-2.0
base_model: Qwen/Qwen2.5-VL-7B-Instruct
datasets:
  - Huang-yating/SID_Set
language:
  - en
library_name: transformers
pipeline_tag: image-classification
tags:
  - image-forensics
  - image-manipulation-localization
  - multi-agent
  - qwen2.5-vl
---

# MADL model weights

This repository is the weight companion for **MADL: Towards Dependable Image
Forgery Detection via Multi-Agent Forensic Reasoning**. The executable code,
configuration files, and evaluation scripts are distributed in the companion
[GitHub code repository](https://github.com/shannima/MADL).

Model repository: [moindy/MADL](https://huggingface.co/moindy/MADL).

MADL produces one of three image-level labels—`real`, `synthetic`, or
`tampered`—and returns a binary localization mask only for locally tampered
images. The release contains the trainable components developed for MADL; it
does not redistribute the Qwen base model or SAM ViT-H.

## Files

| Directory | Component | Purpose |
|---|---|---|
| `agent_a_qwen_lora/` | Agent A LoRA adapter | Adapts Qwen2.5-VL-7B-Instruct to three-class semantic verification and region proposals. |
| `agent_b_dualstream/` | Agent B dual-stream checkpoint | RGB/SRM ConvNeXt-Tiny classifier and continuous manipulation heatmap. |
| `agent_b_visual_ranker/` | Agent B visual-ranker checkpoint | Selects a final SAM candidate using RGB, mask, heatmap, and geometry evidence. |
| `MANIFEST.json` | Integrity manifest | Records the byte size and SHA-256 of every released model artifact. |

The dual-stream release checkpoint contains only the model state, architecture,
class names, and a compact training summary. Optimizer state and local training
paths were intentionally removed. The visual-ranker checkpoint uses the
strict four-field runtime schema required by the companion loader.

## External dependencies

1. Download `Qwen/Qwen2.5-VL-7B-Instruct` through Transformers or the
   Hugging Face Hub. PEFT reads this identifier from `adapter_config.json`.
2. Download the official SAM ViT-H checkpoint
   `sam_vit_h_4b8939.pth` from Meta's Segment Anything release.
3. Place this repository under the runtime weight root and place SAM at
   `external/sam_vit_h_4b8939.pth`, or override the relative paths in the MADL
   YAML configuration.

Expected layout:

```text
models/
├── agent_a_qwen_lora/
├── agent_b_dualstream/
│   └── madl_agent_b_dualstream_v1.pt
├── agent_b_visual_ranker/
│   └── madl_agent_b_visual_ranker_v1.pt
└── external/
    └── sam_vit_h_4b8939.pth
```

Verify the files before inference:

```bash
python scripts/verify_weights.py /path/to/models
```

## Training data and evaluation scope

The released components were trained for the SID-Set task, which distinguishes
authentic, fully synthetic, and locally tampered images and provides masks for
the locally tampered class. The release does not include SID-Set images. Users
must obtain the dataset under its own terms.

## Intended use

The weights are intended for academic research on image forgery detection,
localization, evidence interaction, and multi-agent forensic reasoning. They
are not a substitute for human forensic examination and should not be used as
the sole basis for legal, disciplinary, or content-removal decisions.

## Limitations

- Very small, fragmented, or weak-trace manipulations remain difficult.
- Cross-dataset behavior has not been established by the final manuscript.
- Multiple model calls increase latency and GPU memory requirements.
- The natural-language field is a structured evidence trace; the release does
  not claim that it has been independently evaluated as explanation quality.
- Results depend on the exact Qwen, SAM, preprocessing, and threshold versions
  documented by the companion code.

## Licenses and attribution

The released MADL code and original weight packaging are provided under
Apache-2.0. Qwen2.5-VL, SAM, SID-Set, and other dependencies retain their own
licenses and terms. The [code README](https://github.com/shannima/MADL#third-party-components)
lists the relevant upstream projects. Dataset-derived use should preserve SID-Set attribution.