TranX-Adapter:
Bridging Artifacts and Semantics within MLLMs for Robust AI-generated Image Detection

Wenbin Wang¹, Yuge Huang², Jianqing Xu², Yue Yu², Jiangtao Yan²,
Shouhong Ding², Pan Zhou³, Yong Luo¹

¹Wuhan University   ²Tencent YouTu Lab   ³Singapore Management University

News

TL;DR

🔧 Installation

1. Clone this repository and navigate into the codebase
   bash

   1git clone https://github.com/DreamMr/TranX-Adapter.git 2cd TranX-Adapter
2. Install Packages
   bash install.sh
   When the environment is created successfully, you will see:
   Conda environment name tranxadapter has been created🎉. Now you can run "conda activate tranxadapter"

📦 Preparation

1. Download the datasets
   • Training Data: GenImage, RRDataset, BFree (SD2.1_selfconditioned_origBG.zip (ai) and COCO_real_512.zip (real)).
   • Evaluation Data: GenImage, Chameleon, RRDataset
   • Our constructed VQA dataset
   Download the above data to ./Dataset, with the structure as follows:
   text

   1Dataset/ 2├── TranXAdapter-Dataset/ 3│ ├── training/ 4│ │ └── GenImage_Sdv1d4.jsonl 5│ │ └── GenImageAll.jsonl 6│ │ └── RRDataset.jsonl 7│ │ └── BFP.jsonl 8│ ├── evaluation/ 9│ │ └── Chamelon.tsv 10│ │ └── GenImage.tsv 11│ │ └── RR.tsv 12├── GenImage/ 13│ ├── ADM 14│ ├── test 15│ ├── ... 16├── Chameleon/test/ 17│ ├── 0_real 18│ ├── 1_fake 19├── RRDataset_final/ 20│ ├── original 21│ ├── redigital 22│ ├── ... 23├── RRDataset_original_train_val 24│ ├── train 25│ ├── val 26├── BFP 27│ ├── ai 28│ ├── real
2. Process Data
   Run python preprocess_data.py to replace the image paths in JSONL/CSV files with absolute paths.
   Note: You need to copy the MD5 values corresponding to the CSV files into DATASET_MD5 in ./VLMEvalKit/vlmeval/dataset/aigc_detection.py

🏋️ Training

1. Merge TranX-Adapter into the MLLM
   First, TranX-Adapter needs to be merged into the MLLM so that it can be directly loaded with from_pretrained(). We provide merge scripts (./llavanpr/merge_model.py and ./qwen3vlnpr/merge_model.py) as well as the merged models: DreamMr/TranXAdapter-LLaVA-next-mistral7B-v0, DreamMr/TranXAdapter-Qwen3VL2B-v0, DreamMr/TranXAdapter-Qwen3VL4B-v0
2. Start training Take training Qwen3Vl-2B on GenImage Sdv1.4 as an example:
   bash

   1cd ms-swift/scripts/training 2bash train_qwen3vl_Chameleon.sh
   i. If you want to train on RRDataset, you need to set the input image resolution to 512x512 (./ms-swift/swift/llm/template/template/qwen.py line 637 and ./ms-swift/swift/llm/template/templatellava.py line192).
   ii. We found that if the model is trained directly on GenImage Sdv1.4, the MLLM tends to overfit to the input image resolution. Therefore, we recommend training with real and fake images that have the same resolution. We use the BiasFree part (SD2.1_selfconditioned_origBG.zip and COCO_real_512.zip) to prevent the model from overfitting to image resolution. We recommend downloading the data from the official link.
   iii. We found that MLLM training converges quickly and also overfits rapidly. Therefore, we recommend using a checkpoint from the middle of training.

📈 Evaluation

1. Modify the LMUData in ./VLMEvalKit/scripts/run_task.sh
   You need to modify LMUData to the absolute path of Dataset.
2. Modify DATASET_URL and DATASET_MD5 in ./VLMEvalKit/vlmeval/dataset/aigc_detection.py.
   Replace DATASET_URL with the absolute path of the CSV file, and fill in DATASET_MD5 with the MD5 value computed earlier.
3. Run code
   bash

   1cd VLMEvalKit/scripts 2bash run_task.sh

📧 Contact

• Wenbin Wang: wangwenbin97@whu.edu.cn

🖊️ Citation

@inproceedings{wang2026tranx,
  title={TranX-Adapter: Bridging Artifacts and Semantics within MLLMs for Robust AI-generated Image Detection},
  author={Wang, Wenbin and Huang, Yuge and Xu, Jianqing and Yu, Yue and Yan, Jiangtao and Ding, Shouhong and Zhou, Pan and Luo, Yong},
  booktitle={Forty-third International Conference on Machine Learning},
  url={https://arxiv.org/abs/2602.21716}
}

🙏 Acknowledgement

• VLMEvalKit
• ms-swift