Research map / Vision-language and multimodal / Cross-modal learning
TranX-Adapter: Bridging Artifacts and Semantics within MLLMs for Robust AI-generated Image Detection
Wenbin Wang, Yuge Huang, Jianqing Xu, Yue Yu, Jiangtao Yan, Shouhong Ding, Pan Zhou, Yong Luo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add artifact encoder (NPR) features to the MLLM (ablation) | LLaVA-1.6-mistral 7B · GenImage | accuracy: +4.6% | Improved |
| add X-Fusion individually (ablation) | LLaVA-1.6-mistral 7B + artifact encoder · GenImage | accuracy: +2.3% | Improved |
| add TOP-Fusion individually (ablation) | LLaVA-1.6-mistral 7B + artifact encoder · GenImage | accuracy: +3.2% | Improved |
| combine X-Fusion and TOP-Fusion (ablation) | LLaVA-1.6-mistral 7B + artifact encoder · GenImage | accuracy: 91.9 | Improved |
| use TOP-Fusion in both transfer directions instead of the asymmetric design (ablation) | TranX-Adapter · Chameleon (Qwen3-VL 2B) | accuracy: 82.3 -> 78.9 | Hurt |
| use X-Fusion in both transfer directions instead of the asymmetric design (ablation) | TranX-Adapter · Chameleon (Qwen3-VL 2B) | accuracy: 82.3 -> 81.3 | Hurt |
| use cosine similarity cost matrix instead of JS divergence (ablation) | TOP-Fusion · GenImage (Qwen3-VL 2B) | accuracy: 88.0 -> 83.7 | Hurt |
| use KL divergence cost matrix instead of JS divergence (ablation) | TOP-Fusion · GenImage (Qwen3-VL 2B) | accuracy: 88.0 -> 86.8 | Hurt |
| use TranX-Adapter instead of direct feature concatenation | Qwen3-VL 4B · RRDataset | accuracy: 85.6 -> 90.9 | Improved |
| use TranX-Adapter instead of direct feature concatenation | LLaVA-1.6-mistral 7B · Chameleon (train all GenImage) | accuracy: 81.9 -> 85.1 | Improved |
| use TranX-Adapter with a VAE-based artifact encoder | LLaVA-1.6-mistral 7B + VAE · Chameleon | accuracy: 79.3 -> 83.8 | Improved |
| train TranX-Adapter (160M params) instead of LoRA or Adapter at the same size | LLaVA-1.6-mistral 7B · Chameleon | accuracy: 74.4 -> 75.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AIGI-Holmes — AIGI-Holmes: Towards Explainable and Generalizable AI-Generated Image Detection via Multimodal Large Language Models (ICCV 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unveiling Perceptual Artifacts: A Fine-Grained Benchmark for Interpretable AI-Generated Image Detection (ICLR 2026 · similarity 0.88)
- SALART-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated Images (NeurIPS 2026 · similarity 0.85)
- Breaking Semantic Artifacts for Generalized AI-generated Image Detection (NeurIPS 2024 · similarity 0.84)
- A Sanity Check for AI-generated Image Detection (ICLR 2025 · similarity 0.84)
- Divide, Conquer and Combine: A Training-Free Framework for High-Resolution Image Perception in Multimodal Large Language Models (AAAI 2025 · similarity 0.84)
- Lifting the Veil on Visual Information Flow in MLLMs: Unlocking Pathways to Faster Inference (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
