Rapid advances in AI-generated image (AIGI) technology enable highly realistic synthesis, threatening public information integrity and security. Recent studies have demonstrated that incorporating texture-level artifact features alongside semantic features into multimodal large language models (MLLMs) can enhance their AIGI detection capability. However, our preliminary analyses reveal that artifact features exhibit high intra-feature similarity, leading to an almost uniform attention map after the softmax operation. This phenomenon causes attention dilution, thereby hindering effective fusion between semantic and artifact features. To overcome this limitation, we propose a lightweight fusion adapter, TranX-Adapter, which integrates a Task-aware Optimal-Transport Fusion that leverages the Jensen-Shannon divergence between artifact and semantic prediction probabilities as a cost matrix to transfer artifact information into semantic features, and an X-Fusion that employs cross-attention to transfer semantic information into artifact features. Experiments on standard AIGI detection benchmarks upon several advanced MLLMs, show that our TranX-Adapter brings consistent and significant improvements (up to +6% accuracy).
Figures & tables
Figure 1: Comparison between the previous fusion method and our TranX-Adapter . (a) Previous Method: Concatenates artifact ( e.g. , from NPR) and semantic features ( e.g. , from CLIP-ViT), resulting in uniform attention and weak interaction. (b) Our TranX-Adapter : Incorporates a lightweight bidirectional fusion mechanism that enhances feature interaction.
Figure 2: Comparison between the processed NPR input (a) and the CLIP-ViT input image 2 2 2 The image is sourced from BiasFree ( Guillaro et al., 2025 ) (b) , where NPR highlights local pixel interdependencies for synthetic image detection.
Figure 3: Distributional comparison of representational variances between CLIP-ViT and NPR: (a) variance of L2 norms and (b) variance of cosine similarities across image patches.
Figure 4: Comparison of cross-encoder interactions between CLIP-ViT and NPR. (a) Relative entropy of attention maps. Higher relative entropy values, approaching 1 , indicate that the distribution is closer to uniform. (b) Attention map for the direction CLIP→NPR , where the query originates from NPR features and the key and value are derived from CLIP-ViT features. (c) Attention map for the direction NPR→CLIP , where the query corresponds to CLIP-ViT features and the key and value are obtained from NPR features. (d) Relationship between the training loss and the information flow metrics S .
Figure 5: Overview of the proposed TranX-Adapter . Our TranX-Adapter consists of two complementary fusion modules. The Task-Aware Optimal-Transport Fusion (TOP-Fusion) aligns artifact and semantic feature predictions by computing JS divergence between their logits and transferring artifact features into the semantic features through optimal-transport, yielding an enhanced semantic feature F^sem . The X-Fusion module transfers semantic features into artifact features via multi-layer cross-attention, producing F^art . The fused representations are finally fed into the Large Language Model (LLM) for detection.
Method
Midjourney
SD v1.4
SD v1.5
ADM
GLIDE
Wukong
VQDM
BigGAN
Mean
F3Net ( Qian et al., 2020 )
50.1
99.9
99.9
49.9
50.0
99.9
49.9
49.9
68.7
DIRE ( Wang et al., 2023b )
60.2
99.9
99.8
50.9
55.0
99.2
50.1
50.2
70.7
PatchCraft ( Zhong et al., 2023b )
79.0
89.5
89.3
77.3
78.4
89.3
83.7
72.4
82.3
UnivFD ( Ojha et al., 2023 )
73.2
84.2
84.0
55.2
76.9
75.6
56.9
80.3
73.3
AIDE ( Yan et al., 2025 )
79.4
99.7
99.8
78.5
91.8
98.7
80.3
66.9
86.9
NPR ( Tan et al., 2024 )
89.8
90.7
90.7
84.6
90.3
90.7
87.0
81.8
88.3
Table 1: Cross-model accuracy performance on the GenImage Dataset. Accuracy (%) of different detectors (rows) in distinguishing real images from those produced by various generative models (columns). The best result are marked in bold . “ † ” indicates direct concatenation of artifact and semantic features following Zhou et al. (2025) .
Method
Training Set
SDv1.4
All GenImage
UnivFD ( Ojha et al., 2023 )
55.6
60.4
DIRE ( Wang et al., 2023b )
59.7
57.8
PatchCraft ( Zhong et al., 2023b )
56.3
55.7
NPR ( Tan et al., 2024 )
58.1
57.8
LLaVA-1.6-mistral 7B † ( Liu et al., 2024 )
69.4
81.9
Table 2: Cross-dataset accuracy performance on the Chameleon testset. “ † ” denotes the setting following Zhou et al. (2025) , where the artifact features and semantic features are directly concatenated.
Method
RRDataset
Ori.
Trans.
Re.
Avg.
MLLMs (Zero-shot)
GPT-4o ( Achiam et al., 2023 )
94.5
84.7
73.1
84.1
Claude-3.7-sonnet ( 35 )
89.9
83.8
73.9
82.5
Gemini-2-flash ( Team et al., 2023 )
85.3
74.8
71.8
77.3
Qwen2VL-72B ( Wang et al., 2024 )
59.9
56.4
59.8
58.7
Table 3: Cross-dataset accuracy performance on the RRDataset. The “ Ori. ” represents the Original, “ Trans. ” represents the Transmission, and “ Re. ” represents the Re-digitization. “ † ” indicates the direct concatenation of artifact and semantic features ( Zhou et al., 2025 ) .
Artifact Encoder
X-Fusion
TOP-Fusion
Accuracy
82.3
✓
86.0
✓
✓
89.3
✓
✓
90.3
✓
✓
✓
91.9
Table 4: Ablation study of different module in TranX-Adapter . The first row shows the results of LLaVA-1.6-mistral 7B. We integrate NPR, apply X-Fusion and TOP-Fusion separately, and then use the full bidirectional fusion strategy.
Method
Params. (M)
Acc.
Full
7261
76.8
LoRA ( Hu et al., )
40
69.1
160
74.4
Adapter ( Houlsby et al., 2019 )
40
69.4
160
72.5
TranX-Adapter
40
73.8
Table 5: Comparison with PEFT methods on Chameleon dataset , where “Params.” denotes the number of learnable parameters measured in millions.
Figure 6: Overview of JS divergence and the evolution of information-flow significance across training. (a) Distribution of JS divergence. (b) Illustrates the relationship between the training loss and the information-flow significance S . (c) Comparison of image and text information-flow significance and (d) layer-wise distribution of S across the LLM, where “Low” indicates layers near the input.
AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.