cs.CVDec 7, 2025

Align, Don't Blend: Orthogonal Supervision Disentanglement As the Key to Generalizable AI-Generated Image Detection

Authors: Ruoxin Chen, Jiahui Gao, Kaiqing Lin, Keyue Zhang, Yandan Zhao, Isabel Guan, Taiping Yao, Shouhong Ding

Organizations: Tencent Youtu Lab · East China University of Science and Technology · Shenzhen University · Hong Kong University of Science and Technology

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Recent advances in AI-Generated Image (AIGI) detection have largely focused on model architectures, loss design, and training strategies, while overlooking a more fundamental factor: how supervision is structured. Existing detectors typically train on benchmark-defined data that implicitly blends semantic and artifact signals, leaving supervision design unexplored. In this paper, we examine supervision structure through controlled experiments and show that such blended supervision is not merely suboptimal but fundamentally counterproductive. Replacing it with disentangled, orthogonal supervision - paired with inductively matched models - suffices to achieve strong cross-generator generalization. We formalize this as Orthogonal Supervision Disentanglement and instantiate it in AlignGemini, a two-branch detector in which a VLM receives purely semantic supervision and a vision expert receives purely artifact supervision. With disentangled supervision coupled with inductively matched models, even the simplest training recipe achieves a significant improvement of 9.5% across five in-the-wild benchmarks - trained on images from only two generators, without complex training strategies or diverse generator coverage. We further introduce AIGI-Now, a diagnostic benchmark covering eight contemporary generators, including five closed-source commercial models, with disentangled semantic and pixel evaluation subsets for fine-grained analysis of detector capabilities. Our code is publicly available at https://github.com/GzZning/Align-Don-t-Blend.

Figures & tables

Figure 1. Top: Performance comparison on five challenging in-the-wild benchmarks. AlignGemini consistently outperforms all existing detectors across all five datasets. Bottom: Comparison of training corpus complexity. AlignGemini relies on a substantially simpler training corpus—only two generators—yet achieves the strongest performance overall. Through orthogonal supervision disentanglement, AlignGemini demonstrates that supervision design, not data scale or generator diversity, is a critical and previously overlooked lever for generalizable AIGI detection.
Figure 2. Illustration of systematic blind spots of semantic and pixel-artifact detectors. Left: semantically-faithful synthetic images evade semantic detectors despite the existence of pixel artifacts. Right: heavy degradations (e.g., compression) destroy the generation pixel trace, reducing artifact detectors’ detection rate by 16%+, even for evidently implausible images.
Figure 3. A controlled study on the effect of supervision structure on models with different inductive strengths. "Semantic/Pixel Val" denotes images generated by the same generator used in training. "Semantic/Pixel Test" denotes images from unseen generators. Top: DINOv2 trained with semantic supervision generalizes poorly to unseen semantic inconsistencies, whereas pixel-artifact supervision yields strong generalization on pixel-artifact tests. Bottom: Qwen-VL trained with semantic supervision robustly detects semantic flaws from unseen generators, while pixel-artifact supervision fails to improve its sensitivity to low-level artifacts. These results reveal a clear inductive preference of each model toward different signal types. Rightmost column: Blended semantic–pixel supervision degrades each model’s inherent strengths. The results indicate that semantic and pixel subtasks align with different inductive biases. Joint supervision cannot reconcile this mismatch, whereas task–model alignment yields superior generalization.
Figure 4. Illustration of the pipeline. AlignGemini adopts a deliberately simple yet effective design by constructing two orthogonal supervision sets. For semantic supervision, a pretrained VLM generates DPO-style labels explicitly constrained to semantic cues. The VLM and the expert are then fine-tuned separately, aligning each model with its corresponding subtask to encourage complementary specialization. At inference, AlignGemini jointly evaluates semantic and pixel-level cues.
Figure 5. Comparison between AIGI-Now and existing benchmarks. Top: Existing benchmarks mainly rely on pre-2023 open-source generators. By contrast, real-world usage is dominated by newer commercial, closed-source models, creating a clear gap between the existing benchmark data and real-world fakes. Bottom: Our proposed AIGI-Now is constructed from the latest generators, including five commercial, closed-source models with unknown architectures, enabling rigorous evaluation of detectors’ cross-architecture generalization.
Figure 6. Ablation on the impact of supervision disentanglement. “Pure” denotes semantic-only supervision for the VLM branch and artifact-only supervision for the Expert branch, while “blend” applies combined semantic and pixel-level supervision. Pure supervision encourages stronger branch specialization and more effective complementarity, whereas blended supervision dilutes specialization and degrades detection performance.
Figure 7. Ablation on the effect of using two models across five in-the-wild benchmarks. Each subfigure reports the in-the-wild performance of a baseline method augmented with one branch of AlignGemini, resulting in capacity-matched two-model–to–two-model comparisons. AlignGemini consistently achieves non-trivial performance improvements, indicating that its gains are not solely attributable to the use of multiple models.
Figure 8. Robustness evaluation on AIGI-Now (pixel), which assesses robustness under common post-processing operations. Double JPEG compression applies the same JPEG quality factor twice in succession. Double resizing downsamples the image and then upsamples it back to the original resolution. AlignGemini demonstrates superior robustness across all post-processing.
Figure 9. The VLM branch of AlignGemini discriminates fine-grained semantic flaws despite being trained exclusively on obviously inconsistent images, suggesting that pure semantic supervision activates the VLM’s internal knowledge rather than teaching new discriminative patterns.
Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 10. Controlled studies on the effect of supervision structure on six models.
Figure 11. Overall pipeline for constructing the AIGI-Now benchmark.
Figure 12. Illustrative prompts used in AIGI-Now.
Figure 13. Representative AIGI-Now examples synthesized by Flux-Kera, Flux-Kontext, and Flux-Dev, illustrating the high visual fidelity and stylistic diversity of recent Flux generators.
Figure 14. Representative AIGI-Now examples generated by five recent closed-source models, showcasing diverse scenes and styles that increase the difficulty of robust AI-generated image detection.
Figure 15. Prompt templates used to construct DPO preference pairs.
Figure 16. System prompt.
MethodBLINKRealWorldVQAMMEMMT-Bench VAL
General MLLM
Qwen-2.5-VL-3B0.47500.658815900.6025
Qwen-2.5-VL-7B0.54810.675816770.5948
LLaVA-1.5-7B0.41710.542414360.4713
MLLM-based AIGI detector
FakeVLM0.37610.538512210.4445
Appendix
Table 11. Comparison of AlignGemini and baseline VLM-based detectors on general multimodal understanding benchmarks. Results are referred from ( Lin et al., 2025b ) .
Figure 17. Examples of general multimodal conversations. It indicates that the general capabilities of AlignGemini are largely preserved after adaptation for AIGI detection.
Figure 18. Qualitative examples of AlignGemini’s predictions on AI-generated images from the AIGI-Now benchmark.
Figure 19. Qualitative examples of AlignGemini’s predictions on images from the Chameleon benchmark.
Figure 20. Qualitative examples of AlignGemini’s predictions on images from the BFree-Online benchmark.
Figure 21. Qualitative examples of AlignGemini’s predictions on images from the CO-SPY-Bench (in-the-wild) benchmark.
Figure 22. Qualitative examples of AlignGemini’s predictions on images from the WildRF benchmark.
Figure 23. Qualitative examples of AlignGemini’s predictions on images from the AIGI-Bench benchmark.
More in this topic

AI-Generated Image Detection

AI-generated image detection identifies synthetic or AI-edited images and develops methods that generalize across generators, domains, and image transformations.

See all 194 papers

Explore similar work

CardsList

  1. Where Detectors Fail: Probing Generative Space for Generalizable AI-Generated Image Detection

    May 24, 2026Zijie Cao, Weijie Tu, Yao Xiao +3Synthetic Data AugmentationAI-Generated Image Detection

    School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Australian National University, Canberra, Australia · Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China +1

  2. Look Closer: Patch-wise Supervision for AI-Generated Image Detection

    Sep 29, 2026Zhida Zhang, Tao Wu, Siyu Liu +1AI-Generated Image Detection

    MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences (CASIA), Beijing, China · ShanghaiTech University, Shanghai, China · Anhui University, Hefei, China

  3. GlobalForge: Towards Robust AI-Generated Image Detection

    Jul 16, 2026Manni Cui, Ruiqi Liu, Dianyuan Zou +8Image Corruption RobustnessNeural Network Robustness

    1Huazhong University of Science and Technology · Institute of Automation, Chinese Academy of Sciences · 3Jilin University +1