- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes EvoKnow, a continual learning framework for AI-generated image detection that does not retain or replay historical generated images or their instance-level features. It preserves a shared forensic representation and a base expert, adds isolated experts for newly arriving generators, and updates an Analytical Incremental Router (AIR) through recursive regularized least squares. Each expert combines LoRA adaptation with global structural and local frequency modeling under a multi-view learning objective. The default incremental setting uses ten generated images and 1,000 task-agnostic real images per stage. Experiments cover cross-generator generalization, few-shot expansion, external transfer, and a 14-stage continual detection benchmark, where the authors report 96.32% mean stage-wise accuracy and 4.32 percentage points of average forgetting.
Critical strengths
- A coherent framework for continual detection without historical generated-image replay. The paper addresses a practically relevant problem: extending detection coverage to emerging generators while preserving previously acquired capabilities without retaining historical generated images. Freezing the shared representation and previous experts prevents direct parameter interference, while accumulated sufficient statistics allow the router to be updated without replaying historical samples. This provides a well-motivated architecture for separating knowledge acquisition from knowledge preservation.
- The evaluation covers several relevant capabilities and includes informative component analyses. The experiments examine base-model generalization, few-shot expansion, transfer without target-benchmark adaptation, and continual learning over a 14-stage sequence. The reported improvements over the Base configuration suggest that incremental experts can provide useful information beyond the shared representation. Ablations of multi-view modeling, structural and frequency cues, auxiliary objectives, routing representations, and adaptation budgets help explain the framework’s behavior. Taken together, these results provide promising evidence for the approach, although the reporting discrepancies below require clarification.
Critical weaknesses
- The GenImage results need clarification. Tables 1 and 10 report 96.21% accuracy on ADM, while the final stage of Table 5 reports 83.85%. The corresponding BigGAN results also differ, at 93.77% and 97.18%. It is unclear whether these tables use different configurations or evaluation protocols. Since these results support the claimed benefit of incremental learning, the differences should be explained.
- The ablations mainly examine the detection architecture, while the contribution of expert expansion and AIR is less clear. A continual baseline using the same backbone and training data would help assess the benefit of the proposed strategy. For AIR, the feature probes and routing visualizations do not show whether it improves detection over a simpler router using the same experts.
Questions for authors
- What causes the differences between Tables 1, 5, and 10? Do they use the same checkpoint, adaptation samples, generator order, and evaluation split?
- Can the authors compare against sequential adaptation using the same backbone and data budget, including the 1,000 real images used at each stage?
- With the trained experts fixed, how does AIR compare with nearest-prototype routing in detection accuracy and forgetting?
7 comments
Thank you for your reply and the additional experiments. The clarification regarding the order of arrival of generators and the new controlled continual learning comparison have resolved some of my doubts. The consistency of experimental settings between Table 1 and Table 5, the differences in real-image budgets (10 vs. 1,000), and the independent contribution of AIR under fixed expert conditions compared to the nearest prototype routing remain unclear.
Response to Reviewer fyq1: Experimental Consistency across Tables 1 and 5
We apologize for the misleading wording in our previous response. Our explanation did not clearly convey that the differences arise from order-dependent sample selection. Both experiments use SD v1.4 pretraining, the same 10-shot training procedure, and the same global random seed. However, sampling follows the generator arrival order: as the random-number state advances, different orders select different samples for each generator. These samples affect both expert training and AIR's accumulated statistics. With per-generator samples and training randomness held fixed, changing only the arrival order does not change the final expert bank or AIR solution.
We supplement this clarification with experiments assessing detection stability and the source information captured by AIR.
Detection stability across random seeds
We repeat the Table 5 protocol with ten random seeds while keeping the generator order fixed. The following results report final detection accuracy as mean ± standard deviation across runs.
| Generator | Final detection accuracy (%) |
|---|---|
| SD v1.4 (base) | 99.20 ± 0.39 |
| BigGAN | 96.39 ± 1.31 |
| VQDM | 96.73 ± 1.53 |
| SD v1.5 | 98.62 ± 0.48 |
| Wukong | 99.67 ± 0.29 |
| ADM | 93.39 ± 4.43 |
| GLIDE | 96.66 ± 3.41 |
| Midjourney | 93.81 ± 3.66 |
| Mean | 96.81 ± 0.83 |
The overall detection accuracy remains high, although individual generators show greater variability, particularly ADM. These results quantify sensitivity to random sampling and training.
AIR attribution and detection on CoReBench
Order-dependent sample selection also changes the feature–label statistics used by AIR and can affect its routing decisions. We therefore examine source attribution as an auxiliary diagnostic of the generator information captured by the routing representation.
We evaluate attribution across 23 generator classes and the Real class on CoReBench [2]. EvoKnow uses 10-shot incremental adaptation, while OmniDFA [1] follows its respective training protocol.
| Class / Metric | Release | OmniDFA | EvoKnow |
|---|---|---|---|
| Real | -- | 38.73 | 98.32 |
| HunyuanImage-3.0 | 2025-09 | 97.50 | 99.44 |
| BLIP3o | 2025-05 | 86.39 | 57.08 |
| GPT-4o | 2025-03 | 86.39 | 63.33 |
| GPT-Image-1.5 | 2025-12 | 69.17 | 60.83 |
| show-o2-7B | 2025-06 | 100.00 | 86.94 |
| Seedream-4.5 | 2025-12 | 75.28 | 90.00 |
| LongCat-Image | 2025-12 | 95.83 | 68.89 |
| GoT-R1-7B | 2025-05 | 96.67 | 52.78 |
| HiDream-I1 | 2025-04 | 81.67 | 24.44 |
| Qwen-Image | 2025-08 | 85.56 | 42.50 |
| BAGEL | 2025-05 | 91.11 | 47.50 |
| Janus-Pro | 2025-01 | 75.83 | 92.08 |
| Infinity-8B | 2025-02 | 85.28 | 30.00 |
| FLUX-main | 2024-08 | 63.06 | 21.39 |
| FLUX-klein | 2026-01 | 67.78 | 51.11 |
| Qwen-Image-2512 | 2025-12 | 25.83 | 17.78 |
| PixArt | 2024-03 | 79.17 | 34.86 |
| OmniGen2-7B | 2025-06 | 47.78 | 60.00 |
| Nano-Banana | 2025-11 | 72.22 | 22.41 |
| Z-Image-Turbo | 2025-11 | 45.00 | 49.44 |
| imagen-4 | 2025-06 | 32.22 | 3.47 |
| Gemini-2.0-Flash | 2025-03 | 76.39 | 22.50 |
| SD-3.5-Large | 2024-10 | 6.67 | 63.06 |
| Mean attribution accuracy, excluding Real | -- | 71.43 | 50.51 |
| Mean attribution accuracy, including Real | -- | 70.06 | 52.51 |
| Chameleon (separate binary-detection evaluation) | -- | 83.48 | 94.48 |
EvoKnow is less effective at exact source attribution on this benchmark. This is consistent with its residual knowledge accumulation: experts learn complementary evidence for distinguishing generated images from real images, and related generators can benefit from the same expert. The selected expert therefore need not identify the generating model.
Despite this attribution limitation, EvoKnow achieves 94.48% detection accuracy on the held-out Chameleon benchmark, compared with 83.48% for OmniDFA under their respective training protocols. This 11% difference supports the usefulness of the accumulated forensic knowledge for binary detection.
References
[1] Shiyu Wu, Shuyan Li, Jing Li, Jing Liu, and Yequan Wang. OmniDFA: A Unified Framework for Open Set Synthesis Image Detection and Few-Shot Attribution. 2025.
[2] Ouxiang Li, Yuan Wang, Xinting Hu, Huijuan Huang, Rui Chen, Jiarong Ou, Xin Tao, Pengfei Wan, Xiaojuan Qi, and Fuli Feng. Easier Painting Than Thinking: Can Text-to-Image Models Set the Stage, but Not Direct the Play? ICLR, 2026.
Rebuttal for Reviewer fyq1
Thank for your valuable input.
W1 & Q1: GenImage results in Tables 1, 5, and 10.
Table 1 reports generalization and expansion results, while Table 5 tracks learning under a different generator arrival order, explaining their different final accuracies. Table 10 compares generator-based and category-based expert organization through per-generator test accuracies after the final incremental stage.
W2 & Q2: Controlled continual learning comparisons.
Please see our response to Reviewer 1 (W2 & W4) for matched sequential fine-tuning comparisons and Reviewer 2 (W1, W2 & W4) for the Base-only ablation without incremental experts. Within each 10-shot comparison, methods share the same backbone initialization and use identical samples: ten generated images and ten real images per stage. The 1,000 real images form the reference pool from which these ten real images are sampled.
W2 & Q3: Routing diagnostics and alternatives to AIR.
We cannot directly apply instance-level nearest-neighbor routing because historical feature vectors are not retained. Table 3(b) assesses feature separability, while Table 10 and Figure 3(a) motivate shared forensic knowledge and cross-generator expert reuse.
Please see our response to Reviewer 2 (W1, W2 & W4) for detection accuracy and forgetting comparisons with jointly trained MoE and generator-associated expert selection. AIR retrieves relevant experts from image features and executes only the selected expert, avoiding the all-head voting used by our frozen-backbone baselines.