acceptodds
Under review as a conference paper at ICLR 2027

EvoKnow: Continual Knowledge Evolution for AI-Generated Image Detection

Abstract

AI-generated image detectors are commonly trained on fixed generator domains and become difficult to maintain as new generative models emerge. Continual adaptation is challenging because replaying historical generated images is costly, whereas updating shared parameters with limited current-domain data can overwrite prior forensic knowledge. We propose **EvoKnow**, a replay-free framework that formulates continual AI-generated image detection as forensic knowledge evolution. EvoKnow preserves a shared forensic basis learned from base domains, incrementally adds isolated residual experts for complementary generator-relevant evidence, and retrieves expertise through an Analytical Incremental Router (AIR) updated in closed form from current-stage generated images and accumulated sufficient statistics. Experiments demonstrate effective cross-generator generalization, few-shot expansion, and long-horizon continual adaptation. With ten generated images per arriving generator, EvoKnow achieves 96.70% average accuracy on non-base GenImage generators and 94.48% accuracy on Chameleon without target-benchmark adaptation. Under a strict replay-free continual learning protocol, EvoKnow achieves state-of-the-art continual learning performance, attaining 96.32% mean stage-wise accuracy and 4.32% average forgetting.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated Image Detection, Continual Knowledge Evolution

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fyq1OpenReview ↗
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
  1. 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.
  2. 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
  1. 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.
  2. 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
  1. What causes the differences between Tables 1, 5, and 10? Do they use the same checkpoint, adaptation samples, generator order, and evaluation split?
  2. Can the authors compare against sequential adaptation using the same backbone and data budget, including the 1,000 real images used at each stage?
  3. With the trained experts fixed, how does AIR compare with nearest-prototype routing in detection accuracy and forgetting?
7 comments
AuthorsOpenReview ↗

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.

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.

AuthorsOpenReview ↗

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.

GeneratorFinal detection accuracy (%)
SD v1.4 (base)99.20 ± 0.39
BigGAN96.39 ± 1.31
VQDM96.73 ± 1.53
SD v1.598.62 ± 0.48
Wukong99.67 ± 0.29
ADM93.39 ± 4.43
GLIDE96.66 ± 3.41
Midjourney93.81 ± 3.66
Mean96.81 ± 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 / MetricReleaseOmniDFAEvoKnow
Real--38.7398.32
HunyuanImage-3.02025-0997.5099.44
BLIP3o2025-0586.3957.08
GPT-4o2025-0386.3963.33
GPT-Image-1.52025-1269.1760.83
show-o2-7B2025-06100.0086.94
Seedream-4.52025-1275.2890.00
LongCat-Image2025-1295.8368.89
GoT-R1-7B2025-0596.6752.78
HiDream-I12025-0481.6724.44
Qwen-Image2025-0885.5642.50
BAGEL2025-0591.1147.50
Janus-Pro2025-0175.8392.08
Infinity-8B2025-0285.2830.00
FLUX-main2024-0863.0621.39
FLUX-klein2026-0167.7851.11
Qwen-Image-25122025-1225.8317.78
PixArt2024-0379.1734.86
OmniGen2-7B2025-0647.7860.00
Nano-Banana2025-1172.2222.41
Z-Image-Turbo2025-1145.0049.44
imagen-42025-0632.223.47
Gemini-2.0-Flash2025-0376.3922.50
SD-3.5-Large2024-106.6763.06
Mean attribution accuracy, excluding Real--71.4350.51
Mean attribution accuracy, including Real--70.0652.51
Chameleon (separate binary-detection evaluation)--83.4894.48

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.

Reviewer_7J6HOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3(+1)
Read review
Paper summary

EvoKnow addresses continual AI-generated image detection, aiming to adapt to emerging generators without replaying historical generated images while preserving existing detection capabilities. It freezes shared forensic features, trains independent LoRA residual experts for new generators, and selects experts using a recursive least-squares (RLS) router. It also incorporates structural and frequency-domain feature enhancement modules to improve detection performance. Experiments demonstrate competitive results in cross-generator detection, few-shot adaptation, and continual learning.

Critical strengths
  1. The paper addresses continual AI-generated image detection, aiming to adapt to emerging generators without replaying historical generated images while preserving prior detection capabilities. This problem has practical relevance.

  2. The method is well structured, freezing shared features and previous experts to prevent parameter interference while using analytical routing for replay-free updates.

  3. The method achieves competitive results in cross-generator detection, few-shot adaptation, and continual learning.

Critical weaknesses
  1. The core framework closely resembles Any-SSR [1], including freezing the early layers of the backbone, training independent LoRA experts, using RLS-based analytical routing, and continual learning without retaining historical samples. These are also the main contributions emphasized in this paper, but they have already been explored in Any-SSR. The proposed method largely transfers this framework from LLM continual learning to AI-generated image detection, with additional task-specific modules, offering limited methodological novelty.

  2. The current experiments do not sufficiently demonstrate the contributions of the structural and frequency modules. The authors should provide ablations removing GPB, LFB, AIR, and incremental experts, particularly in continual learning settings rather than only reporting static balanced accuracy.

  3. In Table 2, EvoKnow achieves 96.32% AA, compared with 95.64% for Wang et al. However, the methods use different backbones, training data budgets, and protocols. Moreover, most baseline results are taken from prior work rather than reproduced under consistent settings. Thus, the current evidence is insufficient to support the claimed SOTA performance under fair comparisons.

  4. The paper aims to select the most suitable expert based on forensic features. However, in Appendix A.5, samples from each new generator are assigned one-hot targets corresponding to the newly added expert. This effectively learns generator-associated routing rather than directly selecting the expert with the best detection performance. The authors should justify this design and compare it with simpler routing methods.

  5. Table 1 and Appendix Table 5 report substantially different results for GenImage incremental learning. For example, ADM accuracy is 96.21% in Table 1 but 83.85% in Table 5. The authors should clarify the differences in experimental settings that lead to this discrepancy.

[1] Any-SSR: How Recursive Least Squares Works in Continual Learning of Large Language Model.

Questions for authors
  1. Appendix A.2 states that each stage uses 10 generated images and 1,000 real images for training, resulting in a substantial class imbalance. Do the authors use balanced sampling or loss weighting? If not, what motivates this setting?

  2. In Table 10, Category-based outperforms the default Generator-based approach by 1.79 percentage points. Why is Generator-based still chosen as the default? Also, when generators from the same category arrive sequentially, how are the category-based experts updated? Does this require historical data or retraining existing experts?

  3. Although the method avoids replaying historical samples, each new task adds expert parameters, and the router stores additional statistics. The authors should report how model size, training time, and inference latency scale with the number of tasks, and evaluate the impact of reducing the routing feature dimension (e.g., to 1,024 or 2,048).

  4. More controlled comparisons are needed, using the same DINOv3 backbone, base detector, and 10-shot adaptation budget. The authors could also compare against few-shot adaptation methods such as FiSeR [1] to assess whether the additional complexity of incremental experts and routing provides clear benefits over simpler feature adaptation approaches.

[1] FiSeR: Fine-Grained Source Representations for Cross-Domain AI Image Detection.

8 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer 7J6H W1, W2 & W4

Thank for your valuable input.

W1, W2 & W4: Relationship to Any-SSR, continual ablations, and routing.

EvoKnow builds on Any-SSR's independent LoRA experts and RLS routing updates. EvoKnow's contribution lies in how it organizes shared knowledge and incremental experts and models complementary forensic evidence for continual detection. We learn and freeze a shared real/fake basis, then add residual experts combining GPB's structural cues and LFB's frequency cues. AIR retrieves complementary knowledge across experts, supporting cross-generator reuse under a common binary detection objective.

We add eight-stage GenImage ablations using identical DINOv3 initialization, SD v1.4 base-training data, and matched 10-shot samples. Removing GPB, LFB, or both lowers final AA by 0.90, 2.19, and 3.88 pp, respectively. Single-expert sequential fine-tuning also yields lower final AA and greater forgetting than the full model.

The one-hot targets provide generator-associated supervision as a proxy for expert relevance; they do not identify the detection-optimal expert for each image. At inference, AIR selects experts from image features without requiring generator identity. The full model achieves 96.31% final AA versus 92.82% for generator-associated expert selection, supporting retrieval beyond a fixed generator–expert assignment. The MoE alternative, which jointly retrains the router and existing experts, achieves 92.73%.

Summary reports mean AA across eight stages and final AF. Generator-associated selection uses a frozen, dedicated expert for each generator, yielding zero AF by construction. It therefore serves as a diagnostic reference and is excluded from the AF ranking. Only the best summary values are bolded.

Stage / GeneratorFull AA ↑Full AF ↓w/o GPB AA ↑w/o GPB AF ↓w/o LFB AA ↑w/o LFB AF ↓w/o GPB and LFB AA ↑w/o GPB and LFB AF ↓w/o AIR (MoE) AA ↑w/o AIR (MoE) AF ↓w/o AIR (Generator-associated expert) AA ↑w/o AIR (Generator-associated expert) AF ↓Base only (w/o incremental experts) AA ↑Base only (w/o incremental experts) AF ↓
SD 1.499.98--99.97--99.98--99.98--99.98--99.98--99.98--
BigGAN98.77-0.3196.630.0096.350.0096.050.0096.420.0096.410.0096.630.02
VQDM98.44-0.1595.76-0.0895.44-0.0994.75-0.0895.77-0.1395.210.0097.710.95
SD 1.598.75-0.0296.570.3096.300.3495.740.4196.680.1796.110.0087.850.59
Wukong98.930.0697.240.2297.020.2696.570.3197.300.1796.860.0087.691.15
ADM96.470.0195.53-0.2094.24-0.1892.67-0.2693.06-0.0292.990.0085.721.25
GLIDE96.530.0895.790.1494.530.1192.960.1093.220.2693.290.0085.951.18
Midjourney96.310.0695.410.1294.120.1092.430.2192.730.2492.820.0086.661.65
Summary98.020.0696.610.1296.000.1095.140.2195.650.2495.460.0091.021.65
AuthorsOpenReview ↗

Rebuttal for Reviewer 7J6H Q1, Q2, Q4, W1 & W5

W3 & Q4: Controlled continual learning comparisons.

Please see our response to Reviewer 1 (W2 & W4) for matched DINOv3-L/16, ProGAN-initialized, 10-shot comparisons. EvoKnow improves both mean AA and final AF over Wang et al. (R0). With FiSeR initialization, it achieves 90.77% accuracy on Chameleon versus 89.29% for frozen linear heads with voting, without target adaptation.

W5: GenImage results in Tables 1 and 5.

The two tables use different generator arrival orders during incremental learning, leading to different final results. We apologize for the ambiguity and will clarify the respective orders in the captions and Appendix B.1.

Q1: Class balance during training.

We use balanced sampling. The 1,000 real images form a reference pool, from which each stage samples ten real images to pair with ten generated images, maintaining a 1:1 class ratio. The larger pool allows real-image samples to vary across experts, providing more diverse examples of the real class. We clarify this procedure in Appendix A.2.

Q2: Generator-based and category-based experts.

We retain generator-based expansion as the default because replay-free sequential adaptation is a central objective: new experts learn from current-stage data while existing experts remain frozen. Reproducing the category-based setting under sequential arrivals would require revisiting historical samples and retraining the corresponding category expert. Moreover, generators with unclear categories may all be assigned to the Unknown expert, concentrating successive updates on the same expert.

AuthorsOpenReview ↗

Rebuttal for Reviewer 7J6H Q3

Q3: Resource costs and routing dimension.

We add resource measurements using DINOv3-L/16 on a single NVIDIA L20, with routing dimension d=10,000d=10{,}000. From one to fourteen tasks, saved model size, including AIR state, increases from 1280.16 to 1527.02 MB, and inference latency from 153.98 to 162.36 ms/image. Expert training takes approximately 86 seconds at the measured incremental stages, with mean AIR updates below one second. Latency is measured at batch size 1 with identical numerical precision; ± denotes standard deviation.

Observed tasksSaved model size (MB)Incremental expert training (s)AIR initialization or update (s)End-to-end inference (ms/image)
1 (base)1280.16--0.32 ± 0.04153.98 ± 0.18
21337.0085.93 ± 0.900.77 ± 0.74160.47 ± 0.18
81432.0185.64 ± 0.770.82 ± 0.81159.40 ± 1.76
141527.0285.80 ± 0.630.58 ± 0.03162.36 ± 0.36

In a separate 14-stage experiment, we vary the routing dimension while keeping the base model and expert bank fixed. The FP32 inverse-Gram matrix RR requires 4d24d^2 bytes. Reducing dd from 10,000 to 2,048 lowers its storage from 400.00 to 16.78 MB, while final AA decreases from 96.32% to 93.32%. Smaller dimensions therefore offer a storage–accuracy trade-off, but do not consistently reduce end-to-end latency.

Routing dimensionFP32 R size (MB)Mean AIR update time (s/stage)Final inference latency (ms/image)Final AA (%) ↑Final AF (pp) ↓
1,0244.190.367164.5092.875.15
2,04816.780.357167.7093.324.95
4,09667.110.495160.8694.174.83
8,192268.440.366168.4795.244.58
10,000400.000.459160.5196.324.32
Reviewer_RSzUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes EvoKnow, a replay-free continual learning framework for AI-generated image detection. The central objective is to adapt a detector to newly emerging image generators using very few generated examples while preserving its ability to detect images from previously encountered generators.

Critical strengths
  1. The need to adapt AI-generated image detectors as new generators emerge is well motivated. The proposed protocol explicitly excludes historical generated-image replay and storage of instance-level historical features, while allowing a fixed real-image reference pool. This is a useful and reasonably well-specified setting. Freezing previously learned components and adding isolated experts provides a straightforward mechanism for limiting parameter interference, while AIR allows the model to retrieve different forensic knowledge at inference time.
Critical weaknesses
  1. The cross-generator comparison in Table 1 uses previously reported results from methods trained under different source-domain protocols. In particular, many baselines use ProGAN as the source domain, whereas EvoKnow uses SD v1.4.

  2. A similar concern applies to the 14-stage continual comparison in Table 2, where the baseline results follow their respective reported protocols. The proposed method uses a strong DINOv3-L/16 backbone.

  3. EvoKnow avoids storing historical generated images, but it introduces a separate residual expert for each incoming generator in the default configuration. The number of experts therefore grows with the number of encountered generators.

  4. Since efficiency and maintainability are central motivations for replay-free continual detection, the absence of a direct comparison against parameter-efficient alternatives limits the practical significance of this work. Could EvoKnow be compared against sequential fine-tuning, frozen-backbone adaptation, and representative replay-free methods using the same backbone settings?

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer RSzU W1

Thank for your valuable input.

W1: Source-domain comparability

We apologize for the misleading labels. Revised Table 1 clarifies the source-domain protocols and adds ProGAN-initialized evaluations on UniversalFakeDetect and Chameleon.

We also compare EvoKnow with Wang et al. using the same DINOv3 backbone, ProGAN base-training data, and 10-shot adaptation samples. As shown below, EvoKnow achieves higher mean AA and lower final AF than Wang et al. both with and without replay.

R0/R500 denote 0/500 replay samples; EvoKnow operates without replay. AA is reported in percent and AF in percentage points. Summary reports mean AA over all 14 stages and final-stage AF. Only the best summary values are bolded.

Stage / GeneratorWang R0 AA ↑Wang R0 AF ↓Wang R500 AA ↑Wang R500 AF ↓EvoKnow AA ↑EvoKnow AF ↓
ProGAN99.98--99.90--99.97--
Deepfake87.330.0596.070.4195.554.44
BigGAN89.750.0095.201.1596.752.52
StyleGAN288.93-0.4294.671.0594.855.20
DDPM88.73-1.6991.720.7098.81-0.08
ADM88.43-2.4091.610.4496.242.85
DALL-E89.38-1.9891.540.4398.390.47
GLIDE89.33-1.6091.190.5596.302.92
SD 1.489.38-1.4190.640.6298.160.89
Midjourney87.01-0.5589.930.6997.391.37
VQDM87.42-0.3389.570.8496.082.89
SD 2.183.882.0486.632.1193.285.80
SDXL 1.081.813.1883.723.4292.406.58
SD 3.077.926.1381.395.4594.314.32
Summary87.816.1390.985.4596.324.32
AuthorsOpenReview ↗

Rebuttal for Reviewer RSzU W2 & W4

W2 & W4: Controlled continual learning comparisons

We add full-parameter sequential fine-tuning (10/100 shots) and frozen-backbone linear-head adaptation alongside Wang et al. (R0). The continual comparisons share DINOv3-L/16 initialization, ProGAN base-training data, and task order, with identical samples for 10-shot runs. EvoKnow achieves higher mean AA and lower final AF than these continual baselines.

FT 10/100 denotes sequential fine-tuning with 10/100 shots. Wang R0 and EvoKnow use 10 shots; R0 denotes no replay. AA is reported in percent and AF in percentage points. Summary reports mean AA over all 14 stages and final-stage AF. Only the best summary values are bolded.

Stage / GeneratorFT 10 AA ↑FT 10 AF ↓FT 100 AA ↑FT 100 AF ↓Wang R0 AA ↑Wang R0 AF ↓EvoKnow AA ↑EvoKnow AF ↓
ProGAN99.98--99.98--99.98--99.97--
Deepfake76.06-0.0275.46-0.0287.330.0595.554.44
BigGAN84.240.7283.760.6489.750.0096.752.52
StyleGAN278.551.6278.331.7588.93-0.4294.855.20
DDPM81.701.5581.601.4988.73-1.6998.81-0.08
ADM81.372.0282.382.7088.43-2.4096.242.85
DALL-E81.881.8081.851.9889.38-1.9898.390.47
GLIDE81.762.6583.192.1189.33-1.6096.302.92
SD 1.473.293.6973.773.1089.38-1.4198.160.89
Midjourney73.273.0773.343.0687.01-0.5597.391.37
VQDM72.394.9472.633.1087.42-0.3396.082.89
SD 2.173.525.5273.725.7883.882.0493.285.80
SDXL 1.073.966.0473.696.6981.813.1892.406.58
SD 3.061.737.0564.566.5677.926.1394.314.32
Summary78.127.0578.456.5687.816.1396.324.32

Frozen-backbone baselines use a separate 10-shot linear head for each generator and all-head voting on held-out Chameleon; EvoKnow uses AIR after incremental learning. EvoKnow improves accuracy on Chameleon by 17.38 pp with DINOv3 and 1.48 pp with FiSeR, which includes additional WildFake pretraining.

In the table below, “Frozen” denotes frozen-backbone linear-head adaptation. Each generator is evaluated using its corresponding head for frozen baselines and AIR selection for EvoKnow. Accuracy is reported in percent; mean accuracy averages the 14 generator results. Only the best summary values are bolded. † denotes additional WildFake pretraining.

Generator / MetricFrozen DINOv3-L/16EvoKnow–DINOv3Frozen FiSeR†EvoKnow–FiSeR†
Shots10101010
ProGAN84.15100.0099.0699.79
Deepfake49.8951.9958.1165.10
BigGAN70.8098.5899.0899.41
StyleGAN271.5698.0298.7798.02
DDPM54.4184.7399.5899.40
ADM59.4379.4799.6699.62
DALL-E73.4799.4599.7099.46
GLIDE62.1994.1397.6897.39
SD 1.470.2698.1299.3599.94
Midjourney56.3253.5797.1398.97
VQDM71.6396.2899.8397.85
SD 2.157.9272.1299.7199.69
SDXL 1.055.9067.5497.1898.96
SD 3.058.7459.8293.8895.76
Mean accuracy ↑64.0582.4295.6296.38
Chameleon accuracy ↑71.3388.7189.2990.77
Chameleon inferenceAll-head votingAIR routingAll-head votingAIR routing
AuthorsOpenReview ↗

Rebuttal for Reviewer RSzU W3

W3: Long-term scalability.

Expert storage grows with task count, while the backbone is shared and only the selected expert is executed at inference. From one to fourteen tasks, saved model size, including AIR state, increases from 1280.16 to 1527.02 MB, and latency from 153.98 to 162.36 ms/image on an NVIDIA L20. Expert training takes approximately 86 seconds at the measured incremental stages (2, 8, and 14). Routing costs still grow with the expert pool.

For long-term deployment, expert addition could be triggered by validated performance gaps. If historical data become available, periodic offline consolidation could produce a new base model and reset AIR and the expert pool after validating knowledge retention. These are prospective maintenance options beyond the replay-free protocol evaluated here.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised Manuscript Submitted

We have submitted a revised manuscript that corrects errors in the original submission and includes additional experiments. The changes are documented in the Summary of Revisions section of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer RSzU: Framework Contributions and Deployment Costs

We thank the reviewer for carefully reading our rebuttal and providing constructive feedback. We address the following two points to clarify your concerns and kindly ask you to reconsider your assessment of our work.

1. The EvoKnow Framework and the Contribution of Pretraining

EvoKnow is a continual-learning framework for AI-generated image detection, supporting different pretrained foundations, including DINOv3 and FiSeR. Its core contribution lies in organizing shared forensic knowledge and incremental residual experts, and modeling complementary evidence for continual detection. Representation learning, head adaptation, and routing are complementary, not competing mechanisms that jointly support knowledge acquisition, retention, and retrieval for the same real/fake detection objective.

FiSeR’s forensic pretraining provides a richer shared knowledge basis. Within the same EvoKnow framework, DINOv3 and FiSeR initialization yield mean accuracies of 82.42% and 96.38%, respectively. With sufficiently comprehensive shared knowledge, incremental experts focus on supplementing less-covered long-tail forensic features. Holding FiSeR initialization fixed, EvoKnow further improves mean accuracy over frozen-encoder linear-head adaptation from 95.62% to 96.38%, and Chameleon accuracy from 89.29% to 90.77%. These results are consistent with pretrained and incremental knowledge playing complementary roles.

2. Deployment Roles, Expansion Benefits, and Resource Costs

In deployment, the shared forensic module provides reusable knowledge, incremental residual experts capture evidence from new generators, and AIR retrieves relevant experts for each input. This division enables adaptation using current-stage samples while freezing existing experts, supporting continual learning without replay.

Sequential FT repeatedly updates the same model and risks forgetting existing capabilities. Using replay to mitigate this requires retaining historical data. Our FiSeR multi-head baseline uses all-head voting for images of unknown origin, with growing prediction-aggregation and maintenance costs as heads accumulate. Further adaptation remains necessary when existing knowledge provides insufficient coverage of new generators.

In the GenImage ablation, final AA is 96.31% for the full model, 86.66% for sequential fine-tuning without expert expansion, and 92.43% without both GPB and LFB. These results support expert expansion and complementary residual modeling.

At inference, each image requires shared feature extraction followed by only one selected expert’s forward pass. The number of expert forward passes stays fixed as the pool grows. From 1 to 14 tasks, inference latency on an NVIDIA L20 increases from 153.98 to 162.36 ms/image, while saved model size, including AIR state, increases from 1280.16 to 1527.02 MB. New-expert training takes approximately 86 seconds at the measured incremental stages.

For long-term deployment, knowledge clustering and expert consolidation could be explored to reduce expert count after validating retention, followed by further expansion as needed. This adjusts knowledge organization while preserving continual acquisition, retention, and retrieval.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.