acceptodds
Withdrawn from ICLR 2027

EviSlot: Evidence Slot-Guided Vision-Language Learning for Generalizable AI-Generated Image Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

AI-generated images have become increasingly realistic and visually indistinguishable from real images, making generalizable AI-generated image detection a challenging task. Existing vision-language detectors usually rely on global visual representations and deterministic textual prompts, which may overlook localized high-frequency artifacts and provide limited flexibility under cross-generator distribution shifts. To address this issue, we propose EviSlot, an evidence slot-guided vision-language learning framework that jointly models global semantics, local forensic evidence, and diverse textual hypotheses. Specifically, we employ DINOv3 as the visual encoder to extract both CLS and patch-level features, and introduce a DCT-guided evidence mining module to identify frequency-abnormal regions with potential artifact cues. A learnable Evidence Slot module then aggregates local artifact cues from patch tokens with frequency-aware attention guidance and fuses them with the global representation for robust visual evidence modeling. On the textual side, frequency-aware probabilistic prompting generates multiple real/fake prototypes from shared, image-specific, and class-specific prompt components. The final prediction is obtained by image-text matching and prompt-ensemble inference. The proposed framework provides a compact, flexible, and robust detection paradigm for generalizable AI-generated image detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-generated image detection, Generalization across generators, Vision-language learning, Frequency-aware evidence mining, Probabilistic prompting.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_f6UcOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper targets generalizable AI-generated image detection. EviSlot uses a frozen DINOv3 ViT-L/16 to extract CLS and patch tokens, computes a band-weighted DCT energy score per patch, and injects that score as an additive attention bias into cross-attention between four learnable “evidence slots” and the patch tokens. The pooled slot output is fused with the CLS feature, passed through an MLP adapter, and projected into CLIP space. On the text side, shared, private, and class-specific prompt tokens receive biases from probabilistic flow modules, giving an ensemble of real/fake prototypes. Prediction is cosine matching with prompt ensembling.

Critical strengths
  1. The evaluation set is well chosen, and one of its findings is genuinely useful. Most work in this area still evaluates on GAN-era generators or early diffusion models, while testing on AIGI-Now’s nine recent generators is a harder and more current setting. The resulting observation is worth reporting: PPM-CLIP reaches 99.40–99.50% on Nano and flux-dev yet falls to exactly 50.00% on flux-pro and jimeng, and NPR shows the same pattern. High in-domain accuracy therefore gives little information about which unseen generators a detector will handle, and averaging hides single-generator collapse. The paper makes this point explicitly (Section 4.2) and frames its own contribution around stability rather than per-generator wins, which is the right framing.

  2. The design intuition is sound, and the frequency prior is used in a sensible form. Treating patch-level DCT energy as a soft attention bias (Eq. 3) rather than a hard top-k patch selection is a reasonable choice: it keeps the pipeline differentiable, lets the model down-weight the prior where it is uninformative, and avoids discarding patches outright when frequency responses are noisy. The stated motivation that local high-frequency traces and global semantics are complementary, is well supported by the prior work the paper cites.

Critical weaknesses
  1. The paper’s own AP numbers contradict its main claim.

Section 4.4 calls DCT Evidence “the most critical component,” because accuracy drops sharply when it is removed. But Table 2 also reports AP, which doesn’t depend on a threshold, and AP goes the other way: on AIGI-Now, the model without DCT Evidence gets 95.85 AP, while the full model gets 93.86. The component the paper calls most critical makes ranking quality worse, not better. The r_acc/f_acc columns show what is really happening. Without DCT Evidence, the model gets 55.97 r_acc and 100.00 f_acc on GenImage, and 63.22 / 98.71 on AIGI-Now. It is calling almost everything fake at the fixed threshold. So the module is moving the decision boundary, not improving how well real and fake are separated. Section 4.5 shows the same thing again. At fusion weights 0.3 and 0.5, AIGI-Now accuracy drops to exactly 50.00% while AP stays high, and the authors attribute this to the threshold being poorly calibrated. The paper notices this twice but never asks what it means for the main result.

This matters because Table 1, which is the whole comparison with prior work, reports accuracy only. No baseline has an AP number, even though Section 4.1 lists AP as a main metric. Several baselines sit at exactly 50.00 on flux-pro and jimeng, which is what single-class collapse looks like. So the 88.62% vs 75.77% gap might just mean EviSlot happens to be better calibrated at the same fixed threshold, not that it tells real from fake any better. Without AP for every method, or accuracy compared after calibrating each method’s threshold separately, the main empirical claim doesn’t hold up.

  1. The backbone was changed without controlling for it, and that alone could explain the gains.

EviSlot uses DINOv3 ViT-L/16. The CLIP-based baselines use CLIP ViT-L/14, and the rest are much smaller CNNs. DINOv3 is newer, stronger, and pretrained differently, and the paper itself cites work showing that self-supervised foundation features transfer well for this task. Nothing in the paper separates the backbone from the proposed modules. There is no DINOv3 linear probe, no version of EviSlot running on CLIP, and no baseline rerun on DINOv3 features. Table 2 removes Evidence Slots, DCT Evidence, and Evidence Fusion, but never the encoder, so even the worst variant still gets the upgrade. The one configuration that would answer the question, DINOv3 with none of the proposed modules, is missing. The claimed contributions are modules stacked on top of a backbone swap, and the simplest explanation for beating PPM-CLIP by 12.85 points is the swap. Put together with the first point, the paper cannot rule out that the gains come from a stronger encoder and from better threshold calibration. Neither is what the paper claims to contribute, and telling them apart needs new experiments, not a clarification.

Questions for authors
  1. On AIGI-Now, the variant without DCT Evidence reaches 95.85 AP while the full model reaches 93.86 (Table 2). Since AP does not depend on a threshold, does this mean the evidence modules improve calibration rather than discrimination? Please report AP for every baseline in Table 1, and accuracy after calibrating each method’s threshold on a held-out split.
  2. What does a linear probe on frozen DINOv3 features achieve under the same training protocol, and what does EviSlot achieve with a CLIP ViT-L/14 visual encoder? Without these two numbers, how should the 12.85-point margin over PPM-CLIP be attributed between the backbone and the proposed modules?
  3. Please report the standard GenImage cross-generator protocol: train on SDv1.4, test on all eight subsets. How does EviSlot compare to the published numbers for AIDE, DRCT, C2P-CLIP, and Li et al. (2026), all of which are cited but not compared against?
  4. Why is EviSlot’s GenImage accuracy 98.51 in Table 1 but 97.58 for the full model in Table 2?
Reviewer_Q84ZOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work adds DCT-biased evidence slots to DINOv3 features and matches them against probabilistic CLIP text prototypes. Trained on GenImage-SDv1.4, it reports 88.62% mean accuracy on nine AIGI-Now pixel subsets, surpassing PPM-CLIP.

Critical strengths

S1. The result in Table 1 shows large gains over the listed baselines on flux-kera, flux-pro and MiniMax (> 25 points). Removing slots lowers target accuracy by 3.36 points and AP by 2.40 points (Table 2), which directly supports the introduced slot module. S2. Clean graphics.

Critical weaknesses

W1. The novelty seems limited when compared to PPM-CLIP CVPR Poster. Compared with which, EviSlot replaces the CLIP ViT with DINOv3 and adds DCT-biased slots. Most losses and settings are reused from the work. DINOv3 is a popular choice for in the NTIRE 2026 AI-generated image detection challenge (CVPRW 2026), and thus the reviewer believes that an ablation using DINOv3 ViT-L/16 + linear should be conducted and the result should be reported. W2. Removing DCT raises target AP from 93.86% to 95.85% while accuracy drops by ~8pp. This, together with Figure 4's one-class collapse to 50% at larger fusion weights, points to a bad threshold rather than the robustness claimed in Section 4.4.

Questions for authors
  1. What causes the difference between the headline numbers, EviSlot full model on GenImage, in Table 1 and 2?
  2. AIGI-Now is not cited, but Section 4.5 uses it to pick the defaults. It is unclear whether the target set was untouched.
  3. Section 3.5 has weird wordings such as "This grouped formulation keeps the objective concise while remaining consistent with the implementation", as if explaining the formulation based on the code. Modifications are recommended.
Reviewer_NbCAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an AI-generated image detection framework that combines a DINOv3 visual encoder, DCT-guided Evidence Slots for local feature aggregation, and a CLIP-based probabilistic prompting module. Experimental results show that the method generalizes well to unseen image generators on the evaluated benchmarks.

Critical strengths

The AIGC detection method proposed in the paper is effective against unknown generators. The combination of DCT-guided local evidence aggregation and probabilistic prompting is conceptually reasonable.

Critical weaknesses

(1) The paper lacks sufficient explanation and validation regarding the specific details of the proposed methods. For instance, the "DCT-guided patch contrastive loss" and "evidence slot diversity regularization" are not described with adequate implementation details. Furthermore, the paper does not include dedicated ablation studies to evaluate the individual contributions of these components.

(3) The comparison with recent state-of-the-art methods is incomplete. In particular, the paper omits AIDE, a closely related method that also combines DCT-guided local features with global semantic representations.

(2) How is the probabilistic prompting module implemented?

Questions for authors

Please refer to questions 1 and 2 under "Critical Weaknesses*."

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

Related papers

Discussion (0)

Sign in to comment.