- 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
-
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.
-
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
- 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.
- 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
- 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.
- 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?
- 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?
- Why is EviSlot’s GenImage accuracy 98.51 in Table 1 but 97.58 for the full model in Table 2?