- Rating
- 2
- Confidence
- 2(−1)
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper addresses the open-set generalization problem in AI-generated image detection, specifically focusing on "blind spots" where detectors trained on limited generators fail on unseen ones. The authors propose Generator Fingerprint Isolation (GFI), a training-free framework that augments frozen detectors with generator-specific evidence without modifying their parameters. The key idea leverages an empirical observation: removing principal components accounting for 95% of real-image variance preserves ~78.8% of generator-discriminative signal in the residual space. GFI constructs residual projections from real images, extracts normalized mean displacement fingerprints for each target generator, calibrates alignment scores against held-out real images, and fuses this evidence with the frozen detector's output.
Critical strengths
- Reframing detection failures as addressable "blind spots" rather than requiring full retraining is practically valuable and well-motivated. GFI requires no gradient updates to the base detector, making it lightweight and practical.
- The key empirical finding that generator-discriminative signal concentrates in low-variance directions after removing dominant real-image variation is interesting.
Critical weaknesses
- The paper does not analyze how many images are needed for reliable fingerprint estimation, nor the sensitivity to the quality/diversity of the fingerprint set. Additionally, the "post-hoc repair" setting assumes the user has identified which generator is the blind spot, this requires additional machinery to detect failures.
- The novelty is limited. The main conceptual that the empirical finding about signal retention in residual space is interesting, but the paper does not provide theoretical justification or deeper analysis of why this holds.
- The comparison is limited to fine-tuning baselines. Given the paper's focus on AIGCDetectBench, it should compare against state-of-the-art cross-generator detection methods.
Questions for authors
- Could you provide a deeper analysis of why removing the top-95% of real-image variance preserves most generator-discriminative signal? Is this specific to CLIP/DINO features, or would it hold for other encoders?
- How many self-generated images are needed for reliable fingerprint estimation? How sensitive is performance to the specific images used in the fingerprint set?
- For the generators that remain below 0.90 after GFI, what specific challenges exist?
- For generators where fingerprint alignment alone achieves 100% AUC (Table 2), what does this imply? Are these generators trivially detectable by any method?
- How does a user know a generator is a "blind spot" before registering it?
4 comments
Thank you again for your thoughtful review. As the discussion period nears its close, we wanted to gently follow up to check whether our response addresses your concerns. Please let us know if you have any further questions we can answer before the deadline.
Thank you for your detailed response. However, regarding Figure 1 and the writing style throughout the paper, I find that the logical progression is lacking and the tone is overly defensive. I suspect this paper may have been generated using automated research tools; therefore, I am maintaining my score but lowering my confidence rating to 3–2. I hope this assists the Area Chair in their assessment.
Thank you for taking the time to follow up. We appreciate your feedback on Figure 1 and the flow of the writing, and we will certainly work on making both clearer. At the same time, we want to respectfully push back on the suggestion that this paper was produced by automated research tools.
To share our side clearly: every part of the research—the core ideas, theoretical derivations, experimental setup, ablations, and analysis—was designed and run by us. Following the guidelines, we declared using AI only to help polish the phrasing of our draft, not to generate the work itself. Since the initial review focused entirely on the technical details, we were quite surprised by this new concern after the rebuttal. Could you share what specifically gave you that impression, or which parts of the logical progression felt disconnected? We would genuinely value understanding your reasoning.
Lastly, can you let us know whether our earlier responses answered your technical questions, or if there are any other doubts we can clarify?
We thank the reviewer for their feedback and address each concern below.
1. Signal Retention & Encoder Generality
Pretrained vision encoders like CLIP capture semantic content such as objects, scenes and lighting. Their dominant feature directions contain substantial natural variation across real images. Generator artifacts may also occur in these directions, but are harder to distinguish against this background. Importantly, artifacts persist in lower-variance directions, where real-image variation is smaller. Removing the top 95% of real-image variance (Eqs. 4–8, App. C.1) reduces this background while retaining 78.8% of generator-discriminative signal, as quantified by Eqs. 7–8 (Table 2, Fig. 3).
Two existing analyses support this:
We add two ViTs (AugReg, MAE) and use unbiased Crossnobis distances to avoid noisy overestimates. All four retain 73–89% of generator signal in 86–90% of directions:
Together, these results show that dominant-variance removal suppresses natural content variation while retaining generator-discriminative signals across encoders.
2. Sample Size and Sensitivity
We use 20 random subsets per sample size to assess sensitivity to image selection.
Just 10 images capture over two-thirds of the improvement. At 50–100 images, gains increase with low variability (SD ≤.008), showing consistent improvements across diverse generators. We use N=100 for experiments.
3. Why Does WFIR Remain Below .90 on Effort and SPAI?
AIGCDetectBench's WFIR pairs heavily web-compressed StyleGAN faces with Flickr real faces, making detection harder. The standalone fingerprint bank (no fused detector) tested on AIGCDetecBench achieves .71–.77 AUC (Table 11). GFI fusion (Table 3) combines fingerprint evidence with the base detector: a strong base like DDA pushes WFIR above .90 (from .87), while weaker bases also improve (Effort: .75→.84; SPAI: .58→.72), but they remain below .90 threshold. Augmenting enrollment samples to improve fingerprint robustness is a natural extension (7Vjw, #3).
4. What Does 100% AUC Mean?
Table 2 measures isolated fingerprint separability: whether a generator's fingerprint distinguishes its clean outputs from held-out real images. AUC 1.00 shows that this generator-specific signal is recoverable after semantic removal, not that existing detectors detect the generator reliably (Table 1). On benchmark images, the standalone fingerprint bank achieves only .71–.77 AUC for WFIR (Table 11), showing that perfect separability in this controlled evaluation does not imply perfect detection in practice.
5. Unknown Generator Blind Spots.
GFI does not require generator identities for enrollment. We demonstrate this by clustering unlabeled images in residual CLIP space, where images with matching fingerprint directions naturally group together.
Both methods recover generator-specific clusters: GMM-BIC estimates group count more accurately; eigengap achieves higher purity. Without isolation γ=0, purity drops to .857/.849, and both estimate six groups. Thus, GFI can discover generator-specific groups without knowing their identities in advance.
6. Comparison With Cross-Generator Detectors
Table 1 benchmarks seven published detectors across all AIGCDetectBench generators, revealing generator-specific blind spots across diverse approaches. GFI addresses these as a training-free repair layer for frozen detectors, rather than a standalone detector. We therefore compare it with fine-tuning, which can also repair blind spots but requires retraining. We select DDA, Effort and SPAI for their distinct detection mechanisms. Improvements across all three demonstrate GFI's applicability across detector types; it can also extend to the remaining Table 1 detectors.