- Rating
- 2(+1)
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper tackles open-set AI-generated image attribution. The authors formulate the problem as determining which known generative model produces an image or rejecting it as being generated by an unseen generator. The central claim is that existing holistic fingerprint methods fail on unknown sources, as new generators reuse patterns from known ones rather than producing entirely novel fingerprints. The authors propose Restricted Dynamic Hypergraph, which models each image as activating a sparse combination of learnable fingerprint atoms.
Critical strengths
Open-set AI-generated image attribution is an important and timely problem that deserves further investigation. I appreciate the authors for choosing a meaningful research topic. However, I have serious concerns about the problem formulation and the overall modeling of the task.
The results reported in Table 1 are significantly higher than those of the existing state-of-the-art methods. This substantial performance gap raises concerns about the fairness and reliability of the comparisons.
Critical weaknesses
AI-generated image attribution is inherently an open-set problem. Redefining AI-generated image attribution as a novel-class discovery problem is not well motivated. The proposed formulation does not address a fundamentally new problem, but instead introduces an unnecessary reformulation of an existing open-set attribution task.
The claim that “existing attribution methods typically represent each source using a holistic fingerprint or a global source-discriminative representation” is largely speculative. The paper does not provide sufficient empirical evidence to establish that this is the primary reason why existing methods struggle with unseen generators.
The authors report results on only a single benchmark, which is insufficient to substantiate this claim.
The paper does not provide a comparison of computational complexity or runtime with existing methods.
The proposed method appears to be highly dependent on the extracted feature representations, yet the paper does not evaluate its robustness to different backbones.
Questions for authors
My main concern is that the authors' redefinition of open-set image attribution is not reasonable, and I am also not convinced by the proposed model. The authors seem to have relied too heavily on GPT in conducting the research. Overall, I find the entire paper provides very little meaningful insight to me.
Dear Area Chair and Reviewers,
Thank you for your feedback. We summarize the additional work we carried out during the discussion.
In RdH, an atom is a learned vector in the same space as the image features. Different generators can activate the same atoms, and RdH uses their joint responses for attribution. To examine what these atoms capture, we changed one decoder operator at a time and compared reconstructions of the same image. We used a separate set of images to select the atoms and establish their expected response directions before evaluation. For example, when we replaced bilinear with nearest-neighbor upsampling, the joint response of five atoms shifted in the expected direction in 95% of paired comparisons on average, compared with 78% for the top-ranked atom alone. Across all four operator changes, individual atoms carried part of the evidence for the change, and their joint response captured it more consistently across decoder configurations (Appendix G.2, Table A.19).
We added four protocols to examine the effect of dataset origin on unknown-source rejection. In the original Cross-Era protocols, known sources came from GenImage and unknown sources from AIGI-Holmes. Two new protocols combined FLUX images from AIGI-Holmes and EvalGEN within the same unknown source. The other two swapped a known and an unknown generator so that GenImage and AIGI-Holmes each contributed both known and unknown sources. RdH achieved over 91% AUROC in each of the four protocols, although the swaps also changed the generator mix. We also examined how well RdH distinguished closely related generators. When SD3.5-Large was unknown, at most 0.20% of its test images were accepted as SD 1.5. When both were known, their bidirectional misclassification rate was 0.00% in the two swap protocols. The two unknown variants Janus-Pro-1B and Janus-Pro-7B proved harder to distinguish by their atom responses. They shared 27 of their 30 most active atoms, and cosine similarity was close to chance at determining whether two images came from the same variant (AUROC 0.55, Appendix G.1, Table A.17).
We also extended our analysis of test batches to single-query inference. RdH processes each new test image with 127 fixed reference images from known sources in the training set. With the same frozen checkpoints, mean AUROC across five protocols was 92.28% for individual queries and 92.63% for batches of 128 test images. On one H100 GPU, an individual query took 6.95 ms, compared with 1.09 ms per image averaged over a full batch. Batching was more efficient per image, while the fixed references allowed each query to be processed as soon as it arrived (Appendix H.2, Table A.21).
Thank you again for your time and consideration.