acceptodds
Under review as a conference paper at ICLR 2027

Not Just a Matter of Scale: Recovering Forensic Traces to Build a Forensic-Enhanced Vision Backbone

Abstract

Foundation models expose useful but uneven forensic evidence. We study whether a compact DINOv3-L backbone can recover evidence absent from its CLS readout, using DINOv3-7B only as a training-time teacher. Rank–nullity and orthogonal projection characterize this bottleneck and show when student patch features can reduce reconstruction error. We propose RELIC, which uses frozen teacher-relative residuals to guide persistent collectors across student layers; a decoder combines the collected evidence with the projected CLS into one Restored Token. A training-only Local Evidence-Aware Director (LEAD) further guides forensic readout. Rather than matching teacher layers directly, RELIC fixes a residual coordinate system first and then trains collection and readout in stages; a controlled ablation shows that this separation avoids unstable joint transfer. At inference, the teacher and LEAD are removed, leaving a representation usable for standalone classification or plug-and-play detector replacement. On AIGIBench Setting II, RELIC reaches 85.3% domain-macro accuracy, compared with 77.6% for AIDE, 82.3% for DGS-Net, and 75.5% for RA-Det with its original DINOv3-L interface; it reaches 91.4% on Chameleon. Probes and corruption tests further show that the recovered residual carries usable complementary evidence.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI-Generated Image Detection, Image Forensics, Knowledge Distillation, Forensic Representation Learning, Residual Evidence Recovery;Vision Foundation Models

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_vgstOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RELIC, a framework for exploiting local forensic cues that may not be adequately captured by a compact vision backbone’s global representation. It uses residual supervision from a frozen teacher, staged training, and cross-layer feature collection to construct an enhanced representation. The teacher and auxiliary guidance modules are removed at inference. The paper evaluates the approach through cross-dataset detection, downstream feature replacement, and corruption tests, examining its effectiveness, generalization, and deployment cost.

Critical strengths
  1. An efficient inference pathway for forensic representation learning. RELIC uses fixed teacher residual supervision and persistent inter-layer collection to produce a single Restored Token. The teacher, Residual Tokenizer, LEAD, and auxiliary readouts are discarded at inference. The forward-only measurements in Table 4, obtained under a common hardware and input protocol, provide direct support for the claimed modest inference overhead.
  2. Useful diagnostic evaluations. The domain-level, downstream-transfer, and corruption results report both improvements and regressions, helping establish the method’s operating limits. These analyses help identify the conditions under which the method is effective.
Critical weaknesses
  1. The independent contribution of the central mechanism is not sufficiently isolated. The theoretical orthogonal residual need not coincide with the calibrated training target. The final RELIC system is unmatched in Table 16, while Table 12 combines independently trained probes with the deployed classifier. These differences make it difficult to attribute the gains specifically to the proposed residual-recovery mechanism.
  2. Result traceability and deployment claims remain incomplete. Balanced accuracy is inconsistent with the class recalls in Tables 16–17, and the ForensicConcept summary differs between the main text and Table 7. The downstream interface mapping lacks necessary training details, while RA-Det uses per-domain optimal thresholds, limiting the scope of the plug-and-play claim.
Questions for authors
  1. Could the authors reconcile the balanced accuracy and class recalls in Tables 16–17, specifying the evaluation set, checkpoint, and aggregation used for each? Which ForensicConcept domain-macro results—the main-text summary or Table 7—correspond to the final evaluation protocol?
  2. Is a matched control available that starts from the same pre-aligned modules and differs only in whether the residual coordinates remain fixed during collector training, with the architecture, training data, and optimization budget otherwise matched?
  3. How is the Projected CLS + collector memory control constructed to match the stated input dimension and capacity of the probes in Table 12? Does the Restored Token’s AIGIBench advantage persist when it is evaluated using the same independent-probe training protocol as the intermediate representations?
  4. What are the parameterization, training objective, and training data for the RELIC-1024 mapping, and is this mapping shared across host detectors or fitted separately? How should the plug-and-play claim be qualified given RA-Det’s use of independently optimized per-domain thresholds?
Reviewer_sDAkOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper looks at AI-generated image (AIGI) detection and addresses a common bottleneck: why compact vision models (like DINOv3-L) tend to miss subtle forensic cues that are actually hidden inside their patch features. Instead of just scaling up the model, the authors propose a method called RELIC. RELIC uses a massive frozen teacher model (DINOv3-7B) strictly during training to establish a baseline residual target. Through a clever multi-layer "collector" and some auxiliary supervision during training, RELIC pools these local patch details and compresses them into a single "Restored Token". Once trained, the heavy teacher and training-only components are stripped away, leaving a lightweight, drop-in backbone that performs remarkably well on benchmarks like AIGIBench and Chameleon without slowing down inference.

Critical strengths

The authors don't just throw heuristics at the problem; they use linear algebra—like rank-nullity, orthogonal projections, and least-squares analysis—to mathematically prove why standard global CLS tokens miss forensic signals and how patch features can recover them.

Unlike methods that force you to retrain an entire pipeline, RELIC outputs a clean, single-token interface that can be dropped straight into existing detectors (like RA-Det or ForensicConcept) to instantly boost performance.

Critical weaknesses
  1. One question I’m curious about: why not directly use DINOv3-7B? It only has 7B parameters, so it does not consume much VRAM and has relatively low computational overhead. Also, reference [1] shows that simply training a linear layer on top of DINOv3-7B can achieve strong performance.

[1] Simplicity Prevails: The Emergence of Generalizable AIGI Detection in Visual Foundation Models

  1. And why is DINOv3-7B not included as a baseline in Table 2?

  2. There is still a noticeable performance gap compared with DINOv3-7B. Also, is there a mistake in Table 7.1? When RELIC is used as an alternative component, we can see clear performance drops across many datasets.

  3. Is the ordering of tables in the supplementary materials incorrect?

  4. Is figure 2 generated by AI? It has many weird mistakes!

Questions for authors

N/A

Reviewer_zNPUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RELIC, a forensic-enhanced vision backbone for AI-generated image detection. Using a frozen DINOv3-7B teacher during training, the method defines residual targets relative to projected CLS features of a compact DINOv3-L student. Persistent collector tokens aggregate patch information across multiple student layers, and a decoder combines the collected features with the projected CLS into a Restored Token. A training-only Local Evidence-Aware Director (LEAD) provides auxiliary forensic supervision. The teacher and LEAD are discarded at inference. Experiments on AIGIBench and Chameleon report improved detection performance, and the resulting representation can replace feature extractors in existing detectors.

Critical strengths
  1. The paper addresses a meaningful problem. Improving the forensic representations of a compact vision backbone without relying on a larger model at inference is relevant for efficient and generalizable AI-generated image detection.

  2. The proposed framework has a coherent overall design. It separates residual-target calibration from cross-layer patch collection and uses training-only auxiliary supervision while retaining a single-token inference interface.

  3. The evaluation is relatively extensive. It includes comparisons on AIGIBench and Chameleon, component ablations, representation probes, cross-detector feature replacement, corruption tests, and runtime measurements. The reported results suggest that the proposed representation can be useful in practice.

Critical weaknesses
  1. The theoretical analysis does not sufficiently justify the central forensic-recovery claim. Equations (1)–(3) characterize the limitations of low-dimensional readouts and show that additional patch features can reduce the least-squares reconstruction error of teacher features. However, these results do not establish that the teacher residual contains forensic-specific information, or that reducing its reconstruction error improves real-versus-generated discrimination. Moreover, the theoretical orthogonal residual and the calibrated residual used in training are not generally identical, as acknowledged in Eq. (2). The analysis also considers a feature span. In contrast, the actual collector is a nonlinear module trained with several objectives. The authors should clarify this equation and implementation.

  2. The ablations do not isolate the benefit of teacher-relative residual learning from ordinary multi-layer patch aggregation. In Table 3, adding the representation projector and collected token improves accuracy from 79.17% to 83.73%, while adding the residual decoder further improves it only to 84.19%. These results suggest that much of the gain may come from the collector rather than the proposed residual-recovery mechanism. The paper should compare against a collector with the same architecture, parameter budget, training schedule, and LEAD supervision but without residual targets. Simpler multi-layer attention pooling and direct patch-feature aggregation are also important controls, particularly given related patch-token-based forensic methods such as TAP.

  3. The claimed advantage of staged residual learning over joint transfer is not established by a controlled comparison. Section 3.2 and Table 16 emphasize the gap between joint transfer (54.52%) and staged RELIC (85.26%). However, Table 16 explicitly notes that staged RELIC is not one of the strictly matched ablation variants, because it uses pre-aligned modules and a separate multi-stage training schedule. Therefore, the difference cannot be attributed solely to the order of optimization or the use of stable residual coordinates.

  4. The paper does not directly verify that the collector reconstructs the intended teacher residual or captures forensic traces. The representation probes in Tables 12 and 17 show that the learned features are discriminative, but a high classification accuracy does not establish faithful residual recovery. Similarly, the attention-diversity statistics in Table 14 show that collector slots attend to different patches, not that those patches contain identifiable forensic evidence. The authors should report held-out residual reconstruction errors, cosine similarities, or explained variance, and examine whether these quantities correlate with detection improvements.

  5. The plug-and-play claim needs a more careful evaluation. The paper reports gains when RELIC replaces the feature interface of existing detectors without retraining their heads. However, the host evaluations use different decision rules: ForensicConcept uses a fixed threshold, whereas RA-Det uses independently optimized thresholds for each test domain. The latter is an oracle-style evaluation condition and does not directly demonstrate deployment on an unknown target domain. Moreover, the gains are not consistent across domains: RA-Det improves on 16 of 25 domains but regresses on 9, with substantial drops on StyleGAN-XL and StyleSwim. Could the author explain this?

  6. Several reporting inconsistencies and reproducibility issues should be resolved. Section 5.4 states that the ForensicConcept macro accuracy increases from 78.9% to 83.3%, while Appendix Table 7 reports 78.5% to 83.7%. Table 2 describes the Chameleon result as using a "historical Large Epoch-2 Simple deployment branch," making it unclear whether the same checkpoint as the main AIGIBench experiment is used. The main results are also largely reported for fixed checkpoints without variation across independent training runs, while bootstrap intervals over test images do not measure training variability.

Questions for authors

Please see the weaknesses above. In particular, the most important questions concern the gain attributable specifically to residual supervision under an otherwise matched collector, the direct evidence for forensic residual recovery, and a fair comparison of joint and staged training.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.