acceptodds
Under review as a conference paper at ICLR 2027

NOVAD: Variable Native Observation for AI-Generated Image Detection

Abstract

AI-generated images appear at widely varying resolutions and aspect ratios, yet fine-grained forensic cues are better preserved at native scale. Resizing large images can suppress these cues, motivating detectors to encode large images as local views at native scale. However, exhaustively processing all native regions makes inference cost grow with image area. This raises a natural question: can a detector preserve access to native detail without requiring full spatial coverage? We introduce **NOVAD** (**N**ative **O**bservation with **V**ariable **A**llocation for **D**etection), a detector that learns to operate on subsets of native local views together with a global view. During training, NOVAD is exposed to different numbers and spatial configurations of local views, teaching a shared detector to combine global features with local evidence from changing view subsets. These subsets reuse features encoded once per view, enabling multi-set supervision without repeated view encoding. Geometry-aware fusion combines local detail, whole-image context, and spatial information, allowing one model to handle changing view sets, resolutions, and aspect ratios. At inference, attribution from the global-only prediction guides local tile selection before local encoding, while the inference budget determines how many tiles are processed. The same model operates across inference budgets without retraining. Experiments show that NOVAD retains strong detection performance with only a few native local views. Trained on OpenFake, it achieves **98.44%** mean accuracy across Chameleon, CommunityAI, SocialRF, and WildRF, and **98.05%** unweighted mean accuracy across the eight resolution strata of HiRes-50K.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AIGI detection, High-Resolution detection, Feature aggregation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_enT3OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NOVAD, a framework for AI-generated image detection. The motivation is to retain the benefits of native-resolution evidence via selected non-resized patches, without exhausting all native-resolution observations. NOVAD fuses the features extracted from the global view and a variable number of native-resolution local views. The detector can adapt to different numbers of local views through the proposed multiple partial-view learning. During inference, the local views can be selected by the global attribution map. Experimental results indicate the strong performance of NOVAD on different in-the-wild benchmarks.

Critical strengths
  1. The proposed method is novel and utilizes the relations between global and native views.
  2. The generalization performance on in-the-wild datasets is strong, and the robustness against image transformation is impressive.
Critical weaknesses
  1. The necessity of the complexity of the proposed framework is not well justified by the ablation studies:
  • Tables 5-7 suggest that increasing the number of local views brings only a small improvement; it seems sufficient to use one global view with one local view.
  • Table 7 and Table 10 show small differences between different selection rules and geometry variants.
  • The effect of the adversarial JEPG regularization is not studied by an ablation experiment.
Questions for authors
  1. The paper emphasizes the significant role of native-resolution evidence and utilizes the local views obtained without resizing. Why is the proposed method extremely robust to rescaling, as suggested by Fig. 4(d)?
Reviewer_CjpKOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NOVAD addresses AI-generated image detection at varying resolutions by combining a global image view with a budget-controlled subset of native-resolution local views through geometry-aware global–local fusion. During training, partial-set supervision exposes the detector to different local-view subsets while reusing encoded features. At inference, attribution from the global-only prediction guides tile selection before local encoding, allowing the same model to operate under different observation budgets without retraining.

Critical strengths
  1. Flexible inference budgets. NOVAD addresses the practical cost of high-resolution image detection by selecting native-resolution local views rather than exhaustively encoding the entire image. A single trained model supports different observation budgets without retraining.

  2. Effective partial-set supervision and global–local fusion. Training on varied local-view subsets encourages reliable predictions from incomplete observations while reusing encoded features. The ablations support the benefits of partial-set supervision and incorporating global context directly into attention, particularly when few local views are available.

Critical weaknesses
  1. The comparison with HiDA-Net does not isolate the methodological contribution. NOVAD uses PE-Core-L14-336 with LoRA adaptation, whereas HiDA-Net uses a frozen CLIP ViT-L/14 backbone. These provide different pretrained representations despite both being ViT-L/14 models. Appendix C also states that the baseline results in Tables 2–4 are taken from published work rather than rerun under the training recipe of NOVAD. Consequently, the 14.33-percentage-point improvement over HiDA-Net on HiRes-50K under all-GenImage training could partly reflect differences in the backbone and adaptation strategy. The component ablations provide useful evidence, but a backbone- and training-protocol-matched comparison is needed to better isolate the contribution of NOVAD.

  2. The improvements over HiDA-Net on GenImage and DRCT-2M are modest. On GenImage (Table 3), NOVAD improves average accuracy by 0.7 and 0.5 percentage points without and with VAE augmentation, respectively. On DRCT-2M (Table 4), the gain is 0.4 percentage points in both settings. These small gains, obtained with a different backbone and adaptation strategy, provide less convincing evidence of a substantial accuracy benefit from the proposed method itself.

  3. The role of partial-set supervision in low-budget accuracy preservation is not clearly established. In Table 5, the full-pool-only baseline already maintains strong accuracy at a low local-view cap, achieving 97.95% at cap 1 and decreasing slightly to 97.51% at cap 16. Partial-set supervision improves accuracy by 0.46–0.93 percentage points, but these results do not establish that it is responsible for preserving accuracy when fewer local views are used. Figure 3 plots only the partial-set model. If the accuracy–latency curve of the full-pool-only baseline were plotted, it might exhibit a similarly favorable low-budget tradeoff. This would suggest that partial-set supervision primarily improves overall accuracy rather than specifically enabling accuracy preservation under reduced observation budgets. Table 6 instead suggests that incorporating global context directly into attention is a key factor in low-budget performance: at cap 1, this variant achieves 96.59% accuracy, compared with 79.81% without a global view and 77.35% with global context incorporated only through concatenation. Taken together, these results suggest that performance saturates at a small local-view cap. These results suggest that adding more local views provides little additional benefit, rather than showing that partial-set supervision is what allows the model to maintain accuracy with fewer views.

Questions for authors
  1. Backbone-controlled comparison. Could you compare NOVAD and HiDA-Net (Mu et al., 2026) using the same pretrained backbone, adaptation strategy, and training protocol, particularly on HiRes-50K? For example, evaluating NOVAD with CLIP ViT-L/14 or HiDA-Net with PE-Core-L14-336 under matched adaptation settings would help separate the contribution of NOVAD’s proposed mechanisms from differences in pretrained representations.

  2. Reliability of the reported improvements. Could you report results across multiple random seeds, including means and standard deviations, for the modest improvements on GenImage and DRCT-2M (Tables 3 and 4) and the local-view budget ablations (Tables 5 and 7)? This would help assess whether the observed differences are consistent rather than attributable to training variability.

  3. Motivation relative to prior work. HiDA-Net (Mu et al., 2026), titled No Pixel Left Behind, emphasizes preserving native-resolution detail through full spatial coverage at inference. Similarly, All Patches Matter, More Patches Better (Yang et al., 2026) emphasizes leveraging distributed patch evidence and avoiding over-reliance on a few regions. NOVAD instead processes a selected subset of native local views. Could the authors clarify how this selective-observation strategy relates to these prior motivations? Beyond computational savings, what insight supports the sufficiency of a small subset of local views, and under what conditions might full coverage still be necessary?

Reference

Mu et al. (2026). No Pixel Left Behind: A Detail-Preserving Architecture for Robust High-Resolution AI-Generated Image Detection. ICLR 2026.

Yang et al. (2026). All Patches Matter, More Patches Better: Enhance AI-Generated Image Detection via Panoptic Patch Learning. ICLR 2026.

Reviewer_gHMeOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

NOVAD addresses high-resolution AI-generated image detection under bounded inference cost by fusing a global view with a variable subset of native-scale local views. It trains a shared detector via partial-set supervision on reused view features and uses global-only gradient attribution to select tiles before local encoding. Evaluated on four in-the-wild benchmarks, HiRes-50K, GenImage, and DRCT-2M, an OpenFake-trained model reaches 98.44% four-domain mean accuracy and 98.05% on HiRes-50K while maintaining strong accuracy with few local views.

Critical strengths
  1. Selecting tiles before local encoding lowers mean latency from 1076.32 ms (full grid, 100% coverage) to 136.10 ms at budget 1 (14.25% coverage) while maintaining stable accuracy across budgets 1–16.
  2. Experiments cover in-the-wild, cross-generator, high-resolution (94.67% on HiRes-50K under all-GenImage training), and perturbed images, with partial-set supervision improving over full-pool training by 0.46–0.93 percentage points.
Critical weaknesses

1.Although the paper argues that resizing suppresses fine-grained artifacts, inference relies on gradients from a downsampled global view (capped at 576 patches) to locate local tiles, risking missed cues at high resolutions. In Table 7, Gradient: fake outperforms backward-free Global attention by only 0.06–0.19 percentage points and Random by 0.24–0.33 percentage points.

2.In Table 5, raising the budget from 1 to 16 (14.25% to 79.49% coverage) changes accuracy only from 98.41% to 98.44%, suggesting near-saturation from the fine-tuned PE-Core-L14-336 global view alone. However, Table 5 omits the OpenFake Global-only baseline, Table 2 only evaluates budget 8 on HiRes-50K without per-stratum budget curves, no backbone-matched comparison with HiDA-Net is provided, and 69D Fourier box encoding adds just 0.15% over raw coordinates (Table 10).

Questions for authors
  1. Across the eight strata of HiRes-50K, what are the accuracies for Global-only (budget 0), budgets 1, 4, 8, and the full grid?
  2. On the 3000–5000px and >5000px strata of HiRes-50K, how much advantage does Gradient: fake retain over Global attention and Random selection?
  3. Using the same backbone as HiDA-Net, how much improvement does NOVAD maintain on GenImage and HiRes-50K?

Comments on the paper

1 comment
Area_Chair_V2AeOpenReview ↗

A Gentle Reminder: About One Day Left in the Discussion Period

Dear Reviewers,

This is a gentle reminder that approximately one day remains in the discussion period. Please review the authors' responses and other reviewers' comments, and engage in the discussion to clarify which concerns have been addressed and which remain unresolved.

Thank you for your time and effort.

Best,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.