acceptodds
Under review as a conference paper at ICLR 2027

DINOv3 Beats Specialized Detectors: Revisiting Image Manipulation Localization from a Foundation Model Perspective

Abstract

Image manipulation localization (IML) identifies regions that have been spliced, copied, removed, or otherwise edited. Its labels describe provenance rather than semantic identity, so existing systems rely on specialized forensic cues. Frozen DINOv3 features, despite strong readouts on general vision tasks, localize manipulations poorly. We ask whether a general backbone can make editing evidence readable without a dedicated forensic front end. A stronger decoder helps only in part: a global Transformer head improves frozen features, but adapting the encoder yields a far larger gain. We call this contrast the forensic readability gap: frozen features contain editing information, but it is difficult to read with one linear rule shared across images. We therefore propose DINOv3-IML, a simple QKV-LoRA adaptation with a shallow convolutional head and no forensic feature extractor. Adaptation makes editing evidence readable through a shared linear rule and aligns editing directions across images; once the encoder is adapted, a global head brings no further gain. DINOv3-IML reaches 0.85 and 0.77 pixel F1 under the CAT-Net and MVSS protocols, surpassing the best prior methods by 12 and 24 points, respectively, and establishes a simple yet strong baseline for IML.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image manipulation localization, image forensics, foundation models, DINOv3, representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Dx8jOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DINOv3‑IML for the IML task. It performs lightweight adaptation on the DINOv3 backbone via QKV‑LoRA without deploying a dedicated forensic feature extractor. The proposed concept of forensic readability gap is defined as follows: frozen DINOv3 features already contain tampering cues, yet these cues cannot be retrieved by cross-image shared linear rules. After LoRA adaptation of the encoder, the feature directions of tampered regions and background become aligned across different images, which substantially improves cross-image linear readability, enabling a shallow head to accomplish localization. Evaluated under two benchmark protocols of CAT-Net and MVSS, the method achieves pixel F1 scores of 0.85 and 0.77 respectively, outperforming existing approaches by a considerable margin.

Critical strengths
  1. The detector for IML can be efficiently fine-tuned via lightweight QKV-LoRA adaptation of the DINOv3 encoder alone.
  2. The concept of forensic readability gap is put forward, revealing that even equipped with a powerful decoder, a frozen backbone cannot bridge this gap; instead, internal patch interactions within the encoder require modification.
  3. It surpasses conventional dedicated tampering localization models by a large margin under the CAT-Net and MVSS evaluation protocols.
Critical weaknesses
  1. The adopted QKV-LoRA is not custom-designed for forensic tasks. The proposed framework is theoretically transferable to other dense prediction tasks, with no task-specific operators, artifact extraction modules tailored for image forensics, nor incorporation of forensic priors such as noise patterns and compression artifacts. Despite the prominent performance gains observed in experiments, such improvements stem from representation reshaping of the pre-trained backbone driven by task-specific data, rather than forensics-oriented architectural innovation.

  2. The base model is inherently trained for semantic tasks. However, the paper fails to elaborate on how it is transferred to learn forensic features, and what exactly these forensic features correspond to (e.g., boundary noise, JPEG compression artifacts, illumination inconsistency, or other forensic cues), leading to limited interpretability.

  3. Although the datasets adopted in experiments are standard benchmarks, they are relatively outdated and do not cover the latest forgery types, such as diffusion-model-based local generation and full-image reconstruction.

Given these notable limitations, I assign a negative score at this stage and look forward to the authors' rebuttal.

Questions for authors

Please refer to the weaknesses.

3 comments
AuthorsOpenReview ↗

Response (1/2): W1 – forensic priors

Thank you for the review and for keeping an open mind for our rebuttal. Following your comments, we have added experiments on forensic priors, the cues the model learns and newer forgery types. We answer the three weaknesses in order.

W1 (no forensic-specific design). "The adopted QKV-LoRA is not custom-designed for forensic tasks ... nor incorporation of forensic priors such as noise patterns and compression artifacts. Despite the prominent performance gains observed in experiments, such improvements stem from representation reshaping of the pre-trained backbone driven by task-specific data, rather than forensics-oriented architectural innovation."

Our contribution is not a forensic module but evidence on whether localization needs one: an adapted general-purpose encoder, with no forensic prior built in, outperforms detectors designed around such priors. Three results support this.

1. Our model outperforms detectors built around the priors you mention. Noise patterns are modelled by Noiseprint++ in TruFor and by the noise branch of MVSS-Net, and compression artifacts by the DCT stream of CAT-Net. Our model has none of these modules and outperforms all three under the CAT-Net protocol, also at a comparable parameter count:

MethodEncoderForensic moduleParametersMean F1
MVSS-NetResNet-50noise branch151M0.536
CAT-NetHRNetJPEG DCT stream117M0.601
TruForSegFormer MiT-B2Noiseprint++69M0.627
OursDINOv3 ViT-B + LoRA r32none90M0.764
OursDINOv3 ViT-L + LoRA r32none312M0.847

It also leads the strongest baseline of each protocol, by 12.2 points under the CAT-Net protocol and 24.0 points under the MVSS protocol (Tables 1 and 2).

2. Adapting the encoder lets it learn these cues from task data, so no module has to supply them. As you note, the gain comes from adapting the pre-trained encoder: with a similar number of trainable parameters, it is far more effective than a larger head on frozen features (Table 3, all models trained on CASIAv2):

EncoderHeadTrainable parametersMean F1
Frozen DINOv3 ViT-LTransformer8.93M0.467
DINOv3 ViT-L + LoRA r32Conv39.05M0.785

What the encoder learns includes the cues that forensic modules are built to extract. As our answer to W2 shows, the adapted encoder responds to noise and compression differences between a region and its surroundings, much more strongly than frozen features.

3. In our test so far, adding a DCT branch to our model did not improve it. We added a DCT branch to the adapted encoder: fixed 8×8 DCT filters followed by a small CNN, fused with the encoder features by cross-attention. Both models below were trained on CASIAv2:

EncoderHeadForensic branchCASIAv1ColumbiaNIST16IMD2020Mean
DINOv3 ViT-L + LoRA r32FPNnone0.8820.9210.6350.6900.782
DINOv3 ViT-L + LoRA r32FPNDCT0.8590.9060.6280.6850.770

Experiments with three other common forensic extractors, Bayar convolution, SRM filters, and TruFor's Noiseprint++, are still running, and we will post the results here within two days.

AuthorsOpenReview ↗

Response (2/2): W2 – forensic cues; W3 – newer forgery types

W2 (which forensic features). "The base model is inherently trained for semantic tasks. However, the paper fails to elaborate on how it is transferred to learn forensic features, and what exactly these forensic features correspond to (e.g., boundary noise, JPEG compression artifacts, illumination inconsistency, or other forensic cues)."

The adapted model uses the noise, compression and illumination cues you mention, and adaptation teaches it to read them as differences between a region and its surroundings. We tested this on 248 authentic photos from COVERAGE and CASIAv1. Each photo receives a region shaped like a real forgery mask, and one trace is applied inside the region, outside it, or to the whole image. In every condition, a 32-pixel gray band is painted over the region's outline, so the trace leaves no visible edge. We report the share of the region's tokens that the model flags as edited.

1. Noise, JPEG compression, blur and resampling are read as differences between a region and its surroundings. For the CAT-Net-protocol model, which flags 0.093 without a trace:

TraceInside onlyOutside onlyEverywhere
Noise σ = 100.1690.6210.114
JPEG q500.2110.1860.114
Blur σ = 20.5020.1790.110
Resampling to 50%0.3270.1450.094

A trace on one side raises the flagged share, whether it is applied inside the region or only outside it. In the second case the region itself is untouched; it is flagged because it now differs from its surroundings. The same trace over the whole image barely changes the share.

2. Illumination changes are read mainly at the region's outline. In a separate test on 237 photos, changing the region's brightness with a gamma of 1.4 raises the AP of the model's output for that region from 0.32 to 0.47 when the outline is covered, and from 0.11 to 0.80 when the outline is visible.

3. Reading traces as differences is learned during adaptation. Frozen features respond to the trace itself: with blur, their flagged share rises when the whole image is blurred, but not when only the surroundings are. Both adapted models show the opposite, including the one trained without any augmentation:

EncoderTraining dataNo traceInside onlyOutside onlyEverywhere
Frozen ViT-LCASIAv20.0750.3030.0800.136
ViT-L + LoRA r32CASIAv2, no augmentation0.0220.3900.1830.013
ViT-L + LoRA r32CAT-Net protocol0.0930.5020.1790.110

Sec. 5 shows how adaptation changes the features: a single linear rule shared across images reads edited tokens with AP 0.501 on frozen features and 0.919 after adaptation, and the edited-versus-authentic direction becomes shared across images, with alignment rising from 0.21 to 0.78. We will add these tests to App. H.

W3 (newer forgery types). "Although the datasets adopted in experiments are standard benchmarks, they are relatively outdated and do not cover the latest forgery types, such as diffusion-model-based local generation and full-image reconstruction."

Evaluated zero-shot, our CAT-Net-protocol model reaches pixel F1 0.684 on CocoGlide and 0.508 on AutoSplice, the highest mean over the two datasets among the compared methods. CocoGlide inpaints one region of each COCO image with the GLIDE diffusion model, and AutoSplice edits local regions with DALL-E 2; we use the Q100 release of AutoSplice. The other rows are from RITA (Zhu et al., 2025), which trains each method under the IMDL-BenCo CAT-Net protocol and tests at 512×512, and from PAL (Zhu et al., NeurIPS 2026).

MethodTraining dataInputCocoGlideAutoSpliceMean
MVSS-NetCAT-Net protocol5120.4430.3850.414
CAT-NetCAT-Net protocol5120.4100.3870.399
PSCC-NetCAT-Net protocol5120.4740.5510.513
TruForCAT-Net protocol5120.2830.3930.338
MesorchCAT-Net protocol5120.4500.4020.426
RITACAT-Net protocol5120.4930.6440.569
ConvNeXt-Small + PALCAT-Net protocol + EditGroup-45K2560.5380.4830.510
Ours, ViT-L + LoRA r32CAT-Net protocol5120.6840.5080.596

On CocoGlide, our model is 14.6 points above the next best method. On AutoSplice, it ranks third, behind RITA and PSCC-Net. We will add this table to the appendix.

We hope these answers address your concerns, and we are happy to clarify anything further.

AuthorsOpenReview ↗

Response (update): W1 – Bayar, SRM and Noiseprint++ results

Here are the results with Bayar, SRM and Noiseprint++ that we promised in our response to W1.

The forensic priors add little beyond the extra parameters of the branch, and switching the forensic branch off barely changes the result.

Setup. We added each extractor to the model of the paper and retrained it with everything else unchanged:

ModelDINOv3 ViT-L/16 + LoRA r64 + Conv3 head, MVSS protocol, 512×512 input
ExtractorsBayar: learnable constrained 5×5 convolution. SRM: three fixed filters. Noiseprint++: TruFor's pre-trained extractor, frozen
How they are addedEach extractor feeds its own ResNet-18 (up to stride 16), whose output is concatenated with the DINOv3 features before the head. The new head weights start at zero
ControlThe same branch fed with the RGB image, to separate the forensic prior from the extra parameters
Kept the sameCASIAv2 training data (5,123 images), augmentation, batch size 80 with 2 accumulation steps, learning rate 3e-4, 100 epochs, seed 43

Pixel F1 at the best epoch on the test sets:

Extra branchTrainable paramsCASIAv1ColumbiaNIST16COVERAGEIMD2020Mean
None (seed 43)12.19M0.89760.89460.62290.86990.65330.7876
None (seed 44)12.19M0.88480.91110.64150.86570.64890.7904
RGB (control)16.15M0.89100.89010.63990.88480.65780.7927
SRM16.17M0.90020.88270.64340.88090.62310.7861
Noiseprint++16.15M0.89010.91210.63080.88590.64350.7925
Bayar16.15M0.89790.90080.64940.88580.66160.7991

1. The extra parameters alone give as much gain as SRM or Noiseprint++. The branch adds about 4M trainable parameters. Fed with the plain RGB image, which carries no forensic prior, it raises the mean F1 from 0.7876 to 0.7927. SRM (0.7861) and Noiseprint++ (0.7925) do no better; only Bayar (0.7991) is higher, by 0.6 points.

2. Switching off the forensic branch barely changes the result. We set the branch output to zero and evaluated each model again at its best epoch. The forensic models lose less than one point, and Bayar only goes from 0.7991 to 0.7987. Switching off the RGB branch, in contrast, costs about six points:

Extra branchBranch onBranch switched off
RGB (control)0.79270.7316
SRM0.78610.7776
Noiseprint++0.79250.7899
Bayar0.79910.7987
Reviewer_zvgXOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DINOv3-IML, combining QKV-LoRA adaptation of DINOv3 with a shallow segmentation head. It reports substantial localization gains under the CAT-Net and MVSS protocols and uses representation probes to show that adaptation makes manipulation labels more consistently readable across images.

Critical strengths
  • The method is simple and uses relatively few trainable parameters. QKV-LoRA adaptation and a shallow segmentation head achieve strong localization without a dedicated forensic feature extractor.

  • The reported improvements are substantial under both CAT-Net and MVSS protocols. Source-validation-selected comparisons also support the benefit of encoder adaptation over increasing decoder capacity.

  • The analysis goes beyond benchmark scores. Linear probes, normalization controls, layer-wise measurements, and controlled copy-move experiments provide useful evidence about representation changes.

Critical weaknesses
  • The headline results use test labels for checkpoint selection. Selecting checkpoints by target-benchmark F1 introduces selection bias. Although the source-validation controls are reassuring, independently selected results are needed for the exact headline configurations under both protocols.

  • The evaluation has limited scope. It covers five established forgery datasets, with post-processing robustness tested only on CASIAv1. Modern generative edits and false positives on pristine images are not evaluated, limiting the broader conclusions about replacing specialized forensic features.

  • The interpretation of the readability analysis remains incomplete. Higher probe AP and stronger alignment after mask-supervised adaptation establish improved label separability, but do not identify which forensic or contextual cues drive it. The copy-move experiment matches interior pixels while leaving surrounding context different, so it does not fully isolate the proposed relational explanation.

Questions for authors
  • Can you provide joint sample counts across the four axes and editing engines? How do correlations between these factors affect the interpretation of the transfer gaps?

  • What quantitative annotation-quality checks were performed? For AI edits, how were changes outside the intended editing region handled in the ground-truth masks?

  • Were dataset size, training steps, and input resolution matched across the training resources in Table 2? Are any existing controlled comparisons available?

  • Which benchmark findings specifically motivated the ForenScope components, and do the existing ablations demonstrate improvements on the corresponding failure cases?

  • What protocol differences explain SparseViT’s Multi-Source F1 of 0.850 in Table 3 versus 0.722 in Appendix Table 13? Please clarify the training partitions and evaluation cohorts.

8 comments
AuthorsOpenReview ↗

Response (1/2): W1 – checkpoint selection; W2 – robustness, generative edits, authentic images

Thank you for the time you spent on our paper and for the detailed comments. Your points on checkpoint selection, robustness and the cues behind the readability gain showed us where the paper needed more evidence, and we have run new experiments for each of them. We answer the weaknesses in the order of the review and the questions at the end.

W1 (checkpoint selection). "The headline results use test labels for checkpoint selection. ... independently selected results are needed for the exact headline configurations under both protocols."

Taking the last epoch instead of the checkpoint selected on the test sets lowers the scores by at most 0.7 points:

ProtocolConfigurationTest-selectedFixed final epoch
MVSSLoRA r320.7700.765
MVSSLoRA r640.7740.770
MVSSFull FT0.6810.677
CAT-NetLoRA r32, seed 430.8510.844
CAT-NetLoRA r32, seed 440.8390.839

For the CAT-Net configuration we report re-trained runs with seeds 43 and 44. At the final epoch, our lead remains above 11 points over IML-ViT at 0.725 under CAT-Net and above 23 points over TruFor at 0.530 under MVSS.

W2 (robustness). "... with post-processing robustness tested only on CASIAv1."

We applied the same corruptions to all four CAT-Net test sets. Pixel F1 of the CAT-Net-protocol model:

CorruptionCASIAv1COVERAGEColumbiaNIST16Mean
Clean0.9070.9050.9410.6360.847
JPEG q500.8050.6860.9300.5600.745
Noise variance 230.8930.8200.9150.5900.804
Blur kernel 230.4790.4010.8990.5810.590

We will add this table to the robustness appendix.

W2 (generative edits). "Modern generative edits ... are not evaluated, limiting the broader conclusions about replacing specialized forensic features."

Evaluated zero-shot, our CAT-Net-protocol model reaches pixel F1 0.684 on CocoGlide and 0.508 on AutoSplice, the highest mean over the two datasets among the compared methods. CocoGlide inpaints one region of each COCO image with the GLIDE diffusion model, and AutoSplice edits local regions with DALL-E 2; we use the Q100 release of AutoSplice. The other rows are from RITA (Zhu et al., 2025), which trains each method under the IMDL-BenCo CAT-Net protocol and tests at 512×512, and from PAL (Zhu et al., NeurIPS 2026).

MethodTraining dataInputCocoGlideAutoSpliceMean
MVSS-NetCAT-Net protocol5120.4430.3850.414
CAT-NetCAT-Net protocol5120.4100.3870.399
PSCC-NetCAT-Net protocol5120.4740.5510.513
TruForCAT-Net protocol5120.2830.3930.338
MesorchCAT-Net protocol5120.4500.4020.426
RITACAT-Net protocol5120.4930.6440.569
ConvNeXt-Small + PALCAT-Net protocol + EditGroup-45K2560.5380.4830.510
Ours, ViT-L + LoRA r32CAT-Net protocol5120.6840.5080.596

On CocoGlide, our model is 14.6 points above the next best method. On AutoSplice, it ranks third, behind RITA and PSCC-Net. We will add this table to the appendix.

W2 (pristine images). "... false positives on pristine images are not evaluated."

On authentic images the pixel false-positive rate is 2.4% on COVERAGE and 0.9% on CASIAv1, and the image-level AUC is 0.959 and 0.982. We use the CAT-Net-protocol model with its unchanged threshold at logit 0, and score each image by the mean of its top 1% pixel logits.

DatasetAuthentic / tamperedPixel FPR on authentic95% CIImage-level AUC
COVERAGE100 / 1002.4%1.6–3.3%0.959
CASIAv1800 / 9200.9%0.7–1.1%0.982
AuthorsOpenReview ↗

Response (2/2): W3 – cues and copy-move experiment; Questions 1–5

W3 (which cues). "Higher probe AP and stronger alignment after mask-supervised adaptation establish improved label separability, but do not identify which forensic or contextual cues drive it."

Both kinds of cue are involved: the model responds to noise, blur, resampling and JPEG traces when they differ between a region and its surroundings, and hardly at all when the same trace covers the whole image. We tested this on 248 authentic photos from COVERAGE and CASIAv1. In each photo we mark a region shaped like a real forgery mask and add one trace inside the region only, outside it only, or over the whole image. In every condition, including the one without a trace, a 32-pixel gray band is painted over the region's outline, so the trace leaves no visible edge. The table gives the share of the region's tokens that the model flags as edited:

EncoderTraining dataTraceNo traceInside onlyOutside onlyEverywhere
ViT-L + LoRA r32CAT-Net protocolNoise σ = 100.0930.1690.6210.114
Blur σ = 20.0930.5020.1790.110
Resampling to 50%0.0930.3270.1450.094
JPEG q500.0930.2110.1860.114
ViT-L + LoRA r32CASIAv2, no augmentationNoise σ = 100.0220.3720.3140.025
Blur σ = 20.0220.3900.1830.013
Resampling to 50%0.0220.5530.1750.027
JPEG q500.0220.1650.1620.026

When the trace covers the whole image, the flagged share changes by at most 0.02, so the trace by itself hardly affects the model. When the trace is on one side only, the share rises to as much as 0.62, even if only the surroundings were changed. The model therefore reads these traces by comparing a region with its surroundings. This is the context dependence discussed in the next answer, and it matches Sec. 1, where we note that editing evidence lies in inconsistent compression statistics and in relationships among regions. The model trained without augmentation behaves the same way, so this behavior does not require the blur and JPEG augmentations. We will add these tests to App. H.

W3 (copy-move). "The copy-move experiment matches interior pixels while leaving surrounding context different, so it does not fully isolate the proposed relational explanation."

We agree: the copy-move experiment in App. H shows that the model relies on context, but not which part of the context it uses. In this experiment, a square of an authentic image is copied to another position, so the copy and its source have identical pixels but different surroundings. The model could therefore tell them apart either by relating the copy to its source or from the copy's own surroundings. Our answer on which cues the model uses shows that the surroundings alone already matter. We will therefore replace "relational provenance inference" with "context-dependent localization" in contribution 1 and App. H, where context means any part of the image outside the region, including the source.

Questions 1–5. "joint sample counts across the four axes and editing engines", "For AI edits, how were changes outside the intended editing region handled in the ground-truth masks?", "matched across the training resources in Table 2", "the ForenScope components", "SparseViT's Multi-Source F1 of 0.850 in Table 3 versus 0.722 in Appendix Table 13"

Thank you for the questions. We could not find the items they refer to in our submission, and we suspect they may have come from the review of another paper, which can easily happen when several papers are reviewed at the same time. Our paper does not use four axes or editing engines, does not include a component called ForenScope, and does not report SparseViT. In case some of the questions were meant for our paper, the closest parts are these: Table 2 gives results under the MVSS protocol, where every method is trained on CASIAv2 and the input size of each method is listed; Table 3 compares encoder adaptation with prediction heads; and our ground-truth masks are the ones supplied with the existing benchmarks. Your weaknesses raised the points we most needed to address, and we have answered all of them above. If you could let us know which parts of our paper you had in mind for these questions, we would be glad to answer them as well.

We hope these answers address your concerns, and we are happy to clarify anything further.

As the discussion period ends today, here is a short summary of our answers.

W1 (checkpoint selection). At the fixed final epoch, instead of the epoch selected on the test sets, the scores drop by at most 0.7 points, and our lead stays above 11 points under the CAT-Net protocol and above 23 points under the MVSS protocol.

W2 (scope). We added corruption results on all four CAT-Net test sets, zero-shot results on generative edits (CocoGlide 0.684, AutoSplice 0.508), and false positives on authentic images (pixel false-positive rate 2.4% on COVERAGE and 0.9% on CASIAv1). The corruption and generative-edit tables are in the revised PDF (Tables 25 and 30).

W2 (replacing forensic features). We have also added Bayar, SRM and Noiseprint++ to our model (results in our thread with Reviewer Dx8j). An RGB branch of the same size gains as much as SRM or Noiseprint++, and switching off the forensic branch changes the mean F1 by less than one point.

W3 (cues). The model responds to noise, blur, resampling and JPEG traces when they differ between a region and its surroundings. We agree that the copy-move test shows that the model relies on context but not which part of it, and we narrow the claim to "context-dependent localization". These changes are in the revised PDF: the trace tests in App. H (Table 29) and the new wording in contribution 1 and App. H.

We hope these answers and the revision address your concerns. If any point still seems unclear or insufficient, we would be grateful to hear about it, and we would be happy to clarify it before the discussion closes.

Reviewer_AwkqOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies image manipulation localization (IML) by predicting which pixels of an image were spliced, copied, removed or otherwise edited. The authors observe that frozen DINOv3 features with a simple head localize edits poorly. They compare two ways of improving performance. First, a stronger decoder on frozen features (a global Transformer head) helps somewhat from 0.32 to 0.47 F1. Adapting the encoder with LoRA helps much more with 0.76-0.79 F1. They call this difference a "forensic readability gap." They measure it as the per-image AP of a single linear probe shared across images, with 0.50 for frozen features and 0.92 for adapted features. The proposed model, DINOv3-IML, is DINOv3 ViT-L with rank-32/64 LoRA on the fused QKV projections and a three-layer convolutional head. It uses no forensic feature extractor, and the results indicate that the IMDL-BenCo CAT-Net protocol has 0.847 pixel F1, vs. 0.725 for IML-ViT (+12.2 points). MVSS protocol (CASIAv2-only training) has 0.774 F1, vs. 0.530 for TruFor (+24.4 points) and the analyses show that after adaptation, edit-vs-background directions become aligned across images (cosine 0.21 to 0.78). The change appears in the middle-to-late blocks and both the Q/K and V updates contribute. Furthermore, a linear head already reaches 0.815, and LoRA beats full fine-tuning, and the model is more robust than prior methods to noise, blur and JPEG on CASIAv1.

Critical strengths
  • DINOv3-IML beats all listed specialized methods on every test set under the MVSS protocol and on three of four under the CAT-Net protocol, with no forensic front end. The gains are not just an artifact of choosing checkpoints on the test sets. In the matched study (Table 3, App. C), the model is trained on 4,098 CASIAv2 images with no augmentation and the checkpoint is selected on held-out CASIAv2 validation data. It still achieves a mean F1 of 0.785 on the four CAT-Net test sets. For comparison, TruFor and IML-ViT reach about 0.58 and 0.57 on the same four sets under the MVSS protocol (Table 2).

  • The paper separates the encoder from the head with matched parameter counts (frozen + Transformer head 8.93M vs. LoRA + Conv3 9.05M). It tests rank, layer coverage, DoRA and AdaptFormer, linear vs. convolutional heads, and three decoder extensions. It reports best-vs-final epochs for the component runs and gives paired bootstrap intervalprovides the spatial diagnostics (Table 19).

Critical weaknesses

1. LoRA on a DINO backbone plus a light or linear head for manipulation localization already exists in DinoLizer (LoRA on DINOv2 Q/V with a linear patch head). The paper cites DinoLizer but does not compare against it. Also, the authors’ own DINOv2 variant (0.807, Table 6) beats all prior methods. So the main finding is “LoRA-adapting a large self-supervised ViT works well for IML,” which the title presents as DINOv3-specific. The shared linear probe is fitted on CASIAv2 tokens (App. E), and CASIAv2 is part of the adapted encoder’s own training data in both the CAT-Net and matched setups. The encoder was trained end-to-end with a near-linear readout on these same labels. It is therefore expected that a linear probe works much better on adapted features than on frozen ones. This is the usual gap between linear probing and fine-tuning. The same applies to the alignment measure, which uses class-mean directions from the training data. The global-affine invariance remark and the affine controls only show that LoRA is not a global affine change of the features, which is also expected.

2. "Editing information is present in frozen features" may reflect object saliency, not forensic evidence and the 0.50 vs. 0.14 AP comparison is used to argue that frozen features contain editing information. But frozen AP is 0.80 on CASIAv1 splicing and only 0.39 on copy-move (Table 5). That pattern fits spliced objects simply being salient objects. No saliency or segmentation control is reported. Relational reasoning is not shown and the claim of "relational provenance inference" (contribution 1) relies mainly on the controlled copy-move test in App. H (role AUC 0.99). But the training augmentations include synthetic copy-move (App. A), and the pasted 128×128 squares are not blended, so their edges are visible to global attention. The test may measure detection of this augmentation’s boundary artifact rather than provenance reasoning. Furthermore, the decoder and fine-tuning conclusions rest on untuned baselines. The Transformer head’s learning rate was tuned with the frozen encoder and reused for the adapted encoder (App. C), so "a global head brings no further gain" may reflect hyperparameters.

3. Full fine-tuning was tried at only one lower learning rate (10^5, decay 0.85). The full fine-tuning rows in Tables 1-2 use the shared 3×10^⁴ recipe. The conclusion that placement of capacity "matters more than how much is added" is therefore too broad. Checkpoints are selected on the test sets. For the main tables, each run’s checkpoint maximizes mean F1 on the target test sets, with evaluations every few epochs (Section 4.1, App. D). This follows IMDL-BenCo practice, but best-vs-final results are not reported for the headline rows. Table 15 covers only component runs. The rank is also chosen per protocol (r=32 for CAT-Net, r=64 for MVSS). Single seed, with unstable results. Every run uses one seed (42). The MVSS results vary widely: ViT-B r=32 scores 0.350 vs. 0.613 for r=64, and ViT-B full fine-tuning scores 0.110, below ManTra-Net’s 0.123 (Table 13). This points to training instability that a single seed cannot reveal.

4. Training differs from the prior methods.

  • Only tampered images are used for CAT-Net-protocol training (Section 4.1).
  • The augmentations include synthetic copy-move, inpainting, JPEG and blur (App. A).
  • The pretrained backbone is much larger.

Prior-method numbers are copied from earlier papers, so it is unclear whether the training conditions match. This especially affects the robustness comparison (Tables 20-23), since the model was trained with JPEG and blur augmentation. Authentic images are never evaluated. Only tampered test images are used, so false positives on authentic images and image-level detection are untested. Tampered-only training makes such an outcome a real risk for practical use. One small dataset drives much of the CAT-Net gain. Datasets are weighted equally, so Coverage (100 images; 0.905 vs. 0.654) contributes about 6.3 of the 12.2-point gain.

Some minor details that may reflect the overall changes in the paper:

  • The title attributes the result to DINOv3, while DINOv2 also beats all priors (App. A.1).
  • The robustness paragraph (Section 4.3) does not say which model it evaluates.
  • The Fig. 2 examples are "pre-selected" with no stated criterion.
  • The Table 3 caption is confusing.
  • Precision is mixed between 3 and 4 decimals.
Questions for authors

1. Please report final-epoch results, or results with source-validation selection, for the headline rows of Tables 1 and 2: ViT-L r=32/64 and full fine-tuning under both protocols. Please also report several seeds for the main ViT-L models. Why do ViT-B r=32 (0.350) and ViT-B full fine-tuning (0.110) collapse under the MVSS protocol?

2. Were the prior CAT-Net-protocol results in Table 1 trained with authentic images and with the same augmentations? Can you retrain at least one strong prior method (for example Mesorch or IML-ViT) with your sampling and augmentations, or train your model with the default IMDL-BenCo augmentations?

3. What is the pixel-level false-positive rate on authentic images, and what is the image-level detection AUC? Both are supported by IMDL-BenCo.

4. The readability probe is fitted on CASIAv2, which the adapted encoder was trained on. Could you fit it on data that was not used for adaptation? How does the "readability gap" differ from the standard gap between linear probing and fine-tuning, for example, compared with a nonlinear probe on frozen features with capacity matched to the LoRA parameters?

5. How much of the frozen-feature AP (especially 0.80 on splicing) is explained by object saliency? A probe that predicts salient-object masks would answer this question.

6. App. H: was the encoder trained with synthetic copy-move augmentation? Can you repeat the test with blended or feathered pastes, or with the matched no-augmentation models from Table 3?

7. Was the Transformer head’s learning rate re-tuned for the adapted encoder?

6 comments
AuthorsOpenReview ↗

Response (1/4): W1 – DinoLizer, title, readability probe

Thank you for the detailed and constructive review. We answer the weaknesses (W) and questions (Q) in the order of the review.

W1 (DinoLizer). "LoRA on a DINO backbone plus a light or linear head ... already exists in DinoLizer. The paper cites DinoLizer but does not compare against it."

DinoLizer and our work address different problems and use different training and test datasets:

DinoLizer v2Ours
Edits localizeddiffusion inpaintingsplicing, copy-move, removal
Starting weightsDINOv2-B already fine-tuned for synthetic-image detectiongeneral-purpose DINOv3, no forensic pretraining
Training dataB-Free, TGIFCAT-Net protocol sources; CASIAv2 for the MVSS protocol
Test dataCocoGlide, TGIF, SAGI, SIDA, MagicBrush, FOSSILCASIAv1, Columbia, NIST16, COVERAGE, IMD2020

We agree that our current sentence on DinoLizer in Sec. 2 is too brief, as it mentions only what the two works share. We will rewrite it with the differences above.

On timing, DinoLizer added LoRA in its second arXiv version, posted on 30 Jul 2026. Its first version, posted on 25 Nov 2025, trains only a linear head on a frozen DINO backbone. Since the LoRA version appeared less than two months before the ICLR 2027 deadline and is so far available only on arXiv, we will cite it as contemporaneous work, in line with the ICLR 2027 guidelines.

W1 (title). "... the authors' own DINOv2 variant (0.807, Table 6) beats all prior methods. So the main finding is 'LoRA-adapting a large self-supervised ViT works well for IML,' which the title presents as DINOv3-specific."

We agree that DINOv2 also beats all prior methods. Under the CAT-Net protocol at 448 input, DINOv2 with the same adaptation reaches 0.807 and DINOv3 reaches 0.833 (App. A.1). In the introduction we will state that the finding holds for both DINOv2 and DINOv3. We keep "DINOv3" in the title because the main experiments and all analyses in the paper use DINOv3.

W1, Q4 (readability probe). "The readability probe is fitted on CASIAv2, which the adapted encoder was trained on." "How does the 'readability gap' differ from the standard gap between linear probing and fine-tuning, for example, compared with a nonlinear probe on frozen features with capacity matched to the LoRA parameters?"

The gap remains when the probe is fitted on data never used for adaptation. None of the four test sets was used to train the encoder. We refitted the probe on three of them and evaluated it on the remaining one, rotating over all four. Mean token AP:

Probe fitted onFrozenAdapted
CASIAv2, as in the paper0.5010.919
The other three test sets0.5500.920

You note that the alignment measure also uses directions from the training data. With directions from the other three test sets, alignment is also unchanged: 0.23 for frozen and 0.78 for adapted features, against 0.21 and 0.78 in the paper.

We agree that the readability gap is a case of the usual gap between linear probing and fine-tuning, and we will describe it this way in the abstract, the introduction and Sec. 5. The capacity-matched comparison you suggest is in Table 3: a nonlinear Transformer head on frozen features reaches 0.467 F1 with 8.93M parameters, against 0.785 for LoRA + Conv3 with 9.05M. We also agree that the affine controls only rule out a global affine change of the features, and we will state this in the revision.

AuthorsOpenReview ↗

Response (2/4): W2 – object saliency, copy-move test, head learning rate

W2, Q5 (object saliency). "'Editing information is present in frozen features' may reflect object saliency, not forensic evidence." "How much of the frozen-feature AP (especially 0.80 on splicing) is explained by object saliency?"

Saliency accounts for much of the frozen-feature AP on Columbia and COVERAGE, and for part of it on CASIAv1 splicing. We fitted the same ridge probe on frozen DINOv3 tokens to human salient-object masks instead of edit masks, using 2,000 DUTS-TR images; it reaches AP 0.966 on DUTS-TE. We then scored this saliency probe against the manipulation masks and compared it with the paper's probe, which is fitted on edit masks. Token AP:

GroupFrozen + saliency probeFrozen + paper's probeAdapted + paper's probe
CASIAv1 splicing0.6060.7960.984
CASIAv1 copy-move0.2410.3930.939
Columbia0.6020.5540.958
COVERAGE0.3990.3320.949
NIST160.4500.5250.807

Two further analyses separate saliency from edit evidence. First, we restricted the evaluation to the salient regions of 345 CASIAv1 splices. There the paper's probe on frozen features reaches AP 0.856, while a constant score reaches 0.380, the share of edited tokens in these regions; the difference of 0.476 has a 95% CI of 0.443–0.507. Second, for 228 CASIAv1 splices the original photo from which the object was copied is also available. We measured the share of the object's tokens scored above zero, in the spliced image and in its original photo:

DINOv3-L featuresReadoutIn the spliced imageIn its original photo
Frozenlinear probe87%51%
LoRA r32linear probe99%9%
Frozentrained Conv3 head73%2.5%
LoRA r32trained Conv3 head98%3.4%

Read with a linear probe, frozen features flag half of the object even in its original photo. A Conv3 head trained on the same frozen features flags 2.5% there, so on these splices frozen features do carry edit information, but a linear probe cannot separate it from saliency.

We will replace "editing information is present in frozen features" in the abstract and Sec. 5 with: frozen features localize spliced objects partly through object saliency; after adaptation, edited regions are readable in all five groups with AP between 0.81 and 0.98.

W2, Q6 (relational claim). "Relational reasoning is not shown ... The test may measure detection of this augmentation's boundary artifact rather than provenance reasoning." "App. H: was the encoder trained with synthetic copy-move augmentation? Can you repeat the test with blended or feathered pastes, or with the matched no-augmentation models from Table 3?"

Yes, the CAT-Net-protocol encoder was trained with synthetic copy-move augmentation, but the model still tells the pasted copy from its source when the paste is feathered or Poisson-blended, and also when it is trained without any augmentation. The Table 3 model was trained on 4,098 CASIAv2 images without any augmentation and selected on CASIAv2 validation. Role AUC with the fixed linear probe:

EncoderHard pasteFeatheredPoisson
LoRA r32 with copy-move augmentation0.9880.9931.000
LoRA r32 without augmentation0.9150.9630.960
Frozen0.7030.7820.842

We agree that this test does not establish "relational provenance inference". With hard pastes, the model tells the pasted copy from its source although their pixels are identical, so the decision depends on the surrounding context. We will therefore replace the phrase with "context-dependent localization" in contribution 1 and App. H.

W2, Q7 (Transformer-head learning rate). "The Transformer head's learning rate was tuned with the frozen encoder and reused for the adapted encoder."

We did not re-tune the head learning rate for the adapted encoder in the submission. With a re-tuned learning rate, the Transformer head is within one point of Conv3, so once the encoder is adapted, the choice of head changes the result very little. We have now trained LoRA + Transformer head at two more head learning rates in the Table 3 setting. Mean F1 over CASIAv1, Columbia, NIST16 and COVERAGE:

HeadHead LRMean F1
Transformer3e-4, submitted0.756
Transformer1e-40.791
Transformer1e-30.774
Conv31e-30.785

The best Transformer head, at 1e-4, scores 0.791, 0.6 points above Conv3 in a single run. By comparison, adapting the encoder raises the same Conv3 head from 0.320 with a frozen encoder to 0.785 (Table 3). In the abstract and Sec. 5, we will replace "a global head brings no further gain" with "a global head adds little once the encoder is adapted", and add this table.

AuthorsOpenReview ↗

Response (3/4): W3 – full fine-tuning, checkpoint selection, rank, seeds, ViT-B

W3 (full fine-tuning). "Full fine-tuning was tried at only one lower learning rate."

At a higher learning rate of 3e-5, full fine-tuning improves from 0.668 to 0.767, but stays below LoRA + Conv3. Table 3 setting, mean F1 over the four test sets:

EncoderEncoder LRHeadTrainable parametersMean F1
Frozen DINOv3 ViT-L–Transformer8.93M0.467
Full fine-tuning1e-5, submittedConv3309M0.668
Full fine-tuning3e-5Conv3309M0.767
LoRA r321e-4Conv39.05M0.785

We will add the 3e-5 row to Table 3 and the introduction, and narrow the conclusion to: at a similar trainable-parameter budget, adapting the encoder is more effective than enlarging the head.

W3, Q1 (checkpoint selection). "best-vs-final results are not reported for the headline rows."

Taking the last epoch instead of the checkpoint selected on the test sets lowers the scores by at most 0.7 points:

ProtocolConfigurationTest-selectedFixed final epoch
MVSSLoRA r320.7700.765
MVSSLoRA r640.7740.770
MVSSFull FT0.6810.677
CAT-NetLoRA r32, seed 430.8510.844
CAT-NetLoRA r32, seed 440.8390.839

For the CAT-Net configuration we report re-trained runs with seeds 43 and 44. At the final epoch, our lead remains above 11 points over IML-ViT at 0.725 under CAT-Net and above 23 points over TruFor at 0.530 under MVSS.

W3 (rank per protocol). "The rank is also chosen per protocol."

Tables 1 and 2 already report both r = 32 and r = 64 under both protocols, and the two ranks differ by at most one point. With r = 32 under both protocols, the leads over the strongest baselines are 12.2 and 24.0 points:

Protocolr = 32r = 64Strongest baselineLead with r = 32
CAT-Net0.8470.837IML-ViT 0.725+12.2
MVSS0.7700.774TruFor 0.530+24.0

The lead therefore does not depend on choosing the rank per protocol. We will state this fixed-rank comparison in the results discussion.

W3, Q1 (seeds). "Please also report several seeds for the main ViT-L models."

Across seeds, the final-epoch mean F1 ranges from 0.770 to 0.788 for MVSS LoRA r64 and from 0.839 to 0.844 for CAT-Net LoRA r32.

ModelSeedPhysical batch × accumulationTest-selectedFinal epoch
MVSS, LoRA r6442, submitted16 × 100.7740.770
4380 × 20.7880.784
4480 × 20.7900.788
CAT-Net, LoRA r3242, submitted24 × 100.847–
4380 × 30.8510.844
4480 × 30.8390.839

The effective batch is the same as in the submitted runs: 160 for MVSS and 240 for CAT-Net.

W3, Q1 (ViT-B). "Why do ViT-B r=32 (0.350) and ViT-B full fine-tuning (0.110) collapse under the MVSS protocol?"

The ViT-B r32 result varies with the seed, and ViT-B full fine-tuning collapses under the MVSS protocol because the shared learning rate of 3e-4 is too large for updating all ViT-B weights. MVSS protocol, mean F1:

ViT-B runSeedLearning rateTest-selectedFinal epoch
LoRA r32, submitted423e-40.350–
LoRA r32433e-40.5520.539
Full FT, submitted423e-40.110–
Full FT433e-40.1970.008
Full FT433e-50.5480.529

The two seed-43 full fine-tuning runs differ only in the learning rate: at 3e-4 the best score is reached at the first evaluation and the final epoch falls to 0.008, while at 3e-5 the run reaches 0.548.

Both LoRA and a small learning rate keep the change to the pretrained encoder small, and a large change is hard to control. LoRA limits the change by design, through a rank-32 update that starts at zero, which makes it easier to tune: it reaches 0.552 with the shared learning rate, while full fine-tuning needs a learning rate ten times lower to reach 0.548. We will report these runs and add this explanation to Sec. 4.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions

We thank all reviewers for their comments. Below we list the changes in the revised PDF by section, with the comment each one answers. Table and appendix numbers refer to the revised PDF. The items marked "zvgX follow-up" were added in a second revision on 19 Nov.

Abstract and introduction

  • The introduction defines the forensic readability gap as "in IML, readouts of frozen features fall far behind fine-tuning the encoder", and the abstract describes it as a task-specific instance of the gap between frozen-feature readout and supervised encoder adaptation; the abstract no longer states that frozen features contain editing information (Awkq W1, W2).
  • The introduction states that the same adaptation applied to DINOv2 also surpasses all prior methods (App. A.1) (Awkq W1).
  • The abstract, introduction and Sec. 5 state that a global head adds little once the encoder is adapted; the introduction no longer compares LoRA with full fine-tuning (Awkq W2, W3).
  • Contribution 1 and App. H use "context-dependent localization" in place of "relational provenance inference" (Awkq W2, zvgX W3).

Sec. 2 (related work)

  • SparseViT is cited, and the DinoLizer sentence gives its backbone and task and describes it as contemporaneous work (Awkq W1, Awkq follow-up 5).
  • Forensim (WACV 2026) is cited (zvgX follow-up).

Sec. 4 (benchmark evaluation)

  • Sec. 4.1 describes the CAT-Net-protocol training set as the IMDL-BenCo training set (Awkq follow-up 2), and points to App. I and J (zvgX follow-up).
  • Sec. 4.2 and 4.3 report final-epoch results below Tables 1 and 2; under the CAT-Net protocol, these come from the rank-32 runs retrained with seeds 43 and 44 (zvgX follow-up).
  • The matched comparison adds full fine-tuning at backbone learning rate 3e-5 (Table 3) (Awkq W3).
  • Table 1 adds a parameter column and the SparseViT row; the Columbia sentence in Sec. 4.2 is updated (Awkq W4, Awkq follow-up 5).
  • Sec. 4.3 reports rank-32 LoRA under the MVSS protocol (0.770), notes that the ViT-B LoRA result depends on the seed and that ViT-B full fine-tuning needs a lower learning rate (Table 14), and names the model used for robustness (Awkq W3, Awkq follow-up 4, Awkq presentation).
  • Sec. 4.3 states that blur with kernel size 23 causes the largest drop (Table 25) (zvgX follow-up).

Sec. 5 (analysis) and conclusion

  • Table 3 has a new caption, reports the checkpoint selected on CASIAv2 validation, and adds the 3e-5 run (Awkq W3, Awkq presentation).
  • Sec. 5.1 points to the head learning-rate study (Table 16) and notes that on spliced images part of the frozen-feature response comes from object saliency (App. E) (Awkq W2).
  • Sec. 5.2 reports AP between 0.807 and 0.984 after adaptation on every test set and manipulation type, describes the affine controls as global affine controls, and states that the Fig. 2 examples are drawn at random within each dataset (Awkq W1, W2, Awkq presentation).
  • The conclusion states that, for forensic localization, adapting the encoder is more effective than enlarging the head (Awkq W3).
  • All tables report three decimals (Awkq presentation).

Appendix

  • App. B: SparseViT row in Table 12; Table 14 gives ViT-B under the MVSS protocol across seeds and learning rates (Awkq W3, Awkq follow-up 4).
  • App. C: Table 16 gives Transformer-head learning rates on the LoRA-adapted encoder; the 3e-5 run is added (Awkq W2, W3).
  • App. D: Table 18 gives seeds 43 and 44 with their batch configuration (Awkq W3, Awkq follow-up 3).
  • App. E: Table 21 compares each spliced object with the same object in its original photo (Awkq W2).
  • App. G: Table 25 gives corruption results on all four CAT-Net test sets (zvgX W2).
  • App. H: trace tests inside a region, outside it and over the whole image (Table 29) (zvgX W3, Dx8j W2). App. H also notes that these tests do not show whether the model identifies the edited side, and cites Forensim (zvgX follow-up).
  • App. I (new): zero-shot results on CocoGlide and AutoSplice (Table 30) (zvgX W2, Dx8j W3).
  • App. J (new): false alarms on authentic images and the evaluation procedure (Table 31) (Awkq W4, zvgX W2, zvgX follow-up).

References

  • Published versions replace arXiv citations where they exist, including IMDL-BenCo (NeurIPS 2024), Mesorch (AAAI 2025), M2SFormer (ICCV 2025), TruFor (CVPR 2023), DINOv2 and DINOv3 (TMLR) and LoRA (ICLR 2022). Following the published Mesorch table, the TruFor clean score in Tables 26–28 is 0.818, as in Table 1 (Awkq follow-up 5).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.