acceptodds
Withdrawn from ICLR 2027

EditScan: Towards Generalizable Detection and Localization of Instruction-Driven Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Instruction-driven image editing models can realistically alter local image regions, posing severe challenges to the accuracy and generalization of image editing detection and localization. Existing forensic benchmarks for generative local editing either cover a limited range of instruction-driven editors or are primarily designed around mask-conditioned inpainting,leaving the generalization of forensic methods to rapidly evolving instruction-driven editing models insufficiently evaluated. To address this gap, we introduce EditScan-Bench together with a strong baseline approach, EditScan. EditScan-Bench is a purpose-built benchmark for instruction-driven image editing forensics, comprising 44K annotated image pairs at 1K resolution, with machine-produced masks that are carefully validated by human experts. It spans 3 widely used open-weight and 5 proprietary model families across four common manipulation types: Add, Remove, Replace, and Attribute Change. Building upon EditScan-Bench, we further propose EditScan, which connects vision foundation models with general-purpose segmentation models through specialized adaptation modules to enable both image-level detection and pixel-level localization. EditScan includes two complementary designs: (1) explicitly capturing local patch evidence to increase sensitivity to subtle edits, improving image-level detection; and (2) converting the foundation model’s global-local features into forensic prompts that steer the segmentation model toward accurate manipulation localization. In cross-generator tests, EditScan raises average image-level detection F1 and Accuracy by () and (), respectively, compared with the strongest baseline. For pixel-level localization, it achieves improvements of F1 () and IoU (), while also showing stronger robustness to post-processing degradations. The benchmark dataset and code are available at https://anonymous.4open.science/r/EditScan.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
generated image detection, image forgery detection, image forensics, content generation, content security

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_R9nNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper tackles the problem of localising image edits in instruction-controlled image editing via generative models. There are two main contributions: a) a new image-editing benchmark, EditScan-bench; and b) EditScan, a baseline method for detecting image edits. The novelty of the benchmark is that it extends beyond simple crop/splicing setups or inpainting benchmarks. The novelty of the method is a two stage architecture that combines global, image-level detection with local patchwise detection to enhance identification of pixel-level and image-level edits. While the performance results are good, I have concerns regarding the evaluation rigour. Specifically, the method appears to not fairly disentangle the performance gains stemming from the benchmark against those stemming from the detection method, prompting me to give it a reject.

Critical strengths
  1. The paper tackles a very relevant problem, e.g., detecting ai-modified images is very relevant for countering deep-fakes.
  2. The paper is generally well
Critical weaknesses
  1. It is not clear if the baseline detection methods were trained on the same EditScan-Bench. The description in Sec. 5.1 only says that all baselines were evaluated on the same test set. The code in the github repository lacks the scripts required to train the baselines on the proposed EditScan-Bench dataset, leading me to believe that those methods are applied off-the-shelf using their pre-loaded checkpoints during the evaluation. If this is the case, the performance results are unfair to the baselines, since it does not indicate whether the better results of EditScan stem from training on a different dataset or from the methodological superiority.

    1. Continuing the above weakness, the paper does not separate the gains stemming from the new Benchmark, versus the methodological gains on the detection model itself. An ablation study comparing the OOD performance of EditScan’s (and the baselines’) when trained on an existing benchmark dataset against a setting where all methods train on the EditScan-Bench, is necessary to isolate the effect of the new benchmark.

    2. The paper uses the DinoV3 model as a backbone in the detection pipeline. However, it appears that the baseline methods do not use the same backbone, which adds an additional confounding layer. It is not clear if the performance gains stem from using a better backbone or from the loss function and other proposed designs.

    3. The paper claims that the benchmark was validated by human experts. However there are insufficient details regarding how this validation was conducted. For example, there is a lack of information on the number of annotators, how annotations across subjects were combined, whether there were labelling differences between annotators, how they were selected, etc.

Questions for authors
  1. Can you provide the exact details on how the baselines were trained? Specifically, were they fine-tuned on EditScan-Bench when reporting the results, or were they pre-trained on their own detection benchmarks?
  2. Can you provide details on how the human evaluation was conducted? Preview
Reviewer_UB8LOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the detection and localization of instruction-driven image editing by introducing the EditScan-Bench dataset and the EditScan framework. EditScan-Bench comprises 44,000 pairs of high-resolution images, spanning eight families of image editing models and four categories of editing operations. During the dataset construction process, the authors employed Qwen3-VL to generate editing instructions and identify modified regions, followed by the generation of pixel-level masks via LISA, with a final stage of manual verification. EditScan uses a shared DINOv3 backbone network and is designed with two task branches. The image-level detection branch enhances the global classification results through local patch evidence, while the pixel-level localization branch converts the global and local features of DINOv3 into SAM prompts to achieve tamper region segmentation. Experimental results demonstrate strong performance in detection, localization, and robustness against degradation across seven previously unseen families of image editing models.

Critical strengths
  1. EditScan-Bench consists of 44,000 pairs of 1K-resolution images, encompassing eight families of open-source and proprietary editing models. It covers four types of operations: Add, Remove, Replace, and Attribute Change. Unlike settings that focus exclusively on mask-conditioned inpainting, this dataset enables a more direct study of image editing tasks where the location of the manipulation is determined by natural language instructions.

  2. Experiments show that EditScan achieves an average Image Accuracy of 0.8252 and a Pixel F1 of 0.5189 on seven unseen editing model families. The authors also conducted evaluations under conditions involving JPEG compression and Gaussian noise, providing further evidence of the method's robustness.

Critical weaknesses
  1. Section 3.1 uses Qwen3-VL to analyze the original image, the edited image, and the editing instructions, and then uses LISA to generate a tampering mask based on the semantic region description. While this process can localize the edited semantic objects, it may not accurately cover the entire area where actual pixel-level changes have occurred. This is particularly true for "Remove" and "Attribute Change" operations, where the actual modifications may extend beyond the original object boundaries or differ significantly from the semantically segmented regions. Although the authors mention that manual review was conducted, they do not report the proportion of samples reviewed, the consistency of the annotations, or a quantitative comparison against independently created manual masks. Consequently, it is difficult to determine whether these automatically generated masks are sufficient for evaluating high-precision pixel-level localization. This issue directly impacts the reliability of the benchmark.

  2. In Section 5.1, existing methods are evaluated using their publicly available weights, whereas EditScan is trained on a newly constructed dataset of Qwen-based edits. Therefore, the performance advantages of Tables 2–3 may come from the training data distribution, the pre-training capabilities of DINOv3 or SAM, and the proposed adaptation modules, and are difficult to attribute directly to the method design itself. Furthermore, the "SAM-only" ablation study in Table 5 removes both DINOv3 and the Localization Prompt Generator; thus, it cannot independently demonstrate that the proposed prompt-generation strategy is superior to other methods of integrating DINOv3 and SAM features.

Questions for authors
  1. Regarding the annotation quality of EditScan-Bench, could you provide further details on the manual review process, its scope, and the criteria used to ensure consistency? Are there any statistical results from independently annotated samples that could validate the accuracy of the automatically generated masks across different types of edits?

  2. Please clarify whether the training, validation, and test subsets for Qwen are entirely independent at the level of the original images. Additionally, please specify the differences in training data between the various baselines and EditScan. This information is essential for evaluating the reported gains in cross-editor generalization.

  3. Table 5 excludes both DINOv3 and the Localization Prompt Generator. Are there any existing experiments that can disentangle the performance gains attributable to pre-trained features from those contributed by the prompt generation module itself? Furthermore, could you explain why methods such as IMDPrompter and Detective SAM were not included in the direct comparison?

Reviewer_t2r9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes EditScan-Bench and EditScan for detecting and localizing instruction-driven local image edits, where edited regions are often small and their forensic traces can be diluted in global representations, while cross-generator generalization remains challenging. EditScan uses a shared DINOv3 backbone and introduces a Patch-Artifact-Aware Detection Enhancement module that scores patch-level forensic evidence, fuses complementary patch statistics, and applies learnable residual scaling to refine image-level detection; it introduces a Localization Prompt Generator that maps DINOv3 global-local features into SAM-compatible forensic prompts to guide pixel-level edited-region localization. Experiments show that EditScan achieves strong in-domain and out-of-domain detection and localization performance, generalizes well to unseen generators, remains robust under JPEG compression and Gaussian noise, and that ablation studies confirm the effectiveness of the proposed modules.

Critical strengths
  1. The paper introduces EditScan-Bench, a multi-generator benchmark spanning both open-weight and proprietary editors with image-level labels and pixel-level masks, enabling unified evaluation of image-level detection, pixel-level localization, and cross-generator generalization.

  2. The paper proposes EditScan, which combines a Patch-Artifact-Aware Detection Enhancement module for patch-level forensic evidence with a Localization Prompt Generator that converts DINOv3 global-local features into SAM-compatible prompts, jointly improving image-level detection and pixel-level localization.

Critical weaknesses
  1. Limited novelty in benchmark/task formulation, and the mask definition is questionable. Locate-Edit, PromptForge-350k, and FragFake[1] already benchmark this problem from pixel level to instance level, so the incremental contribution of EditScan-Bench is unclear. More seriously, instruction-driven mask-free edits are closer to conditional generation: the reference image may serve only as conditioning, while the whole image is regenerated. The pixel mask may therefore capture global regeneration or semantic change rather than the actual edited region. This is consequential because it undermines the task definition and the validity of the pixel-level localization claims.

  2. The baseline comparison in Table 3 is unusually weak and potentially unfair. The reported Pixel F1 values of several baselines are extremely low for a segmentation/localization task, and it is unclear whether they were trained on EditScan-Bench under the same protocol. If they were trained on the benchmark, the low scores require explanation.

[1] Sun et, al. Can VLMs Detect and Localize Fine-Grained AI-Edited Images? Arxiv 2505.15644.

Questions for authors

1: What is the essential contribution of EditScan-Bench relative to Locate-Edit, PromptForge-350k, and FragFake, beyond scale or generator coverage?

2: How is the pixel-level mask defined and verified for instruction-driven edits where the whole image may be regenerated from a reference condition?

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

Related papers

Discussion (0)

Sign in to comment.