{"id":"3319e847-07c4-4d9c-a957-254db40d4c19","arxiv_id":"2505.11753","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"X-Edit uses Stable Diffusion inversion features with a U-Net and attention to predict edited-region masks, and contributes a paired 167,026-image dataset for the task.","lead":"This paper introduces X-Edit, a method that finds which pixels of an image were changed by text-guided diffusion editing, plus a dataset of 167,026 original and edited image pairs. It is a forensic tool for pinpointing manipulated regions rather than only labeling an image as fake.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The localization claim rests on a raw pixel-difference mask and PSNR/SSIM metrics that are also the training objective; thresholded segmentation metrics on semantic masks are needed.","rationale":"The paper targets a real gap and the pipeline is plausible: inversion features from Stable Diffusion have known forensic utility, and a U-Net with CBAM can in principle regress a mask. The strongest part is the dataset size and the idea of framing edit localization as mask regression. My concern is narrower but load-bearing: the headline claim is 'localizes edits', yet both the training target (Eq. 6) and the reported metrics (PSNR/SSIM in Table 1) are directly tied to the raw pixel-difference mask. PSNR is just -10log10(MSE); SSIM is the second term of Eq. (8). Reporting these metrics is therefore closer to reporting training loss than to independent evidence of localization. The mask y=|x_o - x_e| conflates semantic edits with global changes, and no thresholded segmentation metrics or human-validated masks are provided. The supplementary out-of-distribution experiments are qualitative and do not close the gap. This is why I agree with the reader's weakest assumption. The concern is addressable: with IoU/Dice/F1 at thresholds and a semantic-mask sanity check, the claim could be either substantiated or rejected. I do not recommend changing the reader's conditional verdict, because the weakness is a missing validation step rather than a discovered internal contradiction.","tokens_in":14364,"tokens_out":5125,"duration_ms":57356,"concrete_test":"Take a subsample of 100-200 images from the test split and have independent annotators mark the edited regions, or use edit-caption-conditioned region annotations; then threshold each model's predicted mask over a sweep and compute IoU, Dice, and F1 against these semantic masks, and also against y. If X-Edit's IoU/Dice are not clearly above the U-Net and SAM baselines, or if y itself has low IoU with human masks, the claimed localization ability is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that X-Edit localizes text-guided diffusion edits depends on Eq. (6), y=|x_o - x_e|, being a valid semantic edit mask, and on PSNR/SSIM measuring localization quality. Both links are weak. PSNR is a monotone transform of MSE and SSIM appears directly in the segmentation loss Eq. (8), so Table 1 largely shows how well the model reconstructs its own training target on the in-distribution test set; it does not demonstrate that predicted masks coincide with edited regions. The target itself is dense and nonlocal: any global change in illumination, tone, or compression makes y nonzero over the whole image, while a style transfer or subtle edit can leave y sparse or misleading. The paper's own failure cases (Fig. 6) concede this. Moreover, PSNR/SSIM do not penalize localization errors such as false positives far from the edit or missing interior regions; a blurred copy of the difference map can score well. No IoU/Dice/thresholded precision-recall is reported, and the out-of-distribution evaluation is qualitative only. Since the model is optimized to minimize MSE and SSIM against y, reporting those same quantities as evidence of 'accurate localization' is circular for the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces X-Edit, a method for localizing text-guided diffusion-based edits in images. The pipeline performs DDIM inversion of a query image through a pretrained Stable Diffusion model, constructs a composite feature volume from the original image, decoded noise map, reconstructed image, and residual, and feeds this into a U-Net with CBAM attention trained to predict a soft mask. The ground-truth mask is defined as the absolute pixel difference between the original and edited images (Eq. 6). Training combines an MSE+SSIM segmentation loss (Eq. 8) with a relevance loss (Eq. 7) aimed at suppressing high-frequency artifacts. The authors contribute a dataset of 167,026 original/edited pairs generated with InstructPix2Pix, and report PSNR/SSIM comparisons against U-Net, ViT, SegFormer, and SAM baselines, plus ablations over input composition and qualitative out-of-distribution results on FPE, MasaCtrl, and PnP edits. The central claim is that X-Edit accurately localizes diffusion-based edits and outperforms baselines on this task.","tokens_in":14535,"tokens_out":2461,"duration_ms":24477,"significance":"If the localization claim were established, the paper would provide a useful forensic tool and a large paired dataset for a relatively new task. The method's core idea of exploiting inversion discrepancies is well motivated and builds sensibly on FakeInversion. The paper ships a concrete architecture, a loss design, and a dataset of nontrivial scale, and the qualitative results do show that the model often highlights plausible edited regions. However, the quantitative evidence as presented does not demonstrate localization accuracy in the forensic sense: the evaluation metric is computed against the same raw difference mask used as the training target, and no segmentation-style metrics (IoU, Dice, thresholded precision/recall) are reported. The claim of 'balanced precision and recall' in the abstract is not backed by any precision/recall numbers in the paper. The significance of the contribution therefore hinges on whether the authors can re-frame the target and evaluation to match the localization claim.","major_comments":[{"comment":"The ground-truth mask y=|x_o - x_e| is a dense, nonlocal pixel-level difference and does not represent a semantic edit mask. Any global change in illumination, tone, or compression produces nonzero values everywhere, while a subtle or style-based edit can yield a sparse or misleading target. Using this same mask as both the training target and the evaluation reference means that Table 1 largely measures how well the model reconstructs its own training target on an in-distribution test set, rather than whether predicted masks coincide with the actually edited regions. The paper's own failure cases (Fig. 6, color/style shifts) concede this limitation, but the central 'accurate localization' claim is not supported by this target definition.","section":"§3, Eq. (6)"},{"comment":"PSNR and SSIM are computed against the same y defined in Eq. (6), and SSIM appears directly in the training loss (Eq. 8). PSNR is a monotone transform of MSE, the first term of L_S. Reporting these quantities as evidence of localization quality is therefore circular: a model optimized to minimize MSE and SSIM against y will naturally score well on those same metrics, and a blurred copy of the difference map can score arbitrarily high in PSNR while being useless for localization. The absence of any IoU, Dice, or thresholded precision/recall computed against a binarized semantic edit mask is a load-bearing gap for the paper's central claim.","section":"§4, 'Inference and Evaluation Metrics'; Eq. (8)"},{"comment":"The abstract claims 'balanced precision and recall,' but no precision or recall numbers are reported anywhere in the main text or supplementary material. Table 1 reports only PSNR and SSIM. Moreover, the best PSNR (X-Edit on φ_FI, 24.946) and the best SSIM (X-Edit+finetuning on φ, 0.954) come from different configurations, and no error bars, confidence intervals, or significance tests are provided, so the statement that X-Edit 'outperforms baselines' is not statistically supported.","section":"Abstract; §5, Table 1"},{"comment":"The out-of-distribution evaluation is purely qualitative. The paper claims robustness to FPE, MasaCtrl, and PnP edits, but only a handful of example images are shown, with no quantitative results on the 100-image Flickr30k set described in the supplementary. Given that the main quantitative evaluation is in-distribution with InstructPix2Pix only, the generalization claim is not substantiated without numbers.","section":"§5, 'Qualitative results'; Supplementary Sec. J"}],"minor_comments":[{"comment":"In the Conclusions, 'paired original end edited image' should be 'paired original and edited images.'","section":"§6"},{"comment":"'LAION-Aestetics V2 6.5+' contains a typo: 'Aestetics' should be 'Aesthetics.'","section":"§3, Dataset"},{"comment":"The notation for the FakeInversion feature is introduced as φ_FI but the subscript is dropped in several places in Table 1 and the text, which can confuse which input configuration is being discussed.","section":"§4, 'Input Data Format'"},{"comment":"The claim that EfficientNet has 'high detection accuracy (99.93%) and precision (99.92%)' is stated without any training details or reference; please provide the source or experimental setup for these numbers.","section":"§2, Figure 2"},{"comment":"SAM is listed with PSNR 23.478 and SSIM 0.506; the paper explains the low SSIM via false positives on original images, but a quantitative comparison of false positive rates on the original-image subset would be more informative than the single aggregate SSIM value.","section":"§5, Table 1"},{"comment":"The relevance loss in Eq. (7) uses λ_flat and λ_edge, but the paper does not report sensitivity of the performance to these hyperparameters or to α; a short sensitivity analysis would strengthen the claim that the chosen values are robust.","section":"§3, 'Finetuning Procedure'"}],"recommendation":"major_revision","confidential_remarks":"The core idea and dataset are valuable, but the evaluation protocol needs substantial rework before the localization claim can be accepted. I would suggest the authors either (i) redefine the task as soft mask prediction with explicit acknowledgment that the target is the pixel difference map, and evaluate with regression metrics only while tempering the 'localization' language, or (ii) adopt semantic edit masks (e.g., human-annotated or derived from editing instruction/object masks) and report IoU/Dice/thresholded precision-recall. The SSIM-in-loss/SSIM-in-evaluation circularity is the most serious technical issue and must be addressed head-on. Also note that the dataset release is promised but not yet available; given the paper is from an industry lab, the editor may want to confirm data availability before accepting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on X-Edit (arXiv:2505.11753). The genuinely new thing here is the task framing and the dataset: nobody has trained a pixel-level localizer specifically for text-guided diffusion edits, and a 167k-pair InstructPix2Pix dataset is a useful resource if it actually ships. The pipeline is sensible: invert with Stable Diffusion, concatenate the reconstructed image, noise map and residual, and run a U-Net with CBAM. The authors are explicit that the architecture is assembled from published components, which I respect.\n\nWhat the paper does well: it identifies a real gap in forensics, builds a large paired dataset rather than hand-waving, and reports failure cases honestly. The out-of-distribution qualitative results on FPE, MasaCtrl and PnP are a nice addition even if only qualitative.\n\nThe soft spot is the evaluation, and it is load-bearing. The ground-truth mask is defined as the absolute pixel difference y=|x_o - x_e|, and the model is trained with MSE plus SSIM against that same y. Then Table 1 reports PSNR and SSIM between the predicted mask and y. PSNR is a monotone transform of MSE, and SSIM is literally in the training loss. So the numbers largely show how well the model reconstructs its own target on an in-distribution test set, not that the predicted mask coincides with edited regions in a forensic sense. For global edits—illumination, tone, style transfer—the difference mask is dense and non-local, so a blurred copy of the difference map can score well. No IoU, Dice, or thresholded precision/recall is reported anywhere, and the best PSNR and best SSIM come from different configurations (X-Edit on phi_FI for PSNR, finetuned X-Edit on phi for SSIM). That split weakens the \"outperforms baselines\" claim.\n\nThere are also minor issues: the input format description is ambiguous in places (phi vs phi_FI, grayscale vs RGB), and the dataset and code are promised but not yet available. The authors do acknowledge failure cases (dark flat objects, subtle low-entropy edits) and the OOD evaluation is qualitative only.\n\nDespite the circularity, I think the paper deserves a serious referee. The task and dataset are valuable enough that the evaluation can be fixed: threshold the predicted masks and report IoU/Dice/F1, ideally on a semantically meaningful mask rather than raw pixel difference, and give error bars across seeds. If the authors do that, this becomes a useful benchmark. If they don't, the localization claim should be scaled back to \"reconstructs the pixel-difference map.\"\n\nWho is this for: forensic analysts and people working on diffusion-based image editing evaluation. I'd bring it to reading group, and I'd be willing to referee it myself.","headline":"A genuinely new task and a large paired dataset, but the evaluation is circular—PSNR/SSIM against the same pixel-difference mask used as the training target—so the localization claim is not yet established.","tokens_in":15133,"tokens_out":1704,"would_cite":false,"duration_ms":15972,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that diffusion-based image edits can be localized pixel-by-pixel by inverting the image with Stable Diffusion and training a U-Net with attention on the inversion features, backed by a 167,026-pair dataset.","keywords":["text-guided image editing","edit localization","diffusion model forensics","DDIM inversion","deepfake detection","segmentation network","Stable Diffusion","InstructPix2Pix"],"falsifier":"Run X-Edit on edits that change global illumination or artistic style (e.g., “make it night” or “turn it into an oil painting”) and compare the predicted masks with human-annotated masks of the semantically changed region; if the predictions track the pixel-difference signal rather than the semantic edit, X-Edit localizes pixel changes, not edits in the forensic sense claimed.","tokens_in":1639,"feed_emoji":"🎯","tokens_out":2634,"duration_ms":77189,"temperature":0.7,"pith_summary":"X-Edit claims that edits introduced by text-guided diffusion models can be located pixel-by-pixel by inverting the image with a pretrained Stable Diffusion model and feeding the inversion features to a segmentation network. The paper argues that discrepancies between an image and its diffusion reconstruction concentrate in edited regions, so a U-Net with channel and spatial attention trained on these features can output a mask matching the changed area. It also introduces a paired dataset of 167,026 original–edited image pairs generated with InstructPix2Pix. If the claims hold, X-Edit offers a forensic tool for pinpointing manipulated regions in diffusion-edited images rather than merely flagging the image as fake, and it sets a baseline for a new localization task.","feed_headline":"Method pinpoints pixels changed by text-guided image edits","feed_subtitle":"X-Edit inverts images with Stable Diffusion, then a U-Net predicts the edited mask, beating baselines on PSNR and SSIM.","key_machinery":"The load-bearing object is the composite inversion-feature tensor $\\varphi$ defined in Eq. (5): $\\varphi = x \\oplus D(\\hat{z}_T) \\oplus D(\\hat{z}_0) \\oplus |x - D(\\hat{z}_0)|$, built from the original image, the decoded DDIM inversion noise, the reconstructed image, and the reconstruction residual. The argument is that discrepancies between the original and its diffusion reconstruction are amplified in edited regions, so a segmentation network can learn to map these discrepancies to the edited mask. The mask target is $y = |x_o - x_e|$, the absolute pixel difference between original and edited images. The model is a U-Net augmented with convolutional block attention modules (CBAM), which apply channel and spatial attention inside each block, and it is trained with a segmentation loss combining mean squared error and SSIM plus a relevance loss that uses integrated gradients and Sobel-filter high/low-frequency maps to suppress attention to edges.","core_discovery":"X-Edit is the paper's proposed answer to the question of where a text-guided diffusion edit actually changed an image. The method starts from the observation, inherited from fake-image detection, that DDIM inversion of an image through a pretrained Stable Diffusion model reconstructs unedited content well and leaves detectable discrepancies where content was altered. The paper constructs an input tensor $\\varphi$ by concatenating the original RGB image, the decoded inversion noise, the reconstructed image, and the residual between them, then trains a U-Net with CBAM attention to predict a mask $\\hat{y}$ that reproduces the absolute pixel difference $y = |x_o - x_e|$ between the original and edited image. A combined MSE-plus-SSIM segmentation loss and a relevance loss that pushes the model's integrated-gradient maps away from high-frequency edges are used to keep predictions focused on edited regions. The authors report that this pipeline outperforms the tested baselines on PSNR and SSIM, and they present it as the first method specifically modeling the localization of diffusion-based edits.","pith_inferences":["One implication the authors leave implicit is that the pixel-difference mask defines “edit” operationally; edits that recolor or relight whole scenes will yield diffuse masks, so a stricter evaluation would compare against human-annotated semantic regions.","A testable extension is to retrain the same U-Net with object-level or human-annotated masks instead of pixel differences and check whether the inversion features still separate edited from unedited regions; this would isolate whether the method learns edit semantics or low-level change.","Because inversion is conditioned on an auto-generated caption, the method’s sensitivity to caption quality could be measured by perturbing the BLIP-3 caption; if localization degrades sharply on wrong captions, the captioner becomes a silent dependency of the whole pipeline."],"forward_implications":["If X-Edit works as claimed, forensic pipelines gain a localization step: instead of only deciding that an image is fake, they can point to the pixels most likely changed by a text-guided edit.","The released dataset of 167,026 paired original–edited images provides a training and evaluation resource for a task that currently has no standard benchmark.","The qualitative out-of-distribution experiments suggest the method transfers, at least partially, to other editing methods such as FPE, MasaCtrl, and Plug-and-Play, which would make it useful beyond the InstructPix2Pix training distribution.","Because the model predicts near-blank masks for original images, it could serve as a false-positive-aware detector as well as a localizer."],"supporting_citations":[{"why":"Supplies the DDIM-inversion feature extraction that X-Edit extends to localization; FakeInversion is the base method.","marker":"[8]"},{"why":"InstructPix2Pix is the text-guided editing method used to generate the 167,026 edited images in the dataset.","marker":"[5]"},{"why":"RobustViT’s relevance-loss finetuning is adapted into X-Edit’s relevance loss steering attention away from high-frequency edges.","marker":"[9]"},{"why":"CBAM provides the channel and spatial attention blocks inserted into the U-Net.","marker":"[57]"},{"why":"U-Net is the segmentation backbone that maps the composite feature tensor to the predicted mask.","marker":"[43]"},{"why":"DDIM inversion is the procedure used to estimate the noise map and reconstruct the image latents.","marker":"[50]"},{"why":"SSIM defines the perceptual component of the segmentation loss and one of the two evaluation metrics.","marker":"[55]"},{"why":"LAION is the source of original images used to construct the paired training dataset.","marker":"[46]"},{"why":"BLIP-3 generates the captions that condition the inversion at training and inference time.","marker":"[60]"}],"fun_headline_variants":["AI tool pinpoints edits made by text-guided diffusion","Spot the diff: new method localizes AI image edits","X-Edit reveals exactly where diffusion models tweaked pixels","First method to localize text-guided diffusion edits","Forensic tool flags pixels altered by text-to-image models"],"cache_read_input_tokens":17280,"weakest_assumption_plain":"The paper assumes that the raw absolute pixel difference between the original and edited image is an adequate ground-truth mask of the edited region, and it uses that same mask both to train the network and to score it.","fun_headline_variants_meta":{"raw":{"variants":["AI tool pinpoints edits made by text-guided diffusion","Spot the diff: new method localizes AI image edits","X-Edit reveals exactly where diffusion models tweaked pixels","First method to localize text-guided diffusion edits","Forensic tool flags pixels altered by text-to-image models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1429,"prompt_tokens":1016,"completion_tokens":413,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":335}},"tokens_in":632,"tokens_out":413,"duration_ms":3889,"temperature":1.0,"reasoning_tokens":335,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:48:38.191394+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run X-Edit on edits that change global illumination or artistic style (e.g., “make it night” or “turn it into an oil painting”) and compare the predicted masks with human-annotated masks of the semantically changed region; if the predictions track the pixel-difference signal rather than the semantic edit, X-Edit localizes pixel changes, not edits in the forensic sense claimed.","supporting_citations":[{"cited_title":"Fakeinversion: Learning to detect images from un- seen text-to-image models by inverting stable diffusion","cited_arxiv_id":null,"evidence_quote":"Supplies the DDIM-inversion feature extraction that X-Edit extends to localization; FakeInversion is the base method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"InstructPix2Pix is the text-guided editing method used to generate the 167,026 edited images in the dataset."},{"cited_title":"Optimizing relevance maps of vision transformers improves robustness","cited_arxiv_id":null,"evidence_quote":"RobustViT’s relevance-loss finetuning is adapted into X-Edit’s relevance loss steering attention away from high-frequency edges."},{"cited_title":"Cbam: Convolutional block attention module","cited_arxiv_id":null,"evidence_quote":"CBAM provides the channel and spatial attention blocks inserted into the U-Net."}],"review_version":1}