{"id":"6e4fe509-f2cf-47da-932d-56a486ec848b","arxiv_id":"2508.12543","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"REVEAL reframes image forgery detection as a prompt-driven visual reasoning task using scene-level and region-level prompts on vision-language models.","lead":"This paper introduces REVEAL, a framework that prompts large vision-language models to detect image forgeries through whole-scene reasoning and region-wise anomaly analysis. It reports tests on Photoshop-edited, DeepFake, and AI-generated edited images, comparing against baseline forgery detectors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim unverifiable: abstract lacks localization metrics and reasoning-grounding analysis, leaving the sufficiency of VLM semantic alignment for pixel-level forgery detection untested.","rationale":"The reader's weakest assumption is that semantic alignment in VLMs is sufficient to transfer general knowledge about physics, perspective, and realism without task-specific training. I agree this is the critical load-bearing premise, but I make it more concrete: the danger is that VLM-generated text may be plausible without being grounded in local visual evidence, and the abstract gives no indication that the paper evaluates this. The abstract promises both detection and localization, yet a detection-only evaluation could mask a failure to localize actual manipulated pixels. My concern is not that the method is wrong, but that the evidence needed to assess the central claim is missing from the abstract. Since the reader already marked the paper UNVERDICTED due to lack of full text, my concern does not change that verdict; it sharpens what to look for in the full paper. I marked agreement as 'partial' because the reader focuses on sufficiency of semantic alignment, while I add the need for localization verification and reasoning-grounding checks as the decisive test.","tokens_in":687,"tokens_out":2318,"duration_ms":28033,"concrete_test":"Retrieve the full text and examine the experiments for (a) a per-region or per-pixel localization metric (e.g., IoU) comparing region-wise predictions to ground-truth manipulation masks, and (b) a human or model-based validation that the natural-language reasons reference the actual manipulated regions. If neither is present, the central detection/localization claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"REVEAL's central claim is that prompt-driven visual reasoning via VLMs can detect and localize forgeries across Photoshop, DeepFake, and AIGC domains without task-specific training. The load-bearing premise is that VLM 'semantic alignment capabilities' (abstract) ground high-level cues (physics, semantics, perspective, realism) in local pixel evidence. This premise is not demonstrated by the abstract: no evaluation metric (e.g., localization IoU), no comparison of reasoning text with ground-truth manipulated regions, and no ablations controlling for language priors are mentioned. If REVEAL's region-wise outputs are not quantitatively assessed against ground-truth masks, or if the generated rationales are not verified to correspond to actual artifacts, the framework could produce correct labels on the basis of surface-level dataset correlations while failing at genuine localization. The claim of cross-domain generalization also requires evidence that prompts are not tuned per dataset; the abstract's 'generalized guidelines' need to be shown domain-agnostic.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes REVEAL, a framework for image forgery detection that uses vision-language models with prompt-driven visual reasoning. It introduces two approaches: (1) holistic scene-level evaluation based on physics, semantics, perspective, and realism of the whole image; and (2) region-wise anomaly detection that splits the image into multiple regions and analyzes each one. The abstract claims experiments are conducted over datasets from Photoshop, DeepFake, and AIGC editing domains, with comparisons against competitive baselines and analysis of reasoning provided by the models. The central assertion is that VLMs' semantic alignment capabilities enable cross-domain generalization without task-specific training, relying on 'generalized guidelines.' However, no quantitative results, dataset names, baseline names, evaluation metrics, or error bars are provided in the abstract, making the central claim unverifiable from the available material.","tokens_in":932,"tokens_out":2778,"duration_ms":32302,"significance":"If the claimed results hold, the contribution would be significant: a single prompt-driven framework for detecting and localizing diverse manipulation types while generating natural-language rationales, with the potential to generalize across Photoshop, DeepFake, and AIGC domains without per-domain training. This would be a practical and interpretable alternative to supervised per-domain detectors. However, the abstract alone provides no evidence to support these claims. The lack of details on evaluation protocol, localization accuracy, and reasoning grounding means the significance is conditional on the full paper supplying the missing support. The idea is worth examining, but the current submission as presented does not demonstrate the claimed advances.","major_comments":[{"comment":"The central claim of cross-domain forgery detection is not accompanied by any quantitative evaluation data. The abstract states that experiments are conducted over datasets from different domains and that VLMs are compared against competitive baselines, but it does not name datasets, baselines, evaluation metrics, or error bars. Consequently, the reader cannot assess whether the proposed framework actually outperforms existing methods or generalizes. This is a load-bearing omission for a paper whose primary contribution is a new detection approach.","section":"Abstract (overall)"},{"comment":"The abstract describes region-wise anomaly detection that splits the image into multiple regions, but it does not state how localization is evaluated. The stress-test concern that the model might produce correct labels via dataset statistics without genuine pixel-level localization is unresolved. The paper should report a localization metric (e.g., IoU with ground-truth masks) or otherwise demonstrate that the region-wise outputs correspond to manipulated areas.","section":"Abstract (region-wise approach)"},{"comment":"The abstract claims that REVEAL incorporates 'generalized guidelines' and leverages the semantic alignment capabilities of VLMs, but it is unclear whether the prompts are fixed across datasets or tuned per domain. Without this specification, the claimed cross-domain generalization cannot be distinguished from prompt overfitting. The paper must clarify the prompt design and whether any dataset-specific adaptation is used.","section":"Abstract (generalized guidelines)"}],"minor_comments":[{"comment":"The phrase 'two tangential approaches' is ambiguous; 'complementary' or 'parallel' would better convey the intended relationship between the holistic and region-wise methods.","section":"Abstract (wording)"},{"comment":"The opening sentence is long and contains redundancy (e.g., 'detecting and interpreting' followed by 'providing reasoning as well as localization'). Consider tightening the phrasing for clarity.","section":"Abstract (style)"}],"recommendation":"uncertain","confidential_remarks":"This review is based solely on the abstract, as the full text was not available. Given that, I cannot reach a soundness determination. The central claim is interesting and potentially significant, but the abstract lacks the essential experimental details needed for a verdict. I recommend obtaining the full manuscript to evaluate the missing evidence, especially localization metrics and prompt-design transparency."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know about this one: the abstract for REVEAL promises a framework that uses vision-language models to detect and localize forgeries across Photoshop, DeepFake, and AIGC via prompt-driven reasoning. That is a sensible extension of recent multimodal reasoning work, and the split into holistic scene-level and region-wise anomaly detection is a reasonable way to approach it. But the abstract contains no numbers, no baseline names, no metrics, and no error bars, so the central claim is entirely unverified.\n\nWhat is actually new, as far as I can tell, is the emphasis on the reasoning output itself as an evaluation signal, plus the claim that generalized guidelines can transfer across manipulation domains without task-specific training. That is worth thinking about, but it is also exactly what needs testing. The stress-test note is right: the load-bearing premise is that VLM semantic alignment can ground high-level cues like physics, semantics, perspective, and realism in local pixel evidence. Nothing in the abstract shows that grounding. There is no IoU for localization, no check that the generated rationales correspond to actual manipulated regions, and no evidence that the prompts are domain-agnostic rather than tuned per dataset. Without those, the framework could generate plausible-sounding text while simply picking up dataset correlations.\n\nThe circularity concern is also legitimate. If the model's own reasoning is used both as detection evidence and as evaluation, you need an independent ground-truth anchor before you can trust any of it. The abstract does not address that.\n\nI would not cite this yet, and I would not bring it to a reading group based on the abstract alone. But the problem is important and the direction is promising enough that a full paper deserves serious referee time. I would send it to peer review with the clear expectation that the authors provide localization metrics, a reasoning-grounding analysis, and evidence that the prompts are not dataset-specific. If those do not show up in the full version, the paper should be rejected; if they are solid, it could be a useful contribution.","headline":"A plausible VLM-prompting framework for cross-domain forgery detection, but the abstract gives no quantitative evidence that semantic alignment actually localizes forgeries.","tokens_in":601,"tokens_out":668,"would_cite":false,"duration_ms":27135,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"REVEAL reframes image forgery detection as a prompt-driven visual reasoning task, arguing that large vision-language models can flag manipulated images by checking scene-level consistency and region-level anomalies, without task-specific tr","keywords":["image forgery detection","vision-language models","prompt-driven reasoning","semantic alignment","DeepFake detection","AIGC editing","region-wise anomaly detection","scene-level evaluation"],"falsifier":"A concrete test: on a held-out forgery dataset, run only the region-wise prompt and compare the model's predicted forged regions against human-annotated masks. If the model produces fluent explanations but its region masks have near-zero overlap with the ground truth while global scene-level accuracy stays high, the assumption that semantic alignment grounds local pixel evidence would be undermined.","tokens_in":646,"feed_emoji":"🔍","tokens_out":2022,"duration_ms":24573,"temperature":0.7,"pith_summary":"This paper claims that image forgery detection can be reframed as a prompt-driven visual reasoning task. Instead of training a detector per manipulation type, the REVEAL framework asks a large vision-language model to judge whether an image respects physics, semantics, perspective, and realism, and to inspect individual regions for inconsistencies. If correct, this would let one prompting scheme generalize across Photoshop edits, DeepFake faces, and AI-generated content while also producing a natural-language reason for each verdict. The paper evaluates this by comparing vision-language models against supervised baselines on datasets from all three domains.","feed_headline":"Prompts alone catch image forgeries across three editing domains","feed_subtitle":"REVEAL uses scene-level and region-wise reasoning to detect Photoshop, DeepFake, and AI-generated edits without retraining.","key_machinery":"The operative mechanism is the semantic alignment of large vision-language models: the model's pretrained ability to connect language descriptions to visual content. REVEAL operationalizes this through two prompt families: holistic prompts that ask the model to assess whole-image consistency with physical and semantic expectations, and region-wise prompts that isolate parts of the image for local anomaly checks. The framework carries the argument by relying on these prompts rather than on task-specific fine-tuning, so any transferable performance across domains must come from the model's general world knowledge.","core_discovery":"The central claim is that the semantic alignment already present in large vision-language models is enough to detect and localize visual forgeries, provided the detection is framed as a reasoning task rather than a classification task. The paper proposes two complementary prompting approaches: Holistic Scene-level Evaluation, which probes the image's overall coherence with physics, semantics, perspective, and realism, and Region-wise Anomaly Detection, which splits the image into regions and examines each for local inconsistencies. Experiments span Photoshop, DeepFake, and AIGC editing datasets, and the authors report that vision-language models are competitive with existing baselines while","pith_inferences":["If prompt-based semantic alignment is sufficient, the practical implication is that forensic systems could be updated for new manipulation types by editing prompts rather than collecting and labeling new training sets.","The region-wise anomaly approach implies a harder, possibly more informative test: whether the model can localize the exact forged pixels, since fluent explanations might mask weak grounding.","The reasoning output also invites a new evaluation dimension—explanation plausibility versus pixel-level accuracy—that the paper's comparison to baselines likely does not fully settle.","A promising extension would be to test the same prompts on unseen manipulation classes or adversarially constructed forgeries, since the claimed transfer depends on the breadth of the model's world knowledge."],"forward_implications":["A single prompting framework could detect and localize multiple manipulation types—Photoshop, DeepFake, and AIGC—without retraining on each domain.","The model's natural-language reasoning output makes each forgery judgment interpretable, allowing users to see why an image was flagged.","Scene-level and region-wise evaluations can be combined to catch forgeries that global coherence checks might miss and vice versa.","Vision-language models could serve as off-the-shelf forensic tools, sidestepping the need for labeled forged images for every new manipulation technique."],"supporting_citations":[],"fun_headline_variants":["No retraining: prompts detect Photoshop, DeepFake, and AIGC edits","Reasoning prompts spot image forgeries across domains","Prompt-based reasoning exposes manipulated images without fine-tuning","Holistic + region-wise prompts reveal visual forgeries","Vision-language models detect fakes via prompt reasoning alone"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The framework assumes that the semantic alignment in large vision-language models is reliable enough to ground high-level cues like physics, perspective, and realism in actual image pixels, so that prompt-based general knowledge transfers across forgery domains without task-specific training.","fun_headline_variants_meta":{"raw":{"variants":["No retraining: prompts detect Photoshop, DeepFake, and AIGC edits","Reasoning prompts spot image forgeries across domains","Prompt-based reasoning exposes manipulated images without fine-tuning","Holistic + region-wise prompts reveal visual forgeries","Vision-language models detect fakes via prompt reasoning alone"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1194,"prompt_tokens":694,"completion_tokens":500,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":433}},"tokens_in":438,"tokens_out":500,"duration_ms":6465,"temperature":1.0,"reasoning_tokens":433,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:24:03.948001+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: on a held-out forgery dataset, run only the region-wise prompt and compare the model's predicted forged regions against human-annotated masks. If the model produces fluent explanations but its region masks have near-zero overlap with the ground truth while global scene-level accuracy stays high, the assumption that semantic alignment grounds local pixel evidence would be undermined.","supporting_citations":[],"review_version":1}