{"id":"26624656-97cc-43ef-839a-ee67b1b2af06","arxiv_id":"2412.04292","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A 300K-image social media deepfake benchmark plus a multimodal assistant, SIDA, that detects, localizes, and explains manipulated images with reported state-of-the-art accuracy.","lead":"The paper introduces SID-Set, a 300,000-image benchmark for detecting, localizing, and explaining deepfakes in social media images, and a vision-language model called SIDA that performs all three tasks. The authors report that SIDA outperforms prior deepfake detectors on the new benchmark and generalizes to an existing diffusion-image dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"If Language-SAM masks don't align with the regions Latent Diffusion actually edits, SIDA's localization metrics (Table 3) measure agreement with noisy labels, not true tampering accuracy.","rationale":"The central claim is that SIDA achieves superior performance in detection, localization, and explanation. The most load-bearing assumption is the validity of the localization ground truth in SID-Set. Section 3.2 describes generating masks with Language-SAM on the original image and then using Latent Diffusion to regenerate the image with a modified caption, but it never establishes that the changed pixels coincide with the SAM mask. If the mask is not used directly as an inpainting constraint, or if the diffusion edit does not respect it, then every localization metric in Table 3 (AUC 87.3, F1 73.9, IoU 43.8) is measured against noisy labels. The explanation ground truth is also affected, because the GPT-4o prompts for tampered images (Figure 14) explicitly include the mask as input, so the expert-reviewed descriptions inherit any mask error. This makes the concern load-bearing for two of the three advertised capabilities. I do not see a fatal flaw in the detection results; SIDA's performance is plausible and the baselines are standard. The Table 2 protocol ambiguity (fine-tuning 'on the validation set') is also worth clarifying, but it is a comparison-fairness issue rather than a validity issue of the dataset itself. The proposed test—comparing SAM masks to pixel-differenced edited regions on a sample—would directly settle the mask alignment question. Until that is done (or code/masks released), the localization and explanation claims should remain conditional.","tokens_in":24820,"tokens_out":6857,"duration_ms":64482,"concrete_test":"Take a random sample of ~200 tampered images from SID-Set. Compute the actual edited region as the pixel-wise difference between the original and tampered image (after alignment; e.g., using a diffusion-tuned difference or human annotation). Compare the Stage-2 Language-SAM mask with this true edited region using IoU. If the mean IoU is substantially below, say, 0.8, the ground truth masks are too noisy to support SIDA's localization AUC/F1/IoU claims, and Table 3 must be re-run on corrected masks. Additionally, release the Stage-4 generation code so the mask usage can be inspected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 (Stages 2–4) constructs tampered-image ground truth by first extracting object masks from original COCO/Flickr30k images with Language-SAM (Stage 2), then using Latent Diffusion to modify captions and regenerate images (Stage 4). The paper never specifies whether the Stage-2 mask is used as an inpainting mask in Stage 4, nor does it verify that the regenerated image's actually-altered pixels coincide with that mask. If the diffusion edit drifts outside the SAM mask (or the mask covers regions the edit leaves unchanged), then the 'ground truth' masks used to train and evaluate SIDA (Eq. 5, Table 3) are noisy. The reported AUC 87.3 / F1 73.9 / IoU 43.8 then measure agreement with these noisy labels, not true localization accuracy. This also contaminates the explanation fine-tuning: the GPT-4o prompts for tampered images (Appendix, Figure 14) feed the mask as input, so the 3,000 descriptions and SIDA's learned explanations inherit any mask error. The central 'superior performance' claim for localization, and the explanation capability itself, therefore rest on an unvalidated alignment assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces SID-Set, a 300K-image benchmark for social-media deepfake detection, localization, and explanation, and proposes SIDA, a LISA-based large multimodal model that adds <DET> and <SEG> tokens to perform three-way detection (real / fully synthetic / tampered), pixel-level tamper localization, and textual explanation. The authors report state-of-the-art detection results against eight baselines on SID-Set (Table 2), the best localization numbers among four comparison methods (Table 3), robustness to JPEG compression, resizing, and Gaussian noise (Table 4), and strong performance on the DMimage benchmark (Table 5). The dataset, model, and code are promised for release.","tokens_in":25033,"tokens_out":5664,"duration_ms":54588,"significance":"If the benchmark's localization labels are valid, SID-Set would be the largest social-media-focused deepfake dataset with masks and explanations, and SIDA would be the first single VLM to unify three-way detection, pixel-level localization, and textual explanation. The paper has notable strengths: extensive retraining of detection baselines on SID-Set, a perturbation robustness study, qualitative failure-case analysis, and a commitment to release the dataset, code, and model. The significance is moderated by unresolved questions about the alignment between Language-SAM masks and the regions actually altered by Latent Diffusion, by an ambiguous evaluation protocol in Table 2, and by an unfair cross-benchmark comparison in Table 5.","major_comments":[{"comment":"The construction pipeline does not state whether the Language-SAM mask from Stage 2 is used as an inpainting or conditioning mask in the Stage 4 Latent Diffusion regeneration, nor does it verify that the pixels actually changed by the diffusion edit coincide with the Stage-2 mask. Because the Stage-2 masks are used both as training targets for Lmask in Eq. (5) and as inputs to the GPT-4o explanation prompts (Appendix Figure 14), any drift between the edited region and the mask makes the reported localization AUC/F1/IoU and the learned explanations measures of agreement with noisy labels rather than true tampering accuracy. Please specify the conditioning in Stage 4 and provide a quantitative alignment check, such as mean IoU between Stage-2 masks and pixel-wise differences of original versus regenerated images, or a human-verified sample.","section":"Section 3.2, Stages 2-4 and Table 3"},{"comment":"The caption says parenthesized values are \"performance changes after fine-tuning the models using the SID-Set on the validation set.\" If the models were fine-tuned on the validation set and evaluated on the same validation set, the comparison is invalid because of train/test leakage; if they were fine-tuned on the training split and evaluated on the validation split, the wording should be corrected. Please report results on the held-out test split and specify the exact protocol for each baseline.","section":"Table 2 caption"},{"comment":"SIDA is compared with baselines using their original pre-trained weights after SIDA has been trained on 300K SID-Set images, so the comparison is not apples-to-apples. The proper generalization test would fine-tune all baselines on the SID-Set training split and report test-split numbers, or evaluate all methods in a zero-shot setting without SID-Set training. As written, the \"superior performance\" on DMimage may reflect a training-data advantage rather than a genuine generalization advantage.","section":"Section 5.4, Table 5"},{"comment":"The loss-weight specification is internally inconsistent. Implementation Details says the localization loss weights lambda_bce and lambda_dice are both set to 1.0, while Section 5.5 states \"for our experiments, we set lambda_det to 1, lambda_bce to 2.0, and lambda_dice to 0.5,\" and Table 7 lists yet another configuration. Please state unambiguously which configuration produced Tables 2-4 and report the corresponding validation or test numbers; otherwise the experimental results are not reproducible.","section":"Section 4.2 / Section 5 Implementation Details / Section 5.5"},{"comment":"The detection, localization, and ablation results are reported as single-run point estimates without standard deviations or the number of seeds. Given that the stated margins in Table 2 are large but the ablation differences in Table 6 are small (e.g., ACC 93.5 vs 91.1 for removing the attention module), please report mean and standard deviation over at least three runs, or otherwise justify that the reported differences are statistically meaningful.","section":"Tables 2-4 and Section 5.5"}],"minor_comments":[{"comment":"The \"Overall Acc and F1\" are described as the average of the values from the three classification categories; this is a macro average, not an overall accuracy, and the caption should say so explicitly to avoid misleading readers.","section":"Table 2 caption"},{"comment":"The 3,000 explanation ground-truth texts are generated by GPT-4o and reviewed by five experts, but the paper reports no quantitative evaluation of explanation quality, only qualitative examples in Section 5.6. Moreover, the prompt for tampered images in Appendix Figure 14 explicitly provides the mask as input, so the explanation stage is conditioned on the same localization labels whose validity is in question; this limitation should be acknowledged.","section":"Section 4.2 / Appendix Figures 7 and 14"},{"comment":"The paper describes SID-Set as having \"comprehensive annotations,\" but explanations are available for only 3,000 of the 300K images; the wording should be tempered to distinguish the mask annotations from the explanation annotations.","section":"Abstract and Section 3"},{"comment":"The ablation study for training weights reports only two configurations and does not compare against the Implementation Details configuration of lambda_bce=lambda_dice=1.0; please include that configuration or justify its omission.","section":"Section 5.5 and Table 7"},{"comment":"There are several minor typos and naming inconsistencies, including \"CnnSpott\" in Table 2 versus \"CnnSpot\" in the text, \"HIFI-Net\" versus \"HiFi-IFDL,\" and \"textural descriptions\" versus \"textual descriptions.\" Please proofread for consistency.","section":"Throughout"},{"comment":"The paper says MagicBrush images were \"integrated\" and \"filtered,\" but provides no details on the filtering criteria or the proportion of MagicBrush images in the final dataset; please specify this, and also cite the OpenImages V7 dataset rather than only a URL.","section":"Section 3.2 and Appendix D"},{"comment":"The expert evaluation is described only qualitatively; please report at least the number of images flagged and excluded, and ideally inter-annotator agreement, so that the quality-control claims can be assessed.","section":"Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The two load-bearing issues are the validity of the localization ground truth and the fairness of the cross-benchmark comparison; both are addressable in revision. I would also verify that the released dataset respects the licenses of the source data (OpenImages, COCO, Flickr30k, MagicBrush) before public release, since this is a dataset-contribution paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on SIDA. The dataset is the real contribution: SID-Set, 300K images with three-way real/synthetic/tampered labels, masks, and 3,000 expert-reviewed explanations. That is genuinely new and likely durable. The model itself is a sensible stitching of LISA with a detection head and a segmentation token — not a big architectural leap, but it works as a strong baseline on the benchmark.\n\nThe headline claim that SIDA achieves superior detection, localization, and explanation in one VLM outruns the evidence. Detection is fine: broad baseline set, clear margins, plausible. Localization is where I'd press. The pipeline extracts Language-SAM masks from original images, then uses Latent Diffusion to regenerate with modified captions. The paper never says whether the Stage-2 mask is used as an inpainting region in Stage 4, and never verifies that the pixels the diffusion actually changed coincide with the SAM mask. If they drift — which text-guided image-to-image editing can easily do — then the AUC/F1/IoU in Table 3 measure agreement with noisy labels, not true tamper localization. This is not a fatal flaw for detection, which doesn't depend on it, but it is a load-bearing assumption for the localization and explanation claims. The stress-test note lands on this exactly.\n\nTwo more soft spots, both real but more minor. First, Table 2's caption says fine-tuned baselines were evaluated on the validation set, while SIDA's own split is unclear; the comparison needs an explicit statement of which split each number comes from. No error bars or seeds anywhere, though the margins are large enough that this is probably a minor issue. Second, the explanation capability is not quantitatively evaluated: the training text is generated by GPT-4o with the mask as input, and the paper offers no measure of explanation faithfulness. As it stands, explanations are illustrative, not validated.\n\nTo be fair, the paper does a lot right: a careful data-generation pipeline, expert quality control, a robustness study, and reported failure cases. That is honest work. The dataset, if released as described, is a resource the field will use.\n\nBottom line: this deserves a serious referee, not a desk reject. A reviewer should ask for (1) verification or release of the tampered-region ground truth to confirm mask alignment; (2) a clear statement of which split the results come from; (3) some quantitative evaluation of explanation quality, even a human-rated one. The paper is for deepfake dataset builders and VLM-forensics researchers, and I'd use the dataset myself after those checks are addressed.","headline":"The dataset is the real contribution; the localization ground truth needs an alignment check before the headline claims are accepted.","tokens_in":25603,"tokens_out":3624,"would_cite":true,"duration_ms":36664,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that one large multimodal model can detect whether a social-media image is real, fully synthetic, or tampered, draw a pixel-level mask of the tampered region, and explain its verdict in text, and that the 300K-image…","keywords":["deepfake detection","tamper localization","large multimodal models","vision-language model","social media images","synthetic image dataset","image forensics","explainable AI"],"falsifier":"Take a random sample of SID-Set tampered images, have human annotators draw the actually changed pixels against the original photo, and recompute SIDA's localization AUC, F1, and IoU; if scores drop materially below the reported 87.3, 73.9, and 43.8, the masks used as ground truth did not match the true edits.","tokens_in":24622,"feed_emoji":"🕵️","tokens_out":4304,"duration_ms":38680,"temperature":0.7,"pith_summary":"This paper tries to establish that a single large multimodal model can do what previously required separate systems: decide whether a social-media image is real, fully synthetic, or tampered; draw a pixel mask over the tampered region; and say in plain language what cues led to that verdict. To test this, the authors build SID-Set, a 300,000-image benchmark of real, FLUX-generated, and locally tampered photos drawn from COCO and Flickr30k, with masks and textual explanations. They then introduce SIDA, a vision-language assistant built on LISA with two added tokens, <DET> and <SEG>, whose hidden features drive a detection head and a mask decoder. The reported result is that SIDA outperforms or matches existing detectors and localization methods on SID-Set and generalizes to the external DMimage benchmark, at 93.5% overall accuracy and 87.3 localization AUC.","feed_headline":"SIDA spots, localizes, and explains deepfakes in one model","feed_subtitle":"A 300K-image social-media benchmark plus a VLM that beats prior detectors on localizing tampered regions","key_machinery":"The load-bearing mechanism is the token-head architecture borrowed from reasoning-segmentation vision-language models: two added tokens, <DET> and <SEG>, whose last-layer hidden states are extracted and fed to a lightweight detection head and, via a single multihead-attention residual module, to a mask decoder conditioned on the frozen image encoder's features. On the data side, the machinery is the four-stage tampering pipeline: GPT-4o extracts objects from captions, Language-SAM produces masks on the original images, replacement dictionaries define object swaps and attribute changes, and Latent Diffusion regenerates the image with the altered caption. The masks from stage two are the training ground truth for localization.","core_discovery":"On the paper's own terms, the central discovery is that detection, localization, and explanation are not competing tasks but can be fused in one vision-language model by letting the detection representation condition the segmentation features. The model adds <DET> and <SEG> tokens to the vocabulary; the <DET> hidden state classifies the image into real, fully synthetic, object-tampered, or part-tampered, and when tampering is found, a multihead-attention module uses detection features as queries over segmentation features, then a frozen-image-encoder decoder produces the mask. The same model generates a textual justification. Trained end-to-end on SID-Set with cross-entropy for detection, BCE+DICE for masks, and cross-entropy for text, SIDA-7B and SIDA-13B reach 93.5-93.6% overall accuracy and set the top localization scores among compared methods.","pith_inferences":["If SID-Set's masks prove accurate, it could become the standard benchmark for non-facial image forensics, much as face-centric datasets were a decade ago.","The reliance on one generator for synthetic images and one diffusion model for tampering means SID-Set may bake in artifacts of those two generators; a testable next step is adding images from other generators and checking whether SIDA's advantage shrinks.","The explanation head is trained only on 3,000 GPT-4o descriptions, so SIDA's explanations likely reflect GPT-4o's judgment criteria rather than SIDA's own internal cues; a direct test would compare model explanations with its attention or occlusion maps."],"forward_implications":["A single SIDA-style model can replace separate detector and localization pipelines for social-media images, reducing the deployment cost of content moderation.","Because SIDA also outputs textual explanations, moderation systems gain a human-readable audit trail for every decision, not just a score.","SID-Set provides a training and evaluation ground for a wider variety of fakes than face-only datasets, covering fully synthetic and object- or part-tampered images.","Fine-tuning on SID-Set improves several prior detectors' performance, indicating the dataset is a useful transfer resource beyond SIDA itself.","The robustness experiments suggest SIDA tolerates JPEG compression, resizing, and Gaussian noise without explicit augmentation."],"supporting_citations":[{"why":"FLUX model that generated the 100K fully synthetic images in SID-Set.","marker":"[45]"},{"why":"Latent Diffusion used in stage four to regenerate tampered images from altered captions.","marker":"[58]"},{"why":"Language-SAM produced object masks on original images that serve as localization ground truth.","marker":"[31]"},{"why":"GPT-4o extracted objects from captions and generated the 3,000 textual explanation ground truths.","marker":"[49]"},{"why":"LISA, the base vision-language model whose reasoning-segmentation design SIDA extends with detection and segmentation tokens.","marker":"[30]"},{"why":"MagicBrush edited-image data added to SID-Set to increase diversity.","marker":"[82]"},{"why":"GenImage, the large prior synthetic-image benchmark that SID-Set compares against and extends.","marker":"[89]"},{"why":"DMimage, the external benchmark used to test SIDA's cross-dataset generalization.","marker":"[8]"}],"fun_headline_variants":["SIDA: one model that detects, localizes, and explains fakes","SIDA fuses deepfake detection, localization, and explanation","SIDA: the deepfake detector that also shows tampered regions and why","SIDA: 300K-image benchmark plus a VLM that excels at deepfake spotting","SIDA: a single VLM for deepfake detection, segmentation, and justification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central load-bearing premise is that Language-SAM masks computed on the original images align with the regions Latent Diffusion actually altered, so the reported localization scores measure true tampered-region accuracy rather than agreement with approximate labels.","fun_headline_variants_meta":{"raw":{"variants":["SIDA: one model that detects, localizes, and explains fakes","SIDA fuses deepfake detection, localization, and explanation","SIDA: the deepfake detector that also shows tampered regions and why","SIDA: 300K-image benchmark plus a VLM that excels at deepfake spotting","SIDA: a single VLM for deepfake detection, segmentation, and justification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1588,"prompt_tokens":995,"completion_tokens":593,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":491}},"tokens_in":611,"tokens_out":593,"duration_ms":6113,"temperature":1.0,"reasoning_tokens":491,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:32:47.483657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of SID-Set tampered images, have human annotators draw the actually changed pixels against the original photo, and recompute SIDA's localization AUC, F1, and IoU; if scores drop materially below the reported 87.3, 73.9, and 43.8, the masks used as ground truth did not match the true edits.","supporting_citations":[{"cited_title":"Flux model","cited_arxiv_id":null,"evidence_quote":"FLUX model that generated the 100K fully synthetic images in SID-Set."},{"cited_title":"High-resolution image syn- thesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Latent Diffusion used in stage four to regenerate tampered images from altered captions."},{"cited_title":"Lang-segment-anything","cited_arxiv_id":null,"evidence_quote":"Language-SAM produced object masks on original images that serve as localization ground truth."},{"cited_title":"GPT-4 technical report","cited_arxiv_id":null,"evidence_quote":"GPT-4o extracted objects from captions and generated the 3,000 textual explanation ground truths."},{"cited_title":"LISA: reasoning segmenta- tion via large language model","cited_arxiv_id":null,"evidence_quote":"LISA, the base vision-language model whose reasoning-segmentation design SIDA extends with detection and segmentation tokens."},{"cited_title":"Magicbrush: A manually annotated dataset for instruction- guided image editing","cited_arxiv_id":null,"evidence_quote":"MagicBrush edited-image data added to SID-Set to increase diversity."},{"cited_title":"Genimage: A million-scale benchmark for de- tecting ai-generated image","cited_arxiv_id":null,"evidence_quote":"GenImage, the large prior synthetic-image benchmark that SID-Set compares against and extends."}],"review_version":1}