{"id":"5426d855-3e0f-4d14-8a23-64f1f08c29f3","arxiv_id":"2608.07863","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LHSDet detects AI-generated images by fusing low-level texture features, SigLIP2 global features, and BLIP-2 captions in a visual question answering framework, reaching 96% average cross-model accuracy.","lead":"This paper introduces LHSDet, an AI-generated-image detector that treats detection as a visual question answering task and combines low-level texture, high-level visual, and caption-derived semantic features. The authors report about 96% average accuracy across diffusion, autoregressive, and social-media image sets, outperforming prior detectors by roughly 9 percentage points.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BLIP-2 is claimed to be LoRA-fine-tuned, but the caption string is non-differentiable and the paper says captions are pre-generated; the reported triple-branch VQA pipeline cannot be trained as stated.","rationale":"The paper's central claim is that LHSDet achieves 95-96% accuracy and beats the second-best baseline by more than 9% using a triple-branch VQA architecture with LoRA-fine-tuned SigLIP2, BLIP-2, and Phi-3-mini. For that claim to be meaningful, the evaluated system must actually be the described architecture. The BLIP-2 inconsistency is the most load-bearing weak point because it is an internal contradiction, not merely a missing confidence interval: the caption path is discrete, so BCE gradients cannot train BLIP-2 through Eq. (9), and the stated use of pre-generated captions contradicts any BLIP-2 fine-tuning. This directly affects reproducibility and the interpretation of Table V. I do not think the lack of repeated trials is the primary threat: with roughly 600 images per cell and a 9-11 point aggregate margin, ordinary sampling noise is unlikely to erase the headline gap, although split selection and code release still matter. The reader's weakest_assumption focused on statistical rigor, but the reader's rationale also flagged the BLIP-2 issue; hence partial agreement. The correct disposition remains CONDITIONAL, since the accuracy numbers could survive clarification but the paper must either justify the differentiable training path or explicitly state that BLIP-2 is frozen.","tokens_in":18936,"tokens_out":11777,"duration_ms":143561,"concrete_test":"Request or release the training code and LoRA adapters, and check whether any LoRA weights are actually attached to BLIP-2's q_proj and v_proj. Then run two identical training runs on the Flickr2K/SDv3 split: (A) BLIP-2 frozen with captions pre-generated once, and (B) BLIP-2 LoRA-fine-tuned with captions regenerated each epoch through a specified differentiable pipeline. Compare the full-model and w/o-Semantic accuracies from Table V. If A and B are indistinguishable, the BLIP-2 fine-tuning claim should be retracted and the semantic branch's contribution re-evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-E states that 'SigLIP2, BLIP2, and LLM are fine-tuned using LoRA,' and Fig. 1 marks BLIP2 with LoRA. However, Eq. (7) generates a discrete caption Tdesc and Eq. (9) tokenizes it, while the only training loss, the BCE loss in Eq. (14), is computed on the LLM's final-token hidden state. Gradients cannot flow back through discrete tokens to BLIP-2's LoRA parameters. The paper itself says in Sec. IV-E, 'Following the pre-generated captions to LHSDet, it requires approximately 20 minutes per epoch,' implying captions are produced once offline with an effectively frozen BLIP-2. If captions are pre-generated, BLIP-2 fine-tuning cannot affect them; if BLIP-2 is truly fine-tuned, the captions must be regenerated each epoch or a differentiable surrogate must be specified. Neither is described. This is not a stylistic detail: the semantic branch is presented as a trainable component of LHSDet, and Table V shows that removing it changes average accuracy by about two points. Without clarification, the reader cannot tell whether the evaluated system is the proposed architecture or a variant with a frozen captioner. The claimed 'fine-tuned vision-language framework' is therefore internally inconsistent and the reported ablation does not establish the contribution attributed to the semantic branch.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes LHSDet, a triple-branch detector for high-resolution AI-generated images. The low-level branch aggregates non-overlapping patches with SRM filters and convolutions to capture texture artifacts; the high-level branch uses a SigLIP2 encoder with LoRA; the semantic branch generates captions with BLIP-2; and the combined multi-modal features are fed to a Phi-3-mini LLM that outputs a binary real/fake answer in a VQA format. The paper reports state-of-the-art accuracy of 95–96% on eleven datasets spanning diffusion models, an autoregressive model (Infinity), and the WildRF social-media benchmark, claiming a >9% margin over the second-best baseline, plus strong robustness to post-processing. The training protocol uses BCE loss on the LLM's final-token hidden state, with LoRA applied to SigLIP2, BLIP-2, and the LLM.","tokens_in":19246,"tokens_out":6021,"duration_ms":66498,"significance":"If the results hold, the paper makes a useful contribution to AI-generated image detection by combining low-level texture cues with high-level semantic and textual cues in a high-resolution VQA framework, and by evaluating on genuinely held-out generators rather than only on training-domain data. The ablation study (Table V) and robustness experiments (Tables III–IV) are informative. However, the current manuscript contains a load-bearing inconsistency about how BLIP-2 is fine-tuned, and the comparison against two CLIP-based baselines is unequal because they are evaluated with official pretrained checkpoints rather than retrained under the same few-shot protocol. The central accuracy claims are likely to survive a corrected comparison given the large margins, but the paper as written does not yet establish the proposed training procedure or the claimed SOTA in a sound manner.","major_comments":[{"comment":"The table captions claim that 'All methods were trained on Flickr2K and SDv3,' yet UnivFD and Fatformer are evaluated using their official pretrained checkpoints, as indicated by the † footnote. This is not a like-for-like comparison under the few-shot training protocol used for the other baselines and for LHSDet. A method that has never seen the training distribution is at a substantial disadvantage, and this asymmetry directly undermines the headline claim that LHSDet surpasses the second-best baseline by more than 9%. The authors should either retrain these baselines on the same training splits (Flickr2K/DIV2K plus SDv3) or explicitly remove the 'all methods trained' claim and qualify the performance comparison as a mixed-protocol benchmark.","section":"Tables I and II, footnote †"},{"comment":"No error bars, confidence intervals, repeated runs, or significance tests are reported, and the test sets contain only 300, 100, or 50 samples per category. In Table IV, with 50 samples per class, a single misclassification changes the reported accuracy by 1%, so the listed differences (e.g., LHSDet 95.33% vs. Antifake 81.17%) are noisy estimates of the true performance gap. The large observed margins make the central conclusions plausible, but the paper should report the variance across multiple random splits or repeated training runs, and should specify the split procedure and any fixed seed, in order to support the claimed margins and the robustness conclusions.","section":"Sec. IV-C, Tables I–IV"}],"minor_comments":[{"comment":"The two convolutions in Eq. (1) are both named Conv2, which is confusing; they should be distinguished (e.g., Conv2 and Conv3), and the patch size h×w used for the non-overlapping patch aggregation is never specified, so the local texture block is under-specified for replication.","section":"Sec. III-A.1, Eq. (1)"},{"comment":"The comparison would be strengthened by including recent high-resolution or training-free detectors beyond the listed ones (e.g., LARE², RIGID, or other 2024–2025 methods) to substantiate the 'state-of-the-art' claim, especially since the paper targets 1024×1024 inputs.","section":"Sec. IV-A.3, Tables I and II"},{"comment":"The robustness protocol is asymmetric: blur, JPEG, and noise are applied after downsampling to 256×256, while brightness, contrast, and rotation are applied at 1024×1024. This is a reasonable design choice, but the motivation for the asymmetry should be stated explicitly in the text.","section":"Sec. IV-D"},{"comment":"The sentence 'By default, SigLIP2 processes 1024×1024 images at 384×384 resolution' is confusing; it should be rewritten to say that SigLIP2 is designed for high-resolution inputs but is applied here at 384×384 resolution, or similar.","section":"Sec. III-A.2"},{"comment":"The manuscript contains numerous typos (e.g., 'evolved' as 'evoled', 'characteristic' as 'characterstic', 'another' as 'anothor', 'Illustation' in the Fig. 3 caption, 'neighboring' as 'neiboring') and a few redundant words; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The BLIP-2 LoRA inconsistency is the key technical blocker and should be resolved before acceptance, either by describing a valid training path through the captioner or by explicitly freezing BLIP-2 and adjusting the ablation interpretation. The unequal baseline protocol (official checkpoints vs. retrained methods) also needs to be addressed. The empirical margins are large, so I believe the paper can be made acceptable after a major revision that clarifies these points and strengthens the evaluation with variance reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid engineering paper with a real empirical gap, but there is a load-bearing inconsistency in the training description that has to be fixed before the results can be trusted. The high-level claim—detecting high-resolution AI-generated images by combining low-level texture, high-level visual, and semantic caption features in a VQA setup—delivers 95–96% average accuracy across diffusion, autoregressive, and WildRF sets, about 9 points above the best baseline. That is a practically meaningful result if it holds.\n\nWhat is genuinely new is the specific combination: non-overlapping patch/SRM texture features, SigLIP2 global features, BLIP-2 captions, cross-attention fusion, and a LoRA-tuned Phi-3 as the reasoning head. None of these parts is new, but I do not know of another paper putting them together exactly this way. The evaluation is also broader than most: 11 generated datasets plus the WildRF social-media benchmark, and a robustness table against blur, JPEG, noise, and rotation. The ablations are informative, showing both visual branches matter and that cross-attention beats gated fusion.\n\nThe soft spots are real. First and most important: the paper says BLIP-2 is LoRA-fine-tuned, but the loss is BCE on the LLM's last-token hidden state, and the caption is a discrete token sequence. Gradients cannot flow back to BLIP-2's LoRA parameters. Section IV-E then says captions are pre-generated, which means BLIP-2 is effectively frozen. Either the captioner is frozen and the claim of fine-tuning is false, or the training procedure is not described. The ablation says removing the semantic branch costs about two points, so this is not cosmetic—if BLIP-2 is frozen, the evaluated system is not the proposed triple-branch trainable architecture. This needs a clear fix: either pre-generate captions and call BLIP-2 frozen, or specify a differentiable captioning loss and regenerate captions each epoch.\n\nSecond, the statistics are thin. Test sets of 300, 100, or 50 per category, one run, no error bars or significance tests. The 9-point margin is large, so I suspect it would survive noise, but small sets with no variance make that a guess. Third, UnivFD and Fatformer are evaluated with official checkpoints, not retrained; the dagger is disclosed, but it is still an unfair comparison for a few-shot training claim.\n\nI do not think the core empirical direction is wrong. The architecture is reasonable, the evaluation is broad, and the gains are consistent. The BLIP-2 issue is fixable, but it must be fixed. My recommendation: this deserves peer review, not desk rejection. The referee should ask for code, clarification of the caption-generation pipeline, and error bars.","headline":"Useful engineering result with a real 9-point cross-model gap, but the claimed trainable BLIP-2 caption branch is not trainable as described; fix that before trusting the ablation.","tokens_in":19767,"tokens_out":2563,"would_cite":false,"duration_ms":27136,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LHSDet claims that high-resolution AI-generated image detection is best done by a triple-branch visual question answering system that fuses low-level texture, high-level global, and semantic caption features, achieving 95–96% accuracy…","keywords":["AI-generated image detection","visual question answering","high-resolution images","diffusion models","autoregressive models","multi-modal feature fusion","LoRA fine-tuning","SigLIP2"],"falsifier":"Re-run the Flickr2K and DIV2K experiments under the exact protocol but with multiple random train/test splits, or enlarge each test set to several thousand samples, and check whether the 9–11 point gap over AEROBLADE and AntifakePrompt persists; a gap that collapses under resampling would falsify the claim of state-of-the-art generalization.","tokens_in":18756,"feed_emoji":"🖼️","tokens_out":5533,"duration_ms":51822,"temperature":0.7,"pith_summary":"This paper proposes LHSDet, a detector for high-resolution AI-generated images that treats the task as a visual question answering problem. The central claim is that combining three complementary feature streams—low-level texture cues, high-level global perception, and semantic-level captions—lets a fine-tuned vision-language model tell real from synthetic images better than existing methods. On a battery of diffusion, autoregressive, and real-world social-media datasets, the method reports 95–96% average accuracy, beating the second-best baseline by roughly 9 to 11 percentage points. This matters because high-resolution synthetic images are increasingly hard to flag, and existing detectors either downsample away fine texture artifacts or rely on reconstruction models that generalize poorly to unknown generators.","feed_headline":"Triple-branch VQA detector beats fake-image baselines by 9 points","feed_subtitle":"LHSDet fuses texture, global, and caption cues to hit 95-96% accuracy on unseen generators.","key_machinery":"The central mechanism is the triple-branch VQA architecture: a local texture branch using non-overlapping patch aggregation with SRM noise filters and neighboring-pixel convolutions, a global perception branch using LoRA-fine-tuned SigLIP2, and a semantic branch producing BLIP-2 captions; these are fused through cross-attention, projected into a Phi-3-mini LLM, and answered as 'Yes' or 'No'. The cross-attention fusion, which uses global features as the query and local features as key and value, is the load-bearing integration step that lets high-level semantics be enriched by fine texture detail.","core_discovery":"LHSDet's core claim is that high-resolution AI-generated image detection is best served by a triple-branch, multi-modal design rather than a single visual encoder. A low-level branch aggregates non-overlapping image patches and computes SRM high-pass noise maps plus neighboring-pixel inconsistency features to capture texture artifacts; a high-level branch uses the SigLIP2 visual encoder, fine-tuned with LoRA, for global semantic inconsistencies; and a semantic-level branch generates a caption with BLIP-2, which is concatenated with a fixed authenticity question. The two visual streams are fused via cross-attention, projected into the Phi-3-mini LLM embedding space, and the LLM outputs a 'Yes' or 'No' answer. The paper reports state-of-the-art accuracy of 95–96% across diffusion models (SDv3, SDXL, Playground, DALL·E 3, MidJourney v5/v6, IF), the autoregressive model Infinity, and the WildRF real-world benchmark, outperforming the second-best baseline by more than 9%.","pith_inferences":["If the small-test-set issue is resolved by larger evaluation, the cross-attention fusion between local and global branches could be reused as a plug-in visual encoder for other forensic tasks, such as deepfake video frame detection or face forgery localization.","The semantic caption branch is not trained for detection, yet it supplies context that helps the LLM reason about implausible content; a testable hypothesis is that swapping in a captioning model with stronger world knowledge would further improve detection of semantically inconsistent images with no change to the visual branches.","The paper does not test against adversarial post-processing such as semantic inpainting or model-specific artifact removal; a natural next experiment is whether the texture branch remains discriminative when artifacts are explicitly minimized.","Since the reported gains are computed against a fixed set of baselines on narrow test sets, an independent benchmark with larger and more diverse splits would clarify whether the 9-point margin is a property of the architecture or an artifact of the evaluation protocol."],"forward_implications":["If LHSDet is right, high-resolution detectors no longer need to downsample inputs; the patch-aggregation branch preserves texture cues at 1024×1024 resolution, and the same architecture should extend to 2K/4K images without modification.","The VQA reformulation means detection inherits the few-shot and reasoning abilities of modern vision-language models; training on only 4,000 images for two epochs suffices, so detectors can be rapidly retargeted to newly emerged generative models.","The 9–11 percentage point accuracy margin over reconstruction-based baselines suggests that learned multi-modal features transfer across unseen generators more reliably than reconstruction-error heuristics.","The robustness results, with an average accuracy drop of about 3.5% under blur, JPEG compression, noise, rotation, brightness, and contrast changes, indicate that the method tolerates common post-processing that degrades CNN-based detectors by tens of points.","The improvement over AntifakePrompt, which also uses a VQA formulation, attributes the gain to the redesigned visual encoder and the added low-level texture branch, implying that visual encoder design matters more than the prompting strategy alone."],"supporting_citations":[{"why":"AntifakePrompt reformulates AI-image detection as a VQA task and is the main VLM baseline; LHSDet builds on the same formulation and surpasses it on nearly all datasets.","marker":"[32]"},{"why":"AEROBLADE is the training-free high-resolution detector that is the second-best baseline in Tables I and II, against which LHSDet claims a 9–11 point improvement.","marker":"[19]"},{"why":"SigLIP2 supplies the high-level visual encoder that LHSDet fine-tunes with LoRA for global perception features.","marker":"[39]"},{"why":"BLIP, the captioning foundation, provides the semantic-level textual branch that generates image descriptions for the VQA prompt.","marker":"[31]"},{"why":"LoRA is the parameter-efficient fine-tuning method applied to SigLIP2, BLIP-2, and Phi-3-mini, allowing adaptation with only 0.23% trainable parameters.","marker":"[45]"},{"why":"DIRE is a representative reconstruction-error baseline that anchors the comparison for the RE-based detector class and is outperformed by LHSDet on generalization and robustness.","marker":"[18]"}],"fun_headline_variants":["VQA detector outperforms baselines by 9 points on high-res fakes","LHSDet fuses texture, global, and caption cues to spot AI fakes","High-res AI fakes detected via triple-branch VQA","Ask the model: is it real? LHSDet answers with 95-96% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy margins rest on test sets of 300, 100, or 50 images per category, evaluated once without repeated trials or confidence intervals; if those small splits are noisy or unrepresentative, the claimed 9-percentage-point lead over the second-best baseline could shrink.","fun_headline_variants_meta":{"raw":{"variants":["VQA detector outperforms baselines by 9 points on high-res fakes","LHSDet fuses texture, global, and caption cues to spot AI fakes","High-res AI fakes detected via triple-branch VQA","Ask the model: is it real? LHSDet answers with 95-96% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00067,"raw_usage":{"total_tokens":3098,"prompt_tokens":1037,"completion_tokens":2061,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":1974}},"tokens_in":653,"tokens_out":2061,"duration_ms":16421,"temperature":1.0,"reasoning_tokens":1974,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:45:04.908415+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Flickr2K and DIV2K experiments under the exact protocol but with multiple random train/test splits, or enlarge each test set to several thousand samples, and check whether the 9–11 point gap over AEROBLADE and AntifakePrompt persists; a gap that collapses under resampling would falsify the claim of state-of-the-art generalization.","supporting_citations":[{"cited_title":"Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error,","cited_arxiv_id":null,"evidence_quote":"AEROBLADE is the training-free high-resolution detector that is the second-best baseline in Tables I and II, against which LHSDet claims a 9–11 point improvement."},{"cited_title":"Dire for diffusion-generated image detection,","cited_arxiv_id":null,"evidence_quote":"DIRE is a representative reconstruction-error baseline that anchors the comparison for the RE-based detector class and is outperformed by LHSDet on generalization and robustness."}],"review_version":1}