{"id":"26acbe09-d278-4d92-ae38-980b970fe910","arxiv_id":"2509.01107","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A frequency-guided layout-to-image generation framework, FICGen, improves fidelity, layout alignment, and detector trainability on degraded scenes across five benchmarks.","lead":"This paper adapts layout-to-image diffusion generation to degraded scenes like low light, underwater, fog, and blur, by splitting images into high-frequency instance details and low-frequency context and injecting both into a frozen Stable Diffusion model. The reported result is better layout alignment and better downstream detector training than prior layout-to-image methods across five degraded benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prototype dictionary in §3.2 injects real training-set instance crops into generated images; alignment/trainability gains may be retrieval leakage, not frequency disentanglement.","rationale":"The reader correctly identifies the untested fixed Fourier mask (Eq. 3) as a weakness, but the most load-bearing threat to the central claim is the prototype construction in §3.2. FICGen's frequency prototypes are derived from real, same-class instances sampled from the training set. This makes every generated image, including those used for alignment evaluation and downstream trainability, conditioned on actual training-set crops. The alignment numbers (e.g., FICGen exceeding the real ExDARK Oracle by 5.3 mAP) become plausible only if the detector recognizes these real crops. The trainability protocol (§4.2.1, Table 4) trains detectors on real plus FICGen-generated images, where the generated set is built from the same real training instances; this is equivalent to increasing the epoch count on the real training set, so the observed +2-3 mAP gains are expected even without any domain-adaptive generation. The paper provides no comparison against a trivial copy-paste baseline and no ablation of the prototype source. This is an internal-validity threat, not a disagreement with consensus. The proposed concrete test—comparing against copy-paste and using held-out prototypes—would settle it. Given the paper's broad empirical coverage and the plausibility of the frequency-disentanglement idea when prototypes are properly sourced, I do not recommend outright rejection; the verdict should remain conditional, with the required control experiments as conditions.","tokens_in":22655,"tokens_out":5465,"duration_ms":62703,"concrete_test":"Run a copy-paste baseline: for each layout, paste the same real instance crops used to build FICGen's prototypes into the target bboxes (optionally using the instance coherence map), and evaluate FID, alignment AP, and trainability under the identical protocol of §4.1. If this baseline matches or exceeds FICGen, the central claim fails; the gains come from training-set retrieval. Also, repeat the trainability experiment with prototypes drawn only from a held-out validation split; if downstream mAP drops substantially, the synthetic set leaks training data into downstream training.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim of consistent gains in alignment and trainability is confounded by the prototype construction in §3.2. FICGen builds class-specific dictionaries by collecting same-class instances from the training set, samples a real degraded instance for each bbox, and derives HF/LF prototypes from its features. At inference (including the trainability protocol in §4.2.1), these prototypes are drawn from the training split and injected into the frozen LDM. Thus the generated 'synthetic' images are conditioned on actual training-set crops of the target classes. Alignment is then measured by pre-trained detectors on synthetic test images that contain real training instances; the detector may recognize these real crops more easily than truly novel generations, which could explain why FICGen exceeds the Oracle on ExDARK (42.5 vs 37.2 mAP). Trainability is measured by training detectors on real plus FICGen-generated images; since the generated set is built from the same real training instances, this doubles the effective exposure to the training data, inflating downstream gains. The paper does not compare against a simple copy-paste baseline (paste the sampled real crops into the layout), nor does it ablate the source of the prototypes. If a copy-paste baseline matches or beats FICGen, the claimed benefits are not due to frequency-inspired contextual disentanglement but to retrieval of real instances.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FICGen, a layout-to-image (L2I) generation method targeting degraded scenes (low-light, underwater, aerial, weather, blur). It is built on a frozen Stable Diffusion v1.5 latent diffusion model and injects frequency prototypes into the denoising process through a dual-query frequency perceiver resampler, a visual-frequency enhanced attention module, an instance coherence map, and an adaptive spatial-frequency aggregation module. The authors evaluate FID fidelity, layout alignment via off-the-shelf detectors, and downstream detector trainability on five degraded benchmarks plus natural VOC 2012, and report consistent improvements over MIGC and CC-Diff, along with robustness to unseen layouts and occlusion levels.","tokens_in":22925,"tokens_out":9054,"duration_ms":93019,"significance":"If the reported gains are truly attributable to the proposed frequency-inspired contextual disentanglement, FICGen would be a practical plug-and-play data-augmentation tool for degraded-scene object detection using a frozen LDM. The empirical surface is broad: five degraded datasets, four detector families, multiple trainability protocols, occlusion-level breakdowns, and efficiency comparisons. The method also includes ablations of its three main modules. However, the central attribution of the gains is compromised by the prototype construction in Sec. 3.2, which injects real training-set instance crops into every generated image, and by an unvalidated fixed frequency-mask hyperparameter. These issues must be resolved before the headline claims can be accepted.","major_comments":[{"comment":"The prototype construction confounds the central claim. FICGen builds class-specific dictionaries by collecting same-class instances from the training set, samples a real degraded instance for each bbox, and derives the HF/LF prototypes used in Eqs. (5)–(6) from those real crops. At inference and in the trainability protocol, these prototypes are drawn from the training split and injected into the frozen LDM. Synthetic test images therefore contain real training instances. This explains suspicious results such as FICGen exceeding the Oracle on ExDARK (42.5 vs. 37.2 mAP in Table 1) and inflates the trainability comparison, since the generated set is built from the same real instances used for training the detector, effectively doubling exposure. The paper does not compare with a copy-paste baseline (simply pasting sampled real crops into the layout) nor ablate the prototype source. Withou","section":"§3.2, §4.2.1, Table 1"},{"comment":"The frequency-disentanglement mechanism depends on a single fixed binary Fourier mask parameterized by γ, but γ is never reported, ablated, or validated. The same mask is applied to all classes and all degradation types, even though the paper argues that degradation alters the HF/LF balance (Fig. 2). If blur spreads high frequencies, backgrounds are textured, or objects themselves are low-frequency blobs, the prototypes encode the wrong cues and the claimed benefit disappears. Please report γ for each dataset, ablate it, and provide a quantitative check that the HF/LF split tracks instance/context (e.g., overlap between HF prototypes and instance masks).","section":"§3.2, Eq. (3)"},{"comment":"The abstract claims FICGen 'consistently surpasses existing L2I methods in terms of generative fidelity.' The paper's own FID numbers contradict this. On DIOR-H (Table 1) FICGen achieves FID 31.25 vs. CC-Diff 30.88, and on natural VOC 2012 (Table 8) FICGen achieves 48.93 vs. CC-Diff 48.70. Thus in two of the reported settings FICGen is worse on fidelity than the closest baseline. The fidelity claim should be qualified, or a fidelity-improving variant should be added.","section":"Abstract, Table 1, Table 8"},{"comment":"The most practically relevant trainability setting in Table 4, 'Train with Real & Synthetic Data,' does not support the claim that FICGen consistently achieves the highest gains. In that block FICGen obtains mAP 50.5, below MIGC (51.7) and CC-Diff (50.9). The text instead highlights per-class gains and states 'the last three rows' of Table 4 demonstrate the benefit, which is misleading. The claims need to be reconciled with the full table, and the authors should explain why FICGen underperforms in the mixed real+synthetic protocol while leading in other protocols.","section":"§4.2.1, Table 4"},{"comment":"No error bars, multiple seeds, or statistical tests are reported. Several headline differences are small (e.g., Table 5 DAWN mAP 25.9 vs. 25.8; Table 4 VOC mAP differences of less than 1 point). With no code or configuration files released and no seed variability, it is impossible to assess whether these differences are significant. At minimum, report standard deviations over at least three seeds for the main fidelity/alignment/trainability tables.","section":"§4.2.1, Tables 1–5"}],"minor_comments":[{"comment":"The text refers to the visual-frequency enhanced attention module as '(Fig.3(c))', but Fig. 3(c) is the Adaptive Spatial-Frequency Aggregation module; the attention module appears to be Fig. 3(b). Please correct the reference.","section":"§3.3"},{"comment":"The phrase 'As shown in the last three rows of Tab. 4' is ambiguous because Table 4 has four blocks (Full Real Data, Pure Synthetic, Synthetic + Finetune, Real & Synthetic), each with different row counts. Please refer to the specific rows by setting name.","section":"§4.2.1"},{"comment":"The logical combination of the two coordinate conditions in the binary mask is unclear: is MF=1 when both u and v are outside the central band, or when either is outside? Also, the text says '|v| and |u| denote the width and height of normalised frequencies,' which is confusing; please define the coordinates precisely.","section":"Eq. (3)"},{"comment":"FID scores are computed with different Inception variants (standard vs. RSICD-finetuned for remote sensing). This detail is only in the supplement; the main text should state that cross-dataset FID comparisons are not meaningful.","section":"Supp. B"},{"comment":"The paper claims FICGen 'can be seamlessly integrated as a plug-and-play component to enhance other diffusion-based generative models,' but only SDv1.5 is used as the base model. This claim is not demonstrated and should be either supported or removed.","section":"§1"},{"comment":"No code or configuration files are released. Given the number of hyperparameters (γ, N=15, classifier-free guidance scale, training schedule), reproducibility would benefit from releasing the exact settings, especially the unreported γ value.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the prototype-retrieval design in Sec. 3.2 is a confound rather than a contribution. If a copy-paste baseline matches FICGen, or if removing training-set prototypes eliminates the gains, the paper's core narrative collapses. I would ask the authors to add these controls before making a final decision. Also, the inconsistency in Table 4 (Real & Synthetic block) should be checked carefully; it is not a minor wording issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The frequency-inspired idea is well-motivated, and the empirical scope is genuinely broad—five degraded datasets, multiple off-the-shelf detectors, trainability protocols. But there is a load-bearing confound the paper never rules out: the prototypes used to condition the frozen diffusion model are built by collecting same-class instances from the training split and sampling a real crop for each bbox. So every generated image is conditioned on actual training-set instances. Alignment is then measured by detectors on synthetic test images that may contain those real crops, and trainability doubles exposure to the real training instances. A simple copy-paste baseline—paste the sampled real crop into the layout, with or without the same frequency filtering—would separate frequency disentanglement from retrieval leakage. It is not in the paper, and neither is an ablation of the prototype source. The central mechanism claim is unproven as it stands.\n\nWhat is new: the specific combination of dual-query frequency perceiver resamplers, visual-frequency enhanced attention, instance coherence map, and adaptive spatial-frequency aggregation is new and reasonable. The paper also adds useful evidence that a frozen SD with an adapter can generate plausible degraded imagery for data augmentation. The downstream gains, if real, are valuable for low-light and remote-sensing detection.\n\nSoft spots, in order: (1) The prototype-source confound above. (2) No seeds or error bars, and no code—a big deal for a systems paper. (3) The Fourier mask parameter gamma in Eq. 3 is never reported or ablated, despite being the one parameter that encodes the HF/LF split. (4) The “consistently surpasses” claim is contradicted by the paper’s own Table 1: on DIOR-H, FICGen's FID (31.25) is worse than CC-Diff's (30.88). (5) Table 4 has internal inconsistencies: in the “finetune on real” block RetinaNet-PVT-M FICGen is 47.5 vs CC-Diff 47.9, and in the “real & synthetic” block FICGen is 50.5 vs MIGC 51.7, yet the text claims FICGen consistently achieves the highest gains.\n\nWho it’s for: researchers doing data-centric detection in degraded domains and anyone working on layout-to-image, especially those worried about training-set leakage in generative augmentation. It deserves a serious referee, but the review should require a prototype-source ablation and a copy-paste comparison before publication. I would not cite it as evidence for the frequency-disentanglement mechanism until that control exists.","headline":"FICGen's frequency-disentanglement story is appealing and the experiments are broad, but the prototype dictionary pulls real training instances into every generated image, so the headline gains are not yet cleanly attributable to the method.","tokens_in":23443,"tokens_out":3287,"would_cite":false,"duration_ms":37306,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Separating high-frequency instance detail from low-frequency context in the Fourier domain, and injecting both into a frozen diffusion model, fixes layout-to-image generation in degraded scenes and improves downstream detection.","keywords":["layout-to-image generation","degraded image synthesis","frequency disentanglement","latent diffusion models","contextual illusion dilemma","low-light and underwater imaging","frequency prototypes","data-centric object detection"],"falsifier":"Swap the two prototype channels in a trained model — route the high-frequency instance prototypes into the context branch and the low-frequency context prototype into the instance branch — and measure detector alignment on the same synthetic test sets; if scores do not drop sharply, the claimed instance/context frequency assignment is not what carries the gain. A complementary check: sweep the mask parameter γ of Eq. 3 over a wide range, including per-class masks; if alignment stays flat, the binary spectral split is not the operative mechanism.","tokens_in":22530,"feed_emoji":"🌆","tokens_out":16345,"duration_ms":155649,"temperature":0.7,"pith_summary":"The paper targets layout-to-image (L2I) generation in degraded scenes — low light, underwater, aerial, adverse weather, blur — where prior generators entangle foreground objects with their surroundings and produce wrong object counts, positions, or interactions, a failure it names the 'contextual illusion dilemma.' Its explanation is spectral: degraded images attenuate the high-frequency detail that defines objects while low-frequency context dominates, so FICGen splits the two in the Fourier domain and feeds both back into a frozen latent diffusion model as conditioning tokens. The design combines frequency prototypes extracted from training exemplars, a learnable dual-query resampler, a visual-frequency enhanced attention with an instance coherence map, and an adaptive spatial-frequency aggregation module. Across five degraded benchmarks the paper reports that FICGen surpasses existing L2I methods in fidelity (FID), layout alignment (detection AP), and downstream trainability, including a ~2.0 mAP gain when synthetic images are added to detector training. The reader should care because a plug-and-play adapter that produces on-demand degraded training data would relieve the severe data scarcity in extreme-condition perception.","feed_headline":"Splitting image frequencies makes low-light scenes obey layouts","feed_subtitle":"FICGen's frequency prototypes make low-light and underwater images obey layouts, lifting detector scores by ~2 points.","key_machinery":"The load-bearing object is the frequency prototype pair (p↑, p↓): training-set instance features are transformed into the Fourier domain and split by a single fixed binary mask whose parameter γ sets the passband, encoding the premise that instance detail lives in high frequencies and context in low frequencies. Around it sit three mechanisms: a learnable dual-query mechanism whose HF and LF frequency perceiver resamplers turn prototypes into frequency-aware tokens; visual-frequency enhanced attention injecting those tokens into the frozen latent-diffusion U-Net, with an instance coherence map (a hard per-box 0/1 mask) blocking feature leakage between instances and background; and the adapti","core_discovery":"FICGen's diagnosis: in degraded scenes, high-frequency object detail is attenuated while low-frequency context dominates, so layout generators entangle instances with surroundings and hallucinate wrong counts, positions, and interactions — the 'contextual illusion dilemma.' Its answer: extract high-frequency instance and low-frequency context prototypes with a binary Fourier mask, refine them through dual-query resamplers, and inject them into a frozen Stable Diffusion U-Net via visual-frequency attention, an instance-coherence mask, and adaptive spatial-frequency aggregation. Across five degraded benchmarks FICGen beats prior L2I methods — 27.6 vs 23.6 mAP over CC-Diff on DIOR-H — and gives","pith_inferences":["The fixed mask parameter γ is never reported, ablated, or validated against image statistics; a natural test the authors left open is to learn or sweep the mask per class and per degradation, which would reveal whether the binary spectral split — rather than the added parameters — carries the gain.","The spectral premise implies a testable ranking: gains should be largest where instance and background spectra are most separated (small crisp objects on flat, low-texture backgrounds) and smallest where blur or textured backgrounds smear frequencies; the paper's own per-class tables could be re-analysed to check this prediction.","If the mechanism is degradation-agnostic as described, it should transfer beyond the five benchmarks to haze, night-vision, medical, or industrial imaging — an extension the paper does not run.","The paper's acknowledged failure case (tiny vehicles lost to latent downsampling) indicates Fourier prototypes cannot restore detail destroyed by the autoencoder, pointing to a companion high-resolution or super-resolution latent branch as the next logical step."],"forward_implications":["One adapter trained once on a frozen diffusion backbone covers multiple degradations — the same FICGen weights are reported to beat per-scenario baselines on low-light, underwater, aerial, weather, and blur benchmarks.","Synthetic degraded images can serve as auxiliary training data: mixing them into detector training yields ~2.0 mAP overall and larger per-class gains (e.g., +5.7 AP for 'bus' on DAWN), directly addressing data scarcity.","Because the design leaves the pretrained diffusion model frozen, FICGen is claimed to be a plug-and-play component for other diffusion-based generators.","Layout control extends beyond the training distribution: flipped, translated, and oriented boxes still generate coherent scenes, so annotated layouts can be augmented to create labeled data on demand.","On ExDARK the synthetic images align with their layouts better than the real test set does (AP 42.5 vs 37.2 oracle), implying synthetic data can be not merely a proxy but a closer match to the layout condition."],"supporting_citations":[{"why":"supplies the frozen latent-diffusion backbone (SDv1.5) that FICGen adapts without retraining","marker":"[33]"},{"why":"the multi-instance generation baseline FICGen must beat, and a source of the per-box instance masks","marker":"[60]"},{"why":"the contextual-coherence baseline whose foreground–background modeling FICGen extends with frequency cues","marker":"[58]"},{"why":"the grounded text-to-image adapter paradigm whose layout-conditioning design FICGen builds on","marker":"[16]"},{"why":"the remote-sensing layout-conditioned baseline and the source of the per-image instance cap","marker":"[41]"},{"why":"the Perceiver-style resampler architecture that inspires the dual-query frequency resamplers","marker":"[53]"},{"why":"supplies the trainability protocol (synthetic/real mixing and bbox augmentation) used for downstream evaluation","marker":"[3]"},{"why":"the ExDARK low-light benchmark, the flagship degraded domain for alignment and trainability claims","marker":"[19]"}],"fun_headline_variants":["Frequency split fixes AI's layout mistakes in low-light images","Disentangling frequencies makes generated scenes obey user layouts","FICGen separates image frequencies to beat prior layout-to-image models","New frequency-aware method improves degraded image generation","Frequency-aware generation lifts detector mAP by 4 points in low light"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole pipeline rests on the assumption that one fixed frequency-filtering setting cleanly separates object detail from background across every object class and every degradation type at once — a single binary mask in the Fourier domain whose parameter γ is never reported, varied, or checked against real image statistics.","fun_headline_variants_meta":{"raw":{"variants":["Frequency split fixes AI's layout mistakes in low-light images","Disentangling frequencies makes generated scenes obey user layouts","FICGen separates image frequencies to beat prior layout-to-image models","New frequency-aware method improves degraded image generation","Frequency-aware generation lifts detector mAP by 4 points in low light"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00044,"raw_usage":{"total_tokens":2094,"prompt_tokens":797,"completion_tokens":1297,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":1215}},"tokens_in":541,"tokens_out":1297,"duration_ms":12936,"temperature":1.0,"reasoning_tokens":1215,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:52:13.368438+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Swap the two prototype channels in a trained model — route the high-frequency instance prototypes into the context branch and the low-frequency context prototype into the instance branch — and measure detector alignment on the same synthetic test sets; if scores do not drop sharply, the claimed instance/context frequency assignment is not what carries the gain. A complementary check: sweep the mask parameter γ of Eq. 3 over a wide range, including per-class masks; if alignment stays flat, the binary spectral split is not the operative mechanism.","supporting_citations":[{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"supplies the frozen latent-diffusion backbone (SDv1.5) that FICGen adapts without retraining"},{"cited_title":"Migc: Multi-instance generation controller for text-to-image synthesis","cited_arxiv_id":null,"evidence_quote":"the multi-instance generation baseline FICGen must beat, and a source of the per-box instance masks"},{"cited_title":"CC-Diff: Enhancing Contextual Coherence in Remote Sensing Image Synthesis","cited_arxiv_id":"2412.08464","evidence_quote":"the contextual-coherence baseline whose foreground–background modeling FICGen extends with frequency cues"},{"cited_title":"Gligen: Open-set grounded text-to-image generation","cited_arxiv_id":null,"evidence_quote":"the grounded text-to-image adapter paradigm whose layout-conditioning design FICGen builds on"},{"cited_title":"Getting to know low- light images with the exclusively dark dataset","cited_arxiv_id":null,"evidence_quote":"the ExDARK low-light benchmark, the flagship degraded domain for alignment and trainability claims"}],"review_version":1}