{"id":"7f268d05-cb01-4b6b-ae02-5e005ac8a24e","arxiv_id":"2412.13815","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GoDiff improves object detection generalization across weather by generating pseudo-target images with a diffusion model and mixing style statistics during training, but the reported gains are modest and depend on knowing the test conditions.","lead":"This paper proposes GoDiff, a method that uses a diffusion model to generate synthetic photos of the same street scenes under different weather and lighting, then trains an object detector on the mix of real and synthetic images. It also mixes style statistics between images during training to make the detector more robust on unseen conditions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CLIP-RBF object filter is internally inconsistent: Eq. 6 defines a similarity kernel, but the text and Eq. 7 treat it as a distance, so the retained boxes may be the least consistent ones, undermining the pseudo-label reliability claim.","rationale":"The reader's weakest assumption correctly targets the CLIP-RBF filter, but the more fundamental problem is that the filter's own equations are mutually inconsistent. The sign error is load-bearing because the entire 'preserved annotations' claim depends on removing low-quality generated objects before their boxes become labels for all pseudo-target domains. If the filter is inverted, the training set could be dominated by the most unreliable pseudo-labels, and any measured gains would have to come from some other component (e.g., CSN or the sheer diversity of generated images), not from the filtering mechanism the paper emphasizes. This is an internal-consistency issue, not a disagreement with community consensus, so it is directly actionable. I would keep the reader's conditional verdict: the paper should not be accepted as is, but the concern can be settled by checking the released code and correcting either Eq. 6 or Eq. 7. I do not see grounds to reject outright because the empirical gains, if reproducible, might survive with the filter sign fixed.","tokens_in":849,"tokens_out":1352,"duration_ms":157723,"concrete_test":"Inspect the released code (anonymous.4open.science/r/GoDiff-FB85) and run the filtering step on a small set of source-to-virtual-source box pairs. For each pair compute D_RBF with a fixed gamma and record whether the box is retained. Verify two things: (1) whether D_RBF is high for visually matched pairs and low for mismatched pairs; (2) whether the code retains boxes with D_RBF > tau or D_RBF <= tau. A correct implementation must retain high-similarity boxes if Eq. 6 is used, or must define D_RBF as 1 - exp(-gamma * ||.||^2) if Eq. 7 is used. If the retained set has lower mean CLIP similarity than the discarded set, the filter is inverted and the pseudo-label reliability claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step in PTDG is the object filter, which is supposed to guarantee that only semantically consistent objects are kept as pseudo-labels. As written, this step is internally inconsistent. Eq. 6 defines D_RBF(fS(bi), fPS(bi)) = exp(-gamma * ||fS(bi) - fPS(bi)||^2), which is an RBF similarity kernel: it is maximal (equal to 1) for identical features and decreases as features diverge. The text then says that a box is filtered out when D_RBF exceeds a threshold tau, and Eq. 7 formalizes the retained set as B_hat = {bi | D_RBF <= tau}. Under Eq. 6, the retained set therefore consists of boxes whose CLIP features are least similar to the source, not most similar. Conversely, if the implementation follows the text's intent and removes low-similarity boxes, then Eq. 7 is incorrect. Since neither tau nor gamma is reported, the direction of the inequality is not a harmless typo: it determines whether the filter removes malformed objects or preferentially retains them. The central claim that PTDG produces reliable, annotation-preserving pseudo-target data rests on this filter, so the method as written does not specify a coherent rule for ensuring pseudo-label quality.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GoDiff, a single-domain generalization method for object detection in urban scenes. It contains two main components: a Pseudo Target Data Generation (PTDG) module that uses an instance-conditioned latent diffusion model (InstanceDiffusion) to generate stylized pseudo-target images from a single annotated source domain, and a Cross-Style instance Normalization (CSN) module that mixes feature statistics across source and pseudo-target domains during training, together with a Covariance Matching Loss. Experiments are reported on the Diverse Weather Dataset (DWD) and Cityscapes-C, with claims that GoDiff improves multiple detectors and, when attached to OA-DG, achieves state-of-the-art performance. The paper also includes ablations, class-wise analyses, and a qualitative appendix.","tokens_in":15817,"tokens_out":6068,"duration_ms":52128,"significance":"If the claims held, the paper would make a useful contribution by showing that modern instance-level controllable diffusion models can generate pseudo-target data for single-domain object detection while preserving annotations, and that feature-level style mixing further helps. The paper ships an anonymous code link, provides ablations in Table 3 that broadly support the contribution of each component, and demonstrates consistent mPC gains when GoDiff is applied to three detectors (Table 4). These are real strengths. However, the central claims contain a technical inconsistency in the object filter (Eqs. 6-7), an unsupported SOTA claim relative to the paper's own Table 1, and a benchmark-awareness concern because the pseudo-target domain descriptors exactly name the test domains. These issues materially affect the validity of the core message as currently written.","major_comments":[{"comment":"The CLIP-RBF filter is internally inconsistent. Eq. (6) defines D_RBF(fS(bi), fPS(bi)) = exp(-gamma * ||fS(bi) - fPS(bi)||^2), which is a similarity kernel: it is maximal at 1 for identical features and decreases as the feature distance grows. The text then says a box is filtered out when D_RBF exceeds a threshold tau, and Eq. (7) retains boxes with D_RBF <= tau. Under the stated definition, Eq. (7) retains the least similar objects, not the most similar ones, exactly reversing the intended semantics. If the implementation actually removes low-similarity boxes, then Eq. (7) is wrong. Because neither tau nor gamma is reported, this is not a harmless typo: the direction of the inequality determines whether the filter removes malformed objects or preferentially keeps them. Since the reliability of the pseudo-labels is the load-bearing claim of PTDG, the method as written does not specify a coherent rule for quality filtering.","section":"Object Filter, Eqs. (6)-(7)"},{"comment":"The paper's claim that GoDiff 'consistently outperforms others across all target domains' and the abstract's claim of state-of-the-art performance are contradicted by Table 1. GoDiff alone achieves an mPC of 29.6, which is lower than OA-DG's 31.8 and equal to SRCD's 29.6; in individual domains OA-DG is substantially better (e.g., Night Sunny 38.0 vs 35.4, Dusk 33.9 vs 32.1, Daytime Foggy 38.3 vs 35.9). The only configuration that sets a new SOTA is OA-DG + GoDiff (32.6 mPC). The claims should be reworded to state that the plug-and-play combination with OA-DG reaches SOTA, and the standalone method should not be described as globally superior.","section":"Comparison Results, Table 1"},{"comment":"The pseudo-target domain descriptors Td exactly name the test target domains: Night-Sunny, Night-Rainy, Daytime-Foggy, and Dusk-Rainy. This means the method uses knowledge of the evaluation domains at training time, which is inconsistent with the stated goal of generalizing to 'unseen domains' from a single source domain. The paper should clarify whether the target descriptors are available during training. If they are, the method is better characterized as target-aware augmentation, not single-domain generalization, and the claim of generalization to unseen domains should be dropped or tested on held-out domains. If they are not, the experimental protocol needs to be described with generic descriptors and validated on domains not named in the descriptors.","section":"Implementation Details / Pseudo-Target Domain Generation"},{"comment":"The object filter compares features from the source image Xs with features from a 'virtual source domain' image XPS (also Daytime-Sunny), and the retained boxes are then used as labels for all pseudo-target domains. This checks only whether the generator is stable under the source domain descriptor; it does not verify that objects survive the stylistic transformation into Night-Sunny, Dusk-Rainy, Night-Rainy, or Daytime-Foggy. The stated conclusion that 'this approach ensures that the generated pseudo-labels are more accurate and consistent' is therefore not supported by the described procedure. The filter should either be applied to each generated pseudo-target domain or an experiment should be added showing that source-to-virtual-source consistency predicts preservation in the actual target-style generations.","section":"Object Filter scope"},{"comment":"The object filter's two key hyperparameters, the threshold tau and the RBF width gamma, are not reported anywhere in the main text or appendix. The number of generated pseudo-target images per source image, and the full CSN insertion scheme (beyond the appendix's 0.1 activation probability and two-layer cap), are also unspecified. This makes the pseudo-target dataset construction and the CSN training procedure impossible to reproduce, which is especially problematic given the Eq. (6)-(7) inconsistency in the same filter.","section":"Reproducibility"}],"minor_comments":[{"comment":"Equation (5) writes Xt as z minus a sum of denoising outputs over all timesteps, which is not the standard reverse-diffusion sampling update. This notation confuses the iterative nature of denoising and should be replaced with the recursive update (e.g., z_{t-1} = ... ) to be technically correct.","section":"Pseudo-Target Domain Generation, Eq. (5)"},{"comment":"The text cites 'InstanceDiffusion (Jia et al. 2023)' but the reference list entry for Jia et al. (2023) is DGInStyle, a different paper. The InstanceDiffusion model is correctly cited as Wang et al. (2024) in the introduction; the in-text citation in the method section should be corrected.","section":"Method, Object Conditioned Image Generation"},{"comment":"The statement that 'omitting any component results in a performance decline' is not strictly true for all domains: the CSN-only row has Night Sunny 29.0 versus the baseline 31.8, and adding CML to PT+CSN slightly decreases Day Sunny from 55.4 to 55.0. The text should acknowledge these domain-specific exceptions or report variance across seeds.","section":"Ablation Study, Table 3"},{"comment":"On Cityscapes-C, GoDiff alone achieves mPC 20.3, which is lower than OA-DG's 21.8 and also lower than FSCE (21.0) and SupCon (20.9). The sentence 'GoDiff is less effective against digital corruption types' should be strengthened to state clearly that the standalone method does not outperform OA-DG on this benchmark.","section":"Comparison Results, Table 2"},{"comment":"The heading 'Conlusion' is a typo for 'Conclusion'. The class-wise numbers in Table 6 contain apparent Unicode artifacts and are very hard to parse; the table should be regenerated with clean formatting.","section":"Conclusion and Table 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has the look of a near-final AAAI submission (copyright line present, anonymous code link). The most serious issues are the Eq. (6)-(7) filter contradiction and the fact that the pseudo-target descriptors are the names of the test domains; both must be addressed before the paper can be considered. The claimed SOTA is also overstated relative to the paper's own table. On the positive side, the ablation design and multi-detector experiments are useful, and the code release is a plus. I would not reject outright because the core idea is promising and the technical issues appear fixable within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nBottom line: GoDiff is a plausible diffusion-based augmentation pipeline for single-domain object detection, with consistent gains across three detectors, but the paper as written has a load-bearing inconsistency in its object filtering equations and overstates its SOTA claim against its own Table 1.\n\nWhat's new: using InstanceDiffusion with dual prompts (image-level + object-level) to synthesize pseudo-target domains while preserving boxes, then filtering objects via a CLIP-RBF consistency check, plus a feature-level cross-style normalization and covariance matching loss. The idea is timely and the ablation in Table 3 supports each component contributing. The plug-and-play experiments on Faster R-CNN, Mask R-CNN, and DINO show consistent gains of roughly 4-4.7 mPC, which is the strongest evidence for the method.\n\nSoft spots: First, the filter is incoherent as written. Eq. 6 defines D_RBF as exp(-γ||Δ||^2), a similarity that is maximal when features match. The text wants to keep boxes with high similarity, but Eq. 7 keeps boxes with D_RBF ≤ τ, i.e., the least similar. Either Eq. 6 should be a distance, or Eq. 7 should use ≥. Since τ and γ are never reported, a reader cannot disambiguate; the pseudo-label reliability claim rests on this filter. Second, the SOTA claim is not supported by the paper's own numbers: in Table 1, OA-DG alone (31.8 mPC) beats GoDiff alone (29.6), and GoDiff's best result is only when stacked on OA-DG (32.6). Same pattern on Cityscapes-C (GoDiff 20.3 vs OA-DG 21.8). The combined gain of +1.0 over OA-DG is small, and no error bars are given. Third, the pseudo-target descriptors ('night', 'rainy', 'foggy', 'dusk') are exactly the test domains, which weakens the 'unseen' story, even if the generation itself does not fit parameters to test data. Finally, hyperparameters are missing (τ, γ, images per source, etc.) and the code link is anonymous, so reproducibility is not yet established.\n\nWho this is for: researchers working on domain generalization for detection or diffusion-based synthetic training data. It deserves a serious referee, but the authors need to fix the filter equation, report τ/γ and variance, temper the SOTA wording, and ideally release the dataset.\n\nRecommendation: send to peer review, but expect major revision. If the filter issue is a typo, the paper could be solid after re-evaluation.","headline":"Plausible diffusion-based augmentation for S-DG detection, but the object filter equations contradict the text and the SOTA claim outruns the paper's own tables.","tokens_in":16355,"tokens_out":3776,"would_cite":false,"duration_ms":32806,"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":"GoDiff re-renders annotated urban images in new weather styles with boxes and labels intact, then swaps style statistics during training, improving object detection in unseen conditions.","keywords":["single-domain generalization","object detection","diffusion models","data augmentation","style normalization","autonomous driving","pseudo-target data generation","urban scene understanding"],"falsifier":"Re-run the DWD experiment with the object filter inverted, keeping boxes the CLIP-RBF check rejects and discarding the boxes it keeps; if mean performance under corruption does not drop, object filtering is not doing the work the paper assigns to it.","tokens_in":15290,"feed_emoji":"🌧️","tokens_out":10746,"duration_ms":86014,"temperature":0.7,"pith_summary":"The paper proposes GoDiff, a training-data pipeline for object detection that requires only one annotated source domain: daytime-sunny urban images. A fine-tuned latent diffusion model re-renders each source image in several target styles (night, fog, rain, dusk) while carrying over the original bounding boxes and labels, and a CLIP-based consistency check filters out objects that were not generated faithfully. During detector training, GoDiff also swaps channel-wise style statistics between real and generated feature maps and aligns their covariances, so the network sees a smoother, broader style distribution. If the method works as reported, it gives consistent gains across Faster R-CNN, Mask R-CNN, and DINO, and it can be added to an existing generalization method to push it further. The practical payoff is that autonomous-driving detectors could be prepared for unseen weather without collecting new labeled data.","feed_headline":"Synthetic styles lift object-detector generalization nearly 4 points","feed_subtitle":"Re-rendering annotated street scenes in fog, rain, and night, then mixing style features, improves detection on unseen conditions.","key_machinery":"The load-bearing machinery is the Pseudo Target Data Generation (PTDG) pipeline built on an InstanceDiffusion model fine-tuned on the source domain. It uses a dual-prompt strategy: an image-level prompt assembled from Tag2Text tags plus target descriptors such as 'foggy' or 'night', and per-box object-level prompts sampled from randomized templates. Generation preserves bounding boxes and labels, and the CLIP-RBF filter, $D_{\\mathrm{RBF}}(f_S(b_i),f_{PS}(b_i))=\\exp(-\\gamma\\|f_S(b_i)-f_{PS}(b_i)\\|^2)$, drops boxes whose CLIP features in a re-rendered 'virtual source' image are too far from the source, with threshold $\\tau$. On the feature side, Cross Style Normalization (CSN) swaps $\\mu$ and $\\sigma$ between feature maps of image pairs, and the Covariance Matching Loss $\\mathcal{L}_{CM}=\\|\\Sigma_{X_A,X_A}^i - \\Sigma_{X_B,X_B}^i\\|_2$ aligns inter-domain covariance statistics, together pushing the detector toward style-invariant representations.","core_discovery":"The central claim is that single-domain generalization in object detection can be achieved by combining image-level and feature-level augmentation. The PTDG module fine-tunes an instance-level latent diffusion model on the source domain, prompts it with an image-level sentence built from extracted tags plus target descriptors and per-box object-level prompts, and generates pseudo-target images whose annotations are inherited from the source. A CLIP-RBF object filter removes boxes whose objects are not semantically consistent between the source image and a re-rendered copy. The CSN module then exchanges the mean and variance of feature maps from different styles during training, together with a covariance matching loss, to make the detector learn style-invariant features. The paper reports that this raises Faster R-CNN mean performance under corruption on DWD from 25.5 to 29.6, improves Mask R-CNN and DINO by about 4 to 5 points, and, when combined with OA-DG, achieves the best reported results on DWD and Cityscapes-C.","pith_inferences":["If the CLIP-RBF consistency check is as transferable as the paper assumes, the same filtering idea could be applied to other box-conditioned generation pipelines for tasks like instance segmentation, where a bad generated mask would otherwise corrupt training.","Because the filter depends on two unstated control settings, $\\tau$ and $\\gamma$, a sensitivity study sweeping those values would show how much of the reported gain comes from filtering rather than from sheer data diversity.","The weaker results on digital corruptions suggest the pipeline is tuned toward natural condition shifts; mixing corruption-specific descriptors into the image-level prompts could be a direct test of whether synthetic noise robustness can be improved the same way.","The randomized object-level templates could be repurposed to rebalance long-tailed detection datasets by sampling rare categories more often, which the paper does not explore."],"forward_implications":["A detector trained on source images plus four generated pseudo-target domains improves mean performance under corruption on DWD from 25.5 to 29.6 for Faster R-CNN, with the largest gains in Dusk-Rainy and Night-Rainy.","The same generated data and style-exchange training can be bolted onto an existing object-aware domain generalization method, raising its DWD mPC from 31.8 to 32.6 and its Cityscapes-C mPC from 21.8 to 22.8.","The benefit transfers across detector architectures: Faster R-CNN, Mask R-CNN, and DINO each gain roughly 4 to 5 mPC points, including the strongest relative gains in adverse conditions like Night-Rainy and Daytime-Foggy.","Each generated pseudo-domain helps most on its matching target condition; for example, generated Daytime-Foggy data raises Daytime-Foggy mAP from 32.0 to 35.0, and training on all generated domains together gives the best overall result.","Fine-tuning the diffusion model on source-domain data lowers the distribution gap between generated and real target data, and the paper reports that lower gap correlates with larger detector gains."],"supporting_citations":[{"why":"supplies InstanceDiffusion, the instance-level diffusion model that accepts per-box prompts and preserves object layout in generated images.","marker":"(Wang et al. 2024)"},{"why":"provides the latent diffusion model (Stable Diffusion 1.5) that is fine-tuned and serves as the backbone of the PTDG generator.","marker":"(Rombach et al. 2022)"},{"why":"supplies Tag2Text, which extracts source-image tags used to build the image-level prompts.","marker":"(Huang et al. 2023)"},{"why":"supplies the CLIP image encoder whose per-box features feed the RBF similarity filter for generated objects.","marker":"(Radford et al. 2021)"},{"why":"provides the CMMD discrepancy measure and RBF kernel idea that inspire both the object filter and the domain-gap analysis.","marker":"(Jayasumana et al. 2024)"},{"why":"defines instance normalization and the channel-wise mean/variance style representation that CSN exchanges between feature maps.","marker":"(Ulyanov, Vedaldi, and Lempitsky 2016)"},{"why":"contributes OA-DG, the object-aware domain generalization baseline that GoDiff augments to set the reported state-of-the-art results.","marker":"(Lee et al. 2024)"},{"why":"supplies the DWD benchmark, the S-DGOD baseline, and the mPC evaluation protocol used for weather-condition generalization.","marker":"(Wu and Deng 2022)"},{"why":"supplies the Cityscapes-C corruption benchmark used to measure robustness against common corruptions.","marker":"(Michaelis et al. 2019b)"}],"fun_headline_variants":["GoDiff: diffusion-generated styles boost detector generalization","Synthetic weather styles sharpen object detection in unseen domains","Pseudo-target diffusion lifts object detector accuracy by 4-5 points","Style-mixing training makes object detectors robust to domain shifts","Plug-and-play GoDiff enhances object detection in novel environments"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole gain rests on the assumption that an object which survives a similarity check between a source image and one re-rendered copy, under filter strictness settings the paper does not report, will be generated faithfully in every other target style, so that its original label and box stay valid.","fun_headline_variants_meta":{"raw":{"variants":["GoDiff: diffusion-generated styles boost detector generalization","Synthetic weather styles sharpen object detection in unseen domains","Pseudo-target diffusion lifts object detector accuracy by 4-5 points","Style-mixing training makes object detectors robust to domain shifts","Plug-and-play GoDiff enhances object detection in novel environments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1654,"prompt_tokens":947,"completion_tokens":707,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":626}},"tokens_in":563,"tokens_out":707,"duration_ms":6444,"temperature":1.0,"reasoning_tokens":626,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:45:28.758729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the DWD experiment with the object filter inverted, keeping boxes the CLIP-RBF check rejects and discarding the boxes it keeps; if mean performance under corruption does not drop, object filtering is not doing the work the paper assigns to it.","supporting_citations":[{"cited_title":"S.; Girdhar, R.; and Misra, I","cited_arxiv_id":null,"evidence_quote":"supplies InstanceDiffusion, the instance-level diffusion model that accepts per-box prompts and preserves object layout in generated images."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"contributes OA-DG, the object-aware domain generalization baseline that GoDiff augments to set the reported state-of-the-art results."}],"review_version":1}