{"id":"8017a7e3-c026-4381-8f46-e4173935283b","arxiv_id":"2508.11165","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"EM-B3DM combines EM-based distribution decoupling and a bidirectional Brownian bridge diffusion model for semi-supervised dehazing, claiming state-of-the-art or comparable performance on synthetic and real datasets.","lead":"This paper proposes a new two-stage training method for image dehazing that pairs expectation-maximization with a Brownian bridge diffusion model. It aims to handle thick real-world haze with limited paired data, which matters for autonomous driving and photography.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world transfer of the stage-two unpaired-data training is asserted but not supported by the available abstract; the core empirical claim is unverifiable without ablations and domain-gap analysis.","rationale":"I agree with the reader's weakest_assumption. The abstract alone cannot support the central empirical claim; the key risk is the transfer of the synthetic conditional and the benefit of unpaired data. The reader's UNVERDICTED verdict is appropriate; my concern does not change it. The proposed tests would settle whether the concern lands.","tokens_in":686,"tokens_out":2488,"duration_ms":26995,"concrete_test":"Obtain the full manuscript and perform two checks: (a) locate the real-world benchmark (e.g., Dense-Haze/NH-Haze) and compare the reported metric for the full EM-B3DM against the same model with stage-2 removed; if the stage-2 increment is absent or within noise, the semi-supervised contribution is unsupported. (b) Inspect the stage-2 objective to verify that unpaired clear images enter only through a well-posed prior/regularizer and not through an implicit assumption that the real hazy formation model matches the synthetic one; if the objective requires matching full conditionals without paired data, test on a synthetic-to-real domain shift benchmark (train on RESIDE, test on Dense-Haze) and report the domain gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that EM-B3DM achieves superior or comparable state-of-the-art dehazing on real-world thick haze via a two-stage scheme: (1) EM decouples the joint distribution of paired synthetic data into conditionals modeled by a Brownian bridge diffusion, and (2) large-scale unpaired hazy/clear images further improve performance. The load-bearing premise is that the conditional p(clear|hazy) learned from synthetic pairs transfers to real thick haze, and that unpaired real images provide domain-aligned gradients for this conditional. This premise is not demonstrated in the abstract: there are no real-world versus synthetic domain-gap statistics, no ablation isolating the stage-2 gain, no analysis of whether unpaired clear images act as a prior or introduce distribution shift, and no details on how the EM decomposition avoids degenerate solutions when only marginals are observed in stage 2. Because the full text was not available for review, this is an unverified empirical claim rather than an observed inconsistency, but it is the single point on which the headline result depends.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes EM-B3DM, a semi-supervised image dehazing method. Stage one uses the EM algorithm to factor the joint distribution of paired hazy/clear images into two conditionals, modeled by a unified bidirectional Brownian bridge diffusion model. Stage two fine-tunes the pre-trained model on large-scale unpaired hazy and clear images to improve real-world thick-haze performance. A Residual Difference Convolution (RDC) block is introduced for detail enhancement. The abstract claims state-of-the-art or comparable results on synthetic and real-world datasets. This review is based solely on the abstract, as no full text, equations, tables, or experimental details were available.","tokens_in":932,"tokens_out":1827,"duration_ms":24146,"significance":"If the claimed results hold, the method would be a meaningful contribution: it addresses the scarcity of real paired dehazing data by exploiting unpaired images, which is an important practical problem. The combination of EM-based distribution decoupling with diffusion models is also a plausible and interesting architectural direction. However, because the manuscript as provided contains no quantitative evidence, no comparisons, no dataset descriptions, no ablation studies, and no derivation of the EM update or the Brownian bridge construction, the significance cannot currently be assessed beyond the plausibility of the proposal. The lack of detail prevents a determination of whether the core claim is supported.","major_comments":[{"comment":"The central empirical claim—'superior or at least comparable performance to state-of-the-art methods on both synthetic and real-world datasets'—is presented without any supporting quantitative results. No datasets, evaluation metrics, baseline names, or numerical scores are given. Because the full text was not available for review, this claim is unverifiable. Please provide tables/figures with comparisons and ablations, and specify the datasets and protocols used.","section":"Abstract"},{"comment":"The second stage assumes that unpaired hazy and clear images provide useful gradients for improving the conditional model p(clear|hazy) learned from synthetic paired data. This is a load-bearing premise for real-world generalization, but the abstract offers no evidence or analysis of the domain gap between synthetic and real haze, no ablation isolating the stage-two contribution, and no discussion of how unpaired clear images act as a training signal without introducing distribution shift. Please include a domain-gap analysis or an ablation that quantifies the effect of stage two on real-world benchmarks.","section":"Abstract (two-stage scheme)"},{"comment":"The abstract states that the EM algorithm decouples the joint distribution into two conditionals, but it does not explain how identifiability or degenerate solutions are avoided when stage two observes only marginals. Without specifying the EM formulation (e.g., the latent variables, the M-step updates, and any regularization), it is unclear whether the unpaired data can actually improve the conditional model or merely reinforce a trivial solution. Please provide the EM derivation and discuss its convergence and degeneracy safeguards.","section":"Abstract (EM decomposition)"}],"minor_comments":[{"comment":"The phrase 'Bidirectional Brownian Bridge Diffusion Models' is not defined. It is unclear what 'bidirectional' refers to (e.g., forward/reverse processes, or hazy-to-clear and clear-to-hazy directions). A brief clarification would help.","section":"Abstract"},{"comment":"The 'Residual Difference Convolution block (RDC)' is introduced as a contribution but no explanation of its operation is given. A short description of how it captures 'gradient-level information' would improve readability.","section":"Abstract"},{"comment":"The abstract claims efficiency ('efficient semi-supervised image dehazing method'), but no runtime, parameter count, or computational complexity is reported. Such information would support the efficiency claim.","section":"General"}],"recommendation":"uncertain","confidential_remarks":"The submission as provided is an abstract only, so no substantive technical evaluation is possible. The soundness of the method cannot be assessed without the full manuscript. I suggest the editor require the full text before further review, or treat this as an incomplete submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the specific combination of EM-based decoupling with a bidirectional Brownian bridge diffusion for semi-supervised dehazing is new as far as I know. The motivation is honest: real paired haze data is expensive, and thick haze is exactly where supervised models fall apart. The two-stage idea—train conditionals on synthetic pairs, then adapt with unpaired real images—is worth taking seriously. The RDC block is a small architectural add-on, probably not the main event.\n\nThe real question, and the one the abstract cannot answer, is whether stage-two unpaired training actually closes the synthetic-to-real gap. The stress-test note is right: the abstract asserts the transfer but gives no ablations, no domain-gap statistics, and no analysis of whether the unpaired clear images act as a useful prior or just pull the model off. That's not a fatal flaw in the paper; it's just the load-bearing claim that needs evidence. If the full paper shows a clean ablation isolating stage-two gains and a comparison on real thick haze, the method would be a solid subfield contribution.\n\nWhat I can't do from here is check the soundness of the EM formulation or the Brownian bridge sampling. The abstract gives no equations or details, so all the technical risk is hidden. That said, nothing in the abstract sets off alarm bells; the components are known and the combination is reasonable.\n\nThis paper is for people working on image dehazing and diffusion-based restoration. As a referee assignment, I'd send it to an expert who can verify the experiments and the EM derivation. The empirical claims are extensive but unverifiable at this stage, so a serious referee is exactly what it needs. I would not cite it yet, but I'll keep an eye on the full version.","headline":"A plausible semi-supervised dehazing pipeline combining EM decoupling and Brownian-bridge diffusion, but the abstract alone doesn't support the real-world transfer claim.","tokens_in":1294,"tokens_out":1588,"would_cite":false,"duration_ms":18329,"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":"Semi-supervised diffusion dehazes thick haze without real paired data","keywords":["image dehazing","semi-supervised learning","expectation-maximization","Brownian bridge diffusion","diffusion models","unpaired data","residual difference convolution","image restoration"],"falsifier":"If a held-out set of real thick-haze images is dehazed with the stage-one model alone and shows no perceptual or metric improvement over the synthetic-only baseline, or if adding unpaired real images degrades performance on synthetic benchmarks, the two-stage transfer premise would fail. A concrete check: compare stage-one-only, stage-one-plus-stage-two, and a fully supervised upper bound on a fixed real-world test set with reference-free haze metrics.","tokens_in":664,"feed_emoji":"🌫️","tokens_out":3681,"duration_ms":38297,"temperature":0.7,"pith_summary":"The paper proposes a semi-supervised image dehazing method, EM-B3DM, that avoids the costly collection of real paired hazy and clear images. In the first stage, expectation-maximization decouples the joint distribution of paired synthetic images into two conditional distributions, both captured by a single bidirectional Brownian bridge diffusion model. In the second stage, large-scale unpaired hazy and clear images refine the pretrained model, and a detail-enhanced Residual Difference Convolution block adds gradient-level information. If correct, the method offers a practical way to dehaze thick real-world haze using abundant unpaired data, achieving results on synthetic and real benchmarks that are at least comparable to fully supervised state-of-the-art methods.","feed_headline":"Semi-supervised diffusion dehazes thick haze without paired data","feed_subtitle":"Two-stage EM plus Brownian bridge diffusion uses unpaired real images to dehaze scenes synthetic pairs cannot.","key_machinery":"The load-bearing mechanism is the combination of the EM algorithm—which decouples the joint distribution $p(x,y)$ of hazy and clear images into two conditional distributions $p(y|x)$ and $p(x|y)$—with a unified Bidirectional Brownian Bridge Diffusion Model that learns both mappings within one generative framework; the Residual Difference Convolution (RDC) block additionally injects gradient-level detail into the learned representation.","core_discovery":"EM-B3DM's central claim is that the dehazing problem can be reformulated so that a unified diffusion model learns both conditional directions—hazy-to-clear and clear-to-hazy—by first using EM to factor the joint distribution of paired data into tractable conditional distributions, then fine-tuning on unpaired real images. The paper argues this two-stage scheme transfers the structural relationship learned on synthetic pairs to real-world thick haze, where paired data is absent, and that the RDC block restores gradient-level detail. The reported outcome is superior or comparable performance to state-of-the-art methods on both synthetic and real-world datasets.","pith_inferences":["The EM-plus-Brownian-bridge recipe is not specific to haze; it could be carried over to other dense-to-clean image restoration tasks (deraining, desnowing, low-light enhancement) where paired data is scarce but unpaired data is plentiful.","The bidirectional nature of the diffusion model suggests the clear-to-hazy direction could serve as a data generator, producing new synthetic paired samples from unpaired real clear images—an implicit data-augmentation loop the paper does not explicitly develop.","If Brownian bridge diffusion reduces the sampling path compared to standard denoising diffusion, the method may also lower inference cost, though the paper's abstract does not report sampling speed."],"forward_implications":["Real-world thick-haze dehazing can be improved without collecting paired hazy and clear images, lowering data cost.","Large-scale unpaired hazy and clear images become a usable training signal for diffusion-based restoration.","The same Brownian bridge model can represent both hazy-to-clear and clear-to-hazy conditionals, so the method is a single generative framework rather than two separately trained networks.","Gradient-level detail injected by the RDC block adds representation capacity that helps recover texture in dense haze.","Performance on synthetic and real benchmarks is at least comparable to fully supervised state-of-the-art methods."],"supporting_citations":[],"fun_headline_variants":["EM plus Brownian bridge diffusion dehazes without paired images","Semi-supervised diffusion clears thick haze using unpaired images","Two-stage EM and diffusion dehaze real-world thick haze","Dehaze thick haze with unpaired images via EM diffusion"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The claim depends on the synthetic-pair-learned conditional distribution of clear given hazy also holding for real-world thick haze, with unpaired real images supplying gradients that push the model toward the true real-world distribution rather than away from it.","fun_headline_variants_meta":{"raw":{"variants":["EM plus Brownian bridge diffusion dehazes without paired images","Semi-supervised diffusion clears thick haze using unpaired images","Two-stage EM and diffusion dehaze real-world thick haze","Dehaze thick haze with unpaired images via EM diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000952,"raw_usage":{"total_tokens":3889,"prompt_tokens":729,"completion_tokens":3160,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":3103}},"tokens_in":473,"tokens_out":3160,"duration_ms":23668,"temperature":1.0,"reasoning_tokens":3103,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:04:26.067148+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a held-out set of real thick-haze images is dehazed with the stage-one model alone and shows no perceptual or metric improvement over the synthetic-only baseline, or if adding unpaired real images degrades performance on synthetic benchmarks, the two-stage transfer premise would fail. A concrete check: compare stage-one-only, stage-one-plus-stage-two, and a fully supervised upper bound on a fixed real-world test set with reference-free haze metrics.","supporting_citations":[],"review_version":1}