{"id":"db7bb4fd-c8e1-4bb4-ad75-337f98beb3a8","arxiv_id":"2501.14265","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-stage BNN-DNN framework models one-to-many image enhancement by sampling diverse coarse illuminations and refining them with a deterministic network.","lead":"This paper proposes a two-stage Bayesian neural network framework for image enhancement that can generate multiple plausible enhanced outputs from a single degraded input. It combines a sampling-based uncertainty model in a low-dimensional space with a deterministic refinement network, reporting strong results on low-light and underwater benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core claim is validated only by output variability: no multi-reference benchmark shows BEM samples cover plausible targets, so one-to-many modeling is not actually demonstrated.","rationale":"The reader's weakest assumption (Stage I/II latent mismatch) is a legitimate engineering risk, but it concerns the performance of the two-stage pipeline, not the existence of a one-to-many mapping. Even if the latent distributions matched perfectly, the paper would still only show that BEM outputs vary under weight sampling; it would not show that this variation corresponds to the space of plausible targets. Conversely, the central claim fails if samples are off-manifold or do not cover expert references, regardless of Stage II. The training loss Eq. (6) is a standard single-target regression (with a KL term); no diversity-covering or multi-reference objective is present, and the paired benchmarks have a single GT per input. Thus the load-bearing condition is that the BNN posterior samples are a meaningful proxy for the true target distribution. The paper provides no benchmark with multiple references to test this, only single-reference metrics and no-reference IQA distributions. A multi-reference coverage and diversity calibration test would settle it. Conditional acceptance is appropriate: the method is plausible and the two-stage design is useful, but the central conceptual claim needs this validation before the title-level claim is accepted.","tokens_in":12226,"tokens_out":10347,"duration_ms":104867,"concrete_test":"Run BEM on a multi-reference enhancement benchmark such as MIT-Adobe FiveK (5 expert retouches per input). For each input, generate K=25 BEM samples and compute (i) the fraction of expert references whose LPIPS distance to the input lies within the 5th-95th percentile of BEM sample LPIPS distances (coverage), and (ii) the mean pairwise LPIPS among BEM samples versus among the 5 expert references. If coverage is near chance or sample diversity is an order of magnitude different from expert inter-rater diversity, the one-to-many claim is not supported. If coverage is high and diversity is comparable, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that BEM, by sampling weights from q(w|theta), maps a single degraded input to a distribution of plausible enhanced targets. The experiments establish that repeated sampling produces different outputs, but variability is not the same as modeling a one-to-many target distribution. All paired benchmarks used (LOL-v1/v2, UIEB-R90) provide exactly one reference per input, so the training objective in Eq. (6) is a single-target regression with a stochastic network; there is no term that encourages the sampled outputs to cover multiple plausible references, nor an evaluation that checks coverage. The diversity statistics in Figs. 8-10 and the comparison with VAE/diffusion only show that samples differ from each other; a model that produces off-manifold or implausible variants would pass these diagnostics. Without a dataset with multiple human references per input, the claim that the BNN posterior samples correspond to the true ambiguity in enhancement targets, rather than to variational noise or label noise, is unsupported. This assumption is load-bearing: if it fails, the core novelty reduces to stochastic augmentation, and the reported benchmark gains could come from the DNN refinement and ranking rather than from Bayesian one-to-many modeling.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Bayesian Enhancement Model (BEM), a two-stage BNN-DNN architecture for low-light and underwater image enhancement. Stage I uses a Bayesian neural network to map a downsampled, low-pass-filtered input to a coarse latent illumination map, sampling weights from a variational posterior; Stage II uses a deterministic DNN to refine fine details given the input and the coarse latent. Inference can either rank a set of latent candidates with a no-reference image quality metric (BEMRank) or average the candidates (BEMMC). The authors report full-reference results on LOL-v1/v2 and UIEB-R90, no-reference results on LIME, NPE, MEF, DICM, VV, C60, and U45, and qualitative and quantitative analyses of output diversity.","tokens_in":12511,"tokens_out":4011,"duration_ms":40250,"significance":"The problem of one-to-many mapping in image enhancement is real and underexplored, and the two-stage BNN-DNN design is a sensible attempt to make Bayesian inference practical for high-resolution images. The paper ships code and evaluates on multiple established benchmarks, which is a strength. The adaptive prior and the ranking-based selection are interesting ideas. However, the central claim that BEM models a true one-to-many target distribution is not directly demonstrated, and the reported state-of-the-art results partly rely on oracle candidate selection that is not a deployable inference procedure. If the one-to-many claim can be supported with multi-reference evaluation and the actual (non-oracle) inference variants are shown to be robust, the contribution would be valuable.","major_comments":[{"comment":"The full-reference state-of-the-art claim is based on the BEMRank rows, which the paper itself marks in gray as an upper bound and 'not directly comparable to the other results.' The sentence 'our BEM ... achieves superior performance across all metrics and datasets' therefore conflates an oracle selection procedure with the actual method. The deployable MC variant is competitive but not uniformly dominant; for example, on LOL-v1 BEM-MC has PSNR 28.30 vs. 27.84 for GlobalDiff, a small margin, and on LOL-v2-real it trails GLARE in PSNR (31.41 vs. 28.98? actually BEM-MC is higher, but the margins are not consistent across all metrics). At minimum, the abstract and Section 4.1 should state that the upper-bound rows are not part of the SOTA claim, and the actual inference modes should be compared separately with error bars.","section":"Sec. 4.1, Tables 1-2"},{"comment":"The central one-to-many claim is not supported by the training objective or the evaluation. Eq. (6) is a single-target regression loss (L2 plus KL) over paired data with one reference per input; it contains no term that encourages the sampled outputs to cover multiple plausible reference images. Figures 8-10 demonstrate output variability, but variability alone does not establish that the samples lie on the manifold of plausible enhancement targets or that they cover the true ambiguity. Without a multi-reference dataset or a task-specific coverage evaluation, the paper has not shown that the BNN posterior corresponds to one-to-many target structure rather than variational noise or label noise.","section":"Sec. 3.1, Eq. (6)"},{"comment":"There is a train/inference mismatch for the coarse latent z. During Stage II training, z is replaced by the analytic solution LP((sqrt(x^2+4αy)-x)/(2α)), but at inference z comes from the BNN in Eq. (7), namely Up(F(Down(LP(x),r);w)), or from the upsampled candidate in Algorithm 1. These two quantities differ in how they are filtered, downsampled, and upsampled, and the paper provides no analysis, visualization, or ablation showing that the inference-time z distribution matches the training-time pseudo-ground-truth distribution. If they are mismatched, the Stage II DNN is trained on inputs it will not encounter at test time, which would directly degrade output quality.","section":"Sec. 3.2, Eqs. (7)-(9)"},{"comment":"The adaptive prior is a claimed contribution, but no experiment isolates its effect. The text says it 'has been shown to achieve faster convergence than fixed or empirical priors,' yet no citation or ablation is provided. Since the prior is defined as an EMA of the posterior parameters, it is self-referential and could in principle collapse the KL term; the paper should include an ablation with and without the adaptive prior, and with fixed priors, to support the claim.","section":"Sec. 3.1, 'Adaptive Prior'"},{"comment":"No error bars or multiple-seed results are reported. Because both BEMRank and BEMMC are stochastic, point estimates from a single run are insufficient to establish significant improvements over deterministic baselines, especially where margins are small. Reporting means and standard deviations over at least three runs, or paired significance tests, would materially strengthen the empirical claims.","section":"Tables 1-3"}],"minor_comments":[{"comment":"The description of BEMRank in the full-reference tables is confusing: Algorithm 1 ranks by a no-reference IQA, but Table 1 states that BEMRank selects the candidate closest to the reference under the chosen metric. The paper should clarify which ranking is used in which table and distinguish 'oracle rank' from 'IQA rank' clearly.","section":"Sec. 3.3"},{"comment":"For the no-reference evaluation, BEMRank ranks candidates with NIQE and then reports NIQE on the selected candidate. This is not circular as a method comparison, but it is a form of test-time optimization; the paper should state this explicitly and also report results with a ranking metric different from the evaluation metric.","section":"Sec. 4.1, Table 3"},{"comment":"Equation (8), tilde-y = (x + α z) ⊙ z, is introduced as a modeling choice with only informal motivation. A brief derivation or a toy experiment justifying why this composition is preferable to x+z or x⊙z would improve reproducibility.","section":"Sec. 3.2"},{"comment":"There are several formatting issues in the references and text, including 'V onikakis' (missing letter), duplicated method names in Figure 7's caption, and inconsistent use of 'VAEs' spacing. These should be cleaned up.","section":"References"},{"comment":"The figure does not clearly show the resolution at which z is produced and where the upsampling happens relative to the ranking operation. Aligning the figure with Algorithm 1 would help readers understand the pipeline.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising core idea and a reasonable experimental setup, but the main claim of one-to-many modeling is not yet demonstrated, and the headline results are partly obtained with an oracle selection procedure. I believe these issues are fixable within the scope of a major revision, provided the authors add a multi-reference evaluation or at least a principled diversity-coverage analysis, separate oracle from actual inference, and add ablations for the adaptive prior and the latent-space mismatch."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a real attempt at a real problem. Perceptual enhancement is genuinely one-to-many, and the paper is the first to treat it with a variational BNN rather than a generative model. The two-stage BNN-DNN design is sensible, the speed result is convincing (22x faster than a full BNN), and the code release means the claims can be checked. The combination of a coarse BNN stage, deterministic refinement, and ranking/MC inference is new to this task and not something I've seen before.\n\nThe empirical results are strong when read honestly. But read honestly means ignoring the gray-box rows. BEM Rank is selected by an oracle, so the top-line numbers are upper bounds, not directly comparable to other methods. The paper does say this. The comparable MC variant is competitive on LOL and underwater sets, though not uniformly SOTA; on LOL-v1 it beats LLFlow and GLARE on SSIM/LPIPS but lags GLARE on PSNR. That is fine, but it is not the breakthrough the abstract implies.\n\nSoft spots, in order of importance:\n\n1. The one-to-many claim is not actually tested. All benchmarks have one reference per input. The paper shows output variability (Figs. 8-10) and compares against VAE/diffusion, but variability alone does not demonstrate that the posterior covers plausible targets. A model that injects noise into its latent could produce diverse but off-manifold outputs. Without a multi-reference dataset or a precision/coverage metric, the core claim is supported only by visual plausibility. I think this is the load-bearing weakness, and the stress-test note lands. It is not fatal, but it is unsubstantiated as written.\n\n2. The coarse-latent training/inference mismatch is real and unexamined. At training, Stage II sees the analytic solution to Eq. (8) with y replaced by ground truth; at inference it sees Up(F(Down(LP(x),r);w)). These are different objects. An ablation with Stage II trained on predicted z, or an analysis showing the distributions overlap, should be required.\n\n3. The adaptive prior is a training heuristic with no supporting experiment. There is no convergence curve and no ablation showing it helps. It might be fine, but right now it is an unsupported claim.\n\n4. No error bars anywhere. For a BNN paper, that is especially hard to excuse. A few repeated runs would help.\n\nThe citation pattern looks fine. Self-citations are to relevant prior work from the same group. Related work is broad enough.\n\nWho is this for? People working on ambiguous enhancement tasks and BNN practitioners. It deserves a serious referee. My recommendation: send to peer review, with a request for a multi-reference evaluation (even a small user study) and the Stage II ablation. The MC variant, honestly reported, is a solid conference paper.","headline":"A genuinely new BNN-based take on a real one-to-many problem; the MC baseline is solid but the headline SOTA is oracle-ranked and the central multi-reference claim is untested.","tokens_in":12987,"tokens_out":2190,"would_cite":true,"duration_ms":20581,"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":"A Bayesian network that samples weights from a learned posterior maps one degraded input to many plausible enhanced images, and a two-stage design delivers state-of-the-art low-light and underwater enhancement fast.","keywords":["Bayesian neural networks","one-to-many mapping","low-light image enhancement","underwater image enhancement","variational inference","uncertainty modeling","two-stage enhancement framework"],"falsifier":"Feed the Stage II refinement network the true inference-time latents from Eq. (7) instead of the analytic pseudo-ground-truth latents on the same paired benchmarks, and compare output quality; equivalently, measure a distributional distance such as FID or MMD between the two latent sets. If quality drops markedly or the distributions diverge, the premise that the two-stage pipeline refines what Stage I actually produces is unsupported.","tokens_in":12054,"feed_emoji":"🌙","tokens_out":12155,"duration_ms":94298,"temperature":0.7,"pith_summary":"This paper argues that low-light and underwater enhancement is inherently a one-to-many problem — one degraded input has many plausible enhanced targets — and that deterministic networks, which learn a single mapping, are the wrong tool for it. The proposed Bayesian Enhancement Model (BEM) instead treats network weights as random variables drawn from a learned variational posterior, so each forward pass returns a different plausible output. To keep repeated sampling affordable, BEM splits the work: a Bayesian network samples many coarse illumination maps in a low-dimensional space, a score-based or Monte Carlo step picks one, and a deterministic network refines it into full-resolution detail. The paper reports that this design beats deterministic and probabilistic baselines on the LOL and UIEB benchmarks while running about 22 times faster than a full-resolution Bayesian network and on par with a plain deterministic network.","feed_headline":"One dark image in, many clean images out","feed_subtitle":"Bayesian weight sampling captures all plausible enhancements in one pass — at plain-network speed.","key_machinery":"The load-bearing machinery is the reparameterized Bayesian weight posterior. Every layer of the Stage I backbone becomes Bayesian, with weights sampled as $w = \\mu + \\sigma \\odot \\epsilon$, $\\epsilon \\sim \\mathcal{N}(0,I)$, so $K$ forward passes yield $K$ distinct coarse outputs; an Adaptive Prior, updated by exponential moving average from the posterior parameters, stabilizes training and speeds convergence. The second load-bearing piece is the two-stage decomposition itself: the BNN works on a low-pass-filtered, downsampled input, and the Stage II refinement network is trained not on the BNN's own latents but on the analytic pseudo-ground-truth $z = \\mathrm{LP}\\big((\\sqrt{x^2 + 4\\alpha y} - x)/(2\\alpha)\\big)$, obtained by solving $\\tilde{y} = (x + \\alpha z) \\odot z$ for $z$ with $\\tilde{y}$ set to the true target $y$. This substitution is what prevents the deterministic stage from regressing the BNN's diverse samples back into a single mode.","core_discovery":"The central discovery claim is that weight-space Bayesian sampling is a practical way to model ambiguity in image enhancement, and that the ambiguity can be resolved cheaply by separating coarse diversity from fine detail. Concretely, Stage I passes a low-pass-filtered, downsampled input through a Bayesian backbone and draws $K$ weight samples $w = \\mu + \\sigma \\odot \\epsilon$, producing $K$ coarse latent illumination maps $\\{z_k\\}$; each candidate is scored through the closed-form proxy $\\tilde{y} = (x + \\alpha z) \\odot z$, and the best-scoring (or the averaged) latent $z^*$ is handed to a deterministic Stage II network that writes the high-frequency detail. The paper shows the same recipe works with Transformer and Mamba backbones, that it improves PSNR, SSIM, and LPIPS simultaneously on LOL-v1/v2 and UIEB-R90 against deterministic baselines (KinD, Restormer, RetinexFormer, SNR-Net) and probabilistic ones (LLFlow, GLARE, diffusion, VAE), and that its pixel-wise output variability is far larger than a VAE's or a diffusion model's while remaining low along object boundaries. The authors report the ranking-based variant as an upper bound in full-reference tables, because ranking can in effect select toward the reference image; the Monte Carlo variant is the directly comparable number and is still competitive or better.","pith_inferences":["The paper trains Stage II on the analytic pseudo-ground-truth latent but runs it on the BNN's own upsampled stochastic latents; a natural testable extension is to train Stage II directly on latents drawn from the real inference distribution and measure whether the reported gains persist or improve.","Because the ranking mode is steered by a user-selected no-reference metric (CLIP-IQA, NIQE, UIQM, UCIQE), the pipeline doubles as a tunable aesthetic prior: the same trained model could be pushed toward different notions of quality at inference time without retraining, a use the paper only begins to illustrate.","The same BNN-DNN decomposition should transfer to other one-to-many restoration problems such as deblurring, deraining, or super-resolution with unknown degradations, where the coarse stage would need to encode a different ambiguity structure than illumination; the paper does not test this."],"forward_implications":["One-to-many modeling can replace one-to-one regression as the default framing for low-light and underwater enhancement: the same two-stage recipe improves PSNR, SSIM, and LPIPS at the same time on LOL-v1, LOL-v2, and UIEB-R90.","Coarse-space Bayesian sampling is enough to capture the diversity, so uncertainty modeling does not have to cost full-resolution latency: BEM is reported to run about $22\\times$ faster than a full-resolution BNN and about as fast as a plain DNN on $512\\times512$ images.","The framework is backbone-agnostic: both the Transformer and Mamba instantiations deliver the reported gains, so the design inherits future backbone advances.","The two inference modes are complementary tools: ranking-based inference ($\\text{BEM}_{\\text{Rank}}$) produces sharper, higher-contrast results and best no-reference scores on C60, U45, and five unpaired low-light sets, while Monte Carlo averaging ($\\text{BEM}_{\\text{MC}}$) yields more conservative, noise-suppressed outputs.","The spread of BEM's repeated predictions carries signal, not just noise: score distributions show a low-scoring tail consistent with noisy reference labels, supporting the claim that weight-level uncertainty absorbs label noise instead of committing to one arbitrary target."],"supporting_citations":[{"why":"Supplies the variational-inference objective over diagonal-Gaussian weight posteriors that every Bayesian layer in Stage I is trained with.","marker":"Blundell et al. 2015"},{"why":"Provides the reparameterization trick used to sample weights in each forward pass.","marker":"Kingma 2014"},{"why":"The momentum-update idea adapted into the Adaptive Prior that stabilizes and accelerates Bayesian training.","marker":"He et al. 2020"},{"why":"Supplies the LOL-v1 paired low-light benchmark used for the full-reference LLIE evaluation.","marker":"Wei et al. 2018"},{"why":"Supplies the LOL-v2 benchmark, the second paired low-light test set.","marker":"Yang et al. 2021"},{"why":"Supplies the UIEB-R90 paired underwater benchmark and its reference pairs.","marker":"Li et al. 2019"},{"why":"One of the two backbone architectures (Transformer) instantiated in the two-stage framework.","marker":"Vaswani et al. 2017"},{"why":"Defines the Mamba backbone used in the default BEM model and in the backbone-agnostic claim.","marker":"Gu and Dao 2023"},{"why":"Restormer, the leading deterministic restoration baseline that BEM compares against on PSNR, SSIM, and LPIPS.","marker":"Zamir et al. 2022"},{"why":"LLFlow, the normalizing-flow one-to-many baseline that BEM must beat on the low-light benchmarks.","marker":"Wang et al. 2022"}],"fun_headline_variants":["One input, many valid outputs via Bayesian sampling","Bayesian neural nets model one-to-many image enhancement","Generate diverse enhanced images with Bayesian sampling","From one degraded image, a range of plausible enhancements","Fast diverse image enhancement with Bayesian weight sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the coarse latent $z$ the Stage I network actually produces at inference — an upsampled, stochastically sampled map — lies on the same manifold as the analytic pseudo-ground-truth $z$ used to train Stage II, and the paper reports no analysis or ablation comparing these two distributions.","fun_headline_variants_meta":{"raw":{"variants":["One input, many valid outputs via Bayesian sampling","Bayesian neural nets model one-to-many image enhancement","Generate diverse enhanced images with Bayesian sampling","From one degraded image, a range of plausible enhancements","Fast diverse image enhancement with Bayesian weight sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001005,"raw_usage":{"total_tokens":4250,"prompt_tokens":944,"completion_tokens":3306,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":3235}},"tokens_in":560,"tokens_out":3306,"duration_ms":21830,"temperature":1.0,"reasoning_tokens":3235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:15:07.876356+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed the Stage II refinement network the true inference-time latents from Eq. (7) instead of the analytic pseudo-ground-truth latents on the same paired benchmarks, and compare output quality; equivalently, measure a distributional distance such as FID or MMD between the two latent sets. If quality drops markedly or the distributions diverge, the premise that the two-stage pipeline refines what Stage I actually produces is unsupported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the variational-inference objective over diagonal-Gaussian weight posteriors that every Bayesian layer in Stage I is trained with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the reparameterization trick used to sample weights in each forward pass."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LOL-v1 paired low-light benchmark used for the full-reference LLIE evaluation."},{"cited_title":"W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F","cited_arxiv_id":null,"evidence_quote":"Restormer, the leading deterministic restoration baseline that BEM compares against on PSNR, SSIM, and LPIPS."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LLFlow, the normalizing-flow one-to-many baseline that BEM must beat on the low-light benchmarks."}],"review_version":1}