{"id":"e4bc2a41-6c2a-4360-b142-999ad79cdc61","arxiv_id":"2506.23537","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"AFUNet reports top PSNR and SSIM on three standard HDR datasets by alternately aligning and fusing multiexposure features over four unfolded stages.","lead":"AFUNet is a deep learning system for reconstructing high dynamic range images from three differently exposed photos by alternating between alignment and fusion across several network stages. It reports top scores on standard HDR benchmarks, and its main intellectual claim is that this alternating architecture is a principled unfolding of a MAP estimation problem.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SAM's update has a generically non-symmetric Jacobian, so it cannot be the gradient-descent operator in Eq. (4) for any differentiable prior; the MAP-unfolding derivation is therefore not actually implemented.","rationale":"The reader's weakest assumption is the right one and is the most load-bearing: the paper's central theoretical novelty is that AFUNet unfolds a MAP optimization, but the named modules are standard attention networks with no verified link to the derived operators. My Jacobian-symmetry test makes this concern falsifiable without retraining. If the test fails, the derivation in Section 3.2 is not what the network implements, and the paper reduces to an empirical architecture with optimization-inspired naming. That does not invalidate the benchmark numbers, but it changes what the paper has shown. The empirical tables have a secondary but real issue: the margins over prior work are thin, no error bars are given, and Table 1's PSNR-l contradicts the 'consistently surpassing' wording. I am not moving the verdict because the reader already identified the same core weakness and issued a conditional verdict; my read keeps that conditional status and sharpens the required evidence: run the Jacobian check, couple or constrain the learned operators if the theory is retained, and qualify the SOTA claim per-metric with error bars.","tokens_in":14852,"tokens_out":10911,"duration_ms":115244,"concrete_test":"Using the released code, instantiate one SAM at the paper's default configuration and compute, by autograd, J = d f_alpha_i^t / d f_alpha_i^{t-1} on 100 random feature tensors of the default shape. Report rho = max_input ||J - J^T||_F / ||J||_F. If rho is not near machine precision (say below 1e-3), the SAM map cannot equal alpha - zeta * grad p(alpha) for any differentiable p, so Eq. (4) is not implemented. The authors could alternatively exhibit a scalar p_i whose gradient field matches the SAM output, which would also settle the question.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the modules in Section 3.3 implement the updates derived in Section 3.2. This is unverified and, for SAM, likely false. Eq. (4) states alpha_i^t = G_i(alpha_i^{t-1}) = alpha_i^{t-1} - zeta_i * grad_{alpha_i} p_i(D_i x^{t-1}, alpha_i^{t-1}). For any twice-differentiable p_i, the Jacobian of G_i with respect to alpha_i^{t-1} is I - zeta_i * Hessian(p_i), which is symmetric. The SAM update in Eqs. (11)-(12) is a windowed cross-attention block plus FFN; its Jacobian with respect to the input feature is generically non-symmetric, so it cannot equal a gradient step for any scalar potential p_i. The same gap appears in Eq. (8): B^{-1} and D2^T are learned by independent MLPs with no constraint B^{-1} = (D2^T D2 + (beta1 + beta3) I)^{-1}, and the derivation is in image space while DCM operates on features f_u, f_v, f_y2, with no invertible image-feature map supplied. The paper also assigns independent parameters to every unfolding stage, so the stages do not correspond to repeated iterations of one fixed objective. Consequently, the experiments validate an empirical architecture, not the MAP-unfolding claim. The empirical claim also needs qualification: on Kalantari PSNR-l, AFUNet (42.59) equals LFDiff and is below FlexHDR (42.60) and SAFNet (43.18), so 'consistently surpassing state-of-the-art' is too strong as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AFUNet, a deep unfolding network for multi-exposure HDR reconstruction. The authors formulate HDR reconstruction as a MAP estimation problem with two spatial-correspondence priors, decompose it into alternating alignment and fusion subproblems, and derive iterative updates involving gradient descent, proximal operators, and a closed-form least-squares step. These updates are then 'unfolded' into a T-stage network whose modules—SAM, SFM, CFM, and DCM—are claimed to implement the derived operators. The method is evaluated on Kalantari, Hu, and Tel datasets, with qualitative results on Tursun, and ablation studies on component contributions, paradigm ordering, and number of stages. The code is publicly available.","tokens_in":15374,"tokens_out":4512,"duration_ms":42564,"significance":"If the empirical results hold, AFUNet is a competitive HDR deghosting model, and the cross-iterative alignment-fusion design is a plausible design pattern; the ablations in Table 4-6 provide some evidence for the contribution of each module. However, the paper's central theoretical claim—that the network is 'theory-grounded' via MAP unfolding—is not substantiated: the priors p1 and p3 are never defined, and no evidence is given that the learned modules implement the derived mathematical operators. The empirical margins over the strongest baselines are small (often below 0.2 dB), with no error bars or significance tests. The contribution is therefore best viewed as an empirical architecture with a MAP-inspired narrative rather than a validated unfolding algorithm.","major_comments":[{"comment":"The central claim that AFUNet is 'theory-grounded' is not supported by the module design. In Eq. (4), G_i is defined as a gradient descent operator on p_i(D_i x, alpha_i); for any differentiable p_i, the Jacobian of G_i with respect to alpha_i is I - zeta_i * Hessian_{alpha_i}(p_i), which is symmetric. SAM uses window-based cross-attention and an FFN (Eqs. 11-12), whose Jacobian is generically non-symmetric, so SAM cannot implement Eq. (4) for any differentiable prior. Similarly, CFM is described as a proximal operator for Eq. (7), but no property of the learned module (e.g., a minimization objective or fixed-point condition) is verified, and the proximal operator of p_i(D_i u, alpha_i) has a specific dependence on D_i and alpha_i that CFM does not reflect. Moreover, the paper states that all stages have independent parameters, so the T stages do not correspond to repeated iterations of a single objective. The authors should either provide constraints or training losses that enforce the operator identities, or explicitly reposition the MAP derivation as an architectural inspiration rather than an implemented optimization.","section":"Section 3.3, Eqs. (11)-(15)"},{"comment":"The closed-form update in Eq. (8) is derived in image space and requires B^{-1} = (D_2^T D_2 + (beta1 + beta3) I)^{-1}. The DCM learns B^{-1} and D_2^T with independent MLPs (Fig. 3d) and applies them to feature maps f_u, f_y2, and f_v, but no invertible mapping between image space and feature space is supplied. Thus Eq. (14) is not the feature-space version of Eq. (8) unless the MLPs are constrained to satisfy the Woodbury relation linking B^{-1}, D_2, and the beta weights. Please add consistency constraints or an experiment demonstrating that the learned matrices satisfy the required relation on the validation set.","section":"Section 3.3, Eqs. (8) and (14)"},{"comment":"The statement that AFUNet 'consistently surpass[es] state-of-the-art methods' is overstated. On Kalantari PSNR-l, AFUNet (42.59 dB) ties LFDiff and is below FlexHDR (42.60 dB) and SAFNet (43.18 dB); the PSNR-mu margins over the strongest baselines are 0.15 dB (LFDiff) and 0.25 dB (SAFNet), respectively, and no error bars, multiple seeds, or significance tests are provided. Please quantify variability and either soften the claim or add a statistical test.","section":"Abstract, Section 4.2, Table 1"},{"comment":"The spatial correspondence priors p1 and p3 are never defined; the paper does not specify their functional form, differentiability, or how they relate to the training loss in Eq. (17). The variables alpha1 and alpha3 are initialized as features, not as aligned images, and no term in the training loss enforces that alpha_i is a spatially aligned version of y_i. Consequently, the MAP derivation is a scaffold that is not instantiated or verified. Please either define the priors and show the corresponding gradient and proximal operators, or state clearly that p1 and p3 are abstract placeholders and that the contribution is an empirical unfolding architecture.","section":"Section 3.2, Eq. (2)"}],"minor_comments":[{"comment":"There is an unmatched parenthesis in the proximal operator definition: 'prox_{lambda1/beta1 p1(·,alpha_t1))}(x^{t-1})' has an extra closing parenthesis; please fix the notation.","section":"Section 3.2, Eq. (7a)"},{"comment":"The text says 'The features f^{t-1}_{alpha1}, f^{t-1}_x, f^{t-1}_{alpha3} represent the outputs from stage t-2', but Eq. (10) indicates they are the outputs from stage t-1; the indexing should be made consistent.","section":"Section 3.3, after Eq. (10)"},{"comment":"The caption says SAM aligns features with 'f^t_x', but Eq. (11) uses f^{t-1}_x; the notation should be aligned.","section":"Figure 3(a) caption"},{"comment":"Table 3 lists only a subset of the baselines included in Tables 1 and 2 (e.g., ADNet, APNT, HyHDR, LFDiff, RFG-HDR, SAFNet are omitted). Please either include the same baseline set or state explicitly why some methods are not compared on Tel's dataset.","section":"Table 3"},{"comment":"The paper does not report the number of parameters, FLOPs, or runtime for the proposed model or the compared methods; please add these for reproducibility and practical assessment.","section":"Section 4.1, Implementation Details"},{"comment":"The dataset name is spelled 'Tursen' in the main text and figure caption; the correct name is 'Tursun'.","section":"Section 4.2 and Figure 5"},{"comment":"The text says the loss is computed in the tone-mapped domain by applying the mu-law function, but the formula applies tau(x) and tau(hat x) without specifying how the predicted HDR values are clamped or normalized before tone mapping; please clarify the preprocessing.","section":"Section 3.4, Eq. (17)"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution may be publishable if the authors reframe the paper as an empirically designed architecture with a MAP-inspired derivation, and if they add robustness analysis for the thin performance margins. As written, the 'theory-grounded' claim is likely to be challenged by reviewers familiar with deep unfolding; the absence of any constraint or verification linking the learned modules to the derived operators is a substantive issue. I would also recommend the editor ask for a clear statement of how the baselines were trained or evaluated, since several strong baselines are missing from Table 3."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: AFUNet is a serious empirical HDR deghosting paper with strong benchmark results, but the MAP-unfolding derivation doesn't survive contact with the architecture. The stress-test's Jacobian point is correct: Eq. (4) is a gradient step with symmetric Jacobian (I - ς·Hessian), while SAM is cross-attention plus FFN, whose Jacobian is generically non-symmetric. So SAM cannot literally be G_i for any differentiable prior. The same looseness applies to Eq. (8): B^{-1} and D2^T are learned by unconstrained MLPs, and the derivation is in image space while DCM works on features. Independent per-stage parameters compound the mismatch. The theory is a scaffold, not an implemented constraint.\n\nWhat is genuinely new: the alternating alignment/fusion schedule inside an unfolded framework, with the SAM/SFM/CFM/DCM module set. That's a useful design pattern, and the paper documents it clearly. The ablations are thorough, the writing is honest about most implementation details, and the code is public. On Kalantari, Hu, and Tel the numbers are competitive—top or near-top on several metrics—so the empirical contribution stands on its own.\n\nThe soft spots are mostly about claims, not data. 'Consistently surpassing state-of-the-art' is too strong: on Kalantari PSNR-l, AFUNet ties LFDiff and trails SAFNet and FlexHDR. Gains over the strongest baselines are often below 0.2 dB with no error bars or significance tests. Missing comparisons—Mai et al.'s unrolled low-rank method and MERF—should be added since both are cited. The 'unfolding' reading is also weakened by stage-independent parameters; the network does not literally iterate a fixed objective.\n\nProportionately: the empirical architecture is solid; the theoretical framing is overstated. This is fixable. Tone down the MAP claims, add the missing baselines and error bars, and either constrain the operators or explicitly call them learned approximations. As written, I'd not reject it—I'd ask for major revision.\n\nRecommendation: yes, send it out for serious peer review. A competent referee can push the authors to align the story with the implementation. I'd cite the architecture if I worked on HDR, but not the theoretical justification.","headline":"A competitive HDR deghosting network whose MAP-unfolding story is contradicted by its own modules; empirically solid, theoretically overstated.","tokens_in":15805,"tokens_out":4100,"would_cite":true,"duration_ms":41962,"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":"AFUNet reconstructs HDR images from multi-exposure LDR inputs by alternating alignment and fusion at every iterative stage, and reports top results on three standard benchmarks.","keywords":["HDR reconstruction","deep unfolding","MAP estimation","multi-exposure imaging","deghosting","alignment-fusion","transformer","image restoration"],"falsifier":"Train AFUNet in two configurations: the full model and a stripped model in which the SAM, CFM, and DCM are replaced by generic attention or convolution blocks of comparable capacity, and also record the value of the objective in Eq. (5) after each stage. If the stripped model matches the full model's accuracy, or if the unfolded objective does not decrease stage by stage, then the claimed correspondence between the modules and the optimization operators is not doing the work; conversely, if the full model strictly outperforms and the objective decreases, the theory is supported.","tokens_in":14637,"feed_emoji":"🌄","tokens_out":6186,"duration_ms":60445,"temperature":0.7,"pith_summary":"The paper proposes AFUNet, a deep network for high-dynamic-range (HDR) reconstruction from three differently exposed low-dynamic-range images. Its central claim is that HDR reconstruction should be treated as a single optimization problem in which alignment and fusion are solved alternately rather than as separate pre-processing and merging steps. The authors cast the problem as maximum a posteriori (MAP) estimation with spatial correspondence priors, derive an alternating minimization scheme, and unfold that scheme into an end-to-end trainable network. They report that AFUNet consistently surpasses previous methods on the Kalantari, Hu, and Tel benchmarks, with gains attributed to the cross-iterative synergy between alignment and fusion. A sympathetic reader would care because the work offers a mathematical storyline for a task whose recent architectures are mostly empirical.","feed_headline":"AFUNet beats prior HDR imaging on three benchmarks","feed_subtitle":"By alternating alignment and fusion across stages, the network cuts ghosting and recovers detail.","key_machinery":"The central object is the cross-iterative Alignment-Fusion Module (AFM), the network block obtained by unrolling one alternating-minimization step. Within an AFM, two window-based cross-attention Spatial Alignment Modules act as the gradient descent operators G1 and G3 of Eq. (4), aligning the non-reference features with the reference feature; a transformer-based Spatial Fusion Module followed by two channel-attention Channel Fusion Modules acts as the proximal operators of Eq. (7); and a Data Consistency Module implements the closed-form update of Eq. (8) as a learned combination of u, v, and the reference feature, with an MLP and residual connection refining the result. The mathematical roles are assigned by construction: the paper identifies SAM with gradient descent on the alignment priors, the SFM/CFM pair with proximal operators for the fusion priors, and DCM with the quadratic-regularized least-squares update. The network uses four default stages, with each stage sharing the same structure but independent parameters.","core_discovery":"The paper's central claim is that multi-exposure HDR reconstruction is a joint alignment-and-fusion problem that should be optimized in alternation, not as sequential stages. Starting from a MAP objective over the reference exposure and two non-reference exposures, the authors introduce spatial correspondence priors p1 and p3, decompose the objective into alignment subproblems (solved by gradient descent operators G1 and G3) and a fusion subproblem (solved with half-quadratic splitting and proximal operators), and obtain a closed-form data-consistency update x_t = (D2^T D2 + (β1+β3)I)^{-1}(D2^T y2 + β1 u_t + β3 v_t). Unfolding this iteration yields AFUNet, whose T stages each contain a Spatial Alignment Module (window-based cross-attention), a Spatial Fusion Module plus two Channel Fusion Modules (channel-attention transformers), and a Data Consistency Module, with the degradation matrices $B^{{-1}}$ and D2^T learned by MLPs. The paper argues that because alignment sees the fusion state and fusion sees the alignment state at every stage, the two subtasks reinforce each other, producing better deghosting and detail than either the alignment-first or fusion-only paradigm. Empirically, the paper reports top PSNR, SSIM, and HDR-VDP2 numbers on the Kalantari, Hu, and Tel datasets, plus qualitative generalization to the Tursun dataset.","pith_inferences":["The mathematical grounding is asserted but not verified: nothing in the training loss or ablations checks that the module outputs satisfy the fixed-point equations of the derived operators, so a cautious reading treats the MAP derivation as a design metaphor whose empirical payoff is what the tables show.","If the cross-iterative pattern is what matters, the same alternating alignment-fusion unfolding could transfer to other multi-image tasks such as burst denoising, video super-resolution, or multi-exposure fusion outside HDR, where alignment and fusion are also coupled.","A direct testable extension would be to add a consistency loss that penalizes deviation from the optimization equations, for example by enforcing that SAM's output equals a gradient step or that DCM matches Eq. (8) with shared B^{-1} and D2^T; such a constraint would either strengthen the theory-based claim or reveal where the analogy breaks.","Interpreting what the learned B^{-1} and D2^T matrices encode, such as whether they converge to expected degradation structure, could clarify how much of the model's behavior is truly unfolded optimization versus learned feature processing."],"forward_implications":["If the reported numbers hold, AFUNet is the best-performing HDR deghosting network on the Kalantari, Hu, and Tel benchmarks across PSNR, SSIM, and HDR-VDP2.","The alternating alignment-fusion pattern becomes a validated architectural template for multi-exposure reconstruction, since ablations show that adding SAM, CFM, and DCM each improves the result and that alignment-before-fusion beats fusion-before-alignment.","Because the degradation matrices B^{-1} and D2^T are learned rather than fixed, the same architecture can adapt to different camera response and degradation settings without re-deriving the model.","The 3-stage model already surpasses previous state-of-the-art in PSNR-µ and SSIM-µ, suggesting that the unfolding design achieves strong results with fewer iterations.","The deep-unfolding formulation means each stage refines a common objective rather than being a purely feed-forward stack, so the number of stages can be traded against quality and complexity."],"supporting_citations":[{"why":"Supplies the Kalantari dataset, the evaluation split, and the tone-mapped loss convention that the paper follows.","marker":"[13]"},{"why":"Provides the half-quadratic splitting method used to decouple the fusion subproblem into proximal updates.","marker":"[6]"},{"why":"The prior deep-unrolled HDR method based on low-rank tensor completion that AFUNet is compared against and situates itself beyond.","marker":"[23]"},{"why":"Supplies the Hu dataset with saturation-heavy scenes used for evaluation.","marker":"[10]"},{"why":"Supplies the Tel dataset and serves as a representative alignment-free HDR baseline.","marker":"[33]"},{"why":"Provides the window-based self-attention design adapted into the Spatial Alignment Module.","marker":"[19]"},{"why":"Provides the channel-attention transformer architecture used in the Channel Fusion Modules.","marker":"[45]"},{"why":"Supplies the VGG-19 features used in the perceptual loss term of the training objective.","marker":"[30]"},{"why":"Supplies the HDR-VDP2 perceptual metric used to evaluate reconstruction quality.","marker":"[24]"},{"why":"Supplies the SSIM metric used alongside PSNR in all quantitative comparisons.","marker":"[37]"}],"fun_headline_variants":["AFUNet unrolls MAP estimation for HDR deghosting","Alternating alignment and fusion improves HDR reconstruction","Deep unfolding couples alignment and fusion for HDR","Joint alignment-fusion via iterative unfolding in AFUNet"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the network's attention modules really perform the mathematical optimization steps they are named after (gradient descent, proximal updates, closed-form fusion); the paper provides no check that they do, so if that correspondence is only a name, the mathematical foundation is a post-hoc story and the real contribution is the empirical design.","fun_headline_variants_meta":{"raw":{"variants":["AFUNet unrolls MAP estimation for HDR deghosting","Alternating alignment and fusion improves HDR reconstruction","Deep unfolding couples alignment and fusion for HDR","Joint alignment-fusion via iterative unfolding in AFUNet"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1451,"prompt_tokens":1103,"completion_tokens":348,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":719,"completion_tokens_details":{"reasoning_tokens":282}},"tokens_in":719,"tokens_out":348,"duration_ms":4096,"temperature":1.0,"reasoning_tokens":282,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:38:41.381862+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train AFUNet in two configurations: the full model and a stripped model in which the SAM, CFM, and DCM are replaced by generic attention or convolution blocks of comparable capacity, and also record the value of the objective in Eq. (5) after each stage. If the stripped model matches the full model's accuracy, or if the unfolded objective does not decrease stage by stage, then the claimed correspondence between the modules and the optimization operators is not doing the work; conversely, if the full model strictly outperforms and the objective decreases, the theory is supported.","supporting_citations":[{"cited_title":"Deep high dynamic range imaging of dynamic scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the Kalantari dataset, the evaluation split, and the tone-mapped loss convention that the paper follows."},{"cited_title":"Nonlinear image recov- ery with half-quadratic regularization","cited_arxiv_id":null,"evidence_quote":"Provides the half-quadratic splitting method used to decouple the fusion subproblem into proximal updates."},{"cited_title":"Deep unrolled low-rank tensor completion for high dynamic range imaging","cited_arxiv_id":null,"evidence_quote":"The prior deep-unrolled HDR method based on low-rank tensor completion that AFUNet is compared against and situates itself beyond."},{"cited_title":"HDR deghosting: How to deal with saturation? In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 1163–1170, 2013","cited_arxiv_id":null,"evidence_quote":"Supplies the Hu dataset with saturation-heavy scenes used for evaluation."},{"cited_title":"Alignment-free HDR deghosting with semantics consistent transformer","cited_arxiv_id":null,"evidence_quote":"Supplies the Tel dataset and serves as a representative alignment-free HDR baseline."},{"cited_title":"Very deep convolutional net- works for large-scale image recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the VGG-19 features used in the perceptual loss term of the training objective."},{"cited_title":"HDR-VDP-2: A calibrated visual metric for visibility and quality predictions in all luminance conditions","cited_arxiv_id":null,"evidence_quote":"Supplies the HDR-VDP2 perceptual metric used to evaluate reconstruction quality."}],"review_version":1}