{"id":"f259de74-d8e5-4874-805f-db1b87c0b5a9","arxiv_id":"2507.18144","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Training a diffusion model on both enhancement and degradation paths, with a shared encoder and a reflection-aware correction module, yields state-of-the-art low-light enhancement on multiple benchmarks.","lead":"Low-light photos are brightened by a diffusion model that trains both the low-to-normal and the normal-to-low mapping at once, sharing an encoder and a learned feature-interaction block. The authors report top scores on standard low-light benchmarks and argue the two-way training makes colors and details more consistent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core bidirectional loss in Eq. (4) is underspecified: \\bar{\\epsilon}_t is never defined, so the central 'implicit symmetry constraint' is not implementable as stated, making the main claimed mechanism unverifiable.","rationale":"The reader's weakest_assumption identifies the same issue, and I agree it is the most load-bearing. The paper's central claim is a novel bidirectional diffusion mechanism; its only formal specification is Eq. (4). If that equation is not well-defined, the mechanism cannot be reproduced or falsified, which undermines the main contribution regardless of the benchmark results. The rest of the paper has compensating strengths: the method is compared against many external baselines, the ablation shows H2L helps, and the cross-task results are suggestive. These would be sufficient for a conditional accept if the loss were clarified. The suspicious numerical coincidence between Table 6's 'w/o L_diff' and Table 5's #2 adds a secondary worry that the loss ablation may not isolate the intended term, but the primary issue remains the undefined \\bar{\\epsilon}_t. Therefore the verdict should remain CONDITIONAL (unchanged), with a request for the expanded loss and code inspection.","tokens_in":15240,"tokens_out":5544,"duration_ms":52602,"concrete_test":"Obtain the official code from the provided GitHub repository (https://github.com/hejh8/BidDiff) and locate the training loss implementation. Verify whether \\bar{\\epsilon}_t is computed from two distinct noise vectors associated with separate forward diffusions of x_h and x_l, or whether both paths share the same x_t (as Fig. 3 suggests). If the implemented loss does not match any well-defined reading of Eq. (4), or if the second term is absent or materially different, then the reported gains cannot be attributed to the claimed bidirectional symmetry constraint, and the paper should state exactly what loss was used.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is the bidirectional diffusion optimization with an 'implicit symmetry constraint' (Section 3.1, abstract). This constraint is implemented only through the second term of Eq. (4), ||\\bar{\\epsilon}_t - \\epsilon_{min}||^2, where \\epsilon_{min} = \\epsilon_\\theta(x_t, x_h, t) - \\epsilon_\\theta(x_t, x_l, t). The text states only that \\bar{\\epsilon}_t 'represents the noise difference \\epsilon used at step t for both paths.' This is not a definition. Figure 3 shows a single noisy latent x_t generated from the normal-light image x_h and then used for both paths, so there is only one ground-truth noise \\epsilon_t. There is no second noise vector whose difference with \\epsilon_t could serve as \\bar{\\epsilon}_t. If the intended meaning is that x_l is also noised separately to produce a different latent, that contradicts the figure and the text. Thus the objective cannot be instantiated, and the claimed symmetry constraint is not a well-defined loss. This is load-bearing because the ablation in Table 5 attributes a 2.071 dB gain to adding the H2L path, and Table 6 attributes a 1.3 dB drop to removing L_diff; if L_diff is not actually the symmetric constraint, these gains may come from the shared encoder, the content loss, or an alternative consistency loss. Additionally, Table 6's 'w/o L_diff' entry (25.648) exactly matches Table 5's #2 configuration (25.648), suggesting the ablation may not cleanly isolate L_diff.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BidDiff, a diffusion-based low-light image enhancement method. During training, a shared-encoder U-Net is run in two directions: low-to-normal (L2H) and normal-to-low (H2L), with separate decoders. An adaptive feature interaction (AFI) block and a reflection-aware correction module (RACM) are added. The core conceptual claim is that minimizing the difference between the noise estimates of the two paths imposes an implicit symmetry constraint on illumination attenuation and noise distribution, improving the L2H enhancement path. The method is evaluated on LOL-v1, LOL-v2-Real, LOL-v2-Syn, UHD-LL, and unpaired datasets, with state-of-the-art or competitive reported PSNR/SSIM/LPIPS/MUSIQ. Ablations are provided for the H2L path, AFI, RACM, and loss terms, plus cross-task results for denoising and inpainting.","tokens_in":15599,"tokens_out":1915,"duration_ms":19912,"significance":"If the proposed bidirectional training mechanism is validated as described, it would be a useful and reasonably novel contribution to diffusion-based low-light enhancement: it is a training-time strategy that does not add inference cost, and the reported gains over strong baselines (e.g., 26.948 PSNR on LOL-v1 vs. 26.316 for DiffLL in Table 1) are meaningful. The paper also reports parameter counts, full comparison tables, and a code repository link, which are strengths. However, the conceptual core of the method—the implicit symmetry constraint implemented by the second term of Eq. (4)—is not specified well enough to be reproduced, and the ablation evidence for that specific term is weakened by a table inconsistency. The paper's central claim therefore needs clarification and re-analysis before its significance can be fully assessed.","major_comments":[{"comment":"The term \\bar{\\epsilon}_t in the second loss term of Eq. (4) is never defined. The sentence '\\bar{\\epsilon}_t represents the noise difference \\epsilon used at step t for both paths' is not a definition, and the text gives no formula or explanation of how this quantity is obtained from the two paths. Since this term is presented as the implementation of the 'implicit symmetry constraint' that is the paper's central novelty, the loss as written is not implementable from the paper, and the claimed mechanism cannot be verified or reproduced. Please provide the exact definition of \\bar{\\epsilon}_t, preferably as an explicit formula in terms of the two noise estimates and the forward-process noise, and clarify how it relates to the shared noisy latent x_t shown in Figure 3.","section":"Eq. (4), Section 3.1"},{"comment":"The ablations do not cleanly isolate the diffusion loss L_diff. The 'w/o L_diff' row in Table 6 reports PSNR 25.648, which is exactly the same as configuration #2 in Table 5 (H2L path without AFI or RACM). If the 'w/o L_diff' model still uses the H2L path, the comparison against the Default row conflates the effect of removing L_diff with the effects of AFI and RACM. Please report the full configuration for each ablation row, and provide an ablation that removes only the second (symmetry) term of Eq. (4) while keeping the first term and the H2L path, so the contribution of the symmetry constraint can be separately assessed.","section":"Table 5 vs. Table 6, Section 4.5"},{"comment":"The cross-task validation on CBSD68 denoising and CelebA-HQ inpainting lacks any training protocol. The model is trained for low-light enhancement, but the text does not state whether the reported denoising/inpainting results use the same LLIE-trained weights without any fine-tuning, or whether the model is retrained or adapted for each task. This is essential for interpreting Table 4: without this information, the 'generalization' claim is not verifiable. Please specify the exact protocol used for each cross-task experiment.","section":"Section 4.4, Cross-Task Validation"},{"comment":"The diffusion objective in Eq. (4) is written as \\|\\epsilon_t - \\epsilon_\\theta(x_t,x_l,t)\\|^2, but the denoising process described in Section 3.1 and Figure 3 starts from a noisy latent x_t generated from the normal-light image x_h. This raises a consistency question: is the ground-truth noise \\epsilon_t associated with the noising of x_h or of x_l, and how is the L2H path's prediction target defined when x_t is not a noised version of x_l? Please clarify this in the text, as it is central to understanding both the forward process and the symmetry constraint.","section":"Eq. (4), noise prediction formulation"}],"minor_comments":[{"comment":"The sentence 'While diffusion-based methods have shown promising performance due to their strong generative capabilities. However, ...' contains a grammatical error: 'While' starts a dependent clause that is followed by a period, and 'However' then begins a new sentence. Please rewrite for clarity.","section":"Abstract"},{"comment":"The caption says 'index H2L AFI RACM' but the table rows show only checkmarks or blank entries for H2L, AFI, and RACM. Please add a legend (e.g., a tick means the component is used) and explicitly indicate that '#4' is the Default configuration, since the text refers to '#4' as the full configuration but the table does not label it as 'Default'.","section":"Section 4.5, Table 5 caption"},{"comment":"In the ablation discussion, the text states that configuration #2 gives gains of '2.071 dB and 0.23' in PSNR and SSIM. The SSIM gain is 0.858 - 0.835 = 0.023, not 0.23. Please correct this typo.","section":"Section 4.5, text"},{"comment":"In the 'Effectiveness of the Loss Function' paragraph, the sentence about removing the content loss says 'PSNR and SSIM improve by 1.571 dB and 0.16, respectively'. Comparing w/o L_content (25.378) with Default (26.948), the PSNR difference is 1.570 dB (not 1.571), and the SSIM difference is 0.869 - 0.853 = 0.016 (not 0.16). Please correct these numbers and clarify whether the sentence refers to degradation or improvement.","section":"Section 4.5, text"},{"comment":"The caption of Figure 3 states that x_t is generated from the normal-light image X_h, but the text of Section 3.1 says 'we first apply forward diffusion to the input real image x_h'. Later, Eq. (4) uses epsilon_theta(x_t, x_l, t) as the L2H prediction. Please make the notation consistent and explain how the L2H path can receive a noisy latent derived from X_h while predicting the noise for X_l.","section":"Figure 3"},{"comment":"Some references appear incomplete or inconsistent: [57] is cited as 'LOL-v2-Real [57]' in the experiments but as Yang et al. 2020 in the reference list, while [58] is the LOL-v2 paper. Please check that all dataset citations point to the correct entries.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical systems paper whose main issue is reproducibility of the core loss term. The undefined \\bar{\\epsilon}_t in Eq. (4) is exactly the kind of load-bearing ambiguity that cannot be left in a published manuscript. The ablation table inconsistency compounds this. I would not reject: the architecture, the RACM module, and the broad experimental results are plausible and potentially useful, and the authors may well have a working implementation (they provide code). But the central claim needs to be made precise, and the cross-task protocol must be specified. If the authors resolve Eq. (4) and the ablation isolation, this could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a plausible incremental extension of diffusion-based low-light enhancement, with a genuinely new training trick—jointly training a normal-to-low path alongside the low-to-normal path, sharing the encoder but keeping separate decoders—plus an AFI feature interaction block and a Retinex-based reflection correction module. The experiments are solid in scope, with many external baselines, standard benchmarks, and ablations that speak for the H2L path. A 2.07 dB gain from adding the reverse path is real evidence the idea has substance. I'd send this to review.\n\nWhere it gets soft: the central loss in Eq. (4) is underdefined. The second term is ||\\bar{\\epsilon}_t - \\epsilon_min||^2, with \\epsilon_min the difference of the two path predictions, but \\bar{\\epsilon}_t is never defined—the text says only that it 'represents the noise difference \\epsilon used at step t for both paths,' which is not a definition. The figure shows a single noisy latent x_t drawn from x_h, so there is only one ground-truth noise \\epsilon. There is no second noise vector whose difference could be \\bar{\\epsilon}_t. If the intended meaning is that both images are noised separately, that contradicts the figure and text. So the claimed implicit symmetry constraint is, as written, not implementable. That matters because the H2L path is the conceptual core and the ablation attributes a large gain to it. This has to be fixed with expanded math or a code-level specification.\n\nAlso, Table 6's 'w/o L_diff' row gives 25.648 dB, exactly matching the H2L-only configuration in Table 5 (#2). If the 'w/o L_diff' model still includes AFI and RACM, the equal PSNR is odd unless those modules contribute nothing at that configuration—or the table has a typo. Either way, the ablation doesn't cleanly isolate L_diff, which is a second load-bearing issue.\n\nSmaller stuff: Section 4.4's cross-task results (inpainting, denoising) give no training protocol—was the model fine-tuned, trained from scratch, or used as-is? There are no error bars, which is common but worth noting when SOTA gaps are under 0.1 dB. Self-citation of [52] is appropriate, as this is a direct extension of that cycle-reconstruction idea.\n\nBottom line: this is a genuine, incremental contribution, and the reported gains are worth checking. It needs a serious referee; the main risk is that the core loss is not verifiable as written. I'd advise the editor to send to review, and ask the authors to provide the full definition of Eq. (4) and clarify the ablation coincidence.","headline":"Bidirectional training idea is plausible and the benchmark gains are real, but Eq. (4)'s undefined noise-difference term is a load-bearing hole that needs fixing before the core claim can be verified.","tokens_in":16102,"tokens_out":3068,"would_cite":false,"duration_ms":30978,"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":"A diffusion model trained with two opposing degradation paths—low-to-normal and normal-to-low—learns where detail and color are lost, and the enhanced images come out cleaner than unidirectional diffusion.","keywords":["low-light image enhancement","diffusion models","bidirectional diffusion","Retinex prior","reflection-aware correction","adaptive feature interaction","image restoration","degradation consistency"],"falsifier":"Inspect the released training code and expand Eq. (4): if $\\bar{\\epsilon}_t$ resolves to something that does not couple the two paths' noise predictions—for example, if it is just the standard noise added to the L2H path—then the bidirectional objective contains no symmetry constraint. A direct experiment is to train the same model with only the first term of $\\mathcal{L}_{\\text{diff}}$; if PSNR on LOL-v1 stays near 26.9 dB, the reported H2L-path gain of 2.071 dB would not be reproducible, and the central claim would be falsified.","tokens_in":15053,"feed_emoji":"🌙","tokens_out":6694,"duration_ms":69941,"temperature":0.7,"pith_summary":"This paper argues that the degradation from a well-lit image to a low-light image and the enhancement back to normal light are two halves of one process, and that a diffusion model should be trained on both directions at once. During training, the model denoises the same noisy latent toward a low-light conditioning image and toward its normal-light counterpart, sharing encoder weights and minimizing the difference between the two paths' predicted noise. The paper claims this implicit symmetry constraint gives more consistent degradation learning, and that the resulting one-directional inference network produces better enhancement than unidirectional diffusion models. It reports top PSNR, SSIM, and LPIPS on LOL-v1, LOL-v2-Syn, and UHD-LL, top SSIM and LPIPS on LOL-v2-Real, and the best MUSIQ scores on DICM, LIME, and MEF. A reflection-aware correction module using Retinex priors adds color fidelity and overexposure suppression.","feed_headline":"Two-way diffusion sharpens low-light image enhancement","feed_subtitle":"Twin diffusion paths share degradation knowledge and beat prior enhancers on four benchmark sets.","key_machinery":"The carrier of the argument is the bidirectional diffusion loss. A standard DDPM forward process creates $x_t$ from the normal-light image; the L2H and H2L paths are conditional denoisers $\\epsilon_\\theta(x_t, x_l, t)$ and $\\epsilon_\\theta(x_t, x_h, t)$ that share the U-Net encoder but have separate decoders. The objective $\\mathcal{L}_{\\text{diff}}$ combines the per-path noise prediction error with a cross-path term $\\|\\bar{\\epsilon}_t - \\epsilon_{\\min}\\|^2$, where $\\epsilon_{\\min}$ is the difference between the two paths' predicted noises; minimizing this term is what is meant to enforce degradation consistency. The adaptive feature interaction block (AFI) passes attention-weighted features between paths with a learnable mixing weight $\\lambda$, and RACM applies channel and spatial attention to a Retinex reflection map.","core_discovery":"The central claim is that degradation is not a one-way street: the normal-to-low-light transition and the low-to-normal-light restoration are inverse views of the same underlying process, so jointly optimizing both diffusion paths makes the model learn where and how detail is lost. The implemented bidirectionality takes a single noisy latent $x_t$ and denoises it under two conditions—low-light image $x_l$ and normal-light image $x_h$—with a shared-encoder U-Net; the training loss in Eq. (4) pushes the L2H path's noise estimate to match the H2L path's estimate, which the paper calls an implicit symmetry constraint on illumination attenuation and noise distribution. At inference only the L2H path runs, followed by the reflection-aware correction module (RACM) that uses a Retinex reflection map to restore color and suppress overexposed areas. On the paired benchmarks the method reports the best PSNR, SSIM, and LPIPS on LOL-v1, LOL-v2-Syn, and UHD-LL, and best SSIM and LPIPS on LOL-v2-Real, with ablations attributing the largest single gain to adding the H2L path.","pith_inferences":["The paper's symmetry constraint is only implicit; an explicit version, such as a cycle-consistency loss between the two paths' outputs, might remove the ambiguity in Eq. (4) and could strengthen the method on unpaired data.","The same bidirectional training recipe could be transferred to reciprocal tasks like haze synthesis and dehazing or rain synthesis and deraining, where paired forward and inverse degradation data are available; the paper does not test this.","Because RACM relies on the Retinex assumption that reflection is illumination-invariant, its overexposure suppression may degrade on images with strong colored light sources or specular highlights, which the current benchmarks only partially cover."],"forward_implications":["At inference the H2L path is discarded, so bidirectional training buys quality without extra sampling cost; only 10 denoising steps are needed.","If the symmetry constraint works as claimed, the same idea should hold for any paired degradation and restoration problem, not only low light.","The shared encoder with independent decoders lets the model learn common degradation features while keeping each direction's specific noise, which supports the reported generalization to denoising and inpainting.","The reported ablation gain of 2.071 dB from adding the H2L path is the direct evidence the paper offers for the bidirectionality claim."],"supporting_citations":[{"why":"Supplies the forward and reverse diffusion formulation and the noise-prediction objective that both the L2H and H2L paths build on.","marker":"[12]"},{"why":"Provides the Retinex decomposition method used to obtain the reflection map prior in RACM.","marker":"[45]"},{"why":"The closest Retinex-guided diffusion baseline whose reported scores the bidirectional loss must beat on LOL datasets.","marker":"[8]"},{"why":"Introduces the UHD-LL ultra-high-definition dataset used for paired evaluation and comparison with diffusion methods.","marker":"[23]"},{"why":"Supplies the LOL-v2 paired low-light benchmarks where the method reports best SSIM and LPIPS.","marker":"[58]"},{"why":"A diffusion-based restoration prior baseline that grounds the comparison against unidirectional diffusion approaches.","marker":"[4]"},{"why":"The diffusion baseline used for cross-task validation on denoising and inpainting.","marker":"[28]"}],"fun_headline_variants":["Bidirectional diffusion uncovers low-light degradation symmetry","Two diffusion directions, one consistent low-light enhancer","Learning degradation both ways improves low-light images","Dual-path diffusion learns what low-light degradation shares"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim stands or falls on the cross-path term in the training loss; the paper never defines one of its two ingredients, so the symmetry constraint may not be implemented as stated.","fun_headline_variants_meta":{"raw":{"variants":["Bidirectional diffusion uncovers low-light degradation symmetry","Two diffusion directions, one consistent low-light enhancer","Learning degradation both ways improves low-light images","Dual-path diffusion learns what low-light degradation shares"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1520,"prompt_tokens":1020,"completion_tokens":500,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":440}},"tokens_in":636,"tokens_out":500,"duration_ms":5977,"temperature":1.0,"reasoning_tokens":440,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:37:44.765621+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released training code and expand Eq. (4): if $\\bar{\\epsilon}_t$ resolves to something that does not couple the two paths' noise predictions—for example, if it is just the standard noise added to the L2H path—then the bidirectional objective contains no symmetry constraint. A direct experiment is to train the same model with only the first term of $\\mathcal{L}_{\\text{diff}}$; if PSNR on LOL-v1 stays near 26.9 dB, the reported H2L-path gain of 2.071 dB would not be reproducible, and the central claim would be falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Retinex decomposition method used to obtain the reflection map prior in RACM."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A diffusion-based restoration prior baseline that grounds the comparison against unidirectional diffusion approaches."}],"review_version":1}