{"id":"e99c8eae-3f9f-420d-b833-7b2480207277","arxiv_id":"2505.01212","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Mono-HDR-3D couples an LDR NeRF/3DGS model with two learned color converters for single-exposure HDR novel view synthesis, but its impressive results depend on HDR ground-truth images during training.","lead":"This paper introduces Mono-HDR-3D, a method for high dynamic range novel view synthesis from single-exposure LDR images, using learned LDR-to-HDR and HDR-to-LDR color converters. The headline claim of unsupervised training is contradicted by the paper's own ablations, which show that HDR ground-truth supervision is essential.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported HDR-NVS gains come from HDR ground-truth supervision, not the claimed LDR-only closed loop; removing L_hdr drops HDR PSNR from 38.57 to 13.50 dB.","rationale":"I read the paper as proposing a new problem: HDR-NVS from single-exposure LDR training images with no HDR ground truth, enabled by the closed-loop H2L-CC module. For this central claim to hold, the H2L-CC loss must be sufficient to train the LDR-to-HDR mapping without HDR targets. The ablations demonstrate the opposite: removing L_hdr collapses HDR PSNR from 38.57 dB to 13.50 dB, and the closed-loop term accounts for only a 0.38 dB gain on top of an already HDR-supervised baseline. Table 6 further shows the paper's best numbers come from a 1/1 LDR/HDR training mixture, not from LDR-only data. The architectural modules are physically motivated and the LDR/NVS components may be useful, but the paper's stated contribution is not what is evaluated. This is a load-bearing internal inconsistency rather than a matter of taste or external consensus. The reader's rejection is therefore well-founded, and I recommend keeping the REJECT verdict as the paper stands; a revised manuscript that honestly reframes the contribution as single-exposure HDR-NVS with sparse HDR supervision could merit reconsideration.","tokens_in":12400,"tokens_out":4965,"duration_ms":48683,"concrete_test":"Run the released Mono-HDR-GS training code with alpha=0 and with no HDR images in the training set, keeping the same single-exposure LDR training views at one exposure time, then evaluate HDR PSNR on held-out views. Per Table 5 row 5 and Table 6 ratio 1/0, this should reproduce roughly 13.50 dB. If it instead recovers close to 38.57 dB, the closed-loop supervision is sufficient and the concern is resolved; if it remains near 13.50 dB, the paper's core claim of LDR-only single-exposure HDR-NVS fails.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that Mono-HDR-3D performs HDR-NVS when only single-exposure LDR images are available, with the H2L-CC closed loop supplying supervision in the absence of HDR ground truth. The paper's own experiments contradict this. The default Mono-HDR-GS result of 38.57 dB in Table 1 corresponds to the 1/1 LDR/HDR training mixture in Table 6, not to an LDR-only setting. Equation (9) always includes L_hdr, and Table 5 row 5 shows that removing L_hdr and training with only Lldr + Lh2l produces 13.50 dB HDR PSNR; L_h2l alone gives 11.87 dB. Thus the closed-loop term contributes only the 0.38 dB improvement shown in Table 4 and cannot substitute for HDR targets. The paper's own Section 4.3 states that L_hdr is the 'foundational component' and calls HDR data 'irreplaceable,' which is consistent with this finding. Consequently, the headline 38.57 dB result is not a demonstration of the claimed LDR-only single-exposure setting; it is an HDR-supervised result, and the abstract's statement that 'only single exposure LDR images are available during training' is unsupported by the reported experiments. This is an internal inconsistency between the stated problem and the evaluated setup, not a disagreement with external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the single-exposure HDR-NVS problem, in which only single-exposure LDR images are assumed available for training, and proposes Mono-HDR-3D, a meta-algorithm built on an LDR-to-HDR color converter (L2H-CC) and an HDR-to-LDR closed-loop converter (H2L-CC). The method is instantiated with NeRF and 3DGS backbones, and experiments report large improvements over HDR-NeRF and HDR-GS on synthetic and real datasets. However, the experimental protocol contradicts the core problem statement: the main results use HDR ground-truth supervision, and the paper's own ablations show that removing the HDR loss collapses HDR PSNR from 38.57 dB to 13.50 dB.","tokens_in":12698,"tokens_out":2935,"duration_ms":28590,"significance":"If the single-exposure, LDR-only claim were supported, the work would be significant: it would remove the multi-exposure capture burden of existing HDR-NVS methods and could serve as a plug-in for NeRF and 3DGS. The paper is clearly written, the camera-imaging motivation is sensible, and the ablation study is unusually thorough, with multiple loss settings and LDR/HDR data ratios reported. What is missing, though, is a demonstration that the method actually works in the setting it advertises. The headline numbers are obtained with HDR ground-truth supervision, and the pure-LDR ablation (Table 6, 1/0) reaches only 13.50 dB PSNR, which is not a usable HDR reconstruction. Because the central claim of the paper is directly contradicted by its own experiments, the significance of the contribution as presented cannot be sustained without major changes to both the method and the evaluation.","major_comments":[{"comment":"The abstract and Section 1 state that 'only single exposure LDR images are available during training' and that the H2L-CC closed loop enables learning 'in the case of no access to HDR ground-truth training data.' However, the total loss in Eq. (9) always includes alpha * L_hdr, and Table 6 shows that the headline result of 38.57 dB HDR PSNR corresponds to the LDR/HDR ratio 1/1, not to an LDR-only setting. The only pure-LDR row in Table 6 (ratio 1/0) yields 13.50 dB PSNR, so the paper's central claim of operating without HDR ground truth is not supported by any reported experiment.","section":"Abstract, Section 1, Eq. (9), Table 6"},{"comment":"The loss ablation directly contradicts the 'unsupervised closed loop' narrative. Removing L_hdr while keeping L_ldr and L_h2l (Table 5, row 5) drops HDR PSNR from 38.57 to 13.50 dB, whereas adding L_h2l to L_ldr + L_hdr (row 4 vs. row 7) only improves PSNR by 0.38 dB. Thus the closed-loop term L_h2l is not a substitute for HDR supervision; it is a minor refinement on top of an HDR-supervised solution. The paper's own observations in Section 4.3, calling L_hdr 'the foundational component' and stating that HDR data is 'irreplaceable,' further confirm this conclusion.","section":"Section 4.3, Table 5, rows 4-7"},{"comment":"The problem in Eq. (1) asks for a model that maps single-exposure LDR images to HDR novel views, but the evaluation protocol always provides HDR ground truth for the training views: Tables 1, 3, 4, and 5 report results on the synthetic dataset where HDR images are available, and Table 6 artificially mixes LDR and HDR supervision. No experiment evaluates the setting in which only single-exposure LDR images are accessible during training and HDR images are held out entirely; the 1/0 row of Table 6 is the closest, and its 13.50 dB PSNR shows that the method does not solve the stated problem.","section":"Section 2 (problem formulation), Section 4.1"}],"minor_comments":[{"comment":"The notation 'LDR / HDR ratio' is ambiguous: it is not clear whether the ratio refers to the number of images, the number of scenes, or the weighting of the respective loss terms. This should be defined explicitly.","section":"Table 6"},{"comment":"Equation (9) says L_hdr is used 'if available,' but the implementation details always set alpha = 0.6, and no experiment with alpha = 0 is reported as the main configuration. The paper should either clarify that the default setting includes HDR supervision or present a fully LDR-only variant as the primary result.","section":"Section 3.3"},{"comment":"There is a stray line 'Mono-HDR-GS-Syn-HDR' appearing directly after Table 1 in the submitted text; it appears to be a leftover fragment or a broken table/figure element and should be removed or properly integrated.","section":"Table 1 and Section 4.1"},{"comment":"The term 'Unsaturation' in Eq. (3) should be written as 'unsaturation' or 'unsaturated pixels' for consistency with the surrounding text, and the definition of I_ideal versus I_overflow could be stated more carefully to avoid confusion about which quantity is physically observed.","section":"Section 3.2, Eq. (3)"},{"comment":"Row 1 of Table 5 reports no numbers for L_ldr alone; since the caption says 'Results are averaged across exposures and scenes,' it would be helpful to state explicitly that training with only L_ldr failed to produce valid HDR outputs, rather than leaving the entry as dashes.","section":"Section 4.3, Table 5 row 1"},{"comment":"The claim that Mono-HDR-GS at 1/0 ratio 'generates marginally better results than HDR-GS trained solely on HDR images' is misleading: the 13.50 dB result is low in absolute terms, and the comparison does not speak to the quality of HDR reconstruction in the intended single-exposure setting.","section":"Section 4.3, observation (IV)"}],"recommendation":"reject","confidential_remarks":"The paper is internally inconsistent in a way that goes to the core of its contribution: the advertised LDR-only, single-exposure setting is never evaluated, and the paper's own ablations show that the proposed closed-loop loss cannot compensate for the missing HDR supervision. The problem formulation is interesting, but the experimental design and the central claim would need to be substantially reworked before a fair assessment is possible. I do not see a path within the current manuscript to support the claim as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the problem formulation is new and the closed-loop idea is worth discussing, but the paper does not demonstrate what its abstract promises. The default 38.57 dB result is trained with HDR ground truth (1/1 LDR/HDR ratio in Table 6), and the closed loop alone—without HDR labels—gives 13.50 dB (Table 5, row 5; Table 6, 1/0). That is a load-bearing inconsistency, not a minor gap.\n\nWhat's genuinely new: the single-exposure HDR-NVS task, and the L2H-CC/H2L-CC modules designed from the physical camera model. The idea of using a differentiable HDR-to-LDR converter to close the loop is clever and could be useful auxiliary supervision when HDR labels are scarce. The meta-algorithm integration with NeRF and 3DGS is clean, and the paper reports reasonable speed and quality comparisons. The authors also honestly report the ablations that sink their main claim—that deserves credit.\n\nThe soft spots are the usual suspects but they matter. The word 'unsupervised' is not supported anywhere; the only unsupervised variant (L_h2l only) gives 11.87 dB. The claim that 'only single-exposure LDR images are available during training' is contradicted by their own loss in Eq. (9) and Table 6. On real data there is no HDR ground truth, and they only report LDR metrics there, so there is zero evidence for single-exposure HDR synthesis in the real world. The 0.38 dB gain from the closed loop is real but modest, and it's measured in a setting that already has HDR supervision. Also, the comparison with HDR-GS when both use 1/1 LDR/HDR is of limited interest if the actual contribution is just a better network for L2H mapping.\n\nWho is this for? Someone working on HDR imaging or novel view synthesis might learn from the architecture and the ablation methodology, but they should not take the abstract at face value. As a claim of a new capability, it fails. As a potential incremental improvement to HDR-GS with HDR supervision, it might have value, but the paper doesn't sell it that way.\n\nRecommendation: I would not desk-reject this because the problem framing is genuinely novel and the experimental section is thorough enough to expose its own flaw. Send it to a serious referee with a clear request to verify the training protocols and to demand a re-framing. If the authors can reposition the paper as 'HDR-supervised NVS with a closed-loop auxiliary loss that slightly improves quality' and remove the unsupervised language, there may be a publishable small contribution. As is, the central claim fails.","headline":"The paper frames a genuinely new problem, but its own ablations show the headline 'single-exposure, no HDR supervision' claim is false: remove the HDR loss and PSNR collapses from 38.57 to 13.50 dB.","tokens_in":13274,"tokens_out":3093,"would_cite":false,"duration_ms":29021,"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":"The paper introduces Mono-HDR-3D, a meta-algorithm that synthesizes high-dynamic-range (HDR) novel views from single-exposure low-dynamic-range (LDR) training images by learning an LDR-to-HDR color converter plus a closed-loop HDR-to-LDR…","keywords":["single-exposure HDR-NVS","HDR novel view synthesis","low dynamic range","closed-loop supervision","camera response model","3D Gaussian splatting","neural radiance fields","LDR-to-HDR conversion"],"falsifier":"Train Mono-HDR-GS on the synthetic scenes using only single-exposure LDR images with no HDR ground truth, then measure HDR PSNR against held-out HDR images; if it stays near 13.5 dB instead of near 38.6 dB, the claim that the closed loop alone enables single-exposure HDR-NVS is refuted.","tokens_in":12203,"feed_emoji":"📷","tokens_out":6728,"duration_ms":60814,"temperature":0.7,"pith_summary":"High-dynamic-range novel view synthesis normally requires multiple exposures per viewpoint to capture the full brightness range. This paper claims that a single exposure per view is enough, by introducing Mono-HDR-3D, a meta-algorithm that first builds an ordinary LDR 3D scene model and then lifts it to HDR with a camera-imaging-aware LDR-to-HDR color converter, while a second HDR-to-LDR converter closes the loop for supervision. The authors report that Mono-HDR-GS, the 3D Gaussian splatting instantiation, reaches 38.57 dB HDR PSNR on synthetic scenes, 3.27 dB above HDR-GS, and that the same machinery works with NeRF. The caveat visible in the paper's own ablations is that the closed-loop loss alone is not sufficient: removing the HDR ground-truth term drops HDR PSNR from 38.57 to 13.50 dB, so the method's practical strength combines LDR geometry with HDR supervision when it is available.","feed_headline":"Closed-loop LDR-to-HDR converter beats multi-exposure baselines","feed_subtitle":"New Mono-HDR-3D lifts single-exposure LDR scenes to HDR radiance, topping HDR-GS by 3.27 dB on synthetic views.","key_machinery":"The load-bearing object is the pair of color converters L2H-CC and H2L-CC, both designed as small MLPs whose layer structure and activations mirror the physical LDR formation equation $I_l = (\\Delta t/g)\\cdot I_h + I_0 + \\epsilon - I_{\\mathrm{overflow}}$ for unsaturated pixels and saturation at $I_{\\max}$ for clipped ones. L2H-CC approximates the inverse mapping (scaling by $g/\\Delta t$, offset correction, noise term) with ReLU-constrained nonnegative outputs, and H2L-CC approximates the forward scaling and offset with ReLU and Tanh activations. The closed loop formed by rendering an HDR image, converting it to LDR, and comparing with the LDR training image is what the paper claims enables optimization without HDR ground truth; the losses are an LDR reconstruction loss, an optional HDR loss in the $\\mu$-law domain, and an H2L consistency loss.","core_discovery":"Mono-HDR-3D's central claim is that the LDR image formation process can be inverted and learned as two color converters, making single-exposure HDR-NVS tractable. Starting from single-exposure LDR training images, the framework learns an LDR 3D model in the usual way, then applies L2H-CC, a per-color MLP whose structure mirrors the inverse camera formula $I_h = (g/\\Delta t)(I_l - I_0 + I_{\\mathrm{overflow}}) - (g/\\Delta t)\\epsilon$, to obtain HDR radiance at each rendered point. The H2L-CC module re-encodes the rendered HDR image back to LDR through the forward formation model, so the HDR branch can be optimized against LDR training views even without HDR ground truth. The authors position this as a meta-algorithm that can wrap any differentiable 3D representation, and they instantiate it with both NeRF and 3DGS. On the synthetic benchmark, Mono-HDR-GS reports the best HDR results, with the closed loop contributing a modest additional gain when HDR ground truth is already present.","pith_inferences":["The paper's 'unsupervised' framing is stronger than its evidence: at LDR/HDR ratio 1/0 (no HDR images), Mono-HDR-GS reaches only 13.50 dB, close to HDR-GS's 10.51 dB, so the closed loop should be described as a regularizer rather than a standalone supervision source.","The same camera-formation-inspired converter pair could be transferred to single-image HDR reconstruction or to inverse rendering tasks where a forward physical model is known, using the loop as a cycle-consistency loss.","A natural next test is exposure generalization: train on one exposure time and evaluate HDR rendering at held-out exposure times, since the current experiments train and test at the same exposure time.","For real deployments, the assumption that a single global color mapping per scene suffices will likely fail in scenes with mixed lighting or severe saturation; a spatially varying or confidence-weighted version of L2H-CC is a testable extension."],"forward_implications":["If valid, the single-exposure setting removes the multi-exposure capture pipeline, avoiding ghosting, alignment, and storage costs in HDR-NVS.","Because Mono-HDR-3D is a meta-algorithm, any future differentiable 3D scene representation can be given HDR capability by adding the two converters.","The reported 38.57 dB HDR PSNR and 0.975 SSIM on synthetic scenes would make single-exposure HDR-NVS competitive with multi-exposure HDR-GS while using less capture effort.","The closed-loop term's +0.38 dB gain when HDR ground truth is available suggests the LDR-to-HDR consistency can serve as a regularizer in other HDR inverse problems."],"supporting_citations":[{"why":"Supplies the LDR image formation equation (exposure, gain, offset, noise, saturation) that L2H-CC and H2L-CC are designed to invert.","marker":"Hasinoff et al. (2010)"},{"why":"HDR-NeRF, the first HDR-NVS baseline and the NeRF-based competitor the single-exposure setting is compared against.","marker":"Huang et al. (2022)"},{"why":"HDR-GS, the Gaussian-splatting HDR-NVS baseline whose evaluation protocol, $\\mu$-law L2 loss, and reported numbers are used for comparison.","marker":"Cai et al. (2024)"},{"why":"3D Gaussian Splatting, the 3D representation that Mono-HDR-GS integrates with.","marker":"Kerbl et al. (2023)"},{"why":"NeRF, the implicit neural representation that Mono-HDR-NeRF integrates with and the backbone of HDR-NeRF.","marker":"Mildenhall et al. (2021)"},{"why":"Provides the $\\mu$-law compression used in the HDR L2 loss of Eq. (11).","marker":"Kalantari et al. (2017)"}],"fun_headline_variants":["Single-exposure HDR views without multi-shot capture","Mono-HDR-3D: single-exposure HDR-NVS beats multi-exposure","Invert the camera: single exposure to HDR radiance","One exposure, full HDR radiance","Single-shot HDR novel views beat multi-exposure"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the closed-loop HDR-to-LDR reconstruction loss can supervise HDR estimation without any HDR ground-truth images; the paper's own ablations instead show that removing the HDR-supervision term drops HDR PSNR from 38.57 dB to 13.50 dB, so the success depends on HDR supervision when it exists.","fun_headline_variants_meta":{"raw":{"variants":["Single-exposure HDR views without multi-shot capture","Mono-HDR-3D: single-exposure HDR-NVS beats multi-exposure","Invert the camera: single exposure to HDR radiance","One exposure, full HDR radiance","Single-shot HDR novel views beat multi-exposure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00076,"raw_usage":{"total_tokens":3407,"prompt_tokens":1009,"completion_tokens":2398,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":2313}},"tokens_in":625,"tokens_out":2398,"duration_ms":17601,"temperature":1.0,"reasoning_tokens":2313,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:23:09.956596+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Mono-HDR-GS on the synthetic scenes using only single-exposure LDR images with no HDR ground truth, then measure HDR PSNR against held-out HDR images; if it stays near 13.5 dB instead of near 38.6 dB, the claim that the closed loop alone enables single-exposure HDR-NVS is refuted.","supporting_citations":[{"cited_title":"W., Durand, F., and Freeman, W","cited_arxiv_id":null,"evidence_quote":"Supplies the LDR image formation equation (exposure, gain, offset, noise, saturation) that L2H-CC and H2L-CC are designed to invert."},{"cited_title":"Hdr-nerf: High dynamic range neural radiance fields","cited_arxiv_id":null,"evidence_quote":"HDR-NeRF, the first HDR-NVS baseline and the NeRF-based competitor the single-exposure setting is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"HDR-GS, the Gaussian-splatting HDR-NVS baseline whose evaluation protocol, $\\mu$-law L2 loss, and reported numbers are used for comparison."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"3D Gaussian Splatting, the 3D representation that Mono-HDR-GS integrates with."},{"cited_title":"K., Ramamoorthi, R., et al","cited_arxiv_id":null,"evidence_quote":"Provides the $\\mu$-law compression used in the HDR L2 loss of Eq. (11)."}],"review_version":1}