{"id":"4bcf1c43-9857-400b-a6f2-50178f402c0c","arxiv_id":"2412.12942","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"On simulated single-photon camera data, Pix2PixHD colorization followed by DeepHDR reconstruction gives the best color HDR output, with a single-stage Pix2PixHD variant close behind.","lead":"This paper evaluates combinations of existing image translation and single-image HDR networks for converting simulated monochromatic single-photon camera images into color high dynamic range images. The best combination, Pix2PixHD plus DeepHDR, outperforms other pairings, and a single-stage variant produces comparable results.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Oracle exposure-time selection and 8-bit quantization make the simulated SPC benchmark idealized; the claimed ranking may not transfer to real sensors.","rationale":"The reader's weakest assumption correctly identifies the simulator's oracle exposure-time selection and 8-bit quantization as the central load-bearing risk. My analysis agrees: these choices make the benchmark an idealized monochrome-LDR-to-color-HDR task rather than a genuine SPC-to-HDR task. The central claim is conditional on this assumption, so the CONDITIONAL verdict is appropriate. I found no additional independent fault that would change the verdict; the paper's internal tables support the ranking only under the stated simulator protocol. The strongest counterfactual test—fixed exposure time and higher-bit-depth inputs—directly targets whether the ranking survives more realistic conditions. If the ranking flips, the claim should be weakened to a benchmark on simulated, oracle-tuned, tone-mapped data; if it survives, the pipeline is more credible. Therefore, UNCHANGED is appropriate, with the condition being the realistic-simulator validation described above.","tokens_in":10538,"tokens_out":2260,"duration_ms":23320,"concrete_test":"Fix the exposure time to a single value determined from the training set only (e.g., the median of training-image exposure times), re-run the full two-stage pipeline on the test images without per-image oracle tuning, and report the HDR-VDP3 ranking. If Pix2PixHD + DeepHDR no longer has the highest score, or the margin over Pix2Pix + DeepHDR shrinks below the metric's reported resolution, the claimed superiority is an artifact of oracle exposure selection. A supplementary check: retrain/evaluate the colorization stage on 16-bit or 32-bit photon-count SPC images instead of 8-bit tone-mapped ones; if the ranking shifts, the pipeline is not actually exploiting SPC dynamic range.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim—that Pix2PixHD + DeepHDR is the best color HDR reconstruction pipeline for SPC images—rests on a simulator protocol that is materially more favorable than any real SPC capture. In Section 3.2, the exposure time T is selected per image based on the ground-truth luminance, i.e., the simulator is given access to the test-time answer. This oracle tuning ensures all scene regions are visible in the simulated SPC frames, removing the very exposure-robustness problem that motivates SPC imaging. A real SPC camera would use a fixed or user-chosen exposure, not one hand-picked from the HDR ground truth to maximize visibility. Additionally, the monochromatic SPC images are saved as 8-bit tone-mapped images (Section 3.2, Section 4), destroying the high-bit-depth photon-count information that constitutes the SPC's extended dynamic range. The pipeline is therefore evaluated as a monochrome-LDR-to-color-HDR task, not as an SPC-to-HDR task. These two choices jointly undermine the external validity of the reported HDR-VDP3 ranking: the ranking may reflect the methods' ability to exploit oracle-exposure, tone-mapped inputs rather than their performance on genuine SPC data. The paper itself concedes in Section 3.3 that current SPAD prototypes are not ready to compete with CMOS, yet the abstract and conclusion imply readiness without real-sensor validation. No real SPC data, no code, and no train/test split are supplied, so the reader cannot independently verify the comparison is leakage-free or that the advantage of Pix2PixHD + DeepHDR is stable across plausible simulator variations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript evaluates computational pipelines for converting simulated monochromatic Single Photon Camera (SPC) images into color high dynamic range (HDR) images. It proposes a two-stage design that first colorizes the SPC image with an image-to-image translation network (Pix2PixHD, Pix2Pix, or CycleGAN) and then reconstructs HDR with a single-image HDR method (DeepHDR, SIHDR, ExpandNet, or LHDR), plus a single-stage variant based on a modified Pix2PixHD. Experiments on the Laval HDR dataset report that Pix2PixHD followed by DeepHDR yields the highest HDR-VDP3 scores, that frame averaging improves results, and that the single-stage method is comparable to the two-stage approach.","tokens_in":10845,"tokens_out":6026,"duration_ms":56555,"significance":"The paper is a systematic comparison of established components for a plausible future SPC processing pipeline. Its strengths include testing many combinations, two resolutions, single-frame versus averaged inputs, and both two-stage and single-stage routes, with HDR-VDP3 as an appropriate metric for HDR reconstruction quality. The authors also include a candid caveat in Section 3.3 that current SPAD prototypes are not yet competitive with CMOS sensors. However, the external validity of the central ranking rests on an idealized simulation protocol: per-image oracle exposure selection and 8-bit tone-mapping of SPC outputs. Because these choices alter the task being evaluated, the paper's claim to validate an SPC-to-color-HDR pipeline is stronger than the evidence supports. With a corrected evaluation protocol or substantially narrowed claims, the comparison could be a useful reference for the community.","major_comments":[{"comment":"The simulator selects the exposure time T based on the ground-truth luminance: the paper states, 'we select the exposure time based on the luminance of the ground truth HDR images.' This is oracle access to the test target. A real SPC camera uses a fixed or user-chosen exposure, not one hand-picked from the scene radiance to maximize visibility. This choice removes exactly the saturation and exposure-robustness problem that motivates SPC HDR imaging, so the reported ranking may reflect how well the methods exploit oracle-exposed inputs rather than how they would perform on genuine SPC captures. Please evaluate with held-out exposure settings, such as one fixed exposure per resolution or a sweep over exposures, and report how the ranking changes.","section":"Section 3.2"},{"comment":"The SPC images are saved and fed to the pipeline as 8-bit tone-mapped monochromatic images. This discards the high-bit-depth photon-count information that constitutes the SPC's extended dynamic range, effectively turning the problem into monochrome-LDR-to-color-HDR rather than SPC-to-HDR. If the chosen architectures require 8-bit input, the claims should be narrowed accordingly, or the paper should include an experiment that preserves higher bit depth and, ideally, validate on real SPC sensor data.","section":"Sections 3.2 and 4"},{"comment":"The paper does not report the train/test split used for the Laval dataset, although the same dataset provides both the simulated SPC inputs and the ground-truth targets. Table 2 reports scores 'averaged over 262 images' but never states that those images were excluded from training. Without this information, the quantitative comparisons are potentially circular. Please specify the exact split, the number of training and test images, and ideally release code and the simulator settings so that the results can be reproduced.","section":"Sections 4 and 5, Tables 1-3"},{"comment":"The central ranking claim that Pix2PixHD + DeepHDR is 'the best possible HDR reconstruction' is based on mean HDR-VDP3 scores without any measure of variability or statistical significance. Since the differences between some competing combinations are relatively small (for example, 3.6388 versus 3.4545 for Pix2PixHD + DeepHDR versus Pix2Pix + DeepHDR at 1024x512), please report standard deviations, confidence intervals, or paired significance tests over the 262 test images to support the claim.","section":"Tables 2 and 3"}],"minor_comments":[{"comment":"There is a notation inconsistency: the text says 'the appropriate exposure time, denoted τd' but the equations use T for exposure time and τd for dead time.","section":"Section 3.2"},{"comment":"The statement that SPAD sensors 'have surpassed' CMOS sensors is too strong given the authors' own caveat in Section 3.3 that current SPAD prototypes cannot compete directly with CMOS image sensors; please soften the wording.","section":"Abstract and Section 1"},{"comment":"The caption says 'moving from Monochromatic HDR to Color HDR,' but the intended task is from monochromatic SPC to color HDR; please correct the wording.","section":"Figure 3 caption"},{"comment":"The row labels 'Low+High' are unclear; please replace them with explicit resolution names or add a note explaining the notation.","section":"Table 2"},{"comment":"The metric name is spelled inconsistently as 'HDR-VDP3' and 'HDR-VDP-3'; please use one form throughout.","section":"Section 5"},{"comment":"The single-stage results in Table 3 are described as 'comparable' to the two-stage results, but the Q-scores are consistently lower at 1024x512 and only close at 2048x1024; please qualify this statement more precisely.","section":"Section 5 and conclusions"}],"recommendation":"major_revision","confidential_remarks":"The paper is an empirical evaluation without code or data release, and the manuscript does not state a train/test split. For a journal publication, I would recommend requiring the split and simulator configuration to be released, and ideally a real-sensor validation or a clearly labeled idealized benchmark. The novelty is modest, but the comparison could be useful if the evaluation protocol is made rigorous and the claims are matched to the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a systematic evaluation, not a real-sensor demonstration. It compares twelve two-stage pipelines (colorization plus single-image HDR) and one single-stage variant on simulated single-photon-camera images from the Laval HDR dataset. Within its own setup, the tables support the headline ranking: Pix2PixHD + DeepHDR wins on HDR-VDP3, 4-frame averaging helps, and higher resolution helps. The authors mostly stay in evaluation mode and even concede in Section 3.3 that current SPAD prototypes are not ready to compete with CMOS. That honesty is worth something.\n\nThe soft spots are real, and one is load-bearing. Section 3.2 says the simulator selects exposure time based on the luminance of the ground-truth HDR image. That is oracle access to the test-time answer: the simulator guarantees all scene regions are visible, which removes the exposure-robustness problem that motivates SPC imaging in the first place. On top of that, the monochromatic SPC images are saved as 8-bit tone-mapped images, discarding the high-bit-depth photon-count information that gives SPADs their dynamic-range advantage. So the task being evaluated is effectively monochrome-LDR-to-color-HDR, not SPC-to-HDR. That undermines the external validity of the ranking.\n\nThere are smaller issues too. No train/test split is reported, no error bars, no code or data. The single-stage “modified Pix2PixHD” is not actually described as an architectural change; Section 3.3 just says it was modified. The abstract and conclusion imply a level of SPAD readiness that the paper's own Section 3.3 contradicts. The simulator equations are taken from prior work, so there is no derivation to audit, but the protocol around them is the problem.\n\nWho is this for? Someone looking for a quick sanity check of which off-the-shelf components might work on synthetic SPC-like inputs. It is a useful benchmark on simulated data, not a guide for real sensor pipeline design. I would not cite it as evidence about actual SPC performance, but I might cite it as a baseline if I were building a similar simulated comparison.\n\nRecommendation: it deserves a serious referee, but only with major revisions. The oracle exposure selection must be disclosed prominently or, better, replaced with a fixed or randomly sampled exposure protocol. The 8-bit quantization should be justified or removed. Code, data splits, and error bars are needed. If the Pix2PixHD + DeepHDR ranking survives a non-oracle protocol, the paper becomes genuinely useful. As is, it is a well-executed but idealized benchmark.","headline":"Useful synthetic benchmark, but oracle exposure selection and 8-bit tone-mapping make the claimed ranking unlikely to transfer to real SPC sensors.","tokens_in":11408,"tokens_out":1820,"would_cite":false,"duration_ms":18094,"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 shows that a two-stage pipeline of Pix2PixHD colorization followed by DeepHDR HDR reconstruction produces the best color HDR images from monochrome single-photon camera captures, with a single-stage variant nearly matching it.","keywords":["Single Photon Avalanche Diode","High Dynamic Range Imaging","Generative Adversarial Networks","image-to-image translation","single image HDR reconstruction","SPAD simulator","HDR-VDP3","colorization"],"falsifier":"Capture the same scenes with a real single-photon camera and a conventional color HDR camera, feed the real monochrome frames through the trained Pix2PixHD + DeepHDR pipeline, and compare the perceptual HDR quality scores against the simulated results; a large drop in score or a change in the method ranking would show that the simulator's representation is not faithful.","tokens_in":10262,"feed_emoji":"📷","tokens_out":16882,"duration_ms":126878,"temperature":0.7,"pith_summary":"The paper asks whether the noisy, monochrome images produced by single-photon cameras—sensors that detect individual photons and therefore have an extremely wide dynamic range—can be turned into color high-dynamic-range (HDR) images using existing deep-learning tools. It answers with a two-stage recipe: first translate the monochrome single-photon-camera image into a color low-dynamic-range image, then reconstruct a color HDR image from that. Across twelve combinations of colorization and HDR networks, the paper reports that Pix2PixHD followed by DeepHDR gives the highest HDR-VDP3 quality scores, and that a one-stage Pix2PixHD variant trained to go directly from monochrome single-photon camera images to color HDR performs comparably. If correct, this provides a computational route to usable color HDR output from single-photon sensors before color-capable SPAD hardware is widely available.","feed_headline":"Pix2PixHD plus DeepHDR turns SPAD images into color HDR","feed_subtitle":"The winning combo tops HDR-VDP3 scores on simulated single-photon camera frames, and a one-stage network comes close.","key_machinery":"The load-bearing object is the passive-mode single-photon camera (SPC) model and the two-stage network stack built around it. In the simulator, photon arrivals per pixel follow Poisson statistics, and each detected photon imposes a dead time $\\tau_d = 150\\,\\mathrm{ns}$, producing a soft-saturating response with mean $\\mathbb{E}[N^{SPC}_T] = q_{SPAD}\\phi T/(1+q_{SPAD}\\phi\\tau_d)$ and variance $q_{SPAD}\\phi T/(1+q_{SPAD}\\phi\\tau_d)^3$; the exposure time is chosen from the ground-truth luminance, and the resulting monochrome frames are saved as 8-bit images. The pipeline then applies a GAN-based image-to-image translator (Pix2PixHD, Pix2Pix, or CycleGAN) for colorization and a single-image HDR network (DeepHDR, SIHDR, ExpandNet, or LHDR) for the HDR stage, with HDR-VDP3 as the evaluation metric.","core_discovery":"Monochromatic single-photon camera images can be converted into color HDR images by a two-stage approach: an image-to-image translation network first colorizes the SPC image, and a single-image HDR network then expands it to HDR. On the indoor HDR dataset used, with SPC images generated by the paper's Poisson/dead-time simulator, the combination of Pix2PixHD for colorization and DeepHDR for HDR reconstruction achieves the highest HDR-VDP3 scores of all twelve two-stage combinations, at both 1024x512 and 2048x1024 resolutions, and averaging four SPC frames improves the results. A single-stage variant—Pix2PixHD modified to translate directly from monochrome SPC images to 32-bit-per-channel HDR images—produces comparable scores, although the paper reports that its colors deviate slightly from the ground truth.","pith_inferences":["A natural extension the paper leaves untested is applying the same two-stage recipe to real SPAD sensor data; the reported ranking may change because the simulator selects exposure time from the known ground-truth luminance, which a real camera cannot do.","The near-parity of the single-stage network suggests the colorization and HDR-expansion tasks share enough structure to be learned jointly; a lighter joint model with a color-aware loss might close the remaining gap.","Because the pipeline's input is tone-mapped to 8-bit monochrome, the paper effectively measures what is achievable from compressed SPC output; feeding higher-bit-depth representations to the first stage could recover more of the camera's native dynamic range.","The final HDR stage is judged only by HDR-VDP3; perceptual color-fidelity metrics or human ratings could rank the method combinations differently."],"forward_implications":["A two-stage pipeline of Pix2PixHD plus DeepHDR gives a working computational path from monochrome single-photon camera frames to color HDR images without color-capable SPAD hardware.","Averaging four SPC frames instead of one raises both the colorization metrics (LPIPS, PSNR, SSIM) and the final HDR-VDP3 score across all tested method combinations.","Higher input resolution (2048x1024) improves HDR reconstruction quality for every combination tested, so resolution is a direct lever on output quality within this pipeline.","A single-stage Pix2PixHD variant trained directly from monochrome SPC to color HDR reaches scores close to the two-stage pipeline, making the conversion possible in one network when slight color deviations are acceptable."],"supporting_citations":[{"why":"Supplies the dead-time/Poisson SPC simulator model and the mean-variance equations used to generate monochrome input images from ground-truth HDR photos.","marker":"[9]"},{"why":"Provides the Pix2PixHD architecture that wins the colorization stage and also serves as the backbone for the single-stage translation variant.","marker":"[28]"},{"why":"Defines DeepHDR, the single-image HDR reconstruction network that, combined with Pix2PixHD, produces the highest HDR-VDP3 scores.","marker":"[34]"},{"why":"Defines HDR-VDP3, the perceptual metric used for all quantitative HDR quality comparisons in the paper.","marker":"[43]"},{"why":"Provides the high-resolution indoor HDR dataset from which ground-truth images are downsampled for simulation, training, and testing.","marker":"[47]"},{"why":"Supplies the frame-averaging method used to generate multi-frame averaged SPC inputs that improve colorization and HDR results.","marker":"[36]"},{"why":"Provides the high-flux passive imaging statistics that the simulator uses to compute photon-count mean and variance.","marker":"[10]"},{"why":"Presents Pix2Pix, the colorization baseline that the paper compares against Pix2PixHD.","marker":"[16]"},{"why":"Presents CycleGAN, the unsupervised colorization baseline used in the comparison.","marker":"[46]"},{"why":"Presents SIHDR, one of the single-image HDR baselines that DeepHDR is compared against in the second stage.","marker":"[33]"}],"fun_headline_variants":["Two-stage Pix2PixHD-DeepHDR tops SPAD color HDR conversion","Single-stage network rivals two-stage for SPAD color HDR","Pix2PixHD plus DeepHDR best for SPAD-to-color HDR","Color HDR from SPAD images: two-stage wins, single-stage close"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire ranking rests on the assumption that the simulated single-photon camera images—generated from ground-truth HDR photos with exposure time chosen from the known scene luminance and then stored as 8-bit monochrome—faithfully represent what a real single-photon camera would capture.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage Pix2PixHD-DeepHDR tops SPAD color HDR conversion","Single-stage network rivals two-stage for SPAD color HDR","Pix2PixHD plus DeepHDR best for SPAD-to-color HDR","Color HDR from SPAD images: two-stage wins, single-stage close"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000731,"raw_usage":{"total_tokens":3288,"prompt_tokens":979,"completion_tokens":2309,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":2235}},"tokens_in":595,"tokens_out":2309,"duration_ms":14994,"temperature":1.0,"reasoning_tokens":2235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:34:10.919710+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Capture the same scenes with a real single-photon camera and a conventional color HDR camera, feed the real monochrome frames through the trained Pix2PixHD + DeepHDR pipeline, and compare the perceptual HDR quality scores against the simulated results; a large drop in score or a change in the method ranking would show that the simulator's representation is not faithful.","supporting_citations":[{"cited_title":"Single-photon camera guided extreme dynamic range imaging","cited_arxiv_id":null,"evidence_quote":"Supplies the dead-time/Poisson SPC simulator model and the mean-variance equations used to generate monochrome input images from ground-truth HDR photos."},{"cited_title":"High-resolution image synthesis and semantic manipulation with condi- tional gans","cited_arxiv_id":null,"evidence_quote":"Provides the Pix2PixHD architecture that wins the colorization stage and also serves as the backbone for the single-stage translation variant."},{"cited_title":"Single Image HDR Reconstruction Using a CNN with Masked Features and Perceptual Loss","cited_arxiv_id":"2005.07335","evidence_quote":"Defines DeepHDR, the single-image HDR reconstruction network that, combined with Pix2PixHD, produces the highest HDR-VDP3 scores."},{"cited_title":"Photon-starved scene inference using single photon cameras","cited_arxiv_id":null,"evidence_quote":"Supplies the frame-averaging method used to generate multi-frame averaged SPC inputs that improve colorization and HDR results."},{"cited_title":"High flux passive imaging with single-photon sensors","cited_arxiv_id":null,"evidence_quote":"Provides the high-flux passive imaging statistics that the simulator uses to compute photon-count mean and variance."},{"cited_title":"Image-to-image translation with conditional adversarial networks","cited_arxiv_id":null,"evidence_quote":"Presents Pix2Pix, the colorization baseline that the paper compares against Pix2PixHD."},{"cited_title":"Unpaired image- to-image translation using cycle-consistent adversarial networks","cited_arxiv_id":null,"evidence_quote":"Presents CycleGAN, the unsupervised colorization baseline used in the comparison."},{"cited_title":"Single-image HDR reconstruction by learn- ing to reverse the camera pipeline","cited_arxiv_id":null,"evidence_quote":"Presents SIHDR, one of the single-image HDR baselines that DeepHDR is compared against in the second stage."}],"review_version":1}