{"id":"ee043ca8-cc30-48e8-b3b9-cd64305defbf","arxiv_id":"2507.17157","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"UNICE trains a two-stage model to generate and fuse a pseudo multi-exposure sequence from one image, generalizing across four contrast-enhancement tasks without human labels.","lead":"The paper trains a two-stage neural network, UNICE, that enhances any image by first generating a fake multi-exposure sequence and then fusing it into one improved picture. The authors claim this single model generalizes across low-light, exposure, backlit, and HDR-reconstruction tasks without human-labeled training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MES-Net's 8-bit L2 training cannot recover HDR information absent from the input, so the claimed HDR-recovery mechanism is unvalidated; the two-stage gain over SD-T lacks a demonstrated cause.","rationale":"The central contribution is a label-free universal contrast enhancer whose design is motivated by the claim that multi-exposure sequences carry HDR information that a single sRGB image lacks (Sec. 1, Sec. 3.3). The empirical generalization claim is supported by extensive cross-task/cross-dataset experiments, including FR-IQA metrics where UNICE also leads, and by the SD-T ablation that isolates the benefit of the two-stage MES+MEF architecture. I do not dispute the empirical generalization result. However, the method's internal logic has a soft spot: MES-Net is trained to map between 8-bit sRGB frames rendered from the same raw (Eq. 2). A single 8-bit frame has already lost out-of-range radiance values via clipping and quantization; the target frame at another EV also only contains the same scene's 8-bit information. Consequently, the learned mapping is a function of the input alone and cannot inject information that was not captured. The generated pseudo-MES is therefore not a real HDR acquisition, and the MEF stage cannot perform true dynamic-range expansion. The paper never validates that generated MES images resemble real multi-exposure captures, nor that they enable detail recovery in clipped regions. If this test fails, the claimed mechanism is wrong, and the method's advantage over the direct SD-T baseline must be attributed to some other inductive bias of the two-stage decomposition, which is currently unexplained. I selected this over the NR-IQA selection-bias concern because the latter only affects the GT-comparison highlight (Tab. 3) and can be addressed by reporting FR metrics; the MES-HDR issue reaches the core design of the method and its stated rationale. The recommendation matches the reader's CONDITIONAL verdict: the empirical claims are credible, but the paper should provide the validation above or substantially reframe the mechanism.","tokens_in":22867,"tokens_out":8773,"duration_ms":90402,"concrete_test":"Hold out a set of raw images from the training sources (e.g., 100 from RAISE/PPR10K). For each raw, render the full 7-exposure MES. Take one exposure (e.g., EV=-2) as input to UNICE's MES-Net and generate the other six exposures. (1) Compute PSNR/SSIM/LPIPS between each generated exposure and the corresponding real rendered exposure; low fidelity indicates the generated MES is not a real MES proxy. (2) Apply the same MEF algorithm used in Sec. 3.2.2 to the real and generated MES; in regions clipped in the input (shadows/highlights), compare detail preservation (e.g., gradient magnitude, or a reference-based metric against the raw-derived fusion). If generated-MES fusion recovers no more detail than directly upscaling/tone-mapping the input, the claimed HDR-recovery mechanism is absent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 3.3 trains MES-Net with an L2 loss on 8-bit sRGB pairs (Eq. 2), mapping (zk/zj)·I_j to I_k. Since both I_j and I_k are 8-bit renders of the same raw scene, the mapping is a deterministic tone/color transform of the input. An 8-bit sRGB image already clips highlights and shadows; those lost values are not present in the input, so no learned transform can place them in the generated MES. Thus the generated 'pseudo-MES' is informationally equivalent to the input up to the learned transform, contradicting the claim in Sec. 3.3 that it 'can acquire the HDR information of the scene.' The MEF-Net stage therefore cannot extend dynamic range beyond the input; any recovered detail is hallucinated by the SD-Turbo prior, not fused from real HDR data. The paper provides no check that generated MES images match real multi-exposure captures from the same raw, nor that fusion of generated MES preserves details that exist only in real well-exposed frames. This matters because the SD-T ablation (Tabs. 6-7) is the key evidence that the MES pipeline helps; if the MES carries no extra information, the two-stage advantage must stem from a different inductive bias, leaving the paper's central mechanism unexplained.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UNICE, a universal contrast enhancement model trained entirely on automatically generated data without human-annotated ground truths. The authors collect 48,361 HDR raw images from public sources, render multi-exposure sequences (MES) through an emulated ISP, generate pseudo-GTs by ensemble multi-exposure fusion (MEF) with NR-IQA-based selection, then train two one-step diffusion networks: MES-Net, which synthesizes an MES from a single sRGB image, and MEF-Net, which fuses the generated MES into an enhanced image. Experiments on LLIE, EC, BIE, and L2HT report strong cross-task and cross-dataset generalization, with UNICE often outperforming task-specific baselines and sometimes surpassing manual ground truths in no-reference IQA metrics.","tokens_in":23225,"tokens_out":5799,"duration_ms":56902,"significance":"If the results hold, this is a valuable contribution: the large-scale automated data pipeline (328,496 pairs) and a single model addressing four contrast-related tasks go beyond typical task-specific designs, and the SD-T baseline provides a useful control for attributing gains to the MES/MEF pipeline rather than the diffusion backbone. The release of code, model, and dataset increases reproducibility. However, two load-bearing issues temper the significance: the claim of outperforming manual GTs is circular because the same NR-IQA metrics used to select pseudo-GTs are used for that comparison, and the HDR-information-recovery mechanism of MES-Net is unvalidated given the 8-bit L2 training objective.","major_comments":[{"comment":"The pseudo-GT selection pipeline in Sec. 3.2.2 ranks and selects candidate MEF outputs using NIQE, BRISQUE, PI, and ARNIQA, and the same four NR-IQA metrics are then used in Table 3 to claim that UNICE outputs 'even exceed manually labeled GTs.' Because the network is trained on pseudo-GTs that were explicitly chosen to score well on these metrics, the comparison in Table 3 does not establish perceptual superiority over manual GTs; it is circular. Please replace or supplement this claim with a human perceptual study comparing UNICE outputs against manual GTs, or with FR-IQA/human-opinion metrics that were not used in pseudo-GT selection, and revise the abstract's claim accordingly if the circular evidence is removed.","section":"Sec. 3.2.2 / Table 3"},{"comment":"MES-Net is trained with an L2 loss on the mapping (z_k/z_j)*I_j -> I_k (Eq. 2), where both I_j and I_k are 8-bit sRGB renders of the same raw scene. Since the 8-bit input has already undergone quantization and clipping, no deterministic learned transform can reintroduce the lost HDR values; the statement in Sec. 3.3 that the synthesized MES 'can acquire the HDR information of the scene' is therefore not supported. The paper provides no comparison between generated MES and real multi-exposure captures from the same raw, nor any experiment showing that the fusion step recovers details that exist only in real well-exposed frames. Because the SD-T ablation (Tabs. 6-7) is the main evidence that the MES/MEF pipeline contributes beyond the backbone, the mechanism behind that gain remains unexplained. Please add a validation study (e.g., compare generated MES with real MES, check detail recovery in clipped regions) or revise the framing to attribute the gain to the learned fusion objective or inductive bias rather than HDR recovery.","section":"Sec. 3.3, Eq. (2)"},{"comment":"The cross-task generalization results in Table 5 compare UNICE only against task-specific models trained on MSEC (24,330 images) and omit the SD-T baseline, which is present in Tables 6 and 7. Without SD-T in the cross-task setting, the large gains in Table 5 over LCDPNet, RF, and CSEC could be attributed primarily to the much larger and more diverse training set (328,496 pairs) rather than to the MES/MEF pipeline. Please include SD-T in the cross-task evaluation, or explicitly report the UNICE-vs-SD-T comparison for the same cross-task protocol, so that the attribution of the generalization advantage to the proposed pipeline is properly controlled.","section":"Sec. 4.3.1 / Table 5"}],"minor_comments":[{"comment":"The text refers to 'MSE-Net' and 'MFE-Net'; these should be 'MES-Net' and 'MEF-Net'.","section":"Sec. 4 (first paragraph)"},{"comment":"Tables 5 and 6 are both titled 'Cross-dataset evaluation on LLIE and EC tasks,' but both tables include results for BIE and L2HT as well; the titles should be revised to reflect the full scope of the tables.","section":"Tables 5 and 6"},{"comment":"The pretrained SD-Turbo model is cited as [71] in Sec. 3.3, but the reference list assigns [50] to the SD-Turbo paper; please correct the citation.","section":"Sec. 3.3"},{"comment":"In the BIE row of Table 4, the BAID dataset is cited as [21], but the BAID dataset is reference [4]; please fix the citation.","section":"Table 4"},{"comment":"The final row of Table 10 is the full UNICE model but is labeled 'SD-T [71] / SD-T [71]'; this label is confusing because 'SD-T' elsewhere denotes the direct-mapping baseline. Please rename the rows to distinguish the backbone choice from the SD-T baseline.","section":"Sec. 4.6.2 / Table 10"}],"recommendation":"major_revision","confidential_remarks":"The central generalization claim is grounded in held-out FR-IQA evaluations and a careful SD-T control, so I do not see grounds for rejection. The main publication risk is the circular 'outperforming manual GTs' claim in the abstract and Sec. 4.2; if the authors add a non-circular human study or remove the claim, the paper would be substantially stronger. The HDR-mechanism concern also needs an experimental response. The manuscript has several table-title and citation errors that are straightforward to fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper actually delivers a single model that generalizes across four contrast-enhancement tasks, and the held-out full-reference evaluations (not the NR-IQA ones) support that. Second, the 'HDR recovery' story for the synthetic multi-exposure sequence is not credible as stated; the paper never shows the generated MES recovers information that wasn't already in the 8-bit input.\n\nWhat is genuinely new: a 328k-image synthetic dataset built from ~47k HDR raws via emulated ISP rendering plus MEF pseudo-GTs, and a two-stage architecture that first synthesizes an exposure-adjusted sequence, then fuses it, fine-tuned from SD-Turbo with LoRA. Each component alone is known, but the combination at this scale is new, and the dataset is a reusable asset.\n\nThe experiments are the strong part. Cross-task and cross-dataset evaluations are thorough: baselines are retrained on their dataset, an SD-T direct-mapping baseline is included, and UNICE wins consistently on PSNR, SSIM, LPIPS, and DISTS across LLIE, EC, BIE, and L2HT. The user study is a reasonable extra. These results make a real empirical case for the universal-enhancer idea.\n\nThe soft spots are real but not fatal. First, the MES-Net HDR claim (Sec. 3.3, Eq. 2) does not survive information-theoretic scrutiny. The network maps an 8-bit render to another 8-bit render; both are tone/color transforms of the same clipped signal. Clipped highlights and shadows are absent from the input, so no learned transform can place them in the output. The generated MES cannot 'acquire the HDR information of the scene.' Any recovered detail is hallucinated by the diffusion prior. That does not automatically sink the method—the two-stage pipeline still beats SD-T—but it means the stated mechanism is wrong, and the ablation does not explain why the two-stage helps. Likely real causes are exposure-conditional generation acting as a strong augmentation, or the fusion stage giving the network a multi-scale way to attend to tone. The authors should either validate the generated MES against real multi-exposure captures or re-frame the contribution as a learned exposure-conditional prior rather than HDR recovery.\n\nSecond, the 'beats manual GTs' claim (Tab. 3) is partly circular: the same NR-IQA metrics used to select pseudo-GTs are used to claim superiority over manual GTs. The full-reference comparisons against task-specific baselines are not circular, and those are the ones that matter. The GT comparison should be treated as illustrative at best. Minor point: a few free parameters (EV range, number of fused images, ARNIQA threshold, style codes) lack sensitivity analysis.\n\nOverall, this paper is for anyone in image enhancement who wants a generalist model rather than a bag of task-specific fixes. The dataset alone justifies attention. Send it to serious peer review: the empirical generalization claim will likely hold up even though the mechanism section needs major revision, and a good referee will push the authors to either prove the HDR-recovery claim or drop it.","headline":"A strong, thoroughly tested generalization result wrapped in an overstated HDR-recovery mechanism; worth serious review, but the mechanism section needs rework.","tokens_in":23710,"tokens_out":3275,"would_cite":true,"duration_ms":35882,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single model trained without human labels outperforms task-specific contrast enhancers on low-light, exposure, backlit, and HDR benchmarks.","keywords":["universal image contrast enhancement","multi-exposure sequence synthesis","multi-exposure fusion","low-light image enhancement","exposure correction","backlit image enhancement","LDR-to-HDR transformation","no-reference image quality assessment"],"falsifier":"Collect static scenes with genuine bracketed exposures at several stop-separated exposure values, fuse them with a trusted HDR algorithm as the reference, and give UNICE a single frame from each scene. Compare UNICE's output with the reference in clipped highlight and shadow regions; if its recovered details do not track the real bracketed exposures, or if UNICE performs no better than direct input-to-ground-truth mapping on such scenes, the pseudo-MES mechanism is falsified.","tokens_in":22675,"feed_emoji":"📸","tokens_out":10429,"duration_ms":114877,"temperature":0.7,"pith_summary":"The paper tries to establish that one network, trained without any human labeling, can handle four distinct contrast problems that until now required separate models. Its central wager is that the common operation underneath low-light enhancement, exposure correction, backlit enhancement, and LDR-to-HDR transformation is exposure and contrast adjustment, and that this can be learned if the network first reconstructs a multi-exposure sequence from a single image and then fuses it. To get data at scale, the paper renders 328,496 sRGB images from 46,928 HDR raw files with an emulated camera pipeline, creates pseudo-ground-truths by multi-exposure fusion, and trains two one-step diffusion networks: MES-Net produces the exposure sequence and MEF-Net fuses it. If correct, the result is a single contrast enhancer that transfers across tasks and datasets, and whose outputs can score better than manually retouched references on no-reference quality metrics. That matters because manual ground-truth collection is the bottleneck that has kept each enhancement task small and task-specific.","feed_headline":"One model beats task-specific enhancers on four contrast tasks","feed_subtitle":"No human labeling: UNICE trains on auto-generated exposure sequences and beats per-task models on held-out data.","key_machinery":"The core mechanism is the synthesized multi-exposure sequence (MES) used as an intermediate representation of the scene's dynamic range. In the data stage, ~47k HDR raw images are rendered by an emulated ISP at seven exposure values, and an ensemble of multi-exposure fusion algorithms, ranked by no-reference metrics, produces pseudo-GTs. MES-Net (a one-step SD-Turbo diffusion network with LoRA adapters) maps an input sRGB image $I_j$ to another exposure $I_k$ using a style code $z$ equal to the mean intensity (Eq. 2, trained with L2 loss), thereby synthesizing a pseudo-MES from a single image. MEF-Net (same backbone) receives the pseudo-MES and learns normalized per-pixel weight maps $W(\\cdot)$ that form an implicit 32-bit HDR representation (Eqs. 3-4), then fuses it into a high-quality 8-bit image matching the pseudo-GT. The two-stage decomposition separates exposure/style change from fusion, which is what lets the model transfer to unseen tasks and datasets.","core_discovery":"On the paper's own terms, the discovery is that a single, label-free model, called UNICE, can generalize across and within four contrast-enhancement tasks because it learns an intermediate multi-exposure representation rather than a direct input-to-output mapping. From a large set of HDR raw images, an emulated ISP renders exposure sequences, and multi-exposure fusion algorithms generate pseudo ground truths, so no human-annotated pairs are needed. MES-Net synthesizes a pseudo multi-exposure sequence from one sRGB image, and MEF-Net fuses that sequence into a high-quality 8-bit image. The authors report that UNICE outperforms task-specific baselines on held-out LLIE, EC, BIE, and L2HT datasets, and that in many cases its outputs surpass manually created ground truths in no-reference image-quality metrics.","pith_inferences":["A direct test of the central mechanism would compare UNICE's output on clipped highlights and shadows against genuine bracketed captures of the same scene; because MES-Net only sees the 8-bit input, any detail it recovers there must be generated rather than measured, and it is an open question how often that generation is faithful.","The style code $z$, defined as mean intensity, could be exposed as a user-controlled dial to produce continuously variable enhancement rather than the three fixed exposure levels tested in the paper.","The same MES-then-fusion decomposition could extend to video contrast enhancement, where the pseudo-exposure sequence would need temporal consistency; this is a neighbouring task the paper does not address.","The NR-IQA-ranked ensemble used to select pseudo-GTs could be moved to inference time, letting the model compare several fusion candidates and pick the one with the best predicted quality, which the paper does not explore."],"forward_implications":["On the tested held-out datasets, a single UNICE model outperforms task-specific baselines on most full-reference and no-reference metrics, including large gains such as 19.40 dB PSNR versus 15.90 dB on LSRW for low-light enhancement.","Task-specific models retrained on UNICE's synthetic data improve over their original training, so the automatically generated dataset itself carries transferable value.","The MES synthesis plus fusion pipeline, not the SD-Turbo backbone alone, drives the gains: the direct input-to-ground-truth mapping baseline SD-T underperforms UNICE.","UNICE can match or beat manually created ground truths in multiple no-reference quality metrics, and wins 68–83% of user-study votes against per-task best models.","Despite a large diffusion backbone (~965M parameters, ~3.3 TFLOPs at 512×512), one-step inference keeps the model usable at 5.58 FPS on an A100 GPU.","A universal contrast enhancer reduces the need to train and deploy a separate model for each lighting-related task, which is the paper's main practical consequence."],"supporting_citations":[{"why":"Supplies the emulated ISP rendering procedure and the MSEC exposure-correction dataset used to build inputs and baseline comparisons.","marker":"[1]"},{"why":"Defines the L2HT task, provides the HDREye and HDR-Real test sets, and supplies the exposure model $I = ISP(S_i \\Delta t)$ that motivates MES synthesis.","marker":"[5]"},{"why":"One of the raw-image sources and evidence that manual retouching annotations are costly and inconsistent.","marker":"[28]"},{"why":"Provides raw portrait images as another source for the automatically generated training data.","marker":"[29]"},{"why":"Contributes the style-transfer, two-stage training idea and the emulated-ISP rendering approach on which the data pipeline builds.","marker":"[32]"},{"why":"Classical Mertens exposure-fusion operator that, with other MEF methods, generates the pseudo-GT candidates in the dataset pipeline.","marker":"[36]"},{"why":"Pretrained SD-Turbo diffusion model used as the frozen generative backbone for MES-Net and MEF-Net.","marker":"[50]"},{"why":"LoRA low-rank adaptation keeps the fine-tuned parameters task-specific and small.","marker":"[51]"},{"why":"Provides the one-step image-translation training setup and the SD-T direct-mapping baseline used in ablations.","marker":"[71]"}],"fun_headline_variants":["Label-free UNICE model masters four contrast tasks at once","Universal contrast enhancer trains itself, no labels needed","One model to fix all contrast issues: UNICE does it","No human labels, one model, four contrast tasks beaten","Multi-exposure trick yields universal enhancer, beats specialists"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that the three generated exposure images carry the same recoverable detail that real multi-exposure photos would, so that fusing them reconstructs true scene information; the paper never validates this against real multi-exposure captures.","fun_headline_variants_meta":{"raw":{"variants":["Label-free UNICE model masters four contrast tasks at once","Universal contrast enhancer trains itself, no labels needed","One model to fix all contrast issues: UNICE does it","No human labels, one model, four contrast tasks beaten","Multi-exposure trick yields universal enhancer, beats specialists"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000387,"raw_usage":{"total_tokens":2052,"prompt_tokens":963,"completion_tokens":1089,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":1008}},"tokens_in":579,"tokens_out":1089,"duration_ms":8800,"temperature":1.0,"reasoning_tokens":1008,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:56:03.254510+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect static scenes with genuine bracketed exposures at several stop-separated exposure values, fuse them with a trusted HDR algorithm as the reference, and give UNICE a single frame from each scene. Compare UNICE's output with the reference in clipped highlight and shadow regions; if its recovered details do not track the real bracketed exposures, or if UNICE performs no better than direct input-to-ground-truth mapping on such scenes, the pseudo-MES mechanism is falsified.","supporting_citations":[{"cited_title":"Learning multi-scale photo exposure correction,","cited_arxiv_id":null,"evidence_quote":"Supplies the emulated ISP rendering procedure and the MSEC exposure-correction dataset used to build inputs and baseline comparisons."},{"cited_title":"Single-image hdr reconstruction by learning to reverse the camera pipeline,","cited_arxiv_id":null,"evidence_quote":"Defines the L2HT task, provides the HDREye and HDR-Real test sets, and supplies the exposure model $I = ISP(S_i \\Delta t)$ that motivates MES synthesis."},{"cited_title":"Learning photo- graphic global tonal adjustment with a database of input/output image pairs,","cited_arxiv_id":null,"evidence_quote":"One of the raw-image sources and evidence that manual retouching annotations are costly and inconsistent."},{"cited_title":"Ppr10k: A large- scale portrait photo retouching dataset with human-region mask and group-level consistency,","cited_arxiv_id":null,"evidence_quote":"Provides raw portrait images as another source for the automatically generated training data."},{"cited_title":"Unsupervised exposure correction,","cited_arxiv_id":null,"evidence_quote":"Contributes the style-transfer, two-stage training idea and the emulated-ISP rendering approach on which the data pipeline builds."},{"cited_title":"Exposure fusion: A simple and practical alternative to high dynamic range photography,","cited_arxiv_id":null,"evidence_quote":"Classical Mertens exposure-fusion operator that, with other MEF methods, generates the pseudo-GT candidates in the dataset pipeline."},{"cited_title":"Lora: Low-rank adaptation of large language models,","cited_arxiv_id":null,"evidence_quote":"LoRA low-rank adaptation keeps the fine-tuned parameters task-specific and small."}],"review_version":1}