{"id":"5025d28e-4fc7-4ade-89f0-ab6d0bd0cfa7","arxiv_id":"2608.08720","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DPEC fine-tunes a pretrained diffusion model for single-step exposure correction and fuses its low-frequency output into a regression network, improving perceptual metrics on LCDP, MSEC, and SICE.","lead":"The paper fine-tunes Stable Diffusion to correct underexposed and overexposed photos in one quick step, then adds a detail-preserving network to keep sharpness. The method shows gains on standard exposure-correction benchmarks, but the gains are small and the claim of consistent top performance is too strong.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table VI shows the full model beats regression-only by only 0.22 dB PSNR and 0.0138 LPIPS, and Tables I–III show CoTF wins MSEC PSNR and NIQE on four benchmarks; without error bars or significance tests, the abstract's claim of consistent superiority is not supported.","rationale":"The reader's verdict is CONDITIONAL with MODERATE confidence, and my independent stress-test agrees. The method is internally coherent: the fine-tuning strategy, v-prediction objective, single-step sampling, and joint cross-attention module are described consistently, and the ablations are internally consistent. However, the abstract's strongest claim—consistent superiority on fidelity, perceptual quality, and visual effects—is contradicted by the paper's own tables: CoTF is better on MSEC PSNR and on NIQE for MSEC, SICE, LIME, and DICM. No error bars, significance tests, code, or checkpoints are provided, and the key ablation margin over regression-only is small (0.22 dB PSNR, 0.0138 LPIPS). The paper itself acknowledges failure cases with color deviation and lost details in large over/underexposed regions. These issues do not invalidate the method or its direction, but they do mean the central claim is not currently supported by the reported evidence. The appropriate verdict remains CONDITIONAL: the authors should provide reproducible artifacts, statistical validation, and revise the overclaim. No change to the reader's verdict is needed.","tokens_in":23487,"tokens_out":1561,"duration_ms":16037,"concrete_test":"Run the full DPEC model and the regression-only baseline on the LCDP test set with at least 5 independent training seeds, and report mean ± std for PSNR, SSIM, and LPIPS plus a paired significance test over the 218 test images. If the LPIPS gap shrinks below 0.005 or fails to reach significance, the claimed benefit of diffusion prior fusion is not supported. In parallel, re-evaluate CoTF and DPEC on the MSEC test set using identical preprocessing and test splits; if CoTF's PSNR remains higher, the abstract's 'consistently outperforms' claim must be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that DPEC consistently outperforms state-of-the-art methods on fidelity, perceptual quality, and visual effects. The paper's own tables contradict this: on MSEC, CoTF has higher PSNR (23.44 vs 23.29) and better NIQE (3.70 vs 3.74); on SICE, CoTF has better NIQE (3.13 vs 3.28); and on LIME and DICM, CoTF has better NIQE. The claimed advantage is mainly in LPIPS and SSIM, not in all three stated dimensions. More fundamentally, all numbers are single runs with no error bars, no standard deviations, and no significance tests. The key load-bearing assumption is that the 0.22 dB / 0.0138 LPIPS gain from fusing diffusion priors (Table VI) is a real effect, not noise or overfitting, and that the fine-tuned U-Net at t=1000 with 64x64 downsampled input provides useful low-frequency content rather than acting as a complex regularizer that a simpler encoder could replace. Without released code or checkpoints, independent verification is impossible, so the headline claim is not established by the reported evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DPEC, a two-stage exposure correction framework built on a pre-trained Stable Diffusion model. In the first stage, the denoising U-Net is fine-tuned with a v-prediction objective at a fixed maximal timestep t=T, and inference is performed in a single denoising step on 64x64 downsampled images. In the second stage, the fine-tuned U-Net is frozen and its multi-scale features are fused into a regression encoder-decoder through a proposed Joint Cross-Attention Module (JCAM), with a regression loss combining L1, perceptual, and SSIM terms. The method is evaluated on LCDP, MSEC, and SICE with PSNR/SSIM/LPIPS/NIQE, on unlabeled LIME/DICM/NPE with NIQE, and compared against a large set of existing methods including diffusion-based restoration baselines and CoTF. Ablations cover the fine-tuning strategy, the noise timestep, the diffusion-regression integration, and the JCAM module.","tokens_in":23659,"tokens_out":6539,"duration_ms":67289,"significance":"If the reported results are reproducible, DPEC makes a useful practical contribution: it shows that a generative diffusion prior can be adapted to exposure correction with a single-step low-resolution denoising pass, avoiding the multi-step sampling cost of typical diffusion-based restoration. The division of labor between low-frequency generation by the diffusion model and high-frequency refinement by the regression network is a sensible design, and the mathematical relations in Eqs. (1)-(4) are correct. The paper also ships a reasonable set of ablations that are internally consistent. The significance is moderated, however, by the lack of any statistical uncertainty quantification and by the fact that the paper's own tables do not support the abstract's claim of consistent state-of-the-art performance across all reported fidelity, perceptual, and no-reference metrics.","major_comments":[{"comment":"The abstract claims that DPEC 'consistently outperforms existing state-of-the-art methods' across fidelity, perceptual quality, and visual effects, but the paper's own tables contradict this. On MSEC, CoTF [60] achieves higher average PSNR (23.44 vs 23.29) and better NIQE (3.70 vs 3.74). CoTF also achieves better NIQE on SICE (3.13 vs 3.28), LCDP (3.2123 vs 3.3592), LIME (3.17 vs 3.25), and DICM (3.10 vs 3.22). The authors should either revise the claim to state specifically which metrics and datasets DPEC improves upon (e.g., SICE PSNR/SSIM/LPIPS, LCDP PSNR/SSIM/LPIPS, NPE NIQE) or provide statistical significance tests demonstrating that the reported differences are meaningful. As written, the headline claim is not supported.","section":"Abstract and Section IV-B, Tables I-III"},{"comment":"The noise-level ablation is confounded. The text states that 'the test input is pure Gaussian noise,' which is only consistent with training at t=1000. For t=500 and t=1, the training-time forward process produces y_t distributions that differ substantially from pure Gaussian noise, so the poor results at those settings can be attributed to a train-test mismatch rather than to the noise level itself. A valid ablation would train and test with matched noise schedules at each t, or otherwise vary the inference-time noise distribution together with the training choice. Without this, the conclusion that t=1000 is the best design choice is not established.","section":"Section IV-C.2, Table V"},{"comment":"The key load-bearing evidence that diffusion priors improve the final model is the comparison in Table VI: the full model beats 'Only Regression' by 0.22 dB PSNR, 0.0048 SSIM, and 0.0138 LPIPS on LCDP. No error bars, standard deviations, or significance tests are reported anywhere in the paper, so this gain could plausibly lie within run-to-run variability. Given that the entire architecture is justified by this improvement, the authors should provide repeated-run statistics (e.g., multiple training seeds with mean and std, or paired significance tests) for at least this central ablation.","section":"Section IV-C.3, Table VI"},{"comment":"The claim that 'increasing the number of time steps provides almost no additional improvement' is not directly supported by the reported experiments. Table IV compares 'Multi-noise 20' (22.03 dB), 'Multi-noise 1' (21.58 dB), and 'Single-noise 1' (21.97 dB), but these differ in both training strategy and sampling steps, so the comparison does not isolate the effect of the number of inference steps for a fixed model. A controlled experiment that varies only the sampling-step count of the same trained model is needed to substantiate the single-step efficiency claim.","section":"Section III-B.2 and Table IV"}],"minor_comments":[{"comment":"The expectation in L_diff is written as E_{x,y,t,eps}, but in the proposed strategy t is fixed to T. Please rewrite the objective with t=T made explicit, or clarify that a degenerate distribution over t is intended.","section":"Equation (3)"},{"comment":"The description of how the MSEC dataset is split into 'Under' and 'Over' subsets would be easier to follow if the text clarified whether the 'Average' column in Table I is the mean of the two subset columns.","section":"Section IV-B.1"},{"comment":"The text says the proposed fine-tuning strategy 'demonstrates better performance in single-step generation,' but the single-noise 1-step setting (21.97 dB) is slightly lower than the multi-noise 20-step setting (22.03 dB). 'Comparable' would be a more accurate descriptor.","section":"Section IV-C.1, Table IV"},{"comment":"The JCAM ablation reports PSNR, SSIM, and LPIPS. Adding a no-reference metric such as NIQE, or inference-time overhead, would strengthen the claim that JCAM improves the overall quality 'across multiple performance metrics.'","section":"Section IV-C.4, Table VIII"},{"comment":"There are several typos and formatting inconsistencies, for example 'V AE' instead of 'VAE' in Sections III-C and IV-C.3, and the phrase 'the effect of the noise level' could be more specific as 'the effect of the noise level on reconstruction quality.'","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main comparison target CoTF appears to be from the same research group as the current paper (shared authors in [60] and this manuscript). This is not a problem by itself, but it increases the importance of independent verification; the manuscript does not release code or checkpoints, and all quantitative results are single-run values. I would encourage the editor to require the authors to provide error bars or multiple-seed statistics, and to consider whether the abstract's claims need to be toned down to match what the tables actually show."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is actually decent. The authors fine-tune Stable Diffusion at a fixed noise level (t=1000) with the VAE removed, run a single denoising step at 64x64, and fuse the U-Net's features into a full-resolution regression network via a symmetric joint cross-attention module. That combination is new, the math in Eqs. (1)-(4) is right, and the ablations are internally consistent. They even show failure cases and discuss limitations, which I appreciate. If the reported single-step training really preserves most of the generative prior, the method is a practical way to use diffusion priors without multi-step sampling cost.\n\nThe problem is that the paper oversells its results. The abstract says DPEC \"consistently outperforms existing state-of-the-art methods\" on fidelity, perceptual quality, and visual effects. The paper's own tables contradict that. On MSEC, CoTF has higher PSNR (23.44 vs 23.29) and better NIQE (3.70 vs 3.74). On SICE, LIME, and DICM, CoTF also wins NIQE. The claimed advantages are mostly LPIPS and SSIM, not all three stated dimensions. The body text is more careful, using words like \"competitive,\" but the abstract is a clear overclaim.\n\nMore substantively, the key evidence that the diffusion prior helps is Table VI: the full model beats regression-only by 0.22 dB PSNR and 0.0138 LPIPS. That's a small gain, and there are no error bars, standard deviations, or significance tests anywhere in the paper. Given the randomness in diffusion and the cost of training, I cannot tell whether that gain is real or just noise on the LCDP test set. No code, checkpoints, or data are released, so independent reproduction is impossible. The efficiency claim is also relative: 0.3096 s per image is fast for a diffusion method, but it is about 44x slower than CoTF's 0.0071 s, so calling the method practical compared to regression baselines is a stretch.\n\nThe direction is worth pursuing, and the architecture choices are sensible, but the evidence as presented is not sufficient to support the paper's headline. This deserves a serious referee, but the authors need to fix the overclaim, provide variance or significance testing, and ideally release code or checkpoints.\n\nRecommendation: send it to peer review with a request for major revision and artifact release.","headline":"A plausible single-step diffusion-prior recipe for exposure correction, but the abstract's consistent-SOTA claim is contradicted by the paper's own tables and the missing artifacts keep the key gain unverified.","tokens_in":24305,"tokens_out":1593,"would_cite":false,"duration_ms":18124,"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":"Diffusion priors fix exposure correction in one denoising step, the paper argues.","keywords":["exposure correction","diffusion model","Stable Diffusion","generative priors","image enhancement","cross-attention","single-step denoising"],"falsifier":"Run DPEC on LCDP with the diffusion branch replaced by a trainable bilinear or VAE-based low-frequency generator at the same 64x64 resolution and identical fusion; if PSNR and LPIPS remain within noise of the full model, the generative prior itself is not contributing. Alternatively, feed the diffusion branch a zero image with no content and check whether the one-step output still resembles plausible image structure; if it does, the prior dominates the condition, which would predict failure on inputs whose extreme exposure regions cover most of the frame.","tokens_in":23176,"feed_emoji":"📸","tokens_out":3972,"duration_ms":34031,"temperature":0.7,"pith_summary":"The paper claims that exposure correction can be recast as a single-step conditional denoising task by fine-tuning a pre-trained Stable Diffusion model, then fusing its low-frequency features into a regression network that preserves high-frequency detail. The method, DPEC, is reported to outperform prior state-of-the-art methods on the LCDP, MSEC, and SICE benchmarks in both fidelity and perceptual metrics. The key move is to let the diffusion model work only on a 64x64 downsampled image, so it supplies believable low-frequency content for blown-out or crushed regions without the artifacts and cost of full-resolution iterative sampling. A joint cross-attention module merges the diffusion and regression features at multiple scales. If the claims hold, DPEC offers a practical recipe for exploiting generative priors in image restoration at regression-like inference cost.","feed_headline":"Diffusion priors fix exposure correction in one step","feed_subtitle":"DPEC fuses a fine-tuned Stable Diffusion with a regressor to beat prior benchmarks on LCDP, MSEC, and SICE.","key_machinery":"The central object is the fine-tuned denoising U-Net used as a low-frequency prior generator. Because the U-Net input is the concatenation of the noisy latent and a 64x64 downsampled version of the input image, the model only needs to refine illumination and large-scale structure rather than synthesize textures; training is restricted to the maximum timestep so the network learns one noise level, and inference is a single DDIM-style step using the v-prediction parameterization. The Joint Cross-Attention Module (JCAM) then lets diffusion features and regression features each query a concatenated key-value set, giving symmetric fusion across scales.","core_discovery":"On its own terms, the paper's central claim is that a pre-trained diffusion model can be adapted to exposure correction by fine-tuning it to denoise at a single fixed timestep ($T=1000$) with the downsampled poor-exposure image concatenated as condition, enabling one-step generation; and that the resulting low-frequency prior, when fused into a regression network through a joint cross-attention module, yields enhanced images that are simultaneously more faithful and more perceptually natural than those from regression-only or diffusion-only baselines. The reported numbers on LCDP (24.09 dB PSNR, 0.8627 SSIM, 0.0905 LPIPS) and on MSEC/SICE averages support the claim that DPEC achieves state-of-the-art performance.","pith_inferences":["A testable implication not explored in the paper is that the same single-step fine-tuning recipe could transfer to other restoration tasks, such as dehazing or deraining, where the input provides most of the structure and only low-frequency content is missing.","The paper's ablation suggests the gain of the diffusion prior over a regression-only model is modest in PSNR (0.22 dB on LCDP) but larger in LPIPS (0.0138); if that pattern generalizes, the practical value of the diffusion prior may be mainly perceptual, which is exactly where regression methods are weakest.","Because the diffusion branch is fed a 64x64 input, the method's ceiling may depend on how well the downsampling preserves the low-frequency content of extreme exposure regions; a multi-scale or adaptive downsampling scheme could test whether the bottleneck is information loss or the denoiser itself."],"forward_implications":["Single-step sampling makes diffusion-based exposure correction practical at 0.31 seconds per 512x512 image, competitive with regression methods and far faster than multi-step diffusion baselines.","Because the diffusion branch operates only on low-frequency content at 64x64, artifacts common in fine-tuned diffusion outputs (such as distorted facial features or textures) are largely avoided.","On LCDP, MSEC, and SICE, DPEC reports the best or near-best PSNR, SSIM, LPIPS, and competitive NIQE, including on unlabeled generalization sets such as LIME, DICM, and NPE.","Removing the VAE and text conditioning from Stable Diffusion reduces computational overhead while retaining the benefit of the generative prior."],"supporting_citations":[{"why":"Stable Diffusion, the public pre-trained latent diffusion model whose generative priors DPEC fine-tunes for exposure correction.","marker":"[67]"},{"why":"StableSR, a diffusion-prior restoration method that DPEC compares against and improves upon by adding regression fusion.","marker":"[6]"},{"why":"PASD, another pixel-aware Stable Diffusion fine-tuning approach used as a diffusion-based baseline.","marker":"[7]"},{"why":"Exposure normalization and compensation method, which supplies the regression-loss recipe and the SICE evaluation setup that DPEC follows.","marker":"[53]"},{"why":"CoTF, the strongest regression-based baseline whose reported results DPEC claims to beat on LCDP and MSEC.","marker":"[60]"},{"why":"The MSEC dataset and its multi-scale exposure correction approach, providing one of the three main evaluation benchmarks.","marker":"[3]"},{"why":"The LCDP dataset with local color distribution priors, providing the non-uniform illumination benchmark and a regression baseline.","marker":"[4]"},{"why":"Diff-Retinex, a diffusion-based low-light enhancement baseline that DPEC contrasts with its own use of the generative prior.","marker":"[71]"}],"fun_headline_variants":["One-step diffusion prior corrects exposure with top scores","Diffusion priors make exposure correction one-step and SOTA","DPEC: single-step exposure correction using diffusion priors","Diffusion priors boost exposure fidelity and perceptual quality","Exposure correction: one-step diffusion prior beats baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a U-Net fine-tuned at a single fixed noise timestep and fed a 64x64 downsampled input yields low-frequency content more useful than a simple interpolation baseline, and that fusing these features into a full-resolution regressor improves fidelity rather than introducing hallucinated structure.","fun_headline_variants_meta":{"raw":{"variants":["One-step diffusion prior corrects exposure with top scores","Diffusion priors make exposure correction one-step and SOTA","DPEC: single-step exposure correction using diffusion priors","Diffusion priors boost exposure fidelity and perceptual quality","Exposure correction: one-step diffusion prior beats baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000324,"raw_usage":{"total_tokens":1815,"prompt_tokens":940,"completion_tokens":875,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":795}},"tokens_in":556,"tokens_out":875,"duration_ms":9257,"temperature":1.0,"reasoning_tokens":795,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:26:11.002963+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DPEC on LCDP with the diffusion branch replaced by a trainable bilinear or VAE-based low-frequency generator at the same 64x64 resolution and identical fusion; if PSNR and LPIPS remain within noise of the full model, the generative prior itself is not contributing. Alternatively, feed the diffusion branch a zero image with no content and check whether the one-step output still resembles plausible image structure; if it does, the prior dominates the condition, which would predict failure on inputs whose extreme exposure regions cover most of the frame.","supporting_citations":[{"cited_title":"Exploiting diffusion prior for real-world image super-resolution,","cited_arxiv_id":null,"evidence_quote":"StableSR, a diffusion-prior restoration method that DPEC compares against and improves upon by adding regression fusion."},{"cited_title":"Exposure normalization and compensation for multiple-exposure cor- rection,","cited_arxiv_id":null,"evidence_quote":"Exposure normalization and compensation method, which supplies the regression-loss recipe and the SICE evaluation setup that DPEC follows."},{"cited_title":"Real- time exposure correction via collaborative transformations and adaptive sampling,","cited_arxiv_id":null,"evidence_quote":"CoTF, the strongest regression-based baseline whose reported results DPEC claims to beat on LCDP and MSEC."},{"cited_title":"Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model,","cited_arxiv_id":null,"evidence_quote":"Diff-Retinex, a diffusion-based low-light enhancement baseline that DPEC contrasts with its own use of the generative prior."}],"review_version":1}