{"id":"b5a99dae-94bc-46fd-b657-004354886df5","arxiv_id":"2608.01720","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An extreme low-light RAW enhancement pipeline combining a domain-conditioned tokenizer, one-step MeanFlow transport, and training-only Retinex loss, evaluated on a new synthetic motion-blur benchmark where it reports 6 to 7 dB gains.","lead":"This paper introduces a way to restore smartphone camera images that are both extremely dark and blurred from hand shake, working directly on the raw sensor data. It also releases a new benchmark, SIDED, built by adding synthetic motion blur to an existing low-light dataset, and reports large quality gains over prior methods on that benchmark.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SIDED's synthetic blur-after-noise construction is the load-bearing weakness: the 6–7 dB SOTA margins rest on an absent degradation recipe whose noise coupling may make the benchmark artificially easy.","rationale":"The paper's central claim is that the method achieves state-of-the-art performance in extremely low-light RAW enhancement and robustly handles coupled motion and noise degradation. The SIED results are more modest and are disclosed honestly (best SAM only at one illumination level, Stage III PSNR drop), so the key evidence for the stronger claim is the SIDED benchmark and its 6.06–7.42 dB improvements over the strongest competing method. The reader correctly identifies the realism of SIDED's motion degradation as the weakest assumption. My reading sharpens that concern in one technical direction: the construction appears to blur the RAW measurements after sensor noise is already present, whereas real camera shake during exposure produces a physically different noise—motion-blurred shot noise is correlated and partially smoothed before readout noise is added. Blurring an already noisy image with a smoothing kernel reduces independent high-frequency noise, which can make the deblurring task easier and may also make the input less representative of real extreme-dark handheld capture. This is not an internal inconsistency in the method, so there is no basis for rejection; the paper has independent support from a released code link, coherent ablations, and honest reporting of SIED limitations. But because the SIDED recipe is deferred to a missing supplementary, the headline margins are not yet interpretable. A targeted re-run with physically ordered blur-before-noise degradation would settle whether the reported gains transfer. Since the reader's verdict is already conditional for essentially this reason, no change to the verdict is needed.","tokens_in":11492,"tokens_out":6429,"duration_ms":70869,"concrete_test":"Using the released repository, regenerate an alternate SIDED variant in which motion blur is applied to the clean low-light signal before adding the SIED sensor noise (or, better, simulate real shake by averaging several shifted noisy low-light frames), keeping splits, training steps, and all other protocols identical. Re-run Table 2 with the same seed/reporting policy. If the PSNR advantage over the best competitor falls materially below the reported 6.06–7.42 dB, the blur-after-noise construction is inflating the headline; if the margin persists, the concern is answered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"All headline SIDED margins (6.06/6.84/7.42 dB PSNR) are measured on a synthetic dataset whose construction is not in the preprint: 'Complete generation details are provided in the supplementary material,' and no supplementary is present (Experimental Setup, SIDED construction). The central claim of robustly handling coupled motion and noise therefore rests on an unverifiable degradation recipe. The available description says blur is applied to both Bayer measurements after the SIED noise is already present. That is the wrong physical order: real camera shake blurs the scene during exposure, so shot noise is integrated along the motion trajectory, and read noise is added afterward. Convolving an already noisy RAW with a unit-sum kernel smooths the noise and reduces its high-frequency variance, effectively denoising the input before the model sees it. This can make SIDED easier than real low-light handheld capture and may disproportionately favor the proposed learned deblurring route while hurting comparators that model independent sensor noise. Because the kernel family, sizes, and whether the blur is spatially varying are unspecified, the 6–7 dB margins cannot yet be interpreted as evidence of real-world robustness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SIDED, a synthetic motion-blurred extension of SIED for extremely low-light RAW enhancement, and a framework combining (i) a domain-adaptive RAW tokenizer with dataset-adaptive scaling and Domain-Adaptive Affine Modulation (DAAM), (ii) a one-step 'MeanFlow' enhancement model in latent space, (iii) an optional blind latent deblurring MeanFlow for motion-degraded inputs, and (iv) a training-only Retinex-based refinement. On SIED, the method reports the best PSNR/SSIM at all three illumination levels (26.03/25.73/25.04 dB PSNR). On the proposed SIDED, it reports PSNR gains of 6.06–7.42 dB over compared methods. The core enhancement objective is the MeanFlow velocity target from Geng et al. (2025), applied in the tokenizer's latent space.","tokens_in":11787,"tokens_out":4476,"duration_ms":55177,"significance":"If SIDED's degradation model is physically faithful, the dataset would fill a real gap: existing RAW low-light benchmarks largely ignore accompanying motion blur, and the paper is the first to propose a controlled RAW benchmark for this coupled degradation. The unified tokenizer with DAAM and the one-step latent MeanFlow design are interesting and potentially useful, and the training-only physics guidance is a clean way to add supervision without inference cost. The paper ships a code link, which is a plus. However, the current evidence for the central claim is substantially weakened by the unverifiable construction of SIDED and by the comparison protocol, which gives the proposed method an extra deblurring module unavailable to baselines. The empirical claims therefore need additional support before the conclusions can be accepted.","major_comments":[{"comment":"The central 6.06–7.42 dB PSNR gains on SIDED are measured on a dataset whose generation is described only as applying 'channel-wise motion degradation' to both Bayer measurements, with 'complete generation details are provided in the supplementary material'; the preprint, however, has no supplementary. As described, blur is applied after the SIED noise is already present. This is not the physical order of real camera shake: blur occurs during exposure, so shot noise is integrated along the motion trajectory, and read noise is added afterward. Convolving an already noisy RAW with a unit-sum kernel smooths the noise and reduces its high-frequency variance, which can make the task artificially easy and may disproportionately favor a learned deblurring module. The kernel family, sizes, spatial variability, and exact pipeline must be specified, and the experiments should be repeated with blur","section":"Experimental Setup, SIDED construction; Table 2"},{"comment":"All seven baselines (SID, LLPackNet, RRT, MCR, DNF, RAWMamba, SIED) are extreme-low-light RAW enhancement methods without any deblurring component, while the proposed method, on motion-degraded inputs, additionally activates a dedicated Blind Latent Deblurring MeanFlow (Methodology, 'Blind Latent Deblurring MeanFlow'). The reported 6.06–7.42 dB advantage therefore conflates the advantage of having a deblurring module with the quality of the proposed architecture. A fairer assessment would require RAW-domain joint enhancement+deblurring baselines, an 'Ours without deblurring module' variant, or a standard deblurring module cascaded with a strong enhancement baseline. As it stands, the large SIDED margin is not an apples-to-apples comparison.","section":"Results on SIDED / Compared Methods"},{"comment":"In the motion-degraded route, the output depends on a fresh sample ε∼N(0,I) with no averaging or seed control described (Methodology, 'Inference'). Since the SIDED results in Table 2 are point estimates, the stochasticity of this route may make the reported margins non-reproducible. Please report mean±std over multiple samples or use a fixed seed/averaging scheme, and state the evaluation protocol explicitly.","section":"Inference"}],"minor_comments":[{"comment":"The abstract contains 'focusonillumination' (missing space). Table 2 lists LLPackNet PSNR as '15.5' at 0.0001 lux while other entries have two decimals; this should be '15.50' for consistency.","section":"Abstract / Typos"},{"comment":"The paper repeatedly refers to the supplementary material for SIDED construction details and for 'full objectives and settings' of cascade adaptation, but no supplementary is appended to the preprint. This needs to be included in the submission.","section":"Supplementary references"},{"comment":"The standalone tokenizer ablation reports PSNR >47 dB in rendered RGB space. This is far outside the range of Tables 1–2. Please state explicitly that these numbers are not comparable to the full-pipeline results because no enhancement or deblurring is applied.","section":"Ablation Studies, Table 3"},{"comment":"The paper explains the PSNR drop by arguing that MSE 'intrinsically favors overly smoothed estimates.' This is plausible, but it would be more convincing to add a perceptual metric (e.g., LPIPS) or a small user study to support the claim that the Stage III output is perceptually better despite lower PSNR.","section":"Stage III Ablation, Table 4"},{"comment":"The qualitative figures would benefit from zoomed crops with a magnified inset; the current global views make it hard for readers to verify the stated differences in thin structures, fence bars, and seams.","section":"Figures 2 and 3"}],"recommendation":"major_revision","confidential_remarks":"The SIDED benchmark is introduced and evaluated by the same group, and the generation details are deferred to a supplementary that is absent from the preprint. Given that the headline margins all depend on this synthetic dataset, the editor may wish to require the supplementary material and the code (including the SIDED generation script) as part of the revision, and to ask for an independent check of the blur-before-noise versus blur-after-noise synthesis. The comparison issue is also structural: the proposed method receives an extra deblurring module, so the current SIDED evaluation is not a clean test of the claimed unified framework."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful adaptation of MeanFlow to extreme low-light RAW enhancement, with a clean one-step formulation and credible SIED gains. But the headline SIDED benchmark is the weak link: the construction defers all details to a missing supplementary, and what the paper does say is physically suspect.\n\nWhat's new and good: the unified tokenizer with dataset-adaptive scaling and DAAM is a sensible way to align the low-light and GT latents, and the ablation (Table 3) shows a clear 2.5 dB benefit. The training-only Retinex guidance is a reasonable way to avoid inference-time decomposition errors, and the paper honestly reports that Stage III lowers PSNR while improving SSIM/SAM. On SIED, the method beats the existing baseline by 1.1–1.9 dB PSNR across illumination levels; those results are on a third-party benchmark and are credible. The abstract's blanket SOTA claim is a bit strong—SAM is worse than SIED at two levels—but the paper itself admits this in the discussion.\n\nThe soft spot is SIDED. The paper says it applies channel-wise motion degradation to both Bayer measurements preserving original noise. That means blur is applied after the sensor noise is already in place. Real camera shake blurs during the exposure, so shot noise is integrated along the motion trajectory and read noise is added afterward. Convolving an already noisy RAW with a unit-sum kernel smooths the noise and lowers its high-frequency variance—effectively denoising the input before the model sees it. That makes SIDED easier than real handheld low-light capture and may disproportionately favor a learned deblurring route. None of the generation details (kernel family, size, spatial variance) are in the preprint; they are all in a supplementary that is absent. So the 6.06–7.42 dB margins on SIDED cannot be interpreted as evidence of real-world robustness. Also, no error bars or seeds are reported, and the deblurring route samples Gaussian noise at inference without stating the sampling policy or variance of the outputs.\n\nWho this is for: if you work in low-light RAW restoration, the SIED result and the MeanFlow formulation are worth studying. The SIDED dataset needs a much more careful description and ideally a real-capture validation before being used as a benchmark.\n\nRecommendation: send it to peer review. The core SIED contribution is solid, and the issues with SIDED are addressable in revision if the authors provide the full construction and correct the physics.","headline":"Worth a serious referee for the SIED results and the MeanFlow adaptation, but the SIDED benchmark—and its 6–7 dB margins—cannot be evaluated until the missing blur recipe appears and the physics are corrected.","tokens_in":12326,"tokens_out":2843,"would_cite":false,"duration_ms":27841,"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 claims that extremely low-light RAW restoration should treat motion blur as a first-class degradation, and that a one-step latent MeanFlow with a domain-conditioned tokenizer beats prior methods on clean and motion-degraded extrem","keywords":["extremely low-light RAW enhancement","motion deblurring","MeanFlow","one-step generative model","domain-adaptive tokenizer","Retinex guidance","RAW restoration","SIDED dataset"],"falsifier":"Collect real paired sharp/blurry extremely-dark RAW images (e.g., tripod reference plus handheld shots, or IMU-measured shake) and compare the released model against the strongest baselines. A drop from the reported 6–7 dB margin to near parity would falsify the transfer claim; alternatively, re-running SIDED with several realistic measured blur kernels and checking whether the gains persist would test the benchmark's difficulty.","tokens_in":11339,"feed_emoji":"📷","tokens_out":7893,"duration_ms":78452,"temperature":0.7,"pith_summary":"Extremely low-light RAW enhancement usually corrects exposure and noise but ignores the camera shake that naturally accompanies handheld dark shots. This paper claims that motion blur should be treated as a first-class degradation and that a one-step latent-flow pipeline can restore both darkness and blur directly from sensor data. To test this, it builds SIDED, a motion-degraded version of the SIED dataset, and reports PSNR gains of 6.06 to 7.42 dB over the strongest compared methods there, along with leading PSNR and SSIM on the clean SIED benchmark. If the claim holds, the framework makes practical handheld extremely-dark RAW restoration feasible in a single network evaluation.","feed_headline":"One-step latent flow beats dark RAW motion-blur rivals by 7.42 dB","feed_subtitle":"Extreme-darkness RAW enhancement usually ignores camera shake; this framework handles both in one evaluation.","key_machinery":"The framework rests on three coupled pieces. A unified RAW tokenizer with domain-adaptive affine modulation (DAAM) maps scaled low-light and well-exposed RAW measurements into one shared latent space, using zero-initialized per-domain scale and shift so pretrained mappings survive. The task-aware MeanFlow group—MeanFlow being a one-step generative model that fits an average velocity field between distributions—transports low-light latents to well-exposed latents in one function evaluation, and optionally runs a blind latent deblurring MeanFlow, initialized on GT-domain motion pairs and cascade-adapted to the enhancement output, without needing a blur kernel at inference. A frozen Retinex net","core_discovery":"The paper's central discovery is that the extreme-dark-to-well-exposed RAW transformation can be carried out in a shared latent space by a one-step MeanFlow, and that the same tokenizer supports a separate blind-latent-deblurring route when motion degradation is present. On the Sony subset of SIED the method reports PSNR 26.03/25.73/25.04 dB at 0.01/0.001/0.0001 lux, the best among compared methods; on SIDED it reports 24.62/24.65/24.23 dB, 6.06–7.42 dB above the strongest competitor. A training-only Retinex-guided refinement stage improves SSIM and SAM at a small PSNR cost, and the paper shows that the motion-aware route remains stable as illumination drops.","pith_inferences":["Inference: if SIDED's synthetic blur approximates real handheld shake, the same latent route should transfer to real blurry RAW captures; a field test with gyro-measured or tripod-vs-handheld pairs would verify this.","Inference: because blur is applied to already-noisy Bayer values, the deblurring network may be exploiting noise correlations specific to the synthetic pipeline, so the 6–7 dB margin could shrink on true noise-motion coupling.","Inference: the DAAM domain-calibration idea is not limited to darkness and could be reused for other RAW domain gaps, such as sensor, ISO, or exposure differences, within one tokenizer.","Inference: the Stage III PSNR-versus-perception trade-off suggests that for motion-degraded extreme darkness, structural and color-direction metrics may be more informative than pixel error alone."],"forward_implications":["Extreme low-light enhancement can be performed in a single function evaluation instead of iterative diffusion, while still modeling the full dark-to-light distribution.","Motion blur and extreme darkness can be handled jointly in the RAW domain, preserving sensor noise statistics rather than pre-ISP information.","SIDED provides a controlled benchmark for motion-degraded extreme low-light RAW restoration, enabling direct comparison of future methods in this setting.","Because the Retinex guidance is training-only, the improved structure and color fidelity it brings costs no extra inference time.","The widening PSNR advantage as illumination falls suggests that exposure-and-noise-only methods are most vulnerable exactly when blur removes scarce structural cues."],"supporting_citations":[{"why":"SID dataset and baseline; defines extreme low-light RAW enhancement and is the primary comparison point.","marker":"Chen et al. 2018"},{"why":"SIED dataset; supplies the paired extreme-low-light RAW samples that SIDED extends with motion degradation.","marker":"Jiang et al. 2025a"},{"why":"MeanFlow one-step generative modeling; provides the average-velocity-field mechanism used for enhancement and deblurring.","marker":"Geng et al. 2025"},{"why":"Retinexformer; provides the Retinex decomposition prior used as training-only physics guidance.","marker":"Cai et al. 2023"},{"why":"DNF; a strong RAW enhancement baseline that the framework must outperform on SIED and SIDED.","marker":"Jin et al. 2023"},{"why":"Retinex-RAWMamba; a recent RAW enhancement baseline compared in the experiments.","marker":"Chen et al. 2026"},{"why":"LEDNet; joint low-light enhancement and deblurring in RGB, the gap the paper targets by moving to RAW.","marker":"Zhou, Li, and Loy 2022"}],"fun_headline_variants":["One-step MeanFlow restores dark RAW under motion blur, up to 7.42 dB","Extreme darkness plus camera shake? MeanFlow does both in one pass","Unified RAW tokenizer and MeanFlow beat motion-degraded dark rivals","Dark RAW enhancement handles motion blur with single-step latent flow","MeanFlow: one step to recover extreme-low-light RAW with blur"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that SIDED's controlled, channel-wise motion blur applied to noisy Bayer RAW values faithfully mimics real handheld camera shake in extreme darkness; if that realism fails, the reported margins may not transfer to real captures.","fun_headline_variants_meta":{"raw":{"variants":["One-step MeanFlow restores dark RAW under motion blur, up to 7.42 dB","Extreme darkness plus camera shake? MeanFlow does both in one pass","Unified RAW tokenizer and MeanFlow beat motion-degraded dark rivals","Dark RAW enhancement handles motion blur with single-step latent flow","MeanFlow: one step to recover extreme-low-light RAW with blur"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1543,"prompt_tokens":761,"completion_tokens":782,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":684}},"tokens_in":505,"tokens_out":782,"duration_ms":8828,"temperature":1.0,"reasoning_tokens":684,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T22:07:17.758988+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect real paired sharp/blurry extremely-dark RAW images (e.g., tripod reference plus handheld shots, or IMU-measured shake) and compare the released model against the strongest baselines. A drop from the reported 6–7 dB margin to near parity would falsify the transfer claim; alternatively, re-running SIDED with several realistic measured blur kernels and checking whether the gains persist would test the benchmark's difficulty.","supporting_citations":[{"cited_title":"2018 , eprint =","cited_arxiv_id":null,"evidence_quote":"SID dataset and baseline; defines extreme low-light RAW enhancement and is the primary comparison point."},{"cited_title":"Advances in Neural Information Processing Systems , volume =","cited_arxiv_id":null,"evidence_quote":"MeanFlow one-step generative modeling; provides the average-velocity-field mechanism used for enhancement and deblurring."},{"cited_title":"Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =","cited_arxiv_id":null,"evidence_quote":"Retinexformer; provides the Retinex decomposition prior used as training-only physics guidance."},{"cited_title":"IEEE Transactions on Circuits and Systems for Video Technology , volume =","cited_arxiv_id":null,"evidence_quote":"Retinex-RAWMamba; a recent RAW enhancement baseline compared in the experiments."},{"cited_title":"Computer Vision -- ECCV 2022 , year =","cited_arxiv_id":null,"evidence_quote":"LEDNet; joint low-light enhancement and deblurring in RGB, the gap the paper targets by moving to RAW."}],"review_version":1}