{"id":"b0690ab7-087f-4b52-bc22-e6abfe17d45d","arxiv_id":"2509.00917","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DarkVRAI combines capture-condition conditioning and a burst-order selective scan to win the AIM 2025 low-light RAW video denoising challenge.","lead":"This paper describes DarkVRAI, a deep learning model for cleaning up dark, grainy smartphone video that won first place in the 2025 AIM low-light RAW video denoising challenge. It works by feeding camera settings into the network and scanning frames in order to borrow details across time.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The +0.31 dB BOSS gain in Table 2 is confounded by parameter count and missing seed/error-bar controls; the paper's causal attribution of SOTA to BOSS is not yet established.","rationale":"The reader's conditional verdict already identifies the capacity/seed confound in the BOSS ablation; I agree that this is the most load-bearing concern about the paper's causal claims. The first-place result itself is credible because it comes from an official CodaBench leaderboard, so I would not move to reject. The soft spot is that Table 2 is the only evidence that BOSS specifically contributes, and the comparison does not control for parameter count, training budget, or stochasticity. This is testable: the challenge's validation set is fixed, and the training schedule is described precisely enough to rerun with a capacity-matched control module. If the control reproduces the +0.31 dB, then BOSS is not the mechanism; if it does not, the authors' attribution would be strengthened. The paper's lack of code and error bars is a related reproducibility limitation, but the capacity-matched ablation is the most direct check and does not require the private test set. Therefore the appropriate verdict remains conditional, with the condition that the authors provide either code or the controlled ablation before the mechanism claim is accepted.","tokens_in":6548,"tokens_out":5712,"duration_ms":77234,"concrete_test":"Run a capacity-matched ablation on the validation set: take Model A and add a control temporal module with the same parameter count and FLOPs as the BOSS block (e.g., a temporal depthwise convolution plus gating) in the same insertion points. Train Baseline, Model A, Model A+control, and DarkVRAI with identical iterations/learning rate and at least 3 random seeds; report mean±std PSNR and per-sequence paired differences. If Model A+control matches DarkVRAI (47.16 dB) within 0.1 dB, the BOSS-specific design is not needed to explain the gain. Also compute a paired t-test between Model A and DarkVRAI if per-sequence validation outputs can be obtained.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DarkVRAI is a state-of-the-art low-light RAW video denoiser and that C3 + BOSS drive the gain. The leaderboard result is externally supported by the official challenge report, so I do not dispute the first-place ranking. The load-bearing weakness is the internal validation of the BOSS contribution. Table 2 compares Baseline, Model A (C3), and DarkVRAI (C3+BOSS). The jump from Model A to DarkVRAI is +0.31 dB PSNR, but DarkVRAI has extra BOSS blocks with more parameters and no capacity-matched control. The paper reports no seeds, no standard deviations, and no validation-set size, so 0.31 dB may be within training/sampling noise. Also, Eqs. (1)-(2) and Fig. 2 never specify the selective-scan state-space operation; without that, the BOSS module could be functioning as a generic gated temporal aggregator, so even the mechanism-level attribution is unclear. This does not invalidate the challenge result, but it leaves the paper's second main contribution unsupported as the cause of the reported improvement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DarkVRAI is a two-stage framework for low-light RAW video denoising, described as the first-place solution to the AIM 2025 Low-light RAW Video Denoising Challenge. It combines a Capture Condition Conditioning (C3) scheme, which injects sensor, illuminance, and frame-rate metadata via adaptive layer normalization into both the alignment and denoising stages, with a Burst-Order Selective Scan (BOSS) mechanism intended to capture long-range temporal dependencies. On the challenge private test set the method reports 48.32 dB PSNR and 0.9879 SSIM (Table 1). A validation-set ablation (Table 2) reports +0.46 dB for C3 and an additional +0.31 dB for BOSS.","tokens_in":6857,"tokens_out":4108,"duration_ms":52715,"significance":"If the results are taken at face value, DarkVRAI is a strong practical result: it wins a competitive multi-sensor benchmark, and the C3 extension from image denoising to video is a plausible and potentially useful idea that receives independent validation in the ablation. The leaderboard numbers are externally sourced from the official challenge report, which is a concrete strength. However, the paper's second contribution, BOSS, is not sufficiently specified or causally validated: the operation is left as a black-box 'BOSS' call in Eq. (2), and the +0.31 dB ablation gain is not controlled for parameter count, seeds, or training noise. The paper also does not provide code or pretrained models, which limits reproducibility beyond the architectural description. Overall this is a promising challenge-report-style manuscript whose central technical claim about BOSS needs additional evidence and precision.","major_comments":[{"comment":"The +0.31 dB PSNR gain attributed to BOSS (Model A -> DarkVRAI) is not causally identified. DarkVRAI adds BOSS blocks and therefore has more parameters than Model A, yet no capacity-matched control is provided. The manuscript also reports no standard deviations, no number of seeds, no validation-set size, and does not state whether the baseline, Model A, and DarkVRAI use the same training budget and random seeds. With single-run validation, a 0.31 dB difference may be within training/sampling noise. Please provide multi-seed means and standard deviations, a capacity-matched baseline, or otherwise control for these factors before claiming that BOSS is the cause of the improvement.","section":"§3.3, Table 2"},{"comment":"The BOSS mechanism is never defined mathematically. Eq. (2) writes xboss = BOSS(phi(Conv2d(x))) but the 'BOSS' operation itself is not specified: there is no state-space recurrence, state dimension, scan order beyond 'natural temporal order', or selectivity rule, and Fig. 2 only labels 'scan & sum'. Consequently, the claim that BOSS 'leverages principles from state-space models to progressively aggregate long-range temporal information' is not verifiable; any temporal pooling/gating operation would satisfy the current description. Please give the explicit operation or substantially soften the SSM claim.","section":"§2.3, Eq. (2), Fig. 2"}],"minor_comments":[{"comment":"The text says 'standard variation'; this should be 'standard deviation'.","section":"Eq. (1)"},{"comment":"'hierachical' should be 'hierarchical'.","section":"Fig. 2 caption"},{"comment":"The title displays 'RA W' (broken LaTeX spacing) and §3.3 uses 'DARKVRAI' in all caps while the rest of the paper uses 'DarkVRAI'. Please unify.","section":"Title and §3.3"},{"comment":"The entry 'Peng [15]' appears to cite the NAFNet paper by Chen et al. as [15], but no separate reference to a challenge participant named Peng is listed. Please correct the citation or add the missing reference.","section":"Table 1 and §3.2"},{"comment":"The notation xz and the split Linear(Xin) = [x∥z] are not explained; please define the split and the role of each branch before the BOSS call.","section":"Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The external leaderboard result is solid and likely sufficient for a challenge-report venue. For a journal-level archival claim, however, the BOSS contribution needs a concrete mathematical definition and a controlled ablation; as written, the paper risks overclaiming. I would also suggest checking whether the manuscript is long enough for the journal's scope, since it reads as an extended workshop paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the thing you should know: this is a genuine first-place challenge solution, and the leaderboard result (48.32 dB PSNR, 0.9879 SSIM on the private test set) is externally confirmed in the AIM 2025 report. The paper is a clear, honest write-up of a winning pipeline, and the extension of camera-parameter conditioning from images to video is a legitimate and practical step. The BOSS block is a reasonable way to add temporal context within a Mamba-like gated scan, and the authors show a consistent +0.77 dB over their own baseline on validation. That is real work and worth a referee's time.\n\nWhere it gets soft: the ablation isolates C3 and BOSS, but the BOSS step is not capacity-controlled. DarkVRAI with BOSS has additional parameters, and there are no seeds, standard deviations, or a stated validation-set size. A +0.31 dB increment over Model A could sit inside training noise. The stress-test note is on target. Also, Eq. (2) and Fig. 2 never define the scan operation itself; what we see is a gated linear unit with a LayerNorm, and 'scan & sum' is depicted but not formalized. So the mechanism-level claim that BOSS is doing state-space modeling is unsupported. It may be working as a generic temporal aggregator, which is fine, but the paper says more than it shows.\n\nThe central challenge result survives these concerns, since it comes from an independent leaderboard. The unsupported part is the internal attribution of the gain to BOSS. That is a fixable weakness, not a fraud. The paper also leans on the authors' prior C3 paper [6]; that is normal in a sequence of work, and the validation here is independent.\n\nWho gets value: someone working on low-light video denoising or challenge-benchmark methods will want to read this for the architecture and the benchmark numbers. For a rigorous research claim about BOSS, wait for a version with controlled ablations and code.\n\nMy recommendation: yes, send it to peer review. It deserves a serious referee, with the expectation that the authors either supply a capacity-matched BOSS ablation and error bars or soften the causal claim. I'd take it to reading group as an example of a credible challenge report.","headline":"Genuine first-place challenge result with a solid write-up; the BOSS contribution is real but not yet causally established.","tokens_in":7308,"tokens_out":2062,"would_cite":true,"duration_ms":24535,"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":"DarkVRAI claims that injecting capture metadata into both alignment and denoising and scanning burst frames in temporal order lifts low-light RAW video denoising to 48.32 dB PSNR, the top score on the challenge's private test set.","keywords":["low-light video denoising","RAW video restoration","capture-condition conditioning","selective state-space scan","temporal fusion","burst denoising","metadata-guided processing","video benchmark"],"falsifier":"Retrain DarkVRAI and the C3-only variant with matched parameter counts (for example, widen the baseline or remove BOSS while adding equivalent convolutional capacity) on the same validation split, and check whether the +0.31 dB gap persists.","tokens_in":6480,"feed_emoji":"🎥","tokens_out":7399,"duration_ms":76811,"temperature":0.7,"pith_summary":"DarkVRAI is a two-stage pipeline that takes ten consecutive noisy RAW frames and produces one clean frame. Its central claim is that explicitly feeding the network capture metadata—sensor identity, illuminance, and frame rate—lets both the alignment and denoising stages adapt to each camera's noise profile, and that a burst-order selective scan aggregates temporal context more effectively than simple concatenation or attention. On the private test set the full model reaches 48.32 dB PSNR and 0.9879 SSIM, first among submitted methods, with multi-frame inputs beating the best single-frame method by about 1.8 dB. Ablations attribute +0.46 dB to the metadata conditioning and +0.31 dB to the selective scan.","feed_headline":"First-place low-light RAW video denoiser hits 48.32 dB PSNR","feed_subtitle":"Fusing camera metadata with burst-order scanning reaches 48.32 dB PSNR on a 14-sensor low-light video set.","key_machinery":"Capture Condition Conditioning (C3): a metadata embedding built from one-hot sensor, illuminance, and frame rate, projected into per-channel scale and bias applied through adaptive layer normalization, tailoring every layer's feature statistics to the degradation context. Burst-Order Selective Scan (BOSS): a state-space-style scan over the ten frame features in their natural temporal order, placed before each encoder and alignment block so that alignment operates on temporally enriched features. These two mechanisms sit inside a two-stage backbone: a burst-alignment stage followed by a convolutional U-shaped denoising stage.","core_discovery":"The paper establishes that low-light RAW video denoising can be treated as a condition-aware temporal fusion problem. It takes ten noisy RAW frames plus the capture conditions under which they were shot, encodes sensor, illuminance, and frame rate into an embedding, and modulates feature statistics throughout the network via adaptive layer normalization. In parallel, it inserts Burst-Order Selective Scan blocks before each alignment and encoder stage so that features are scanned in temporal order and enriched with long-range context before motion alignment happens. The combination is the winning solution of the 2025 low-light RAW video denoising benchmark, achieving 48.32 dB PSNR and 0.9879","pith_inferences":["If the BOSS gain survives capacity-matched retraining, the 'scan before alignment' design is a transferable recipe for other multi-frame tasks such as burst HDR, video super-resolution, or spatio-temporal restoration with large inter-frame motion.","Because conditioning relies on one-hot metadata, a natural extension is replacing the one-hot encoder with a learned continuous estimator of illuminance and gain, which could generalize to unseen sensors without explicit metadata at test time.","A natural next check is whether the first-place margin holds when entire sensor types, rather than one scene, are held out from training.","The same conditioning mechanism could be tested for robustness by corrupting or omitting the metadata at inference time, revealing how much of the gain depends on perfect metadata availability."],"forward_implications":["Explicit capture metadata improves video denoising: adding C3 to a blind baseline yields +0.46 dB PSNR on the validation set.","The BOSS scan adds a further +0.31 dB PSNR on top of C3, bringing the total gain over the blind baseline to +0.77 dB.","Temporal fusion is the dominant factor: the best multi-frame method beats the best single-frame method by about 1.8 dB PSNR on the private test set.","Combining metadata conditioning with burst-order scanning yields the leading private-test scores of 48.32 dB PSNR and 0.9879 SSIM, ahead of all submitted methods.","The framework is reproducible from the described architecture and training recipe: 300k iterations, 256x256 patches, Adam with cosine schedule, and L1 plus MS-SSIM loss."],"supporting_citations":[{"why":"Defines the challenge dataset, the private test protocol, and the leaderboard where the first-place result is established.","marker":"[5]"},{"why":"Source of the capture-parameter conditioning scheme that the paper extends from single-image denoising to video.","marker":"[6]"},{"why":"Provides the selective state-space scanning principle that BOSS adapts for temporal aggregation.","marker":"[10]"},{"why":"Supplies the burst-alignment backbone that the alignment stage is built on.","marker":"[13]"},{"why":"Provides the U-shaped denoising structure and the single-frame UNet baseline.","marker":"[14]"},{"why":"Supplies the NAFBlocks used as convolutional restoration blocks and a single-frame baseline method.","marker":"[15]"},{"why":"Supplies the multi-frame transformer baseline compared in the reported results.","marker":"[21]"}],"fun_headline_variants":["Condition-aware temporal fusion wins low-light RAW denoising","Metadata-guided denoising tops low-light RAW video benchmark","Burst-order scan + camera metadata: 48.32 dB PSNR win","DarkVRAI: low-light RAW video denoising champion with BOSS","Winning low-light denoiser fuses capture conditions with burst scanning"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The BOSS module's measured gain is attributed to its scan design, but the comparison adds parameters without controlling for model capacity or training budget; if extra capacity alone explains the improvement, the scan mechanism is not the cause.","fun_headline_variants_meta":{"raw":{"variants":["Condition-aware temporal fusion wins low-light RAW denoising","Metadata-guided denoising tops low-light RAW video benchmark","Burst-order scan + camera metadata: 48.32 dB PSNR win","DarkVRAI: low-light RAW video denoising champion with BOSS","Winning low-light denoiser fuses capture conditions with burst scanning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000546,"raw_usage":{"total_tokens":2423,"prompt_tokens":696,"completion_tokens":1727,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":440,"completion_tokens_details":{"reasoning_tokens":1632}},"tokens_in":440,"tokens_out":1727,"duration_ms":14803,"temperature":1.0,"reasoning_tokens":1632,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:03:42.587935+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain DarkVRAI and the C3-only variant with matched parameter counts (for example, widen the baseline or remove BOSS while adding equivalent convolutional capacity) on the same validation split, and check whether the +0.31 dB gap persists.","supporting_citations":[{"cited_title":"A physics-based noise formation model for extreme low-light raw denoising,","cited_arxiv_id":null,"evidence_quote":"Defines the challenge dataset, the private test protocol, and the leaderboard where the first-place result is established."},{"cited_title":"Toward efficient deep blind raw image restoration,","cited_arxiv_id":null,"evidence_quote":"Source of the capture-parameter conditioning scheme that the paper extends from single-image denoising to video."},{"cited_title":"Towards Controllable Real Image Denoising with Camera Parameters","cited_arxiv_id":"2507.01587","evidence_quote":"Provides the selective state-space scanning principle that BOSS adapts for temporal aggregation."},{"cited_title":"Cbam: Convolutional block attention module,","cited_arxiv_id":null,"evidence_quote":"Supplies the NAFBlocks used as convolutional restoration blocks and a single-frame baseline method."}],"review_version":1}