Pith. sign in

REVIEW 2 major objections 5 minor 25 references

DarkVRAI: Capture-Condition Conditioning and Burst-Order Selective Scan for Low-light RAW Video Denoising

T0 review · 2 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read 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.

desk verdict Genuine first-place challenge result with a solid write-up; the BOSS contribution is real but not yet causally established. read the letter →

arxiv 2509.00917 v1 pith:NX4TBQAB submitted 2025-08-31 cs.CV

classification cs.CV
keywords low-lightvideodenoisingRAWrestorationcapture-conditionconditioningselectivestate-spacescantemporalfusionburstmetadata-guidedprocessingbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

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.

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 (2)
  1. [§3.3, Table 2] 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.
  2. [§2.3, Eq. (2), Fig. 2] 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.
minor comments (5)
  1. [Eq. (1)] The text says 'standard variation'; this should be 'standard deviation'.
  2. [Fig. 2 caption] 'hierachical' should be 'hierarchical'.
  3. [Title and §3.3] 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.
  4. [Table 1 and §3.2] 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.
  5. [Eq. (2)] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: C3 and BOSS are empirical components validated by external benchmark and internal ablation; the sole self-citation is not load-bearing.

full rationale

The paper's central claim is a first-place result on the AIM 2025 Low-light RAW Video Denoising Challenge private test set (Table 1), with scores sourced from the official leaderboard. The two proposed components, C3 and BOSS, are introduced as architectural mechanisms, not as quantities derived from the target metric. C3 is inspired by the authors' prior work [6], but in this paper it is implemented in a new video-denosing pipeline and its contribution is measured by an ablation (Table 2), not deduced from the conditioning equations. BOSS is specified by Eq. (2) and its gain is likewise established empirically by comparing Model A and DarkVRAI. No equation reduces the reported PSNR/SSIM values to a fitted parameter, to the conditioning metadata itself, or to a self-cited uniqueness claim. The only self-citation is the origin of the conditioning idea; the present validation is independent of whether that prior paper is correct. The ablation's lack of capacity matching, seeds, and error bars is a rigor concern about causal attribution, not a circularity. Therefore the derivation chain is self-contained for circularity purposes.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a set of hand-chosen architectural hyperparameters and on the trustworthiness of the challenge benchmark. No independent theoretical principle fixes the architecture; the gains are empirical.

free parameters (5)
  • Number of input frames T = 10
    Set to balance temporal context against memory and compute; not justified by analysis.
  • Initial feature channels C = 48
    Architectural width chosen by hand; no sensitivity study is provided.
  • Encoder/decoder/bottleneck block counts = 4/4/8
    Depth configuration of the U-shaped denoiser, stated in Section 2.4; not ablated.
  • Training hyperparameters = patch 256, stride 192, batch 4, 300k iterations, LR 2e-4 cosine decay
    Chosen for the challenge; no sensitivity analysis is reported.
  • BOSS block placement = before each encoder and alignment block
    A placement choice described qualitatively in Section 2.3; exact number and locations are not specified precisely.
assumptions (4)
  • domain assumption The challenge dataset and private test set are representative of low-light RAW video noise.
    The paper relies on the benchmark (ref [5]) as the sole grounding for its SOTA claim; no external validation is provided.
  • domain assumption Training on the provided challenge data generalizes to the private test set.
    The model is trained only on the public part of the dataset and evaluated on an undisclosed test set; no distribution analysis is given.
  • standard math The building blocks (EDA, NAFBlocks, CBAM, selective scan) perform as claimed in their original papers.
    The framework is assembled from cited modules; the paper does not re-derive or verify their properties.
  • domain assumption Capture metadata (sensor, illuminance, frame rate) is available and accurate at inference.
    C3 conditioning requires this metadata; the paper does not evaluate robustness to missing or noisy metadata.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DarkVRAI: Capture-Condition Conditioning and Burst-Order Selective Scan for Low-light RAW Video Denoising." pith.science (2026). https://pith.science/paper/NX4TBQAB

@misc{pith2026250900917,
  author       = {Pith},
  title        = {Pith review of: DarkVRAI: Capture-Condition Conditioning and Burst-Order Selective Scan for Low-light RAW Video Denoising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NX4TBQAB}},
  note         = {Machine review of arXiv:2509.00917}
}
read the original abstract

Low-light RAW video denoising is a fundamentally challenging task due to severe signal degradation caused by high sensor gain and short exposure times, which are inherently limited by video frame rate requirements. To address this, we propose DarkVRAI, a novel framework that achieved first place in the AIM 2025 Low-light RAW Video Denoising Challenge. Our method introduces two primary contributions: (1) a successful application of a conditioning scheme for image denoising, which explicitly leverages capture metadata, to video denoising to guide the alignment and denoising processes, and (2) a Burst-Order Selective Scan (BOSS) mechanism that effectively models long-range temporal dependencies within the noisy video sequence. By synergistically combining these components, DarkVRAI demonstrates state-of-the-art performance on a rigorous and realistic benchmark dataset, setting a new standard for low-light video denoising.

Figures

Figures reproduced from arXiv: 2509.00917 by the authors.

Figure 1
Figure 1. The framework accepts a sequence of 10 consecutive [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 19 canonical work pages

  1. [6]

    Toward efficient deep blind raw image restoration,

    Marcos V Conde, Florin Vasluianu, and Radu Timofte, “Toward efficient deep blind raw image restoration,” in 2024 IEEE International Conference on Image Process- ing (ICIP). IEEE, 2024, pp. 1725–1731

  2. [1]

    DarkVRAI: Capture-Condition Conditioning and Burst-Order Selective Scan for Low-light RAW Video Denoising

    INTRODUCTION The ability to capture high-quality video in low-light con- ditions is critical for a wide range of applications, from consumer photography to autonomous systems. However, the physics of image capture imposes significant constraints. Video acquisition requires maintaining a target frame rate (e.g., 24 to 120 fps), which places a strict upper ...

  3. [2]

    Overall Framework DarkVRAI is designed as a two-stage pipeline consisting of frame alignment, followed by denoising, as illustrated in Fig

    METHOD 2.1. Overall Framework DarkVRAI is designed as a two-stage pipeline consisting of frame alignment, followed by denoising, as illustrated in Fig. 1. The framework accepts a sequence of 10 consecutive noisy RAW frames as input and produces a single, clean RAW frame corresponding to the final frame in the sequence. Both stages are modulated by our Cap...

  4. [3]

    Experimental Setup Dataset

    EXPERIMENTS 3.1. Experimental Setup Dataset. Experiments are conducted using the AIM 2025 Low-light RAW Video Denoising Challenge dataset [5]. The Table 1. Quantitative results on the AIM 2025 Low-light RAW Video Denoising Challenge private test set. Data is sourced from the official challenge leaderboard. Method Type PSNR↑ SSIM↑ Noisy - 36.06 0.8093 UNet...

  5. [4]

    We have addressed the fundamental difficulties of this task—severe noise, motion, and sensor-specific degra- dations—through a novel two-stage framework

    CONCLUSION This paper introduced DarkVRAI, our winning solution to the AIM 2025 Low-light RAW Video Denoising Chal- lenge. We have addressed the fundamental difficulties of this task—severe noise, motion, and sensor-specific degra- dations—through a novel two-stage framework. The key innovations of our method are the synergistic use of the C 3 strategy fo...

  6. [5]

    A physics-based noise formation model for extreme low-light raw denoising,

    Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang, “A physics-based noise formation model for extreme low-light raw denoising,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 2758–2767

  7. [7]

    A high-quality denoising dataset for smart- phone cameras,

    Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown, “A high-quality denoising dataset for smart- phone cameras,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 1692–1700

  8. [8]

    Benchmarking denois- ing algorithms with real photographs,

    Tobias Plotz and Stefan Roth, “Benchmarking denois- ing algorithms with real photographs,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1586–1595

Show all 25 references
  1. [9]

    AIM 2025 low-light raw video denois- ing challenge: Dataset, methods and results,

    Alexander Yakovenko, George Chakvetadze, Ilya Khrapov, Maksim Zhelezov, Dmitry Vatolin, Radu Tim- ofte, et al., “AIM 2025 low-light raw video denois- ing challenge: Dataset, methods and results,” in Pro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICC...

  2. [10]

    Towards controllable real image de- noising with camera parameters,

    Youngjin Oh, Junhyeong Kwon, Keuntek Lee, and Nam Ik Cho, “Towards controllable real image de- noising with camera parameters,” arXiv preprint arXiv:2507.01587, 2025

  3. [11]

    Scalable diffu- sion models with transformers,

    William Peebles and Saining Xie, “Scalable diffu- sion models with transformers,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4195–4205

  4. [12]

    Layer normalization,

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton, “Layer normalization,” arXiv preprint arXiv:1607.06450, 2016

  5. [13]

    Attention is all you need,

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin, “Attention is all you need,” Ad- vances in neural information processing systems , vol. 30, 2017

  6. [14]

    Mamba: Linear-time sequence modeling with selective state spaces,

    Albert Gu and Tri Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752, 2023

  7. [15]

    Cbam: Convolutional block attention module,

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon, “Cbam: Convolutional block attention module,” in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 3–19

  8. [16]

    Mambair: A simple baseline for image restoration with state-space model,

    Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia, “Mambair: A simple baseline for image restoration with state-space model,” in European conference on computer vision. Springer, 2024, pp. 222– 241

  9. [17]

    Burstormer: Burst image restoration and enhancement transformer,

    Akshay Dudhane, Syed Waqas Zamir, Salman Khan, Fahad Shahbaz Khan, and Ming-Hsuan Yang, “Burstormer: Burst image restoration and enhancement transformer,” in 2023 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) . IEEE, 2023, pp. 5703–5712

  10. [18]

    U-net: Convolutional networks for biomedical image segmentation,

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Med- ical image computing and computer-assisted interven- tion. Springer, 2015, pp. 234–241

  11. [19]

    Simple baselines for image restoration,

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun, “Simple baselines for image restoration,” in Eu- ropean conference on computer vision . Springer, 2022, pp. 17–33

  12. [20]

    Codabench: Flexible, easy-to-use, and reproducible meta-benchmark platform,

    Zhen Xu, Sergio Escalera, Adrien Pav ˜ao, Magali Richard, Wei-Wei Tu, Quanming Yao, Huan Zhao, and Isabelle Guyon, “Codabench: Flexible, easy-to-use, and reproducible meta-benchmark platform,” Patterns, vol. 3, no. 7, 2022

  13. [21]

    Adam: A method for stochastic optimization,

    Diederik P Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

  14. [22]

    Sgdr: Stochastic gradient descent with warm restarts,

    Ilya Loshchilov and Frank Hutter, “Sgdr: Stochastic gradient descent with warm restarts,” arXiv preprint arXiv:1608.03983, 2016

  15. [23]

    Multiscale structural similarity for image quality as- sessment,

    Zhou Wang, Eero P Simoncelli, and Alan C Bovik, “Multiscale structural similarity for image quality as- sessment,” in The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003 . Ieee, 2003, vol. 2, pp. 1398–1402

  16. [24]

    Yond: Practical blind raw image denoising free from camera-specific data dependency,

    Hansen Feng, Lizhi Wang, Yiqi Huang, Tong Li, Lin Zhu, and Hua Huang, “Yond: Practical blind raw image denoising free from camera-specific data dependency,” arXiv preprint arXiv:2506.03645, 2025

  17. [25]

    Restormer: Efficient transformer for high- resolution image restoration,

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang, “Restormer: Efficient transformer for high- resolution image restoration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5728–5739

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.