Pith. sign in

REVIEW 4 major objections 5 minor 121 references

A single blind 4K video restorer can handle haze, rain, noise, and darkness together, without labels or optical flow.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 20:08 UTC pith:Q4HNYKKN

load-bearing objection Useful 4K all-in-one benchmark and an efficient architecture, but the headline accuracy claim rests on a benchmark that encodes the band-split premise, and the one external fine-band test goes the other way. the 4 major comments →

arxiv 2608.01829 v1 pith:Q4HNYKKN submitted 2026-08-03 cs.CV

MoCRA: Mixture of Compositional Rank-1 Atoms for 4K All-in-One Video Restoration

Yongcong Wang , Pu Wang , Hingchin Chen , Runci Bai , Yucheng Xin , Chen Wu , Chengchao Shen , Guangwei Gao
show 3 more authors
Siyuan Yao Pengwen Dai Zhuoran Zheng
This is my paper
classification cs.CV
keywords all-in-one video restoration4K restorationblind restorationrain streak removaldenoisinglow-light enhancementdehazingrank-1 atoms
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

MoCRA tries to establish that a single blind video-restoration model can serve four degradations at native 4K if it stops trying to identify the degradation globally and instead splits the work by resolution band. The paper's premise is that haze and low light are smooth fields that survive aggressive downsampling, while rain streaks and sensor noise live only in native pixels. On that basis, a lightweight coarse branch restores global photometry from a 512-pixel view, and a shallow native-resolution refiner cleans high-frequency structure, with one shared dictionary of rank-1 atoms routed sparsely per frame. Trained once on a new paired 4K benchmark with all four degradations over shared clean clips, the model is reported to lead eleven retrained baselines on task-mean PSNR, match flow-based video methods on temporal stability without ever computing motion, and restore a 4K frame in under half a second. The paper would matter because it turns the joint 4K all-in-one restoration problem from a capacity and memory problem into a band-allocation problem.

Core claim

The central claim is that the four degradations separate cleanly by scale, and that allocation by scale is enough: haze and low light are low-frequency fields, so they can be corrected once per clip on a downsampled view; rain and noise are native-pixel phenomena, so they must be removed at full resolution. MoCRA embodies the split in a two-branch design. A once-per-clip coarse branch processes a 4.2x-decimated view and emits both a coarse restoration and a routing tensor; a shallow native-resolution refiner adds high-frequency residual corrections using cues that distinguish transient artifacts from persistent texture. Everything is conditioned by a per-frame sparse mixture of rank-1 atoms

What carries the argument

The central object is a shared dictionary of rank-1 atoms with per-frame sparse routing. At each of 68 linear injection sites, a dictionary of 96 factor pairs (u_k v_k^T) supplies rank-1 weight updates; a small Atom Composer reads a 256-pixel view of the downsampled clip and emits a sparse coefficient vector per frame, choosing r=12 active atoms. This routing tensor is the model's entire degradation judgment: it adapts the operator rather than the activation, threads one routing space through both the coarse and fine branches, and is regularized for temporal smoothness. The second load-bearing mechanism is band-matched supervision: a degradation-evidence weighting (DEW) map concentrates the

Load-bearing premise

The claim rests on the assumption that haze and low light are always smooth enough to survive a 4.2x downsampling, while rain and noise are always fine-scale and pixel-local; if a real degradation has structure coarser than about eight pixels in the fine band, the two bands coincide and the model loses its advantage.

What would settle it

Take a real 4K video with visible rain streaks that are several pixels wide or with spatially correlated sensor noise, run MoCRA against a flow-based video derainer, and check whether MoCRA still holds task-mean PSNR and temporal stability; the paper's own RainMotion result, where the short side falls below the coarse view's 512 and MoCRA trails by about 10 dB, already provides a concrete test of this condition.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single checkpoint can serve multiple deployment conditions without a degradation label, per-frame prompts, or optical flow, removing a major obstacle to shipping one restoration model.
  • 4K restoration cost decouples from the coarse analysis: the expensive part is a shallow refiner, so high-resolution output can be affordable on memory-constrained devices.
  • Per-frame routing should make the model robust to degradations that switch within a clip, since switching only re-mixes atoms for one frame instead of breaking a fused window.
  • The conditioning vocabulary is compositional, suggesting unseen compound degradations could be handled by mixing atoms rather than retraining.
  • The design premise gives a simple testable rule for when to use the model: only when rain and noise are genuinely native-scale phenomena.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If real rain streaks are wider than the roughly eight-pixel threshold set by the 4.2x decimation, the band split collapses; the paper's own external RainMotion result at 640x480, where MoCRA falls well behind flow-based models, is a concrete demonstration of the premise withdrawn.
  • The benchmark's synthetic renders encode the band split by construction: streaks are rasterized as one-pixel lines and noise is drawn per pixel, so the model's advantage on UHV-4K-AIO may overstate its edge on real footage where these assumptions soften.
  • One natural extension is to make the coarse-view threshold adaptive, so the model could decide per clip whether a degradation is low-band or high-band rather than hard-coding a fixed decimation factor.
  • The per-frame atom routing could be probed as an interpretable degradation sensor: if routing vectors correlate with physical parameters like haze thickness or noise variance, the same mechanism could serve blind quality assessment.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes MoCRA, a 3.6M-parameter blind model for joint 4K video dehazing, deraining, denoising, and low-light enhancement, together with a new benchmark, UHV-4K-AIO, that renders these four degradations over a shared set of 100 clean 4K clips with common depth and motion. The method is organized around a band split: haze and low light survive a 4.2× antialiased downsampling to short side 512, while rain and noise live only at native scale. MoCRA uses a per-frame sparse mixture of rank-1 atom dictionaries to condition a coarse branch plus a shallow native-resolution refiner, with no optical flow. On the proposed benchmark, trained once for all four tasks, MoCRA reports the best task-mean PSNR (27.23 dB versus 27.11 dB for AirNet) and 0.48 s per 4K frame, while also reporting worse mean SSIM and LPIPS than several baselines. On the external RainMotion benchmark at 640×480 the ranking inverts, which the paper explicitly acknowledges as "the design premise withdrawn."

Significance. If the claims hold, the paper makes a useful contribution: UHV-4K-AIO is the only 4K all-in-one video benchmark with a shared clean ground truth across four degradations, and the released physical-state modalities plus consistency checks are valuable assets for the community. The architecture is an efficient and non-obvious way to combine coarse photometric restoration with a native-scale refiner, and the authors are unusually transparent in retraining all baselines under one recipe and in reporting the RainMotion failure. However, the headline advantage is small (0.12 dB mean PSNR over AirNet, from single runs), the benchmark is constructed so that the central band-split premise is true by construction, and the only external fine-band evaluation contradicts the model's design assumption. The paper therefore demonstrates a competitive and fast model on its own benchmark, but the real-world generalization claim is not yet supported.

major comments (4)
  1. [Tables 1 and 4] The same "full" MoCRA model is reported with different numbers in the main comparison and the ablation table. Table 1 gives 27.23 dB mean PSNR with per-task values 23.47, 28.70, 32.07, 24.66; Table 4 gives 27.43/0.900 with per-task PSNRs 23.86, 28.71, 32.33, 24.83. No explanation is provided. The 0.20 dB discrepancy between these two rows is larger than the paper's headline margin over AirNet (0.12 dB). This is load-bearing: the reader cannot tell which numbers correspond to the released checkpoint, and the discrepancy suggests seed/checkpoint sensitivity. Please reconcile the tables, identify the official model, and report results over multiple seeds.
  2. [Table 1 / Experimental Setup] The central claim "takes the best task-mean PSNR of eleven retrained baselines" rests on a single run with no error bars or significance test for a 0.12 dB margin. Given the Table 1/Table 4 inconsistency, run-to-run variability is evidently non-negligible. Please report mean and standard deviation over at least 3 seeds for MoCRA and the two closest baselines (AirNet, AverNet), and state whether the PSNR difference is statistically reliable. Without this, the claimed advantage over AirNet is not established.
  3. [§Dataset, §A2.3, §A11] The benchmark encodes the band split rather than testing it. Rain is rasterized at 4K as one-pixel anti-aliased line segments and noise is drawn per pixel (Eq. 2, A2.3), while the coarse branch operates on a 4.2× decimation that erases structure finer than roughly eight native pixels. Consequently, Table 1 cannot validate the premise that rain and noise exist only at native scale; it is true by construction. The only external fine-band test, RainMotion at 640×480, inverts the ranking (MoCRA 21.79 dB vs. 31.53–31.77 dB for flow-based baselines), and the paper itself calls this "the design premise withdrawn." This is honest but leaves the real-world claim unprotected. Please add a controlled 4K test in which the fine-band assumption is violated while staying at native resolution—for example, rain streaks wider than ~8 native pixels, motion-blurred streaks, or spatially correlated sensor no
  4. [§A6 / Table 1 fairness note] The comparison is carefully retrained under one recipe, which is a strength, but one detail should be clarified. The appendix states that EDVR is the official EDVR-M variant, while other video baselines use their full published models. EDVR-M is a reduced-capacity version. Since the paper's efficiency and accuracy comparisons depend on these choices, please justify why EDVR-M is used and report the official EDVR full-model result if it can be trained under the recipe, or state explicitly that the full model does not fit the training budget.
minor comments (5)
  1. [Figure 4 caption] The caption text "Rain concentrates won the structure it altered" appears garbled; presumably "Rain concentrates [weight] on the structure it altered" is intended.
  2. [Eq. (5)] The notation meanc is used before it is defined in Eq. (6). Define it at first use.
  3. [Abstract] The Abstract says "takes the best task-mean PSNR"; this is accurate, but given the substantial SSIM/LPIPS losses, consider adding the qualifier "on UHV-4K-AIO" and a pointer to the multi-metric results to avoid implying overall superiority.
  4. [Experiments / Table 1 caption] Runtime is reported as 0.48 s per 4K frame, but the caption does not state the GPU. Figure 7 implies V100; please state the hardware in the caption or table notes.
  5. [§A4] The tiling fallback description is clear, but it would help to state the tile overlap and cosine-taper parameters used in the reported runs, since tiled baselines are timed with these settings.

Circularity Check

1 steps flagged

The central band-split premise is encoded into MoCRA's own benchmark: rain is rasterized at 4K as one-pixel lines and noise is per-pixel, so the claim that rain/noise live only at native scale is true by construction, and the sole external fine-band test (RainMotion) inverts the ranking.

specific steps
  1. self definitional [Abstract; UHV-4K-AIO Dataset / Degradation Synthesis; Methodology (Dual-Band Execution); A11 Results on Public Benchmarks]
    "its construction exposes the split MoCRA is built on: haze and low light survive aggressive downsampling, while rain and noise exist only at native scale. ... rain streaks are rasterized at 4K, where upsampling would attenuate them ... n_sr_t ~ N(0, sigma_s^2 g_t x_t + sigma_r^2) ... an antialiased 4.2x decimation that erases structure finer than roughly eight native pixels: it is the operational form of the band split MoCRA is organized around ... On RainMotion the ranking inverts and MoCRA falls 10dB behind ... That is the design premise withdrawn."

    The fine-band half of MoCRA's central premise is not independently established; it is written into the benchmark by the renderer. Rain is deliberately synthesized as one-pixel anti-aliased line segments at 4K and noise is drawn per-pixel, so these degradations are guaranteed to disappear under the 4.2x downsampling that defines the coarse band. MoCRA is then organized around that same split and evaluated on that same benchmark, so the Table 1 margin partially returns the premise that was inserted into the data. The paper's only external fine-band test, RainMotion at 640x480, sits below the coarse view's 512 short side, and the authors themselves describe the failure as 'the design premise withdrawn.' This does not invalidate the method, but it means the claim 'rain and noise exist only at

full rationale

The main circularity is one of benchmark construction, not of fitted parameters or self-citation chains. The dataset's rain and noise are synthesized so that they exist only at native scale—rain as one-pixel rasterized lines, noise as per-pixel draws—and the model's coarse branch downsampling is defined to erase exactly such structures. Evaluating the model on that benchmark therefore partly measures whether the model matches the renderer's assumptions, not whether real rain and sensor noise are inherently native-scale phenomena. The paper is unusually honest about the boundary of its claim: REVIDE real haze, which is near the coarse scale, is won independently, and the RainMotion failure at 640x480 is reported as the design premise withdrawn. These external results substantially mitigate the circularity; they show the coarse-band half of the premise has independent support and that the fine-band half is at least openly conditional. No load-bearing self-citation or uniqueness argument was found, and the efficiency and stability results are not circular. Because the central state-of-the-art claim still rests in part on a premise encoded into the benchmark by construction, a score of 6 is appropriate: one core prediction reduces by construction, with partial independent support.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 1 invented entities

Everything the central claim rests on: loss weights and DEW constants are grid-searched on held-out validation (disclosed); the band split is assumed from the authors' rendering choices; pretrained depth/flow models supply shared geometry and the temporal ground truth; sensor and rain models are borrowed from cited work. Benchmark severity tiers are hand-set. No physical entity is postulated; the rank-1 atom routing is an internal construct evidenced only on the authors' benchmark.

free parameters (7)
  • Reconstruction and auxiliary loss weights omega_g, omega_c, omega_p, omega_b, omega_i, omega_alpha, omega_Y = 0.15, 0.8, 0.025, 0.01, 0.03, 0.08, 0.02
    Grid-searched once on held-out validation clips, then frozen for every task, ablation, and baseline (Eq. 7). All reported performance is conditional on this search.
  • DEW weighting parameters lambda and s = 24, 0.03
    Chosen to shape the degradation-evidence map w (Eq. 6); lambda=24 sets the fine-band gradient budget and s the presence-gate width.
  • Sparse routing capacity r, temperature schedule tau, exploration noise sigma = r=12, tau 5 to 0.4, sigma 0.3 to 0
    Hand-chosen; annealing breaks the atom/routing feedback loop described in the Methodology, and r=12 sets the mixture size.
  • Coarse-view scale and composer scale = short side 512; composer short side 256
    This is the operational form of the band split: the claim that rain and noise live in the fine band is literally defined as structure finer than roughly eight native pixels at the 512 decimation.
  • Temporal gate width rho = 0.05
    Eq. 5 gate that decides which static regions enforce coarse-output temporal consistency.
  • Dictionary size K=96, active atoms r=12, sites L=68 (56 backbone + 12 refiner) = 96/12/68
    Manual architectural capacities; the ablation table varies rank and prototype memory but not K or L.
  • UHV-4K-AIO severity tier parameters
    Hand-set per-clip draws for haze beta and A, rain density/speed/opacity/veil, noise ISO tiers, exposure, gamma, vignette, color cast (Table A2). These define task difficulty, and the low-light brightness guard shifts the tier balance to 37/42/21.
axioms (6)
  • standard math Top-r sparse gating with temperature-scaled softmax (Shazeer et al. 2017) produces routing coefficients alpha; treated as established.
    Eq. 4 uses the sparsely-gated MoE convention without modification.
  • standard math Low-rank weight adaptation: adapting a linear map by a sum of rank-1 factor pairs is an effective conditioning mechanism.
    Assumed from the LoRA literature the paper cites (Hu et al. 2022, LoRA-IR, UIR-LoRA); the paper contributes the per-frame sparse composition, not the low-rank premise.
  • domain assumption Depth and optical flow estimated by pretrained Depth-Anything-V2 and RAFT on a 960p proxy are accurate enough to serve as shared geometry and motion for all four degradations and as reference motion for temporal metrics.
    Appendix A2.1; errors propagate into rain particle motion, veil smoothing, and the warping-error metric that all methods are scored against.
  • domain assumption The band split: haze and low light are smooth fields robust to 4.2x antialiased decimation; rain streaks and sensor noise are sub-8-pixel native-only structure.
    Assumed from the authors' 4K rendering choices; contradicted in the low-resolution regime by the paper's own RainMotion results (Section A11).
  • domain assumption The sensor noise model (Foi et al. 2008, calibrated by Yue et al. 2020) and the rain particle synthesis (RDD-Net) are faithful enough that rankings on synthetic UHV-4K-AIO transfer to real footage.
    No real 4K rain or noise data is used anywhere; REVIDE covers only real haze.
  • ad hoc to paper Low-light synthesis adjusts tier assignments by content brightness; near-black scenes are forced to the mildest tier and heavy assignments demoted.
    Appendix A2.3; this moderates low-light difficulty for dark content and shifts the tier balance to 37/42/21, directly affecting the low-light task where MoCRA records its largest win.
invented entities (1)
  • Per-frame rank-1 atom dictionaries with sparse routing tensor alpha_{t,l} no independent evidence
    purpose: Recompose weight updates at 68 sites across coarse and fine branches from one routing space, replacing per-degradation experts or dense prompts.
    An architectural construct; Table 4 ablations on the authors' own benchmark are the only evidence that sparse rank-1 composition, rather than the dual-band setup or the benchmark's encoding of the premise, causes the gains.

pith-pipeline@v1.3.0-daily-deepseek · 5613 in / 5710 out tokens · 325543 ms · 2026-08-04T20:08:19.766994+00:00 · methodology

0 comments
read the original abstract

Real-world video arrives hazy, rainy, dark, or noisy, and a deployable restorer faces three demands at once: no degradation label, native 4K output, and stability in playback. Existing methods answer them separately and break on the joint problem, because per-frame degradation readings flip between frames, downsampled proxies erase the rain and noise they are meant to remove, and dense temporal alignment does not fit 4K memory. No paired benchmark even poses that problem, so we build one. UHV-4K-AIO renders physically modeled haze, rain, sensor noise, and low light over the same 100 clean 4K clips with shared depth and motion, and its construction exposes the split MoCRA is built on: haze and low light survive aggressive downsampling, while rain and noise exist only at native scale. Band-matched compositional conditioning follows, spending conditioning capacity, computation, and supervision in the band where each degradation lives. One dictionary of rank-1 atoms, recomposed sparsely per frame, conditions both a once-per-clip coarse branch and a shallow native-resolution refiner, in 3.6M parameters and with no optical flow. Trained once for all four tasks, MoCRA takes the best task-mean PSNR of eleven retrained image and video baselines, holds warping error at the level of the flow-based video models while never estimating motion, and restores native 4K in under half a second, against 1.7 seconds for the fastest baseline.

Figures

Figures reproduced from arXiv: 2608.01829 by Chengchao Shen, Chen Wu, Guangwei Gao, Hingchin Chen, Pengwen Dai, Pu Wang, Runci Bai, Siyuan Yao, Yongcong Wang, Yucheng Xin, Zhuoran Zheng.

Figure 1
Figure 1. Figure 1: Task-mean PSNR against per-frame 4K run￾time; bubble area is proportional to parameter count. No baseline is both faster and more accurate than MoCRA. prompt or an embedding (Li et al. 2022; Potlapalli et al. 2023; Cui et al. 2025); repeated per frame, the reading can flip between neighbors, and the flip lands in the output as a color step. Ultra-high-definition restoration estimates its correction on a do… view at source ↗
Figure 2
Figure 2. Figure 2: Samples of UHV-4K-AIO. Two clips, each with clean frame and native-resolution crops of four degradations. UHV-4K-AIO Dataset Paired restoration data is all-in-one only at 1080p and below (Yang et al. 2023; Zhao et al. 2024), 4K for one degrada￾tion at a time (Deng et al. 2021; Zhao et al. 2025; Wang et al. 2026), or real captured pairs below 4K (Zhang et al. 2021; Wang et al. 2021); Table A1 places UHV-4K-… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of MoCRA. The window is reduced once to I ↓ , where the Atom Composer emits α and the coarse branch restores global photometry; only the Fine Refiner reads native pixels. Dashed arrows carry α. Insets: (a) composer, (b) backbone block, (c) refiner. keep: the degraded center frame with its Sobel magnitude, and the two adjacent frames with their absolute differences. Transience marks the degradation… view at source ↗
Figure 4
Figure 4. Figure 4: Degradation-evidence weighting on one held-out clip rendered under rain and under noise. Rain concentrates w on the structure it altered; noise leaves it near uniform. punishes. Under that loss our refiner passes 0.81 of the high￾frequency energy inside rain-streak masks through to the output, against 0.16 for noise with the same weights. Capacity is not the bottleneck; the loss is spent where the degradat… view at source ↗
Figure 5
Figure 5. Figure 5: Restoration at native 4K, one held-out clip per degradation, all methods blind; boxes are 1:1 pixel crops. Each baseline settles on its own white balance in the photometric rows, while MoCRA tracks the reference [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Temporal stability on one scanline. A fixed band of a low-light test clip, shared luminance scale; the scene itself dims briefly early on. 540p 1080p 4K output resolution 0.1 1 10 latency / frame (s) coarse branch ∝ pixels 540p 1080p 4K output resolution 1 10 peak memory (GiB) V100, 32 GiB Ours AverNet BasicVSR++ Shift-Net RVRT [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: checks that the routing is not degenerate. It com￾mits without collapsing: per-frame entropy averages 2.23 [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 8
Figure 8. Figure 8: Phone demos on real degraded videos. Left: 4K [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

121 extracted references · 9 linked inside Pith

  1. [2]

    C.; Zhou, S.; Xu, X.; and Loy, C

    Chan, K. C.; Zhou, S.; Xu, X.; and Loy, C. C. 2022. Basicvsr++: Improving video super-resolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 5972--5981

    2022

  2. [3]

    Charbonnier, P.; Blanc-F \'e raud, L.; Aubert, G.; and Barlaud, M. 1994. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of 1st International Conference on Image Processing, volume 2, 168--172. IEEE

    1994

  3. [4]

    Chen, H.; Chen, X.; Wu, C.; Zheng, Z.; Pan, J.; and Fu, X. 2026. Towards Ultra-High-Definition Image Deraining: A Benchmark and an Efficient Method. IEEE Transactions on Multimedia, 1--13

    2026

  4. [5]

    Chu, M.; Xie, Y.; Mayer, J.; Leal-Taix \'e , L.; and Thuerey, N. 2020. Learning temporal coherence via self-supervision for GAN-based video generation. ACM Transactions on Graphics (TOG), 39(4): 75--1

    2020

  5. [6]

    W.; Khan, S.; Knoll, A.; Shah, M.; and Khan, F

    Cui, Y.; Zamir, S. W.; Khan, S.; Knoll, A.; Shah, M.; and Khan, F. 2025. Adair: Adaptive all-in-one image restoration via frequency mining and modulation. In International Conference on Learning Representations, volume 2025, 101306--101327

    2025

  6. [7]

    Darcet, T.; Oquab, M.; Mairal, J.; and Bojanowski, P. 2024. Vision Transformers Need Registers. In International Conference on Learning Representations

    2024

  7. [8]

    Deng, S.; Ren, W.; Yan, Y.; Wang, T.; Song, F.; and Cao, X. 2021. Multi-scale separable network for ultra-high-definition video deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 14030--14039

    2021

  8. [9]

    Foi, A.; Trimeche, M.; Katkovnik, V.; and Egiazarian, K. 2008. Practical Poissonian-Gaussian noise modeling and fitting for single-image raw-data. IEEE transactions on image processing, 17(10): 1737--1754

    2008

  9. [10]

    Geman, S.; and McClure, D. E. 1987. Statistical methods for tomographic image reconstruction. Bulletin of the International Statistical Institute, 52(4): 5--21

    1987

  10. [11]

    K.; Salameh, M.; and Niu, D

    Ghasemabadi, A.; Janjua, M. K.; Salameh, M.; and Niu, D. 2024. Learning truncated causal history model for video restoration. Advances in Neural Information Processing Systems, 37: 27584--27615

    2024

  11. [12]

    W.; and He, S

    Guo, Y.; Gao, Y.; Lu, Y.; Zhu, H.; Liu, R. W.; and He, S. 2024. Onerestore: A universal restoration framework for composite degradation. In European conference on computer vision, 255--272. Springer

    2024

  12. [13]

    He, K.; Sun, J.; and Tang, X. 2013. Guided Image Filtering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(6): 1397--1409

    2013

  13. [14]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations

    2022

  14. [15]

    Jiang, L.; Dai, B.; Wu, W.; and Loy, C. C. 2021. Focal Frequency Loss for Image Reconstruction and Synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 13899--13909

    2021

  15. [16]

    Ke, J.; Wang, Q.; Wang, Y.; Milanfar, P.; and Yang, F. 2021. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 5148--5157

    2021

  16. [17]

    Lai, W.-S.; Huang, J.-B.; Wang, O.; Shechtman, E.; Yumer, E.; and Yang, M.-H. 2018. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), 170--185

    2018

  17. [18]

    Li, A.; Liu, X.; Li, S.; Du, Y.; Long, Z.; Luo, L.; Zhang, L.; and Zhu, C. 2026. DRNet: All-in-One Image Restoration via Prior-Guided Dynamic Reparameterization. IEEE Transactions on Multimedia

    2026

  18. [19]

    Li, B.; Liu, X.; Hu, P.; Wu, Z.; Lv, J.; and Peng, X. 2022. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 17452--17462

    2022

  19. [21]

    C.; See, S.; Wang, X.; Qin, H.; and Li, H

    Li, D.; Shi, X.; Zhang, Y.; Cheung, K. C.; See, S.; Wang, X.; Qin, H.; and Li, H. 2023 b . A simple baseline for video restoration with grouped spatial-temporal shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9822--9832

    2023

  20. [22]

    Liang, J.; Cao, J.; Fan, Y.; Zhang, K.; Ranjan, R.; Li, Y.; Timofte, R.; and Van Gool, L. 2024. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 33: 2171--2182

    2024

  21. [23]

    Liang, J.; Fan, Y.; Xiang, X.; Ranjan, R.; Ilg, E.; Green, S.; Cao, J.; Zhang, K.; Timofte, R.; and Gool, L. V. 2022. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35: 378--393

    2022

  22. [24]

    Lin, J.; Morris, C.; Lin, R.; Zhang, F.; Bull, D.; and Anantrasirichai, N. 2025. Towards a General-Purpose Zero-Shot Synthetic Low-Light Image and Video Pipeline. In Proceedings of the 3rd International Workshop on Multimedia Content Generation and Evaluation: New Methods and Practice, McGE '25, 3–11. New York, NY, USA: Association for Computing Machinery...

    2025

  23. [25]

    Liu, Y.; Li, D.; Fu, X.; Lu, X.; Huang, J.; and Zha, Z.-J. 2025. Uhd-processer: Unified uhd image restoration with progressive frequency learning and degradation-aware prompts. In Proceedings of the Computer Vision and Pattern Recognition Conference, 23121--23130

    2025

  24. [26]

    R.; Waller, L.; and Koltun, V

    Monakhova, K.; Richter, S. R.; Waller, L.; and Koltun, V. 2022. Dancing under the stars: video denoising in starlight. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 16241--16251

    2022

  25. [27]

    Oquab, M.; Darcet, T.; Moutakanni, T.; Vo, H. V.; Szafraniec, M.; Khalidov, V.; Fernandez, P.; Haziza, D.; Massa, F.; El-Nouby, A.; Assran, M.; Ballas, N.; Galuba, W.; Howes, R.; Huang, P.-Y.; Li, S.-W.; Misra, I.; Rabbat, M.; Sharma, V.; Synnaeve, G.; Xu, H.; J \'e gou, H.; Mairal, J.; Labatut, P.; Joulin, A.; and Bojanowski, P. 2024. DINOv2: Learning Ro...

    2024

  26. [28]

    Perez, E.; Strub, F.; de Vries, H.; Dumoulin, V.; and Courville, A. C. 2018. FiLM: Visual Reasoning with a General Conditioning Layer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 3942--3951

    2018

  27. [29]

    W.; Khan, S

    Potlapalli, V.; Zamir, S. W.; Khan, S. H.; and Shahbaz Khan, F. 2023. Promptir: Prompting for all-in-one image restoration. Advances in neural information processing systems, 36: 71275--71293

    2023

  28. [30]

    V.; Hinton, G

    Shazeer, N.; Mirhoseini, A.; Maziarz, K.; Davis, A.; Le, Q. V.; Hinton, G. E.; and Dean, J. 2017. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. In International Conference on Learning Representations

    2017

  29. [31]

    Sun, S.; Zhang, Y.; Wu, C.; Lu, D.; Zhang, G.; Wen, Y.; and Zheng, Z. 2025. Uniflowrestore: A general video restoration framework via flow matching and prompt guidance. In Proceedings of the 33rd ACM International Conference on Multimedia, 7757--7765

    2025

  30. [32]

    Teed, Z.; and Deng, J. 2020. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, 402--419. Springer

    2020

  31. [33]

    Valanarasu, J. M. J.; Yasarla, R.; and Patel, V. M. 2022. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2353--2363

    2022

  32. [34]

    Wang, C.; Zhang, K.; and Yang, J. 2026. Retrieve-to-Restore: Efficient All-in-One Image Restoration with a Retrieval-Based Degradation Bank. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1277--1287

    2026

  33. [35]

    Wang, R.; Xu, X.; Fu, C.-W.; Lu, J.; Yu, B.; and Jia, J. 2021. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In Proceedings of the IEEE/CVF international conference on computer vision, 9700--9709

    2021

  34. [36]

    Wang, S.; Zhu, L.; Fu, H.; Qin, J.; Sch \"o nlieb, C.-B.; Feng, W.; and Wang, S. 2022. Rethinking video rain streak removal: A new synthesis model and a deraining network with video rain prior. In European Conference on Computer Vision, 565--582. Springer

    2022

  35. [37]

    Wang, T.; Zhang, K.; Shen, T.; Luo, W.; Stenger, B.; and Lu, T. 2023. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 2654--2662

    2023

  36. [38]

    C.; Yu, K.; Dong, C.; and Change Loy, C

    Wang, X.; Chan, K. C.; Yu, K.; Dong, C.; and Change Loy, C. 2019. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 0--0

    2019

  37. [40]

    Yang, L.; Kang, B.; Huang, Z.; Zhao, Z.; Xu, X.; Feng, J.; and Zhao, H. 2024. Depth anything v2. Advances in Neural Information Processing Systems, 37: 21875--21911

    2024

  38. [41]

    Yang, S.; Wu, T.; Shi, S.; Lao, S.; Gong, Y.; Cao, M.; Wang, J.; and Yang, Y. 2022. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1191--1200

    2022

  39. [42]

    I.; Fu, H.; Liu, Y.; Wang, W.; and Zhu, L

    Yang, Y.; Aviles-Rivero, A. I.; Fu, H.; Liu, Y.; Wang, W.; and Zhu, L. 2023. Video adverse-weather-component suppression network via weather messenger and adversarial backpropagation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 13200--13210

    2023

  40. [43]

    Yue, H.; Cao, C.; Liao, L.; Chu, R.; and Yang, J. 2020. Supervised raw video denoising with a benchmark dataset on dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2301--2310

    2020

  41. [44]

    W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F

    Zamir, S. W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F. S.; and Yang, M.-H. 2022. Restormer: Efficient Transformer for High-Resolution Image Restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5718--5729

    2022

  42. [46]

    A.; Shechtman, E.; and Wang, O

    Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, 586--595

    2018

  43. [47]

    Zhang, X.; Dong, H.; Pan, J.; Zhu, C.; Tai, Y.; Wang, C.; Li, J.; Huang, F.; and Wang, F. 2021. Learning to restore hazy video: A new real-world dataset and a new method. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9239--9248

    2021

  44. [48]

    Zhao, H.; Tian, L.; Xiao, X.; Hu, P.; Gou, Y.; and Peng, X. 2024. AverNet: All-in-one video restoration for time-varying unknown degradations. Advances in Neural Information Processing Systems, 37: 127296--127316

    2024

  45. [50]

    Zheng, Z.; Ren, W.; Cao, X.; Hu, X.; Wang, T.; Song, F.; and Jia, X. 2021. Ultra-High-Definition Image Dehazing via Multi-Guided Bilateral Learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 16180--16189

    2021

  46. [51]

    arXiv preprint arXiv:2605.11508 , year=

    LiBrA-Net: Lie-Algebraic Bilateral Affine Fields for Real-Time 4K Video Dehazing , author=. arXiv preprint arXiv:2605.11508 , year=

    Pith/arXiv arXiv

  47. [52]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    All-in-one image restoration for unknown corruption , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  48. [53]

    Advances in neural information processing systems , volume=

    Promptir: Prompting for all-in-one image restoration , author=. Advances in neural information processing systems , volume=

  49. [54]

    European conference on computer vision , pages=

    Onerestore: A universal restoration framework for composite degradation , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  50. [55]

    Ultra-High-Definition Image Dehazing via Multi-Guided Bilateral Learning , year=

    Zheng, Zhuoran and Ren, Wenqi and Cao, Xiaochun and Hu, Xiaobin and Wang, Tao and Song, Fenglong and Jia, Xiuyi , booktitle=. Ultra-High-Definition Image Dehazing via Multi-Guided Bilateral Learning , year=

  51. [56]

    Towards Ultra-High-Definition Image Deraining: A Benchmark and an Efficient Method , year=

    Chen, Hongming and Chen, Xiang and Wu, Chen and Zheng, Zhuoran and Pan, Jinshan and Fu, Xianping , journal=. Towards Ultra-High-Definition Image Deraining: A Benchmark and an Efficient Method , year=

  52. [57]

    arXiv preprint arXiv:2302.11831 , year=

    Embedding fourier for ultra-high-definition low-light image enhancement , author=. arXiv preprint arXiv:2302.11831 , year=

    Pith/arXiv arXiv

  53. [58]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  54. [59]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages=

    Uniflowrestore: A general video restoration framework via flow matching and prompt guidance , author=. Proceedings of the 33rd ACM International Conference on Multimedia , pages=

  55. [60]

    Advances in Neural Information Processing Systems , volume=

    Depth anything v2 , author=. Advances in Neural Information Processing Systems , volume=

  56. [61]

    European conference on computer vision , pages=

    Raft: Recurrent all-pairs field transforms for optical flow , author=. European conference on computer vision , pages=. 2020 , organization=

    2020

  57. [62]

    European Conference on Computer Vision , pages=

    Rethinking video rain streak removal: A new synthesis model and a deraining network with video rain prior , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  58. [63]

    IEEE transactions on image processing , volume=

    Practical Poissonian-Gaussian noise modeling and fitting for single-image raw-data , author=. IEEE transactions on image processing , volume=. 2008 , publisher=

    2008

  59. [64]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Supervised raw video denoising with a benchmark dataset on dynamic scenes , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  60. [65]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Dancing under the stars: video denoising in starlight , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  61. [66]

    Proceedings of the 3rd International Workshop on Multimedia Content Generation and Evaluation: New Methods and Practice , pages =

    Lin, Joanne and Morris, Crispian and Lin, Ruirui and Zhang, Fan and Bull, David and Anantrasirichai, Nantheera , title =. Proceedings of the 3rd International Workshop on Multimedia Content Generation and Evaluation: New Methods and Practice , pages =. 2025 , isbn =. doi:10.1145/3746278.3759376 , abstract =

    arXiv 2025

  62. [67]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  63. [68]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Transweather: Transformer-based restoration of images degraded by adverse weather conditions , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  64. [69]

    International Conference on Learning Representations , volume=

    Adair: Adaptive all-in-one image restoration via frequency mining and modulation , author=. International Conference on Learning Representations , volume=

  65. [70]

    International conference on learning representations , volume=

    Controlling vision-language models for multi-task image restoration , author=. International conference on learning representations , volume=

  66. [71]

    European Conference on Computer Vision , pages=

    Instructir: High-quality image restoration following human instructions , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  67. [72]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Complexity experts are task-discriminative learners for any image restoration , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  68. [73]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Foundir: Unleashing million-scale training data to advance foundation models for image restoration , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  69. [74]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Correlation matching transformation transformers for uhd image restoration , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  70. [75]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Uhd-processer: Unified uhd image restoration with progressive frequency learning and degradation-aware prompts , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  71. [76]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Basicvsr++: Improving video super-resolution with enhanced propagation and alignment , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  72. [77]

    IEEE Transactions on Image Processing , volume=

    Vrt: A video restoration transformer , author=. IEEE Transactions on Image Processing , volume=. 2024 , publisher=

    2024

  73. [78]

    Advances in Neural Information Processing Systems , volume=

    Recurrent video restoration transformer with guided deformable attention , author=. Advances in Neural Information Processing Systems , volume=

  74. [79]

    Advances in Neural Information Processing Systems , volume=

    Learning truncated causal history model for video restoration , author=. Advances in Neural Information Processing Systems , volume=

  75. [80]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Seedvr: Seeding infinity in diffusion transformer towards generic video restoration , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  76. [81]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Video adverse-weather-component suppression network via weather messenger and adversarial backpropagation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  77. [82]

    Advances in Neural Information Processing Systems , volume=

    AverNet: All-in-one video restoration for time-varying unknown degradations , author=. Advances in Neural Information Processing Systems , volume=

  78. [83]

    IEEE Transactions on Circuits and Systems for Video Technology , year=

    Removing Multiple Hybrid Adverse Weather in Video via a Unified Model , author=. IEEE Transactions on Circuits and Systems for Video Technology , year=

  79. [84]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Grounding degradations in natural language for all-in-one video restoration , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  80. [85]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Learning to restore hazy video: A new real-world dataset and a new method , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Showing first 80 references.