Organizations: School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China · Chinese Aeronautical Establishment, Beijing, China · Westlake University, Hangzhou, China · MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, Shanghai, China
RAW video restoration is fundamental to high-quality low-level perception and serves as the basis for a wide range of downstream vision applications. While binary neural networks (BNNs) enable efficient lightweight deployment for image enhancement, their deficiencies in modeling temporal coherence and activation value distributions hinder their effectiveness when applied to video scenarios. In this paper, we propose BinRVR, a binarized RAW video restoration framework that reduces computation and parameters by approximately 96% while incurring only about 4% performance degradation. Specifically, we present a Binarized Information Interaction Module (BIIM) to jointly model spatial and temporal information in an efficient and unified manner. Moreover, we develop a Distribution-Aware Binarized Convolution (DAB-Conv) that leverages the statistics of full-precision activations to mitigate quantization errors. The proposed framework further supports multi-bit quantization, enabling flexible accuracy-efficiency trade-offs across different hardware constraints. Extensive experiments demonstrate that our BinRVR achieves competitive performance compared with state-of-the-art binarized methods on RAW video restoration tasks, including low-light enhancement, denoising, deblurring, and super-resolution. We further explore the potential of our method on downstream video applications, including object detection and monocular depth estimation.
Figures & tables
Fig. 1: Performance and model complexity comparison with state-of-the-art binary neural networks (BNNs) and lightweight full-precision methods in RAW video restoration. The size of each circle reflects the model’s parameter count. Our method outperforms both the BNN-based methods and lightweight full-precision methods with comparable model complexity.
Fig. 2: Illustration of the proposed BinRVR. (a) Overall architecture of our binarized RAW video restoration framework. The recurrent arrows indicate the propagated features between adjacent sliding windows. Efficient inter-frame and intra-frame information interaction under binarization is achieved by the proposed Binarized Information Interaction Module (BIIM), whose detailed structure is shown in Fig. 3 . (b) Distribution-Aware Binarized Convolution (DAB-Conv), which leverages feature distribution information to reduce quantization error and improve representation capability in binarized networks.
Fig. 3: Detailed illustration of the proposed Binarized Information Interaction Module (BIIM). It is designed to enable efficient interaction between inter-frame temporal information and intra-frame spatial information under binarization through lightweight operations. (a) Binarized Information Interaction Encoder. (b) Binarized Information Interaction Decoder.
Fig. 5: Illustration of spatial information interaction module. (a) Illustration of the spatial information interaction module, where feature channels are split into identity, 3×3 , 1×11 , and 11×1 binarized depthwise convolution branches. (b) FLOPs per spatial location versus kernel size. (c) Parameter complexity versus kernel size.
Method
Params (M) ↓
FLOPs (G) ↓
SMOID [ 37 ]
LLRVD [ 10 ]
Gain0
Gain15
Gain30
PSNR ↑
SSIM ↑
ST-RRED ↓
PSNR ↑
SSIM ↑
ST-RRED ↓
PSNR ↑
SSIM ↑
ST-RRED ↓
PSNR ↑
SSIM ↑
ST-RRED ↓
Linear Scaled
-
-
31.26
0.7665
0.4965
32.35
0.8108
0.2541
32.27
0.8358
0.4156
30.04
0.6941
0.0863
Full-precision methods
SMOID [ 37 ]
12.21
15.11
39.74
0.9733
0.0446
39.43
0.9751
0.0461
40.06
0.9769
0.0431
37.07
0.9580
0.0391
RViDeNet [ 38 ]
8.57
65.77
41.09
0.9778
0.0510
40.94
0.9803
0.0494
41.60
0.9818
0.0500
37.39
0.9620
0.0400
TABLE I: Quantitative comparison on SMOID [ 37 ] and LLRVD [ 10 ] for low-light RAW video enhancement. For binarized methods, the best and second-best results are shown in bold and underlined , respectively. Lightweight full-precision methods are additionally included for reference, with the best result shown in bold .
Fig. 6: Visual comparisons of binarized methods on the SMOID [ 37 ] and LLRVD [ 10 ] datasets for low-light RAW video enhancement (left: SMOID; right: LLRVD). For each example, the left half shows zoomed-in RGB results obtained by applying an ISP pipeline to the RAW outputs, while the right half presents the corresponding error maps with respect to the ground truth.
Method
Noisy
FastDVD [ 40 ]
EMVD-S [ 42 ]
RViDeNet [ 38 ]
LLRVD [ 10 ]
FloRNN [ 13 ]
BNN [ 19 ]
Bireal [ 61 ]
IRNet [ 62 ]
ReActNet [ 59 ]
BTM [ 69 ]
BBCU [ 20 ]
BiSRNet [ 84 ]
FRBC [ 85 ]
BHViT [ 72 ]
BinRVR (Ours)
Params(M) ↓
-
2.48
0.81
8.57
6.29
10.49
0.26
0.26
0.26
0.28
0.25
0.27
0.28
0.26
0.27
0.22
FLOPs(G) ↓
-
10.50
1.66
65.77
46.14
24.57
1.38
1.38
1.38
1.56
1.33
1.47
1.44
1.38
1.47
1.23
PSNR ↑
32.02
39.84
42.63
43.97
44.23
45.15
42.05
42.06
41.97
41.92
43.11
43.21
43.16
41.48
43.17
43.53
SSIM ↑
0.7333
0.9703
0.9851
0.9874
0.9879
0.9907
0.9804
0.9809
0.9800
0.9805
0.9847
0.9850
0.9848
0.9779
0.9849
0.9868
ST-RRED ↓
0.0171
-
-
-
-
-
0.0118
0.0109
0.0119
0.0116
0.0103
0.0100
0.0100
0.0119
0.0100
0.0080
TABLE II: Quantitative comparison on CRVD [ 53 ] for RAW video denoising. For binarized methods, the best and second-best results are shown in bold and underlined , respectively. Lightweight full-precision methods are additionally included for reference, with the best result shown in bold . The results of full-precision methods are reported from the original papers.
Fig. 7: Visual comparisons of binarized methods on the CRVD [ 53 ] dataset for RAW video denoising. For each example, the left half shows zoomed-in RGB results obtained by applying an ISP pipeline to the RAW outputs, while the right half presents the corresponding error maps with respect to the ground truth.
Method
Blurry
MIMO [ 91 ]
ESTRNN [ 92 ]
LIEDNet [ 93 ]
BNN [ 19 ]
Bireal [ 61 ]
IRNet [ 62 ]
ReActNet [ 59 ]
BTM [ 69 ]
BBCU [ 20 ]
BiSRNet [ 84 ]
FRBC [ 85 ]
BHViT [ 72 ]
BinRVR (Ours)
Params(M)
-
6.81
2.53
4.76
0.26
0.26
0.26
0.28
0.25
0.27
0.28
0.26
0.27
0.22
FLOPs(G)
-
16.00
12.77
5.09
1.38
1.38
1.38
1.56
1.33
1.47
1.44
1.38
1.47
1.23
PSNR ↑
36.52
38.15
38.12
38.61
36.88
36.81
36.95
37.33
37.35
37.50
37.71
36.88
37.33
38.89
SSIM ↑
0.9493
0.9663
0.9673
0.9708
0.9531
0.9520
0.9546
0.9583
0.9589
0.9607
0.9626
0.9548
0.9587
0.9724
ST-RRED ↓
0.0180
0.0151
0.0157
0.0098
0.0180
0.0158
0.0179
0.0169
0.0179
0.0175
0.0167
0.0168
0.0169
0.0106
TABLE III: Quantitative comparison on Deblur-RAW [ 90 ] for RAW video deblurring. For binarized methods, the best and second-best results are shown in bold and underlined , respectively. Lightweight full-precision methods are additionally included for reference, with the best result shown in bold .
Fig. 8: Visual comparisons of binarized methods on the Deblur-RAW [ 90 ] dataset for RAW video deblurring.
Method
Params (M) ↓
FLOPs (G) ↓
2×
3×
4×
PSNR ↑
SSIM ↑
ST-RRED ↓
PSNR ↑
SSIM ↑
ST-RRED ↓
PSNR ↑
SSIM ↑
ST-RRED ↓
Bicubic
-
-
35.23
0.9434
0.2702
33.40
0.9121
0.2827
31.25
0.8558
0.7312
Full-precision methods
TDAN [ 95 ]
1.86
13.76
34.19
0.9344
0.4757
33.06
0.9203
0.5424
32.13
0.8882
0.9604
BasicVSR++ [ 29 ]
7.18
113.13
35.71
0.9579
0.3684
35.42
0.9345
0.3096
33.91
0.9121
0.5131
Realviformer [ 94 ] †
5.82
40.46
37.53
0.9729
0.1720
-
-
-
34.57
0.9305
0.3654
TABLE IV: Quantitative comparison on Real-RawVSR [ 9 ] for RAW video super-resolution. For binarized methods, the best and second-best results are shown in bold and underlined , respectively. Lightweight full-precision methods are additionally included for reference, with the best result shown in bold . Params and FLOPs are computed under the 2× scaling setting. † indicates that Realviformer [ 94 ] and FRBC [ 85 ] fail to train at certain scales.
Fig. 9: Visual comparisons of binarized methods on Real-RawVSR [ 9 ] for RAW video super-resolution (top: 2× ; bottom: 3× ). For each example, the left half shows zoomed-in RGB results obtained by applying an ISP pipeline to the RAW outputs, while the right half presents the corresponding error maps with respect to the ground truth.
Tasks
Metrics
Linear Scaled
BNN [ 19 ]
Bireal [ 61 ]
IRNet [ 62 ]
ReActNet [ 59 ]
BTM [ 69 ]
BBCU [ 20 ]
BiSRNet [ 84 ]
FRBC [ 85 ]
BHViT [ 72 ]
BRVE [ 86 ]
BinRVR (Ours)
Object Detection
AP ↑
61.98
62.85
65.02
63.49
65.46
66.83
66.61
66.60
53.39
66.97
68.52
70.62
AP 50 ↑
64.93
65.63
67.91
66.29
68.15
69.77
69.44
69.59
56.31
69.84
71.05
73.24
AP 75 ↑
63.34
64.22
66.67
64.98
66.96
68.49
68.27
68.27
54.67
68.66
69.97
72.11
Depth Estimation
RMSE ↓
0.1308
0.1251
0.1270
0.1255
0.1224
0.1136
0.1115
0.1149
0.2215
0.1127
0.1067
0.1029
SSIM ↑
0.9324
0.9335
0.9313
0.9323
0.9342
0.9382
0.9392
0.9372
0.8770
0.9386
0.9463
0.9482
TABLE V: Quantitative results of downstream applications of low-light RAW video enhancement using the SMOID [ 37 ] dataset. We provide results for object detection and depth estimation. The best and second-best values are Bold and Underline , respectively.
Method
Bits (w/a)
Params ↓
FLOPs ↓
PSNR ↑
SSIM ↑
ST-RRED ↓
LSQ [ 18 ]
4/4
0.80M
5.81G
37.10
0.9585
0.0461
QuantSR [ 73 ]
4/4
0.80M
5.81G
37.09
0.9585
0.0460
Q-SCI [ 96 ]
4/4
0.80M
5.81G
37.13
0.9590
0.0433
Ours
4/4
0.80M
5.81G
37.71
0.9646
0.0382
LSQ [ 18 ]
3/3
0.63M
3.39G
37.05
0.9580
0.0465
QuantSR [ 73 ]
3/3
0.63M
3.39G
37.06
0.9581
0.0464
TABLE VI: Multi-Bit quantization results on LLRVD [ 10 ] dataset for low-light RAW video enhancement.
Fig. 10: Qualitative comparisons of object detection results produced by GroundingDINO [ 97 ] on enhanced videos.
Fig. 11: Qualitative comparisons of monocular depth estimation results produced by Depth Anything V2 [ 100 ] on enhanced videos.
Fig. 12: Effect of the sliding window stride on recurrent propagation. A larger stride improves efficiency but compromises temporal consistency in restored videos across frames.
Method
PSNR ↑
SSIM ↑
ST-RRED ↓
RAW2RAW+ISP
30.59
0.8408
0.2002
RAW2RGB
27.53
0.8143
0.2938
RGB2RGB
24.91
0.7994
0.5941
RAW2RAW+ISP+H264
30.24
0.8421
0.1810
RGB2RGB+H264
24.88
0.8170
0.5610
H264+RGB2RGB
20.15
0.6138
1.4199
TABLE VII: Comparison of different low-light video enhancement settings on LLRVD [ 10 ] dataset. Metrics are computed in the RGB domain after ISP processing.
Arch.
Window
Rec.
Params ↓
FLOPs ↓
PSNR ↑
SSIM ↑
ST-RRED ↓
Pure sliding window
3
✗
0.22M
0.74G
36.83
0.9567
0.0648
Pure recurrent
1
✓
0.22M
1.05G
36.98
0.9579
0.0506
Hybrid (Ours)
3
✓
0.22M
1.23G
37.21
0.9599
0.0436
TABLE VIII: Architecture-level ablation study on the temporal modeling paradigm. All variants use the same backbone, 1-bit quantization setting, and training configuration.
Partition
Ablation
PSNR ↑
SSIM ↑
ST-RRED ↓
Params ↓
FLOPs ↓
BIIM Encoder
w/o TII
37.06
0.9585
0.0468
0.22M
1.23G
Same as Decoder-TII
37.15
0.9586
0.0450
0.24M
1.43G
BIIM Decoder
w/o TII
37.00
0.9581
0.0470
0.21M
1.12G
Same as Encoder-TII
37.14
0.9595
0.0451
0.21M
1.12G
w/o SII
37.04
0.9580
0.0469
0.21M
1.17G
Strip Kernel →1×9
37.16
0.9592
0.0453
0.22M
1.23G
TABLE IX: Ablation studies on our Binarized Information Interaction (BIIM) Module .
Sequence Length
Avg. over Gain0 / Gain15 / Gain30
PSNR ↑
SSIM ↑
ST-RRED ↓
30 frames
40.86
0.9789
0.0610
60 frames
40.88
0.9790
0.0606
90 frames
40.88
0.9790
0.0606
120 frames
40.88
0.9790
0.0605
TABLE X: Long-sequence evaluation on the SMOID [ 37 ] test set. The metrics are averaged over Gain0, Gain15, and Gain30, and computed on the same target interval.
Mean(⋅)
Std(⋅)
Mean(∣⋅∣)
PSNR ↑
SSIM ↑
ST-RRED ↓
Params ↓
FLOPs ↓
✗
✗
✗
37.04
0.9581
0.0458
0.222M
1.21G
✓
✗
✗
37.12
0.9589
0.0447
0.222M
1.21G
✓
✓
✗
37.17
0.9595
0.0444
0.223M
1.21G
✓
✓
✓
37.21
0.9599
0.0436
0.223M
1.23G
TABLE XI: Ablation studies for our Distribution-Aware Binarized Convolution.
Fig. 13: Visual ablation comparison of representative BinRVR variants. (a) Ablation of temporal modeling components. (b) Ablation of spatial modeling components. Each comparison includes the degraded input, ablation variants without key components, the full model, the ground truth, and the corresponding error maps.
Fig. 14: Visualization of activation distributions and DAB-Conv scaling factors. (a)–(c) Channel-wise activation statistics before binarization under different gain settings. (d)–(f) Representative activation distributions from different channels. (g)–(i) Predicted DAB-Conv scales versus activation statistics. The results show that activation distributions vary across channels and degradations, motivating input-dependent channel-wise scaling.
Platform
Model
Mode
FPS ↑
Latency (ms/f.) ↓
PSNR ↑
NVIDIA RTX 4060 GPU Local laptop
ShiftNet
FP32
3.74
267.34
37.87
Ours
FP32
93.22
10.73
37.81
NVIDIA Ampere GPU NVIDIA Jetson Orin NX
ShiftNet
FP32
1.30
770.39
37.87
Ours
FP32
19.94
50.16
37.81
Ours
W8A8
35.18
28.42
37.77
Intel Core Ultra 9 CPU Local laptop
ShiftNet
FP32
0.87
1144.17
37.87
TABLE XII: Actual-runtime deployment comparison with the representative full-precision method ShiftNet [ 8 ] on different hardware platforms under currently supported inference modes. PSNR is evaluated on the LLRVD [ 10 ] test set.
Enhancing videos under extreme low-light conditions remains challenging due to the difficulty of balancing restoration quality and computational efficiency in resource-constrained settings. This paper introduces EeveeDark, a low-light video enhancement framework that combines the spatial richness of sensor-level RAW data with the temporal precision of event streams. Central to our model is a Binary Neural Network (BNN) architecture that reduces computational overhead by quantizing weights and activations while preserving detail. EeveeDark incorporates (i) modality-specific binary encoders for processing RAW frames and event data, (ii) a lightweight fusion block for integrating spatial and temporal cues, and (iii) an event-guided skip gating mechanism for dynamic spatiotemporal refinement. Experiments on synthetic and real-world datasets show that EeveeDark outperforms prior BNN-based methods and offers a favorable performance-efficiency trade-off compared to full-precision models. The project page is available at https://cyberiada.github.io/EeveeDark.
Onur Eker, Erkut Erdem, Aykut Erdem
Department of Computer Engineering, Hacettepe University, TR-06800 Ankara, Turkey, and also with HAVELSAN Inc., TR-06510, Ankara, Turkey · Department of Computer Engineering, Hacettepe University, TR-06800 Ankara, Turkey, and also with the Koc University Is Bank AI Center, TR-34450, Istanbul, Turkey · Department of Computer Engineering, Koc University, TR-34450 Istanbul, Turkey, and also with the Koc University Is Bank AI Center, TR-34450, Istanbul, Turkey
Diffusion-based video restoration recovers realistic details, but its practical deployment is limited by two efficiency bottlenecks: costly VAE encoding and decoding, and the quadratic cost of full self-attention in diffusion transformers (DiTs). This paper presents FastVR, a streaming video restoration framework built on a one-step diffusion model, which delivers strong restoration quality and temporal consistency while processing 1080p video at 11 FPS on a single H20 GPU. To improve inference efficiency, FastVR combines a lightweight VAE with chunk-wise causal attention, which substantially reduces the computational cost. During training, it further adopts velocity consistency regularization and continuous trajectory learning, which improve restoration quality. Extensive experiments show that FastVR is more efficient than the evaluated diffusion baselines while achieving state-of-the-art performance on synthetic and real-world benchmarks. We hope that this work supports further progress in the community.
Raw images preserve linear sensor measurements and high bit-depth information crucial for advanced vision tasks and photography applications, yet their storage remains challenging due to large file sizes, varying bit depths, and sensor-dependent characteristics. Existing learned lossless compression methods mainly target 8-bit sRGB images, while raw reconstruction approaches are inherently lossy and rely on camera-specific assumptions. To address these challenges, we introduce RAWIC, a bit-depth-adaptive learned lossless compression framework for Bayer-pattern raw images. We first convert single-channel Bayer data into a four-channel RGGB format and partition it into patches. For each patch, we compute its bit depth and use it as auxiliary input to guide compression. A bit-depth-adaptive entropy model is then designed to estimate patch distributions conditioned on their bit depths. This architecture enables a single model to handle raw images from diverse cameras and bit depths. Experiments show that RAWIC consistently surpasses traditional lossless codecs, achieving an average 7.7% bitrate reduction over JPEG-XL. Our code is available at https://github.com/chunbaobao/RAWIC.
Chunhang Zheng, Tongda Xu, Mingli Xie +2
School of Electronics, Peking University, Beijing, 100871, China · Institute for AI Industry Research (AIR), Tsinghua University, Beijing, 100084, China · the University of Chinese Academy of Sciences, Beijing, 100190, China