Prevailing High Dynamic Range (HDR) video reconstruction methods are fundamentally trapped in a fragile alignment-and-fusion paradigm. While explicit spatial alignment can successfully recover fine details in controlled environments, it becomes a severe bottleneck in unconstrained dynamic scenes. By forcing rigid alignment across unpredictable motions and varying exposures, these methods inevitably translate registration errors into severe ghosting artifacts and temporal flickering. In this paper, we rethink this conventional prerequisite. Recognizing that explicit alignment is inherently vulnerable to real-world complexities, we propose LoCAtion, a Long-time Collaborative Attention framework that reformulates HDR video generation from a fragile spatial warping task into a robust, alignment-free collaborative feature routing problem. Guided by this new formulation, our architecture explicitly decouples the highly entangled reconstruction task. Rather than struggling to rigidly warp neighboring frames, we anchor the scene on a continuous medium-exposure backbone and utilize collaborative attention to dynamically harvest and inject reliable irradiance cues from unaligned exposures. Furthermore, we introduce a learned global sequence solver. By leveraging bidirectional context and long-range temporal modeling, it propagates corrective signals and structural features across the entire sequence, inherently enforcing whole-video coherence and eliminating jitter. Extensive experiments demonstrate that LoCAtion achieves state-of-the-art visual quality and temporal stability, offering a highly competitive balance between accuracy and computational efficiency.
Figures & tables
Fig. 1: Motivation of LoCAtion. (a) Existing HDR video reconstruction pipelines (alternating-exposure or dual-stream) are largely formulated as frame-centric local reconstruction, recovering each HDR frame from a short reference-centered window. (b) LoCAtion reformulates dual-stream HDR reconstruction as backbone-guided compensation and propagation. (c) Temporal profile visualization: intensity values sampled along a fixed scanline are stacked across consecutive frames, with horizontal axis as spatial position and vertical axis as time. Broken/stripe-like profiles indicate frame-to-frame intensity inconsistency or temporal instability, while smooth and coherent profiles reflect improved temporal stability.
Fig. 2: Overall architecture of LoCAtion for dual-stream HDR video reconstruction. Given a continuous medium-exposure sequence and sparse low-/high-exposure anchors, LoCAtion organizes reconstruction around the medium-exposure backbone. In the Collaborative Feature Attention (CFA) stage, Temporal Prior Aggregator (TPA) estimates reliability guidance for the low-/high-exposure anchors, and Content-Faithful Fusion (CFF) selectively injects reliable cross-exposure cues in the low-frequency wavelet domain while preserving the structural content of the medium-exposure stream. The intermediate HDR estimates are further refined by Global Sequence Consistency (GSC), where Bidirectional Context Aggregation (BCA) propagates reconstruction context from both temporal directions and Long-Range Temporal Modeling (LRTM) models temporal dependencies within the selected reconstruction segment. Red and blue dashed arrows denote forward and backward temporal propagation, respectively.
Input
Framework
Method
Params (M) ↓
Time (ms) ↓
Slow/Simple Scenes (6)
Dynamic/Complex Scenes (10)
PSNR μ ↑
SSIM μ ↑
t -PSNR ↑
t -SSIM ↑
FovVideoVDP ↑
STD ↓
PSNR μ ↑
SSIM μ ↑
t -PSNR ↑
t -SSIM ↑
FovVideoVDP ↑
STD ↓
AE
Frame- centric
DeepHDRVideo(ICCV’21) [ 8 ]
2.76
167.60
36.78
0.8674
39.26
0.8814
9.7251
4.30
38.71
0.9516
41.22
0.9405
9.0058
3.95
LAN-HDR(ICCV’23) [ 13 ]
7.39
230.70
36.82
0.8617
39.44
0.8827
9.7265
4.23
38.64
0.9539
41.46
0.9475
9.0568
3.52
HDRFlow(CVPR’24) [ 10 ]
5.33
17.60
36.93
0.8696
39.39
0.8869
9.7611
3.90
38.52
0.9514
40.37
0.9499
8.9856
4.29
DS
Frame- centric
AHDNet(CVPR’19) [ 28 ]
1.44
158.90
36.45
0.8663
40.27
0.8852
9.6958
0.39
39.32
0.9630
43.35
0.9527
9.2846
1.68
DomainPlus(MM’22) [ 40 ]
24.78
116.90
36.54
0.8588
40.34
0.8833
9.7260
0.42
39.58
0.9608
43.49
0.9510
9.2416
1.71
TABLE I: Quantitative comparison on the 16-scene HDR video benchmark. The benchmark is divided into two subsets: six slow-motion/simple scenes and ten dynamic/complex scenes. AE and DS denote alternating-exposure inputs and dual-stream inputs, respectively. Best and second-best results are highlighted in bold and underline .
Fig. 3: Speed-performance trade-off comparison with recent state-of-the-art HDR methods on the Cinematic Video dataset. LoCAtion attains high reconstruction quality with moderate model size and substantially lower latency than heavy Transformer-based or frame-centric dual-stream baselines.
Fig. 4: Qualitative comparisons on representative synthetic and real HDR video scenes. Each row shows a challenging case involving saturated highlights, complex lighting conditions, or fast motion. Image-based methods tend to leave residual ghosts or blurred textures due to insufficient temporal reasoning, while video-based alignment methods may introduce noise or motion-induced artifacts. LoCAtion better suppresses ghosting and noise while preserving fine structures in both highlight and shadow regions.
Input
Framework
Method
Real-Captured Quality
Efficiency
LSD ↓
t -SSIM ↑
t -PSNR ↑
Params (M) ↓
Time (ms) ↓
Rel. Time ↓
AE
Frame- centric
DeepHDRVideo (ICCV’21) [ 8 ]
14.0
0.7092
19.55
2.76
167.60
2.04 ×
LAN-HDR (ICCV’23) [ 13 ]
13.1
0.6527
20.26
7.39
230.70
2.81 ×
HDRFlow (CVPR’24) [ 10 ]
11.9
0.7591
20.33
5.33
17.60
0.21 ×
DS
Frame- centric
AHDNet (CVPR’19) [ 28 ]
1.7
0.8765
29.05
1.44
158.90
1.93 ×
HDR-Trans (ECCV’22) [ 30 ]
1.7
0.8729
28.98
1.46
1531.30
18.64 ×
TABLE II: Quantitative comparison with state-of-the-art methods on public real-captured HDR video datasets without ground-truth HDR references. AE and DS denote alternating-exposure inputs and dual-stream inputs, respectively. Rel. Time denotes the runtime ratio relative to our method. Best and second-best results are highlighted in bold and underline , respectively.
Fig. 5: Comparison of cross-exposure routing strategies. All variants share the same dual-stream input and only replace the first-stage compensation module. Temporal profiles visualize scanline intensity evolution over time, where smoother profiles indicate better temporal consistency. Our CFA reduces artifacts and yields a more coherent temporal reconstruction.
Variant
STD ↓
PSNRμ↑
ΔPSNRμ↓
SSIMμ↑
Params (M) ↓
Time (ms) ↓
Component removal
w/o DWT
0.74
35.67
0.08
0.9048
4.50
220.8
w/o TPA
0.75
35.44
0.31
0.9004
3.78
61.3
w/o BCA
0.78
35.16
0.59
0.8969
2.64
80.2
w/o LRTM
0.73
35.51
0.24
0.9021
4.71
73.0
w/o GSC
0.82
34.62
1.13
0.8629
2.57
66.9
TABLE III: Ablation studies of network components and design alternatives. All variants are trained and evaluated under the same dual-stream input protocol. The component-removal variants evaluate the contributions of DWT-domain fusion, TPA, BCA, and LRTM, while the design-alternative variants replace the proposed CFA stage with flow-based alignment, local attention, or cross-attention.
Fig. 6: Ablation study on the impact of Temporal Prior Aggregator (TPA). Without TPA, the reconstructed details around moving regions tend to show duplication and ghosting-like artifacts due to unreliable exposure cue selection. With TPA, the motion region is reconstructed more consistently.
(a) Mid-Exposure Sequence Length
Setting
STD ↓
PSNRμ↑
SSIMμ↑
Memory (GB) ↓
T=1
0.78
35.31
0.9012
4.79
T=5 (Ours)
0.71
35.75
0.9051
19.06
T=7
0.59
35.81
0.9107
24.96
T=10
0.62
35.73
0.9096
34.01
T=15
0.61
35.66
0.9115
49.42
TABLE IV: Ablation study on temporal input configuration. (a) Effect of the medium-exposure sequence length T , which controls the temporal context available for propagation. (b) Effect of the medium-frame-to-anchor-pair sampling ratio. A ratio of 5:1 means that one low-/high-exposure anchor pair is shared by a 5-frame medium-exposure window. The selected configuration is highlighted in gray.
Fig. 7: Feature-level analysis of CFA cross-exposure routing. We visualize normalized feature responses from the low-/high-exposure branches, their fused representation, and the final prediction. In R1 , CFA suppresses inconsistent branch responses around motion, preventing duplicated structures. In R2 , reliable complementary cues are retained to enhance subtle textures. Overall, CFA adaptively routes cross-exposure information for targeted fusion rather than naive feature merging.
Fig. 8: Reliability maps and GSC correction analysis. The low- and high-exposure anchors exhibit complementary reliability: αlow favors highlight-preserving features, while αhigh is less trusted in saturated regions. Stage-1 (local-only) results are refined by GSC, where the residual concentrates on motion boundaries and other locally unstable areas. Compared with uniform smoothing, this targeted correction yields lower and temporally smoother brightness errors over the sequence.
Reconstructing High Dynamic Range (HDR) videos from sequences of alternating-exposure Low Dynamic Range (LDR) frames remains highly challenging, especially under dynamic scenes where cross-exposure inconsistencies and complex motion make inter-frame alignment difficult, leading to ghosting and detail loss. Existing methods often suffer from inaccurate alignment, suboptimal feature aggregation, and degraded reconstruction quality in motion-dominated regions. To address these challenges, we propose F2HDR, a two-stage HDR video reconstruction framework that robustly perceives inter-frame motion and restores fine details in complex dynamic scenarios. The proposed framework integrates a flow adapter that adapts generic optical flow for robust cross-exposure alignment, a physical motion modeling to identify salient motion regions, and a motion-aware refinement network that aggregates complementary information while removing ghosting and noise. Extensive experiments demonstrate that F2HDR achieves state-of-the-art performance on real-world HDR video benchmarks, producing ghost-free and high-fidelity results under large motion and exposure variations.
Huanjing Yue, Dawei Li, Shaoxiong Tu +1
School of Electrical and Information Engineering, Tianjin University · State Key Laboratory of Smart Power Distribution Equipment and System, Tianjin University · Huawei Technologies Co., Ltd.
High dynamic range (HDR) video reconstruction from al-ter-na-ting-exposure sequences remains challenging, especially in regions with extreme luminance variation. We propose a novel HDR reconstruction framework based on dual-stream registration and accurate pyramid fusion. Given three consecutive frames, our method computes optical flow directly with the central frame, while introducing a complementary midpoint displacement strategy to handle cases with severe overexposition. A pyramid fusion stage then merges the resulting radiance and LDR images into a final HDR output. Experimental results demonstrate that our approach consistently outperforms state-of-the-art methods.
Onofre Martorell, Ivan Pereira-Sánchez, Antoni Fuentes +1
Dept. of Mathematics and Computer Science, Universitat de les Illes Balears, Cra. de Valldemossa km 7.5, Palma, 07122, Illes Balears, Spain
Recent generative methods for single-shot high dynamic range (HDR) image reconstruction show promising results, but often struggle with preserving fidelity to the input image. They require separate models to handle highlights and shadows, or sacrifice interpretability by directly predicting the final HDR image. We address these limitations by re-casting single-shot HDR reconstruction as conditional video generation and fusing the generated frames into an HDR image. We finetune a video diffusion model to generate an exposure bracket, conditioned on a low dynamic range (LDR) input. We fuse this image bracket using per-pixel weights predicted by a light-weight UNet. This formulation is simple, interpretable, and effective. Rather than directly hallucinating an HDR image, it explicitly reconstructs the intermediate exposure stack and fuses it into the final output. Our method eliminates the need for separate models across exposure regimes and produces HDR reconstructions with high input fidelity. On quantitative benchmarks, we outperform state-of-the-art generative baselines with comparable model capacity on several reconstruction metrics. Human evaluators further prefer our results in 72% of pairwise comparisons against existing methods. Finally, we show that this input-conditioned sequence generation and fusion framework extends beyond HDR to other image reconstruction tasks, such as all-in-focus image recovery from a single defocus-blurred input.
Chinmay Talegaonkar, Jinshi He, Christopher McKenna +1