CompMVR: Compression-Aware Multi-View Restoration Using Diffusion Models for Geometrically Consistent 3D Reconstruction
Authors: Dong-hwi Kim, Chaewon Moon, Hojun Song, Dongbeom Kim, Junyeong Jang, Aro Kim, Gahyeon Kim, Heejung Choi, Jehee Kim, Gianella Cravioto, Sohyun Lee, Gyeongjin Choi, EunHye Jeong, Soo Ye Kim, Jaehyup Lee, Sang-hyo Park
Organizations: Kyungpook National University, Republic of Korea · Adobe Research, USA
Real-world 3D reconstruction pipelines often rely on multi-view images degraded by lossy compression. While compression artifacts have been extensively studied in single-image or video restoration, their impact on multi-view geometry remains largely overlooked. In this work, we identify compression-induced cross-view geometric inconsistency as a critical failure mode, where view-dependent artifacts disrupt feature matching, dense correspondence, camera parameter estimation, and downstream 3D reconstruction tasks such as novel view synthesis. To address this problem, we propose CompMVR, a compression-aware multi-view diffusion restoration framework that jointly refines compressed views while preserving cross-view consistency, and construct compressed datasets for systematic evaluation. Unlike conventional restoration methods that optimize each image independently for 2D perceptual quality, CompMVR leverages learned compression priors and cross-view correspondence to recover visual details while improving geometric reliability across views. Experiments across diverse datasets, codecs, and compression levels demonstrate gains in restoration quality and downstream 3D tasks, including novel-view synthesis, camera pose estimation, and view matching. These results establish compressed multi-view restoration as a distinct problem and highlight its importance for 3D reconstruction under lossy compression. Code and data for this paper are available at https://mchaewon.github.io/CompMVR-project-page/.
Figures & tables
Figure 1. Effect of HEVC compression on VGGT feature matching in the Free/Lab scene. Lossy compression degrades cross-view correspondence: as the quantization parameter (QP) increases from 27 to 42, the number of matched features decreases compared to the uncompressed input. While existing diffusion-based compression-aware restoration methods recover visual quality, they do not consistently restore geometry-sensitive correspondences. Our method produces more reliable feature matches under both compression levels, demonstrating improved cross-view consistency for downstream 3D reconstruction.
Method
MeanMatches ↑
AUC@5 ↑
Uncompressed
GT
2,001
24.55
Compressed
HEVC, QP42
1,655
22.59
Restored
CODiff ∗
1,625
23.36
DiQP ∗
1,643
23.45
SIR-Diff ∗
1,627
23.09
Ours
1,659
24.34
Table 1. Effect of lossy compression on VGGT view matching on RealEstate10K. We assess cross-view consistency restoration using the number of matched correspondences and AUC@5 for pose accuracy. Our method achieves the best view matching performance, outperforming diffusion-based image, video, and multi-view restoration baselines. The Best and Second-best results are highlighted in red and orange, respectively.
Figure 2. Overview of our proposed CompMVR. (a) In Stage 1, we learn a compression-aware latent representation by training a compression prior embedder. The latent is optimized through image reconstruction and supervision of coding parameters (codec type and QP). (b) In Stage 2, we perform multi-view restoration using a diffusion model conditioned on the learned compression prior. The latent is injected via cross-attention and fused with the UNet input, while a compression artifact estimator (CAE) predicts spatially varying residual maps for degradation-aware restoration. To enforce cross-view geometric consistency, we further apply a cross-view correspondence loss during training.
Figure 3. Energy maps computed from uncompressed images and artifact maps compressed at identical QP settings, for low and high complexity scenes. Higher scene complexity yields larger compression residuals.
Figure 4. Failure cases of a video diffusion restoration model ( Dehaghi et al., 2025 ) on multi-view images. Examples from the Free dataset (hydrant and pillar scenes, left to right) show that the model fails to maintain cross-view consistency. Highlighted regions reveal inconsistent restoration and geometric misalignment caused by large viewpoint differences.
Codec
QP
PSNR ↑
SSIM ↑
LPIPS ↓
Compressed
CODiff ∗
DiQP ∗
SIR-Diff ∗
Ours
Compressed
CODiff ∗
DiQP ∗
SIR-Diff ∗
Ours
Compressed
CODiff ∗
DiQP ∗
SIR-Diff ∗
Ours
AVC
27
22.35
20.83
20.50
19.72
22.90
0.601
0.541
0.625
0.512
0.642
0.309
0.378
0.393
0.426
0.236
42
22.01
21.39
19.95
19.61
23.21
0.644
0.586
0.571
0.517
0.688
0.382
0.433
0.472
0.462
0.325
HEVC
27
23.45
22.11
20.54
20.07
24.08
0.638
0.580
0.607
0.516
0.670
0.300
0.382
0.425
0.447
0.242
42
22.96
22.49
21.10
19.90
24.49
0.687
0.605
0.595
0.517
0.718
0.285
0.403
0.451
0.464
0.265
AV1
27
22.37
20.87
20.27
19.63
22.84
0.605
0.550
0.404
0.512
0.647
0.321
0.379
0.618
0.431
0.238
Table 2. Quantitative novel-view synthesis results using 3D Gaussian Splatting on the Free dataset under different codec and QP settings. We report PSNR, SSIM, and LPIPS, where higher PSNR/SSIM and lower LPIPS indicate better rendering quality. The Best and Second-best results are highlighted in red and orange, respectively.
Codec
QP
AUC@15 ↑
AUC@30 ↑
Compressed
CODiff ∗
DiQP ∗
SIR-Diff ∗
Ours
Compressed
CODiff ∗
DiQP ∗
SIR-Diff ∗
Ours
AVC
27
64.72
72.27
70.05
73.28
77.52
77.36
81.04
81.30
81.71
85.00
42
61.37
69.36
66.16
69.50
73.22
76.20
79.02
78.69
79.26
82.79
HEVC
27
64.12
72.91
70.11
73.35
79.32
76.84
81.38
81.33
81.26
87.52
42
63.77
71.44
69.70
69.61
75.26
76.00
79.82
80.75
79.07
83.70
VP9
27
61.90
73.14
70.85
72.65
77.56
75.38
81.45
81.95
80.92
85.32
Table 3. VGGT camera pose estimation results on the RealEstate10K dataset under different codec and QP settings. AUC@15 and AUC@30 are reported, where higher values indicate better pose estimation accuracy. The Best and Second-best results are highlighted in red and orange, respectively.
Codec
QP
AUC@5 ↑
Compressed
CODiff ∗
DiQP ∗
SIR-Diff ∗
Ours
AVC
27
23.90
26.68
27.06
20.89
27.42
42
20.51
21.52
20.27
20.40
21.34
HEVC
27
26.03
25.96
25.26
21.81
29.46
42
22.59
23.36
23.45
23.09
24.34
VP9
27
24.58
25.03
25.86
22.13
25.18
Table 4. VGGT view matching results on the RealEstate10K dataset under different codec and QP settings. We report AUC@5, where higher values indicate better pose estimation accuracy. The Best and Second-best results are highlighted in red and orange, respectively.
PSNR ↑
SSIM ↑
LPIPS ↓
NIQE ↓
MUSIQ ↑
CLIPIQA ↑
MEt3R ↓
CODiff ∗
23.74
0.662
0.277
3.40
69.18
0.681
0.135
DiQP ∗
27.23
0.803
0.225
4.06
64.99
0.393
0.136
SIR-Diff ∗
20.91
0.794
0.426
3.83
60.99
0.565
0.141
Ours
23.89
0.897
0.281
3.26
68.08
0.673
0.132
Table 5. Quantitative comparison on the Free dataset under a matched fine-tuning protocol. We report 2D restoration metrics and MEt3R for evaluating image quality and multi-view geometric consistency. The Best and Second-best results are highlighted in red and orange, respectively.
Method
PSNR ↑
NIQE ↓
CLIPIQA ↑
CODiff ∗
23.4970
3.3412
0.6917
Ours
23.6228
3.2610
0.6981
Table 6. Generalization to the unseen VVC/H.266 codec on the Free dataset. VVC is not included during training. We report PSNR, NIQE, and CLIPIQA, where higher PSNR/CLIPIQA and lower NIQE indicate better restoration quality. The Best and Second-best results are highlighted in red and orange, respectively.
Codec
QP
PSNR ↑
SSIM ↑
LPIPS ↓
NIQE ↓
MUSIQ ↑
CLIPIQA ↑
CODiff
DiQP
SIR-Diff*
Ours
CODiff
DiQP
SIR-Diff*
Ours
CODiff
DiQP
SIR-Diff*
Ours
CODiff
DiQP
SIR-Diff*
Ours
CODiff
DiQP
SIR-Diff*
Ours
CODiff
DiQP
SIR-Diff*
Ours
AVC
27
23.29
22.38
20.92
23.45
0.896
0.847
0.799
0.906
0.356
0.374
0.390
0.224
5.06
3.37
3.59
3.10
65.60
50.65
62.60
69.84
0.647
0.395
0.577
0.704
42
23.29
21.65
20.73
23.03
0.869
0.802
0.778
0.857
0.458
0.484
0.461
0.391
5.37
4.47
4.07
3.25
58.48
40.66
61.20
65.479
0.529
0.315
0.552
0.610
HEVC
27
24.42
23.12
21.27
25.98
0.897
0.831
0.796
0.916
0.355
0.382
0.447
0.218
5.06
3.49
3.92
3.24
65.02
46.49
58.36
68.39
0.634
0.360
0.555
0.689
42
24.50
22.70
21.20
25.61
0.889
0.811
0.789
0.900
0.409
0.451
0.477
0.320
5.27
4.06
4.21
3.42
62.54
41.93
58.18
67.11
0.581
0.330
0.547
0.640
AV1
27
23.30
22.34
20.91
23.74
0.897
0.845
0.798
0.906
0.359
0.382
0.397
0.236
5.05
3.47
3.62
3.20
65.82
50.48
62.89
69.51
0.645
0.401
0.577
0.709
Table 7. Quantitative comparison of 2D restoration methods on the Free dataset. We evaluate various 2D diffusion-based restoration models under different codec and QP settings. The Best and Second-best results are highlighted in red and orange, respectively.
Figure 5. t-SNE visualization of the Stage 1 latent representations learned by using Lcodec and LQP , and their combination, respectively. Color denotes codec type (AVC / HEVC / AV1 / VP9) and point size denotes QP level.
Lcodec
LQP
PSNR ↑
SSIM ↑
LPIPS ↓
NIQE ↓
MUSIQ ↑
CLIPIQA ↑
✓
23.49
0.887
0.304
3.40
63.69
0.613
✓
23.77
0.896
0.282
3.32
68.42
0.676
✓
✓
23.89
0.897
0.281
3.26
68.08
0.673
Table 8. Ablation study on Stage 1 supervision losses, Lcodec and LQP , evaluated by the final Stage 2 restoration performance. The Best and Second-best results are highlighted in red and orange, respectively.
Conditioning
Sampson ↓
MEt3R ↓
Null embedding
0.081
0.1361
Learned prior
0.078
0.1357
Table 9. Geometric ablation of the learned Stage-1 compression prior. The Best and Second-best results are highlighted in red and orange, respectively.
Conditioning
Inlier Ratio ↑
RANSAC Inliers ↑
Sampson ↓
AUC@5 ↑
MEt3R ↓
Random
0.7747
97.7
0.0960
0.5164
0.1494
Learned
0.8047
130.1
0.0806
0.5275
0.1404
Table 10. Robustness analysis of the learned compression prior against random conditioning. The Best and Second-best results are highlighted in red and orange, respectively.
Compression Prior
CAE
Lcorr
PSNR ↑
SSIM ↑
LPIPS ↓
NIQE ↓
MUSIQ ↑
CLIPIQA ↑
✓
23.79
0.893
0.305
3.16
62.16
0.595
✓
✓
23.71
0.892
0.293
3.27
67.42
0.672
✓
✓
23.82
0.896
0.287
3.39
67.68
0.665
✓
✓
23.87
0.898
0.267
3.54
67.11
0.669
✓
✓
✓
23.89
0.897
0.281
3.26
68.08
0.673
Table 11. Ablation study on Stage 2 components, including the compression prior extracted from Stage 1, the compression artifact estimator (CAE), and the cross-view correspondence loss Lcorr . The Best and Second-best results are highlighted in red and orange, respectively.
Method
Params (M) ↓
Time (s) ↓
Memory (GB) ↓
DiQP
79.373
0.422
1.02
CODiff
93.113
0.699
2.89
SIR-Diff
815.545
0.826
2.48
Ours
276.171
1.073
11.01
Table 13. Computational complexity of the compared restoration methods. We report the number of trainable parameters, forward inference time, and peak GPU memory usage.
Figure 6. Qualitative comparison on the Free dataset. The top rows correspond to an indoor scene (Lab), and the bottom rows correspond to an outdoor scene (Sky). The Lab scene is compressed using HEVC with QP 42, while the Sky scene is compressed using AV1 with QP 42. Highlighted regions reveal that our method better preserves fine details and structures under severe compression.
Figure 7. Visualization of estimated artifact maps on the Free dataset. The estimated artifact maps highlight compression-sensitive regions, such as edges and texture-rich structures, demonstrating that our CAE effectively captures spatially varying degradation cues.
Figure 8. Qualitative comparison of synthesized novel 3D views from restored multi-view inputs on the Free dataset. We evaluate global consistency using scene renderings synthesized from multiple viewpoints and local structural fidelity using zoomed-in regions. Views 1 and 2 correspond to HEVC QP42, and View 3 corresponds to AV1 QP42. Our method preserves sharper geometry-critical details and reduces blur and distortion compared with diffusion-based restoration baselines. Red boxes indicate representative regions for comparison.
Figure 9. Qualitative novel-view synthesis results rendered at ground-truth target camera poses. The top two rows correspond to HEVC QP42, and the bottom two rows correspond to AVC QP42. We compare each method with the corresponding GT view using full-scene renderings and zoomed-in regions. Our method better preserves object structures, text regions, and high-frequency details, producing results closer to GT than diffusion-based restoration baselines. Red boxes indicate representative regions for comparison.
Multi-view consistency studies how visual models keep geometry, appearance, identity, and predictions aligned across different views of the same scene.
Multi-view 3D reconstruction has achieved remarkable progress with the advent of feed-forward 3D reconstruction models. However, these models are typically trained and evaluated under ideal, degradation-free imaging conditions, whereas real-world observations often contain degradations that differ significantly from such settings. Improving robustness for multi-view 3D reconstruction under degraded conditions therefore remains an important challenge. We present Geometry-Aware Representation Denoising (GARD), a novel framework that performs diffusion-based multi-view restoration directly in the feature space of a feed-forward 3D reconstruction model. This design exploits the geometry-aware feature representations of the 3D reconstructor to effectively recover accurate scene geometry. Furthermore, by employing an additional RGB image decoder, the refined representations can also be used to restore high-quality RGB images, thereby enabling the simultaneous recovery of 3D scene geometry and high-quality imagery. Comprehensive experiments on the Depth Anything 3 (DA3) benchmark demonstrate the effectiveness of the proposed GARD framework.
Reconstructing 3D scenes from a single image is a fundamental challenge in computer vision, with broad applications in virtual reality, robotics, and content creation. Recent methods achieve outstanding performance by leveraging camera-controlled video diffusion models, but rely on iterative diffusion sampling, which greatly limits their practical deployment. We observe that geometric forward warping alone can cover the majority of a target view directly from the input image, with only a compact residual left for the encoder to correct. Motivated by this observation, we propose PRISM, a feed-forward framework that decomposes multi-view latent prediction into a parameter-free geometric prior and a learned residual correction, with no diffusion sampling required at inference. To enable generalization from purely synthetic training data, we devise a two-stage training strategy combining latents supervised distillation for geometric generalization and perceptual fine-tuning for appearance quality optimization. Extensive experiments on three benchmarks demonstrate that PRISM achieves competitive reconstruction quality compared with diffusion-based methods, while reducing inference time dramatically to only 36 seconds per scene. Our code is available at https://github.com/microjie372/PRISM.
We present SyncFix, a framework that enforces cross-view consistency during the diffusion-based refinement of reconstructed scenes. SyncFix formulates refinement as a joint latent bridge matching problem, synchronizing distorted and clean representations across multiple views to fix the semantic and geometric inconsistencies. This means SyncFix learns a joint conditional over multiple views to enforce consistency throughout the denoising trajectory. Our training is done only on image pairs, but it generalizes naturally to an arbitrary number of views during inference. Moreover, reconstruction quality improves with additional views, with diminishing returns at higher view counts. Qualitative and quantitative results demonstrate that SyncFix consistently generates high-quality reconstructions and surpasses current state-of-the-art baselines, even in the absence of clean reference images. SyncFix achieves even higher fidelity when sparse references are available.