Vision Foundation Models (VFMs) have become the cornerstone of modern computer vision, offering robust representations across a wide array of tasks. While recent advances allow these models to handle varying input sizes during training, inference typically remains restricted to a single, fixed scale. This prevalent single-scale paradigm overlooks a fundamental property of visual perception: varying resolutions offer complementary inductive biases, where low-resolution views excel at global semantic recognition and high-resolution views are essential for fine-grained refinement. In this work, we propose Multi-Resolution Fusion (MuRF), a simple yet universally effective strategy to harness this synergy at inference time. Instead of relying on a single view, MuRF constructs a unified representation by processing an image at multiple resolutions through a frozen VFM and fusing the resulting features. The universality of MuRF is its most compelling attribute. It is not tied to a specific architecture, serving instead as a fundamental, training-free enhancement to visual representation. We empirically validate this by applying MuRF to a broad spectrum of critical computer vision tasks across multiple distinct VFM families - primarily DINOv2, but also demonstrating successful generalization to contrastive models like SigLIP2.
Figures & tables
Figure 1: The “Recognition vs. Refinement” Dynamic. The feature map obtained when the input is resized to 266, 518, 784. At lower resolutions, the representation is globally coherent, enabling robust recognition . At higher resolutions, boundary details are sharper, enabling precise refinement , but the object’s interior becomes noisy, risking incomplete segmentation. Our work is motivated by synergizing these two roles.
Figure 2: Overview of Mu lti- R esolution F usion ( MuRF ). An input image is resized to multiple resolutions and each view is processed by a frozen DINOv2 encoder to produce separate feature maps. These features are upsampled to a shared spatial resolution and fused into a single multi-resolution representation, which can then be used by lightweight task-specific heads for semantic segmentation, depth estimation, visual question answering, visual grounding, and other downstream tasks. Background is removed from the PCA figures.
Semantic Segmentation ↑
Depth Estimation ↓
Method
Arch.
ADE20K
PASCAL VOC
NYU Depth V2
SUN RGB-D
lin. 1
lin. 4
lin. 1
lin. 4
OpenCLIP ( Cherti et al., 2023 )
ViT-G/14
39.3
71.4
0.541
0.510
0.537
0.476
MAE ( He et al., 2022 )
ViT-H/14
33.3
67.6
0.517
0.483
0.545
0.523
DINO ( Caron et al., 2021 )
ViT-B/8
31.8
66.4
0.555
0.539
0.553
0.541
iBOT ( Zhou et al., 2022 )
ViT-L/16
44.6
82.3
0.417
0.387
0.447
0.435
Table 1: Combined downstream evaluation. Semantic segmentation: performance in mIoU (%) on ADE20K and PASCAL VOC (higher is better). Depth estimation: performance on NYU Depth V2 and SUN RGB-D reporting RMSE (lower is better). MuRF significantly outperforms single-scale baselines, highlighting the benefit of our multi-scale representation for dense prediction and geometric reasoning. Bold indicates the best performance, and underline indicates the second best. "Rel. Improv" measures the relative improvement over original DINOv2 (high resolution for semantic segmentation, and medium resolution for depth estimation)
Figure 3: Qualitative comparison of semantic segmentation results on ADE20K (top) and PASCAL VOC (bottom) with different input resolutions. All images are resized to a square shape before being fed into DINOv2, and the subtitle above each image indicates the corresponding input resolution (side length in pixels).
Method
Resolution
Latency (ms/iter)
Head Params (M)
VRAM (GB)
RMSE ↓
Single-Scale
0.5×
11.36
0.39
0.36
0.423
Single-Scale
1.0×
22.61
0.39
0.43
0.389
Single-Scale
1.5×
32.55
0.39
0.54
0.394
MuRF (Ours)
{0.5,1.0,1.5}×
58.35
1.18
0.56
0.361
Table 2: Computational Cost and Efficiency Analysis on DINOv2-ViT-B/14. We compare the single-scale baselines against our multi-scale MuRF in terms of training latency, downstream head parameter count, and performance (RMSE on NYU Depth V2). MuRF achieves the best performance with a reasonable trade-off in computation.
Figure 4: Qualitative depth estimation results on NYUd (left) and SUN RGB-D (right). We compare single-scale DINOv2 predictions at 0.5×, 1.0×, and 1.5× input resolutions with our MuRF fusion. By aggregating multi-resolution features, MuRF better preserves global scene structure while sharpening local geometry, producing smoother and more accurate depth maps. Labels 0.X× indicate that the image fed into DINOv2 is resized to 0.X of the original image height and width.
Vision Encoder
Res.
MME
Bias
V ∗
RW
MR
GQA
MMB
POPE
Percept.
Cogn.
CLIP (official LLaVA 1.5)
336
1511.4
347.1
16.2
50.3
56.1
26.5
62.0
195.4
86.9
DINOv2
336
1291.6
278.6
17.3
38.7
52.3
26.2
62.1
172.4
87.1
224+336 ( Ours )
1357.1 [-1pt] (+65.5)
366.4 [-1pt] (+87.8)
17.7 [-1pt] (+0.4)
40.3 [-1pt] (+1.6)
53.6 [-1pt] (+1.3)
26.1 [-1pt] (-0.1)
62.4 [-1pt] (+0.3)
173.1 [-1pt] (+0.7)
87.1 [-1pt] (0.0)
Clip+DINOv2
336
1403.4
243.9
15.8
48.7
53.6
31.5
62.2
194.2
86.4
224+336 ( Ours )
1471.2 [-1pt] (+67.8)
281.4 [-1pt] (+37.5)
16.3 [-1pt] (+0.5)
48.2 [-1pt] (-0.5)
56.7 [-1pt] (+3.1)
31.8 [-1pt] (+0.3)
62.9 [-1pt] (+0.7)
198.8 [-1pt] (+4.6)
87.4 [-1pt] (+1.0)
Table 3: LLaVA-MuRF VQA Performance on MME ( Fu et al., 2025 ) , VLMsAreBiased (Bias) ( Vo et al., 2026 ) , V ∗ ( Wu and Xie, 2024 ) , MME RealWorld (MR) ( Zhang et al., 2025b ) , RealWorld QA (RW) ( xAI, 2024 ) , GQA ( Hudson and Manning, 2019 ) , the total score obtained on four subset of mmbench (MMB) ( Liu et al., 2025 ) ("cn_cc", "cn_dev", "en_dev", "ru_dev"), and the accuracy metric of POPE ( Li et al., 2023 ) . Bold indicates the best performance.
Method
Training?
TEST priv
TEST priv,mix
PatchCore ( Roth et al., 2022 )
✗
62.3
52.6
SuperAD ( Zhang et al., 2025a )
✗
61.2
59.3
RoBiS ( Li et al., 2025 )
✓
67.3
59.7
MuRF (Ours)
✗
66.0
62.3 ↑ +2.6
Table 4: Anomaly detection performance (AU-PRO 0.05 in %) on the MVTec AD 2 dataset. MuRF demonstrates state-of-the-art results on TEST priv,mix subset, showcasing its robustness in a challenging training-free scenario. "Training" means if the method involves parameter tuning within a neural network. Bold indicates the best performance, and underline indicates the second best.
Figure 9Table 10Table 11
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Additional segmentation visualizations from ADE20K and PASCAL VOC.
Figure 8: Additional visualizations of depth estimation results on NYUd and SUN RGB-D.
Object
Original Resolution
# of Pixels
New Resolution
New # of Pixels
Can
1024x2232
2,285,568
1536x3348
5,142,528
Fruit Jelly
1520x2100
3,192,000
1900x2625
4,987,500
Vial
1900x1400
2,660,000
2470x1820
4,495,400
(… other categories with no change are omitted for brevity …)
Appendix
Table 9: Resolution conversion table for MVTec AD v2 categories to standardize the total number of pixels for our experiments.
Swiss Federal Institute of Technology Lausanne, Switzerland · Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI) Technische Universität Dresden, Germany