MuRF: Unlocking the Multi-Scale Potential of Vision Foundation Models

Authors: Bocheng Zou, Mu Cai, Mark Stanley, Dingfu Lu, Yong Jae Lee Date: 2026-03-26 Paper ID: arxiv:2603.25744

Summary

This paper introduces Multi-Resolution Fusion (MuRF), a training-free inference-time strategy designed to overcome the limitations of single-scale processing in Vision Foundation Models (VFMs). MuRF leverages the complementary nature of visual information across scales by processing an input image at multiple resolutions using a frozen VFM. The resulting features from these different views are then fused to create a unified, richer representation. The authors validate MuRF’s universal applicability by successfully enhancing performance across various computer vision tasks using different VFM architectures, such as DINOv2 and SigLIP2.

Key Contributions

• Proposed MuRF, a training-free, architecture-agnostic inference-time strategy for Vision Foundation Models that fuses features derived from multi-scale processing of the input image.
• Demonstrated MuRF’s universal effectiveness across a broad spectrum of computer vision tasks using diverse VFM families, including DINOv2 and SigLIP2.
• Showcased that combining low-resolution (global semantics) and high-resolution (fine-grained detail) views via MuRF unlocks complementary inductive biases inherent in visual perception.

Limitations

The primary limitation is the increased inference cost due to processing the same image multiple times at different resolutions, which must be balanced against the performance gains.

Key Concepts

• Multi-Resolution Fusion (MuRF): A training-free, universally applicable strategy that processes an image at multiple resolutions through a frozen Vision Foundation Model (VFM) and fuses the resulting features to enhance visual representations.

Limitations

The primary limitation is the increased inference cost due to processing the same image multiple times at different resolutions, which must be balanced against the performance gains.

Links

• ArXiv Abstract
• PDF