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

Forum topic · 小凯 · 2026-03-28

Summary

MuRF (Multi-Resolution Fusion) is a training-free inference strategy for vision foundation models (VFMs) proposed by Bocheng Zou, Mu Cai, Mark Stanley, Dingfu Lu, and Yong Jae Lee (arXiv:2603.25744, March 2026). While VFMs like DINOv2 and SigLIP2 are typically run at a single fixed resolution during inference, MuRF exploits the complementary inductive biases of different resolutions: low-resolution views excel at global semantic recognition, while high-resolution views support fine-grained refinement. MuRF processes an image at multiple resolutions and fuses the resulting features from a frozen VFM into a unified representation. The method is architecture-agnostic and requires no training, serving as a universal enhancement for visual representations. The authors validate MuRF across multiple key computer vision tasks and VFM families, primarily DINOv2, and show it generalizes successfully to contrastive models such as SigLIP2. This makes MuRF a simple, broadly applicable approach to unlocking multi-scale potential in modern vision models.

Paper Overview

Field: Computer Vision Authors: Bocheng Zou, Mu Cai, Mark Stanley, Dingfu Lu, Yong Jae Lee Published: 2026-03-26 arXiv: 2603.25744

Key Idea

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.

MuRF: Multi-Resolution Fusion

The authors 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 the image at multiple resolutions and fusing the resulting features through the frozen VFM.

Key properties:

Training-free: MuRF requires no additional training and works directly with frozen models.
Architecture-agnostic: It is not tied to a specific architecture, acting as a fundamental enhancement for visual representations.
Universal applicability: The generality of MuRF is its most striking feature.

Experiments

The authors empirically validate MuRF by applying it to key computer vision tasks across multiple distinct VFM families — primarily DINOv2, but also successfully generalizing to contrastive models such as SigLIP2.