English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

MuRF: Unlocking Multi-Scale Potential of Vision Foundation Models via Inference-Time Multi-Resolution Fusion

Forum topic · 小凯 · 2026-03-28

Summary

MuRF (Multi-Resolution Fusion) is a training-free, architecture-agnostic strategy that improves vision foundation models (VFMs) at inference time. While VFMs like DINOv2 and SigLIP2 provide robust visual representations, inference typically uses a single fixed input scale, ignoring the fact that low-resolution views favor global semantic recognition while high-resolution views enable fine-grained detail. MuRF instead processes each image at multiple resolutions and fuses the resulting features from the frozen VFM into a unified representation. The authors, Bocheng Zou, Mu Cai, Mark Stanley, Dingfu Lu, and Yong Jae Lee, demonstrate that this simple enhancement yields consistent gains across key computer vision tasks and generalizes across VFM families, primarily DINOv2 and contrastive models such as SigLIP2, without any retraining. Paper: arXiv 2603.25744.

Paper Overview

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

Summary

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, 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 multiple resolutions of an image and fusing the resulting features through the frozen VFM.

MuRF's universality is its most compelling property — it is not tied to a specific architecture and serves as a fundamental, training-free enhancement for visual representations. The authors empirically validate this by applying MuRF to key computer vision tasks across multiple distinct VFM families — primarily DINOv2, but it also successfully generalizes to contrastive models such as SigLIP2.

---

*Auto-collected on 2026-03-28*

Tags

#computer-vision#vision-foundation-models#multi-resolution-fusion#inference-time#dinov2#siglip2#arxiv#paper

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177169363