Scholarlink

<p>Text-guided image style transfer edits images using natural language descriptions. Current CNN and Transformer methods struggle to balance inference speed and long-range dependency modeling. This trade-off leads to structural loss or high computational costs. We propose VisionMamba-Style (VMS), a hierarchical framework based on State Space Models (SSMs). It provides global modeling with linear computational complexity. VMS consists of four main modules. First, the Multi-Scale Pyramid Mamba (MSP-Mamba) extracts hierarchical features. Second, the Style-Conditioned Mamba (SC-Mamba) modulates parameters dynamically to inject text semantics. Third, the Spatial-Aware Conv Mamba (SAC-Mamba) applies parallel dilated convolutions to reduce spatial tearing. Fourth, the Bi-Directional Mamba (BD-Mamba) extracts multi-axis spatial features. We evaluated our model on the MS-COCO dataset. VMS achieves a CLIP Score of 0.22 and a 7.9% SSIM improvement over the ZeCon baseline. Per-instance optimization requires about 50&#xa0;s. The forward-pass inference takes 34.56 ms on a single GPU. The results demonstrate that VMS preserves image structures well during text-guided style transfer.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

VMS: hierarchical vision mamba for text-guided stylization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Shengchang Luo,
  • Zhixiong Chen,
  • Chengzhi Zhang,
  • Ming Yin,
  • Haoyuan Chen,
  • Yiru Zhang

摘要

Text-guided image style transfer edits images using natural language descriptions. Current CNN and Transformer methods struggle to balance inference speed and long-range dependency modeling. This trade-off leads to structural loss or high computational costs. We propose VisionMamba-Style (VMS), a hierarchical framework based on State Space Models (SSMs). It provides global modeling with linear computational complexity. VMS consists of four main modules. First, the Multi-Scale Pyramid Mamba (MSP-Mamba) extracts hierarchical features. Second, the Style-Conditioned Mamba (SC-Mamba) modulates parameters dynamically to inject text semantics. Third, the Spatial-Aware Conv Mamba (SAC-Mamba) applies parallel dilated convolutions to reduce spatial tearing. Fourth, the Bi-Directional Mamba (BD-Mamba) extracts multi-axis spatial features. We evaluated our model on the MS-COCO dataset. VMS achieves a CLIP Score of 0.22 and a 7.9% SSIM improvement over the ZeCon baseline. Per-instance optimization requires about 50 s. The forward-pass inference takes 34.56 ms on a single GPU. The results demonstrate that VMS preserves image structures well during text-guided style transfer.