Scholarlink

Neural Style Transfer (NST) enables artistic image synthesis, yet controllable multi-style blending remains under-specified in most pipelines. We introduce a hybrid framework that explicitly parameterizes each style’s contribution via normalized weights in the optimization objective, and uses a style-conditioned generator to provide a strong initialization for refinement. A lightweight encoder maps arbitrary styles to compact 256-dimensional codes. The multi-style loss aligns each style’s Gram statistics according to user weights, yielding predictable control over mixtures. We evaluate on MS-COCO 2017 (content) and WikiArt (style). At \(512\times 512\) with 250 L-BFGS steps on MPS (35 runs), we report mean ± std: PSNR \(10.252\pm 0.934\) dB, SSIM \(0.259\pm 0.057\) , LPIPS \(0.515\pm 0.058\) , and runtime \(28.182\pm 0.739\) s. Weight sweeps confirm smooth controllability across blends. The resulting framework provides an interpretable, optimization-grounded foundation for controllable multi-style transfer and interactive creative synthesis.

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

Adaptive Multi-style Transfer with Hybrid Neural Encoding

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jay Vinit Lunia,
  • Saad B. Ahmed

摘要

Neural Style Transfer (NST) enables artistic image synthesis, yet controllable multi-style blending remains under-specified in most pipelines. We introduce a hybrid framework that explicitly parameterizes each style’s contribution via normalized weights in the optimization objective, and uses a style-conditioned generator to provide a strong initialization for refinement. A lightweight encoder maps arbitrary styles to compact 256-dimensional codes. The multi-style loss aligns each style’s Gram statistics according to user weights, yielding predictable control over mixtures. We evaluate on MS-COCO 2017 (content) and WikiArt (style). At \(512\times 512\) with 250 L-BFGS steps on MPS (35 runs), we report mean ± std: PSNR \(10.252\pm 0.934\) dB, SSIM \(0.259\pm 0.057\) , LPIPS \(0.515\pm 0.058\) , and runtime \(28.182\pm 0.739\) s. Weight sweeps confirm smooth controllability across blends. The resulting framework provides an interpretable, optimization-grounded foundation for controllable multi-style transfer and interactive creative synthesis.