Scholarlink

<p>Controllable human image synthesis is a fundamental visual computing problem involving non-rigid deformation, correspondence learning, texture preservation, and semantic part composition. Online fashion retail increasingly requires virtual try-on systems that can replace garments while preserving human identity and supporting flexible pose control. Existing multi-pose methods often rely on intermediate semantic-layout prediction and pixel-level garment warping, which can introduce error accumulation, pixel squeezing, texture distortion, and inaccurate alignment between appearance and target pose. This paper proposes a feature-space alignment framework for person-to-person multi-pose virtual try-on. Instead of deforming RGB pixels directly, the proposed framework estimates appearance flow in a latent feature space and uses the learned flow to warp body-part and garment-related representations . A progressive flow optimization module builds multi-scale correspondences between source and target poses by iteratively refining flow fields across a pyramid of feature resolutions. A mask-guided fusion module then assigns warped attribute features to appropriate target regions and integrates them with pose features before decoding the fused representation into the final image. Experiments on DeepFashion (8570 test pairs) and a DIOR subset (2000 pairs) show that the method improves realism and structural fidelity compared with five representative alternatives, achieving FID of 10.31 (vs. 10.45 for Slot-VTON) and SSIM of 0.834 (vs. 0.827), with ablation studies confirming the respective contributions of progressive flow optimization and mask-guided fusion, particularly in pose-changing scenarios. Complete code, pretrained models, and a reproducible evaluation pipeline are released at <a href="https://github.com/2585443/feature-space-alignment-vton">https://github.com/2585443/feature-space-alignment-vton</a>.</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

Progressive feature-space alignment for pose-controllable virtual try-on

  • 期刊论文

  • 发表日期：

    2026年7月26日

查看全文
  • Yang Yang,
  • Erwei Yin

摘要

Controllable human image synthesis is a fundamental visual computing problem involving non-rigid deformation, correspondence learning, texture preservation, and semantic part composition. Online fashion retail increasingly requires virtual try-on systems that can replace garments while preserving human identity and supporting flexible pose control. Existing multi-pose methods often rely on intermediate semantic-layout prediction and pixel-level garment warping, which can introduce error accumulation, pixel squeezing, texture distortion, and inaccurate alignment between appearance and target pose. This paper proposes a feature-space alignment framework for person-to-person multi-pose virtual try-on. Instead of deforming RGB pixels directly, the proposed framework estimates appearance flow in a latent feature space and uses the learned flow to warp body-part and garment-related representations . A progressive flow optimization module builds multi-scale correspondences between source and target poses by iteratively refining flow fields across a pyramid of feature resolutions. A mask-guided fusion module then assigns warped attribute features to appropriate target regions and integrates them with pose features before decoding the fused representation into the final image. Experiments on DeepFashion (8570 test pairs) and a DIOR subset (2000 pairs) show that the method improves realism and structural fidelity compared with five representative alternatives, achieving FID of 10.31 (vs. 10.45 for Slot-VTON) and SSIM of 0.834 (vs. 0.827), with ablation studies confirming the respective contributions of progressive flow optimization and mask-guided fusion, particularly in pose-changing scenarios. Complete code, pretrained models, and a reproducible evaluation pipeline are released at https://github.com/2585443/feature-space-alignment-vton.