Scholarlink

<p>Facial albedo generation is a key component of relightable 3D face reconstruction from a single image. Existing methods either rely on low-dimensional texture models that miss high-frequency details or employ high-capacity image generation pipelines that may entangle facial reflectance with illumination effects. To address this gap, we propose an end-to-end coarse-to-fine framework for UV albedo generation. Our method first uses a UV Albedo Parametric Model (UVAPM), driven by low-dimensional coefficients, to generate coarse albedo maps with stable skin tones and low-frequency texture structure. It then refines the coarse result with a detail generator trained on decoupled albedo residuals to recover local facial details. Rather than targeting unconstrained image synthesis, the proposed framework is designed for controllable albedo estimation that is compatible with differentiable 3D face reconstruction. Experiments on three datasets show competitive qualitative and quantitative performance and improved recovery of facial details such as beards, wrinkles, and spots. The code and pre-trained model are publicly available at <a href="https://github.com/MVIC-DAI/UVAPM">https://github.com/MVIC-DAI/UVAPM</a> to support reproducibility and further research.</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

Facial albedo generation for 3D face reconstruction from a single image via a coarse-to-fine approach

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Jiashu Dai,
  • Along Wang,
  • Binfan Ni,
  • Tao Cao

摘要

Facial albedo generation is a key component of relightable 3D face reconstruction from a single image. Existing methods either rely on low-dimensional texture models that miss high-frequency details or employ high-capacity image generation pipelines that may entangle facial reflectance with illumination effects. To address this gap, we propose an end-to-end coarse-to-fine framework for UV albedo generation. Our method first uses a UV Albedo Parametric Model (UVAPM), driven by low-dimensional coefficients, to generate coarse albedo maps with stable skin tones and low-frequency texture structure. It then refines the coarse result with a detail generator trained on decoupled albedo residuals to recover local facial details. Rather than targeting unconstrained image synthesis, the proposed framework is designed for controllable albedo estimation that is compatible with differentiable 3D face reconstruction. Experiments on three datasets show competitive qualitative and quantitative performance and improved recovery of facial details such as beards, wrinkles, and spots. The code and pre-trained model are publicly available at https://github.com/MVIC-DAI/UVAPM to support reproducibility and further research.