Scholarlink

<p>We present <b>DuoLit</b>, a prior-guided and lighting-aware framework for estimating human materials from a single image. This problem is severely ill-posed due to material-lighting ambiguity and becomes more challenging when external priors are noisy or domain-shifted, which can cause illumination leakage in albedo and error accumulation during training. Our method addresses these issues with two complementary designs. First, the Dual-level priors guidance module integrates priors at both the feature level and the pixel level: feature-level semantic alignment stabilizes global structure, while pixel-wise adaptive fusion corrects unreliable normal/albedo priors locally, improving robustness under cross-domain shifts. Second, the Gated lighting injection module injects lighting cues into the latent representation through gated fusion, enabling controlled conditioning that improves appearance–illumination disentanglement and reduces lighting baked into diffuse albedo. Extensive evaluations on OpenHumanBRDF and Synthetic CG dataset demonstrate consistent improvements over strong baselines in material estimation and relighting quality.</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

DuoLit: dual-level priors and gated lighting injection for human material estimation

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Yu Jiang,
  • Jianchi Sun,
  • Xiangqian Shen,
  • Chunxia Xiao

摘要

We present DuoLit, a prior-guided and lighting-aware framework for estimating human materials from a single image. This problem is severely ill-posed due to material-lighting ambiguity and becomes more challenging when external priors are noisy or domain-shifted, which can cause illumination leakage in albedo and error accumulation during training. Our method addresses these issues with two complementary designs. First, the Dual-level priors guidance module integrates priors at both the feature level and the pixel level: feature-level semantic alignment stabilizes global structure, while pixel-wise adaptive fusion corrects unreliable normal/albedo priors locally, improving robustness under cross-domain shifts. Second, the Gated lighting injection module injects lighting cues into the latent representation through gated fusion, enabling controlled conditioning that improves appearance–illumination disentanglement and reduces lighting baked into diffuse albedo. Extensive evaluations on OpenHumanBRDF and Synthetic CG dataset demonstrate consistent improvements over strong baselines in material estimation and relighting quality.