Scholarlink

<p>Monte Carlo denoising plays a pivotal role in enhancing image quality in physically based rendering. Despite the widespread adoption of learning-based methods, achieving high denoising accuracy while minimizing GPU memory usage remains a formidable challenge. In this paper, we introduce Denoising Mamba, a novel approach that leverages a memory-efficient state-space model architecture, circumventing the limitations of attention-based and diffusion-based methods. Our framework incorporates an optimized Fast Fourier Convolution Extractor for seamless local–global feature extraction and a Multi-Scale Residual Multiway Mamba Encodethat captures long-range spatial dependencies through linear-complexity multi-directional scanning with progressively increasing kernel sizes. This culminates in a lightweight decoder for reconstructing high-resolution, denoised images. Compared to mainstream CNN-base), attention-based, and diffusion-based learning methods, experimental results on real datasets demonstrate that our method achieves superior quantitative performance while requiring low computational resources. Our code and datasets are publicly available at [<a href="https://github.com/YUJIAJING0408/DenoisingMamba">https://github.com/YUJIAJING0408/DenoisingMamba</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

Efficient Monte Carlo denoising via state-space model with low GPU memory overhead

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Rui Zhu,
  • Jiajing Yu,
  • Tongzhou Zhao

摘要

Monte Carlo denoising plays a pivotal role in enhancing image quality in physically based rendering. Despite the widespread adoption of learning-based methods, achieving high denoising accuracy while minimizing GPU memory usage remains a formidable challenge. In this paper, we introduce Denoising Mamba, a novel approach that leverages a memory-efficient state-space model architecture, circumventing the limitations of attention-based and diffusion-based methods. Our framework incorporates an optimized Fast Fourier Convolution Extractor for seamless local–global feature extraction and a Multi-Scale Residual Multiway Mamba Encodethat captures long-range spatial dependencies through linear-complexity multi-directional scanning with progressively increasing kernel sizes. This culminates in a lightweight decoder for reconstructing high-resolution, denoised images. Compared to mainstream CNN-base), attention-based, and diffusion-based learning methods, experimental results on real datasets demonstrate that our method achieves superior quantitative performance while requiring low computational resources. Our code and datasets are publicly available at [https://github.com/YUJIAJING0408/DenoisingMamba].