Scholarlink

<p>Recent advancements in diffusion models have significantly improved the capability of scene understanding and generation. However, existing approaches still struggle to represent 3D scenes using panoramic images, mainly due to the lack of large-scale panoramic datasets and the inherent difficulty in maintaining consistency across multiple panoramic views. To address these limitations, we introduce PanoVT-HM3D, a panoramic video-text dataset comprising numerous consecutive panoramic frames paired with corresponding textual descriptions. Building upon this dataset, we propose DiffPano, a novel framework for generating multi-view panoramas from textual descriptions. Specifically, we fine-tune a text-to-panorama diffusion model and incorporate a Spherical Epipolar-Aware Module to enhance inter-view consistency. Furthermore, we extend this framework to DiffPano++, which enables image-conditioned multi-view panoramic generation. In addition, to mitigate the image quality limitations of PanoVT-HM3D that could affect model performance, we construct MPano-3D, a higher-quality multimodal panoramic dataset. Extensive experiments demonstrate that both DiffPano and DiffPano++ generate scalable, consistent, and diverse multi-view panoramas.</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

DiffPano++: Scalable and Consistent Multi-View Panorama Generation with Spherical Epipolar-Aware Diffusion

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Chenhao Ji,
  • Weicai Ye,
  • Zheng Chen,
  • Junyao Gao,
  • Xiaoshui Huang,
  • Xuekuan Wang,
  • Guofeng Zhang,
  • Songhai Zhang,
  • Tong He,
  • Wanli Ouyang,
  • Cairong Zhao

摘要

Recent advancements in diffusion models have significantly improved the capability of scene understanding and generation. However, existing approaches still struggle to represent 3D scenes using panoramic images, mainly due to the lack of large-scale panoramic datasets and the inherent difficulty in maintaining consistency across multiple panoramic views. To address these limitations, we introduce PanoVT-HM3D, a panoramic video-text dataset comprising numerous consecutive panoramic frames paired with corresponding textual descriptions. Building upon this dataset, we propose DiffPano, a novel framework for generating multi-view panoramas from textual descriptions. Specifically, we fine-tune a text-to-panorama diffusion model and incorporate a Spherical Epipolar-Aware Module to enhance inter-view consistency. Furthermore, we extend this framework to DiffPano++, which enables image-conditioned multi-view panoramic generation. In addition, to mitigate the image quality limitations of PanoVT-HM3D that could affect model performance, we construct MPano-3D, a higher-quality multimodal panoramic dataset. Extensive experiments demonstrate that both DiffPano and DiffPano++ generate scalable, consistent, and diverse multi-view panoramas.