AuraTracer智迹闻
中文

EVENT DOSSIER

Reflection-aware Generative Novel View Synthesis

2026-09-07 12:00 Models across 2 days 🔥 47.2 heat score
2sources
2days unfolding
47.2heat score
1mentions
SummaryAI generated

To address the shortcomings of multi-view diffusion models in recognizing mirrors and utilizing reflected content in mirror scenarios, researchers proposed a new generative synthesis method called Ref-GeNVS. This method treats mirrors as two complementary views, constructs virtual views by estimating the mirror surface and reflecting the camera pose, and then proposes a two-stage generation strategy involving mirror-gated attention and reflection injection. Ref-GeNVS inherits the strong generalization ability of multi-view diffusion models without the need for fine-tuning. In synthetic and real mirror scenario tests, its generated new views outperform recent methods in terms of reflection consistency and context coherence, and can reveal the scene structure visible only through mirrors.

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
Ref-GeNVS

Integrated timelineUNIFIED TIMELINE

  1. 2026-09-04

    Reflection-aware Generative Novel View …

    提出了一种名为 Ref-GeNVS 的无需训练、具备反射感知的生成式新视角合成方法,用于解决镜像场景中现有多视图扩散模型无法识别镜子及利用反射内容的缺陷。该方法将镜面图像视为两个互补视图,通过估算镜平面并反射相机姿态构建虚拟视图,进而提出…

  2. 2026-09-07

    Reflection-aware Generative Novel View …

    提出了一种名为 Ref-GeNVS 的无需训练、具备反射感知的生成式新视角合成(NVS)方法,用于解决多视图扩散模型在镜场景下无法识别镜子及利用反射内容的缺陷。该方法将镜像视为两个互补视图,通过估计镜面并反射相机姿态构建虚拟视图,进而提出…

SignalsSIGNALS

Keyword heat
  • Ref-GeNVS2

All reports (2)SOURCES

A arXiv cs.CV en 2026-09-05 01:33

Reflection-aware Generative Novel View Synthesis

提出了一种名为 Ref-GeNVS 的无需训练、具备反射感知的生成式新视角合成方法,用于解决镜像场景中现有多视图扩散模型无法识别镜子及利用反射内容的缺陷。该方法将镜面图像视为两个互补视图,通过估算镜平面并反射相机姿态构建虚拟视图,进而提出包含镜面门控注意力与反射注入的两阶段生成策略,以显式利用多视图扩散模型中的反射关系实现反射一致的新视角合成。Ref-GeNVS 继承了多视图扩散骨干的强泛化能力且无需微调,在包含镜子的合成及真实场景中表现优于近期生成式新视角合成方法,能够生成反射一致且上下文连贯的新视角,揭示仅通过镜子可见的场景结构。

A arXiv cs.AI en 2026-09-07 12:00

Reflection-aware Generative Novel View Synthesis

提出了一种名为 Ref-GeNVS 的无需训练、具备反射感知的生成式新视角合成(NVS)方法,用于解决多视图扩散模型在镜场景下无法识别镜子及利用反射内容的缺陷。该方法将镜像视为两个互补视图,通过估计镜面并反射相机姿态构建虚拟视图,进而提出包含镜像门控注意力与反射注入的两阶段生成策略。Ref-GeNVS 继承了多视图扩散骨干的强泛化能力且无需微调,在合成及真实镜场景测试中,其生成的新视角在反射一致性和上下文连贯性上优于近期方法,并能揭示仅通过镜子可见的场景结构。