Ref-GeNVS: Training-Free Reflection-Aware Generative Novel View Synthesis

Forum topic · 小凯 · 2026-09-09

Summary

Ref-GeNVS is a training-free, reflection-aware method for generative novel view synthesis (NVS) in mirror scenes, proposed by GeonU Kim, Shin Dong-Yeon, and Tae-Hyun Oh (arXiv 2609.05382). Existing multi-view diffusion models often fail to recognize mirrors and cannot exploit reflected content for scene generation. Ref-GeNVS addresses this without additional training by treating a mirror image as two complementary views: it estimates the mirror plane from input images and reflects camera poses to form virtual views. Building on this virtual-view setup, a two-stage generation method — Mirror-gated attention and Reflection injection — explicitly leverages reflection relationships within a multi-view diffusion model to achieve reflection-consistent novel view synthesis. The method inherits the strong generalizability of its multi-view diffusion backbone and requires no fine-tuning. On both synthetic and real scenes containing mirrors, Ref-GeNVS outperforms recent generative NVS methods by producing reflection-consistent, contextually coherent views, revealing scene structures visible only through the mirror.

论文概

Research Area: Computer Vision Authors: GeonU Kim, Shin Dong-Yeon, Tae-Hyun Oh Published: 2026-09-04 arXiv: 2609.05382

Introduction

We propose Ref-GeNVS, a training-free, reflection-aware method for generative novel view synthesis (NVS) in mirror scenes. Existing multi-view diffusion models often fail to recognize the mirror in the scene and cannot exploit reflected content for scene generation.

Key Idea

To fix this issue without additional training, the key idea is to treat a mirror image as two complementary views. From input images, Ref-GeNVS estimates the mirror plane and reflects camera poses to form virtual views.

Method

Based on this virtual view setup, a two-stage generation method is proposed:

1. Mirror-gated attention 2. Reflection injection

These components enable reflection-consistent NVS by explicitly leveraging reflection relationships in a multi-view diffusion model.

Results

Ref-GeNVS inherits the strong generalizability of the multi-view diffusion backbone — no fine-tuning required.
On synthetic and real scenes containing mirrors, Ref-GeNVS outperforms recent generative NVS methods by generating reflection-consistent and contextually coherent novel views.
It reveals scene structures visible only through the mirror.