Reflection-aware Generative Novel View Synthesis

TL;DR

Ref-GeNVS is a training-free method for generating reflection-consistent novel views in mirror scenes.

cs.CV 🔴 Advanced 2026-09-05 93 views
GeonU Kim Shin Dong-Yeon Tae-Hyun Oh
generative view synthesis mirror reflections multi-view diffusion training-free image generation

Key Findings

Methodology

Ref-GeNVS treats mirror images as two complementary views, using multi-view diffusion with Mirror-gated attention and Reflection injection to achieve reflection-consistent novel view synthesis. It estimates mirror planes and reflects camera poses to form virtual views without additional training.

Key Results

  • In synthetic and real scenes, Ref-GeNVS outperforms existing methods in reflection consistency and contextual coherence, with PSNR improvements of about 1.5-2.5 points.
  • On real scene datasets, Ref-GeNVS achieves a CLIP similarity of 0.951, significantly outperforming baseline methods.
  • Ablation studies show that Mirror-gated attention and Reflection injection significantly enhance image quality.

Significance

Ref-GeNVS is significant for academia and industry, addressing limitations of existing methods in handling mirror reflections. It generates reflection-consistent images, revealing scene structures visible only through mirrors, applicable in VR and AR to enhance user experience.

Technical Contribution

Technically, Ref-GeNVS introduces Mirror-gated attention and Reflection injection, overcoming limitations of existing multi-view diffusion models. It leverages mirror information directly for generation without model finetuning, offering new engineering possibilities.

Novelty

Ref-GeNVS is the first to treat mirror images as two complementary views, significantly improving reflection consistency in generated images. Unlike existing methods, it leverages mirror information directly without additional training.

Limitations

  • In complex mirror scenes, mirror plane estimation may be inaccurate, affecting generation quality.
  • Handling of occlusions in mirror scenes still needs improvement.

Future Work

Future work could explore more complex scene mirror plane estimation methods and improve adaptability to dynamic mirror scenes.

AI Executive Summary

In computer vision, generating novel views of scenes, especially those with mirrors, is challenging. Existing methods often fail to correctly interpret mirrors, leading to inconsistent reflections in generated images. Ref-GeNVS addresses this by treating mirror images as two complementary views, using multi-view diffusion with Mirror-gated attention and Reflection injection to achieve reflection-consistent novel view synthesis.

Ref-GeNVS requires no additional training, estimating mirror planes and reflecting camera poses to form virtual views directly from input images. Experiments show that in synthetic and real scenes, Ref-GeNVS outperforms existing methods in reflection consistency and contextual coherence, with PSNR improvements of about 1.5-2.5 points and a CLIP similarity of 0.951.

This method has broad applications in virtual and augmented reality, enhancing user experience. Future work could explore more complex scene mirror plane estimation methods and improve adaptability to dynamic mirror scenes.

Deep Analysis

Background

Generating novel views of scenes is a key challenge in computer vision. Traditional methods like NeRF and 3DGS struggle with reflective objects, failing to leverage mirror information correctly. Recent attempts to decompose reflections and diffuse components have not fully resolved reflection consistency issues.

Core Problem

Existing methods treat mirrors as ordinary views, leading to inconsistent reflections in generated images. This issue is particularly significant in VR and AR applications, where users expect realistic and consistent reflections.

Innovation

Ref-GeNVS treats mirror images as two complementary views, using multi-view diffusion with Mirror-gated attention and Reflection injection to achieve reflection-consistent novel view synthesis. It estimates mirror planes and reflects camera poses to form virtual views without additional training.

Methodology

  • �� Estimate mirror planes from input images and reflect camera poses to form virtual views.
  • �� Use Mirror-gated attention to restrict attention to mirror regions.
  • �� Apply Reflection injection during denoising to ensure reflection consistency.

Experiments

Experiments were conducted on synthetic and real scene datasets, using metrics like PSNR, SSIM, and CLIP similarity. Compared to baselines like MVGenMaster and SEVA, Ref-GeNVS shows superior reflection consistency and contextual coherence.

Results

In synthetic and real scenes, Ref-GeNVS outperforms existing methods in reflection consistency and contextual coherence, with PSNR improvements of about 1.5-2.5 points and a CLIP similarity of 0.951.

Applications

Ref-GeNVS can be applied in VR and AR to enhance user experience. Its training-free nature offers high flexibility in practical applications.

Limitations & Outlook

In complex mirror scenes, mirror plane estimation may be inaccurate, affecting generation quality. Handling of occlusions in mirror scenes still needs improvement. Future work could explore more complex scene mirror plane estimation methods and improve adaptability to dynamic mirror scenes.

Plain Language Accessible to non-experts

Imagine you're in a room with a large mirror on the wall. Traditional image generation methods only see the reflection in the mirror, ignoring the world behind it. Ref-GeNVS is like having a magic pair of glasses that lets you see both the reflection and the real world behind the mirror. This way, we can generate more realistic and consistent images, seamlessly combining the mirrored world with reality.

ELI14 Explained like you're 14

Hey there! Imagine you're playing a super cool VR game with a big mirror in the room. Regular games might just show you the reflection, but our Ref-GeNVS is like putting on a magic pair of glasses that lets you see the secret world behind the mirror! It's like discovering a hidden level in the game, super cool, right?

Glossary

Multi-view diffusion model

A method for generating novel views by exchanging information across multiple views.

Used as the backbone model for Ref-GeNVS.

Mirror-gated attention

A technique that restricts attention to mirror regions.

Ensures reflection consistency.

Reflection injection

A technique that injects features from reflected views during denoising.

Used to generate reflection-consistent images.

PSNR

Peak Signal-to-Noise Ratio, a metric for measuring image quality.

Used to evaluate the quality of generated images.

CLIP similarity

A metric for measuring the similarity between images and text.

Used to evaluate contextual coherence of generated images.

Open Questions Unanswered questions from this research

  • 1 How to accurately estimate mirror planes in dynamic scenes? Current methods perform poorly in dynamic settings, requiring advanced estimation techniques.
  • 2 How to improve adaptability to complex mirror scenes? Existing methods may introduce errors when handling complex scenes.

Applications

Immediate Applications

Virtual Reality

Can be used to generate more realistic VR scenes, enhancing user experience. No additional training required, highly adaptable.

Long-term Vision

Augmented Reality

Achieve more realistic reflection effects in AR applications, potentially transforming user interactions.

Abstract

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. To fix this issue without additional training, our key idea is to treat a mirror image as two complementary views. From input images, we estimate the mirror plane and reflect camera poses to form virtual views. Based on this virtual view setup, we propose a two-stage generation method consisting of Mirror-gated attention and Reflection injection, which enables reflection-consistent NVS by explicitly leveraging reflection relationships in a multi-view diffusion model. Ref-GeNVS inherits the strong generalizability of the multi-view diffusion backbone, while it does not require finetuning. On synthetic and real scenes including mirrors, Ref-GeNVS outperforms recent generative NVS methods by generating reflection-consistent and contextually coherent novel views, revealing scene structure visible only through mirrors. Project page: https://kim-geonu.github.io/Ref-GeNVS/

cs.CV cs.AI