Reflection-aware Generative Novel View Synthesis
Ref-GeNVS is a training-free, reflection-aware method for generative novel view synthesis in mirror scenes. It treats a mirror image as two complementary views, estimates the mirror plane, reflects camera poses to form virtual views, and uses a two-stage generation method with Mirror-gated attention and Reflection injection. It outperforms recent generative NVS methods on synthetic and real scenes including mirrors.
Researchers propose Ref-GeNVS, a training-free method for generative novel view synthesis that explicitly handles mirrors. By estimating the mirror plane and reflecting camera poses, it creates virtual views and uses Mirror-gated attention and Reflection injection to generate reflection-consistent novel views. The method inherits the generalizability of multi-view diffusion models without finetuning and outperforms recent generative NVS methods on synthetic and real scenes with mirrors.
The key technical contribution is treating a mirror image as two complementary views and injecting reflection relationships into a multi-view diffusion model via Mirror-gated attention and Reflection injection. This allows the model to exploit reflected content for scene generation without additional training, improving reflection consistency and contextual coherence.
This work addresses a known limitation of multi-view diffusion models in handling reflective surfaces, which is relevant for applications in 3D reconstruction, AR/VR, and robotics where mirrors are common. The training-free approach lowers adoption barriers.
Potential business value lies in improving 3D scene generation and reconstruction for industries such as real estate, interior design, autonomous driving, and virtual production, where accurate handling of mirrors can reduce artifacts and improve user experience.
Next signals include whether the method is adopted in downstream 3D reconstruction pipelines, whether it extends to other reflective surfaces (e.g., windows, water), and whether similar reflection-aware techniques appear in commercial generative models.