Ref-GeNVS: Training-Free Reflection-Aware Generative Novel View Synthesis
Original title:Reflection-aware Generative Novel View Synthesis
Multi-view diffusion models typically struggle with mirrors, treating reflections as visual noise rather than coherent spatial cues. Ref-GeNVS addresses this limitation without retraining or fine-tuning the underlying backbone. By estimating the mirror plane, it mirrors camera poses to construct virtual viewpoints, treating reflection as an explicit geometric counterpart to the primary scene. A combination of mirror-gated attention and reflection injection then guides the diffusion process, allowing the model to consistently generate novel views and recover occluded structures visible only inside the glass.
Why it's worth reading
It solves a longstanding failure mode in novel view synthesis through pure geometric reframing, offering a practical, training-free way to exploit mirror reflections in existing multi-view diffusion backbones.