MirrorWorld: Taming Video Diffusion Models for Mirror Reflection Generation
The paper targets a specific failure mode in video generation: mirrors that do not match the scene around them.
MirrorWorld is presented as a reflection-aware video inpainting framework for keeping mirror content semantically and spatially consistent. Its SRD component models what visible scene content should appear in the mirror, while GTA guides how that reflected content should be arranged. The authors also build a benchmark from four existing video mirror datasets and report better reconstruction quality than image-based reflection methods and strong video inpainting baselines. HF Daily Papers' note
MirrorWorld is presented as a reflection-aware video inpainting framework for keeping mirror content semantically and spatially consistent. Its SRD component models what visible scene content should appear in the mirror, while GTA guides how that reflected content should be arranged. The authors also build a benchmark from four existing video mirror datasets and report better reconstruction quality than image-based reflection methods and strong video inpainting baselines. HF Daily Papers' note
score 4