Mirror reflection generation requires modeling two distinct challenges—semantic consistency (what reflects) and geometric accuracy (how it's arranged)—which can be addressed by distilling relational knowledge and learning spatial transformations.
MirrorWorld tackles the problem of generating realistic mirror reflections in videos by teaching diffusion models to understand what scene content should appear in mirrors and how it should be spatially arranged. The method uses semantic guidance from visual foundation models and geometric transformation learning to ensure reflections are consistent with their surroundings.