CVPR '26 Highlight: Tsinghua University Creates Mirror Illusion Art Generator, Supports 3D Printing
Researchers from Tsinghua University have developed AutoMIA, an AI system that automatically generates mirror illusion art. Given two arbitrary input images, AutoMIA designs a 3D object that, when placed in front of a mirror, displays one image when viewed directly and the other as its mirror reflection. The system models this as a dual-view inverse design problem, using a voxel representation and differentiable volume rendering to optimize for both shape and color fidelity from two specified viewpoints.
Key innovations include Projection-Aligned Connectivity pruning (PAC) to remove noise, Position-Weighted Adaptive denoising (PWA), Internal Voxel Protection (IVP) to prevent structural fractures for 3D printing, and Shape-Color Decoupled optimization (SCD) to balance geometric and color learning. Experiments show AutoMIA outperforms prior methods like Shadow Art in smoothness, noise reduction, and shape consistency, while remaining computationally efficient (76 seconds avg. on an RTX 3090).
The method supports diverse inputs (letters, logos, emojis, etc.) and successfully transitions from digital simulation to physical artifacts via 3D printing, demonstrating practical applications in computational art and fabrication. This work, a CVPR 2026 Highlight and Efficient CVPR award winner, lowers the barrier for creating complex mirror illusion art.
marsbit2 days ago 07:07
