Fast and Lightweight Novel View Synthesis with Differentiable Multiplane Image
Researchers have developed a new method for novel view synthesis that aims to improve speed and reduce model size compared to existing techniques like NeRF and 3D Gaussian Splatting. Their approach revisits the Multiplane Image (MPI) representation, incorporating predicted point maps for geometric initialization and a one-step diffusion process to handle sparse-view conditions and reduce artifacts. This new method is reportedly 30.7% faster and uses 14.8% of the model size of a comparable GS-based method, while maintaining competitive quality. AI
IMPACT Offers a more efficient approach to generating novel views, potentially enabling wider deployment of 3D rendering technologies on resource-constrained devices.