CRISP Decoder Reduces Flying Pixels in Latent LiDAR Generation
October 9, 2026
CRISP introduces a pixel-space diffusion decoder for Latent LiDAR pipelines to correct radial depth discontinuities. Replacing standard decoders with this DiT-based approach reduces FSVD/FPVD by 50.5% to 74% across major datasets like nuScenes and KITTI-360.
HOW THIS AFFECTS YOU
●
builderYou can improve the fidelity of simulated LiDAR data for training autonomous driving perception stacks.
●
researcherYou can implement diffusion-based decoding to resolve edge artifacts in latent-space generative models.