Researchers found a cheap fix for one of AI-generated LiDAR's ugliest artifacts, and it doesn't require retraining the whole system.
The problem is called flying pixels. When autonomous-driving systems generate synthetic LiDAR point clouds using latent diffusion models, the decoder stage often blurs sharp depth edges. That blur produces points that back-project into empty space between surfaces, floating where nothing should be. A team describes CRISP, a pixel-space diffusion decoder built with a backbone-agnostic latent adapter, a DiT-based denoiser, and a support mask predictor. It swaps in for existing video-VAE or LiDAR-native decoders while leaving the encoder and latent generator untouched. Tested on KITTI-360, SemanticKITTI, and nuScenes, the decoder swap alone cut two standard error metrics by 50.5 percent on average, with reductions up to 74 percent on generic video VAEs and 71 percent on a LiDAR-native decoder called LiDM.
This matters because synthetic LiDAR data is increasingly used to train and test self-driving perception systems, and flying pixels are exactly the kind of artifact that teaches a model the wrong thing about where object boundaries are. A decoder-only fix is also a cheap intervention: no need to retrain the costly latent generator that actually learns scene structure. The team reports the swap even improved a pretrained world model's simulation accuracy by 15.5 percent, nudging synthetic data closer to real sensor behavior.
It's a narrow fix for a narrow problem, but narrow problems compound in safety-critical systems. One less hallucinated point floating in a parking lot is one less reason a downstream model learns to trust the wrong edge.