A new method that generates varied multi-object point cloud scenes using latent flow matching and a novel class-partition VQ-VAE. Accepted to AAAI, 2026