This paper presents "Efficient Virtuoso," an efficient method for generating diverse and plausible future path distributions for autonomous vehicle planning systems. This method utilizes a goal-conditional latent diffusion model (LDM) that maintains geometric aspect ratio and ensures stable training targets through a two-stage regularization pipeline. It performs efficient denoising using a simple MLP denoiser in a low-dimensional latent space, and conditions rich scene context using a Transformer-based StateEncoder. It achieves state-of-the-art performance (minADE 0.25) on the Waymo Open Motion Dataset. An ablation study of the goal representation demonstrates that while a single-endpoint objective resolves strategic ambiguity, multi-stage sparse routes enable precise and high-fidelity tactical execution.
