UniDBO: A Unified Dual-Branch One-Step Denoising Framework for Autonomous Driving Scenario Generation

Scenario generation is essential for training, testing, and safety validation of autonomous vehicles (AVs), especially when real-world data coverage is limited and long-tail safety-critical events are rare. Existing methods face a three-way trade-off among open-loop prediction accuracy, closed-loop simulation robustness, and inference efficiency. Autoregressive methods are typically efficient and strong in open-loop fitting, but they are prone to error accumulation and policy-induced state-distribution shift in long-horizon closed-loop rollouts. Diffusion-based methods provide strong multimodal behavior modeling and competitive generation quality, but iterative denoising substantially increases inference latency and limits simulation throughput. To address this challenge, we propose UniDBO, a Unified Dual-Branch One-step denoising framework for autonomous driving scenario generation. UniDBO uses a shared scene encoder and two jointly optimized complementary branches with distinct roles: a continuous-scale branch (CS) for open-loop trajectory fitting and a discrete high-noise branch (DHN) for robust closed-loop simulation. Through joint training, UniDBO coordinates open-loop fitting and closed-loop robustness while preserving one-step inference efficiency. We evaluate UniDBO on multiple benchmarks, including in-domain open-loop and closed-loop evaluation, zero-shot cross-domain generalization, and system-level closed-loop AV testing in unseen scenarios. Results show that UniDBO improves the open-loop/closed-loop balance while maintaining high inference efficiency, and achieves competitive zero-shot generalization on the evaluated benchmarks.