Low-Noise Direct Denoising Sampling for DPM++ Samplers with Denoised Outputs

Diffusion models have been widely used in image generation and structured trajectory generation. DPM++ samplers improve few-step inference by recursively propagating current and historical denoised states, but their behavior in the low-noise regime remains less examined. This paper studies whether the recursive low-noise update is always necessary when the model-predicted denoised output has become relatively stable. Based on this analysis, we propose Low-Noise Direct Denoising Sampling (LNDDS), a training-free modification that replaces the recursive DPM++ update with the current denoised prediction during selected low-noise steps. LNDDS keeps the model architecture, training objective, noise schedule, number of sampling steps, and number of function evaluations unchanged. Experiments on MJHQ-30K image generation and HiLo-nuPlan traffic scenario generation show that the effect of LNDDS is domain-dependent. LNDDS does not provide consistent quality improvements for image generation under the evaluated setting, but it shows clear benefits for traffic scenario generation by improving trajectory quality, motion smoothness, and physical validity. Mechanism analysis further shows that low-noise direct updates affect sampling trajectories differently across domains. These results suggest that LNDDS is a task-dependent low-noise state-update strategy that is more suitable for structured trajectory generation than for general text-to-image generation.