Adaptive Noise-Powered Diffusion Model for Efficient and Accurate Object Detection

Recent advancements in object detection, particularly with DiffusionDet, have demonstrated impressive performance. However, its reliance on numerous random noise-based object candidates limits its efficiency. To overcome this limitation, we propose DifAda, a novel object detection model that incorporates adaptive noise into the diffusion framework. DifAda employs an adaptive noise mechanism that blends random noise with latent codes, enhancing object candidate utilization and improving object localization. Our model features a simplified decoder structure by employing a single decoder layer and utilizes an adaptive interaction mechanism to further refine feature representations, leading to improved performance with fewer parameters. In addition, DifAda supports flexible speed–accuracy trade-offs through adjustable sampling and iteration steps, requiring no retraining. Experimental results across multiple benchmarks, including COCO and RUOD, demonstrate that DifAda achieves competitive performance with significantly fewer object candidates and parameters. Our findings suggest that DifAda represents a step forward in efficient and scalable object detection.