Abstract
Diffusion models have shown remarkable flexibility for solving inverse problems without task-specific retraining. However, existing approaches such as Manifold Preserving Guided Diffusion (MPGD) apply only a single gradient update per denoising step, limiting restoration fidelity and robustness, especially in embedded or out-of-distribution settings. In this work, we introduce a multistep optimization strategy within each denoising timestep, significantly enhancing image quality, perceptual accuracy, and generalization. Our experiments on super-resolution and Gaussian deblurring demonstrate that increasing the number of gradient updates per step improves LPIPS and PSNR with minimal latency overhead. Notably, we validate this approach on a Jetson Orin Nano using degraded ImageNet and a UAV dataset, showing that MPGD, originally trained on face datasets, generalizes effectively to natural and aerial scenes. Our findings highlight MPGD's potential as a lightweight, plug-and-play restoration module for real-time visual perception in embodied AI agents such as drones and mobile robots.
Abstract (translated)
扩散模型在解决逆向问题时表现出惊人的灵活性,无需针对特定任务进行重新训练。然而,现有的方法如保流形引导扩散(MPGD)仅在每次去噪步骤中应用一次梯度更新,这限制了恢复的精度和鲁棒性,尤其是在嵌入式或分布外设置中更为明显。在这项工作中,我们介绍了一种在每个去噪时间步内进行多步优化策略的方法,从而显著提高了图像质量、感知准确性和泛化能力。我们在超分辨率和高斯模糊去除上的实验表明,增加每一步的梯度更新次数可以提高LPIPS(Learned Perceptual Image Patch Similarity)和PSNR(峰值信噪比),同时仅增加了微小的时间延迟开销。 值得注意的是,我们使用Jetson Orin Nano设备在降级后的ImageNet数据集和无人机数据集上验证了这一方法的有效性。实验表明,最初在面部数据集中训练的MPGD模型可以有效地推广到自然场景和空中图像中。我们的研究结果强调了MPGD作为轻量级、即插即用恢复模块用于实时视觉感知的潜力,特别是在如无人机和移动机器人等具身人工智能代理中的应用。
URL
https://arxiv.org/abs/2506.07286