Multi-Step Guided Diffusion for Image Restoration on Edge Devices: Toward Lightweight Perception in Embodied AI
This work addresses the need for lightweight, plug-and-play image restoration for real-time visual perception in embodied AI agents like drones and mobile robots, though it is incremental as it builds on an existing method.
The paper tackled the problem of limited restoration fidelity and robustness in diffusion models for image restoration on edge devices by introducing a multistep optimization strategy per denoising step, resulting in improved LPIPS and PSNR with minimal latency overhead, as validated on a Jetson Orin Nano with degraded ImageNet and a UAV dataset.
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.