Fast image inpainting using accelerated diffusion models

Abstract

Image inpainting has gained increasing attention due to its crucial role in editing and restoring visual content. Despite significant advances through diffusion-based mod- els such as Stable Diffusion, these methods remain computationally expensive, often requiring dozens of iterative denoising steps to produce high-quality results. In this paper, we explore how Latent Consistency Models (LCM) and Low-Rank Adaptation (LoRA) can be adapted from text-to-image pipelines to an inpainting setting, aiming to reduce the number of inference steps while maintaining fidelity. We first review the evolution of inpainting methods, from classical patch-based and PDE approaches to modern diffusion-based pipelines, and highlight the computational overhead asso- ciated with large-scale generation. We then detail our methodology for integrating LCM and LoRA into pre-trained diffusion backbones, placing emphasis on masking strategies, dataset preparation, and distillation workflows. Our results show that LCM holds the potential to effectively reduce the number of inference steps required for inpainting while maintaining acceptable visual fidelity, whereas LoRA—under our constraints—yields less consistent outcomes. These findings suggest that LCM pro- vides a promising avenue for faster, high-resolution inpainting, particularly when a balance of speed and quality is paramount.