Mask Inpaints Using Downsampled Fast Fourier Convolution

In recent years, image inpainting has seen significant advancements, with the LaMa model., introduced in 2021, standing out for its improvements in large-area inpainting. Yet, challenges such as inpainting complex geometries, processing high-resolution images swiftly, and achieving realism in filled areas persist. Existing solutions that incorporate diffusion models to enhance realism are hardware-intensive and unsuitable for high-resolution images. We propose a new image inpainting method, Downsampled Fast Fourier Convolution (DFFC), which addresses these challenges. Unlike LaMa's reliance on Fast Fourier Convolution (FFC) for early-stage global context capture without explicit downsampling, DFFC introduces a deep learning-driven downsampling module to reduce computational complexity and improve high-resolution image processing efficiency. This module allows for data volume reduction during processing while preserving global context, comprising: i) deep learning-based image downsampling, ii) an FFC-based inpainting architecture, iii) a high-perceptual-domain loss function, and iv) dynamic large-area mask training. Our method maintains the original model's performance, enhances processing speed, and reduces computational load.