RealFill is a novel generative approach for image completion that fills in missing regions of an image with content that should have been there. Unlike previous models, RealFill is personalized using only a few reference images of a scene, allowing for drastically varying viewpoints, lighting conditions, camera apertures, or image styles. The personalized model is created through fine-tuning a pre-trained inpainting diffusion model on the reference and target images. RealFill outperforms existing approaches on a diverse and challenging image completion benchmark. However, it does have limitations, such as being relatively slow and struggling with large viewpoint changes or challenging cases for the base model.
https://realfill.github.io/