RealFill is a generative image inpainting model that can fill in missing areas of images using a small number of reference images from the same scene, generating visually consistent content with the original scene. RealFill creates personalized generative models by fine-tuning a pre-trained image inpainting diffusion model on both reference and target images. The model not only retains good image priors but also learns the content, illumination, and style of the input image. Subsequently, the fine-tuned model is used to fill in the missing regions of the target image through a standard diffusion sampling process. RealFill has been evaluated on a new image inpainting benchmark containing diverse complex scenes, demonstrating significantly superior performance compared to existing methods.