Skip to content

RealFill

Fill missing image regions with reference photos

Visit official site

What RealFill is

The model is described as a generative image completion model that fills missing regions using a few reference images of a scene and produces content matching the original scene. The material also mentions an evaluation on a new image completion benchmark covering many complex scenes.

What you can do with it

  • Fill missing regions with content matching the reference images
  • Repair missing parts of an image
  • Generate visual content consistent with the original scene

How it works

  • Fine-tune a pretrained image completion diffusion model on reference and target images
  • Create a personalised generative model that keeps the image prior
  • Learn content, lighting and style from the input images
  • Fill the target image through a standard diffusion sampling process

Pros & cons

✓ What we like

  • The fine-tuned model keeps the image prior while learning content, lighting and style
  • Completion runs through a standard diffusion sampling process

! What to watch out for

  • Fine-tuning needs both reference images and a target image
  • The workflow relies on a pretrained image completion diffusion model

FAQ

What is used to fill the gaps?

A few reference images of the scene and a fine-tuned pretrained image completion diffusion model.

What does the model learn from the inputs?

The material states it learns content, lighting and style while keeping a good image prior.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review