What are Inpainting and Outpainting?

Inpainting

Inpainting, in the context of image generation, is a process where we try to fill in parts of an image with new or tweaked content.

In a technical sense, inpainting methods use the available information in an image (such as edges, textures, colors) to predict what the areas should look like, and then use the selected model to regenerate the image. Invoke also then does this again over any seams between the new and existing image to ensure that the generation fits in well.

On Invokes Unified Canvas, you can Inpaint by using the Brush (B) tool on the Mask layer to define an area that you would like to have regenerated. If you invoke with a masked area, the masked area will be inpainted on Invoking.

How Inpainting works is significantly impacted by the model you have selected.

Inpainting will regenerate the portion you have selected using the same logic that Image to Image uses unless you use a specialized model - Based on your denoising strength, the image will use existing colors/forms to generate new content. If you are attempting to manually draw/add content into your work, it is recommended to use this process, rather than a specialized inpainting model.

Outpainting

Outpainting (also known as image extrapolation) is a process where we try to extend the content of an image beyond its original boundaries. It's like an artist taking a small painting and extending the scene onto a larger canvas.

Imagine you have a picture of a beautiful beach but the picture only shows a small part of it. Wouldn't it be great if you could expand the image to see more of the beach, the ocean, the horizon? That's what outpainting is for!

On Invokes Unified Canvas, you can outpaint by expanding the Bounding Box (Press V to display the Handles) and extending the box to capture an area that you would like to have generated. If you invoke with the bounding box covering an area that has no image data, the missing will be outpainted on Invoking.

Generally, it’s advisable to only outpaint with about ~25% of the image missing in the area currently selected by the bounding box, and to keep the bounding box to no more than ~768x768 in size - This is because the models are typically trained on smaller images, and need sufficient examples of what are already in the image to generate something useful outside of it.

-----

You can learn more about other features of the Unified Canvas by clicking here.