MMOOZZGG.com

Glossary

What is outpainting

In short

Outpainting is the extension of an image beyond its borders: a neural network continues the scene to the right, left, up or down, so that the new part matches the original in style and light.

Images, video, audio · outpainting

MMOOZZGG.com editorsChecked: 1 min read

For a child

Imagine a postcard that shows only the window of a little house. You want to know what is around it, so you glue blank sheets to its sides. On them you draw the wall, the roof, a garden and a path, trying to use the same colors. You get a big picture, and the postcard ends up in its middle. A computer can also stretch a picture in every direction like this.

In plain words

You have a photo, but it is too narrow: for example, a vertical photo when you need a horizontal banner. The program puts the image on a large empty canvas and studies the edges of the original. Then it fills the empty margins: it continues the floor, sky, wall and trees, matching the perspective and the light. The original part stays as it was. You can repeat the process, moving the frame each time, and turn a small picture into a panorama. If there are few clues around the object, the neural network sometimes invents strange details, so someone reviews the result.

For business

Outpainting solves a common marketing task: one shot is needed in a dozen formats. A square for the feed, a vertical for mobile screens, a wide cover for the website and a marketplace card with space for text all come from one photo without a new shoot. The price depends on the number of formats, the resolution and the amount of manual finishing: in a cloud editor you pay per generation, on your own server for the hardware and setup. The risks: the added part may contain extra objects, other brands’ logos or perspective errors, and the model must never change the product itself. So the product is placed on top of the result as a separate layer. Test the process on one advertising campaign.

For a specialist

Outpainting (image extension) is the generation of content beyond the original frame. Technically, the task comes down to inpainting with an inverted mask: the known area is in the center, the unknown one at the edges. But here the context is on one side only, and errors build up as you move away from the original. In Boundless (Teterwak and coauthors, 2019), a generative adversarial network (GAN) did the extension, and the discriminator got a semantic condition so that the new part matched in meaning; the authors also showed extreme extensions up to panoramas. In diffusion pipelines, the source is scaled and placed on an empty canvas, a white mask covers the margins, and ControlNet holds the shape of the object, for example through a depth map. Then an inpainting model runs with a blurred mask, and the original is laid over the result. Long panoramas are extended in overlapping steps.

Example

A furniture store got vertical photos of armchairs from the factory. For a wide banner on the home page, a neural network drew a room with a window and a floor lamp on both sides, and the armchair stayed untouched.

Related terms

Questions and answers

How do you extend a photo with a neural network? You need an editor with an outpainting mode: you enlarge the canvas and describe what should appear on the new margins. It works most reliably when the background at the edges is simple: a wall, sky, floor or nature.
Can you use extended photos in advertising? Yes, if you hold the rights to the original photo, and the new part has no other brands and does not mislead the buyer about the product. A person checks the frame before it is published.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us