Extend an image with AI

The same wide shot gives a vertical story and a square for the feed, without cropping and without cutting the subject. The full workflow, no account needed.

Interactive demo

Real workflow, real results. Pan and zoom freely.

Use this workflow

The real workflow: the wide 16:9 shot is extended top and bottom to fit 9:16, and sideways to fit a square, without the subject moving.

At a glance

What the workflow produces: 2 images in 9:16, 1:1
Canvas structure: 6 nodes wired by 4 connections
Models wired in: Flux Kontext
Cost of one full run: about 12 credits, roughly $0.14

A wide image does not fit a story. Cropping a 16:9 into a 9:16 keeps only a third of the surface and almost always cuts the subject: that is the problem this workflow removes. The photo enters the Media node and feeds two Image nodes in parallel, one extending it vertically, the other into a square.

The instruction asks to EXTEND the scene, not to invent another one. The material added top and bottom has to continue what already exists: same sky, same ground, same light, same grain. That is what separates a clean extension from a collage, and it is also why a photo with a regular background extends better than a busy one.

The instruction also reserves the top of the frame, where the title and the platform's interface will land. A subject centred vertically in a 9:16 ends up half under the account name: better to plan for it at generation than at edit time. Both formats cost fourteen cents.

How it works

1
Start from your wide shot
A horizontal photo, a render, a capture: anything too wide for a vertical format. The more regular the scene around the subject, the more invisible the extension.
2
Run the story version
The 9:16 node adds material top and bottom. The instruction says what the model must continue, and reserves the upper area for the title and the interface.
3
Run the square version
The 1:1 node runs in parallel on the same source, this time widening the sides. One pass gives both formats a post needs.
4
Check the joins
Look at the junction between the original image and the added material. A horizon line that shifts or a texture that changes grain is fixed by describing the scene better in the instruction.

Variations and use cases

The same canvas, tuned for different needs: each variation is two or three lines of prompt away.

Story and Reel from a wide shot

The direct use case: a horizontal photo that has to be published in 9:16 without cutting the subject. The extension adds scenery top and bottom instead of trimming the sides, and the subject keeps its proportions.

The feed square and the story, in one pass

Both nodes run in parallel on the same source. A post needs both formats, and getting them in one gesture avoids the manual crop that shifts the subject from one version to the next.

Site banner and cover image

The problem is reversed: a vertical photo to stretch into a very wide banner. The instruction changes direction but not principle, you extend the scene sideways while leaving the subject where it is.

Rescuing a framing that is too tight

A photo whose subject touches the edge gets room to breathe with a few centimetres of scenery all around. That is often the difference between an image usable as a cover and an image that suffocates.

Reserving room for the text

The instruction keeps the top of the frame free for the title and the platform's interface. A subject centred in a 9:16 ends up half under the account name, and it is simpler to plan at generation than at edit time.

Adapting a campaign visual to every format

The same shot in 9:16, 1:1 and 16:9 keeps the same image and the same subject, which gives a coherent campaign. Three crops give three different images, which is not the same thing.

Common mistakes

Asking it to invent instead of extend

The instruction must say continue this scene, not add a sky. As soon as the model is left free on the added area, it composes something else, and the join shows even when both halves are beautiful.

Extending a subject cut off at the edge

If an arm or a leg touches the edge of the frame, the model thinks it should complete it and invents a limb. What is added must be scenery: crop slightly before extending rather than leaving it that choice.

Extending too far at once

16:9 to 9:16 is already a big gap. Beyond that, the invented part outweighs the real one and stops being believable. Two moderate extensions in a row beat one very wide.

Forgetting the text area

Generating a beautiful 9:16 then finding the title covers the face is the most frequent scenario. The instruction already reserves the top of the frame; if your edit puts the text elsewhere, change the reserved area before generating.

Recommended models

Flux Kontext

Seedream 5 Pro

Nano Banana

Frequently asked questions

How is this better than cropping?

Cropping 16:9 to 9:16 throws away two thirds of the image and cuts the subject or its setting. Here you add instead of removing: the subject keeps its place, its proportions and its surroundings, and the format changes anyway.

How far can you extend?

Going from 16:9 to 9:16 is already a big gap, and that is what the demo shows. Beyond that the model invents more than it extends, and the added part starts to show. Two moderate extensions in a row beat one very wide.

Does it work on a photo of a person?

Yes, as long as the person stays inside the original area. What is added should only be scenery: asking the model to complete a body cut off at the edge of the frame gives you spare hands and legs, and that is the case to avoid.

What do both formats cost?

Fourteen cents for the story and the square. The price shows on each node before you click, and a single format costs half if the other is no use to you.

How is this different from an automatic crop?

A crop chooses which part of the image to throw away. Here none is thrown: the subject keeps its place and its setting, and the missing surface is rebuilt in continuity with what exists.

Is the subject distorted?

No, that is exactly what extending avoids: nothing is stretched. The original part stays intact, and the model only works on the material added around it.

Which formats can I get?

The template starts on 9:16 and 1:1, the two formats a post needs. The Image nodes also accept 4:3, 3:4 and 16:9: changing the node's format is enough, the instruction stays the same.

Does it work on a screenshot?

On a capture with a regular background, yes. On an interface, no: the model will extend the buttons and the text by inventing them, and the result is unreadable. For that case, place the capture on a separately generated background.

Extend an image with AI

Use this workflow

Free account, no card required. The workflow opens pre-filled in your canvas.