Making a photograph bigger than it was taken

The job this solves

A photograph shot in portrait is needed as a wide banner. A square product shot needs to become 16:9 for a slide. A portrait cropped too tight needs headroom for a magazine’s masthead. The image does not extend that far.

The old answers were to crop something else, to mirror a strip of the edge, or to place the photograph on a coloured panel. Outpainting extends the canvas and generates plausible content into the new area, conditioned on the existing pixels along the boundary.

Mechanically it is inpainting with the mask outside the original image instead of inside it. All the properties from the first lesson apply: it is softer, cleaner, wrong about text and regular geometry, and non-deterministic.

Technique that works

Extend in stages, not in one jump. Asking for 100 per cent more width in one operation gives the model a thin strip of context along one edge to condition a very large region. It drifts. Extending by 20 to 30 per cent, then again, gives each pass substantial context and a much more coherent result.

Keep overlap. Include a band of the existing image inside the generated region; most tools do this automatically, and where the setting is exposed, 64 to 128 pixels is a reasonable overlap.

Describe the scene, not the extension. “Continues the wooden table and grey wall, soft window light from the left” works. “More table” does not.

Expect the horizon to drift. A horizon or a strong architectural line will bend slightly across the new region. Check it with a straight edge and correct by transforming the generated area or by cloning the real line outward.

Then unify. Grain, vignette, black point and colour across the whole extended image, exactly as in compositing. The generated strip is cleaner than the photograph.

Aspect-ratio conversion, which is the real use

One photograph has to appear as:

• 1:1 for a marketplace grid
• 4:5 for a social feed
• 9:16 for a story or reel
• 16:9 for a slide or a website hero
• 3:1 for a banner

Cropping to all five destroys the composition in at least three of them. Outpainting the background outward, then cropping, keeps the subject where it belongs in each format.

The practical workflow: extend the master file generously in both directions once, keeping the extension on its own layers, then crop that single master to each ratio. One generation, five outputs, and the subject sits correctly in all of them.

This is the most useful application of generative imaging in everyday commercial work, and it is the least discussed.

Alternatives that need no model

For a modest extension, three older techniques still work and are exact:

Mirror the edge. Copy a strip, flip it, and blend it in with a soft mask. Invisible on texture (grass, water, sky, bokeh) and obvious on anything with recognisable content.

Content-aware scale (Edit > Content-Aware Scale, and GIMP’s Liquid Rescale plugin, which is the free implementation of seam carving). This stretches a photograph while protecting important regions, by removing or duplicating low-energy vertical seams. It works remarkably well on a landscape with a lot of sky and fails visibly on anything with structure.

Extend with a graphic panel. Put the photograph on a coloured or textured field and use the space for type. This is design rather than illusion, it is honest, and for many layouts it looks better than a generated extension.

Consider these first. They are faster, free, deterministic, and they invent nothing.

Free tools

ComfyUI, AUTOMATIC1111 and Fooocus all do outpainting with local Stable Diffusion models. Krita’s AI diffusion plugin does it on a canvas you can then keep editing. Several phone apps offer it.

Adobe’s Generative Expand is the integrated paid version, available from the crop tool.

The honesty question, which is different here

Outpainting raises a question that inpainting does not, and it is worth stating plainly.

Removing an object is a subtraction that a viewer may or may not notice. Extending a photograph adds content that the viewer will read as part of the scene: more of the room, more of the street, more of the landscape. They will believe it was there, because everything else in the frame was.

So:

• Advertising, book covers, banners, design. Unremarkable.
• Editorial and documentary. Not acceptable. Extending a news photograph fabricates the scene around the event.
• Property. An extended room shows more room than exists. This is the sky-replacement problem with a sharper edge to it.
• Portraits and people. Extending the frame around a person often generates additional limbs, objects or figures at the boundary. Check.

The delivery note habit applies: “canvas extended by generative fill on left and right; original frame supplied.”

The limit

The further you extend, the less the original constrains the result, and the more the output reflects the training data rather than your photograph. A 20 per cent extension of a textured background is usually indistinguishable. A photograph extended to three times its width is mostly a generated image with a photograph in the middle of it, and it will look like one: the generated regions share a smoothness, a colour character and a compositional blandness that becomes obvious at scale.

When the extension exceeds roughly half the original dimension, stop and consider whether the honest answer is a different photograph.