Image Generation, in PracticeLesson 46 of 69

Contents

Image Generation, in Practice

Lesson 46 of 69. Sign in to keep your place.

Module 5 · Repair, extend, assemble

Extending a frame when the crop changes

8 min read

The call that makes you need this

The image is approved at 4:5. Marketing now needs a 16:9 version for the website banner. Cropping a 4:5 frame to 16:9 removes most of the picture, and the subject goes with it.

Outpainting adds canvas and fills it with content that continues the image. Done properly it is invisible. Done in one big jump it produces a seam, a change of light, and a background that does not agree with itself.

The method

Extend in steps of 128 to 256 pixels, not in one leap.

  1. Enlarge the canvas by 256 pixels on the side you need.
  2. Mask the new empty area plus 64 to 128 pixels of the existing image: the overlap is the important part.
  3. Inpaint with denoise 0.8 to 1.0 over the new area, using a prompt describing what continues there, not the original scene prompt.
  4. Repeat until you reach the size you need.

Why steps: the model sees only its context window. A 512-pixel extension in one pass means most of the new region has no nearby real pixels to match against, so it invents freely: different light, different perspective, sometimes a different room. A 256-pixel step keeps every new pixel close to something real.

Why the overlap: without it, the model is generating a strip that must meet the original exactly at a hard boundary, and it will not. With 100 pixels of overlap it is continuing an image rather than abutting one, and the transition happens inside the regenerated area where nobody can see it.

Prompting an extension

Describe what should be there, not what the image is of.

  • Extending upward from a courtyard scene: whitewashed wall continuing upward, overcast sky above the roofline, not the full original prompt.
  • Extending sideways from a portrait: plain studio background, soft falloff, not the description of the person, which will get you a second person.

That second-person failure is the most common outpainting mistake. The prompt still describes a woman, the model has empty canvas, and so it puts a woman there.

What breaks

Perspective drift. Each step can shift the vanishing point slightly, and after four steps a straight wall curves. Check long straight lines after every step rather than at the end. If drift starts, a control (canny on the existing edges, or a depth map) holds the geometry.

Light drift. The new area gradually becomes brighter or flatter. Same cause, same remedy: check each step, and correct the grade as you go rather than at the end.

Repetition. Extending a patterned surface produces obvious repeats, because the model continues what it sees. Vary the prompt slightly on each step, or break it up with a repair pass afterwards.

The tell at the join. Look specifically for a faint vertical or horizontal line where the seam was, visible as a slight change in noise or contrast. A low-denoise pass at 0.2 over a band spanning the join clears it.

Doing it in one operation

Some tools handle this well in a single step: Photoshop’s generative expand, InvokeAI’s canvas, and several hosted services all extend a frame cleanly in one action. If you have one and the extension is modest, use it.

The manual method still matters for two reasons. Large extensions (doubling a frame’s width) fail in one shot regardless of tool, because the problem is context rather than implementation. And the manual method gives you control over the prompt at each step, which is how you decide what is in the new space rather than accepting what the model thinks belongs there.

The better answer, when there is time

Outpainting is a rescue. The better outcome is not needing it, and that is the brief block’s job: ask at the start which ratios are required, and either generate wider than any of them or generate each ratio separately.

There is also a real ceiling. Outpainting extends a scene; it cannot extend a composition. If the original frame worked because the subject was placed against the edge, adding 400 pixels of room changes that relationship, and the wider version can be technically seamless and compositionally worse. Look at the extended image as a new picture and judge it as one, rather than only checking whether the join shows. Sometimes the honest answer to the banner request is a fresh generation at 16:9 using the approved frame as a reference, and it takes less time than four careful extension steps.

Every lesson here is open. Your progress is recorded once you pass the Vectors, Logos and Things That Get Printed exam.

Check your understanding

Why does outpainting in 256-pixel steps with overlap work better than one 512-pixel extension?

Pick the one you would defend. Your answer stays private.

Next lesson · 8 minFour ways to remove an object, and when each is right

Stuck on something?Ask Neema, your AI tutor

How to cite: Addaly (2026) ‘Extending a frame when the crop changes’, in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/outpainting-the-crop (Accessed: 13 October 2026).