You have a perfect vertical photo and need it as a horizontal banner. Or an image where the top got cut off and there's no room for the title. Or a square artwork that needs to become a phone wallpaper. In all three cases the traditional solution is bad: stretching distorts, cropping takes away even more space, and redoing it costs time.

Outpainting solves this by generating what would be outside the frame. The AI analyzes the existing image, understands the scene, the perspective and the lighting, and continues the setting beyond the original edges. The result is a bigger image — not upscaled, but extended.

This guide covers what it is, when to use it, when not to, how to write the expansion request, the most common problems, and how to fix them. It also includes six ready-made prompts for the most frequent cases.

What outpainting is and how it works

Outpainting is the generation of new content outside the boundaries of an existing image. The model uses the pixels that are already there as context — colors, shapes, light direction, perspective lines — and synthesizes the logical continuation of the scene in the empty areas.

The term has a sibling that causes confusion: inpainting is the opposite, that is, generating content inside the image, replacing a selected area. Inpainting erases an object and fills the hole. Outpainting adds space around it.

An even more important distinction: outpainting isn't upscaling. Upscaling an image increases the number of pixels while keeping the same framing. Outpainting keeps the pixel density and increases the framing. These are different operations that solve different problems, and swapping one for the other is the most common mistake beginners make.

Outpainting, resizing, cropping, or upscaling: which one you need

Many people who look for outpainting actually need something else. This table settles the question in ten seconds.

Your problemWhat you need
The image is too small in pixelsIncrease resolution (upscaling)
The image is too big in file size or dimensionsResize
There's extra content at the edges you want to removeCrop
There's missing scenery at the edges you want to addOutpainting
The aspect ratio is wrong and cropping would lose something importantOutpainting
The aspect ratio is wrong but there's margin to spareCrop

Notice that only the two rows with "missing scenery" and "cropping would lose something important" actually call for outpainting. In the others, conventional tools are faster, more predictable, and don't introduce invented content: you can resize, crop, or increase the resolution without any generation involved.

💡 Quick test: if you can solve it with cropping without losing anything essential, crop. Outpainting always invents content, and invented content is content that can come out wrong.

When to expand and when not to

Works very well on: landscapes, skies, textures, uniform surfaces, wide indoor spaces, blurred backgrounds, and scenery with a predictable pattern. Anything the AI can continue without having to invent complex structure.

Works poorly on: partially cropped faces, hands at the edge, text that continues off-frame, architecture with rigid geometry, repeated patterns with exact alignment, and anything where the viewer knows what the continuation should look like. In these cases the flaw is obvious.

Don't use it when the image represents something real that needs to be faithful — an ad product photo, a document, a receipt, a news image. Expanding those cases creates content that never existed, with all the problems that implies.

Step by step for expanding an image

  1. Start with the highest available resolution. Expanding a small image makes everything worse: the new area is born with the same low density and the seam becomes more visible. If the original is small, increase the resolution first.
  2. Set the final aspect ratio. Decide in advance whether you're going to 16:9, 9:16, 1:1, or another. Expanding "a little" and deciding later usually results in two expansions, and each expansion degrades things a bit.
  3. Choose the directions. Expanding only to the sides is easier than expanding in all four directions. The less new area, the lower the chance of error.
  4. Write the request describing the continuation. Don't describe the whole image: describe what should exist outside of it.
  5. Generate and compare in batches. Two or three attempts quickly show which continuation is most coherent.
  6. Check the seam at full size. Zoom in on the line where the original ends and the new part begins. That's where the flaws show up.

How to write the expansion request

Describe the continuation, not the image

The most common mistake is re-describing what's already visible. The model already sees the image; what it doesn't know is what should exist around it. "Continue the beach to the right, with the sea following through to the horizon and wet sand in the foreground" works much better than "a pretty beach at dusk".

Lock what can't change

Explicitly ask to preserve the original. The phrasing makes a real difference:

Expand this image to the horizontal 16:9 format. Keep the original area intact, with nothing inside it changed. Naturally continue the scenery on both sides, preserving the same lighting, perspective and color palette.

Name the lighting and perspective

These are the two elements that most give away a poorly done expansion. If the light comes from the left in the original and the new area has shadows going the wrong way, the eye notices instantly even without being able to explain why. Saying "same light direction" and "same horizon line" reduces this risk a lot.

Say what should keep existing

List the elements that are part of the scene and should continue outward: "continue the wooden table, the brick wall in the background, and the same blurred depth of field". Without this, the model sometimes invents a different setting right past the edge.

Also say what shouldn't appear

In the new area, the model tends to add elements on its own — more people, more objects, a building on the horizon. If you want clean space for text, ask for: "clean, uniform expanded area, no new objects or people".

Six ready-made expansion prompts

Vertical to horizontal

Expand to the horizontal 16:9 format. Keep the original area intact. Continue the scenery on both sides preserving the same lighting, perspective and palette. Don't add any new objects or people.

Horizontal to vertical for mobile

Expand to the vertical 9:16 format. Keep the original area intact and centered. Continue the sky upward and the ground downward, keeping the same horizon line and light direction. Clean, uniform upper area.

Create space for text

Expand the image to the left by about 40% of the width. Keep the original area intact. The new area should be a smooth, uniform continuation of the background, with no objects or details, suitable for adding text.

Cropped-off portrait

Expand downward and to the sides, showing more of the body and the surroundings. Keep the face and the original area completely intact. Preserve the same side lighting, the same background blur, and the same clothing.

Product with more space

Expand on all sides while keeping the product centered and intact. Continue the support surface and the background with the same color, texture and lighting. No new shadows or additional objects.

Panoramic banner

Expand to the panoramic 3:1 format, widening only to the sides. Keep the original area intact and centered. Continue the landscape laterally preserving the horizon line, the perspective and the color palette.

Expand in steps, not all at once

This is the technique that most improves the result and that almost nobody uses. Instead of asking for the image to triple in width all at once, expand in increments of 20% to 30% and repeat.

The reason is simple: the bigger the new area relative to the existing one, the less context the model has to work from. Expanding gradually, each step starts from an image that already contains the previous expansion, and the model always has enough material to continue coherently.

In practice, three 25% expansions give a much better result than one 75% expansion. The cost is time, and the gain is perspective and lighting coherence — exactly where large expansions fail.

💡 Practical limit: above roughly double the original size, the image starts to look like a different scene with a piece of the original photo pasted in the middle. If you need much more space than that, it's probably better to generate the whole image from scratch in the right format.

Most frequent use cases

Reusing an image in another format

The most common use. An artwork made for a square feed post that needs to become a horizontal cover, or a vertical photo that will become a banner. Expanding preserves the original composition instead of cutting it.

Opening up space for text

Very useful for campaign pieces and thumbnails. Instead of placing the text over the subject, you create a clean side band. Explicitly ask for the new area to be uniform and detail-free.

Fixing tight framing

Photos where the top of the head is glued to the edge, or where there's no room in front of someone looking to the side. Expanding by a few percent solves it and usually goes unnoticed, because the new area is small.

Adapting for print

Printing requires bleed margin beyond the final area. Expanding a few millimeters on each side keeps the cut from eating into the image — and this is a case where the new content won't even show up in the final result, which makes the risk practically zero.

Recomposing a generated image that came out too tight

This happens a lot: the AI generates a good scene but with the wrong framing. Expanding is faster and safer than generating again and hoping the new version turns out just as good.

Where to do outpainting

Three paths, from simplest to most controlled.

Conversational tools. You upload the image and ask for the expansion in natural language. This is the fastest path and covers most cases well. The limitation is control: you don't set exactly how many pixels in each direction. The ChatGPT guide for creating images shows this workflow in detail.

Editors with an expandable canvas. You drag the frame's edges and the tool fills in the empty space. This is the method with the most control over the exact area to be generated, and the preferred one for people working with a fixed format.

Open models with masking. Maximum control, including generation strength and negative prompt, at the cost of setup. Worth it for high volume or specific requirements.

Since the options change frequently, the practical criterion is to test with your own image and compare the seam. The guide to the best AI tools to create images helps you decide where to start.

The most common problems and how to fix them

Visible seam at the joint

A noticeable line where the original ends. Usually comes from a difference in sharpness or color between the two areas. Fix: expand in smaller increments and, if it persists, explicitly ask for "continuous, imperceptible transition, same sharpness and same color temperature".

Duplicated elements

The model repeats an object from the original in the new area — two identical trees, two matching poles. This is the most frequent flaw. Fix: ask for "no repeating existing elements, varied natural continuation" and, when it persists, expand less each time.

Broken perspective

Lines that should converge start to diverge, or the horizon changes height. Appears mostly in architecture and interiors. Fix: mention "same horizon line and same vanishing point" in the request. If it doesn't fix it, this is a case where cropping is the better path.

Style drift

The new area comes out looking slightly different — more saturated, sharper, with a different texture. Fix: ask for "same palette, same grain and same level of detail as the original image".

Invented objects

People, animals, or buildings show up that you didn't ask for. Fix: include in the instruction "expanded area with no new objects, people, or buildings".

Loss of resolution in the new area

The generated part comes out softer than the original. Fix: start with a higher-resolution image and, in the finishing step, apply sharpness adjustment to the whole image to even things out.

After expanding: finalizing the file

The expanded image rarely comes out at the exact size you need, and almost never at the right file size.

Adjust the dimensions. Expanding generates a new aspect ratio, but not necessarily the final resolution you want. The resizer gets the image to the destination's exact size — 1920 by 1080 for a banner, 1080 by 1920 for stories.

Trim the edges. If the expansion left a little extra in some direction, cropping a few pixels is better than resizing and distorting the aspect ratio.

Even out the sharpness. When the new area came out softer, sharpening the whole image helps disguise the difference between the original and the generated area.

Convert and compress. Expanded images end up bigger in area and therefore heavier. Use the converter for the destination format and the compressor as the last step, always after resizing.

💡 Correct order: increase resolution if needed → expand in steps → crop the excess → resize to the final size → adjust sharpness → convert → compress.

A complete example, start to finish

A real, frequent situation: you have a vertical photo of a café, taken on a phone, and need it as a 1920-by-1080 horizontal banner at the top of a website, with space on the left for the title.

Step 1 — check the resolution. Does the phone photo have enough width? If the original image is under 1080 pixels tall, increase the resolution before anything else. Expanding from small material multiplies the flaws.

Step 2 — decide the strategy. Going from vertical straight to 16:9 means generating a lot more new area than exists in the original. Better path: two or three successive side expansions, each adding about 30% width.

Step 3 — first expansion. The request describes the continuation, not the photo:

Expand this image about 30% on both sides. Keep the original area intact. Continue the wooden counter to the right and the brick wall to the left, preserving the same warm side lighting, the same perspective, and the same background blur. Don't add any people or new objects.

Step 4 — repeat until you reach the aspect ratio. Each new expansion starts from the previous result. Check the seam at every step: it's easier to fix a small error now than to discover on the third round that the perspective broke on the first.

Step 5 — expand a bit more to the left. To open up the text band, one more expansion just on that side, asking for a clean area:

Expand another 25% only to the left. The new area should be a smooth continuation of the wall, uniform and detail-free, with the same color and lighting, suitable for placing text over it.

Step 6 — finalize the file. Crop the excess to the exact aspect ratio, resize to 1920 by 1080, adjust the sharpness if the new area came out soft, convert to WebP, and compress last.

The whole process takes ten to fifteen minutes and preserves the original photo at the center of the composition — which is exactly what a crop would have destroyed.

Expanded image checklist

If any item fails, the fix is almost always the same: expand less each time and repeat. Smaller increments solve most expansion flaws.

Get your expanded image to the exact size

Adjust the final resolution for a banner, stories, wallpaper, or print, right in your browser.

Resize image

Frequently asked questions

What's the difference between outpainting and increasing resolution?
These are different operations that often get confused. Increasing resolution multiplies the number of pixels while keeping exactly the same framing — the image gets bigger in pixels, but shows the same thing. Outpainting keeps the pixel density and expands the framing, generating new content beyond the original edges. If your image is too small, you need upscaling. If it's in the wrong format, you need outpainting.
How much can you expand without ruining it?
As a practical reference, up to roughly double the original area the result tends to stay convincing, as long as the expansion is done in steps of 20% to 30% at a time. Beyond that, the original becomes a small part of a mostly invented scene, and the image starts to look like a collage. When the need is much greater than that, it usually pays off to generate the whole image from scratch already in the right format.
Why do duplicated elements show up in the expanded area?
Because the model uses what already exists as a reference and, when context is missing, it tends to repeat the patterns it recognizes. One tree becomes two identical trees, one pole becomes two. Two things reduce the problem a lot: expanding in smaller increments, giving the model more proportional context at each step, and explicitly asking for a varied continuation, without repeating elements already in the image.
Does outpainting work on photos of people?
It works well when the new area is scenery — more background around them, more headroom above, more environment to the side. It works poorly when the expansion needs to complete the body, especially hands, feet and arms cut off by the edge, which remain the weak point of these models. The practical recommendation is to expand only the surroundings and keep any part of the person that already appears in the original image intact.
Can I use outpainting on a product photo for sale?
Only with care, and never over the product itself. Expanding the background around a real product, keeping the object completely intact, tends to be acceptable and useful for adapting the photo to different formats. What you shouldn't do is use the expansion to complete parts of the product the photo didn't show, or to create context that suggests something isn't included that isn't actually sold with it. The listing image needs to match what the buyer will receive.
Is there an alternative to outpainting for changing the aspect ratio?
Yes, and in many cases it's the better choice. If there's margin left at the edges, cropping solves it with nothing invented and no risk of error. Another common option for graphic pieces is placing the image on a solid or blurred background in the desired format, keeping the original photo whole at the center. Outpainting only pays off when cropping would mean losing something important from the composition.