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 problem | What you need |
|---|---|
| The image is too small in pixels | Increase resolution (upscaling) |
| The image is too big in file size or dimensions | Resize |
| There's extra content at the edges you want to remove | Crop |
| There's missing scenery at the edges you want to add | Outpainting |
| The aspect ratio is wrong and cropping would lose something important | Outpainting |
| The aspect ratio is wrong but there's margin to spare | Crop |
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
- 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.
- 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.
- 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.
- Write the request describing the continuation. Don't describe the whole image: describe what should exist outside of it.
- Generate and compare in batches. Two or three attempts quickly show which continuation is most coherent.
- 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:
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
Horizontal to vertical for mobile
Create space for text
Cropped-off portrait
Product with more space
Panoramic banner
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:
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:
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
- Did the original area stay intact, with no changes?
- Is the seam imperceptible at 100% zoom?
- Is the light direction the same in both areas?
- Does the horizon line stay at the same height?
- Was no element from the original duplicated?
- Did no object or person show up that you didn't ask for?
- Is the sharpness uniform between the original and the new area?
- Does the final aspect ratio match the image's destination?
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