How to extend an image with AI: three outpainting models compared

By the Pickframe team · Published 31 October 2026 · 8 min read

Turning a square or landscape picture into a vertical story or a wide thumbnail usually means cropping most of it away. Outpainting goes the other way: an AI model keeps your image and paints the missing space around it. On 31 October 2026 we gave three outpainting models the same three images and the same frames, and kept the first result from each.

How we tested

Three images: a landscape photograph, and a portrait and a product shot from our first model comparison.
Each image was placed on its new frame exactly as the Pickframe editor places it, and the same frame went to all three models: Bria Expand, Bria FIBO Expand and FLUX.2 Pro Outpaint.
There was no prompt. These models continue the picture they are given.
One result per model and image: no retries, no picking the best of several, no retouching. Select any picture to open the untouched original file.
They ran through the same fal.ai routes and request settings as the Pickframe editor.

Test	Original	New frame	Where the original sits
Landscape to vertical cover	960 × 636	960 × 1707 (9:16)	Low in the frame: 857 px of new sky, 214 px of new ground
Portrait to thumbnail	1024 × 1024	1820 × 1024 (16:9)	At the right edge: 796 px added on the left
Product to vertical story	1024 × 1024	1365 × 2427 (9:16)	Centred at three quarters size: new space on all four sides

One result per model cannot rank models. It shows what each one did with these pictures on one run, which is the kind of difference worth seeing on your own image before you choose.

The results at a glance

All nine results came back at exactly the size asked for, and none has a visible join between the original and the new area, even at full size. The differences are in what each model invented, and in whether it left the original alone.

Test	Bria Expand	Bria FIBO Expand	FLUX.2 Pro Outpaint
Kept the original	Softened it in all three tests	Unchanged	Very close; slightly softer in two tests
Landscape	Blue sky; dark clutter along the old bottom edge	Deep blue sky; a tidy fence line	Turned most of the new sky cloudy
Portrait	A second window frame, then a plain wall	A plain wall, the most room for a headline	A large window onto a blurred city
Product	A taller plinth	A second band of sunlight on the wall	A lower step under the plinth

Test 1: a landscape photograph to a vertical cover

A wide photograph cropped to 9:16 keeps only about a third of its width. Expanding keeps all of it and asks the model for sky above and ground below, leaving room for a headline at the top. The bottom edge of this photograph cuts through a dark strip, which each model had to make sense of.

The desert photograph placed low in a tall 9:16 frame, with the rest of the frame hatched — The hatched area in the first picture is what each model filled. One unedited result per model.

Bria Expand: clear blue sky with thin cloud above the buttes; below, scrub and a fence with dark shapeless heaps beside it — The hatched area in the first picture is what each model filled. One unedited result per model.

Bria Expand continued the clear blue sky with thin cloud, closest to the original’s colour. Below, it turned the dark strip into a fence with dark, shapeless heaps beside it, the one part of these results that looks wrong.
Bria FIBO Expand deepened the blue towards the top, like a photograph taken with a polarising filter. It made the strip a tidy fence line, with open sand and bushes below.
FLUX.2 Pro Outpaint covered most of the new sky in grey-white cloud, so the picture reads as a different, duller day. It made the strip a road and added a sandy gully below. White text would need a darker panel behind it on that sky.

Test 2: a square portrait to a 16:9 thumbnail

Thumbnails often put the person on one side and the title on the other. Here the portrait sits at the right of the frame, and each model had to continue the window and wall to its left: straight lines, where a bad join is easy to spot.

The square portrait at the right of a 16:9 frame, with the left side hatched — The same portrait extended to the left by each model.

Bria Expand: the window gains a second frame, and a plain grey wall fills the left side — The same portrait extended to the left by each model.

Bria Expand added a second window frame and then a plain grey wall.
Bria FIBO Expand filled almost all of the new space with a plain grey wall, which leaves the most room for a headline.
FLUX.2 Pro Outpaint opened the window into a large pane with a blurred city view. It is the most photographic result and explains the soft light on her face, but it is a busy, bright background for text.

Test 3: a product shot extended on every side

Shrinking the image in the frame makes the model add space on all four sides. The plinth, the branch and the shadows all cross the edges, so each one has to carry on convincingly.

The square product shot at three quarters size in the middle of a tall 9:16 frame, hatched all round — The same product shot with new space on all four sides.

Bria Expand: the plinth continues down as a tall stone block and the branch reaches the left edge — The same product shot with new space on all four sides.

Bria Expand continued the plinth down to the bottom of the frame as a tall stone block, and the branch to the left edge.
Bria FIBO Expand added a second band of sunlight across the wall above, a natural repeat of the window light, and gave the branch a few more leaves.
FLUX.2 Pro Outpaint stood the plinth on a lower step and added a pale wall edge at the far left. It invented the most, but nothing in it looks out of place.

Did they keep the original?

An expand model is meant to paint only the new area. We compared the part of each result where the original sits with the original itself.

Bria FIBO Expand returned it unchanged, within one or two levels of brightness out of 255 on average.
FLUX.2 Pro Outpaint came very close. Its copies were slightly softer in the landscape and the portrait.
Bria Expand redrew the whole picture. Its colours match, but by a standard edge-sharpness measure it kept between a fifth and a half of the original’s fine detail in the three tests. At thumbnail size that is hard to see. Up close it is not.

Original: a close-up of the cheek and nose — The same 300 × 200 pixels of the face, from the original and from each result, enlarged 2× without smoothing. Bria Expand’s copy has lost most of the freckles and skin texture.

Bria Expand: a close-up of the cheek and nose — The same 300 × 200 pixels of the face, from the original and from each result, enlarged 2× without smoothing. Bria Expand’s copy has lost most of the freckles and skin texture.

What we would take from this

To keep your image untouched, Bria FIBO Expand did it exactly here and FLUX.2 Pro Outpaint nearly so.
For plain space to put a headline on, Bria FIBO Expand made the plainest continuations: an empty wall, an open sky.
For the most photographic scene, FLUX.2 Pro Outpaint invented the most, from a city view to a change in the weather. Check that it still matches the mood you want.
Bria Expand continued each scene sensibly but softened the original, so it suits pictures that will only be seen small.
Try more than one. The models disagree about what lies outside the frame, and you will usually like one result much more than the others.

Extend your own image in Pickframe

Open an image saved in your account in the editor and switch to the layout you want to fill, such as 4:5, 9:16 or 16:9.
In the AI edit tab, choose Expand canvas. The preview shades the area to fill. Use Image size and the position sliders to place your image.
Tick the models to compare and run the edit. The button shows the credit cost first, and each model makes its own version.
Choose the result you like under Image for this layout. Your text and logo layers stay on top, and every social size drawn from that layout uses it.

The full workflow, from one image to a thumbnail, a square post and a vertical cover, is in how to turn one image into a YouTube thumbnail, square post and vertical cover. Every original, the exact requests, the canvas sizes, the checksums and the measurements for these nine results are in the public manifest.

More guides

How to inpaint with AI: five models compared
We replaced a cap, removed a table and changed a jumper with five AI inpainting models. See every unedited result, which models left the rest of the picture alone, and the wording each one needed.
How to make an SVG logo with AI
We asked an AI vector model for three logos and measured the SVGs it made: spelling, colours, shapes and file size. See what came out clean, what needs tidying and how to finish it in a vector editor.
How to make a character turnaround with AI: five edit models compared
We turned two original characters around with five AI edit models. See every untouched back and side view, which models moved accessories to the wrong side, and the wording that stops three-quarter side views.
FLUX Schnell vs Krea 2 Turbo vs Recraft V4.1: one prompt, three results
We gave three AI image models the same product, portrait and poster prompts. See the unedited outputs, the exact prompts, and where each model followed or missed the brief.
How to get correct text in AI-generated images
Why AI image models misspell words or print your instructions, how to word a prompt for lettering, and when to add the text yourself instead.
How to write an AI image prompt that works across models
A practical structure for image prompts (subject, setting, light, camera and exclusions), with three real prompts and what different models did with them.
How to turn one image into a YouTube thumbnail, square post and vertical cover
Make 16:9, 1:1 and 9:16 versions of one image with an editable headline, extend it with AI where a crop cuts too much, and export every social size at once.