Grow the canvas to the shape a placement needs and let the model fill only what falls outside your photo. The original keeps its pixel size and is composited back untouched, so a product, a label or a horizon survives the extension exactly as you shot it.
Extend your image
Upload a photo, choose the canvas ratio and where the original sits, then check the marked area before you pay. JPG, PNG or WebP up to 20MB, 256px minimum.
This run costs
10 credits
Your balance
--
Each case is the same operation on a different input: the frame grows, one region is generated, and the part of the photo you already had is copied back byte for byte.
The wordmark and the batch code on the left stay exactly where the photographer put them, because that region is not generated at all.
The linen weave and the cutlery shadows run through the join, so the extra height reads as more table rather than a stretched plate.
The horizon stays on one line across the seam; the new sky and water are generated, the photographed right half is untouched.
It adds canvas, not detail. You start with a photo and a target shape; the tool builds a larger frame in that shape, places your photo inside it at the position you chose, and generates only the band of new pixels that the larger frame exposes. Your subject is never resized to fit the new shape, so a vertical packshot stays a vertical packshot inside a wider banner instead of being stretched or cropped.
That is the difference between extending and enlarging. An upscaler invents detail inside the frame you already have; an extender invents pixels outside it. The two jobs are often confused because both "make the image bigger", but only one of them changes the shape of the frame, and only one of them has to continue a scene past an edge that did not exist before.
On IMG AI the canvas comes from the ratio you pick and the anchor you set. Three ratios (16:9, 1:1, 9:16) and nine anchor positions cover a banner, a square post and a story. The ratio decides which side receives the extra space; the anchor decides which edge your photo keeps. When the ratio already matches your photo there is no new canvas, so there is nothing to fill and nothing to charge.
The last step is the one that makes the result usable: the original region is composited back from your own file after generation, so a label, a logo or a face does not get redrawn. Only the new band is a generation.
Pick the anchor that keeps the side you care about, then let the ratio do the rest. A bottle framed on the left of a vertical photo belongs on the left of a 16:9 banner, so the anchor is centre-left and the new space opens to the right. The photographed side is copied through; the invented side is the empty one.
Choose a JPG, PNG or WebP up to 20MB with both sides at least 256px. The frame keeps its own pixels; nothing is stretched on upload.
Pick 16:9 for a banner, 1:1 for a square post or 9:16 for a story, then set which of the nine positions holds your photo. Both controls only redraw the preview: the marked band is exactly the area that will be generated.
The preview states the final canvas in pixels and what share of it is new. The button shows 10 credits before you press it, and a run that would add no canvas stays disabled instead of charging for an empty mask.
The generated band continues the scene next to it. Look along the join at full size: the light direction, the surface and any straight line in the photo should carry across without a step or a repeat.
The file you download is the full canvas with your original region pasted back from your upload, so the product side is bit-for-bit what you shot. Save it to My Creations and re-download it later if you need the same file again.
Marketplaces, ad slots and email headers ask for a wide frame; the photographer delivered a tall one. Extending right or left keeps the label and the cap in their original pixels while the surface and the background continue past the old edge.
A 1:1 overhead shot of a dish is the wrong shape for a 9:16 placement. Extending up and down adds table, not plate, so the food keeps its true size and the linen continues at the same scale.
A horizon photo used as a page header often needs empty sky or water where a headline will sit. Extending one side gives the type a calm area and keeps the horizon on its own line.
Interiors are usually shot as tight as the space allows. Extending a wall or a floor continues the perspective already in the frame, which is far more convincing than stretching the photo to a wider aspect.
A print or a thumbnail may need a squarer frame than the camera produced. Setting the anchor protects the subject side and lets the background absorb the change.
Five tools get searched for the same problem, and only one of them adds canvas while keeping your photo intact. The rest solve genuinely different jobs, so the honest answer depends on what has to survive.
You need more canvas around a photo that must survive exactly as shot.
Upload, pick a ratio and one of nine anchors, read the final dimensions and the share of new area, confirm the price, generate the new band, check the seam, download the composited PNG.
The extension is generated, so it can only be as convincing as the context next to it; a scene with no continuation cue at the edge will show a seam.
Something inside the existing frame has to be replaced or removed.
Paint a region that already exists, describe what belongs there, and redraw only that region. The frame never changes size.
It cannot give you a new shape. If the job is a wider banner, in-frame fill has no canvas to work with.
You have a message and a placement, and the artwork can be invented from a brief.
Describe the campaign, set a headline and a format, and let the model compose the whole frame from scratch.
Your product photo becomes a suggestion rather than a constraint. A real package with real type will not come back pixel-identical.
A designer needs frame-level control over every pixel and has the time to paint the new area by hand.
Enlarge the canvas in an editor, clone or paint the new region, retouch the join, and export at the target size.
The most control and the slowest route. Straight lines such as horizons and table edges need careful manual work to continue convincingly.
The frame is right and the problem is resolution.
Feed the existing image to a model that predicts detail at a larger pixel count, keeping the same aspect ratio.
The shape does not change and no new scene appears. Upscaling enlarges the subject; it does not extend the world around it.
Pick by what has to stay true. If the photo is the constraint, extend it. If a region inside it is wrong, fill in-frame. If the design is open, generate a banner. If only the resolution is short, upscale.
The three tools sit next to each other on purpose. The difference is where the generated pixels land relative to your photo.
| The difference | AI Image Extender | Generative Fill | AI Banner Generator |
|---|---|---|---|
| Where the new pixels go | Outside the original frame, on the band the canvas exposes | Inside the original frame, under the brush | Everywhere; the frame is invented from a brief |
| What happens to your photo | Kept at its own pixel size and composited back untouched | Kept outside the painted region | Used at most as a reference for the design |
| What you choose | A canvas ratio and one of nine anchors | A brush stroke and a replacement description | A brief, a headline and a format |
| Resulting shape | A different aspect ratio from the source | The same aspect ratio as the source | Whatever the placement needs, designed from scratch |
| The job it solves | A photo in the wrong shape for a placement | A wrong or unwanted region inside a photo | A campaign graphic that does not exist yet |
| Cost on IMG AI | 10 credits per run | See the generative fill page | See the banner generator page |
One canvas in the ratio you picked, one generated band, and one PNG at the dimensions the preview stated. The photo's own region is copied through from your upload rather than redrawn, so the pixels you care about are the pixels you supplied.
No. Only the band outside your photo is generated. When the run finishes, your original region is composited back from the file you uploaded, at the same pixel offset the preview showed, so those pixels are identical to your source. The new area beside it is generated and will not match your file because there was nothing there to match.
Because the ratio you picked adds no canvas. If your photo is already 16:9 and you choose 16:9, the mask would be empty, so there is nothing to fill and nothing to charge for. Choose a different ratio and the button enables with the new dimensions and the share of new area shown.
You never stretch it. The tool builds a larger frame and places your photo inside at its own pixel size, then generates the empty band around it. Changing the ratio or the anchor only moves where your photo sits in that frame; it never resamples the subject to fill the frame.
The generated band is asked to continue the light, perspective, colour and texture already in the frame, so the join should read as more of the same scene. Straight features such as a horizon or a table edge are the real test: check that the line carries across at the same height and angle before you publish. A scene with no continuation cue at the edge is the hardest case and can show a visible step.
Yes. Set the anchor to an edge position and the spare space opens on the opposite side. Centre-left keeps your photo against the left edge and adds all of the new canvas to the right; top-centre keeps it against the top and adds everything below. The centre anchors split the new space evenly between two opposite sides.
Choose the shape the placement requires, not the shape of your photo. 16:9 suits a banner, a video frame or a web header; 1:1 suits a feed post; 9:16 suits a story, a short or a tall print. A ratio close to your photo adds less canvas and therefore a smaller generated band, which is usually the safer result.
They are generated from the edge of your photo and the anchor you set, then returned in the same canvas. Because they are new, they are the part to review: look for a repeated texture, a broken straight line or a light direction that changes across the join.
The canvas is the smallest frame of your chosen ratio that contains the whole photo, and it is scaled once onto a 32-pixel grid with a 1440-pixel cap on the long side for the run. A 1200 × 1600 photo extended to 16:9 produces a 2133 × 1200 canvas, which is scaled to 1440 × 800 for generation and returned at the canvas size shown in the preview.
Yes. Upload and generation both require a signed-in IMG AI account, and each run reserves 10 credits from it. The price is shown before you press the button, a dropped connection recovers the same run instead of charging twice, and a failed run returns its reserved credits.
Move to the workflow that matches the asset you actually need.
Upload a photo, pick the canvas ratio and the anchor, check the marked area, and download a PNG with your original pixels still in place.
Open the image extender