Guide
How far AI can expand an image
Every AI image extender has a point past which it starts inventing. Here is ours, as a table, with the reason for every refusal — generated from the same rules the tool applies to your photo.
Six photos, eight shapes
| From | 16:9 | 9:16 | 1:1 | 4:5 | 3:2 | 2:3 | 21:9 | 4:1 |
|---|---|---|---|---|---|---|---|---|
| Phone still 9:161,080 × 1,920 | refusedarea ×3.2 | already this shape | in reach6 passes, ~15 s | in reach8 passes, ~19 s | refusedarea ×2.7 | in reach8 passes, ~19 s | refusedarea ×4.1 | refusedarea ×7.1 |
| Phone photo 3:43,024 × 4,032 | refusedarea ×2.4 | in reach6 passes, ~15 s | in reach8 passes, ~19 s | in reach8 passes, ~19 s | in reach4 passes, ~10 s | in reach8 passes, ~19 s | refusedarea ×3.1 | refusedarea ×5.3 |
| Square 1:12,048 × 2,048 | in reach4 passes, ~10 s | in reach4 passes, ~10 s | already this shape | in reach8 passes, ~19 s | in reach4 passes, ~10 s | in reach4 passes, ~10 s | refusedarea ×2.3 | refusedarea ×4.0 |
| Camera 3:26,000 × 4,000 | in reach8 passes, ~19 s | refusedarea ×2.7 | in reach6 passes, ~15 s | in reach4 passes, ~10 s | already this shape | refusedarea ×2.3 | refusededge 1,667 px deep | refusedarea ×2.7 |
| Phone photo 4:34,032 × 3,024 | in reach6 passes, ~15 s | refusedarea ×2.4 | in reach8 passes, ~19 s | in reach4 passes, ~10 s | in reach8 passes, ~19 s | in reach4 passes, ~10 s | refusededge 1,512 px deep | refusedarea ×3.0 |
| Widescreen 16:91,920 × 1,080 | already this shape | refusedarea ×3.2 | in reach6 passes, ~15 s | refusedarea ×2.2 | in reach8 passes, ~19 s | refusedarea ×2.7 | in reach4 passes, ~10 s | refusedarea ×2.3 |
Refused does not mean impossible. It means the new region is bigger than the evidence beside it.
Margins instead of shapes
Even margins are where the model is happiest, because every new pixel sits next to real photograph. On a 4032 × 3024 phone photo:
- 5% on every side — 4,436 × 3,326, in reach: 8 passes, about 19 s on one thread
- 10% on every side — 4,838 × 3,628, in reach: 6 passes, about 15 s on one thread
- 20% on every side — 5,644 × 4,234, in reach: 6 passes, about 15 s on one thread
How the three limits are measured
Area. The canvas may be at most 2 times the photograph. Past that, more than half of the file would be generated, and a continuation larger than the thing it continues stops being one.
Depth. The deepest new edge, times 2.5, must fit inside the canvas's short side: the model needs real picture beside every strip it fills.
Passes. The model works in 512-pixel squares; one run may take at most 9. Times are for WebAssembly on one thread, about 2.4 s a square; where the browser runs it on WebGPU it may be faster, and we do not print a figure we have not measured.
The awkward questions
Why will the AI not expand my phone photo to 16:9?
A 9:16 photo taken to 16:9 needs more than its own width again on each side. The model continues an edge from the photograph beside it, and it needs about one and a half times the depth of the new edge in real picture next to it. There is not that much picture, so the tool refuses rather than produce a smooth, confident, wrong frame — and offers the blurred cover, which is honest at that size.
What are the limits, exactly?
Three, all mechanical. The canvas may be at most 2 times the area of the photograph. The deepest new edge times 2.5 must fit inside the canvas's short side. And one run may take at most 9 passes of the 512-pixel model. The table is those three rules applied to six common photos and eight shapes.
Is the AI free, and does it upload my photo?
It is free and it runs on your device: a 26.6 MB model is downloaded once, on request, from our model file server, and your photo never goes anywhere. That is also why the limits are strict — it is a small model doing one thing well.
What do I use when the table says refused?
Blurred cover for busy photos, mirror for foliage and water, smart extend for skies and walls, one colour for prints and forms. All seven are instant and need no download, and every row's link opens the studio at that shape.
Open the tool
- Continue the edge with a modelAn inpainting model, 26,639,447 bytes, downloaded once on request, asked what carries on past the edge of your photograph.
- Smart extendThe seam's texture carried across and its colours run outward along the lines meeting the frame. No model, same every time.
- 9:16 to 16:9The tall phone frame opened straight into the wide one, with the width doing the growing.
- Extend an image on one sideWiden one side, or two, to make room for a title or a crop-safe margin — the rest of the frame stays put.
