Resize, crop and AI upscale
Set the output size by percent or exact pixels, crop to the subject with padding, and rebuild small photos at 2x or 4x with Real-ESRGAN before the cutout.
The cutout keeps your photo's full resolution by default. Three settings change that on the way out: Resize sets the output dimensions, Crop to the subject trims the empty space around it, and AI upscale rebuilds a small or blurry photo at two or four times its size before the background is removed. None of them lowers quality unless you ask for a smaller size.
How do I remove a background without losing quality?
Leave Resize on Original. The segmentation model works on a small copy of the photo (512 or 1024 pixels square), but only to decide which pixels are subject. Its answer is scaled up to your photo's real size and applied to the original pixels, so the colours and detail in the output are the ones you started with. A 24 megapixel photo comes back as a 24 megapixel cutout.
The one ceiling is the output canvas: 16,777,216 pixels (16.7 megapixels), the largest canvas Safari will create. An export that would be bigger is scaled down proportionally to fit. That only affects very large originals or big upscales, and the Download button shows the final size in pixels before you click it.
What do the Resize options do?
| Option | Control | Result |
|---|---|---|
| Original | none | Output is the size of the photo (or of the crop) |
| Percent | Slider, 10% to 400% in 5% steps | Both sides scaled by the same factor. Switching to Percent starts you at 50% |
| Pixels | Width and height boxes, up to 12,000 each | Exact dimensions, see below |
With Pixels, the link button between the boxes decides how the two sides relate:
- Linked (the default): type one side and the other follows the photo's aspect ratio.
- Unlinked: type both sides. The subject is fitted inside that box and centred, never stretched. The spare space takes your background, or stays transparent. This is what marketplaces that demand a fixed square, such as 2000×2000, want.
Resize never runs an AI model. Going above 100% makes the photo bigger but not sharper; for that, use AI upscale.
How does Crop to the subject work?
Tick Crop to the subject and the output frame shrinks to the subject's bounding box, found from every pixel that is more than about 3% opaque (alpha above 8 out of 255). A Padding slider then adds space back, from 0% to 40% of the subject's longer side, 8% by default. Padding can reach past the edge of the original photo; the extra area gets your chosen background.
Crop happens before Resize. So "crop to subject, 10% padding, 1000×1000 pixels" gives a square product image with the item centred and an even margin, whatever the original framing was.
What does AI upscale do, and when can I use it?
AI upscale runs Real-ESRGAN general x4v3, a 4.9 MB restoration network (BSD-3-Clause), on the photo before the background is removed. Unlike a plain resize, it rebuilds edges and texture and cleans JPEG blockiness, so it helps most on small, old or heavily compressed photos.
- 2× and 4× are the choices. The network always works at 4×; for 2× the 4× result is scaled down by half, which still looks sharper than a 2× resize.
- It needs the GPU path: Chrome or Edge on a computer with WebGPU. Elsewhere the switch is disabled with the reason underneath.
- It is offered for photos up to 1.2 megapixels (about 1265×950). Larger photos already have the pixels it would invent; use Resize for them.
- The photo is processed in 192 pixel tiles with a 12 pixel overlap that is thrown away afterwards, so tile seams do not show, and a progress bar shows the tiles going through.
Changing the upscale factor after a photo is done re-runs only the upscale and the final compositing. The segmentation from the first run is kept in memory for that photo, so the subject outline does not change and the model does not run again.
In what order are these applied?
In the worker, once per photo:
- The segmentation model finds the subject on your original photo.
- If upscale is on, the photo is upscaled and the subject mask is scaled to the new size.
- Edge colours are recovered and the transparent PNG is written.
Then, each time you download or copy: crop, resize, the background, and encoding in the format you picked. That split is why changing the output settings is instant, while changing the upscale factor takes a moment.
If the downloaded file is larger than you want on disk, run it through the image compressor, or use Format converter to switch a PNG to WebP; both run client-side.