Cutting a subject out of a photo used to be careful work with a selection tool. A model now does it in a second, and for most pictures the result is better than a hurried manual attempt.
But the failures are predictable, and knowing them in advance is the difference between one clean export and twenty frustrating attempts.
The model looks for a subject that is distinguishable from what is behind it. The more your photo looks like that, the better the edge:
Fine hair, fur, glass, smoke and transparent fabric are hard for the same reason: at the edge, each pixel is genuinely part subject and part background. Tools approximate it, and on a plain background they approximate it well; against a busy one the halo gives it away.
If a photo like that matters — a product shot for a listing — reshooting against a plain wall takes less time than repairing the cutout.
Transparency needs a format that can store it. JPEG cannot, so saving a cutout as JPEG fills the transparent area with white — which looks fine until you place it on a coloured page. PNG and WebP both keep transparency; PNG is the safer choice because everything accepts it.
If your destination needs a plain background anyway, such as a passport-style photo, export the cutout and place it on a solid colour rather than skipping the cutout.
Most background removers upload your picture to their servers, which for photos of people is worth a second thought. Abzar's runs the segmentation model inside your browser: the model downloads once, then every cutout happens on your own device. The first run takes a moment while the model loads; after that it is quick, and your photo never goes anywhere.