How background removal really works (and why it sometimes fails)

"Remove background" buttons feel like magic. Behind the curtain is a small neural network that has been trained to look at an image and decide, pixel by pixel, what counts as the subject and what counts as backdrop.

Segmentation in one sentence

The model outputs a mask — a grayscale image the same size as your photo — where white pixels mean "keep" and black pixels mean "drop". The original image gets multiplied by that mask, and the result is composited onto a transparent canvas.

Why hair, fur and glass are hard

Soft edges (hair strands, smoke, transparent objects) require the model to output not just a binary mask but a fractional one. Older models snap edges to black or white, which is why you sometimes see crunchy outlines on portraits. Modern matting networks predict that fractional alpha directly.

Practical tips

• Shoot against a contrasting backdrop when you can — even modern models love high contrast.
• Re‑run the result on a coloured background to spot fringing fast.
• If the subject blends into the backdrop, manually paint a quick foreground hint before re‑running.