How does smartphone Portrait Mode produce bokeh?

What's really being asked

The interviewer checks whether you understand Portrait Mode as a combination of depth estimation, segmentation, and physically motivated rendering, and that the hard parts are at object boundaries.

The full answer

First you need a depth signal. Phones get it from dual-pixel sensors that yield tiny disparity between the left and right halves of each pixel, from a second rear camera giving stereo baseline, or from a learned monocular depth network when only one view exists. Second you need subject segmentation or matting, often a neural network that isolates the person, refined to capture fine boundaries like hair. Third you render the blur: rather than a flat Gaussian, you apply a spatially varying blur whose radius grows with depth behind the focal plane and synthesize disk-shaped bokeh from bright highlights, mimicking a real lens. The depth map and the segmentation mask are typically fused so the subject stays sharp while the background is progressively blurred.

The mistakes people make

Saying it is just a Gaussian blur of everything outside a mask, with no depth gradient. Ignoring depth estimation entirely. Forgetting the boundary and hair challenges that make matting hard.

What usually comes next

Why dual-pixel disparity is small and noisy, how learned depth is supervised, handling translucent or thin structures, and rendering convincing bokeh highlights rather than mush.

A concrete example

The Pixel approach combines a person-segmentation network with dual-pixel disparity to build a depth map, then renders a depth-dependent blur so a face stays crisp, the shoulders blur slightly, and the distant background blurs heavily with rounded highlight disks, approximating an optical large aperture.