Why Your Portraits Look Flat: Depth Mapping and Bokeh on Smartphone Cameras

How Smartphones Fake Shallow Depth of Field

When you tap portrait mode, your phone doesn't optically focus light differently — it takes a standard photograph and then mathematically blurs selected regions. The separation between sharp subject and soft background that traditionally required a wide-aperture lens (f/1.4–f/2.8) on a full-frame camera is replicated algorithmically. This is why the effect is called computational bokeh.

The entire process depends on a reliable answer to one question: how far away is each pixel? That answer is encoded in a depth map. If the depth map is accurate, the resulting blur looks convincing. Where the map is wrong, the illusion breaks — and you see the telltale signs of flat, artificial-looking portraits.

For a broader view of how software transforms the entire imaging pipeline, see our guide to computational photography.

How Depth Maps Are Built: Four Main Methods

Smartphones use several hardware and software strategies to generate depth data, and each has distinct accuracy characteristics:

• Dual-camera stereo: Two cameras separated by a small baseline compare the slight offset between their views — the same principle as human binocular vision. Depth accuracy improves with larger baseline separation, but fails on textureless surfaces where there's nothing to match between frames.
• Phase-detection autofocus (PDAF): Masked pixel pairs embedded in the sensor measure focus phase error across the image. This data doubles as shallow depth information, useful for identifying the focus plane but limited in spatial resolution across the full frame.
• Time-of-Flight (ToF) sensors: A dedicated IR emitter pulses light at the scene and measures how long each reflection takes to return, producing active, per-pixel depth readings. ToF is reliable in low light and independent of scene texture, but its native resolution is typically far lower than the main sensor.
• Single-camera machine learning: Models trained on large paired datasets infer depth from monocular cues — object semantics, relative scale, blur gradients. Accuracy depends entirely on whether the scene resembles the training data.

Most flagship phones combine two or more of these inputs, fusing them to build a higher-quality composite depth map. Understanding the underlying sensor stack helps explain why depth accuracy varies so significantly between devices — a detail worth noting when reading a camera spec sheet.

“Depth estimation from a single camera is fundamentally an ill-posed problem — the same 2D image can be consistent with infinitely many 3D scenes. Machine learning solves this by learning strong priors from data, but those priors fail when the scene departs from the training distribution.”

— Marc Levoy, Computational photographer and former principal engineer on Google's computational photography team

Why Edges Fail and Portraits Look Flat

Depth map errors concentrate at object boundaries — precisely the most visually prominent part of any portrait. Several conditions push algorithms toward mistakes:

• Fine or semi-transparent subjects: Hair, eyelashes, and eyeglass frames occupy the same pixel space as background elements, making depth classification inherently ambiguous.
• Low contrast edges: When the subject's clothing blends tonally with the background, the depth model has insufficient contrast gradients to locate the boundary reliably.
• Moving subjects: In phones that capture sequential frames for depth (some use brief burst sequences), subject motion between frames causes geometric misalignment that distorts the depth map.
• Depth discontinuities: Objects very close to the subject — a hand raised near a face, an object held at arm's length — sit on a sharp depth boundary that algorithms frequently misplace.

When the depth map misclassifies a foreground pixel as background, that pixel gets blurred as if it were far away. The result is the characteristic halo artifact around subjects, or conversely, background areas that remain sharp when they should be soft. This is the mechanical explanation for portraits that look flat or unconvincing even on capable hardware.

It's worth noting these limitations exist separately from sensor physics — a larger sensor won't fix a poor depth map. For what sensor size actually affects, see our sensor explainer.

Shooting Conditions That Improve Depth Accuracy

Because computational bokeh depends on scene geometry, small changes in how you compose a shot can meaningfully improve depth map quality — and therefore portrait output.

Increase subject-to-background separation. The more physical distance between your subject and the background, the larger the actual depth discontinuity for the algorithm to detect. One to two meters of separation helps most depth systems produce cleaner masks.

Ensure edge contrast. A background that's tonally different from the subject gives the edge-detection component of the depth pipeline more information to work with. Avoid matching colors between outfit and background.

Use good lighting. Even, well-lit scenes reduce noise in PDAF and stereo matching data. Low light forces higher ISO, which introduces luminance noise that degrades depth map precision.

Minimize motion. Keep the subject still during capture, particularly important on phones that use multi-frame depth capture. Some camera apps indicate this with a stabilization cue before shutter release.

Computational photography continues to evolve rapidly, with machine learning models improving depth estimation through semantic understanding — recognizing what an object is in addition to where it is. For a deeper look at AI's role in this pipeline, see AI camera features explained.