Smartphone portrait modes combine optical information, computational depth estimation, segmentation and synthetic blur. Their output can look convincing in simple scenes yet fail around hair, transparent objects, fine structures, reflective surfaces and moving subjects. This article sets out a reproducible method for comparing portrait-mode depth and bokeh across Apple, Samsung and Google devices. It does not report a completed head-to-head test or claim that one manufacturer leads.
For a manufacturer example of depth capture and adjustable portrait blur, see Apple’s guide to depth information and adjustable Portrait mode blur.
For a broader explanation of computational depth and portrait blur, see our guide to computational bokeh.
A valid comparison requires named phone models, camera modules, software versions, capture settings and repeated test runs. Without those records and the resulting image set, vendor rankings and numerical scores would be speculation.
What a portrait-mode benchmark should measure
Separate the pipeline into observable outcomes rather than assuming access to proprietary depth maps or internal processing stages. The principal dimensions are subject-boundary accuracy, foreground-background ordering, blur transition quality, consistency across focal distances, temporal stability where live preview or video portrait effects are tested, and capture-to-output latency.
Boundary and segmentation quality
Inspect hair, glasses, fine fabric, foliage, fingers, thin objects and gaps between limbs. Record edge leakage, missing subject regions, background fragments included in the subject mask, and halos. Use a reference mask created independently of the tested phone if numerical segmentation scores are required.
For a comparison of smartphone and professional camera image processing, see smartphone computational photography versus professional cameras.
Depth ordering and occlusion
Construct scenes with overlapping objects at known distances. Check whether the system preserves the correct front-to-back order and whether blur is applied consistently across occlusion boundaries. Do not infer metric depth accuracy from a visually pleasing blur effect.
Blur rendering
Assess blur smoothness, highlight shape, transition bands, foreground blur, and whether blur strength changes plausibly with distance. Synthetic bokeh is an aesthetic rendering choice as well as a depth-dependent effect; compare it against a declared reference, not an assumed single “correct” blur.
Controlled capture protocol
- Record exact phone model, rear/front camera used, lens equivalent, OS version, camera-app version and portrait settings.
- Use the same scene layout, subject distance, lighting, framing and output orientation. Record any unavoidable differences in focal length or camera selection.
- Capture a static scene and a separate moving-subject scene. Include hair, glasses, transparent objects, fine mesh, reflective surfaces, low texture and complex backgrounds.
- Repeat each condition at least several times, including changes in subject distance and illumination. Preserve original outputs and metadata.
- Evaluate at native resolution and at a consistent display size. Keep image scaling and sharpening consistent when making crops for review.
Scoring without pretending to measure hidden internals
Use a published rubric and report the individual dimensions. A simple ordinal scale can be used for human review, but the rubric must define each score and reviewers should assess images without knowing the device identity where practical. If numerical segmentation metrics are used, publish the reference masks, calculation method and test set. Report inter-reviewer agreement and the number of samples.
For another controlled comparison of depth estimation and bokeh rendering, see our article on computational photography versus traditional capture.
| Test scene | Failure to inspect | Evidence to retain |
|---|---|---|
| Hair and fine fabric | Edge cut-outs, halos, missing strands | Original output and reference crop |
| Glass and transparent objects | Incorrect masking or blur | Scene setup and repeated captures |
| Moving subject | Temporal inconsistency or motion artefacts | Video sequence or burst, where supported |
| Low light | Noisy boundaries and unstable depth | Lighting, exposure and device settings |
| Reflective highlights | Unnatural blur discs or clipped edges | Full-resolution output and review notes |
Limitations and interpretation
Portrait modes may use different lenses, crop factors, sensor data and processing paths. Matching the scene does not guarantee identical input data. Proprietary software updates can change results, and a test conducted on one phone generation cannot establish performance across an entire brand. State these limits alongside any findings.
Do not describe a proposed protocol as a completed benchmark. A comparative conclusion requires the named devices, dated software, captured sample set, scoring results and documented limitations.
Conclusion
A useful portrait-mode comparison is built from controlled scenes, repeatable captures and transparent scoring. The methodology here can support a future test of Apple, Samsung and Google devices, but no vendor winner is claimed until the evidence has been collected and published.