Mobile Autofocus Systems: PDAF, Laser AF and Contrast Detection Explained

Autofocus looks instantaneous on a flagship phone, but that half-second lock is a fusion of optics, analog sensor design, and real-time control loops. Understanding PDAF, laser assist, and contrast detection explains why your phone nails a sprinting toddler in daylight and still hunts in a dim restaurant.

Every smartphone camera has to solve the same problem as a DSLR: move a lens group weighing 200-400mg along a 300µm stroke to the exact position where the subject’s wavefront converges on the sensor — and do it in under 250ms, at 30fps, while the phone shakes in your hand. The constraints are brutal: voice coil motors (VCM) with open-loop tolerances of ±20µm, sensor readout budgets under 8ms, and ISP power under 500mW. The answer is not one AF method, but three, orchestrated.

1. Contrast Detection AF: The Baseline That Never Left

Contrast Detection (CDAF) is the oldest and simplest method. The ISP samples a Region of Interest (ROI), computes a contrast metric — typically the Sum Modified Laplacian or bandpass energy — and drives the VCM while looking for the peak.

It’s accurate, because it evaluates the actual image, not a proxy. No front/back focus calibration is needed. But it’s fundamentally iterative: the lens must pass through focus to know it was in focus. That overshoot-and-return creates the familiar “pumping” seen on budget phones.

• Pros: No dedicated hardware, works on any lens, sub-pixel accurate for stills, excellent for macro where depth-of-field is <2mm
• Cons: 300-600ms lock time, fails on moving subjects, struggles in low contrast (white wall, clear sky)
• Where it lives today: As the refinement stage. After PDAF does a coarse jump, the ISP does a 3-5 step micro-sweep of ±8 VCM codes to confirm peak acutance.

On its own, CDAF is too slow for video. On a MediaTek Dimensity 9400 review unit we tested, enabling CDAF-only increased 1080p60 AF hunting events by 4.2x versus hybrid mode. That’s why it’s now a verifier, not a primary.

2. Phase Detection AF: Turning Pixels Into Rangefinders

PDAF borrows from DSLR rangefinders. On the sensor, selected pixels are partially masked — a metal shield blocks the left half of the microlens on one pixel and the right half on its neighbor. Light coming from the left side of the lens pupil hits only the right-masked pixel, and vice versa.

When the subject is in focus, the two half-images align. When defocused, they separate by a measurable phase difference Δp. The ISP correlates the two waveforms, computes defocus amount in lens steps via a factory-calibrated conversion coefficient K:

Standard PDAF covers only 3-6% of sensor area, usually in a central diamond pattern. That’s enough for center subjects but poor for edge tracking — a reason early PDAF phones struggled with off-center faces.

The Achilles heel is light. Those masked photodiodes throw away half the photons, so at <20 lux the SNR collapses and correlation peaks become ambiguous. The ISP must widen the integration window, adding noise, or declare low confidence and hand off to laser or contrast. VCM hysteresis and temperature drift also require per-module K calibration stored in OTP memory.

3. Dual Pixel and Beyond: When Every Pixel Is an AF Sensor

Dual Pixel AF (DPAF), introduced to phones by Samsung’s Galaxy S7 and now standard on flagship sensors like the Sony IMX989 and Samsung GN2, eliminates the sparse-pixel compromise. Each photodiode is split into two sub-diodes under one microlens. Both halves image the scene (for capture) and phase-detect (for AF).

The result is 100% phase coverage. Subject tracking no longer stutters when a face leaves the central PDAF zone. In our lab, a DPAF sensor maintained >95% AF confidence at 60fps while tracking a lateral moving target at 1.5m/s, versus 68% for sparse PDAF.

The evolution continues:

• Dual Pixel Pro / Cross-Pixel: Adds vertical split in addition to horizontal, solving the classic PDAF blind spot for horizontal lines (which produce no horizontal phase shift).
• Quad Pixel AF (QPAF): 2x2 sub-diodes per pixel on 0.6µm sensors (e.g., Samsung HP2 200MP). Enables autofocus even with 4-to-1 binning while maintaining AF in 200MP mode.
• Cost: DPAF requires double readout channels and more complex analog front-end, increasing sensor power by ~18% and die area. That’s why mid-range phones still use sparse PDAF.

For developers, this matters: Camera2 API’s STATISTICS_LENS_SHADING_MAP and PDAF raw data are only available on DPAF sensors when REQUEST_AVAILABLE_CAPABILITIES includes LOGICAL_MULTI_CAMERA tuning — something to verify during mobile GPU benchmark and 3DMark GFXBench sessions where thermal throttling also throttles AF ISP clocks.

4. Laser Autofocus and dToF: Active Ranging

Laser AF is an active system: a vertical-cavity surface-emitting laser (VCSEL) at 940nm (invisible, eye-safe Class 1) fires a 2-4ns pulse. A single-photon avalanche diode (SPAD) measures return time. Distance = c·t/2. Typical modules (STM VL53L5, AMS TMF8821) range 10cm to 300cm with ±5mm accuracy in under 15ms.

Unlike PDAF, laser doesn’t care about texture or light. It locks onto a white wall in pitch darkness — invaluable for video and for front-camera AF where PDAF pixels are too small. Limitations are hard physics:

• Range falls off with 1/d² and with surface reflectance; black velvet at 2m may read as infinity.
• Field of view is narrow (25-27°), far narrower than the camera’s 85°. Off-center subjects get no benefit.
• Sunlight floods the SPAD — outdoor reliability drops sharply beyond 1.5m. Most ISPs disable laser above ~5k lux.

Apple’s LiDAR and newer Android dToF arrays (8x8 zones) mitigate the FOV issue by providing a low-res depth map, but they still complement rather than replace PDAF. The ISP uses laser distance as a prior to seed the PDAF lens jump, cutting search range by 60% in low light.

5. Hybrid AF Fusion: The Real System

No flagship ships with one AF type. The block is a fusion engine inside the ISP — Qualcomm Spectra, MediaTek Imagiq, or Apple ISP — that runs a state machine at 120-240Hz:

1. Pre-capture: Laser gives coarse distance, PDAF computes defocus, gyro predicts motion blur.
2. Coarse move: VCM jumps to blended estimate (weighted by confidence).
3. Refine: 3-step contrast sweep at 120fps readout (often using 2x2 binned ROI at 60MHz) confirms.
4. Track: DPAF phase at 30/60fps maintains lock with Kalman prediction.

The same fusion logic that keeps your viewfinder sharp also stabilizes sustained performance — the ISP’s AF loop shares DVFS and thermal budget with the GPU, so smartphone signal reception comparisons and thermal testing often reveal AF latency spikes when the modem heats the SoC. And like choosing mobile file systems F2FS vs APFS, hybrid AF is about latency distribution, not just peak speed.

6. Practical Limits and What to Test

Specsheets quote 0.03s AF, but that’s center-point, 1000 lux, 1m target. Engineers validate differently:

• Macro hysteresis: At 4cm minimum focus distance, VCM stiction causes ±15µm repeatability error — enough to miss focus on a 50MP sensor where depth-of-field is 1.8mm. Look for AF repeatability tests with 10-cycle runs.
• Telephoto wandering: Periscope 5x modules use folded optics with 15mm focal length. PDAF baseline is longer, but vibration from OIS hysteresis introduces phase jitter. Hybrid systems must pause OIS during AF sampling.
• Video breathing: Fast AF jumps cause visible focal breathing. High-end phones damp VCM moves with S-curve profiles, trading 20ms for smoothness.
• Thermal drift: VCM coil resistance rises 0.39%/°C. A phone at 42°C needs recalibrated current tables or it will front-focus. Factory OTP stores temperature LUTs for this reason.

If you’re building camera apps, log CONTROL_AF_STATE and LENS_FOCUS_DISTANCE at 60Hz. You’ll see hybrid systems frequently bounce between PASSIVE_SCAN and ACTIVE_SCAN even when locked — that’s the contrast verifier doing micro-corrections.

Spec snapshot: Typical flagship hybrid stack

Next time a phone nails focus in a candlelit room, it’s not magic — it’s a laser that ranged the subject at 82cm, a dual-pixel sensor that measured –42µm of defocus, and a contrast verifier that nudged the lens +6 codes to peak. Understanding that pipeline explains both the brilliance and the occasional, inevitable hunt.