Under-Display Camera Technology: How UDC Works and Current Limitations

The under-display camera promises the final frontier of all-screen design: no notch, no punch-hole, no motorized popup. Yet putting a camera behind an active OLED panel means shooting through a diffraction grating. Here is how engineers make it work — and why physics still wins.

1. The Architecture: Making Pixels Disappear

A conventional OLED display is opaque by design. The cathode is a continuous sheet of low-work-function metal (MgAg), reflective and light-blocking. The TFT backplane, pixel definition layer, and encapsulation all absorb or scatter light. For a UDC window — typically a 3-5mm circular region — the entire stack must be re-engineered for transmission.

There are two primary approaches in mass production today:

Transparent Cathode + Sparse Pixel Layout

Used by ZTE (Axon 30/40 Ultra), Xiaomi MIX 4, and RedMagic 8 Pro. The metal cathode over the UDC area is patterned or thinned to ~70-80% transparency and replaced with a transparent conductive oxide like ITO or IZO. Simultaneously, pixel density is reduced. Instead of 400 PPI, the UDC region runs at ~200 PPI, with larger gaps between subpixels. This raises the aperture ratio from ~15% to ~45%, dramatically increasing photon throughput to the sensor below.

To avoid a visible square, manufacturers use a demultiplexed pixel arrangement. Four real pixels drive one larger UDC-area pixel via time-division or subpixel sharing. When the camera is inactive, the sparse pixels are driven at higher luminance to match surrounding brightness. When active, they are switched off entirely (0 nits) to prevent OLED emission from flooding the sensor.

High-Transparency OLED on Flexible Substrate

Samsung Display's approach for the Galaxy Z Fold series uses a laser-drilled flexible PI substrate and reorganizes the TFT traces. Rather than removing pixels, it routes the TFT circuitry to the periphery of the UDC zone, leaving a near-transparent window with micro-lenses over each subpixel. Transmission is lower than ZTE's brute-force method (~38% vs ~55%) but display uniformity is superior, critical for foldables where the UDC doubles as a main display region.

2. The Physics Problem: Diffraction Is Inevitable

Even with 50% open area, the UDC region is a periodic 2D grating. The pixel pitch (80–120 µm) is orders of magnitude larger than visible wavelengths (0.4–0.7 µm), creating strong Fraunhofer diffraction. A point light source — a streetlamp at night — becomes a starburst with haze, flare, and reduced contrast. Mathematically, the captured image is the convolution of the ideal image with the display's point spread function (PSF).

This PSF is content-dependent because the grating changes when pixels are on versus off, but during capture pixels are off, so the PSF is fixed and can be measured in the factory with a collimated laser. The modulation transfer function (MTF) typically drops to 35–50% at Nyquist compared to an unobstructed lens, with severe loss in the blue channel where diffraction is strongest.

Two additional artifacts compound the problem:

• Fresnel reflections: The air-gap between display and camera lens creates multiple reflections. Without a polarizer (removed over the UDC window to improve transmission), 4–6% of incoming light bounces between cathode and lens, producing ghosting.
• Color imbalance: Transparent cathodes and polyimide substrates absorb more blue (380–450 nm) than red. Uncorrected, UDC images have a yellow-green cast and desaturated blues.

This is fundamentally different from a periscope telephoto lens engineering challenge, where optics are constrained by z-height but remain unobstructed. UDC trades optical purity for form factor, pushing complexity into computation.

3. Display Compensation: Hiding the Window

Users notice a UDC not in photos, but on the screen itself. A sparse region looks dimmer or grainier. Compensation is handled at three layers:

1. Optical index matching. A transparent filler with refractive index ~1.52 replaces the air gap precisely over the camera cutout, minimizing internal reflections and equalizing haze. Some panels add an anti-reflection coating on the inner glass.

2. Subpixel rendering & dithering. The display driver IC applies a spatial filter that borrows luminance from neighboring full-density pixels. A diamond PenTile-like arrangement with dual-driver transistors maintains perceived resolution. At typical viewing distances (>25 cm), most users cannot distinguish the 200 PPI island.

3. Dynamic luminance calibration. Each panel is factory-calibrated with a camera measuring luminance uniformity. A 7×7 compensation matrix per UDC pixel is burned into the DDIC LUT. This adds ~$3–5 to panel cost but is essential; without it, the UDC appears as a faint rectangle on white backgrounds.

Durability engineering also matters. Removing the polarizer reduces impact resistance and moisture ingress protection. That is why UDC devices require rethinking waterproof smartphone IP68 engineering seals — the filler must be hermetic and the edge adhesive widened to prevent capillary ingress around the transparent window.

4. The Computational Imaging Pipeline

Raw UDC frames are unusable without restoration. The pipeline, running on the ISP and NPU at 30+ FPS, typically executes:

// Simplified UDC ISP pipeline (per-frame, 16-bit RAW)
RAW_16 raw = sensor.read();          // T-compensated exposure, +1.5 to +2 EV gain
RAW_16 dehazed = wienerDeconv(raw, factoryPSF); // PSF from laser calibration
RGB_16 rgb = demosaic(dehazed);
rgb = colorCorrect(rgb, UDC_CCM);      // 3x3 CCM compensates for PI yellowing
rgb = flareSuppress(rgb, irFlareMap);  // IR model removes diffraction ghosts
RGB_8 out = hdrMerge(rgb, dlssrNet(rgb)); // Learned restoration (e.g., ZTE UDC-Net)

The key step is deconvolution with a measured PSF. Classical Wiener filtering is fast but amplifies noise. Modern flagships use a lightweight U-Net (typically 2–4M parameters, int8 quantized) trained on paired datasets: the same scene captured simultaneously through a UDC panel and a reference lens. ZTE's Visible-UDC dataset and Samsung's SID dataset are industry standards for training.

Performance hinges on accurate PSF estimation. Some vendors embed a micro-prism sensor to re-measure PSF at runtime to account for temperature drift and panel aging — OLED transmissivity degrades ~2% per year as organic layers oxidize.

Importantly, this processing must be secured and low-latency. Unlike cloud photo enhancement, UDC restoration runs in the trusted execution environment to prevent raw frame leakage. Our analysis of mobile app network security TLS practices shows similar principles apply: sensitive sensor data should never leave the SoC unencrypted.

5. Where UDC Stands in 2026: Limitations That Remain

Despite three generations of refinement, UDC image quality still lags a conventional 12–16 MP front camera by a measurable margin. The compromises are systematic:

Resolution vs. transparency. Pushing open ratio beyond 60% requires removing more OLED material, making the window visible. The industry has settled at 40–55% as the sweet spot. This caps effective light to the sensor at roughly 1/2 to 1/3 of normal, raising the noise floor by 4–6 dB. In daylight this is masked by multi-frame stacking; in low light, UDC selfies exhibit watercolor-like smearing.

Video is harder than photos. Deconvolution is temporal. Per-frame U-Nets flicker if not temporally consistent. Current NPUs can run a 720p restoration at 30 FPS within a 1.2W power budget, but 4K60 UDC video remains out of reach without throttling. Most devices limit UDC video to 1080p30, while punch-hole phones offer 4K60.

Interaction with display workload. Unlike discrete sensors, UDC shares resources with the GPU and display pipeline. The mobile ray tracing hardware that powers modern gaming competes for NPU and memory bandwidth with the UDC restoration net. Under heavy load, some devices dynamically downscale restoration quality to maintain frame rate.

*Illustrative scores extrapolated from lab MTF and flare tests. Not official DXOMARK.

6. Should You Design Around UDC Today?

For engineers, the answer depends on the product definition. If the goal is a true full-screen media device where selfie quality is secondary — a gaming phone, a foldable tablet — UDC is justifiable. RedMagic's audience tolerates softer front video because the uninterrupted 120 Hz OLED is the selling point.

If the device is creator-focused, UDC is still a compromise. No computational pipeline can fully reconstruct high-frequency detail destroyed by diffraction. Information theory sets a hard limit: frequencies beyond the first diffraction null are lost, not merely attenuated. AI hallucination can invent plausible skin texture, but it is synthesis, not capture.

The next leap will not be a better neural net but a better optical window. Two paths are promising: micro-LED islands where emissive area is an order of magnitude smaller, enabling >70% transmission with invisible gaps, and electrochromic shutters that make the UDC region fully transparent on demand and fully opaque otherwise, eliminating the need for a sparse pixel compromise. Both are lab-stage in 2026, with pilot production forecast for 2028.

Until then, UDC remains an elegant engineering trade: we finally removed the notch by making the display itself the lens filter — and then spent a whole ISP pipeline trying to undo what the display did to the light.