There's a specific moment in hardware history that made modern beauty AR viable. It wasn't a software breakthrough, an AI paper, or a new rendering technique. It was the year smartphone manufacturers stopped treating the front camera as an afterthought. When the front camera crossed 12 megapixels and manufacturers added depth-sensing infrastructure for Face ID and portrait selfie modes, they inadvertently built the hardware foundation that no-code beauty try-on needed. We've been watching this convergence since before we wrote our first line of rendering code.
Why Resolution Alone Wasn't Enough
Early front cameras on phones were, frankly, not good. From 2012 to roughly 2018, they topped out at 5–8 megapixels with fixed focus, poor dynamic range, and no depth information. You could build rudimentary face filters on them — the kind that stick cartoon ears on your head — but you couldn't do precision cosmetic rendering. The pixel density wasn't there to accurately resolve lip contours, eye socket geometry, or the exact boundary where foundation coverage would meet bare skin.
Resolution was part of the problem, but it wasn't the only constraint. Face mesh accuracy depends on being able to detect and track hundreds of discrete facial landmarks at video frame rates. Earlier front cameras produced enough JPEG noise and motion blur that landmark detection models had to run at lower confidence thresholds — meaning the mesh would drift noticeably when the subject moved, causing the virtual product layer to "swim" on the face. This is the tell-tale sign of low-quality AR: the product doesn't stay locked to the face during natural movement.
The Depth Sensor Factor
The real inflection point was depth. Face ID, introduced in 2017, required accurate 3D face reconstruction for secure authentication. To build that, Apple developed a structured-light infrared projector array — the TrueDepth camera system — that projects 30,000 invisible infrared dots onto the face and reads the deformation to compute a precise depth map at 60 fps. This wasn't primarily a consumer camera feature. It was a security system that happened to produce the most accurate 3D facial geometry data ever made available at the consumer level.
The side effect for AR developers was enormous. For the first time, you could get a per-pixel depth estimate at 60 fps from a front-facing camera on a widely distributed consumer device. That depth data enables proper 3D surface registration — the AR layer can follow the actual 3D contours of the face, not just a 2D approximation. Foundation that follows the nose bridge looks different from foundation that's painted flat over a 2D face image, and shoppers can feel the difference even if they can't articulate it.
Android manufacturers followed in subsequent generations. By 2021, time-of-flight and structured-light depth cameras had spread to flagship and upper-midrange Android phones from Samsung, Google, and others. By 2023, depth sensor capabilities had reached enough of the mid-range market that we estimated over 60% of active US smartphones had usable depth data available to WebAR contexts.
On-Device ML: What Changed When NPUs Arrived
Hardware for the rendering side wasn't the only piece. Running an ML-based facial landmark model at 30–60 fps requires real compute. The shift that made this viable without draining the battery or requiring a server round-trip was the arrival of dedicated Neural Processing Units in smartphone SoCs.
Apple's A12 Bionic (2018) was the first consumer chip with an integrated Neural Engine capable of 5 trillion operations per second. The subsequent generations pushed this to 15.8 TOPS in the A14. Google's Tensor chips brought similar on-device ML throughput to Pixel devices starting in 2021. Qualcomm's Snapdragon 8-series integrated Hexagon DSPs with dedicated AI acceleration from 2019 onward.
The practical result for beauty AR: a lightweight facial landmark model can run inference in under 3 milliseconds on an NPU, compared to 18–25 milliseconds on the CPU alone. That difference is the margin between try-on that feels live and try-on that feels like a slide show. Our own face mesh runs at 4.2ms median inference time on a Snapdragon 888 — well within the budget needed to hit 60 fps compositing in the browser.
Every flagship phone since 2019 ships with more on-device ML compute than what powered entire cloud inference clusters in 2016. That hardware democratization is what made client-side beauty AR practical — not just theoretically possible.
— Yemi Adebayo, CTO & Co-Founder, Lumeglint
WebGL and the Browser Rendering Budget
One more hardware unlock was required: GPU capability exposed through the browser. WebGL has been available since 2011, but early mobile GPU implementations had significant limitations. Many mobile browsers restricted shader complexity, blocked certain texture formats, or throttled GPU access to preserve thermal budget.
Between 2020 and 2023, mobile GPU performance scaled dramatically — driven largely by mobile gaming demand — and browser vendors improved their WebGL implementations correspondingly. Safari on iOS added WebGL 2.0 support in 2021. Chrome's mobile WebGL performance on Android improved substantially through the Chromium 90–100 release cycle. The result is that a modern mobile browser can run a reasonably complex fragment shader at 60 fps without thermal throttling on any flagship from 2022 onward and on mid-range devices from 2023.
This matters because it eliminated the need for a native app. Early AR beauty tools required app downloads because the rendering stack couldn't run performantly in a browser. That friction killed adoption — shoppers wouldn't install an app to try on a lipstick. Browser-based WebAR removes that barrier entirely. The try-on lives on the product page, three taps away from add-to-cart.
Where the Hardware Is Today — and What's Still Constrained
As of late 2025, around 78% of US smartphone users are on devices capable of running real-time AR beauty try-on in a browser with acceptable quality. The remaining 22% are primarily on older low-end Android devices with weak GPUs and no depth sensing. For most beauty DTC brands, this coverage rate is sufficient to justify the investment — especially since shoppers who are buying color cosmetics online tend to skew toward the more current device population.
| Hardware Feature | Year Became Widespread | AR Benefit |
|---|---|---|
| 12MP+ front camera | 2019–2020 | Accurate landmark detection; less mesh drift |
| Depth sensor (TrueDepth / ToF) | 2017 (iOS), 2021 (Android mid-range) | 3D surface registration for accurate product overlay |
| Dedicated NPU / Neural Engine | 2018 (iOS), 2019–2020 (Android) | Sub-5ms ML inference; 60 fps landmark tracking |
| Mobile WebGL 2.0 in browser | 2021 (iOS Safari), 2020 (Chrome Android) | No native app required; browser-native rendering |
The hardware story isn't finished. Front cameras continue improving — depth sensing is reaching more mid-range devices every year, and Apple's newer structured-light systems provide even higher-density depth maps. Each hardware generation expands the coverage floor and raises the quality ceiling for in-browser beauty try-on. For the first time in the category's history, the hardware constraints are largely behind us. The remaining work is on the software and content side — and that's where we focus now.
