Most descriptions of face AR start with the effect and work backwards. Teams see a sticker that tracks a face, decide they want one, and only later discover that the property they actually need is consistency: the same look behaving predictably in the viewfinder, on a recorded clip, and on a stream going out live. Effects are easy to demo. Consistency across those three moments is the engineering problem.
This page is a plain account of what the Meishe Face AR SDK is, what it is made of, what it runs on, and who should not adopt it. It is written for teams early in an evaluation, which is usually when the wrong questions get asked — “how many stickers does it ship with” rather than “who owns the camera interface and what happens to our interface when the catalogue doubles”.
We will also state the limits early, because they shape the decision more than the capability list does.
1. Product Positioning
Meishe Face AR SDK is the face analysis and AR rendering module of the cross-platform audio/video processing engine used across the product family. It covers face detection and tracking frame by frame, gesture detection, avatar-driven props and stickers, and beauty and filters evaluated on the same render path, plus optional AI segmentation, intelligent color correction, and AI subtitles.
The same engine serves iOS, Android, Web, Desktop (Windows), smart hardware, React Native, and Flutter (supported through the short-video solution). It does not ship a finished camera UI: your team owns recording controls, permissions, asset browsing, and publishing.
2. What the module is made of
| Part | What it does | Where you meet it |
|---|---|---|
| Face detection and tracking | Finds faces frame by frame and follows them | Prop anchoring, subject framing |
| Gesture detection | Reads hand raise, switches props, advances template steps | Touch-free interaction |
| Avatar-driven props and stickers | Anchors 2D/3D content to the face or the frame | Campaign and entertainment libraries |
| Beauty and filters | Skin smoothing, reshaping, makeup, adjustable filter intensity | Same render path, no second pipeline |
| AI segmentation and correction | Separates subject from background, intelligent color correction, subtitles | Optional module set |
| Three AR points | Capture, timeline editing, live streaming, one effect stack | Anywhere AR appears in your product |
Each part is usable on its own. A team can adopt tracking and props without beauty, or adopt beauty without gesture interaction, and request the corresponding modules only.
3. The engine underneath
Meishe brings over 20 years of development experience in the field of video and audio, holds more than 70 video-related patents, and has won many national awards. What matters more for an integration team is narrower: the Face AR module is not a separate component. It is a consumer of the same engine that handles timeline evaluation, GPU rendering, audio mixing, and encode elsewhere in the product family. The practical consequence is continuity — adding face effects later does not introduce a second render pipeline with its own colour behaviour, export semantics, and set of bugs.
Because the engine already exists per target, adding the face layer is mostly a matter of which modules you request and where you mount the surface.
One engine across clients means effect definitions and export semantics do not have to be re-derived per platform. Device capability still differs; definitions do not.
4. Seven properties that shape an integration
Some of these are easy to verify in a demo; others only appear months later. All seven are worth asking about before signing anything.
- Real-time rendering without pre-processing. Effects evaluate on frames as they arrive rather than requiring a transcode or proxy stage. What the user approves in preview is what gets written, which removes an entire class of disagreement between previewed and exported media.
- Shared detection work. One face analysis per frame feeds tracking, beauty, reshaping, makeup, and prop anchoring. AR props use the resulting reference points as an attachment anchor, which is why a prop holds through turns and brief occlusion — and why adding a prop does not add a second analysis to the loop.
- On-demand module loading. The core is around 3 MB, and heavier optional modules load only when requested, so a build using tracking and props does not carry the AI set.
- Resolution headroom. Up to 8K video editing and 16K image processing are supported, bounded by whatever the host device’s decoder allows.
- HDR-aware pipeline. HDR preview, mixed HDR and SDR material on one timeline, and export following either the HLG or the ST 2084 transfer curve.
- Materials as content, not code. Props, stickers, filters, and beauty looks ship as packages describing what an effect does. Updates are content releases, not engineering projects.
- Acceleration through standard browser facilities. On the web there is no plug-in and no driver install, which keeps deployment simple and ties available headroom to the browser and device in front of the user.
5. AR at three points, one effect stack
Face AR tends to enter a product at one point and spread to the others. Capture is usually first: the user frames themselves and sees the effect live. Timeline editing comes next when recorded clips need the same look applied to a segment. Live streaming arrives when the effect needs to render on frames leaving the device in real time.
Keeping one effect stack across all three is what prevents the most reported class of AR bug — a prop that behaves correctly during capture and shifts once applied to recorded material. It also means a single fix or tune lands everywhere rather than being ported three times.
The snippet below is schematic. Module names, method names, and signatures are illustrative and are not the shipped API surface.
// Illustrative: one session description usable at capture, edit, and stream.
import { createFaceArSession } from "@meishe/face-ar"; // placeholder name
const session = await createFaceArSession({
mount: await resolveSurface({ mode: "capture" }), // or "timeline" / "stream"
modules: ["detect", "props", "beauty"],
effects: [
{ id: "prop.avatar.fox", anchor: "face", intensity: 0.7 },
{ id: "skin.smooth", intensity: 0.35 },
],
});
await session.setPoint("timeline"); // same stack, decoded frames instead of camera frames
await session.setPoint("stream"); // same stack, frames leaving in real time6. Who it suits, and who should look elsewhere
It suits teams that already own a camera or editing surface and want face capability underneath it: short-form creation apps, live streaming products, social features, template-driven marketing tools, onboarding and training recorders, and desktop tools doing preview and batch asset work.
It suits teams that need the same behaviour on several targets, because the engine rather than the interface carries the definitions.
It is a poor fit for teams expecting a finished camera application. The SDK does not ship one — recording controls, permission prompts, upload flows, and publishing remain your work, and if those are exactly what you hoped to buy, this is the wrong shape of product. It is also unsuitable if your evaluation rests on promises nobody can test: no engine recovers tracking quality from a face the lighting never revealed.
7. What arrives on day one
Alongside the modules themselves, integration includes production-ready demo source code structured to be productized rather than read once and abandoned, 1V1 technical support during integration, and authoring tools plus training for the designers who will produce effect content.
Support also extends to automation: Meishe publishes Integration Skills for use inside AI coding assistants, covering Codex, Qoder, Claude Code and Cursor workflows, for teams whose integration work already happens with an assistant in the loop.
FAQ
What exactly does the Face AR SDK include?
Face detection and tracking frame by frame, gesture detection, avatar-driven props and stickers, beauty and filters on the shared render path, and optional AI segmentation, intelligent color correction, and AI subtitles — all delivered as modules you can adopt selectively.
Do we have to take beauty if we only want props?
No. Modules load on demand, so a props-only build leaves beauty out entirely, along with its cost, and adding it later is a configuration decision rather than a rebuild.
Does adding more content slow the pipeline down?
Adding props does not add face analyses, because one analysis per frame feeds tracking, beauty, reshaping, makeup, and prop anchoring. Material cost is separate and depends on what each asset asks the GPU to draw.
Do we get a camera UI with it?
No. The SDK does not ship a finished camera UI. Your team builds recording controls, permission handling, asset browsing, and upload or publishing flows; everything beneath those is the SDK’s job.
Which platforms and frameworks are covered?
iOS, Android, Web, Desktop (Windows), and smart hardware natively, plus React Native, and Flutter supported through the short-video solution for cross-framework teams.
Ready to see it in your own stack? Request a free trial or talk to our integration team — we’ll help you scope the modules you actually need.