How Kuaishou Optimizes Mobile AI Effects with Dynamic Device Grading

Abstract

Kuaishou’s face‑filter stickers are popular because of the AI effect algorithms behind them. Since Kuaishou’s user base spans many device types with large performance differences, a device‑based model grading scheme is essential to balance effect quality and performance on both high‑end and low‑end phones.

This article describes how the Y‑tech engineering team designed a dynamic delivery and execution strategy for mobile AI algorithms, taking into account diverse hardware characteristics and multiple information dimensions to achieve a stable yet flexible device‑grading solution.

Design of the Grading Scheme

The main difficulty in mobile grading lies in the variety of device models and hardware, as well as evolving rule standards and compatibility with underlying runtimes. The scheme therefore emphasizes universal and extensible hardware evaluation criteria, introduces dynamic dispatch and real‑time adjustment mechanisms, and closes the loop through data collection and analysis.

Grading Information Dimensions

AI effect models mainly consume CPU, GPU, NPU and memory resources, so the grading criteria are based on these components. The core dimensions for Android and iOS are illustrated below.

Android

Android’s open ecosystem results in thousands of device models. Grading principles are built from three dimensions:

Chip Architecture : Most chips follow the ARM architecture. Newer generations (e.g., Dimensity 1200 – A78, Snapdragon 8 Gen 1 – X2) generally offer better performance, but a single architecture may cover dozens of chip variants, so grading must be coarse.

AI Acceleration Features (NPU, SNPE, APU, OpenCL, Vulkan) : Manufacturers provide AI‑specific hardware (e.g., Huawei’s HiAI NPU, Qualcomm’s SNPE). Models optimized for these features receive higher grades; otherwise the generic chip‑architecture grade applies.

Third‑Party Scoring Platforms : Benchmarks such as AnTuTu, CPU‑ladder, and Kuaishou’s internal device‑portrait platform provide objective scores that help set grade boundaries.

iOS

iOS devices have a more uniform hardware landscape, making grading simpler. Two main dimensions are considered:

Device Model : iPhone models are ordered by performance; thresholds are set to separate low, medium, and high tiers.

System Features – Metal & CoreML : GPU‑based Metal acceleration (available from iPhone 7 / iOS 8) and newer CoreML/NPU support (iPhone XR and above) are used as additional grading criteria.

Grading Application Logic

SDK Side

Model Grading Loading Mechanism : A zip package contains multiple model levels; the SDK loads from high to low, checks device info, and falls back to a low‑end model to guarantee results.

Dynamic Grading Configuration : Different algorithm types (e.g., background removal, key‑point tracking, GAN‑based effects) have distinct hardware demands. Remote configuration and feature‑switch mechanisms allow hot‑updates without code changes.

Real‑Time Strategy Adjustment : Third‑party scores define grade thresholds that can be adjusted dynamically as user devices evolve.

Service Side

Resource Dispatch : Device hardware info is collected via system APIs and sent to a backend service that determines which graded model to deliver, reducing download size and improving retention.

Configuration Dispatch : Graded configuration data is centrally managed and distributed to all business lines.

Score Information Dispatch : Internal device‑portrait scores (e.g., AnTuTu) are used as long‑term grading references.

Grading Data Monitoring

When reporting business metrics, the grade information is included, enabling backend monitoring of grade distribution, QoS/QoE statistics (frame rate, usage time), and early detection of abnormal devices for targeted fixes.

Practice Results

More than ten grading levels have been deployed across multiple Kuaishou apps (Kuaishou, Kwai Lite, Yidian, Kuaishou Pro, etc.).

Dynamic grading configuration and hot‑update capabilities have accelerated client releases.

Graded model delivery reduces app package size and improves retention on low‑end devices.

Numerous algorithms are in use online, with GAN‑based effects showing especially strong results (see “Fairy Prince” effect).

Summary and Outlook

The mobile AI model‑grading system significantly impacts user experience by balancing latency, visual quality, and resource consumption. As mobile hardware continues to evolve, Y‑tech will keep monitoring industry trends and refining the grading mechanism to deliver optimal experiences.

References

[1] https://baike.baidu.com/item/ARM%E6%9E%B6%E6%9E%84 [2] https://baike.baidu.com/item/OpenCL [3] https://www.antutu.com/ranking/rank1.htm [4] https://baike.baidu.com/item/Metal/10917053