Xiaohongshu Deploys On‑Device Super‑Resolution with Huawei HMS Core for High‑Quality Short Videos

Xiaohongshu, in collaboration with Huawei HMS Core, has applied on‑device super‑resolution (SR) technology to its short‑video platform, achieving high‑quality playback with the same network conditions while reducing bandwidth usage and stutter.

The joint effort solved multiple technical challenges, such as multi‑instance SR, flexible selection of different View types (SurfaceView, TextureView, etc.), and leveraged the Huawei HiAI Foundation to fully utilize GPU and NPU resources. The solution delivers real‑time SR without noticeable performance loss, converting 540p video to 1080p instantly and enhancing 720p video to a 720p+ experience without increasing bitrate.

Term Definition

Super‑resolution (SR) is the process of reconstructing a high‑resolution (HR) image from a low‑resolution (LR) input.

2.1. Limitations of On‑Device Quality Enhancement

Few devices have NPU; most rely on CPU/GPU, limiting compute power, especially on low‑end phones.

Power consumption remains a major issue; higher quality often leads to overheating.

Android device fragmentation makes hardware adaptation difficult.

Traditional algorithm models are monolithic, tightly coupled with business code, increasing APK size.

2.2. Industry Approaches

Client‑cloud co‑processing: cloud performs pre‑processing (sharpening, denoising, bitrate boost) while the client restores the video.

Flexible methods such as machine learning to improve computational efficiency and reduce CPU load.

On‑device post‑rendering techniques.

By fully exploiting Huawei HMS Core’s powerful HiAI Foundation, Xiaohongshu utilizes NPU/GPU with different training models, achieving high performance with low power consumption. The integration replaces the traditional Java‑level MediaCodec with NDK MediaCodec, merging decoding and SR entirely in the native layer to eliminate redundant JNI calls.

Additional technical measures include:

Pre‑video preprocessing (denoising, filtering) to prepare optimal SR conditions.

Adaptive SR selection: only videos with rich colors, moderate frame rates (<60 fps), and resolutions ≤720p are upscaled; HDR content is excluded.

Dynamic SO loading and hot‑swap mechanism to avoid APK bloat and enable real‑time updates based on device model and app version.

Resource‑management module that supports multiple View types, allowing seamless rotation, scaling, and post‑processing after SR.

Since version 7.48, the “video SR” feature has been deployed on over 90 Huawei devices (Mate X2, P50, Mate 40 series, etc.) and more than 10 tablets, serving tens of millions of users daily.

The collaboration provides a reference for large‑scale on‑device SR in short‑video scenarios and demonstrates the potential of soft‑hardware integration for future on‑device quality‑enhancement technologies.