How Tencent Cut Over $1B in Bandwidth Costs with Smart Image & Video Compression

Bandwidth Optimization Case Study – Image Compression

We present Tencent SNG's real‑world cost‑optimization experience, focusing on how technical measures saved more than 1 billion CNY in cash flow while keeping product quality healthy.

(1) QQ Album Image Compression

QQ Album stores tens of billions of images, with daily uploads of millions and downloads of hundreds of billions, leading to peak external bandwidth of over 100 GB per day.

1.1 WebP Image Compression

Since 2015, uploaded JPEG/PNG/GIF images are compressed to WebP on the backend, achieving roughly 30 % size reduction at equal quality.

1.2 Adaptive Resolution

Images are served at resolutions appropriate for the client device, saving about 20 % of traffic without affecting visual clarity.

1.3 GIF Animation Compression

In 2016, GIF‑based status updates caused a 70 % bandwidth surge; only 5 % of downloads accounted for 20 % of bandwidth. GIFs, being essentially lossless bitmap sequences, are large (2‑4 MB) and costly.

1.4 New TPG Image Compression

Using the AVS2‑based TPG format, which applies video‑style inter‑frame prediction, we achieve up to 90 % size reduction for GIFs and an additional 21 % over WebP, saving roughly 43 % compared to JPEG. TPG was fully deployed in 2017.

Unified backend configuration for all clients

Tiered compression and OC point caching

Compatibility handling for GIF URLs

Gray‑scale rollout with asynchronous compression for large GIFs

1.5 Cheap Transcoding Resources

We built an offline platform that schedules idle CPU cycles from over‑sell pools, department pools, and low‑load online devices via Docker, achieving hour‑level scaling.

FPGA transcoding cards provide ten‑fold performance over standard servers, reducing encoding latency by 80 % and cutting thousands of servers from the fleet.

(2) Small Video On‑Demand Optimization

We limited download speed for QQ Space short videos to only 20 seconds ahead of playback, reducing redundant traffic while keeping buffering low.

For long videos, we introduced progressive download (buffer first 20 seconds, then stream at bitrate), cutting average wait time from 12.6 s to 1.77 s and reducing redundant downloads from 65 % to 35 %.

Disabled auto‑play for low‑traffic rooms

Limited forwarding of large files (>1 GB) to curb piracy

Improved safety detection accuracy using AI models

Reduced bitrate for low‑complexity videos, saving bandwidth

(3) Real‑Time Audio‑Video Mixing

Moving mixing from client to server halves downstream bandwidth. We also limit mixing for rooms with fewer than three participants to save compute resources.

(4) Live Streaming Optimization

For high‑traffic live rooms (e.g., NOW live quiz with up to 900 k concurrent viewers), we applied H.265 encoding and GPU‑accelerated transcoding, achieving up to 40 % bandwidth savings and reducing hardware cost by millions of yuan annually.

(5) Fine‑Grained Bitrate Optimization

We employ dynamic bitrate (VBR), predictive network quality adaptation, content‑aware bitrate selection, and scene‑aware encoding using deep learning to further trim bandwidth usage.

(6) AI‑Powered Image Super‑Resolution

QQ Album uses a deep‑neural‑network super‑resolution model that halves pixel dimensions, compresses size to 25 % of the original, yet restores visual quality indistinguishable to the human eye, saving 75 % of traffic and storage.

(7) AI & Security Enforcement

Convolutional neural networks improve detection of obfuscated pornographic content, dropping yellow‑tagged clicks from 25 % to 5 % of top‑1000 items. OCR extracts text from illicit images to combat prostitution ads. Traditional rule‑based methods are complemented by these AI techniques.

Methodology Summary

Bandwidth optimization follows the “one small, two less, three unchanged” principle: make files smaller, reduce download counts, eliminate redundant downloads, while preserving quality. The five‑step process includes resource consumption modeling, focusing on top‑cost modules, deep architectural analysis, joint product‑technology actions, and continuous dynamic operation.