How veImageX Cuts Image Bandwidth by 30% with AI‑Powered HEIF Compression

Introduction

Recently the fifth CLIC (Compression Learning Image Challenge) announced its results, and the Volcano Engine Video Cloud multimedia lab won first place in the video compression track. Image and video compression are critical for internet delivery; reducing file size while preserving perceived quality is a continuous research focus.

veImageX Overview

veImageX is Volcano Engine’s end‑to‑end image solution covering upload, storage, processing, distribution, display and quality monitoring.

An image’s lifecycle consumes bandwidth, storage and compute roughly in a 7:2:1 ratio, with bandwidth often being the dominant cost.

Architecture

The system consists of three core components: a CDN distribution layer, a storage layer, and a media‑processing layer. To lower bandwidth, veImageX adopts a self‑developed HEIF codec that compresses images without degrading user‑perceived quality.

Dual‑End Image Compression

On iOS (JPEG) and Android (WebP) devices, converting images to the custom HEIF format saves more than 30% of the original volume.

Quality Evaluation

While size reduction is the primary goal, preserving visual quality is essential. Objective metrics and a proprietary no‑reference VQScore algorithm were used to assess sharpness and aesthetics, showing negligible quality loss after compression.

Decoding Performance

To keep user‑side load time low, the custom decoder was tuned so that its added latency is smaller than the time saved by reduced file size. Benchmarks show the custom decoder’s latency comparable to WebP and better than many open‑source alternatives.

Experiment Design & Data

A controlled A/B test compared bandwidth usage between two domains (p‑xx‑a vs p‑xx‑b). The experiment showed a 32.4% reduction in bandwidth (2.53 PB vs 1.71 PB), confirming at least a 30% saving versus WebP.

General “Intelligent Slimming”

For H5 scenarios, a DNS‑only integration called “Intelligent Slimming” applies deep‑learning‑based compression to traditional formats (WebP, JPEG, PNG) without requiring a client SDK, delivering 15‑20% additional bandwidth savings.

Conclusion

Beyond the two main scenarios, veImageX’s AI‑driven compression reduces image size by over 30% compared with WebP, adaptive quality compression adds another 5‑10%, and “Intelligent Slimming” contributes 15‑20%, achieving up to 80% overall cost reduction while improving load speed and lowering user data consumption.