EVOLVE: Efficient Learned Volume Compression with Variable-Rate Encoding

Forum topic · 小凯 · 2026-07-22

Summary

EVOLVE is an autoencoder-based volume compression framework presented by Kaiyuan Tang, Maizhe Yang, and Chaoli Wang (arXiv:2607.18187). It targets large-scale scientific simulation data, whose growth outpaces advances in storage and network bandwidth, making effective lossy compression critical. The framework is designed for high compression ratios (CR) in offline settings, addressing two limitations of prior approaches: conventional compressors lose fine structural details at high CR, and implicit neural representations require costly per-volume optimization and produce fixed-CR models. EVOLVE makes three key contributions: (1) a large cross-domain database of 6,376 volumes from 21 scientific simulations, carefully curated with perceptual hashing to ensure diversity, enabling learned features that generalize across volumes; (2) a systematic re-examination of AE compressor design choices, incorporating multiple macro- and micro-design elements into a vanilla autoencoder to substantially improve expressiveness and compression capability; and (3) a learnable gain mechanism enabling variable-rate encoding, so a single trained model supports multiple compression ratios. The authors report substantially higher compression ratios than conventional compressors. The work spans computer graphics (cs.GR), databases (cs.DB), and machine learning (cs.LG).

Overview

Field: Computer Vision / Graphics (CV)
Authors: Kaiyuan Tang, Maizhe Yang, Chaoli Wang
Release date: 2026-07-20
arXiv: 2607.18187
Categories: cs.GR, cs.DB, cs.LG

Background

Large-scale scientific simulations generate volumetric data at rates far outpacing advances in storage and network bandwidth, making effective lossy compression increasingly critical. However:

Traditional compressors often struggle to preserve fine structural details at high compression ratios (CR).
Implicit neural representations (INR) require expensive per-volume optimization and produce fixed-CR models.

EVOLVE

The authors propose EVOLVE, an autoencoder (AE)-based volume compression framework targeting high CR for offline compression, with three key contributions:

1. Large-scale cross-domain database: 6,376 volumes from 21 scientific simulations, carefully curated via perceptual hashing to ensure diversity. This allows the optimized model to extract features that generalize across volumes within the covered scientific simulation domains. 2. Re-examined AE design space: multiple macro-design and micro-design choices are incorporated into a vanilla AE to develop EVOLVE, significantly improving expressiveness and compression capability. 3. Learnable gain mechanism: enables variable-rate encoding, avoiding the fixed-CR limitation of INR-based methods.

Abstract (refined)

> We present EVOLVE, an AE-based volume compression framework with a cross-domain database of 6,376 volumes, improved AE designs, and a learnable gain mechanism for variable-rate encoding, achieving substantially higher compression ratios than conventional compressors.

*Auto-collected on 2026-07-22.*