Paper Overview
- Field: Computer Vision / Scientific Visualization
- Authors: Kaiyuan Tang, Maizhe Yang, Chaoli Wang
- Published: 2026-07-20
- arXiv: 2607.18187
- Categories: cs.GR, cs.DB, cs.LG
- arXiv page: https://arxiv.org/abs/2607.18187
Abstract
Large-scale scientific simulations generate volumetric data at rates far outpacing advances in storage and network bandwidth, making effective lossy compression increasingly critical. However, conventional compressors often struggle to preserve fine structural details at high compression ratios (CR), while implicit neural representations (INR) require costly per-volume optimization and produce fixed-CR models.
To address this, the authors propose EVOLVE, an autoencoder (AE)-based volume compression framework targeting high-CR offline compression, with three key contributions:
1. Large-scale cross-domain database: A curated dataset of 6,376 volumes from 21 scientific simulations, carefully filtered via perceptual hashing to ensure diversity. This enables the optimized model to extract features that generalize across volumes within the covered scientific simulation domains.
2. Improved AE design space: The authors revisit the design space of AE-based compressors, incorporating multiple macro-design and micro-design choices into a vanilla AE to develop EVOLVE, significantly improving expressive power and compression capability.
3. Variable-rate encoding: A learnable gain mechanism allows a single trained model to encode volumes at multiple compression rates, avoiding the fixed-rate limitation of INR-based approaches.
The result is a framework achieving substantially higher compression ratios than conventional compressors while preserving structural detail.