Efficient Learned Volume Compression with Variable-Rate Encoding on a Cross-Domain Database
Large-scale scientific simulations generate volumetric data at rates that far outpace advances in storage and network bandwidth, making effective lossy compression increasingly critical. Conventional compressors often struggle to preserve fine structural details at high compression ratios (CRs), and implicit neural representations (INRs) require costly per-volume optimization and produce models with fixed CRs. We present EVOLVE, an autoencoder-based volume compression framework with three key contributions: a large-scale cross-domain database of 6,376 volumes from 21 simulations curated via perceptual hashing; a redesigned autoencoder that incorporates carefully chosen macro- and micro-designs to substantially improve expressive power and compression capability; and a learnable gain mechanism with a three-stage training strategy that enables variable-rate encoding, allowing a single model to support continuous CR adjustment at inference time. EVOLVE achieves substantially higher CRs than conventional compressors at comparable reconstruction quality, while offering compression speeds orders of magnitude faster than INR-based methods.
6,376 volumes drawn from 21 scientific simulations, curated by perceptual hashing for diversity, so the optimized model learns features that generalize across unseen volumes within the covered domains.
We re-examine the design space of AE-based compressors and fold a set of macro- and micro-designs into a vanilla AE, sharply improving expressive power and rate–distortion performance.
A learnable gain mechanism trained with a three-stage strategy lets one model cover a wide, continuous range of compression ratios — no retraining, no per-rate checkpoints.
Compare any two methods on unseen datasets. Pick a render mode and dataset, choose what goes on each side of the divider, drag to wipe, and toggle the per-voxel error map.
EVOLVE against per-volume INR codecs — SIREN, NeurComp, ECNR, Instant-NGP, fV-SRN, AMGSRN++ and IDLat. Unlike these, EVOLVE needs no per-volume optimization: a single pretrained model encodes any unseen volume in seconds, at substantially higher compression ratios.
Scanned CT volumes (Open SciVis) are a harder, out-of-domain test: they are near-lossless targets (≈ 50 dB) and lie entirely outside EVOLVE's training distribution, which contains only simulation data. Even so, on volumes whose statistics fall close enough to that distribution — chameleon and stag beetle — EVOLVE clears the near-lossless bar and compresses far harder than any baseline (stag beetle: 50.5 dB at 3,830×, 3.6× the next best). We also include foot as a failure case, where quality saturates at 33.7 dB — out-of-domain success tracks how far the data drifts from the training distribution, not a blanket guarantee.
A single EVOLVE model traces full rate–distortion curves in both fidelity (PSNR) and perceptual quality (LPIPS), staying on top across the entire compression-ratio range while conventional and INR-based compressors fall away.
