ECoNGS compresses volumetric data 6x while boosting rendering quality

Volume visualization of large scientific datasets has long struggled with the trade-off between rendering speed and storage efficiency. Existing primitive-based methods like Gaussian splatting offer interactive performance but produce redundant, non-compact representations due to isolated per-scene optimization. ECoNGS addresses this by combining implicit neural networks with explicit primitives: lightweight networks dynamically predict editable Gaussian splats from sparse anchor points, merging the parameter efficiency of implicit representations with the high-performance rendering of explicit primitives. A joint learning strategy clusters geometrically similar scenes and shares parameters across them, reducing training time and model size without sacrificing fidelity. Further compression is achieved via a neural entropy model that estimates probability distributions of anchor attributes, enabling compact storage through entropy coding. Tailored Gaussian initialization strategies improve reconstruction accuracy and accelerate convergence.

ECoNGS was evaluated on multiple univariate and multivariate volume visualization scenes, showing significant gains over prior methods. Compared to iVR-GS, it delivers up to 2.2 dB higher PSNR while cutting model size by 6.1× and training time by 5.9×. The framework achieves high-quality interactive rendering of large datasets with drastically reduced storage and computational overhead. The code is publicly available, and the paper is set to be published in IEEE VIS 2026 (IEEE Transactions on Visualization and Computer Graphics).