Rendering 2025 - Symposium Track

Permanent URI for this collection

Now showing 1 - 1 of 1

Differentiable Block Compression for Neural Texture
(The Eurographics Association, 2025) Zhuang, Tao; Liu, Wentao; Liu, Ligang; Wang, Beibei; Wilkie, Alexander
In real-time rendering, neural network models using neural textures (texture-form neural features) are increasingly applied. For high-memory scenarios like film-grade games, reducing neural texture memory overhead is critical. While neural textures can use hardware-accelerated block compression for memory savings and leverage hardware texture filtering for performance, mainstream block compression encoders only aim to minimize compression errors. This design may significantly increase neural network model loss.We propose a novel differentiable block compression (DBC) framework that integrates encoding and decoding into neural network optimization training. Compared with direct compression by mainstream encoders, end-to-end trained neural textures reduce model loss. The framework first enables differentiable encoding computation, then uses a compressionerror- based stochastic sampling strategy for encoding configuration selection. A Mixture of Partitions (MoP) module is introduced to reduce computational costs from multiple partition configurations. As DBC employs native block compression formats, inference maintains real-time performance.