Neural Graphics Texture Compression Supporting Random Access

💡 Research Summary

**
This paper addresses three fundamental challenges in texture compression for real‑time rendering: (1) the need for random‑access decoding, (2) support for arbitrary numbers of channels beyond the traditional four, and (3) efficient handling of multiple mip‑map levels. Existing neural image compression (NIC) methods are unsuitable because they assume full‑image reconstruction, rely on entropy coding that prevents random access, and are limited to three‑channel RGB inputs. Conventional block‑based texture codecs (ASTC, ETC, etc.) provide random access but only compress up to four channels and treat each mip‑level independently, missing cross‑channel and cross‑scale redundancies.

The authors propose a novel asymmetric auto‑encoder architecture specifically designed for texture sets. The encoder, called the Global Transformer (E), takes a texture set T of size c × h × w (c = total channels) and maps it to a bottleneck latent tensor Zₛ궤 of size c_z × h/8 × w/8. This transformation uses a pure convolutional network without attention blocks and applies a tanh activation to bound the latent values to