开发者发布神经纹理压缩项目：采用进化策略无需反向传播

开发者richgel999于2026年9月4日在GitHub发布了一个神经纹理和PBR材质贴图压缩研究项目1。该项目使用进化策略（Evolution Strategies）优化神经纹理解码器和物理基础渲染（PBR）材质贴图，无需反向传播或自动微分1。

该项目实现了受限的足迹归因机制，允许同时优化低分辨率潜在纹理和小型多层感知机解码器，并支持多层纹理材质的联合压缩1。以128×128×8配置为例，在32线程CPU上的训练时间约为150秒1。在8位量化潜在纹理的情况下，量化成本仅为0.04 dB1。针对包含4层材质的PavingStones070示例，项目实现了2.64 bpp的总码率，每个纹理平均0.66 bpp1。

项目采用C++17编写，主要依赖为stb_image、stb_image_write和OpenMP1，支持在MSVC 2022/2026（Windows）以及gcc 13（WSL2）环境下编译1。

A developer has released an open-source research project on GitHub for compressing neural textures and physically-based rendering (PBR) material maps using evolution strategies, eliminating the need for backpropagation or automatic differentiation 1. The project, published on September 4, 2026, introduces a novel optimization approach that simultaneously refines low-resolution latent textures and compact multilayer perceptron decoders while supporting joint compression of multi-layer material textures 1.

The implementation demonstrates practical performance metrics across different configurations 1. Training a 128×128×8 setup requires approximately 150 seconds on a 32-thread CPU, with 8-bit quantization of latent textures incurring only 0.04 dB of cost 1. In a four-layer material example using PavingStones070 textures, the approach achieves a total bitrate of 2.64 bits per pixel, averaging 0.66 bits per pixel per texture 1. The project is written in C++17 and supports MSVC 2022/2026 on Windows and gcc 13 on WSL2, with dependencies limited to stb_image, stb_image_write, and OpenMP 1.