← All resources

MiniDXNN v0.4.0: Interactive neural texture compression on DirectX 12

Auto-discovered

Library / Codebase Advanced

Takahiro Harada

Visit resource ↗

Description

MiniDXNN v0.4.0: Interactive neural texture compression on DirectX 12 is AMD GPUOpen coverage of Takahiro Harada's open-source MiniDXNN library for running and training small MLPs entirely on the GPU via DirectX 12. Version 0.4.0 highlights Shader Model 6.10 linear-algebra intrinsics and a hands-on ImGui sample that trains a neural texture live each frame.

The release notes explain input encodings used for neural texture compression, how inference and training passes are scheduled in D3D12 command lists, and what changed since earlier MiniDXNN drops on GPUOpen. Readers see how a graphics programmer can experiment with neural materials without leaving the DXGI swap chain.

Expect comfort with HLSL or DXIL, D3D12 resource binding, and basic neural field concepts. The library targets advanced graphics engineers exploring neural graphics prototypes on Windows rather than ML researchers using PyTorch alone.

Source and binaries are distributed through GPUOpen and linked GitHub repositories under open licenses. The ImGui demo provides an interactive entry point before integrating MiniDXNN into larger engines.

Harada's GPU-based ML work and earlier MiniDXNN talks are well known in real-time rendering circles; GPUOpen's neural texture compression articles are a common starting point for teams evaluating on-GPU training in shipping renderers.