NVIDIA compresses material textures into 30% less memory
NVIDIA Research has developed a novel neural compression technique, 'Random-Access Neural Compression of Material Textures,' designed to address increasing storage and memory demands in photorealistic rendering by compressing multiple material textures and their mipmap chains together using a small neural network. This method allows for on-demand, real-time decompression with random access, providing 4x higher resolution (16X texels) using 30% less memory compared to traditional GPU texture formats, with image quality superior to AVIF and JPEG XL. The research was accepted to Siggraph 2023.
Key Takeaways
- The method compresses multiple material textures and their mipmap chains together with a small neural network.
- NVIDIA says the approach delivers 16X more texels, or 4x higher resolution, than BC high texture formats.
- The paper reports 30% less memory use than traditional GPU texture formats.
- The system supports on-demand, real-time decompression with random access similar to GPU block texture compression.
- NVIDIA says its custom training implementation runs more than 10x faster than general frameworks like PyTorch.
Why It Matters
For photorealistic rendering pipelines, the immediate gain is less texture memory pressure without giving up random access or real-time decompression. That matters because the paper combines disk and memory compression in a way that still fits GPU-style workflows, while targeting material textures and mipmap chains rather than isolated images. The competitive benchmark is explicit: NVIDIA says quality is better than AVIF and JPEG XL, and the method was accepted to SIGGRAPH 2023. Next to watch is the paper’s image viewer and whether the reported 16X texel density holds across texture comparisons.
Additional Context
Following its initial research debut, Nvidia expanded on Neural Texture Compression (NTC) during GTC 2026, showcasing its practical impact on video memory requirements. Per VideoCardz (April 2026), Nvidia demonstrated a "Tuscan Villa" scene where VRAM usage plummeted from 6.5GB using traditional BCn compression to just 970MB using NTC. This baseline represents an 85% reduction in memory overhead while maintaining or improving visual fidelity. Tom’s Hardware (April 2026) noted that the technology is designed to run in multiple modes, including "Inference on Load," which transcodes NTC textures to standard formats during map loading to reduce disk footprints without hurting real-time performance. This shift toward neural rendering is part of a broader strategy to bridge the gap between high-end visual demands and local hardware constraints. At CES 2026, Nvidia introduced DLSS 4.5, which uses second-generation transformer models to improve upscaling and frame generation. According to SMBTech (January 2026), these advancements aim to generate up to five frames for every one rendered, further offloading traditional compute tasks to AI cores. By combining NTC with frame generation, Nvidia is essentially rebuilding the rendering pipeline around deterministic AI inference rather than raw rasterization. The competitive landscape is also reacting to these memory-saving breakthroughs. While Nvidia currently holds a dominant position in integrating neural networks directly into the shader pipeline, competitors like AMD and Intel have released their own AI-focused fidelity updates. Per Tom's Hardware, Nvidia has released an NTC SDK on GitHub that uniquely supports AMD and Intel GPUs in specific modes, signaling an attempt to establish NTC as a cross-platform standard for 3D asset delivery and storage.
Read full article at research.nvidia.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source