Nvidia can significantly reduce VRAM usage or achieve much better image quality on the same budget.
During the GTC presentation titled “Introduction to Neural Rendering,” Nvidia showcased its Neural Texture Compression (NTC) technology once again. First introduced nearly three years ago, NTC has been available via the SDK since early 2026. However, thus far, no game developer has adopted it. Perhaps that is why Nvidia seized the opportunity to highlight the technology’s potential benefits once again. Alexey Bekin, a senior DevTech engineer, described NTC as a machine learning approach for more efficient texture storage. Rather than storing every texel directly, NTC compresses the texture into compact, learned latent features that capture its essential visual information. At runtime, a small neural network running on the GPU reconstructs the texel values from these features, computing them on demand instead of loading large textures from memory. NTC is deterministic rather than generative, so it reconstructs the same texture every time.
The system consists of two components. The latent texture is a significantly smaller-scale representation of the original element. In this representation, each texel stores a feature vector describing the material’s properties rather than the final color. To ensure fine details can be recovered, positional encoding is applied to the UV coordinates before reaching the decoder. This feeds high-frequency spatial information into the system, helping the network accurately reconstruct sharp details and repetitive patterns that would otherwise be lost in the compressed representation. Training operates as a standard neural optimization cycle. The network takes position-coded UV coordinates and latent codes as input. It then generates a reconstruction and compares it to the original texture. Next, it calculates the reconstruction loss and iteratively updates the MLP weights and latent codes until the output converges to an exact reproduction of the source material.
Nvidia’s NTC offers three structural advantages over traditional formats such as BCN: higher compression ratios, support for a large number of channels, and practical storage and bandwidth savings. Higher compression ratios allow significantly more texture data to fit into the same amount of video memory. Support for a large number of channels allows complex assets with many packed material channels (normals, roughness, albedo, AO, and more in a single material set) to be compressed cleanly without splitting or simplifying the data. Practical storage and bandwidth savings mean less disk space, smaller game sizes, smaller updates, and faster downloads.
Bekin presented a demo scene depicting a Tuscan villa in which NTC reduced VRAM usage by 85% (970 MB) compared to textures compressed with traditional BCN (6.5 GB). This will be useful for VRAM-intensive games, but NTC can also improve texture quality within the same VRAM budget, avoiding the compression artifacts typical of BCN.
As we previously reported, according to Kepler_L2, Sony could use NTC (developed by Nvidia and supported by AMD and Intel hardware) to reduce the installation size of PlayStation 6 games while keeping costs low with a 1 TB SSD. The NTC SDK is currently in beta and available on GitHub.
Leave a Reply