Intel Texture Set Neural Compression Shrinks Textures by Up to 18x with Minimal Quality Loss

Intel has recently released a new video showcasing its latest Texture Set Neural Compression (TSNC) technology, which delivers textures up to 18 times smaller while maintaining visual quality with little to no noticeable difference compared to the industry-standard compression. Using AI-based neural networks, Intel’s graphics team processes input data from industry-standard BCn textures. These textures are compressed through an AI model encoder, encoded in the latest space values, and then decoded by a network decoder to decompress the textures. The result is output data textures that are up to 18 times smaller, with some quality loss at maximum compression settings. As with any neural technology, TSNC is trained on millions of standardized textures to create an AI model that can replace traditionally compressed textures in the BCn format. This results in new, much smaller game textures that load faster, use less VRAM, and perform better thanks to modern GPU technology.

There are several ways to apply TSNC neural compression, depending on the desired outcome, whether it’s saving game installation size, reducing VRAM usage, or improving performance. Variant A, as Intel calls it, can achieve up to 9 times compression of the standard texture set, with little to no difference in visual quality—almost an unnoticeable drop. However, when the goal is maximum efficiency and requires up to 18 times texture compression, Intel offers Variant B of the TSNC neural network. This variant provides a significant performance boost, with the trade-off being a modest visual change. Using NVIDIA’s FLIP tool to measure quality drop in generated images, Intel notes that Variant A experiences a 5% visual quality drop, while Variant B sees up to a 7% quality drop, which is noticeably more.

You can judge for yourself by viewing the comparison images below.Finally, Intel benchmarked the performance of its Texture Set Neural Compression technology using the latest “Panther Lake” system with Arc B390 integrated graphics, which also includes XMX cores to accelerate these technologies seamlessly. The AI model is producing the first texture pixel in about 0.194 nanoseconds, which is fast enough for users not to notice any additional latency or experience oddities in their game rendering. We can expect this technology to be shipped later this year as an alpha version, with beta and full stable releases expected later on, though no concrete timelines have been provided.