At GDC 2026, Intel graphics engineer Marissa du Bois stepped onto the stage to reveal a major update for video game developers. She presented Intel’s new Texture Set Neural Compression technology. This system acts much like NVIDIA’s version and offers a smarter way to shrink massive game files. Intel originally showed this idea as a research project at GDC 2025. Now, the company has turned that early research into a real software tool that developers can actually use to make games take up less hard drive space.
Traditional graphics tools use fixed math rules to crush texture sizes. While these older formats run fast and work everywhere, they leave a lot of wasted space behind. Intel takes a totally different route. The new tool trains a tiny neural network to learn the exact details of a specific set of textures. The program stores a small version of the data and then uses the neural network to rebuild the full image while you play the game. The tool notices when a brick wall material shares the same pattern across its color, bump, and shadow maps. It uses that shared layout to save massive amounts of digital space.
Intel gives game makers 2 different options to balance file size and visual quality. Variant A serves as the standard choice. It takes a massive 256 MB set of 4K textures and crushes it down to just 26.8 MB. That means developers get an impressive 9x reduction in file size. For comparison, older standard compression tools only reach a 4.8x reduction. Intel tested this method and found roughly a 5% drop in visual quality. Most players will only notice subtle changes in how light falls on the surface, making it a great option for high-quality game environments.
Variant B offers a much more aggressive approach. This option pushes the file sizes down to 1/2, 1/4, and 1/8 of their original size, delivering a massive 17x compression rate. However, the graphics take a larger hit. The visual error rate jumps to around 7%. Intel admits players will easily spot blocky artifacts and messy shadows if they look closely. Developers will likely use this super-compressed option for background mountains or distant buildings where players rarely focus their attention.
Since last year, Intel has completely rebuilt the software from the ground up. The team threw out the old PyTorch code and rewrote everything using Slang compute shaders. This means developers can use the tool easily inside the popular Unreal Engine or drop it right into their own custom game engines. Intel also designed the code to run well on standard computer processors and graphics cards. Players do not need an Intel-brand graphics card to play games that use this new tool.
Developers can choose from 4 ways to deploy the tool, depending on their needs. First, they can let the game decompress textures while the player installs it, saving internet bandwidth during the download. Second, the game can leave files squished on the hard drive and unpack them right as a level loads. Third, the game can stream and unpack files on demand as the player walks around. Finally, the most aggressive method keeps everything compressed in video memory permanently and decodes the colors pixel by pixel just as they hit the screen.
Intel tested the toughest per-pixel decoding method on a standard Panther Lake laptop running simple B390 integrated graphics. The standard math processing path took 0.661 nanoseconds per pixel to load. However, when Intel turned on its special XMX hardware cores, the speed dropped to an incredibly fast 0.194 nanoseconds per pixel. This 3.4x speedup proves the tech works well even on basic laptop chips. Dedicated gaming graphics cards will run it even faster. Intel plans to release an early test version of the tool later this year, giving developers a chance to try it before the final public release.
