AMD PEPS research points to smaller neural texture models, but games will not benefit soon

AMD has presented new research that could make neural texture compression more efficient, although the work is still far from becoming something you can use in today’s games. The research, called PEPS, short for Positional Encoding Projected Sampling, was shown at the I3D Symposium and focuses on improving how neural networks represent texture data.

Neural texture compression is a possible future solution for reducing the memory footprint of game assets. Instead of storing texture data only in a traditional format, the method trains a small neural model to learn how texture coordinates map to color or signal values. In simple terms, the model learns the texture and then recreates it when needed.

That approach can save memory, which is becoming more important as modern games use larger textures while many graphics cards still ship with limited VRAM. The tradeoff is that neural methods add compute cost, since the GPU has to do extra work to reconstruct the texture.

PEPS changes how positional encoding feeds information into the model

The main change in AMD’s PEPS research is how positional encoding is handled. Positional encoding usually takes low dimensional coordinates and projects them into a higher dimensional sine and cosine representation. PEPS builds on this idea by treating each sine and cosine projection as a point on a Lissajous curve, then sampling an encoder or grid at those projected points.

That gives the neural model more useful information to work with, which can reduce the number of model parameters needed at similar quality. According to the reference material, AMD’s approach can cut model parameters by about 25 percent while keeping comparable image quality.

The cost is performance. AMD’s testing on a Radeon RX 9070 XT showed that generating a 1024 by 1024 three channel texture took 4.32 ms with the BI grid baseline. Grid PEPS increased that to 5.47 ms, while a more optimized Grid PinkPEPS version brought the time down to 4.86 ms.

PEPS may also have value outside normal texture compression. AMD tested the approach with signed distance functions, which are used in 3D rendering to represent shapes and surfaces. These can require large high resolution grids, so reducing memory use through neural compression could be useful. In the Pitted Stonefish SDF test, Grid PEPS reportedly came close to matching non PEPS methods that used eight times more encoder parameters.

The research is promising, but it should not be mistaken for a feature that is ready for Radeon owners today. Neural texture compression is still early in consumer gaming. NVIDIA has public demos and tooling in this area, but there are still no major games using a full neural texture compression pipeline. AMD also has not turned this research into a branded consumer feature inside its FSR suite.

For now, PEPS is best seen as a research step. It shows that AMD is exploring ways to reduce texture memory demands while keeping quality stable. That matters as game assets grow and VRAM limits remain a concern, especially for lower memory graphics cards. The real test will come later, when methods like this can run fast enough, fit into game engines cleanly, and deliver clear benefits without adding too much latency or complexity.