AMD has introduced a new research technique called PEPS that could make neural texture compression more memory efficient without causing a major loss in image quality.
PEPS stands for Positional Encoding Projected Sampling. The method changes how positional information is fed into the neural networks used to represent textures and other visual data.
AMD presented the research at the I3D Symposium. Its tests suggest that PEPS can reduce the number of model parameters required for neural texture compression by around 25 percent while maintaining comparable reconstruction quality.
The approach may eventually help graphics developers reduce memory use in games and real time rendering. However, it also increases computation requirements, and the technology is still at the research stage.
Neural texture compression stores images inside a small neural model
Traditional texture compression saves an image using a fixed pixel based format. Neural texture compression instead trains a compact neural network to reproduce the texture when given a pair of coordinates.
These systems often use Implicit Neural Representations, commonly known as INRs. An INR learns a function that connects a coordinate, such as an X and Y position, to the colour or other information stored at that location.
Before the coordinates are passed to the neural network, they are usually transformed through positional encoding. This converts a small coordinate into a larger collection of sine and cosine values, giving the network more information about fine visual detail.
The trained model and its supporting feature data can then act as a compressed representation of the original texture.
| Method or term | Role in the compression process |
|---|---|
| Implicit Neural Representation | Learns how coordinates map to texture values |
| Positional encoding | Expands coordinates into a richer representation |
| Feature grid | Stores information sampled by the neural model |
| Multi layer perceptron | Reconstructs colour or other texture data |
| PEPS | Samples the grid using projected positional encoding points |
PEPS changes how encoded coordinates sample the feature grid
PEPS builds on ordinary sine and cosine positional encoding. Instead of treating the encoded values only as a vector sent into the network, the technique interprets each projection as a point along a Lissajous curve.
Those projected points are then used to sample a feature grid. This allows the network to gather more useful information from the encoded coordinates without relying on a larger set of model parameters.
AMD tested a version called Grid PEPS for neural texture compression. It was able to achieve similar quality to an existing bidirectional grid baseline while using about 25 percent fewer parameters.
A further variation, Grid PinkPEPS, was designed to improve the performance balance. It retained much of the memory benefit while reducing some of the additional processing cost introduced by the new sampling method.
Smaller models require more processing time
The main limitation is that PEPS needs extra calculations and memory accesses. This makes reconstruction slower than the simpler baseline.
In one AMD test using a Radeon RX 9070 XT, producing a 1024 by 1024 texture with three colour channels took 4.32 milliseconds with the baseline method.
Grid PEPS increased the time to 5.47 milliseconds. The optimized Grid PinkPEPS implementation reduced it to 4.86 milliseconds.
| Compression method | Time to generate a 1024 by 1024 texture |
|---|---|
| Bidirectional grid baseline | 4.32 milliseconds |
| Grid PEPS | 5.47 milliseconds |
| Grid PinkPEPS | 4.86 milliseconds |
This shows the central tradeoff. PEPS can reduce storage and parameter requirements, but the GPU must perform more work when reconstructing the data.
Whether that tradeoff is worthwhile will depend on the application. A game limited by graphics memory may benefit from smaller models, while one already struggling with shader performance may prefer the faster baseline.
The technique may also reduce memory use for 3D distance fields
AMD also tested PEPS with signed distance functions, known as SDFs. These mathematical representations describe how far a point is from the surface of a 3D shape.
SDFs are useful for rendering, collision systems, geometry processing, and visual effects. High quality SDFs often require dense grids, which can consume a large amount of memory.
During testing with a Pitted Stonefish model, Grid PEPS produced a reconstruction with roughly similar Intersection over Union accuracy to non PEPS methods that used eight times more encoder parameters.
Intersection over Union measures how closely the reconstructed shape overlaps with the original. A higher result generally indicates a more accurate representation.
Reducing the parameter requirement for SDFs could make detailed geometry representations more practical on hardware with limited graphics memory.
PEPS is unlikely to appear in consumer games soon
The research does not mean that a new Radeon feature or FSR update is ready for release. Neural texture compression remains an emerging area, and full implementations are not yet common in commercial games.
Developers would need tools, runtime support, reliable training pipelines, and hardware friendly decoding methods before the technology could become widely adopted.
AMD has not announced a consumer brand for its neural texture compression research. There is also no confirmed timetable for integrating PEPS into FSR or Radeon drivers.
The work is still relevant because graphics memory capacity continues to affect game development. Many mainstream GPUs are expected to remain available with 8GB of VRAM for several more years, while modern textures and geometry continue to grow.
PEPS shows one possible way to store visual information more efficiently. Its current performance cost means further optimization will be needed, but the research could contribute to future rendering systems that trade additional computation for lower memory use.
Discussion (0)
Be the first to comment.