At the SIGGRAPH 2026 conference, Daniel Craig, chief software engineer of Sony Interactive Entertainment, introduced in detail the updated enhanced PSSR upconverter model of the PS5 Pro game console.
Image source: Sony
One of the most important architectural changes in the enhanced PSSR upconverter relates to how neural network computing resources are used. According to Sony, the original Color Prediction Network (CPN) used about 30% of the parameters at 540p resolution. In the new Core Prediction Network (KPN), up to 90% of parameters already use 540p and lower resolutions, and the neural network wastes almost no resources in 4K.
Compared to the original PSSR, the enhanced PSSR pipeline
The new KPN neural network also does not directly predict the final color. Instead, it generates blend weights and uses normal filtering to do the actual blending. This avoids storing and mixing HDR colors in 8-bit format and reduces the load on the neural network when performing these operations, thus improving overall efficiency.
Sony also notes improved timing stability, sharper details, and reduced memory consumption and GPU runtime. The company says image quality will improve at any resolution, but the effect is especially noticeable at lower resolutions. The company also reports that neural network training time has been reduced from approximately 4 GPU months to less than 1 GPU week.
At SIGGRAPH 2026, Sony also confirmed that Enhanced PSSR replaces the version launched with PS5 Pro, rather than operating as a separate mode. The new version was developed by Sony and AMD as part of the Project Amethyst research program, and its implementation is suitable for PlayStation hardware. Existing PS5 Pro games can use enhanced PSSR through system overlay without updating the game itself. Developers who integrate the new library directly will have full access to the improved algorithms.
If you find an error, select it with your mouse and press CTRL+ENTER.
