Sony PSSR 2.0: How a Kernel-Predicting Network and Hybrid AI-GPU Architecture Overhaul PS5 Pro Upscaling

Sony's PlayStation Spectral Super Resolution (PSSR) 2.0, detailed in a SIGGRAPH 2026 talk by principal software engineer Daniel Craig and covered by Digital Foundry, represents a fundamental architectural shift from a fully neural approach to a hybrid system that blends AI inference with conventional GPU shader math. The result is noticeably more stable image reconstruction, lower performance overhead, and training time slashed from four GPU-months to under one GPU week . Here is a complete breakdown of every technical improvement.

From Color-Predicting Network to Kernel-Predicting Network

The original PSSR used a convolutional neural network (CNN) to reconstruct the entire final image, including HDR color blending — a task that neural networks handle poorly. The model had to "learn" blending logic that existing GPU hardware already performs more efficiently, which caused a "vanishing gradient" problem that produced instability in foliage and low-frequency lighting details like ray-traced global illumination (RTGI) .

PSSR 2.0 replaces this with a Kernel-Predicting Network (KPN). The neural network now has a narrower job: it strictly identifies edges and distinguishes detail from noise. Conventional GPU shader operations then handle the HDR color blending and final image reconstruction. This division of labor fixes the instability issues and lets the AI focus on what it does best .

Contrary to earlier rumors, the KPN architecture in PSSR 2.0 does not include a transformer element .

Elliptical Gaussian Filtering at Input Scale

The original PSSR defined its 3×3 reconstruction kernel at output resolution. At high scaling ratios, this produced a soft, slightly blurry appearance. PSSR 2.0 overhauls the scaling mechanism by defining the kernel at input scale using elliptical Gaussian filtering. This preserves sharper detail when upscaling from low internal resolutions, such as 720p or 900p .

Processing at Lower Resolution

Another key efficiency gain: the majority of neural network processing in PSSR 2.0 now runs at 540p or below. The previous model spent significant computational capacity on high-resolution operations, which limited its performance ceiling. By pushing most inference down to a lower-resolution internal buffer, PSSR 2.0 frees up GPU time and memory bandwidth .

Training Time: From Four GPU-Months to Under One GPU-Week

Training the original PSSR model required approximately four GPU-months. PSSR 2.0 training finished in under one GPU week — a reduction of roughly 94% — driven largely by changes in quantization . This makes it far more practical for Sony to iterate on the model and push updates via system software rather than per-game patches.

Performance: 100 Microseconds Faster Frame-Time

According to Mark Cerny in a Digital Foundry interview, the new PSSR 2.0 model is roughly 100 microseconds faster than the original . That may sound negligible, but it was the critical margin that made the system-wide override feasible. Sony could force-upgrade all PSSR games to version 2.0 without introducing new frame drops .

Sony also states that both frame-time and memory metrics improve versus the original PSSR, though specific figures were not disclosed in the SIGGRAPH talk .

Firmware Update: Default "Enhance PSSR Image Quality" Toggle

The "Enhance PSSR Image Quality" setting was first added in a March 2026 system update as an opt-in toggle that force-upgraded PSSR 1.0 games to the new 2.0 upscaler . The August 2026 beta firmware now enables this toggle by default for all PS5 Pro owners in the beta program .

Once the beta update is installed, any game with PSSR support automatically runs the PSSR 2.0 pipeline without requiring a per-game patch. Sony has indicated the toggle will eventually be deprecated once all games are migrated . Beta testers can still manually disable the toggle if desired, but Sony recommends leaving it on for a "sharper and clearer visual experience" .

What This Means for PS5 Pro Owners

PSSR 2.0 is not a minor incremental update. It is a ground-up rethinking of how AI upscaling should work on console hardware. By handing color blending back to the GPU, running inference at 540p, and using elliptical Gaussian filtering at input scale, Sony has produced an upscaler that is more stable, more efficient, and vastly faster to train. And with the August 2026 beta making it the default for every PSSR game, it is quickly becoming the standard PS5 Pro experience .