by 
We already know that PlayStation Spectral Super Resolution (PSSR) 2.0 produces much better results on PS5 Pro than the first-gen model, but a new SIGGRAPH 2026 talk from Sony principal software engineer Daniel Craig goes into a lot more detail about how the upgraded model works under the hood. In short, it represents a significant shift in philosophy, moving away from a primarily neural network-based approach to a more efficient hybrid design. By asking the neural network to do less, the Sony and AMD engineers behind PSSR managed to address key limitations found in the original implementation, where the system attempted to handle everything from temporal accumulation to complex HDR colour blending.
In the original PSSR model, a colour-predicting convolutional neural network was tasked with reconstructing the final image. While neural networks are brilliant at pattern recognition, it turned out that they often struggled with HDR colour blending - a task for which conventional GPU mathematics is significantly better suited. Furthermore, the original approach suffered from a "vanishing gradient" problem, where the model became biased towards short-term associations and struggled to maintain stability over time, particularly in foliage and low-frequency lighting details like ray-traced global illumination (RTGI).
To solve these issues, the new PSSR architecture transitions from a colour-predicting network to a kernel-predicting network (KPN). The neural network now focuses strictly on identifying edges and distinguishing detail from noise, while conventional GPU operations handle the actual HDR colour blending and final reconstruction. By offloading tasks to standard mathematical operations, the model no longer needs to "learn" blending logic that existing hardware handles more efficiently.
A key efficiency gain in this new iteration is that the majority of neural network processing can now be performed at a lower resolution - specifically at 540p or below. The previous model spent considerable capacity on high-resolution operations, which limited its performance ceiling. Additionally, the scaling mechanism has been overhauled - while the original PSSR defined its 3x3 kernel at the output resolution, resulting in a soft appearance at high scaling ratios due to sampling a small number of rendered pixels, the new PSSR is defined at the input scale using elliptical filtering.
These architectural changes have also yielded massive improvements in development efficiency. The training time for the new PSSR model has been reduced from four GPU months to under a single GPU week, largely due to significant changes in quantisation. While the presentation doesn't provide specific frame-time or memory figures, Sony states that both metrics improve upon those seen in the original PSSR implementation.
Looking ahead, it is worth noting the limits of this new tech as described in the talk. The presentation focused exclusively on first-party test content, and contrary to some earlier rumours, there appears to be no transformer element involved in this specific update. While Kernel Predicting Networks are not a novel concept in themselves, their refined application here demonstrates a smarter, more targeted use of machine learning that promises better stability and efficiency for the next generation of upscaling.