PlaySR: Real-time Super Resolution in Gaming

Real-time super-resolution (SR) for gaming is constrained by computational budgets and the need for temporal consistency. Traditional upscaling methods like TAAU leverage sub-pixel information but often suffer from blurring and ghosting. While motion vectors (MVs) provided by game engines offer a lightweight alternative for frame alignment, they are unreliable in challenging regions such as shadows, reflections, and transparency, leading to ghosting and misalignment artifacts. We propose PlaySR, a deep-learning-based SR framework that integrates neural networks into the deferred rendering pipeline, leveraging game buffers — including colour, depth, and motion vectors — for high-quality upscaling. PlaySR dynamically balances motion-vector-based warping and optical flow estimation for improved temporal alignment. To mitigate dataset biases that favour MVs, we introduce a novel augmentation strategy to enhance optical flow generalisation. Additionally, a lightweight SR model with parameter-free attention ensures real-time efficiency. Experiments show that PlaySR improves perceptual quality and temporal stability under a practical quality-efficiency trade-off.

To learn more and read the full paper, please visit OpenReview.net.