Neural Motion Blur Rendering with Deformable Kernel Prediction

We present a neural post-processing method for generating high-quality motion blur from sequences of statically rendered images and associated inexpensive by-products, such as feature buffers and motion vectors. Unlike traditional postprocessing approaches that apply fixed, local blur filters based on per-pixel motion, our method learns to synthesize motion blur in a content-aware, non-local manner. At the core of our approach is a U-Net-based convolutional neural network that extends the kernel-predicting network (KPN) paradigm: rather than predicting only blending weights for local neighborhoods, our network also predicts spatial offsets, allowing each pixel to gather information from a dynamically selected set of distant pixels along motion trajectories. This design enables the network to model complex motion blur patterns, including those caused by large displacements, occlusions, or nonlinear motion, with high visual fidelity. Our method operates entirely in image space, making it suitable for integration into offline production rendering pipelines without modifying the renderer itself. We demonstrate that our approach produces motion blur effects comparable to or exceeding those generated by physically accurate rendering, while requiring only a fraction of the computational cost. We evaluate our results across multiple representative scenes and show improved quality over equal-time path tracing and existing learning-based methods.