DISK: Differentiable Sparse Kernel Complex for Efficient Spatially-Variant Convolution

DISK proposes a differentiable way to represent dense, spatially variant complex kernels using sparse kernel samples. The interesting part for game tech is that it targets the usual pain points of image-space effects: dense convolution is expensive, and many approximations either lose quality or struggle with non-convex kernel shapes.

The paper says the method includes a decomposition that can be optimized end-to-end, a special initialization strategy to avoid bad local minima on non-convex kernels, and a kernel-space interpolation scheme that extends single-kernel filtering to spatially varying filtering without retraining or extra runtime overhead. In experiments on Gaussian and non-convex kernels, it reportedly beats simulated annealing on fidelity and is much cheaper than low-rank decompositions. That makes it relevant for mobile imaging, real-time rendering, and any pipeline where you want learned or procedural filtering that still stays differentiable.