Pixel-Light Attention for Neural Many-Light Importance Sampling

We present a neural approach for many-light importance sampling built on a pixel-light attention mechanism for guided hierarchical traversal. Prior learning-based methods confine themselves to a single coarse level of the light hierarchy and entangle pixel and light information in a single learned object, limiting how finely the sampling distribution can adapt. Our attention scores each branch via a dot product between a learnable pixel feature and a direction-aware light node feature based on spherical-harmonic bases — factorizing the two sides into separate learnable representations while remaining lightweight enough to run at every traversal step. We train all features online during rendering by minimizing a KL-divergence loss against the ideal distribution, with gradients accumulated along each path via an explicit re-traversal pass. Our method achieves consistently lower error than state-of-the-art techniques across a range of challenging scenes.