Wave-optical adaptation of feed-forward 4D Gaussian splatting for real-time dynamic holographic rendering

Abstract Dynamic holographic rendering requires a scene representation that preserves motion while supporting coherent wave propagation. Feed-forward 4D Gaussian splatting predicts dynamic radiance primitives without test-sequence optimization, but its camera-space renderer cannot generate phase-only holograms. We present FF4D-HoloGS, a shared wave-optical adapter that maps each predicted primitive to a wavelength-conditioned complex weight. A depth-aware Gaussian wave renderer converts these weights into quantized phase patterns for reconstruction across focal planes and shifted pupils. The adapter is trained once on Kubric-4D and evaluated on four synthetic and real dynamic-scene benchmarks, together with repeated optical captures on two N3DV sequences. At 520 nm with one 8-bit phase pattern, FF4D-HoloGS reaches 27.4 dB at 23.6 frames per second. It improves feed-forward per-frame Gaussian wave splatting by 1.0 dB and reduces speckle, temporal, and angular errors, while remaining within 0.9 dB of scene-optimized 4D Gaussian splatting that requires 1800 s of test-sequence fitting. The resulting model generates phase-only holograms directly from feed-forward dynamic Gaussian representations.