Neural Radiance Fields (NeRF) have become a cornerstone of novel view synthesis, yet their classical implementations often struggle with the computational intensity required for high-fidelity rendering. While the recent development of Quantum Neural Radiance Fields (QNeRF) demonstrated the potential of using gate-based quantum computers for these tasks, the framework's reliance on classical sinusoidal positional encoding has proven to be a bottleneck, scaling poorly as scene complexity increases.
To address these limitations, researchers have introduced Hash-QNeRF, a hybrid architecture that replaces traditional sinusoidal encoding with multiresolution hash encoding, a technique derived from Instant-NGP. This modification is designed to enhance the efficiency of spatial coordinate processing without altering the fundamental quantum components of the pipeline, such as the amplitude embedding, parameterized quantum circuits (PQCs), and parity-based measurements.
By leveraging learnable hash grids, Hash-QNeRF achieves significantly improved memory efficiency and faster convergence rates. In tests conducted on synthetic Blender scenes, the model reached a training loss of 0.003534, yielding a peak signal-to-noise ratio (PSNR) of approximately 24.5 dB. This demonstrates that the integration of hash encoding can effectively represent complex spatial data within a quantum-classical hybrid framework.
Furthermore, the research highlights the model's robustness against the inherent noise of current quantum hardware. Testing via Qiskit's FakeKyiv and FakeTorino backends revealed state fidelities between 0.93 and 0.98. These results indicate that the introduction of multiresolution hash encoding does not degrade the noise tolerance of the underlying quantum circuits, paving the way for more scalable quantum-enhanced computer vision applications.