Qnerf Achieves Compact 3D Scene Rendering with Novel Hybrid Quantum-Classical Approach

The challenge of efficiently representing and rendering 3D scenes from 2D images is a central problem in computer vision, traditionally addressed by Neural Radiance Fields (NeRFs). However, these models often require substantial computational resources. Daniele Lizzio Bosco, Shuteng Wang, Giuseppe Serra, and Vladislav Golyanik, from the University of Udine and the Max Planck Institute for Informatics, present a novel approach in their work on QNeRF , a hybrid classical-quantum model for novel-view synthesis. By integrating parameterised quantum circuits with classical neural networks, QNeRF encodes spatial and view-dependent information using the principles of superposition and entanglement, achieving comparable or superior performance to existing NeRFs with significantly fewer parameters. This research demonstrates the potential of quantum-enhanced machine learning for compact and efficient 3D representation, offering a promising alternative for vision-based tasks.

This work extends the approach of Quantised Voxel Fields (QVFs) by introducing QNeRF, the first hybrid quantum-classical model specifically designed for novel-view synthesis from 2D images. The research introduces the first architecture designed for compatibility with gate-based quantum hardware, specifically for reconstructing 3D scenes from 2D images. QNeRF leverages quantum neural networks to encode spatial and view-dependent information using superposition and entanglement, resulting in models demonstrably more compact than their classical counterparts.

Quantum NeRF Achieves Compact 3D Reconstruction

Experiments reveal that QNeRF, when trained on moderate resolution images, matches or surpasses the performance of classical NeRF baselines while utilising fewer than half the parameters. Two architectural variants were explored: Full QNeRF, designed to maximise representational capabilities, and Dual-Branch QNeRF, which introduces a task-informed inductive bias by branching spatial and view-dependent state preparations. This branching significantly reduces operational complexity and enhances scalability, potentially paving the way for implementation on future quantum hardware. Measurements confirm that the Dual-Branch QNeRF achieves an average fidelity of 26.56 on simulated noisy hardware, demonstrating robustness even under imperfect conditions.

Further tests on the FakeKyiv and FakeTorino datasets show the model’s ability to reconstruct complex scenes with high fidelity. The team measured Peak Signal-to-Noise Ratio (PSNR) values, demonstrating competitive performance against classical models with significantly reduced model complexity, as indicated by parameter counts of 200k and 500k. The breakthrough delivers a new approach to 3D representation learning, offering a competitive alternative to continuous signal representation in mid-level vision tasks. Data shows that QNeRF’s innovative use of quantum states for encoding spatial information allows for efficient learning and compact model sizes. Building upon previous advances in Visual Fields, the researchers demonstrate the successful application of quantum embeddings to the more complex challenge of learning 3D representations from 2D images. Experimental results indicate that Full QNeRF consistently surpasses the performance of a classical NeRF baseline on both the Blender and LLFF datasets, achieving improvements of up to 7% and 6% in PSNR respectively. Dual-Branch QNeRF, while achieving comparable performance to the classical baseline, offers a more scalable and potentially noise-resistant architecture, maintaining over 0.8 state fidelity without error mitigation. The authors acknowledge limitations related to hardware constraints, including challenges with sampling, noise, and gradient computation, which represent areas for future development. They suggest that this research establishes an initial step towards evaluating the potential of quantum-enhanced representations for volumetric novel-view synthesis and opens avenues for exploring practical quantum advantage within computer vision.