Novel-view synthesis results on the Blender dataset. Q-NeRF (ours) produces high-quality renderings with accurate material representation and fine details, while using significantly fewer parameters with respect to classical NeRF models. Full vs Dual-Branch Q-NeRF represents a trade-off between quantum encoding complexity and performance.

Abstract

Recently, Quantum Visual Fields (QVFs) have shown promising improvements in model compactness and convergence speed for learning 2D images. Meanwhile, novel-view synthesis has seen major advances with Neural Radiance Fields (NeRFs), where models learn a compact representation from 2D images to render 3D scenes, albeit at the cost of large models and intensive training. In this work, we extend the approach of QVFs by introducing Q-NeRF, the first hybrid quantum-classical model designed for novel-view synthesis from 2D images. QNeRF leverages parameterized quantum circuits to encode spatial and view-dependent information via quantum superposition and entanglement, resulting in more compact models. We present two architectural variants. Full QNeRF maximally exploits all quantum amplitudes to enhance representational capabilities. In contrast, Dual-Branch QNeRF introduces a task-informed inductive bias by branching spatial and view-dependent quantum state preparations, drastically reducing the complexity of this operation and ensuring scalability and potential hardware compatibility. Our experiments demonstrate that—when trained on images of reduced resolution—QNeRF matches or outperforms classical NeRF baselines while using less than half the number of parameters. These results suggest that Quantum Machine Learning can serve as a competitive alternative for continuous signal representation in high-level tasks in Computer Vision, such as 3D representation learning.

Quantum Neural Radiance Fields (Q-NeRF)

In this work, we propose Q-NeRF, a novel hybrid quantum-classical architecture for neural radiance fields. We design two variants of Q-NeRF: Full Q-NeRF (left) and Dual-Branch Q-NeRF (right). Both architectures leverage parameterized quantum circuits to encode 3D spatial coordinates and 2D view directions into quantum states, which are then processed to predict color and density values for novel-view synthesis. Full Q-NeRF utilizes a single amplitude embedding step to encode both spatial and view-dependent information, maximizing the use of quantum amplitudes for enhanced representational capacity. In contrast, Dual-Branch Q-NeRF employs separate quantum encodings for spatial coordinates and view directions, introducing an inductive bias that simplifies state preparation and improves scalability. This design choice allows Dual-Branch Q-NeRF to maintain competitive performance while (exponentially) reducing quantum resource requirements.

Q-NeRF requires significantly fewer parameters than classical NeRF models while achieving comparable or superior performance in novel-view synthesis tasks. The trade-off between Full and Dual-Branch Q-NeRF architectures allows for flexibility in balancing quantum resource demands and model accuracy, making Q-NeRF a promising approach for efficient 3D scene representation and rendering. The reduced complexity of quantum encoding in Dual-Branch Q-NeRF allows also for better potential deployment on near-term quantum hardware, due to smaller depth and higher noise tolerance on real devices.

MLP-Based Quantum Embedding

In the Full embedding, the features obtained after positional encoding are mapped by an MLP to a real-valued vector \(\mathcal{M}(x) \in \mathbb{R}^{2^n}\), which is normalized and directly amplitude-encoded into an \(n\)-qubit quantum state:

\[ \lvert \phi(x) \rangle = \sum_{i=0}^{2^n-1} \alpha_i(x)\,\lvert i \rangle, \qquad \sum_i \lvert \alpha_i(x) \rvert^2 = 1. \]

In the Dual-Branch embedding, positional and view-dependent features are treated separately, similarly to what is done in classical NeRFs. Two MLPs produce amplitude vectors of sizes \(2^{n_p}\) and \(2^{n_v}\), which are independently encoded and combined into a tensor-product state:

\[ \lvert \phi(x) \rangle = \lvert \phi_p(x_p) \rangle \otimes \lvert \phi_v(x_v) \rangle . \]

This construction uses \(n_p+n_v=n\) qubits while requiring only \(2^{n_p}+2^{n_v}\) amplitudes instead of \(2^n\), resulting in an exponential reduction of the state preparation complexity.

Output Scaling

Visualisation of the effect of output scaling for a Full QNeRF on the drums scene.

Variational quantum circuits exhibit exponential concentration, where output distributions collapse around the mean as the number of qubits grows. To mitigate this, each channel output \(\mathcal{O}_c(\mathbf{x})\) is multiplied by a learnable scaling factor \(\alpha_c\) and clipped to \([0,1]\):

\[ \text{output}_c = \text{clip}\big(\alpha_c \mathcal{O}_c(\mathbf{x}), 0, 1\big) \]

This scaling widens the range of outputs, counteracting the collapse due to exponential concentration and improving trainability, particularly for high-variance quantities such as predicted densities.

Experimental Results

Mesh Extraction and 3D Model Visualization

Our model can be combined with each technique that can be applied to classical Radiance Fields. Here, we demonstrate the extraction of 3D meshes from the learnt radiance fields using the Marching Cubes algorithm.

BibTex

            @article{LizzioBosco2026QNeRF, 
                title={QNeRF: Neural Radiance Fields on a Simulated Gate-based Quantum Computer}, 
                author={{Lizzio Bosco}, Daniele and Wang, Shuteng and Serra, Giuseppe and Golyanik, Vladislav}, 
                journal={arXiv preprint arXiv:2601.05250}, 
                year={2026} 
            }
            

Contact information