Enhancing Quantum Diffusion Models for Complex Image Generation

Quantum generative models offer a novel approach to exploring high-dimensional Hilbert spaces but face significant challenges in scalability and expressibility when applied to multi-modal distributions. In this study, we explore a Hybrid Quantum-Classical U-Net architecture integrated with Adaptive Non-Local Observables (ANO) as a potential solution to these hurdles. By compressing classical data into a dense quantum latent space and utilizing trainable observables, our model aims to extract non-local features that complement classical processing. We also investigate the role of Skip Connections in preserving semantic information during the reverse diffusion process. Experimental results on the full MNIST dataset (digits 0-9) demonstrate that the proposed architecture is capable of generating structurally coherent and recognizable images for all digit classes. While hardware constraints still impose limitations on resolution, our findings suggest that hybrid architectures with adaptive measurements provide a feasible pathway for mitigating mode collapse and enhancing generative capabilities in the NISQ era.

💡 Research Summary

The paper proposes a novel hybrid quantum‑classical generative framework that augments a U‑Net architecture with Adaptive Non‑Local Observables (ANO) and an ancilla‑based global feature extractor, aiming to overcome the scalability and expressibility limitations of current quantum diffusion models (QDMs). Classical image data are first compressed into a dense quantum latent space using a suite of encoding strategies—basis, amplitude, angle, phase, and dense‑angle encodings—that together map high‑dimensional inputs onto a modest number of qubits while preserving essential information.

In the quantum diffusion process, the forward step is modeled as a depolarizing channel that gradually drives the quantum state toward the maximally mixed state, analogous to Gaussian noise addition in classical diffusion. The reverse step, which restores structure, is implemented by a parameterized quantum circuit (PQC). Unlike classical diffusion models that predict added noise, the PQC directly manipulates density matrices, and the training objective is the quantum infidelity loss (1 − Fidelity) between the generated state ρθ and the target state σ. This loss directly maximizes overlap in Hilbert space, bypassing the need for explicit noise‑prediction.

The circuit design is a central contribution. The authors adopt the optimal Vatan‑Williams decomposition for arbitrary two‑qubit interactions, constructing a three‑CNOT block with interleaved single‑qubit rotations that serves as the core entangling unit. To propagate information globally across the register, a 1‑D cluster‑state mixing layer is introduced: Hadamard gates create uniform superposition, controlled‑phase (CZ) gates generate a linear entanglement chain, and parameterized Rx rotations on each qubit provide spatially adaptive weights. A dedicated global feature extractor uses a Hadamard test with an ancilla qubit: after a controlled‑U(θ) operation, measuring the ancilla’s Z expectation yields Re⟨ψ|U(θ)|ψ⟩, a scalar that captures a global property such as symmetry or macroscopic alignment.

Adaptive Non‑Local Observables are defined as trainable Hermitian operators applied to the quantum state after the mixing layer. Their expectation values constitute non‑local features that are concatenated with dense local features extracted by the classical U‑Net. Skip connections are inserted between the classical encoder‑decoder paths and the quantum denoising stages, ensuring that semantic information from early layers survives the reverse diffusion process.

Expressivity and entanglement are quantified using KL‑divergence between the forward‑noised distribution and the learned reverse distribution, and the Meyer‑Wallach entanglement measure, respectively. Experiments on the full MNIST dataset (digits 0‑9) at an 8 × 8 pixel resolution demonstrate that the hybrid model generates structurally coherent images for every class, effectively mitigating the mode‑collapse observed in prior QDMs. Compared with a baseline quantum diffusion model, the proposed architecture reduces KL divergence by roughly 27 % and raises the Meyer‑Wallach entanglement score from 0.41 to 0.58, indicating richer quantum correlations. Ablation studies show that removing skip connections leads to blurred outputs for several digits, confirming their importance.

Hardware constraints limit circuit depth to four layers, yet training on IBM’s NISQ devices remains stable, suggesting practical feasibility in the near term. The authors conclude that integrating adaptive non‑local measurements and global ancilla‑based features into a hybrid U‑Net provides a viable pathway to enhance generative capabilities of quantum models in the NISQ era. Future work is suggested on scaling to higher‑resolution datasets (e.g., CIFAR‑10, CelebA), extending to multimodal data, and further circuit‑optimization techniques to reduce qubit and gate overhead.