Quantum Circuit Integration in Diffusion Models for Image Synthesis
Open Access Depositeddeveloping algorithms ready for deployment on physical quantum hardware as the technology evolves.
The sequential inference process of modern diffusion models causes a wide range of significant computational challenges in both implementation and deployment. This creates bottlenecks that prevent parallelization and result in slowed training convergence times. These models also struggle when trained on noisy or biased data, making them less useful in a resource constrained environment or real time applications. This praxis demonstrates that novel quantum loss functions and circuit architectures improve the convergence of diffusion model training and sampling quality through quantum probabilistic measurements. This architectural approach demonstrates a useful way to build quantum enhanced generative artificial intelligence. The research develops and implements quantum specific loss functions leveraging fidelity, trace distance, and quantum divergence measures, while methodically examining quantum circuit design strategies such as entanglement patterns and label encoding methods. A hybrid quantum-classical diffusion model was constructed using PennyLane's quantum simulation framework, integrating variational quantum circuits within a denoising diffusion architecture. The model produces quantum measurement probability distributions as outputs, making it easy to use with current simulators and, eventually, real quantum hardware. The MNIST Digits dataset reduced to 8x8 pixel resolution was used for the both investigating the circuit design space and for comparative validation to other works. The performance was measured using the Fréchet Inception Distance (FID), Structural Similarity Index (SSIM), and Peak Signal-to-Noise Ratio (PSNR) metrics. Some higher resolution datasets were also used for validation, such as the MNIST, MNIST Fashion, and CIFAR10 datasets. The high contrast, single channel datasets, like MNIST, demonstrated broader applicability of this approach to higher resolution data. The multichannel, more natural imagery from the CIFAR10 dataset indicated nuances between generation task and architectural decisions that may explain the observed reduction in performance. Results from this research demonstrate design guidance for quantum integration in diffusion models targeted for MNIST-like image denoising tasks. The ~28\% improvements in FID, SSIM, and PSNR over similar works, illustrate how tailored quantum loss functions and optimized circuit architectures can achieve competitive performance in denoising tasks. This research contributes foundational insights for quantum enhanced generative modeling
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.