Unitary Scrambling and Collapse: A Quantum Diffusion Framework for Generative Modeling

Abstract: Quantum computing, with its promise of exponential speedups, is rapidly emerging as a powerful paradigm for advancing artificial intelligence. We propose QSC-Diffusion, the first fully quantum diffusion-based framework for image generation. Our method integrates classical Gaussian noise with quantum scrambling in the forward process, and employs parameterized quantum circuits with measurement-induced collapse for reverse denoising -- enabling end-to-end sampling without reliance on classical neural architectures or preprocessing modules. To address optimization challenges in deep quantum models, we introduce a hybrid loss that balances fidelity and diversity, coupled with a divide-and-conquer training strategy to mitigate barren plateaus. Remarkably, QSC-Diffusion achieves competitive FID scores across multiple datasets while using orders of magnitude fewer parameters, outperforming even some quantum-classical hybrid baselines in efficiency. These results highlight the potential of quantum-native generative modeling and mark a foundational step toward scalable quantum machine learning.