Resource-aware local quantum generative modeling for few-shot visual representations

Few-shot visual recognition requires learning new classes from only a handful of labeled examples, often in settings where annotation is costly. Generative augmentation is a natural way to expand the usable evidence for each class. Quantum generative models provide a possible route toward generative priors beyond standard classical parameterizations; however, current hardware constraints make full image synthesis a demanding and indirect target, especially when useful augmentation must produce variations that improve the downstream decision rule rather than merely increase sample counts. We propose PCQGAN, a prototype-conditioned quantum–classical generative pathway for few-shot visual representation augmentation. A frozen visual backbone maps support images into a semantic feature space, where class prototypes define local coefficient coordinates. A shallow expectation-value circuit then produces a compact measurement representation that modulates bounded, class-conditioned coefficient displacements. These displacements are reconstructed as synthetic features, locally filtered, and used by the downstream probe. To isolate the role of the quantum generative prior, we compare PCQGAN with a matched classical prior that shares the same coefficient interface, decoder, selection rule, and probe. Across few-shot visual benchmarks, PCQGAN yields consistent gains over this matched reference, indicating that the quantum prior contributes useful coefficient-level variation rather than merely increasing the amount of augmentation. Compared with a FullQGAN-Image counterpart under the same per-sample circuit budget, PCQGAN preserves comparable downstream utility through a lighter representation-level generation route. These results show that PCQGAN offers an efficient quantum–classical generative pathway for few-shot visual augmentation by synthesizing prototype-local feature variations.