SOTA Quantum Generative Adversarial Networks have trouble generating high-quality images and usually tend to use lower-resolution images. A lot of the recent QGANs use a smaller version (usually 8x8 or 16x16) version of the MNIST dataset. We need the ability for GANs to be able to produce higher-quality images as well.
Additionally, we want the ability to generate images even when we have a limited number of available qubits to work with. In this project we try to adapt the PatchGAN [1] model to a larger dataset.
We follow the same Quantum ansatz as proposed in the original PatchGAN paper, but we adapt it for a larger dataset using the number of qubits in a range from 5 to 9.
We start with a random noise vector and iteratively train the Generator and Discriminator to produce better-quality images.
Generating the Digit 0 using 5 qubits
Generating the Digit 3 using 5 qubits
Generating the Digit 9 using 6 qubits
The images are not clear and still have some noise in the background. Additionally, the images are very similar and we would like the generator to produce different-looking images. The next steps could be to experiment with the architecture (changing gate types and adding more layers). We could also try using a quantum discriminator instead of a classical discriminator, to fully realize the quantum advantage.
[1] Huang, He-Liang, et al. "Experimental quantum generative adversarial networks for image generation." Physical Review Applied 16.2 (2021): 024051.