Cytological microscopy is a critical tool in medical diagnostics, but noise and artifacts often degrade image quality and affect the accuracy of analysis. Discrete latent variable models such as Vector Quantized Variational Autoencoders (VQVAEs) and their adversarial generative equivalent (VQGANs) offer promising approaches to reconstruct high-fidelity images. In this study, we present a comprehensive comparison between a VQVAE and a VQGAN enriched with a multi-scale discriminator for the task of reconstructing cervical cell images from the SIPaKMeD and Herlev datasets. We evaluate both models quantitatively (PSNR, SSIM, LPIPS, FID) and qualitatively (visual fidelity) and discuss trade-offs between pixel level accuracy and perceptual quality. Our results show that VQVAE achieves superior PSNR (29.64 dB), SSIM (0.8913) for the SIPaKMeD dataset and PSNR (27.17 dB), SSIM (0.7402) for the Herlev dataset, while VQGAN generates more realistic texture details, despite lower PSNR and SSIM, LPIPS(0.2272) and FID(3.06) for SIPaKMeD dataset and LPIPS(0.2201) and FID(3.5) for Herlev dataset. We conclude that VQ-VAE achieves higher reconstruction accuracy by preserving structural details, while VQ-GAN delivers more realistic textures and perceptual quality.