OS-QLR: One-Shot Quantized Latent Refinement for Fast and Efficient Image Generation

Abstract

This paper introduces One-Shot Quantized Latent Refinement (OS-QLR), a novel two-stage generative framework designed for high-quality image generation and improved computational efficiency. OS-QLR first learns a compact, discrete latent representation using a Vector Quantized Variational Autoencoder (VQ-VAE). It then employs a single-step refinement network within this latent space to produce clean, plausible samples from noisy or random inputs. Experimental results on the FashionMNIST and CIFAR-10 datasets show that OS-QLR consistently delivers superior image quality, featuring sharper details, fewer artifacts, and significantly lower Fréchet Inception Distance scores compared to unrefined VQ-VAE models. Additionally, OS-QLR demonstrates strong performance even with various levels of latent space corruption. Importantly, the training process for OS-QLR is greatly accelerated, taking only hours instead of the days or even weeks required by Diffusion Models, Generative Adversarial Networks (GANs), and Autoregressive image generation models. The non-iterative sampling method allows for rapid image generation, making OS-QLR a compelling and efficient alternative to current computationally intensive generative models. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.