Generative Refinement Network for Visual Synthesis

While diffusion models dominate the field of visual genera-tion, they remain computationally inefficient, as they allocate uniformcomputational effort to samples with varying levels of complexity. In con-trast, autoregressive (AR) models are inherently complexity-aware, asevidenced by their variable likelihoods, but are often hindered by lossydiscrete tokenization and error accumulation. In this work, we introduceGenerative Refinement Networks (GRN), a next-generation visual syn-thesis paradigm to address these issues. At its core, GRN addresses thediscrete tokenization bottleneck through a theoretically near-lossless Hi-erarchical Binary Quantization (HBQ), achieving a reconstruction qual-ity comparable to continuous counterparts. Built upon HBQ’s latentspace, GRN fundamentally upgrades AR generation with a global re-finement mechanism that progressively perfects and corrects artworks —like a human artist painting. Besides, GRN integrates an entropy-guidedsampling strategy, enabling complexity-aware, adaptive-step generationwithout compromising visual quality. On the ImageNet benchmark, GRNestablishes new records in image reconstruction (0.56 rFID) and class-conditional image generation (1.81 gFID). We also scale GRN to morechallenging text-to-image and text-to-video generation, delivering supe-rior performance on an equivalent scale. We release all models and codeto foster further research on GRN.