RadAR: Efficient Visual Generation with Radial Autoregression

This paper addresses the inefficiency of autoregressive models in visual generation by proposing RadAR, a framework that leverages spatial relationships in images to enable parallel generation. The core idea is to reorder the generation process using a radial topology, allowing for parallel prediction of tokens within concentric rings. The introduction of a nested attention mechanism further enhances the model's robustness by correcting potential inconsistencies during parallel generation. This approach offers a promising solution to improve the speed of visual generation while maintaining the representational power of autoregressive models.