MoSs: Mixture of Scales for Efficient High-Resolution Autoregressive Image Generation

Since next-scale prediction was introduced as a new paradigm for autoregressive image generation, it has attracted extensive research interest. By progressively increasing resolution in a draft-to-refinement process, next-scale prediction demonstrates great potential in both generation quality and efficiency. However, at high resolutions, this paradigm faces a fundamental challenge: token sequences grow quadratically and accumulate across multiple scales, resulting in a critical performance bottleneck. Our systematic study reveals two key observations: (1) most image regions stabilize during early drafting stages, rendering subsequent full-scale refinement token-inefficient; and (2) different scales inherently present efficiency-fidelity trade-offs, suggesting that adaptive token dispatch across scales can concentrate computational resources where they yield the greatest quality gains. Motivated by these insights, we propose a training-free Mixture of Scales (MoSs) method for efficient high-resolution autoregressive image generation. MoSs breaks the strict causal dependency across scales in the final refinement steps by parallelizing scales of different resolutions, with each scale responsible for a subset of spatial regions. A lightweight frequency-based token dispatcher analyzes the drafted image and assigns regions to the appropriate scale. The outputs are then composited over the draft to produce the final high-resolution image. The scale-mixture method achieves substantial efficiency improvements with minimal impact on generation quality across various models. For instance, our implementation achieves 2.05-4.96× speedup on transformer backbone, up to 85.62% KV cache reduction, while incurring only 0.1-2.4% loss on GenEval quality metrics, as demonstrated on the state-of-the-art Infinity model.