Diffusion Step Annealing: Accelerating Autoregressive Image Generation

Overview

Recent autoregressive models like MAR, FlowAR, xAR, and Harmon have adopted diffusion sampling to enhance image generation quality. However, this approach typically requires 50-100 diffusion steps per token, leading to significant computational overhead.

We introduce Diffusion Step Annealing (DiSA), a training-free method that addresses this efficiency challenge by:

Gradually reducing diffusion steps as more tokens are generated, achieving 2.5× speedup while maintaining generation quality.

Key Insights

Our approach is motivated by the observation that as more tokens are generated, subsequent tokens follow more constrained distributions:

• Next tokens become more predictable in later generation stages
• Token variance decreases in later autoregressive steps
• Diffusion paths become straighter in later stages

Results

DiSA achieves significant speedups while maintaining generation quality:

• 2.5× speedup for FlowAR
• 1.6× speedup for xAR
• Similar quality images with fewer diffusion steps

Citation

@inproceedings{your2024paper,
    title={Diffusion Step Annealing: Accelerating Autoregressive Image Generation},
    author={Your Name},
    year={2024},
    booktitle={Conference Name},
}