<ul><li>Autoregressive (AR) models have achieved state-of-the-art performance in text and image generation.</li><li>Existing methods to speed up AR generation by generating multiple tokens at once are limited in capturing the output distribution due to token dependencies.</li><li>Distilled Decoding (DD) uses flow matching to create a deterministic mapping from Gaussian distribution to the output distribution, enabling few-step generation.</li><li>DD achieves promising results on ImageNet-256, enabling one-step generation with a speed-up of 6.3x for VAR and 217.8x for LlamaGen.</li></ul>

Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching

Discover more