Autoregressive Image Generation

Autoregressive image generation models images as sequences of visual tokens, predicting them one by one or in coarse-to-fine scales conditioned on text, layouts, or other images. Work on discrete and continuous tokenizers—including compact 1D semantic and hierarchical representations—aims to make those sequences easier to model while preserving detail and global structure. To improve speed, methods such as speculative and Jacobi decoding, token pruning, and KV-cache compression exploit spatial correlations or allocate computation adaptively; reinforcement learning and test-time guidance target prompt alignment and composition. Models now perform competitively on image benchmarks, but balancing fidelity, diversity, and efficiency remains difficult, especially when errors accumulate across long sequences or models must handle ambiguous prompts and safety constraints.