The model looks ahead, just like LLMs look ahead. An LLM outputs token by token but can still output a fully coherent and consistent story for example. This new crop of auto-regressive image models does the same.