Models and architecture

    Autoregressive model

    An autoregressive model generates one piece at a time, each piece conditioned on everything produced before it, rather than refining a whole output at once.

    This is how language models work — one token after another, each one shaped by the run of tokens already written. Applied to media, the pieces become image patches or chunks of audio, and the model builds the output in order rather than resolving all of it simultaneously.

    The strength is control over sequence. Because each piece is conditioned on the actual output so far rather than on a plan, the process handles variable length naturally and can be steered mid-stream. The weakness is the other side of the same coin: errors are inherited. A drift introduced early is context for everything after it, and there is no later pass that revisits the decision.

    Diffusion took the opposite bet — refine everything together, repeatedly — which is why the two families have such different failure signatures. Autoregressive output degrades along its length; diffusion output degrades in its details.

    In practice

    • Generation is sequential, so length and latency are closely tied.
    • Errors compound forward; the model cannot go back and fix an earlier choice.
    • Common in speech and language, less common in current image generation.

    The mistake to avoid

    Assuming every generative model works this way because chatbots do. Most image generation is not sequential at all.

    Related terms

    The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.