Scale invariance may fix why autoregressive image generation lags diffusion models

burkov · x · 2026-11-08

Andriy Burkov breaks down the core problem of autoregressive image generation: generating images left-to-right token by token (like LLMs generate text) is highly attractive for unified text/speech/vision systems, yet with continuous latents from standard image autoencoders, autoregressive models consistently trail diffusion models and bidirectional masked generators.

The root cause: standard continuous latents vary widely in scale across dimensions and tokens; these discrepancies amplify step by step during sequential generation, made worse by steering methods like classifier-free guidance, ultimately causing generative failure. The article argues enforcing scale invariance across the autoregressive process resolves the instability.

Original post →

More from Multimodal

Multimodal channel →