In autoregressive models such as GPT-Image-2, the word order and structure of the prompt are more important than in diffusion models, because the model processes tokens sequentially and earlier descriptions tend to receive higher attention weights.