Also, simply because HART uses an autoregressive design to perform the bulk from the get the job done — the exact same sort of product that powers LLMs — it is more compatible for integration with the new course of unified vision-language generative models. Accelerating components dev