AIM

Researchers at Apple have used the Autoregressive Image Model (AIM) to verify that the more parameters a visual model has, the better its performance. This further demonstrates that as the capacity or amount of pre-trained data increases, the model can continue to improve its performance. AIM can effectively utilize large amounts of unstructured image data, and its training method and stability are similar to those of recent large language models (LLMs). This observation is consistent with previous research results on scaling large language models. Although the model used in this experiment is limited in size, further exploration is needed to verify this rule on models with larger parameter magnitudes. The researchers used pre-trained...