AIM
A family of autoregressive vision models from Apple's AI/ML research team, with released code and pre-trained checkpoints for AIMv1 and AIMv2, aimed at large-scale visual and multimodal representation learning.
This repository is the entry point for AIM, a family of autoregressive vision models. AIMv2 is pre-trained with a multimodal autoregressive objective and, according to the accompanying papers, outperforms OAI CLIP and SigLIP on most multimodal understanding benchmarks and DINOv2 on open-vocabulary object detection and referring expression comprehension, with the largest variant reaching 89.5% on ImageNet using a frozen trunk. AIMv1 is the earlier, image-only autoregressive pre-training method the same authors published at ICML 2024.
Pre-trained checkpoints are released at multiple capacities and pre-training resolutions, including a distilled ViT-Large variant recommended for multimodal applications. The code supports PyTorch, JAX and, for research and experimentation on Apple Silicon, MLX, and installs with pip install.