*Equal Contributors
A dominant paradigm in large multimodal models is to pair a large language de- coder with a vision encoder. While it is well-known how to pre-train and tune language decoders for multimodal tasks, it is less clear how the vision encoder should be pre-trained. A de facto standard is to pre-train the vision encoder with a discriminative objective, such as contrastive loss. This causes a mismatch between pre-training and the generative autoregressive downstream task. At the same time, following their success in the language domain, autoregressive image models have been shown to be capable of pre-training strong and scalable vision encoders. This paper presents AIMv2, a family of large, strong vision encoders pre-trained with a multimodal autoregressive objective. Thanks to a multimodal decoder that gen- erates both raw patches and text tokens. Our models excel not only at multimodal tasks but also in visual recognition benchmarks such as localization, grounding, and classification. In addition, we show that AIMv2 models are efficient to train, outperforming the current state of the art with significantly fewer samples seen during pre-training.
Model weights available on HuggingFace.
