Multimodal Autoregressive Pre-training of Large Vision Encoders

Open paper
Benchmark Model Rank Results
image-classification-on-inaturalistAIMv2-3B (448 res)#1Top 1 Accuracy: 85.9
image-classification-on-inaturalistAIMv2-3B#5Top 1 Accuracy: 81.5
image-classification-on-inaturalistAIMv2-1B#8Top 1 Accuracy: 79.7
image-classification-on-inaturalistAIMv2-H#9Top 1 Accuracy: 77.9
image-classification-on-inaturalistAIMv2-L#10Top 1 Accuracy: 76
image-classification-on-cifar-10AIMv2 ViT-3B/14 448px#3Percentage correct: 99.5
image-classification-on-cifar-100AIMv2 ViT-3B/14 448px#5Percentage correct: 94.5
visual-question-answering-vqa-on-seed-seedbenchAIMv2 ViT-3B/14 + Llama 3.0 8B#17Accuracy: 72.9
image-classification-on-imagenetAIMv2-3B (448 res)#22Top 1 Accuracy: 89.5%
image-classification-on-imagenetAIMv2-3B#49Top 1 Accuracy: 88.5%
image-classification-on-imagenetAIMv2-1B#65Top 1 Accuracy: 88.1%Number of params: 1200M
image-classification-on-imagenetAIMv2-H#93Top 1 Accuracy: 87.5%Number of params: 600M
image-classification-on-imagenetAIMv2-L#144Top 1 Accuracy: 86.6%Number of params: 300M
image-classification-on-imagenetAIMv2-2B#1108Number of params: 2700M
visual-question-answering-vqa-on-textvqa-valAIMv2 ViT-3B/14 + Llama 3.0 8B#23Accuracy: 58.2
visual-question-answering-vqa-on-infographicvqaAIMv2 ViT-3B/14 + Llama 3.0 8B#28ANLS: 23.0

Rank counts only results with a code link.