List your company

Image Classification on ImageNet 1k (v2)

84Top-1 AccuracySwinV2-G
42.08852.96963.8574.731Oct 1, 2021Jun 14, 2022Feb 25, 2023Nov 8, 2023Jul 21, 2024Apr 3, 2025Dec 16, 2025

Evaluation results

MethodLinks
SwinV2-Gparam=3.0B, pre-train images=IN-22K-ext-70M, pre-train length (#im)=3.5B, pre-train im size=192x192, pre-train time=<0.5k†, fine-tune im size=640x6402021.1184
ViT-Gparam=1.8B, pre-train images=JFT-3B, pre-train length (#im)=164B, pre-train im size=224x224, pre-train time=>30k, fine-tune im size=518x5182021.1183.33
SwinV2-Lparam=197M, pre-train images=IN-22K-14M, pre-train length (#im)=1.3B, pre-train im size=192x192, pre-train time=<20†, fine-tune im size=384x3842021.1178.31
SwinV2-Bparam=88M, pre-train images=IN-22K-14M, pre-train length (#im)=1.3B, pre-train im size=192x192, pre-train time=<30†, fine-tune im size=384x3842021.1178.08
SwinV1-Lparam=197M, pre-train images=IN-22K-14M, pre-train length (#im)=1.3B, pre-train im size=224x224, pre-train time=<10†, fine-tune im size=384x3842021.1177.46
SwinV1-Bparam=88M, pre-train images=IN-22K-14M, pre-train length (#im)=1.3B, pre-train im size=224x224, pre-train time=<30†, fine-tune im size=384x3842021.1176.58
Ours (DINOv2)Data=INet-22k, Arch.=L/142024.0674.8
SuperCLIPPretrain=L-12.8B, Zero-shot=true2025.1272.8
CLIPPretrain=L-12.8B, Zero-shot=true2025.1272
ResNet-50 (A1)Training Procedure=A1, Number of epochs=600, Training resolution=224 x 224, Training time=110h, Peak memory by GPU (MB)=22,095, Numbers of GPU=42021.1068.7
ResNet-50 (A2)Training Procedure=A2, Number of epochs=300, Training resolution=224 x 224, Training time=55h, Peak memory by GPU (MB)=22,095, Numbers of GPU=42021.1067.9
MAEData=INet-1k, Arch.=H/142024.0666.5
ResNet-50 (A3)Training Procedure=A3, Number of epochs=100, Training resolution=160 x 160, Training time=15h, Peak memory by GPU (MB)=11,390, Numbers of GPU=42021.1066.1
I-JEPAData=INet-1k, Arch.=H/142024.0666
SuperCLIPPretrain=L-512M, Zero-shot=true2025.1262.5
I-JEPAData=INet-22k, Arch.=g/162024.0659.7
BEITData=INet-1k, Arch.=L/162024.0659
CLIPPretrain=L-512M, Zero-shot=true2025.1257.4
SuperCLIPPretrain=B-512M, Zero-shot=true2025.1255.2
CLIPPretrain=B-512M, Zero-shot=true2025.1253
AIMData=DFN-2B+, Arch.=H/142024.0643.7