Image Classification on STL-10 (Top-1 Accuracy)
99.6Top-1 AccuracyEVA-CLIP-18B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| EVA-CLIP-18BZero-shot=true2024.02 | 99.6 | — | |
| EVA-01-CLIP-g/14+Zero-shot=true2024.02 | 99.5 | — | |
| InternVL-CZero-shot=true2024.02 | 99.4 | — | |
| EVA-02-CLIP-E/14+Zero-shot=true2024.02 | 99.3 | — | |
| EVA-CLIP-8BZero-shot=true2024.02 | 99.3 | — | |
| DFN5B-CLIP-H/14+Zero-shot=true2024.02 | 99.1 | — | |
| EVA-01-CLIP-g/14Zero-shot=true2024.02 | 98.9 | — | |
| DFN5B-CLIP-H/14Zero-shot=true2024.02 | 98.9 | — | |
| Mimetic init.Backbone=ViT-B, Pre-training=ImageNet-1K, Initialization/Warm-up=Mimetic init. [51], Evaluation Protocol=Fine-tuned2025.11 | 98.81 | — | |
| Procedural warm-upBackbone=ViT-B, Pre-training=ImageNet-1K, Initialization/Warm-up=Procedural warm-up (ours), Evaluation Protocol=Fine-tuned2025.11 | 98.66 | — | |
| Random init.Backbone=ViT-B, Pre-training=ImageNet-1K, Initialization/Warm-up=Random init. [10], Evaluation Protocol=Fine-tuned2025.11 | 98.55 | — | |
| OpenCLIP-G/14Zero-shot=true2024.02 | 98.5 | — | |
| FractalDBBackbone=ViT-B, Pre-training=ImageNet-1K, Initialization/Warm-up=FractalDB [23] warm-up, Evaluation Protocol=Fine-tuned2025.11 | 98.46 | — | |
| xCLIPLinear probing=true, Backbone=ViT-B/16, Pre-train=IT35M2022.10 | 98.1 | — | |
| OpenCLIPImage Encoder=ViT-B/16, Zero-shot=true2026.02 | 97.9 | — | |
| CLIPBackbone=CLIP ViT-L-14, zero-shot=true2023.05 | 97.7 | — | |
| POS PCABackbone=CLIP ViT-L-14, zero-shot=true2023.05 | 97.7 | — | |
| POS PGABackbone=CLIP ViT-L-14, zero-shot=true2023.05 | 97.7 | — | |
| CLIPImage Encoder=ViT-B/16, Zero-shot=true2026.02 | 97.6 | — | |
| nCLIPLinear probing=true, Backbone=ViT-B/16, Pre-train=IT35M2022.10 | 97.4 | — | |
| TinyCLIPImage Encoder=ViT-39M/16, Zero-shot=true2026.02 | 97.3 | — | |
| CLIP-MapbaseImage Encoder=ViT-39M/16, Zero-shot=true2026.02 | 97.3 | — | |
| CLIPLinear probing=true, Backbone=ViT-B/16, Pre-train=IT35M2022.10 | 97.1 | — | |
| Noun SubmanifoldBackbone=CLIP ViT-L-14, zero-shot=true2023.05 | 96.9 | — | |
| Noun SubmanifoldBackbone=CLIP ViT-B-32, Evaluation Protocol=Zero-shot2023.05 | 96.3 | — | |
| CLIPBackbone=CLIP ViT-B-32, Evaluation Protocol=Zero-shot2023.05 | 96.1 | — | |
| POS PCABackbone=CLIP ViT-B-32, Evaluation Protocol=Zero-shot2023.05 | 96.1 | — | |
| POS PGABackbone=CLIP ViT-B-32, Evaluation Protocol=Zero-shot2023.05 | 95.8 | — | |
| IL-CLIPPre-training Dataset=CC12M, Backbone=ViT-B/32, Zero-shot=true2024.04 | 94.1 | — | |
| †TinyCLIPImage Encoder=ViT-8M/16, Zero-shot=true2026.02 | 92.5 | — | |
| SAMix-CBackbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=SAMix-C, Predefined Cluster Information=true2021.11 | 92.45 | — | |
| SAMix-IBackbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=SAMix-I2021.11 | 92.41 | — | |
| CLIP-MapsmallImage Encoder=ViT-8M/16, Zero-shot=true2026.02 | 92.3 | — | |
| MoCo.V2(C) + Inter-Intra*Backbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=Inter-Intra*, Predefined Cluster Information=true, Modified by Authors=true2021.11 | 92.24 | — | |
| MoCo.V2 + MixupBackbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=Mixup2021.11 | 92.2 | — | |
| B-12/Wi-1024 + DA + TTABlocks=12, Width=1024, Data Augmentation=true, Test-Time Augmentation=true, Pre-training=ImageNet21k, Evaluation Protocol=Fine-tuning2023.06 | 92.2 | — | |
| BYOLBackbone=ResNet-18, Epochs=1000, linear=true2023.12 | 91.99 | — | |
| MoCo.V2(C) + PuzzleMix*Backbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=PuzzleMix*, Predefined Cluster Information=true, Modified by Authors=true2021.11 | 91.98 | — | |
| Ours (IE-CL)Backbone=ResNet-18, Batch Size=256, Evaluation Protocol=Linear probe2026.03 | 91.9 | — | |
| MoCo.V2 + ManifoldMixBackbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=ManifoldMix2021.11 | 91.65 | — | |
| MoCo.V2 + FMix*Backbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=FMix*, Modified by Authors=true2021.11 | 91.56 | — | |
| MinEntBackbone=ResNet-18, Batch Size=256, Evaluation Protocol=Linear probe2026.03 | 91.5 | — | |
| MoCo-v3Backbone=ResNet-18, Batch Size=256, Evaluation Protocol=Linear probe2026.03 | 91.4 | — | |
| S3OCBackbone=ResNet-18, Batch Size=256, Evaluation Protocol=Linear probe2026.03 | 91.4 | — | |
| OURS_GCProtocol=Linear probing2023.03 | 91.31 | — | |
| BYOLBackbone=ResNet-18, Batch Size=256, Evaluation Protocol=Linear probe2026.03 | 91.3 | — | |
| Mixed Barlow TwinsBackbone=ResNet-18, Epochs=1000, linear=true2023.12 | 91.04 | — | |
| Mixed Barlow TwinsBackbone=ResNet-18, Epochs=2000, linear=true2023.12 | 91.02 | — | |
| OURS_SepProtocol=Linear probing2023.03 | 90.96 | — | |
| MoCo.V2 + ResizeMix*Backbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=ResizeMix*, Modified by Authors=true2021.11 | 90.83 | — | |
| MoCo.V2 + CutMixBackbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=CutMix2021.11 | 90.81 | — | |
| MoCo.V2 + SaliencyMixBackbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=SaliencyMix2021.11 | 90.77 | — | |
| OURS_BrProtocol=Linear probing2023.03 | 90.69 | — | |
| SimCLRBackbone=ResNet-18, Epochs=1000, linear=true2023.12 | 90.51 | — | |
| SimSiamBackbone=ResNet-18, Batch Size=256, Evaluation Protocol=Linear probe2026.03 | 90.5 | — | |
| W-MSEBackbone=ResNet-18, Epochs=1000, linear=true2023.12 | 90.36 | — | |
| Codebook-CLIPPre-training Dataset=CC12M, Backbone=ViT-B/32, Zero-shot=true2024.04 | 90.3 | — | |
| SimCLRBackbone=ResNet-18, Batch Size=256, Evaluation Protocol=Linear probe2026.03 | 90.1 | — | |
| B-12/Wi-1024 + DABlocks=12, Width=1024, Data Augmentation=true, Test-Time Augmentation=false, Pre-training=ImageNet21k, Evaluation Protocol=Fine-tuning2023.06 | 89.9 | — | |
| NegCLIPPre-training Dataset=CC12M, Backbone=ViT-B/32, Zero-shot=true2024.04 | 89.7 | — | |
| MoCo.V2Backbone=ResNet-50, Pre-training Epochs=800, SSL Method=MoCo.V22021.11 | 89.68 | — | |
| OursDistortion=Random Missing Pixels (RM), Evaluation Protocol=Linear Probing2021.10 | 89.44 | — | |
| CLIPPre-training Dataset=CC12M, Backbone=ViT-B/32, Zero-shot=true2024.04 | 89.3 | — | |
| Mixed Barlow TwinsBackbone=ResNet-18, Epochs=2000, k-NN=true2023.12 | 88.94 | — | |
| SAMix-CBackbone=ResNet-50, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=SAMix-C, Predefined Cluster Information=true2021.11 | 88.91 | — | |
| MoCo-v2Backbone=ResNet-18, Batch Size=256, Evaluation Protocol=Linear probe2026.03 | 88.9 | — | |
| SAMix-IBackbone=ResNet-50, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=SAMix-I2021.11 | 88.87 | — | |
| SWAV (C)Backbone=ResNet-50, Pre-training Epochs=800, SSL Method=SWAV, Predefined Cluster Information=true2021.11 | 88.79 | — | |
| BYOLBackbone=ResNet-18, Epochs=1000, k-NN=true2023.12 | 88.64 | — | |
| SAMix-CBackbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=SAMix-C, Predefined Cluster Information=true2021.11 | 88.63 | — | |
| SAMix-IBackbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=SAMix-I2021.11 | 88.58 | — | |
| MoCo.V2(C) + PuzzleMix*Backbone=ResNet-50, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=PuzzleMix*, Predefined Cluster Information=true, Modified by Authors=true2021.11 | 88.4 | — | |
| MoCo.V2(C) + Inter-Intra*Backbone=ResNet-50, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=Inter-Intra*, Predefined Cluster Information=true, Modified by Authors=true2021.11 | 88.33 | — | |
| MoCo.V2 + MixupBackbone=ResNet-50, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=Mixup2021.11 | 88.24 | — | |
| Mixed Barlow TwinsBackbone=ResNet-18, Epochs=1000, k-NN=true2023.12 | 88.23 | — | |
| Barlow TwinsBackbone=ResNet-18, Epochs=1000, linear=true2023.12 | 88.21 | — | |
| MoCo.V2 + FMix*Backbone=ResNet-50, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=FMix*, Modified by Authors=true2021.11 | 88.14 | — | |
| MoCo.V2(C) + PuzzleMix*Backbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=PuzzleMix*, Predefined Cluster Information=true, Modified by Authors=true2021.11 | 88.07 | — | |
| MoCo.V2 + ManifoldMixBackbone=ResNet-50, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=ManifoldMix2021.11 | 88.06 | — | |
| MoCo.V2 + MixupBackbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=Mixup2021.11 | 87.93 | — | |
| MoCo.V2(C) + Inter-Intra*Backbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=Inter-Intra*, Predefined Cluster Information=true, Modified by Authors=true2021.11 | 87.85 | — | |
| MoCo.V2 + ManifoldMixBackbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=ManifoldMix2021.11 | 87.7 | — | |
| IL-CLIPPre-training Dataset=CC3M, Backbone=ViT-B/32, Zero-shot=true2024.04 | 87.7 | — | |
| W-MSEBackbone=ResNet-18, Batch Size=256, Evaluation Protocol=Linear probe2026.03 | 87.7 | — | |
| MoCo.V2 + FMix*Backbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=FMix*, Modified by Authors=true2021.11 | 87.68 | — | |
| CyCLIPProtocol=Linear probing2023.03 | 87.61 | — | |
| MoCo.V2 + CutMixBackbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=CutMix2021.11 | 87.6 | — | |
| MoCo.V2 + CutMixBackbone=ResNet-50, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=CutMix2021.11 | 87.51 | — | |
| MoCo.V2 + SaliencyMixBackbone=ResNet-50, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=SaliencyMix2021.11 | 87.35 | — | |
| MoCo.V2 + SaliencyMixBackbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=SaliencyMix2021.11 | 87.27 | — | |
| MoCo.V2 + ResizeMix*Backbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V2, Mixup Variant=ResizeMix*, Modified by Authors=true2021.11 | 87.25 | — | |
| CLIPProtocol=Linear probing2023.03 | 87.22 | — | |
| W-MSEBackbone=ResNet-18, Epochs=1000, k-NN=true2023.12 | 87.1 | — | |
| MoCo.V2 + ResizeMix*Backbone=ResNet-50, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=ResizeMix*, Modified by Authors=true2021.11 | 86.88 | — | |
| OursDistortion=Additive Gaussian Noise (GN), Evaluation Protocol=Linear Probing2021.10 | 86.3 | — | |
| SimCLRBackbone=ResNet-18, Epochs=1000, k-NN=true2023.12 | 85.68 | — | |
| MoCo.V2Backbone=ResNet-18, Pre-training Epochs=800, SSL Method=MoCo.V22021.11 | 85.64 | — | |
| SAMix-CBackbone=ResNet-18, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=SAMix-C, Predefined Cluster Information=true2021.11 | 85.6 | — | |
| SWAV (C)Backbone=ResNet-18, Pre-training Epochs=800, SSL Method=SWAV, Predefined Cluster Information=true2021.11 | 85.56 | — | |
| SAMix-IBackbone=ResNet-18, Pre-training Epochs=400, SSL Method=MoCo.V2, Mixup Variant=SAMix-I2021.11 | 85.44 | — |