Classification on 15 Downstream Datasets
90.7Food-101LaCLIP
Evaluation results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LaCLIPModel Architecture=ViT-B/16, Pre-training Data=LAION-400M, Evaluation Protocol=Linear Probing2023.05 | 90.7 | 96.7 | 85.5 | 78.7 | 92.8 | 63.1 | 81.3 | 92.8 | 96.2 | 98.8 | 99.1 | 96.4 | 94.6 | 89.5 | 27.5 | 85.6 | |
| CLIPModel Architecture=ViT-B/16, Pre-training Data=LAION-400M, Evaluation Protocol=Linear Probing2023.05 | 90.5 | 96.9 | 85 | 78.1 | 92.1 | 57.2 | 80 | 90.9 | 95.7 | 98 | 98.7 | 96.7 | 94.7 | 90.3 | 27 | 84.8 | |
| LaCLIPModel Architecture=ViT-B/16, Pre-training Data=RedCaps, Evaluation Protocol=Linear Probing2023.05 | 90.1 | 94.3 | 78.5 | 66.6 | 77.6 | 53.6 | 73.9 | 90.8 | 91.5 | 97.9 | 97.6 | 96.6 | 92.7 | 80.8 | 17.2 | 80 | |
| CLIPModel Architecture=ViT-B/16, Pre-training Data=RedCaps, Evaluation Protocol=Linear Probing2023.05 | 89.1 | 94.1 | 78.8 | 65.6 | 74 | 52.5 | 73.2 | 91.5 | 91.4 | 97.7 | 98 | 96.3 | 93.5 | 80.8 | 17 | 79.6 | |
| LaCLIPPre-training Data=LAION-400M, Model Architecture=ViT-B/16, Evaluation Protocol=Zero-shot2023.05 | 86.5 | 93.5 | 73.9 | 67.9 | 87.1 | 24.2 | 58.9 | 90.9 | 92.4 | 73.1 | 98.4 | 48.3 | 65.8 | 46.1 | 19.6 | 68.4 | |
| CLIPModel Architecture=ViT-B/32, Pre-training Data=LAION-400M, Evaluation Protocol=Linear Probing2023.05 | 85.8 | 95.8 | 83.6 | 75.1 | 89.2 | 54.3 | 79.7 | 86.9 | 94.5 | 96.8 | 97.9 | 96.3 | 93.5 | 88.6 | 23.1 | 82.7 | |
| CLIPPre-training Data=LAION-400M, Model Architecture=ViT-B/16, Evaluation Protocol=Zero-shot2023.05 | 85.5 | 93 | 71.7 | 66.8 | 83.5 | 16.7 | 52.8 | 90.1 | 91.2 | 63.9 | 97.3 | 42.4 | 63.3 | 46.2 | 17.8 | 65.5 | |
| LaSLIPModel Architecture=ViT-B/16, Pre-training Data=CC12M, Evaluation Protocol=Linear Probing2023.05 | 85.2 | 94.6 | 80.8 | 75.1 | 77 | 53.8 | 78.5 | 85.6 | 93.7 | 96.5 | 97.9 | 96.8 | 93.5 | 76.1 | 21.1 | 80.4 | |
| LaCLIPModel Architecture=ViT-B/32, Pre-training Data=LAION-400M, Evaluation Protocol=Linear Probing2023.05 | 85.1 | 96.2 | 84.2 | 75.6 | 90.1 | 56.1 | 79.6 | 89.1 | 94.8 | 97.7 | 98.4 | 95.8 | 93.6 | 88.6 | 22.9 | 83.2 | |
| LaCLIPPre-training Data=RedCaps, Model Architecture=ViT-B/16, Evaluation Protocol=Zero-shot2023.05 | 85 | 74.8 | 40.7 | 40.3 | 21.3 | 2.2 | 23.9 | 78.2 | 76.4 | 59 | 91.4 | 27.1 | 41.3 | 5.6 | 7.6 | 45 | |
| SLIPModel Architecture=ViT-B/16, Pre-training Data=CC12M, Evaluation Protocol=Linear Probing2023.05 | 84.4 | 94.2 | 79.1 | 73.5 | 74.2 | 54.6 | 76.5 | 86.1 | 92.7 | 95.7 | 97.6 | 96.8 | 93.7 | 74 | 20.6 | 79.6 | |
| LaCLIPModel Architecture=ViT-B/16, Pre-training Data=CC12M, Evaluation Protocol=Linear Probing2023.05 | 82.9 | 94.7 | 79.7 | 73.8 | 79.9 | 54.5 | 75.7 | 87.7 | 93 | 96.4 | 98 | 96.4 | 93 | 81.9 | 19.7 | 80.5 | |
| CLIPModel Architecture=ViT-B/16, Pre-training Data=CC12M, Evaluation Protocol=Linear Probing2023.05 | 81.6 | 93.8 | 79.3 | 72 | 75.1 | 52.6 | 75.6 | 86.2 | 92.2 | 95.3 | 97.3 | 96.7 | 93.1 | 80.6 | 19.7 | 79.4 | |
| CLIPPre-training Data=RedCaps, Model Architecture=ViT-B/16, Evaluation Protocol=Zero-shot2023.05 | 81.5 | 70.4 | 39.9 | 33.2 | 19.2 | 1.9 | 19.7 | 82.7 | 72.8 | 53.9 | 92.8 | 23.3 | 33.6 | 8.3 | 6.2 | 42.6 | |
| CLIPPre-training Data=LAION-400M, Model Architecture=ViT-B/32, Evaluation Protocol=Zero-shot2023.05 | 79.9 | 91.8 | 72 | 64.6 | 77 | 15.8 | 49.9 | 84.8 | 89.3 | 64.4 | 95.3 | 43.2 | 60.6 | 36.9 | 14.5 | 62.7 | |
| LaCLIPPre-training Data=LAION-400M, Model Architecture=ViT-B/32, Evaluation Protocol=Zero-shot2023.05 | 79.7 | 92.4 | 73 | 64.9 | 81.9 | 20.8 | 55.4 | 87.2 | 91.8 | 70.3 | 97.3 | 50.6 | 61.5 | 49.4 | 16 | 66.1 | |
| LaCLIPModel Architecture=ViT-B/16, Pre-training Data=CC3M, Evaluation Protocol=Linear Probing2023.05 | 63.8 | 87.7 | 69.5 | 60.2 | 32.4 | 42.7 | 64 | 71.1 | 83.3 | 90.2 | 93.4 | 95.8 | 89.7 | 74.6 | 13.2 | 68.8 | |
| LaSLIPPre-training Data=CC12M, Model Architecture=ViT-B/16, Evaluation Protocol=Zero-shot2023.05 | 62.9 | 82 | 50.2 | 59.6 | 32.2 | 4.4 | 30.1 | 70.6 | 82.4 | 37.4 | 95 | 20.4 | 45.6 | 10.1 | 9.2 | 46.1 | |
| CLIPModel Architecture=ViT-B/16, Pre-training Data=CC3M, Evaluation Protocol=Linear Probing2023.05 | 62.6 | 86.8 | 68.1 | 58.5 | 32.8 | 40.9 | 63.4 | 69.6 | 82 | 89.4 | 91.7 | 95.9 | 89 | 71.9 | 13.3 | 67.7 | |
| LaCLIPPre-training Data=CC12M, Model Architecture=ViT-B/16, Evaluation Protocol=Zero-shot2023.05 | 60.7 | 75.1 | 43.9 | 57 | 36.3 | 5.6 | 31 | 72.4 | 83.3 | 39.9 | 95.1 | 27.3 | 44.3 | 12.7 | 8.9 | 46.2 | |
| SLIPPre-training Data=CC12M, Model Architecture=ViT-B/16, Evaluation Protocol=Zero-shot2023.05 | 52.5 | 80.7 | 46.3 | 48.8 | 24.9 | 2.3 | 25.1 | 58.6 | 77.6 | 29.2 | 89.1 | 25.8 | 36.6 | 6 | 5.7 | 40.6 | |
| CLIPPre-training Data=CC12M, Model Architecture=ViT-B/16, Evaluation Protocol=Zero-shot2023.05 | 50.8 | 64.9 | 38.5 | 44.7 | 24.1 | 2.4 | 19.4 | 64.1 | 77.4 | 33.2 | 91 | 20.1 | 38.9 | 7.3 | 5.1 | 38.8 | |
| LaCLIPPre-training Data=CC3M, Model Architecture=ViT-B/16, Evaluation Protocol=Zero-shot2023.05 | 14.2 | 57.1 | 27.5 | 35.1 | 1.6 | 1.6 | 16.6 | 15.6 | 52.7 | 14.7 | 86.2 | 15 | 24.3 | 6.4 | 1 | 24.6 | |
| CLIPPre-training Data=CC3M, Model Architecture=ViT-B/16, Evaluation Protocol=Zero-shot2023.05 | 10.3 | 54.9 | 21.8 | 25 | 0.8 | 1.4 | 10.5 | 12.8 | 43.3 | 10.2 | 77.6 | 14.1 | 19.1 | 6.9 | 0.6 | 20.6 |