Text-to-image retrieval on IMDB
84.4Recall@1CLIP (Cosine)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| CLIP (Cosine)Backbone=CLIP, Alignment Strategy=Cosine2025.06 | 84.4 | 92.3 | 94.7 | |
| CLIP (Manhattan)Backbone=CLIP, Alignment Strategy=Manhattan2025.06 | 78.1 | 90.5 | 92.4 | |
| BLIP (Cosine)Backbone=BLIP, Alignment Strategy=Cosine2025.06 | 77 | 89.5 | 93.1 | |
| CLIP (Euclidean)Backbone=CLIP, Alignment Strategy=Euclidean2025.06 | 46.3 | 63.6 | 69.2 | |
| CLIP (MLP Contr.)Backbone=CLIP, Alignment Strategy=MLP (Contr.)2025.06 | 29 | 56.3 | 65.8 | |
| BLIP (Transformer Contr.)Backbone=BLIP, Alignment Strategy=Transformer (Contr.)2025.06 | 28.3 | 54.8 | 68.6 | |
| CLIP (Transformer Contr.)Backbone=CLIP, Alignment Strategy=Transformer (Contr.)2025.06 | 18.9 | 44.5 | 56 | |
| BLIP (MLP Contr.)Backbone=BLIP, Alignment Strategy=MLP (Contr.)2025.06 | 4.4 | 17.2 | 26 | |
| ConViT/RoBERTa (Transformer Contr.)Backbone=ConViT/RoBERTa, Alignment Strategy=Transformer (Contr.)2025.06 | 1.7 | 3.7 | 8.8 | |
| ConViT/RoBERTa (MLP Contr.)Backbone=ConViT/RoBERTa, Alignment Strategy=MLP (Contr.)2025.06 | 1.5 | 5.7 | 11.3 | |
| CLIP (Transformer MSE)Backbone=CLIP, Alignment Strategy=Transformer (MSE)2025.06 | 1.2 | 5.4 | 9.8 | |
| Meta-Transf. (Transformer Contr.)Backbone=Meta-Transf., Alignment Strategy=Transformer (Contr.)2025.06 | 1.2 | 3.4 | 7.6 | |
| BLIP (Euclidean)Backbone=BLIP, Alignment Strategy=Euclidean2025.06 | 1.1 | 4.4 | 6.3 | |
| CLIP (MLP MSE)Backbone=CLIP, Alignment Strategy=MLP (MSE)2025.06 | 1 | 3.2 | 6.1 | |
| ResNet/BERT (MLP Contr.)Backbone=ResNet/BERT, Alignment Strategy=MLP (Contr.)2025.06 | 1 | 3.4 | 7.6 | |
| ConvNeXt/XLNet (Transformer Contr.)Backbone=ConvNeXt/XLNet, Alignment Strategy=Transformer (Contr.)2025.06 | 1 | 3.7 | 7.1 | |
| BLIP (Transformer MSE)Backbone=BLIP, Alignment Strategy=Transformer (MSE)2025.06 | 0.7 | 2.2 | 3.9 | |
| ResNet/BERT (Transformer Contr.)Backbone=ResNet/BERT, Alignment Strategy=Transformer (Contr.)2025.06 | 0.7 | 6.6 | 10.6 | |
| BLIP (Manhattan)Backbone=BLIP, Alignment Strategy=Manhattan2025.06 | 0.5 | 3.2 | 5.8 | |
| Meta-Transf. (Transformer MSE)Backbone=Meta-Transf., Alignment Strategy=Transformer (MSE)2025.06 | 0.5 | 1 | 2 | |
| ConvNeXt/XLNet (MLP Contr.)Backbone=ConvNeXt/XLNet, Alignment Strategy=MLP (Contr.)2025.06 | 0.5 | 2.9 | 5.7 | |
| ResNet/BERT (Cosine)Backbone=ResNet/BERT, Alignment Strategy=Cosine2025.06 | 0.3 | 0.6 | 1 | |
| ConvNeXt/XLNet (Manhattan)Backbone=ConvNeXt/XLNet, Alignment Strategy=Manhattan2025.06 | 0.3 | 0.8 | 1 | |
| BLIP (MLP MSE)Backbone=BLIP, Alignment Strategy=MLP (MSE)2025.06 | 0.2 | 1.5 | 2.9 | |
| Meta-Transf. (MLP MSE)Backbone=Meta-Transf., Alignment Strategy=MLP (MSE)2025.06 | 0.2 | 1.2 | 2.5 | |
| ResNet/BERT (MLP MSE)Backbone=ResNet/BERT, Alignment Strategy=MLP (MSE)2025.06 | 0.2 | 1.2 | 2.2 | |
| ResNet/BERT (Transformer MSE)Backbone=ResNet/BERT, Alignment Strategy=Transformer (MSE)2025.06 | 0.2 | 1 | 2.5 | |
| ConViT/RoBERTa (Manhattan)Backbone=ConViT/RoBERTa, Alignment Strategy=Manhattan2025.06 | 0.2 | 0.6 | 1.1 | |
| ConViT/RoBERTa (Chi-Square)Backbone=ConViT/RoBERTa, Alignment Strategy=Chi-Square2025.06 | 0.2 | 0.8 | 1.5 | |
| ConViT/RoBERTa (MLP MSE)Backbone=ConViT/RoBERTa, Alignment Strategy=MLP (MSE)2025.06 | 0.2 | 1 | 1.7 | |
| ConViT/RoBERTa (Transformer MSE)Backbone=ConViT/RoBERTa, Alignment Strategy=Transformer (MSE)2025.06 | 0.2 | 1.5 | 2.7 | |
| ConvNeXt/XLNet (Cosine)Backbone=ConvNeXt/XLNet, Alignment Strategy=Cosine2025.06 | 0.2 | 0.6 | 0.9 | |
| ConvNeXt/XLNet (Transformer MSE)Backbone=ConvNeXt/XLNet, Alignment Strategy=Transformer (MSE)2025.06 | 0.2 | 1 | 2.5 | |
| BLIP (Chi-Square)Backbone=BLIP, Alignment Strategy=Chi-Square2025.06 | 0.1 | 0.6 | 1.4 | |
| Meta-Transf. (Euclidean)Backbone=Meta-Transf., Alignment Strategy=Euclidean2025.06 | 0.1 | 0.3 | 0.9 | |
| Meta-Transf. (Cosine)Backbone=Meta-Transf., Alignment Strategy=Cosine2025.06 | 0.1 | 0.5 | 1.1 | |
| Meta-Transf. (Manhattan)Backbone=Meta-Transf., Alignment Strategy=Manhattan2025.06 | 0.1 | 0.5 | 1.1 | |
| ResNet/BERT (Euclidean)Backbone=ResNet/BERT, Alignment Strategy=Euclidean2025.06 | 0.1 | 0.5 | 1 | |
| ResNet/BERT (Manhattan)Backbone=ResNet/BERT, Alignment Strategy=Manhattan2025.06 | 0.1 | 0.7 | 1 | |
| ConViT/RoBERTa (Euclidean)Backbone=ConViT/RoBERTa, Alignment Strategy=Euclidean2025.06 | 0.1 | 0.4 | 1 | |
| ConViT/RoBERTa (Cosine)Backbone=ConViT/RoBERTa, Alignment Strategy=Cosine2025.06 | 0.1 | 0.6 | 1.1 | |
| ConvNeXt/XLNet (Euclidean)Backbone=ConvNeXt/XLNet, Alignment Strategy=Euclidean2025.06 | 0.1 | 0.6 | 1.2 | |
| CLIP (Chi-Square)Backbone=CLIP, Alignment Strategy=Chi-Square2025.06 | 0 | 0.1 | 0.4 | |
| Meta-Transf. (Chi-Square)Backbone=Meta-Transf., Alignment Strategy=Chi-Square2025.06 | 0 | 0.3 | 1 | |
| Meta-Transf. (MLP Contr.)Backbone=Meta-Transf., Alignment Strategy=MLP (Contr.)2025.06 | 0 | 3.2 | 6.9 | |
| ResNet/BERT (Chi-Square)Backbone=ResNet/BERT, Alignment Strategy=Chi-Square2025.06 | 0 | 0.8 | 1.6 | |
| ConvNeXt/XLNet (Chi-Square)Backbone=ConvNeXt/XLNet, Alignment Strategy=Chi-Square2025.06 | 0 | 0.6 | 0.9 | |
| ConvNeXt/XLNet (MLP MSE)Backbone=ConvNeXt/XLNet, Alignment Strategy=MLP (MSE)2025.06 | 0 | 0 | 0 |