Image-Text Retrieval on MSCOCO (R@1, R@5, nDCG@10)
53.4Recall@1DFN-CLIP
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DFN-CLIPArchitecture=Single-vector2026.07 | 53.4 | 77.8 | 69 | |
| SigLIP2Architecture=Single-vector2026.07 | 52.1 | 75.2 | 66.9 | |
| SaMer (w/ FT)Architecture=Multi-vector, Backbone=ColPali, Token Compression (K)=64, Compression Strategy=Merging, Adapted=true2026.07 | 51.6 | 75.5 | 67.1 | |
| EVA-CLIPArchitecture=Single-vector2026.07 | 51 | 74.6 | 66.1 | |
| OpenCLIPArchitecture=Single-vector2026.07 | 48.5 | 73.2 | 64.4 | |
| SaMer (w/ FT)Architecture=Multi-vector, Backbone=ColQwen2, Token Compression (K)=64, Compression Strategy=Merging, Adapted=true2026.07 | 47.5 | 72.3 | 63.5 | |
| ColPaliArchitecture=Multi-vector, Backbone=ColPali, Compression=None2026.07 | 47.4 | 71.7 | 63.3 | |
| GMEArchitecture=VLM-based2026.07 | 47.3 | 71.1 | 62.7 | |
| SaMer (w/o FT)Architecture=Multi-vector, Backbone=ColPali, Token Compression (K)=64, Compression Strategy=Merging, Adapted=false2026.07 | 45.1 | 69.7 | 61.1 | |
| H-PoolArchitecture=Multi-vector, Backbone=ColPali, Token Compression (K)=64, Compression Strategy=Pooling2026.07 | 44.9 | 69.6 | 60.9 | |
| ColQwen2Architecture=Multi-vector, Backbone=ColQwen2, Compression=None2026.07 | 43.3 | 68 | 59.5 | |
| VisRAGArchitecture=VLM-based2026.07 | 42.5 | 67.5 | 58.9 | |
| H-PoolArchitecture=Multi-vector, Backbone=ColQwen2, Token Compression (K)=64, Compression Strategy=Pooling2026.07 | 42.2 | 67.2 | 58.6 | |
| SaMer (w/o FT)Architecture=Multi-vector, Backbone=ColQwen2, Token Compression (K)=64, Compression Strategy=Merging, Adapted=false2026.07 | 42.1 | 67 | 58.4 | |
| SAPArchitecture=Multi-vector, Backbone=ColQwen2, Token Compression (K)=64, Compression Strategy=Structural anchor pruning2026.07 | 40.7 | 65.8 | 57.2 | |
| HPCArchitecture=Multi-vector, Backbone=ColQwen2, Token Compression (K)=64, Compression Strategy=Attention-based pruning2026.07 | 40.1 | 64.8 | 56.5 | |
| SAPArchitecture=Multi-vector, Backbone=ColPali, Token Compression (K)=64, Compression Strategy=Structural anchor pruning2026.07 | 39.6 | 63.7 | 55.4 | |
| MetaCLIPArchitecture=Single-vector2026.07 | 35.7 | 61.7 | 52.8 | |
| CLIPArchitecture=Single-vector2026.07 | 30.8 | 56 | 47.7 | |
| HPCArchitecture=Multi-vector, Backbone=ColPali, Token Compression (K)=64, Compression Strategy=Attention-based pruning2026.07 | 30.6 | 52.5 | 45.4 | |
| VLM2Vec-v2Architecture=VLM-based2026.07 | 20.2 | 44.1 | 36.5 |