Zero-shot Composed Image Retrieval on Efficiency Comparison
0.2Training Time (h)DiCE-CIR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DiCE-CIRBackbone=ViT-L, Training GPUs=RTX A6000 × 2, Evaluation Environment=Same as DiCE-CIR2026.07 | 0.2 | 0.01 | |
| DiCE-CIRBackbone=ViT-G, Training GPUs=RTX A6000 × 2, Evaluation Environment=Same as DiCE-CIR2026.07 | 0.3 | 0.02 | |
| LinCIRBackbone=ViT-L, Training GPUs=A100 × 8, Evaluation Environment=Original paper2026.07 | 0.5 | 0.02 | |
| LinCIRBackbone=ViT-G, Training GPUs=A100 × 8, Evaluation Environment=Original paper2026.07 | 0.8 | 0.047 | |
| LinCIRBackbone=ViT-L, Training GPUs=RTX A6000 × 2, Evaluation Environment=Same as DiCE-CIR2026.07 | 2.1 | 0.05 | |
| Pic2WordBackbone=ViT-L, Training GPUs=A100 × 8, Evaluation Environment=Original paper2026.07 | 3 | 0.02 | |
| SEARLEBackbone=ViT-L, Training GPUs=A100 × 8, Evaluation Environment=Original paper2026.07 | 4.2 | 0.02 | |
| LinCIRBackbone=ViT-G, Training GPUs=RTX A6000 × 2, Evaluation Environment=Same as DiCE-CIR2026.07 | 7.3 | 0.095 | |
| Pic2WordBackbone=ViT-G, Training GPUs=A100 × 8, Evaluation Environment=Original paper2026.07 | 13.4 | 0.05 | |
| SEARLEBackbone=ViT-G, Training GPUs=A100 × 8, Evaluation Environment=Original paper2026.07 | 14.4 | 0.047 | |
| PrediCIRBackbone=ViT-L, Training GPUs=A100 × 4, Evaluation Environment=Original paper2026.07 | 28 | 0.12 |