Composed Image Retrieval on CIRCO (test)
43.86mAP@10UniCVR (Stage II)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| UniCVR (Stage II)Query Reasoner=Qwen3-VL-32B, Retriever=PEcore-G2026.04 | 43.86 | 43.09 | 46.24 | 47.31 | — | |
| PDFBackbone Type=ViT-L/14, Training-free=true2026.05 | 43.73 | 43.04 | 46.03 | 46.86 | — | |
| MMRet-MLLMBackbone=LLaVA-1.6, # Params=7.57B2024.12 | 43.4 | 42.2 | 46.5 | 47.6 | — | |
| ParacosmBackbone=OpenCLIP ViT-G/14, venue&year=ours2026.01 | 40.86 | 39.82 | 43.96 | 45.05 | — | |
| MMRet-LargeBackbone=CLIP-L, # Params=428M2024.12 | 40.2 | 39.2 | 42.9 | 44 | — | |
| PDFBackbone Type=ViT-B/32, Training-free=true2026.05 | 40.12 | 40.02 | 41.75 | 42.5 | — | |
| WISERBackbone=ViT-G/14, Training-free=true2026.02 | 38.14 | 36.53 | 40.46 | 41.26 | — | |
| SEIZE+SC4CIRBackbone=ViT-G/14, Training-free=true2026.06 | 36.81 | 34.48 | 38.93 | 39.88 | — | |
| SEIZE+SC4CIRBackbone=ViT-G/14, Training-free=true2026.06 | 36.81 | 34.48 | 38.93 | 39.88 | 37.53 | |
| UniCVR (Stage I)Query Reasoner=Qwen3-VL-32B, Retriever=PEcore-G2026.04 | 36.4 | 34.97 | 39.11 | 40.29 | — | |
| WISERBackbone=ViT-L/14, Training-free=true2026.02 | 36.3 | 35.1 | 38.46 | 39.15 | — | |
| STiTchArch=ViT-G/14, Train=false2026.05 | 35.56 | 34.4 | 38.07 | 40.02 | — | |
| STiTchArch=ViT-G/14, Train=false2026.05 | 35.56 | 34.4 | 38.07 | 40.02 | — | |
| MagicLens-L# Total Params=613M, Backbone Network=CoCa-L2024.03 | 35.4 | 34.1 | 38.1 | 39.2 | — | |
| MagicLens-LBackbone=CoCa-L, # Params=613M2024.12 | 35.4 | 34.1 | 38.1 | 39.2 | — | |
| MagicLensQuery Reasoner=CoCa-L, Retriever=CoCa-L2026.04 | 35.4 | 34.1 | 38.1 | 39.2 | — | |
| MMRet-BaseBackbone=CLIP-B, # Params=149M2024.12 | 35 | 34.3 | 37.6 | 38.7 | — | |
| CVSIBackbone=ViT-G/142026.01 | 34.96 | 33.51 | 37.86 | 38.98 | — | |
| IP-CIRBackbone=CLIP-G, # Params=43.8B†2024.12 | 34.3 | 32.8 | 36.9 | 38 | — | |
| IP-CIRBackbone=ViT-G/14, Training-free=true2026.02 | 34.26 | 32.75 | 36.86 | 38.03 | — | |
| SEIZEQuery Reasoner=GPT-3.5-Turbo, Retriever=CLIP-G2026.04 | 33.77 | 32.46 | 36.46 | 37.55 | — | |
| SEIZEArch=ViT-G/14, Train=false2026.05 | 33.77 | 32.46 | 36.46 | 37.55 | — | |
| SEIZEArch=ViT-G/14, Train=false2026.05 | 33.77 | 32.46 | 36.46 | 37.55 | — | |
| SEIZEBackbone=ViT-G/14, Training-free=true2026.06 | 33.77 | 32.46 | 36.46 | 37.55 | — | |
| SEIZEBackbone=ViT-G/14, Training-free=true2026.06 | 33.77 | 32.46 | 36.46 | 37.55 | 35.06 | |
| WISERBackbone=ViT-B/32, Training-free=true2026.02 | 33.18 | 32.23 | 34.82 | 35.35 | — | |
| XRBackbone Type=ViT-L/14, Training-free=true2026.05 | 32.88 | 31.38 | 35.46 | 36.5 | — | |
| CoTMRBackbone=ViT-G/14, Training-free=true2026.02 | 32.72 | 32.23 | 35.6 | 36.83 | — | |
| CoTMRQuery Reasoner=Qwen2-VL-72B, Retriever=CLIP-G2026.04 | 32.72 | 32.23 | 35.6 | 36.83 | — | |
| LDREBackbone=OpenCLIP ViT-G/14, venue&year=SIGIR’242026.01 | 32.24 | 31.12 | 34.95 | 36.03 | — | |
| LDREBackbone=ViT-G/142026.01 | 32.24 | 31.12 | 34.95 | 36.03 | — | |
| LDREBackbone=ViT-G/14, Training-free=true2026.02 | 32.24 | 31.12 | 34.95 | 36.03 | — | |
| LDREQuery Reasoner=GPT-3.5-Turbo, Retriever=CLIP-G2026.04 | 32.24 | 31.12 | 34.95 | 36.03 | — | |
| LDREArch=ViT-G/14, Train=false2026.05 | 32.24 | 31.12 | 34.95 | 36.03 | — | |
| LDREArch=ViT-G/14, Train=false2026.05 | 32.24 | 31.12 | 34.95 | 36.03 | — | |
| LDREBackbone=CLIP-G, # Params=10.3B†2024.12 | 32.2 | 31.1 | 35 | 36 | — | |
| FiREQuery Reasoner=BLIP-3, Retriever=BLIP-32026.04 | 32.08 | 31.03 | 34.4 | 35.5 | — | |
| MagicLens-B# Total Params=267M, Backbone Network=CoCa-B2024.03 | 32 | 30.8 | 34.5 | 35.6 | — | |
| MagicLens-BBackbone=CoCa-B, # Params=267M2024.12 | 32 | 30.8 | 34.5 | 35.6 | — | |
| OSrCIRArch=ViT-G/142024.12 | 31.14 | 30.47 | 35.03 | 36.59 | — | |
| OSrCIR w/ GPT-4oBackbone=OpenCLIP ViT-G/14, venue&year=CVPR’252026.01 | 31.14 | 30.47 | 35.03 | 36.59 | — | |
| OSrCIRBackbone=ViT-G/14, Training-free=true2026.02 | 31.14 | 30.47 | 35.03 | 36.59 | — | |
| OSrCIRQuery Reasoner=GPT-3.5-Turbo, Retriever=CLIP-G2026.04 | 31.14 | 30.47 | 35.03 | 36.59 | — | |
| OSrCIRArch=ViT-G/14, Train=false2026.05 | 31.14 | 30.47 | 35.03 | 36.59 | — | |
| OSrCIRArch=ViT-G/14, Train=false2026.05 | 31.14 | 30.47 | 35.03 | 36.59 | — | |
| MagicLens-L# Total Params=465M, Backbone Network=CLIP-L2024.03 | 30.8 | 29.6 | 33.4 | 34.4 | — | |
| MagicLens-LBackbone=CLIP-L, # Params=465M2024.12 | 30.8 | 29.6 | 33.4 | 34.4 | — | |
| CoVR-BLIP-2Year=2024, Pretraining Data=WV-CC-CoVIR, Zero-shot=true2023.08 | 29.55 | 28.29 | 32.18 | 33.26 | — | |
| ImageScopeVLM=CLIP-ViT-L/142025.03 | 29.23 | 28.36 | 30.81 | 31.88 | — | |
| MRA-CIRBackbone Type=ViT-L/14, Training-free=false2026.05 | 28.85 | 27.14 | 31.54 | 32.63 | — | |
| XRBackbone Type=ViT-B/32, Training-free=true2026.05 | 28.33 | 27.51 | 30.28 | 30.95 | — | |
| TFCIRYear=2023, Pretraining Data=Not Specified, Zero-shot=true2023.08 | 28.25 | 26.52 | 31.23 | 31.99 | — | |
| CVSIBackbone=ViT-L/14, Training time(h)=1.5, BLIP2+LLM inference time(s)=24.7, Retrieval time(s)=0.05, Zero-shot=true2026.01 | 28.24 | — | — | — | — | |
| CVSIBackbone=ViT-L/142026.01 | 28.24 | 26.89 | 30.79 | 31.89 | — | |
| CoTMRBackbone=ViT-L/14, Training-free=true2026.02 | 28.22 | 27.61 | 30.61 | 31.7 | — | |
| CoTMRBackbone Type=ViT-L/14, Training-free=true2026.05 | 28.22 | 27.61 | 30.61 | 31.7 | — | |
| CIReVL*Arch=ViT-G/142024.12 | 28.01 | 27.12 | 30.35 | 31.39 | — | |
| SEIZE+SC4CIRBackbone=ViT-L/14, Training-free=true2026.06 | 27.93 | 26.87 | 31.28 | 32.03 | — | |
| SEIZE+SC4CIRBackbone=ViT-L/14, Training-free=true2026.06 | 27.93 | 26.87 | 31.28 | 32.03 | 29.53 | |
| CIREVL# Total Params=14.6B*, Backbone Network=CLIP-G2024.03 | 27.6 | 26.8 | 30 | 31 | — | |
| CIReVLBackbone=CLIP-G, # Params=14.6B†2024.12 | 27.6 | 26.8 | 30 | 31 | — | |
| CIReVLArch=ViT-G/142024.12 | 27.59 | 26.77 | 29.96 | 31.03 | — | |
| CIReVLYear=2024, Pretraining Data=Not Specified, Zero-shot=true2023.08 | 27.59 | 26.77 | 29.96 | 31.03 | — | |
| CIReVLBackbone=OpenCLIP ViT-G/14, venue&year=ICLR’242026.01 | 27.59 | 26.77 | 29.96 | 31.03 | — | |
| CIReVLBackbone=ViT-G/142026.01 | 27.59 | 26.77 | 29.96 | 31.03 | — | |
| CIReVLBackbone=ViT-G/14*2023.10 | 27.59 | 26.77 | 29.96 | 31.03 | — | |
| CIReVLBackbone=ViT-G/14, Training-free=true2026.02 | 27.59 | 26.77 | 29.96 | 31.03 | — | |
| CIReVLQuery Reasoner=GPT-3.5-Turbo, Retriever=CLIP-G2026.04 | 27.59 | 26.77 | 29.96 | 31.03 | — | |
| CIReVLArch=ViT-G/14, Train=false2026.05 | 27.59 | 26.77 | 29.96 | 31.03 | — | |
| CIReVLArch=ViT-G/14, Train=false2026.05 | 27.59 | 26.77 | 29.96 | 31.03 | — | |
| CIReVLBackbone=ViT-G/14, Training-free=true2026.06 | 27.59 | 26.77 | 29.96 | 31.03 | — | |
| CIReVLBackbone=ViT-G/14, Training-free=true2026.06 | 27.59 | 26.77 | 29.96 | 31.03 | 28.84 | |
| DiCE-CIRBackbone=ViT-G/142026.07 | 27.52 | 26.23 | 29.82 | 30.88 | — | |
| IP-CIRBackbone=ViT-L/14, Training-free=true2026.02 | 27.41 | 26.43 | 29.87 | 31.07 | — | |
| IP-CIRBackbone Type=ViT-L/14, Training-free=true2026.05 | 27.41 | 26.43 | 29.87 | 31.07 | — | |
| OSrCIR w/ GPT-4oBackbone=OpenCLIP ViT-G/14, venue&year=reproduced2026.01 | 27.18 | 25.62 | 29.66 | 30.73 | — | |
| FoCoBackbone=ViT-G/14, Zero-shot=true2026.07 | 27 | 25.5 | 27.8 | 28.9 | — | |
| GRB+LCRYear=2023, Pretraining Data=Not Specified, Zero-shot=true2023.08 | 26.93 | 25.38 | 29.82 | 30.74 | — | |
| X-AlignerBackbone=BLIP (ViT-L)2026.01 | 26.86 | 25.72 | 29.16 | 30.2 | — | |
| X-AlignerBackbone=BLIP-2 (ViT-L)2026.01 | 26.49 | 25.11 | 28.83 | 29.98 | — | |
| STiTchArch=ViT-L/14, Train=false2026.05 | 26.27 | 25.55 | 28.81 | 29.99 | — | |
| STiTchArch=ViT-L/14, Train=false2026.05 | 26.27 | 25.55 | 28.81 | 29.99 | — | |
| ImageScopeVLM=CLIP-ViT-B/322025.03 | 25.82 | 25.26 | 27.15 | 28.11 | — | |
| ImageScopeVLM=CLIP-ViT-L/14, CLIP weights=from [48]2025.03 | 25.82 | 25.39 | 27.07 | 27.98 | — | |
| SEIZEArch=ViT-L/14, Train=false2026.05 | 25.82 | 24.98 | 28.24 | 28.35 | — | |
| SEIZEArch=ViT-L/14, Train=false2026.05 | 25.82 | 24.98 | 28.24 | 28.35 | — | |
| ImageScopeBackbone Type=ViT-L/14, Training-free=true2026.05 | 25.82 | 25.39 | 27.07 | 27.98 | — | |
| SEIZEBackbone=ViT-L/14, Training-free=true2026.06 | 25.82 | 24.98 | 28.24 | 29.35 | — | |
| SEIZEBackbone=ViT-L/14, Training-free=true2026.06 | 25.82 | 24.98 | 28.24 | 29.35 | 27.1 | |
| OSrCIRArch=ViT-L/142024.12 | 25.33 | 23.87 | 27.84 | 28.97 | — | |
| OSrCIRBackbone=ViT-L/142025.12 | 25.33 | 23.87 | 27.84 | 28.97 | — | |
| OSrCIRBackbone=ViT-L/14, Training-free=true2026.02 | 25.33 | 23.87 | 27.84 | 28.97 | — | |
| OSrCIRArch=ViT-L/14, Train=false2026.05 | 25.33 | 23.87 | 27.84 | 28.97 | — | |
| OSrCIRArch=ViT-L/14, Train=false2026.05 | 25.33 | 23.87 | 27.84 | 28.97 | — | |
| OSrCIRBackbone Type=ViT-L/14, Training-free=true2026.05 | 25.33 | 23.87 | 27.84 | 28.97 | — | |
| AutoCIRBackbone=ViT-L/14, Training-free=true2026.02 | 25.14 | 24.05 | 27.35 | 28.36 | — | |
| AutoCIRQuery Reasoner=GPT-4o-Mini, Retriever=CLIP-L2026.04 | 25.14 | 24.05 | 27.35 | 28.36 | — | |
| AutoCIRBackbone Type=ViT-L/14, Training-free=true2026.05 | 25.14 | 24.05 | 27.35 | 28.36 | — | |
| Ours (m=0)Training Setting=Zero-shot, Interaction Turns=m=02026.05 | 24.8 | 23.5 | — | — | — | |
| CIReVL + SoFTBackbone=ViT-L/142025.12 | 24.72 | 23.9 | 26.94 | 27.93 | — |