List your company

Composed Image Retrieval on CIRCO

56.54mAP@5OSCAR
6.786419.703232.6245.5368Mar 28, 2024Aug 12, 2024Dec 28, 2024May 15, 2025Sep 29, 2025Feb 14, 2026Jul 2, 2026

Evaluation results

MethodLinks
OSCARType=Agentic, Training Free=true2026.0256.54—58.5362.6761.92
V-Retrver-7BZero-shot=true, Query format=(q_i, q_t) -> c_i2026.0248.2————
U-MARVEL+Reranker=true2025.0746————
QQMM-embed-v2Type=Multimodal Embedding, Training Free=true2026.0245.92—47.1351.4550.39
LamRAreranking=pointwise/listwise2024.1242.8————
LamRA-7BZero-shot=true, Query format=(q_i, q_t) -> c_i2026.0242.8————
LamRAReranker=true2025.0742.8————
ParacosmBackbone=OpenCLIP ViT-G/142026.0139.82————
ORACLE-QPolicy=ORACLE-Q, Turn budget (T)=T32026.0538.1————
U-MARVELReranker=false2025.0736.2————
Ours (EIG)Policy=Ours (EIG), Turn budget (T)=T32026.0536————
ORACLE-QPolicy=ORACLE-Q, Turn budget (T)=T22026.0535.6————
MM-Embed-7BZero-shot=true, Query format=(q_i, q_t) -> c_i2026.0235.5————
MagicLens-LBackbone=CoCa-L, # Total Params=613M2024.0334.170.9———
Ours (EIG)Policy=Ours (EIG), Turn budget (T)=T22026.0533.4————
LamRA-Ret2024.1233.2————
LamRA-RetReranker=false2025.0733.2————
RzenEmbed-v2-7BType=Multimodal Embedding, Training Free=true2026.0232.2—34.1938.6137.41
X^RType=Agentic, Training Free=true2026.0231.38—32.8836.535.46
MLLM-QPolicy=MLLM-Q, Turn budget (T)=T32026.0531.2————
FiREType=CIR Dedicated, Training Free=false2026.0231.03—32.0835.534.4
ORACLE-QPolicy=ORACLE-Q, Turn budget (T)=T12026.0531————
MagicLens-BBackbone=CoCa-B, # Total Params=267M2024.0330.869.3———
MagicLens-LBackbone=CLIP-L, # Total Params=465M2024.0329.668.1———
MagicLens-L2024.1229.6————
MagicLens-LZero-shot=true, Query format=(q_i, q_t) -> c_i2026.0229.6————
MagicLens-LReranker=false2025.0729.6————
MagicLensBackbone=ViT-L/14, Training Costs=128 TPU, 6h2026.0729.6—30.834.433.4
Ours (EIG)Policy=Ours (EIG), Turn budget (T)=T12026.0529.1————
MLLM-QPolicy=MLLM-Q, Turn budget (T)=T22026.0529————
FIXED-QPolicy=FIXED-Q, Turn budget (T)=T32026.0528.1————
MRA-CIRType=Agentic, Training Free=false2026.0227.14—28.8532.6331.54
CIReVLBackbone=OpenCLIP ViT-G/142026.0126.77————
FIXED-QPolicy=FIXED-Q, Turn budget (T)=T22026.0526.7————
MLLM-QPolicy=MLLM-Q, Turn budget (T)=T12026.0526.2————
OSrCIRBackbone=OpenCLIP ViT-G/14, reproduced=true2026.0125.62————
RANDOM-QPolicy=RANDOM-Q, Turn budget (T)=T32026.0525.6————
FIXED-QPolicy=FIXED-Q, Turn budget (T)=T12026.0525————
RANDOM-QPolicy=RANDOM-Q, Turn budget (T)=T22026.0524.9————
E5-V2024.1224.8————
E5-VZero-shot=true, Query format=(q_i, q_t) -> c_i2026.0224.8————
E5-VReranker=false2025.0724.8————
DQE-CIRsetting=zero-shot2026.0324.27—25.6828.1327.47
RANDOM-QPolicy=RANDOM-Q, Turn budget (T)=T12026.0524.1————
AutoCIRType=Agentic, Training Free=true2026.0224.05—25.1428.3627.35
Baseline (ViT-G)Pretraining Data=OACIRR, Pretraining Scale=127 K, Zero-shot=true2026.0423.96—24.69—26.58
RANDOM-QPolicy=RANDOM-Q, Turn budget (T)=T02026.0523.5————
FIXED-QPolicy=FIXED-Q, Turn budget (T)=T02026.0523.5————
MLLM-QPolicy=MLLM-Q, Turn budget (T)=T02026.0523.5————
Ours (EIG)Policy=Ours (EIG), Turn budget (T)=T02026.0523.5————
ORACLE-QPolicy=ORACLE-Q, Turn budget (T)=T02026.0523.5————
LDREType=CIR Dedicated, Training Free=true2026.0223.35—24.0327.526.44
QuResetting=zero-shot2026.0323.22—24.2327.2426.26
MagicLens-BBackbone=CLIP-B, # Total Params=166M2024.0323.166.7———
MagicLensBackbone=ViT-B/32, Training Costs=64 TPU, 6h2026.0723.1—23.826.725.8
OSrCIRSetting=Zero-shot, Turns=m=02026.0622.4—23.6——
CLARASetting=Zero-shot, Turns=m=02026.0622.1—23.9——
CoAlign (ViT-G)Pretraining Data=CIRHS, Pretraining Scale=535 K, Zero-shot=true2026.0421.6—23.38—25.98
Baseline (ViT-L)Pretraining Data=OACIRR, Pretraining Scale=127 K, Zero-shot=true2026.0421.51—23.14—25.47
CoVR-BLIPPretraining Data=WebVid-CoVR, Pretraining Scale=1,644 K, Zero-shot=true2026.0421.43—22.33—24.47
CIReVLBackbone=CLIP-L, # Total Params=12.5B*2024.0318.659.5———
CIReVLType=CIR Dedicated, Training Free=true2026.0218.57—19.0121.820.89
CoVRsetting=zero-shot2026.0318.35—19.2521.8821.02
MCL (LLaMA2-7B)Backbone=ViT-L/14, Training Costs=NA2026.0717.7—18.921.720.8
SPRCsetting=zero-shot2026.0317.57—18.4820.9820.14
Qwen3-Embed-8BType=Text Embedding, Training Free=true2026.0216.45—17.2319.8818.98
CompoDiffPretraining Data=ST18M + LAION-2B, Pretraining Scale=18,000 K, Zero-shot=true2026.0415.33—17.71—19.45
CIReVLBackbone=CLIP-B, # Total Params=12.3B*2024.0314.960.2———
FlowCIRBackbone=ViT-L/14, Training Costs=RTX 3090, 0.5h2026.0714.9—15.718.217.3
Ops-MM-v1-7BType=Multimodal Embedding, Training Free=true2026.0213.56—15.9619.6818.61
Qwen3-Embed-4BType=Text Embedding, Training Free=true2026.0213.55—14.617.216.33
SEARLEBackbone=OpenCLIP ViT-G/142026.0113.2————
FlowCIRBackbone=ViT-B/32, Training Costs=RTX 3090, 0.5h2026.0713.1—13.415.314.6
LinCIR + CIGBackbone=ViT-L/14, Training Costs=NA2026.0713—13.61615.1
CompoDiffBackbone=CLIP-L, # Total Params=568M2024.0312.657.4———
LinCIRBackbone=CLIP-L, # Total Params=442M2024.0312.657.1———
CompoDiffSetting=Zero-shot, Turns=m=02026.0612.6—15.3——
LinCIRBackbone=ViT-L/14, Training Costs=8 A100, 0.5h2026.0712.6—13.615.915
Compo-DiffBackbone=ViT-L/14, Training Costs=128 A100, 231h2026.0712.6—13.416.415.8
LinCIRType=CIR Dedicated, Training Free=false2026.0212.59—13.5815.8515
UniIR-CLIP2024.1212.5————
UniIR-CLIPZero-shot=true, Query format=(q_i, q_t) -> c_i2026.0212.5————
UniIR-CLIP_SFReranker=false2025.0712.5————
SEARLEBackbone=CLIP-L, # Total Params=442M2024.0311.753.8———
SEARLESetting=Zero-shot, Turns=m=02026.0611.7—13.1——
SEARLEBackbone=ViT-L/14, Training Costs=8 A100, 4h2026.0711.7—12.715.114.3
SEARLEType=CIR Dedicated, Training Free=false2026.0211.68—12.7315.1214.33
iSEARLEBackbone=ViT-L/14, Training Costs=122026.0711.3—12.715.314.5
iSEARLEBackbone=ViT-B/32, Training Costs=A100, 12h2026.0710.6—11.213.312.5
CLIP4Cirsetting=zero-shot2026.0310.58—11.1812.9612.32
PLIBackbone=CLIP-L, # Total Params=428M+2024.0310.455.6———
SEARLE + CIGBackbone=ViT-B/32, Training Costs=NA2026.0710.2—10.612.511.8
SEARLE-XL-OTIType=CIR Dedicated, Training Free=false2026.0210.18—11.0313.6712.72
Qwen3-Embed-0.6BType=Text Embedding, Training Free=true2026.029.55—10.3512.3311.61
Pic2WordSetting=Zero-shot, Turns=m=02026.069.5—10.6——
SEARLEBackbone=CLIP-B, # Total Params=165M2024.039.454.9———
SEARLEBackbone=ViT-B/32, Training Costs=8 A100, 4h2026.078.9—9.411.210.6
Pic2WordType=CIR Dedicated, Training Free=false2026.028.72—9.5111.2910.46
Pic2WordBackbone=CLIP-L, # Total Params=429M2024.038.7————
Pic2WordBackbone=ViT-L/14, Training Costs=8 A100, 16h2026.078.7—9.511.310.6