List your company

Image Captioning on Flickr30k (BLEU-4, METEOR, CIDEr, SPICE)

84.5CIDErInstructBLIP
-1.320.97543.2565.525Oct 16, 2021Jul 8, 2022Mar 30, 2023Dec 20, 2023Sep 10, 2024Jun 2, 2025Feb 23, 2026

Evaluation results

MethodLinks
InstructBLIPzero-shot=true, backbone=FlanT5-XL2023.0584.5———
InstructBLIPzero-shot=true, backbone=FlanT5-XXL2023.0583.5———
InstructBLIPzero-shot=true, backbone=Vicuna-13B2023.0582.8———
InstructBLIPzero-shot=true, backbone=Vicuna-7B2023.0582.4———
CleanModel=LLaVa1.5-7B2026.0277.5———
CleanModel=LLaVa1.5-13B2026.0277.1———
BLIP-2zero-shot=true, backbone=FlanT5-XL2023.0576.1———
BLIP-2zero-shot=true, backbone=Vicuna-7B2023.0574.9———
BLIP-2zero-shot=true, backbone=FlanT5-XXL2023.0573.7———
Uni-PerceiverTuning protocol=fine-tuning, Percentage of data used in tuning=100%2021.1272.730.124.518.2
BLIP-2zero-shot=true, backbone=Vicuna-13B2023.0571.6———
Unified VLP2021.1267.430.12317
Flamingozero-shot=true, backbone=80B2023.0567.2———
Uni-PerceiverTuning protocol=prompt-tuning, Percentage of data used in tuning=10%2021.1261.728.822.116.8
Flamingozero-shot=true, backbone=9B2023.0561.5———
Flamingozero-shot=true, backbone=3B2023.0560.6———
CleanModel=OF-9B2026.0260.1———
Uni-PerceiverTuning protocol=prompt-tuning, Percentage of data used in tuning=1%2021.1259.128.121.615.6
Uni-PerceiverTuning protocol=fine-tuning, Percentage of data used in tuning=10%2021.1252.126.921.614.5
Uni-PerceiverTuning protocol=fine-tuning, Percentage of data used in tuning=1%2021.1245.318.919.714.3
MIX.AttackModel=LLaVa1.5-7B, Perturbation (ε)=2/2552026.0242.1———
Uni-PerceiverTuning protocol=none2021.1241.21716.211.2
MIX.AttackModel=LLaVa1.5-13B, Perturbation (ε)=2/2552026.0241———
VT-AttackModel=LLaVa1.5-13B, Perturbation (ε)=2/2552026.0240.7———
VT-AttackModel=LLaVa1.5-7B, Perturbation (ε)=2/2552026.0240———
VT-AttackModel=OF-9B, Perturbation (ε)=2/2552026.0238.6———
FEWVLMlargeModel size=740M, zero-shot=true2021.1036.5——10.7
MIX.AttackModel=LLaVa1.5-7B, Perturbation (ε)=4/2552026.0235.9———
MIX.AttackModel=OF-9B, Perturbation (ε)=2/2552026.0233.7———
MIX.AttackModel=LLaVa1.5-13B, Perturbation (ε)=4/2552026.0232.2———
FEWVLMbaseModel size=224M, zero-shot=true2021.1031——10
AttackVLM-iiModel=LLaVa1.5-7B, Perturbation (ε)=2/2552026.0230.2———
AttackVLM-iiModel=LLaVa1.5-13B, Perturbation (ε)=2/2552026.0227.5———
VT-AttackModel=LLaVa1.5-13B, Perturbation (ε)=4/2552026.0225.2———
Unified VLPModel size=122M, zero-shot=true2021.1024.9——7.2
VT-AttackModel=OF-9B, Perturbation (ε)=4/2552026.0223———
AttackVLM-iiModel=LLaVa1.5-7B, Perturbation (ε)=4/2552026.0220.3———
AttackVLM-iiModel=OF-9B, Perturbation (ε)=2/2552026.0220.1———
VT-AttackModel=LLaVa1.5-7B, Perturbation (ε)=4/2552026.0219.9———
MIX.AttackModel=OF-9B, Perturbation (ε)=4/2552026.0218———
AttackVLM-iiModel=LLaVa1.5-13B, Perturbation (ε)=4/2552026.0217.9———
VEAttackModel=LLaVa1.5-7B, Perturbation (ε)=2/2552026.0210.7———
AttackVLM-iiModel=OF-9B, Perturbation (ε)=4/2552026.029.3———
VEAttackModel=LLaVa1.5-13B, Perturbation (ε)=2/2552026.029.1———
VEAttackModel=OF-9B, Perturbation (ε)=2/2552026.028.7———
VEAttackModel=LLaVa1.5-7B, Perturbation (ε)=4/2552026.026.3———
VEAttackModel=LLaVa1.5-13B, Perturbation (ε)=4/2552026.025.7———
PA-AttackModel=OF-9B, Perturbation (ε)=2/2552026.025.4———
PA-AttackModel=LLaVa1.5-7B, Perturbation (ε)=2/2552026.024.7———
PA-AttackModel=LLaVa1.5-13B, Perturbation (ε)=2/2552026.024.1———
PA-AttackModel=OF-9B, Perturbation (ε)=4/2552026.023.7———
PA-AttackModel=LLaVa1.5-7B, Perturbation (ε)=4/2552026.023.3———
VEAttackModel=OF-9B, Perturbation (ε)=4/2552026.023.2———
VL-T5no-vqaModel size=224M, zero-shot=true2021.102.6——2
PA-AttackModel=LLaVa1.5-13B, Perturbation (ε)=4/2552026.022———