Image Captioning on nocaps (val)
126.9CIDEr (Overall)GIT2
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| GIT22022.05 | 126.9 | 126.9 | 16.1 | 130.6 | 15.8 | 16.1 | 125.8 | 16.2 | — | — | |
| GIT2022.05 | 125.5 | 129.8 | 16.3 | 127.1 | 15.7 | 16 | 124.1 | 16 | — | — | |
| CoCa2022.05 | 122.4 | — | — | — | — | 15.5 | — | — | — | — | |
| UniversalCap2022.05 | 122.1 | 123.2 | 15 | 123.4 | 14.4 | 15 | 121.5 | 15.3 | — | — | |
| LEMONhugePre-training data=ALT200M2021.11 | 117.3 | 118 | 15.4 | 120.2 | 14.5 | 15 | 116.3 | 15.1 | — | — | |
| LEMON2022.05 | 117.3 | 118 | 15.4 | 120.2 | 14.5 | 15 | 116.3 | 15.1 | — | — | |
| SimVLM2022.05 | 115.2 | 113.7 | — | 115.2 | — | — | 110.9 | — | — | — | |
| mPLUG2022.05 | 114.8 | — | — | — | — | 14.8 | — | — | — | — | |
| LEMONlargePre-train #Images=200M2022.01 | 113.4 | 116.9 | 15.8 | 111.3 | 14 | 15 | 113.3 | 15.1 | — | — | |
| LEMONlargePre-training data=ALT200M2021.11 | 113.4 | 116.9 | 15.8 | 111.3 | 14 | 15 | 113.3 | 15.1 | — | — | |
| BLIPViT-LPre-train #Images=129M2022.01 | 113.2 | 114.9 | 15.2 | 115.3 | 14.4 | 14.8 | 112.1 | 14.9 | — | — | |
| SimVLMhugePre-train #Images=1.8B2022.01 | 112.2 | 113.7 | — | 115.2 | — | — | 110.9 | — | — | — | |
| SimVLMhugePre-training data=1.8B2021.11 | 112.2 | 113.7 | — | 115.2 | — | — | 110.9 | — | — | — | |
| BLIPCapFilt-LPre-train #Images=129M2022.01 | 109.6 | 111.8 | 14.9 | 111.5 | 14.2 | 14.7 | 108.6 | 14.8 | — | — | |
| Qwen3-VL-Instruct RGBPScale=8B2026.04 | 108.66 | 104.5 | 13.54 | 103.94 | 13.3 | 13.76 | 113.76 | 14.36 | — | — | |
| GITLModel Variant=Large2022.05 | 106.9 | 107.7 | 14.9 | 102.5 | 13.7 | 14.4 | 107.8 | 14.5 | — | — | |
| LEMONbasePre-train #Images=200M2022.01 | 106.8 | 107.7 | 14.7 | 107.9 | 13.1 | 14.1 | 106.2 | 14.3 | — | — | |
| LEMONbasePre-training data=ALT200M2021.11 | 106.8 | 107.7 | 14.7 | 107.9 | 13.1 | 14.1 | 106.2 | 14.3 | — | — | |
| BLIPPre-train #Images=129M2022.01 | 106.3 | 109.1 | 14.8 | 105.7 | 13.7 | 14.3 | 105.8 | 14.4 | — | — | |
| BLIPPre-train #Images=14M2022.01 | 105.1 | 111.3 | 15.1 | 102.4 | 13.7 | 14.4 | 104.5 | 14.4 | — | — | |
| LEMONlargePre-training data=CC12M2021.11 | 101.8 | 103.6 | 14.4 | 102.7 | 12.6 | 13.6 | 101.1 | 13.8 | — | — | |
| LEMONbasePre-train #Images=12M2022.01 | 100.4 | 104.5 | 14.6 | 96.7 | 12.4 | 13.8 | 100.7 | 14 | — | — | |
| LEMONbasePre-training data=CC12M2021.11 | 100.4 | 104.5 | 14.6 | 96.7 | 12.4 | 13.8 | 100.7 | 14 | — | — | |
| VinVLlargePre-training data=5.65M combined2021.11 | 97.3 | 106.3 | 14.5 | 88.8 | 12.6 | 13.8 | 98 | 14 | — | — | |
| P2CSCST optimization=true, Constrained Beam Search=true2022.09 | 97.2 | 101.4 | 15.1 | 95.4 | 12.9 | 14.2 | 96.8 | 14.5 | — | — | |
| GITBModel Variant=Base2022.05 | 96.6 | 100.7 | 13.8 | 89.6 | 12.5 | 13.4 | 97.7 | 13.5 | — | — | |
| VinVLPre-train #Images=5.7M2022.01 | 95.5 | 103.1 | 14.2 | 88.3 | 12.1 | 13.5 | 96.1 | 13.8 | — | — | |
| VinVLbase+Pre-training data=5.65M combined2021.11 | 95.5 | 103.1 | 14.2 | 88.3 | 12.1 | 13.5 | 96.1 | 13.8 | — | — | |
| VinVL+VIVOSCST=true2021.01 | 94.3 | 103.7 | 13.7 | 83.8 | 11.9 | 13.1 | 95.6 | 13.4 | — | — | |
| VinVL2022.05 | 94.3 | 103.7 | 13.7 | 83.8 | 11.9 | 13.1 | 95.6 | 13.4 | — | — | |
| UFO2022.05 | 94.3 | 103.9 | 14.5 | 83.5 | 12.3 | 13.6 | 95.5 | 13.8 | — | — | |
| ViTCAPCBS decoding=true2021.12 | 93.8 | 98.7 | 13.3 | 95.4 | 12.7 | 13 | 92.3 | 13.3 | — | — | |
| LEMONbasePre-training data=CC3M2021.11 | 91.6 | 96 | 13.8 | 88.1 | 11.8 | 13 | 91.7 | 13.2 | — | — | |
| VinVL + VIVOVIVO pre-training=true, SCST optimization=true, Constrained Beam Search=true2022.09 | 91.4 | 94.8 | 13.3 | 88.7 | 11.6 | 12.7 | 91.4 | 13 | — | — | |
| VinVLSCST=true, CBS=true2021.01 | 90.9 | 96.8 | 13.5 | 87.4 | 11.6 | 12.8 | 90.7 | 13.1 | — | — | |
| VinVLbase*Pre-training data=N/A2021.11 | 90.9 | 96.8 | 13.5 | 87.4 | 11.6 | 12.8 | 90.7 | 13.1 | — | — | |
| VinVLSCST optimization=true, Constrained Beam Search=true2022.09 | 90.9 | 96.8 | 13.5 | 87.4 | 11.6 | 12.8 | 90.7 | 13.1 | — | — | |
| Enc-DecPre-train #Images=15M2022.01 | 90.2 | 92.6 | 12.5 | 94.5 | 11.9 | 12.1 | 88.3 | 12.1 | — | — | |
| Encoder-DecoderPre-training data=CC3M+CC12M2021.11 | 90.2 | 92.6 | 12.5 | 94.5 | 11.9 | 12.1 | 88.3 | 12.1 | — | — | |
| GRITV.E. Type=R+G2022.07 | 90.2 | 105.9 | 13.6 | 72.6 | 11.1 | 12.8 | — | — | — | — | |
| GRITV.E Type=R+G2022.07 | 90.2 | 105.9 | 13.6 | 72.6 | 11.1 | 12.8 | 92.16 | 13.05 | — | — | |
| GRITV.E Type=R+G2022.07 | 90.2 | 105.9 | 13.6 | 72.6 | 11.1 | 12.8 | — | — | — | — | |
| ViTCAPCBS decoding=false2021.12 | 89.2 | 99.3 | 13.2 | 78.1 | 11.9 | 12.7 | 90.4 | 12.9 | — | — | |
| VIVOSCST=true, CBS=true2021.01 | 88.3 | 92.2 | 12.9 | 87.5 | 11.5 | 12.4 | 87.8 | 12.6 | — | — | |
| VIVOCBS decoding=true2021.12 | 88.3 | 92.2 | 12.9 | 87.5 | 11.5 | 12.4 | 87.8 | 12.6 | — | — | |
| Oscar + VIVOModel Scale=Base, VIVO pre-training=true, SCST optimization=true, Constrained Beam Search=true2022.09 | 88.3 | 92.2 | 12.9 | 87.5 | 11.5 | 12.4 | 87.8 | 12.6 | — | — | |
| VIVO2022.05 | 88.3 | 92.2 | 12.9 | 87.5 | 11.5 | 12.4 | 87.8 | 12.6 | — | — | |
| SMALLCAPVariant=w+H2022.09 | 87.5 | 90.5 | — | 91.5 | — | — | 85.6 | — | — | — | |
| Encoder-DecoderPre-training data=CC12M2021.11 | 87.4 | 88.3 | 12.3 | 91.3 | 11.2 | 11.8 | 86 | 11.8 | — | — | |
| Human2021.01 | 87.1 | 84.4 | 14.3 | 95.7 | 14 | 14.2 | 85 | 14.3 | — | — | |
| Human2021.12 | 87.1 | 84.4 | 14.3 | 95.7 | 14 | 14.2 | 85 | 14.3 | — | — | |
| Human2022.09 | 87.1 | 84.8 | 14.3 | 95.7 | 14 | 14.2 | 85 | 14.3 | — | — | |
| Human2022.05 | 87.1 | 84.4 | 14.3 | 95.7 | 14 | 14.2 | 85 | 14.3 | — | — | |
| VIVOCBS decoding=false2021.12 | 85.3 | 90.4 | 13 | 83 | 10.7 | 12.2 | 84.9 | 12.5 | — | — | |
| OSCAR_Lsize=large, SCST=true, CBS=true2021.01 | 83.4 | 85.4 | 11.9 | 80.3 | 10 | 11.4 | 84 | 11.7 | — | — | |
| OscarModel Scale=Large, SCST optimization=true, Constrained Beam Search=true2022.09 | 83.4 | 85.4 | 11.9 | 80.3 | 10 | 11.4 | 84 | 11.7 | — | — | |
| OSCAR2022.05 | 83.4 | 85.4 | 11.9 | 80.3 | 10 | 11.4 | 84 | 11.7 | — | — | |
| OSCAR_Bsize=base, SCST=true, CBS=true2021.01 | 81.1 | 83.4 | 12 | 77.6 | 10.6 | 11.7 | 81.6 | 12 | — | — | |
| OSCARCBS decoding=true2021.12 | 81.1 | 83.4 | 12 | 77.6 | 10.6 | 11.7 | 81.6 | 12 | — | — | |
| OscarModel Scale=Base, SCST optimization=true, Constrained Beam Search=true2022.09 | 81.1 | 83.4 | 12 | 77.6 | 10.6 | 11.7 | 81.6 | 12 | — | — | |
| LEMONbasePre-training data=N/A2021.11 | 79 | 91.4 | 13.3 | 62.6 | 10.6 | 12.3 | 81.4 | 12.5 | — | — | |
| OSCARModel size=Large2022.09 | 78.6 | 78.8 | — | 77.4 | — | — | 78.9 | — | — | — | |
| I-TuningModel size=Large2022.09 | 78.5 | 89.6 | — | 64.8 | — | — | 80.4 | — | — | — | |
| SMALLCAP2022.09 | 77.9 | 87.6 | — | 68.9 | — | — | 78.6 | — | — | — | |
| I-TuningModel size=Medium2022.09 | 75.4 | 89.6 | — | 58.8 | — | — | 77.4 | — | — | — | |
| M² TransformerCBS=true2019.12 | 75 | 81.2 | 12 | 69.4 | 10 | 11.4 | — | — | — | — | |
| UpDown+ELMo=true, CBS=true2021.01 | 74.3 | 79.3 | 12.4 | 71.7 | 9.9 | 11.2 | 73.8 | 11.4 | — | — | |
| UpDownSCST optimization=false, Constrained Beam Search=true2022.09 | 74.3 | 79.3 | 12.4 | 71.7 | 9.9 | 11.2 | 73.8 | 11.4 | — | — | |
| Encoder-DecoderPre-training data=CC3M2021.11 | 73.2 | 81.8 | 11.6 | 65.3 | 10.1 | 11 | 73.7 | 11.1 | — | — | |
| Up-DownCBS=true2019.12 | 73.1 | 80 | 12 | 66.4 | 9.7 | 11.1 | — | — | — | — | |
| UpDownCBS decoding=true2021.12 | 73.1 | 80 | 12 | 66.4 | 9.7 | 11.1 | 73.6 | 11.3 | — | — | |
| UpDownCBS decoding=true, Language Model=ELMO2021.12 | 73.1 | 80 | 12 | 66.4 | 9.7 | 11.1 | 73.6 | 11.3 | — | — | |
| ExpansionNet v2loss=Cross-Entropy Loss, beam size=3, pre-training data=none2022.08 | 72.9 | 83.8 | 12.6 | 54 | 9.3 | 11.4 | 79.2 | 12.4 | — | — | |
| TransformerCBS=true2019.12 | 66.9 | 74.3 | 11 | 62.5 | 9.2 | 10.3 | — | — | — | — | |
| ClipCap2022.09 | 65.8 | 84.9 | — | 49.1 | — | — | 66.8 | — | — | — | |
| M² TransformerCBS=false2019.12 | 64.5 | 85.7 | 12.1 | 38.9 | 8.9 | 11.1 | — | — | — | — | |
| M²Trans.V.E. Type=R2022.07 | 64.5 | 85.7 | 12.1 | 38.9 | 8.9 | 11.1 | — | — | — | — | |
| M2TransformerV.E Type=R2022.07 | 64.5 | 85.7 | 12.1 | 38.9 | 8.9 | 11.1 | — | — | — | — | |
| M²Trans.V.E Type=R2022.07 | 64.5 | 85.7 | 12.1 | 38.9 | 8.9 | 11.1 | — | — | — | — | |
| OSCARCBS decoding=false2021.12 | 63.8 | 79.6 | 12.3 | 45.3 | 9.7 | 11.2 | 66.1 | 11.5 | — | — | |
| NBTCBS=true2019.12 | 60.2 | 62.1 | 10.1 | 62.4 | 8.9 | 9.5 | — | — | — | — | |
| UpDownCBS decoding=false2021.12 | 55.3 | 78.1 | 11.6 | 31.3 | 8.3 | 10.1 | 57.7 | 10.3 | — | — | |
| Up-downV.E. Type=R2022.07 | 55.3 | 78.1 | 11.6 | 31.3 | 8.3 | 10.1 | — | — | — | — | |
| Up-downV.E Type=R2022.07 | 55.3 | 78.1 | 11.6 | 31.3 | 8.3 | 10.1 | 57.7 | 10.3 | — | — | |
| Up-downV.E Type=R2022.07 | 55.3 | 78.1 | 11.6 | 31.3 | 8.3 | 10.1 | — | — | — | — | |
| Up-Down2022.08 | 55.3 | 78.1 | 11.6 | 31.3 | 8.3 | 10.1 | 57.7 | 10.3 | — | — | |
| TransformerCBS=false2019.12 | 54.7 | 78 | 11 | 29.7 | 7.8 | 9.8 | — | — | — | — | |
| Trans.V.E. Type=R2022.07 | 54.7 | 78 | 11 | 29.7 | 7.8 | 9.8 | — | — | — | — | |
| TransformerV.E Type=R2022.07 | 54.7 | 78 | 11 | 29.7 | 7.8 | 9.8 | — | — | — | — | |
| Trans.V.E Type=R2022.07 | 54.7 | 78 | 11 | 29.7 | 7.8 | 9.8 | — | — | — | — | |
| Enc-Dec2022.08 | 54.7 | 72.8 | 11.1 | 34.1 | 8.3 | 10 | 57.1 | 10.2 | — | — | |
| NBTV.E. Type=R2022.07 | 53.9 | 62.7 | 10.1 | 54 | 8.6 | 9.2 | — | — | — | — | |
| NBTV.E Type=R2022.07 | 53.9 | 62.7 | 10.1 | 54 | 8.6 | 9.2 | 51.9 | 9.2 | — | — | |
| NBTV.E Type=R2022.07 | 53.9 | 62.7 | 10.1 | 54 | 8.6 | 9.2 | — | — | — | — | |
| ProCapBackbone=OpenLLaMA-3B2026.04 | 44.72 | 58.22 | 8.66 | 27.1 | 5.2 | 7.06 | 48 | 7.34 | — | — | |
| ProCapBackbone=Vicuna-1.5-7B2026.04 | 43.94 | 58.7 | 8.84 | 27.32 | 5.3 | 7.1 | 43.46 | 7.18 | — | — | |
| ProCapBackbone=OPT-2.7B2026.04 | 34.18 | 51.7 | 7.48 | 16.5 | 3.82 | 5.68 | 32.22 | 5.78 | — | — | |
| ProCapBackbone=TinyLlama-1.1B2026.04 | 32.36 | 42.32 | 6.1 | 19.7 | 3.96 | 5.1 | 33.64 | 5.26 | — | — | |
| Qwen3-VL-InstructScale=2B2026.04 | 19.6 | 17.96 | 13.34 | 19.16 | 12.04 | 12.84 | 21.5 | 13.24 | — | — | |
| Qwen3-VL-InstructScale=8B2026.04 | 19.2 | 18.42 | 12.3 | 18.06 | 11.64 | 12.34 | 21.46 | 13.04 | — | — |