Image Captioning on COCO (test) (B@4, METEOR, CIDEr, SPICE, ROUGE-L, WMD)
50.21BLEU@4Llama3 + GMAIL
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Llama3 + GMAILBase Model=Llama3, Alignment=GMAIL2026.02 | 50.21 | 38.59 | 168.53 | 32.58 | 73.29 | 80.25 | |
| Llama3Base Model=Llama32026.02 | 47.36 | 35.21 | 158.13 | 28.35 | 68.32 | 75.13 | |
| LLaVA + GMAILBase Model=LLaVA, Alignment=GMAIL2026.02 | 43.26 | 34.89 | 146.38 | 27.23 | 65.25 | 71.39 | |
| LLaVABase Model=LLaVA2026.02 | 39.67 | 32.38 | 134.29 | 24.17 | 61.36 | 65.78 | |
| IFCap + GMAILBase Model=IFCap, Alignment=GMAIL2026.02 | 39.32 | 32.07 | 127.86 | 23.98 | 59.83 | 63.51 | |
| ClipCap + GMAILBase Model=ClipCap, Alignment=GMAIL2026.02 | 38.12 | 31.67 | 119.53 | 23.75 | 56.27 | 62.16 | |
| IFCapBase Model=IFCap2026.02 | 33.25 | 28.6 | 115.27 | 21.58 | 51.35 | 56.72 | |
| ClipCapBase Model=ClipCap2026.02 | 32.15 | 27.1 | 108.35 | 20.12 | — | — | |
| Qwen3-VL-Instruct RGBPScale=8B2026.04 | 5.64 | 13.64 | 4.82 | 9.4 | — | — | |
| Qwen3-VL-InstructScale=8B2026.04 | 5.32 | 17.32 | 3.68 | 12.26 | — | — | |
| Qwen3-VL-InstructScale=4B2026.04 | 5.06 | 16.6 | 3.44 | 9.08 | — | — | |
| Qwen3-VL-InstructScale=2B2026.04 | 4.48 | 15.42 | 2.36 | 9.32 | — | — | |
| ProCapBackbone=OPT-2.7B2026.04 | 4.1 | 14.46 | 5.8 | 9.44 | — | — | |
| ProCapBackbone=TinyLlama-1.1B2026.04 | 2.48 | 11.38 | 5.02 | 5.53 | — | — | |
| FastVLMScale=0.5B2026.04 | 2.16 | 10.98 | 1.44 | 5.42 | — | — | |
| FastVLMScale=7B2026.04 | 2.1 | 11.64 | 1.24 | 6 | — | — | |
| ProCapBackbone=Vicuna-1.5-7B2026.04 | 1.9 | 11.66 | 2.94 | 7.5 | — | — | |
| FastVLMScale=1.5B2026.04 | 1.88 | 10.9 | 0.98 | 5.36 | — | — | |
| ProCapBackbone=OpenLLaMA-3B2026.04 | 1.5 | 11.73 | 3.78 | 7.77 | — | — |