Image Classification on 10 Fine-Grained Datasets Generalization Suite
75.39Flower102 AccuracyPromptKD + MERGETUNE
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| PromptKD + MERGETUNEBackbone=ViT-B/16, Merging Strategy=MERGETUNE2026.01 | 75.39 | 55.74 | 92.01 | 74.17 | 77.12 | 93.95 | 89.32 | 68.61 | 26.58 | 63.94 | 71.68 | |
| PromptKDBackbone=ViT-B/16, Training=16-shot ImageNet, External Knowledge=true2026.01 | 75.33 | 55.08 | 91.59 | 73.93 | 76.39 | 93.61 | 88.84 | 68.57 | 26.24 | 63.74 | 71.33 | |
| PromptKD + DAREBackbone=ViT-B/16, Merging Strategy=DARE2026.01 | 74.26 | 54.26 | 90.36 | 73.26 | 75.19 | 93.26 | 88.36 | 67.47 | 25.14 | 63.03 | 70.46 | |
| POMPBackbone=ViT/B-162023.04 | 72.4 | 46.2 | 89.5 | 66.8 | 68.5 | 95 | 86.3 | 67.7 | 25.6 | 52.1 | 67 | |
| MaPLeBackbone=ViT/B-162023.04 | 72.2 | 46.5 | 90.5 | 65.6 | 68.7 | 93.5 | 86.2 | 67 | 24.7 | 48.1 | 66.3 | |
| CoCoOpBackbone=ViT/B-162023.04 | 71.9 | 45.7 | 90.1 | 65.3 | 68.2 | 94.4 | 86.1 | 67.4 | 22.9 | 45.4 | 65.7 | |
| CoCoOpBackbone=ViT-B/16, Training protocol=16-shot ImageNet tuning2022.09 | 71.59 | 45.48 | 90.2 | 65.17 | 68.77 | 94.15 | 84.83 | 67.07 | 22.95 | 42.13 | 65.23 | |
| MMA + MERGETUNEBackbone=ViT-B/16, Merging Strategy=MERGETUNE2026.01 | 71.46 | 46.22 | 90.43 | 66.11 | 69.23 | 94.28 | 86.26 | 67.59 | 25.26 | 46.16 | 66.3 | |
| Bayesian Prompt Learning2022.10 | 70.9 | 46.1 | 90.63 | 65 | 68.67 | 93.67 | 86.3 | 67.47 | 24.93 | 45.87 | 65.95 | |
| VPTBackbone=ViT/B-162023.04 | 70.9 | 46.1 | 90.6 | 65 | 68.7 | 93.7 | 86.3 | 67.5 | 24.9 | 45.9 | 66 | |
| KgCoOp + MERGETUNEBackbone=ViT-B/16, Merging Strategy=MERGETUNE2026.01 | 70.85 | 47.46 | 90.41 | 66.15 | 70.42 | 94.24 | 86.57 | 66.63 | 24.12 | 48.41 | 66.53 | |
| CLIPBackbone=ViT-B/16, Training=Zero-shot2026.01 | 70.7 | 44.3 | 89.1 | 65.7 | 67.6 | 93.3 | 85.9 | 62.6 | 24.9 | 48.3 | 65.24 | |
| LASPBackbone=ViT/B-162023.04 | 70.5 | 45.5 | 89.4 | 64.8 | 68.2 | 94.5 | 86.3 | 67 | 23 | 48.3 | 65.8 | |
| CoOp + MERGETUNEBackbone=ViT-B/16, Merging Strategy=MERGETUNE2026.01 | 70.4 | 46.53 | 89.97 | 65.67 | 68.71 | 93.96 | 86.43 | 66.35 | 23.88 | 46.12 | 65.8 | |
| CoCoOp2022.10 | 70.2 | 45.73 | 90.14 | 65.5 | 68.21 | 94.43 | 86.06 | 67.36 | 22.94 | 45.37 | 65.59 | |
| KgCoOpBackbone=ViT-B/16, Training=16-shot ImageNet2026.01 | 70.01 | 46.35 | 89.83 | 65.41 | 68.5 | 93.92 | 86.36 | 66.16 | 22.51 | 46.04 | 65.51 | |
| MMABackbone=ViT-B/16, Training=16-shot ImageNet, Reproduction=true2026.01 | 69.86 | 45.26 | 89.65 | 65.41 | 68.82 | 93.8 | 85.79 | 67.31 | 24.47 | 43.7 | 65.41 | |
| PromptKD + TIESBackbone=ViT-B/16, Merging Strategy=TIES2026.01 | 69.36 | 50.36 | 89.25 | 70.35 | 72.54 | 91.35 | 86.26 | 64.75 | 23.74 | 60.68 | 67.86 | |
| KgCoOp + DAREBackbone=ViT-B/16, Merging Strategy=DARE2026.01 | 69.29 | 45.09 | 88.15 | 64.81 | 66.06 | 93.66 | 86.16 | 64.64 | 22.24 | 45.38 | 64.55 | |
| TPTEnsemble=false, Training-free=false2024.05 | 68.86 | 46.99 | 87.22 | 66.36 | 68 | 94.12 | 84.63 | 65.41 | 23.1 | 42.93 | 64.76 | |
| TPTBackbone=ViT-B/16, Test-time adaptation strategy=Prompt Tuning2022.09 | 68.79 | 46.79 | 87.09 | 66.38 | 67.86 | 94.13 | 84.67 | 65.41 | 23.44 | 42.78 | 64.73 | |
| CoOp2022.10 | 68.71 | 41.92 | 89.14 | 64.51 | 66.55 | 93.7 | 85.3 | 64.15 | 18.47 | 46.39 | 63.88 | |
| CoOpBackbone=ViT-B/16, Training=16-shot ImageNet2026.01 | 68.71 | 41.92 | 89.14 | 64.51 | 66.55 | 93.7 | 85.3 | 64.15 | 18.47 | 46.39 | 63.88 | |
| CoOpBackbone=ViT/B-162023.04 | 68.7 | 41.9 | 89.1 | 64.5 | 66.6 | 93.7 | 85.3 | 64.2 | 18.5 | 46.4 | 63.9 | |
| MTAEnsemble=false, Training-free=true2024.05 | 68.26 | 45.59 | 88.22 | 68.05 | 68.11 | 94.13 | 84.95 | 64.98 | 25.32 | 38.71 | 64.63 | |
| CoOpBackbone=ViT-B/16, Training protocol=16-shot ImageNet tuning2022.09 | 68.25 | 42.34 | 89.38 | 63.35 | 67.17 | 92.82 | 83.74 | 64.51 | 19.99 | 40.22 | 63.18 | |
| MTA + E.Ensemble=true, Training-free=true2024.05 | 68.06 | 45.9 | 88.24 | 68.47 | 68.69 | 94.21 | 85 | 66.67 | 25.2 | 45.36 | 65.58 | |
| CLIP + E.Ensemble=true, Training-free=true2024.05 | 67.8 | 46.04 | 88.42 | 66.14 | 66.77 | 93.87 | 83.83 | 66.02 | 23.88 | 48.8 | 65.16 | |
| CLIPBackbone=ViT-B/16, Training protocol=zero-shot2022.09 | 67.44 | 44.27 | 88.25 | 65.48 | 65.13 | 93.35 | 83.65 | 62.59 | 23.67 | 42.01 | 63.58 | |
| CLIPEnsemble=false, Training-free=true2024.05 | 67.44 | 44.27 | 88.25 | 65.48 | 65.13 | 93.35 | 83.65 | 62.59 | 23.67 | 42.01 | 63.58 | |
| hard promptBackbone=ViT/B-162023.04 | 67.4 | 44.3 | 88.2 | 65.6 | 65.1 | 93.3 | 85.3 | 62.6 | 23.7 | 42 | 63.7 | |
| CoOp + TIESBackbone=ViT-B/16, Merging Strategy=TIES2026.01 | 67.4 | 44.39 | 88.25 | 65.51 | 65.24 | 93.31 | 85.23 | 62.56 | 23.91 | 42.22 | 63.8 | |
| KgCoOp + TIESBackbone=ViT-B/16, Merging Strategy=TIES2026.01 | 66.92 | 41.75 | 87.56 | 64.33 | 65.74 | 91.94 | 85.53 | 63.66 | 21.69 | 46.36 | 63.55 | |
| MMA + DAREBackbone=ViT-B/16, Merging Strategy=DARE2026.01 | 66.45 | 40.44 | 87.14 | 63.25 | 65.35 | 91.24 | 84.43 | 62.12 | 19.46 | 42.89 | 62.28 | |
| CoOp + DAREBackbone=ViT-B/16, Merging Strategy=DARE2026.01 | 66.32 | 38.55 | 86.4 | 62.87 | 64.68 | 89.67 | 84.27 | 60.27 | 18.07 | 45.65 | 61.67 | |
| MMA + TIESBackbone=ViT-B/16, Merging Strategy=TIES2026.01 | 65.35 | 38.35 | 86.21 | 62.46 | 64.13 | 90.54 | 84.21 | 60.33 | 17.98 | 41.69 | 61.13 | |
| TPTBackbone=RN50, Test-time adaptation strategy=Prompt Tuning2022.09 | 65.11 | 41.43 | 84.42 | 58.53 | 60.64 | 87.23 | 75.02 | 61.46 | 17.6 | 28.46 | 57.76 | |
| CoCoOpBackbone=RN50, Training protocol=16-shot ImageNet tuning2022.09 | 62.8 | 39.14 | 87.83 | 56.4 | 58.57 | 86.95 | 76.18 | 60.62 | 15.13 | 28.79 | 57.47 | |
| CLIPBackbone=RN50, Training protocol=zero-shot2022.09 | 61.75 | 40.37 | 83.57 | 55.7 | 58.84 | 85.88 | 73.97 | 58.8 | 15.66 | 23.69 | 53.82 | |
| CoOpBackbone=RN50, Training protocol=16-shot ImageNet tuning2022.09 | 61.62 | 37.77 | 87.24 | 55.72 | 59.89 | 87.23 | 75.86 | 59.28 | 15.2 | 25.43 | 56.52 |