Image Captioning on COCO (test) (L-infinity Metrics)
120.4Clean Metric ValueDiffCAP
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DiffCAPBackbone=LLaVA 1.5-7B2025.06 | 120.4 | 119.6 | 116.9 | |
| No defenseBackbone=LLaVA 1.5-7B2025.06 | 115.5 | 4 | 3.1 | |
| CLIPureBackbone=LLaVA 1.5-7B2025.06 | 115.1 | 4.9 | 3.4 | |
| JPEG-DLBackbone=LLaVA 1.5-7B2025.06 | 113.3 | 106.4 | 77.2 | |
| FARE2Backbone=LLaVA 1.5-7B, fine-tuning perturbation=ℓ∞_2/2552025.06 | 109.9 | 53.6 | 31 | |
| DiffPureBackbone=LLaVA 1.5-7B2025.06 | 106.5 | 108.4 | 105 | |
| FARE4Backbone=LLaVA 1.5-7B, fine-tuning perturbation=ℓ∞_4/2552025.06 | 102.4 | 57.1 | 40.9 | |
| TeCoA2Backbone=LLaVA 1.5-7B, fine-tuning perturbation=ℓ∞_2/2552025.06 | 98.4 | 44.2 | 30.3 | |
| TeCoA4Backbone=LLaVA 1.5-7B, fine-tuning perturbation=ℓ∞_4/2552025.06 | 88.3 | 50.9 | 35.3 | |
| DiffCAPBackbone=OF-9B2025.06 | 81.4 | 79.3 | 78.4 | |
| CLIPureBackbone=OF-9B2025.06 | 80.8 | 6.6 | 5.3 | |
| No defenseBackbone=OF-9B2025.06 | 79.7 | 1.5 | 1.1 | |
| FARE2Backbone=OF-9B, fine-tuning perturbation=ℓ∞_2/2552025.06 | 79.1 | 34.2 | 19.5 | |
| JPEG-DLBackbone=OF-9B2025.06 | 78.2 | 66.1 | 43.9 | |
| DiffPureBackbone=OF-9B2025.06 | 74.9 | 73.4 | 72.3 | |
| FARE4Backbone=OF-9B, fine-tuning perturbation=ℓ∞_4/2552025.06 | 74.1 | 30.9 | 22.8 | |
| TeCoA2Backbone=OF-9B, fine-tuning perturbation=ℓ∞_2/2552025.06 | 73.5 | 31.6 | 21.2 | |
| TeCoA4Backbone=OF-9B, fine-tuning perturbation=ℓ∞_4/2552025.06 | 66.9 | 28.5 | 21.6 |