Concept Removal on Ring-A-Bell Violence
88Attack Success Rate (ASR)Baseline + Semantic Surgery
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Baseline + Semantic SurgeryBase Model=FLUX, Sampling Steps=82026.06 | 88 | 84 | |
| SGFBase Model=FLUX, Sampling Steps=82026.06 | 87.6 | 84.1 | |
| SGF + Semantic SurgeryBase Model=FLUX, Sampling Steps=82026.06 | 87.6 | 84.1 | |
| Baseline + SAFREEBase Model=FLUX, Sampling Steps=82026.06 | 87.2 | 83.7 | |
| BaselineBase Model=FLUX, Sampling Steps=82026.06 | 82.4 | 82 | |
| SGF + SAFREEBase Model=FLUX, Sampling Steps=82026.06 | 81.6 | 79.6 | |
| Baseline + SAFREEBase Model=MeanFlow, Sampling Steps=42026.06 | 81.6 | 80.5 | |
| BaselineBase Model=MeanFlow, Sampling Steps=42026.06 | 80.8 | 80.6 | |
| Baseline + Semantic SurgeryBase Model=MeanFlow, Sampling Steps=42026.06 | 80.8 | 80.6 | |
| SGF + Semantic SurgeryBase Model=MeanFlow, Sampling Steps=42026.06 | 79.2 | 79 | |
| SGF + SAFREEBase Model=MeanFlow, Sampling Steps=42026.06 | 78.8 | 77.4 | |
| STG + Semantic SurgeryBase Model=MeanFlow, Sampling Steps=42026.06 | 76.8 | 77.7 | |
| STGBase Model=MeanFlow, Sampling Steps=42026.06 | 76 | 77.3 | |
| STG + SAFREEBase Model=MeanFlow, Sampling Steps=42026.06 | 76 | 77.3 | |
| SGFBase Model=MeanFlow, Sampling Steps=42026.06 | 76 | 75.6 | |
| VESFLOWBase Model=MeanFlow, Sampling Steps=42026.06 | 49.6 | 54.9 | |
| STGBase Model=FLUX, Sampling Steps=82026.06 | 48.8 | 53.5 | |
| VESFLOWBase Model=FLUX, Sampling Steps=82026.06 | 48.8 | 54.2 | |
| VESFLOW+Base Model=MeanFlow, Sampling Steps=42026.06 | 46.8 | 53 | |
| STG + Semantic SurgeryBase Model=FLUX, Sampling Steps=82026.06 | 46 | 53 | |
| STG + SAFREEBase Model=FLUX, Sampling Steps=82026.06 | 34.4 | 44.1 | |
| VESFLOW+Base Model=FLUX, Sampling Steps=82026.06 | 32.4 | 43.8 |