Red-teaming Image Safety Detectors on Red-teaming evaluation set
100ASRRedEdit
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| RedEditDetector=Gemini-3.1-Pro†, Type=General VLM, Proposer=Qwen3.6-35B2026.06 | 100 | 94.7 | 1 | 16 | |
| RedEditDetector=FalconsaiNSFW, Type=Conventional, Proposer=Qwen3.6-35B2026.06 | 95.2 | 95.6 | 1.1 | 62 | |
| RedEditDetector=Qwen3.6-35B, Type=General VLM, Proposer=Qwen3.6-35B2026.06 | 76.2 | 93 | 1.26 | 164 | |
| RedEditDetector=NudeNet, Type=Conventional, Proposer=Qwen3.6-35B2026.06 | 75.9 | 95.3 | 1.37 | 83 | |
| RedEditDetector=GPT-5.5, Type=General VLM, Proposer=Qwen3.6-35B2026.06 | 73.2 | 93.2 | 1.31 | 250 | |
| RedEditDetector=Claude-Opus-4-7, Type=General VLM, Proposer=Qwen3.6-35B2026.06 | 58.4 | 93.1 | 1.28 | 262 | |
| RedEditDetector=MultiHeaded, Type=Conventional, Proposer=Qwen3.6-35B2026.06 | 56.2 | 95.4 | 1.38 | 242 | |
| RedEditDetector=SD_Filter, Type=Conventional, Proposer=Qwen3.6-35B2026.06 | 50 | 95.8 | 1.26 | 130 | |
| RedEditDetector=NSFW_Detector, Type=Conventional, Proposer=Qwen3.6-35B2026.06 | 46.3 | 95.7 | 1.39 | 121 | |
| RedEditDetector=Q16, Type=Conventional, Proposer=Qwen3.6-35B2026.06 | 44.3 | 95.7 | 1.28 | 357 | |
| RedEditDetector=LlavaGuard, Type=Safety-tuned, Proposer=Qwen3.6-35B2026.06 | 41.1 | 95.4 | 1.3 | 348 | |
| RedEditDetector=ImageGuard, Type=Safety-tuned, Proposer=Qwen3.6-35B2026.06 | 20.1 | 95.2 | 1.31 | 354 |