Research map / Vision-language and multimodal / CLIP and zero-shot
SafeR-CLIP: Mitigating NSFW Content in Vision-Language Models While Preserving Pre-Trained Knowledge
Adeel Yousaf, Joseph Fioresi, James Beetham, Amrit Singh Bedi, Mubarak Shah
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add proximity-aware redirection and relative redirection loss over Safe-CLIP | SafeR-CLIP · zero-shot classification, 11 datasets | accuracy (%): 52.2 -> 60.2 | Improved |
| add relative redirection loss (ablation) | Safe-CLIP · ViSU / zero-shot | T*->V R@1 and ZS avg: 14.5 -> 28.0; ZS 52.2 -> 56.4 | Improved |
| add proximity-based alignment (ablation) | SafeR-CLIP · zero-shot avg | accuracy (%): 56.4 -> 58.6 | Improved |
| add progressive training (ablation) | SafeR-CLIP · zero-shot avg | accuracy (%): 58.6 -> 60.2 | Improved |
| apply proximity alignment to uni-modal losses as well as cross-modal (ablation) | SafeR-CLIP · ViSU V*->T | R@1: 18.7 -> 24.6 | Improved |
| average top-2 or top-3 safe targets instead of top-1 (ablation) | SafeR-CLIP · ViSU T*->V and V*->T | R@1: T*->V 27.9 -> 26.6 (top-2), 25.5 (top-3) | Hurt |
| use JINA-CLIP instead of CLIP for caption matching (ablation) | SafeR-CLIP · zero-shot avg | accuracy (%): 60.2 -> 59.0 | Hurt |
| use SafeR-CLIP on real NSFW images | SafeR-CLIP · SMID V->T retrieval | % NSFW retrieved: 14.2 -> 3.1 | Improved |
| replace Safe-CLIP with SafeR-CLIP in Stable Diffusion | SD v1.4 · I2P | avg NSFW score: 16.1 -> 16.0 | No clear effect |
| apply SafeR-CLIP to SigLIP | SigLIP-SO400M · NSFWCaps | T*->V R@1: 46.7 -> 72.1; ZS 71.3 -> 76.6 | Improved |
| use SafeR-CLIP encoder in LLaVA-OneVision-7B | LLaVA-OneVision-7B · MMBench, NudeNet | benign task score vs NSFW %: MMBench 81.8 -> 78.0; NudeNet 19.6 -> 9.4 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Safe-CLIP — Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models (ECCV 2024) · better on 14 of 14 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Hyperbolic Safety-Aware Vision-Language Models (CVPR 2025 · similarity 0.91)
- ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.90)
- VLSU: Mapping the Limits of Joint Multimodal Understanding for AI Safety (ICLR 2026 · similarity 0.89)
- Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility (ICML 2026 · similarity 0.88)
- Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models (ICLR 2026 · similarity 0.88)
- Layer-wise Alignment: Examining Safety Alignment Across Image Encoder Layers in Vision Language Models (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
