Research map / Vision-language and multimodal / CLIP and zero-shot
Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models
Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara
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 inappropriate content redirection losses to the fine-tuning (ablation) | Safe-CLIP · ViSU test set, T*-to-V∪V* | R@10: 30.4 -> 46.9 | Improved |
| add negative cosine similarity (single-modality redirect) losses (ablation) | Safe-CLIP · ViSU test set, safe T-to-V | R@1: 41.9 -> 45.9 | Improved |
| add negative cosine similarity losses (ablation) | Safe-CLIP · ViSU test set, V*-to-T∪T* | R@1: 13.7 -> 19.1 | Improved |
| add inappropriate content redirection losses (ablation) | Safe-CLIP · ViSU test set, safe T-to-V | R@1: 49.9 -> 45.9 | Hurt |
| replace CLIP text encoder with Safe-CLIP in Stable Diffusion v1.4 | Stable Diffusion v1.4 · ViSU prompts | probability of unsafe image (%): 26.2 -> 3.6 | Improved |
| add Safe-CLIP on top of SLD-Strong | SLD-Strong · I2P prompts (avg) | probability of unsafe image (%): 13.5 -> 13.0 | Improved |
| use Safe-CLIP visual encoder in LLaVA | LLaVA · NudeNet NSFW images | % NSFW text: 62.6 -> 26.7 | Improved |
| add SFT of Llama 2-Chat on safe-unsafe pairs (ablation) | NSFW text generator · ViSU text, GPT-3.5 NSFW | % NSFW: 9.3 -> 75.0 | Improved |
| add DPO stage after SFT (ablation) | NSFW text generator · ViSU text, GPT-3.5 NSFW | % NSFW: 75.0 -> 79.1 | Improved |
| fine-tune with Safe-CLIP versus DataComp-1B cleaned-data CLIP | CLIP · Real NSFW images, text-to-image (NudeNet) | % NSFW retrieved: 55.6 -> 8.4 | Improved |
Later papers that compared with this method
2 later accepted papers report Safe-CLIP as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- SafeR-CLIP: Mitigating NSFW Content in Vision-Language Models While Preserving Pre-Trained Knowledge (AAAI 2026 · better on 14 of 14 shared results)
- Hyperbolic Safety-Aware Vision-Language Models (CVPR 2025 · better on 7 of 8 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 0.90)
- Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Better Safe than Sorry: Pre-training CLIP against Targeted Data Poisoning and Backdoor Attacks (ICML 2024 · similarity 0.88)
- ToxicTextCLIP: Text-Based Poisoning and Backdoor Attacks on CLIP Pre-training (NeurIPS 2025 · similarity 0.88)
- Image Safeguarding: Reasoning with Conditional Vision Language Model and Obfuscating Unsafe Content Counterfactually (AAAI 2024 · similarity 0.87)
- SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025 · similarity 0.87)
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.
