Research map / Image generation / Text-to-image diffusion
Latent Guard: a Safety Framework for Text-to-image Generation
Runtao Liu, Ashkan Khakzar, Jindong Gu, Qifeng Chen, Philip Torr, Fabio Pizzati
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 safe prompts as negatives in the contrastive loss (ablation) | Latent Guard · CoPro ID Synonym | AUC: 0.607 -> 0.914 | Improved |
| use cross-attention instead of MLP in the Embedding Mapping Layer (ablation) | Latent Guard · CoPro ID Adversarial | AUC: 0.818 -> 0.908 | Improved |
| use cross-attention instead of MLP (ablation) | Latent Guard · CoPro OOD Explicit | AUC: 0.947 -> 0.944 | Hurt |
| concept-matching setup instead of binary safe/unsafe classifier (ablation) | Latent Guard · Unsafe Diffusion | accuracy: 0.737 -> 0.794 | Improved |
| use a larger blacklist at test time (10% to 100% of concepts) (ablation) | Latent Guard · Unsafe Diffusion | accuracy: 0.548 -> 0.794 | Improved |
| increase attention heads from 1 to 16 (d=128) (ablation) | Embedding Mapping Layer · CoPro fine-grained concept classification | accuracy: 0.874 -> 0.931 | Improved |
| train Latent Guard embedding vs plain CLIP/BERT similarity | concept-prompt similarity · CoPro ID Explicit | AUC: 0.697 -> 0.985 (CLIPScore vs ours) | Improved |
| Latent Guard prompt filtering vs LLM filtering before generation | T2I pipeline · Unsafe Diffusion | NudeNet+Q16 inappropriate rate: 0.138 -> 0.029 | Improved |
| Latent Guard vs LLM baseline on adversarial prompts | CoPro Adversarial accuracy · CoPro ID Adversarial | accuracy: 0.867 vs 0.829 | Hurt |
Later papers that compared with this method
1 later accepted paper reports Latent Guard as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- UniNDM: A Unified Noise-Driven Detection and Mitigation Framework Against Sexual Content in Text-to-Image Generation (TPAMI 2026 · better on 1 of 1 shared result)
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
- GuardT2I: Defending Text-to-Image Models from Adversarial Prompts (NeurIPS 2024 · similarity 0.90)
- SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025 · similarity 0.88)
- Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.87)
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · similarity 0.86)
- Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails (NeurIPS 2026 · similarity 0.86)
- Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control (CVPR 2025 · similarity 0.86)
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.
