Research map / Image generation / Text-to-image diffusion
MacPrompt: Maraconic-Guided Jailbreak Against Text-to-Image Models
Xi Ye, Yiwen Liu, Lina Wang, Run Wang, Geying Yang, Yufei Hou, Jiayi Yu
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 |
|---|---|---|---|
| Replace sensitive words with macaronic cross-lingual substitutes vs best baseline | SD v2.1 · NSFW-200 Sex | ASR-1 (%): 30 -> 56 (SurrogatePrompt vs MacPrompt) | Improved |
| Use MacPrompt instead of DACA | SD v2.1 · NSFW-200 Violence | ASR-1 (%): 60 -> 42 | Hurt |
| Use MacPrompt instead of DiffZOO | PromptGuard defense · NSFW-200 Sex | ASR-1 / ASR-5 (%): 8/26 -> 24/34 | Improved |
| Use MacPrompt instead of SurrogatePrompt | Blacklist filter · NSFW-200 Sex | BPR (%): 100 -> 100 | No clear effect |
| Use MacPrompt instead of DiffZOO | LatentGuard filter · NSFW-200 Sex | BPR (%): 56 -> 82 | Improved |
| Use MacPrompt instead of MMP-Attack | SD v2.1 · Object-200 Dog | ASR-1 (%): 66 -> 96 | Improved |
| Use MacPrompt instead of MMP-Attack | ESD · Object-200 Dog | ASR-1 (%): 78 -> 64 | Hurt |
| Use MacPrompt instead of MMP-Attack | Concept removal models · Object-200 (4 classes, SD/ESD/FMN/EAP) | ASR-1 / ASR-5 (%): SD, FMN gains; EAP/ESD often lower | Mixed |
| Use MacPrompt instead of MMP-Attack | MacPrompt · Image-prompt consistency | BLIPScore: 0.414 -> 0.6953 | Improved |
| Use MacPrompt prompts against SafeGen | SafeGen defense · NSFW-200 Violence | ASR-1 / ASR-5 (%): 14/46 (SurrogatePrompt) -> 49/90 | Improved |
Closest papers by meaning
- Perception-Guided Jailbreak Against Text-to-Image Models (AAAI 2025 · similarity 0.90)
- GuardT2I: Defending Text-to-Image Models from Adversarial Prompts (NeurIPS 2024 · similarity 0.89)
- T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks (NeurIPS 2025 · similarity 0.88)
- ColJailBreak: Collaborative Generation and Editing for Jailbreaking Text-to-Image Deep Generation (NeurIPS 2024 · similarity 0.88)
- Direct Unlearning Optimization for Robust and Safe Text-to-Image Models (NeurIPS 2024 · similarity 0.87)
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · 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.
