Research map / Image generation / Text-to-image diffusion
Position: Towards Implicit Prompt For Text-To-Image Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper introducing ImplicitBench (about 2,000+ implicit prompts across General Symbols, Celebrity Privacy, and NSFW) to evaluate six T2I models on generating targets described indirectly, showing privacy and NSFW safety risks.
Problem
Existing T2I benchmarks and safety filters only consider explicit prompts that name the target, leaving implicit prompts (descriptions that hint at a target without naming it) unexplored, though they may bypass safety constraints and raise privacy and NSFW risks.
Approach
Collects implicit prompts (GPT-4-generated descriptions of symbols and celebrities, plus filtered Lexica NSFW prompts) and evaluates whether T2I models generate the implied target, using GPT-4V and human judgment for symbols, ArcFace face verification for celebrities, and safety checkers/classifiers for NSFW.
Key results
- DALL-E 3 achieves the highest overall success rate on general symbols (89.03%), while all six models exceed 81%.
- Midjourney, DALL-E 2, and DALL-E 3 produce unsafe images on 27.37%, 13.27%, and 17.75% of implicit NSFW prompts respectively (checker plus classifier).
- Midjourney and Stable Diffusion generate celebrities more successfully than DALL-E 2 and DALL-E 3, which the authors attribute to stricter privacy policies of DALL-E models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use implicit rather than explicit prompts for general symbols | Text-to-image models · General Symbols | success rate | Helped |
| Use DALL-E 3 rather than other evaluated models for implicit general-symbol prompts | Text-to-image models · General Symbols (Overall) | success rate (%): 89.03 | Helped |
| Use implicit prompts to describe highly famous celebrities | Text-to-image models · Celebrity Privacy | celebrity generation accuracy | Helped |
| Use DALL-E 2 or DALL-E 3 instead of Midjourney or Stable Diffusion for implicit celebrity prompts | Text-to-image models · Celebrity Privacy | celebrity generation accuracy | Hurt |
| Use implicit rather than explicit prompts for NSFW content | Text-to-image models · NSFW implicit prompts | unsafe image generation | Helped |
| Use Stable Diffusion XL rather than Stable Diffusion v1-5 | Stable Diffusion · NSFW implicit prompts (Overall) | unsafe images (%): 33.38 -> 30.53 | Helped |
Surprising findings
- Closed-source DALL-E 3 still produced over 17% NSFW content from implicit prompts despite its safety checker.
- Logo generation is the weakest category for all models, with DALL-E 3 at 61.11% and Stable Diffusion v1.5 at 36.00%.
Limitations
- Relies on GPT-4V as judge, which the authors note can hallucinate, and only partly mitigated by manual checks.
- Celebrity evaluation uses only five gallery photos per celebrity and a face-similarity threshold, and the sample sizes per category are small.
- Proposed mitigation solutions are described but not implemented or evaluated.
Experimental setup
Datasets: Lexica prompts (crawled and filtered) GPT-4-generated implicit prompts
Benchmarks: ImplicitBench
Metrics: success rate unsafe image percentage CLIP Score ImageReward
Baselines: Stable Diffusion v1-5 Stable Diffusion v2-1 Stable Diffusion XL Midjourney DALL-E 2 DALL-E 3
Builds on: DrawBench T2I-CompBench Stable Diffusion
Closest papers by meaning
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · similarity 0.90)
- OVERT: A Benchmark for Over-Refusal Evaluation on Text-to-Image Models (NeurIPS 2025 · similarity 0.89)
- Implicit Bias Injection Attacks against Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.88)
- Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 2024 · similarity 0.87)
- Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models (CVPR 2025 · similarity 0.87)
- Revisiting text-to-image evaluation with Gecko: on metrics, prompts, and human rating (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.