Research map / Image generation / Text-to-image diffusion
When and How do negative prompts take effect?
Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, Cho-Jui Hsieh
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 |
|---|---|---|---|
| apply negative prompt only in steps 5-15 instead of all steps (ablation) | Stable Diffusion v2 object removal · Object removal, GPT-4V eval (avg of 6 datasets) | Removal Success Rate (%): 54.16 -> 63.46 | Improved |
| apply negative prompt only in steps 5-15 instead of all steps (ablation) | Stable Diffusion v2 object removal · Object removal, human eval (avg of 6 datasets) | Removal Success Rate (%): 65.11 -> 80.35 | Improved |
| start the negative prompt later (around step 5 rather than step 1) (ablation) | Stable Diffusion v2 object removal · Timing sweep (Figure 9) | steps required to remove object: about 10 steps -> about 4 steps | Improved |
| start the negative prompt after step 11 (ablation) | Stable Diffusion v2 object removal · Timing sweep (Figure 9) | object removal | Hurt |
| apply the negative prompt in the first few diffusion steps only (ablation) | Stable Diffusion v2 · Reverse activation examples | object appears in output | Hurt |
| apply the negative prompt late rather than throughout (ablation) | Stable Diffusion v2 · Object removal | background preservation: Comparison Rate 82.84 (GPT-4V), 88.45 (human) | Improved |
Closest papers by meaning
- Devil is in the Detail: Towards Injecting Fine Details of Image Prompt in Image Generation via Conflict-free Guidance and Stratified Attention (CVPR 2025 · similarity 0.85)
- Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 2024 · similarity 0.84)
- Anti-Diffusion: Preventing Abuse of Modifications of Diffusion-Based Models (AAAI 2025 · similarity 0.83)
- Dynamic Negative Guidance of Diffusion Models (ICLR 2025 · similarity 0.83)
- Learning to Sample Effective and Diverse Prompts for Text-to-Image Generation (CVPR 2025 · similarity 0.83)
- Decomposing how prompting steers behavior (NeurIPS 2026 · similarity 0.83)
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.
