Research map / Image generation / Text-to-image diffusion
VLM-Guided Adaptive Negative Prompting for Creative Generation
Shelly Golan, Yotam Nitzan, Zongze Wu, Or Patashnik
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 static GPT-4o negative lists with per-step adaptive VLM negatives (ablation) | VLM-Guided Adaptive Negative Prompting · Creative generation (4 categories) | relative typicality: 1.043 -> 2.190 | Improved |
| apply negatives dynamically instead of replaying accumulated list from step 0 per seed (ablation) | VLM-Guided Adaptive Negative Prompting · qualitative bag/building | creativity (qualitative) | Improved |
| use per-seed adaptive negatives instead of cross-seed reuse (ablation) | VLM-Guided Adaptive Negative Prompting · Creative generation (4 categories) | Vendi score: 4.584 -> 5.794 | Improved |
| add accumulation of negative prompts across steps (ablation) | VLM-Guided Adaptive Negative Prompting · Creative generation (4 categories) | GPT novelty score: 0.060 -> 0.370 | Improved |
| use stronger VLM (BLIP-2 or Qwen2.5 instead of ViLT) (ablation) | VLM-Guided Adaptive Negative Prompting · Creative generation (4 categories) | GPT novelty score: 0.1575 -> 0.370 (BLIP-2), 0.401 (Qwen2.5) | Improved |
| limit VLM guidance to first 10-15 of 28 steps (ablation) | VLM-Guided Adaptive Negative Prompting · Building, Bag, Jacket (qualitative) | creative quality / runtime: Qwen2.5-3B runtime 71 -> 56 s | No clear effect |
| use adaptive VLM-guided negative prompting versus creative prompting | SD3.5-large · User study | novelty (1-5): 1.753 -> 4.550 | Improved |
| use VLM-guided method instead of ConceptLab | creative generation · User study | validity (1-5): 3.950 -> 4.503 | Improved |
| add VLM-guided negative prompting versus standard SD3.5 prompting | SD3.5-large · User study | validity (1-5): 4.886 -> 4.503 | Hurt |
| remove creativity adjective from positive prompt (ablation) | VLM-Guided Adaptive Negative Prompting · qualitative | creativity (qualitative) | Hurt |
| add VLM negative prompting; runtime overhead with ViLT | SD3.5-large · Runtime | seconds per image: 22 -> 35 | Hurt |
| add creativity prompt words (Creative Prompting) alone | SD3.5-large · Creative generation (4 categories) | relative typicality: 1.640 -> 1.645 | No clear effect |
Closest papers by meaning
- Enhancing Creative Generation on Stable Diffusion-based Models (CVPR 2025 · similarity 0.89)
- IDEA-Bench: How Far are Generative Models from Professional Designing? (CVPR 2025 · similarity 0.89)
- Seeking the Unfamiliar but Memorable: Conceptual Creativity as Meta-Learning (NeurIPS 2026 · similarity 0.88)
- 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 0.88)
- Imagine Before You Draw: Visual Prompt Engineering for Image Generation (NeurIPS 2026 · similarity 0.87)
- LLM Blueprint: Enabling Text-to-Image Generation with Complex and Detailed Prompts (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.
