Research map / Image generation / Text-to-image diffusion
Supercharged One-step Text-to-Image Diffusion Models with Negative Prompts
Viet Nguyen, Anh Nguyen, Trung Dao, Khoi Nguyen, Cuong Pham, Toan Tran, Anh Tran
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 standard CFG with a negative prompt to a one-step model | FLUX.1-schnell 1 step · unwanted feature removal | success rate: 44% -> 0% | Hurt |
| apply NASA-I negative prompting training-free | FLUX.1-schnell 1 step · unwanted feature removal | success rate: 44% -> 99% | Improved |
| apply CFG=2.5 to DMD one-step | DMD (PixArt-alpha) · HPSv2 | HPSv2 average: 28.71 -> 25.10 | Hurt |
| add NASA-I at inference | SBv2* (PixArt-alpha) · HPSv2 | HPSv2 average: 30.34 -> 30.98 | Improved |
| add NASA-I at inference | SBv2*, DMD2, InstaFlow (SD1.5) · HPSv2 | HPSv2 average: 27.02 -> 27.03 | No clear effect |
| add NASA-T distillation with negative prompt | SBv2* (PixArt-alpha) · HPSv2 | HPSv2 average: 30.34 -> 30.61 | Improved |
| add NASA-T on SD2.1 backbone | SBv2* (SD2.1) · HPSv2 | HPSv2 average: 27.35 -> 27.84 | Improved |
| add NASA-T and NASA-I together with alpha=1 | SBv2* (PixArt-alpha) · HPSv2 | HPSv2 average: 30.34 -> 31.21 | Improved |
| add CFG=1.5 on top of NASA-T model (ablation) | NASA-T (PixArt-alpha) · HPSv2 | HPSv2 average: 30.61 -> 27.97 | Hurt |
| put negative prompts into teacher or LoRA teacher during distillation without NASA in student (ablation) | SBv2* distillation (PixArt-alpha) · HPSv2 | HPSv2 average: 30.34 -> 29.84 | Hurt |
| add negative prompt to LoRA teacher in addition to teacher (ablation) | NASA-T distillation · HPSv2 | HPSv2 average: 30.61 -> 29.89 | Hurt |
| set NASA scale alpha too high (ablation) | NASA-I · HPSv2 | HPSv2 | Hurt |
Closest papers by meaning
- Dynamic Negative Guidance of Diffusion Models (ICLR 2025 · similarity 0.91)
- R&B: Region and Boundary Aware Zero-shot Grounded Text-to-image Generation (ICLR 2024 · similarity 0.88)
- Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.88)
- CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.88)
- SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation (CVPR 2024 · similarity 0.88)
- AdaDiff: Adaptive Step Selection for Fast Diffusion Models (AAAI 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.
