Research map / Image generation / Text-to-image diffusion
VSF: Simple, Efficient, and Effective Negative Guidance in Few-Step Image Generation Models By Value Sign Flip
Wenqi Marshall Guo, Shan Du
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 |
|---|---|---|---|
| add VSF negative guidance (Strong) | SD-3.5-Large-Turbo · NegGenBench | negative score: 0.195 -> 0.545 | Improved |
| add VSF negative guidance (Quality) | SD-3.5-Large-Turbo · NegGenBench | negative score: 0.195 -> 0.420; quality 0.968 -> 0.986 | Improved |
| add VSF to Flux Schnell | Flux Schnell · NegGenBench | negative score: 0.22 -> 0.41 | Improved |
| use VSF instead of NASA | SD-3.5-Large-Turbo · NegGenBench | quality score: 0.867 -> 0.952 (VSF Strong) | Improved |
| flip whole text embedding (WEF) instead of attention values (ablation) | VSF · NegGenBench trade-off | negative score | Hurt |
| add attention masking (ablation) | VSF · NegGenBench trade-off | positive score | Improved |
| add attention bias -beta (ablation) | VSF · NegGenBench trade-off | quality score | No clear effect |
| use VSF instead of CFG (28 steps) | SD-3.5-Large-Turbo · NegGenBench | negative score: 0.300 -> 0.420 | Improved |
| VSF vs NAG compute cost | SD-3.5-Large-Turbo · runtime | time (s): 2.98s (NAG) vs 3.00s (VSF) | No clear effect |
Closest papers by meaning
- Dynamic Negative Guidance of Diffusion Models (ICLR 2025 · similarity 0.85)
- Referee Can Play: An Alternative Approach to Conditional Generation via Model Inversion (ICML 2024 · similarity 0.85)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.85)
- Visual Generation Without Guidance (ICML 2025 · similarity 0.85)
- 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)
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · similarity 0.85)
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.
