VSF: Simple, Efficient, and Effective Negative Guidance in Few-Step Image Generation Models By Value Sign Flip

I. Overview

Few-step image and video generators are fast, but conventional classifier-free guidance is often ineffective at removing concepts named in a negative prompt. Existing alternatives can also require retraining or add substantial inference cost.

VSF introduces negative guidance directly inside attention, making it compatible with modern few-step diffusion and flow-matching architectures while keeping the implementation compact.

II. Key Contributions

• Introduces value sign flipping as a training-free mechanism for suppressing negative-prompt concepts.
• Uses attention masking and token handling to localize negative guidance and reduce unintended changes.
• Evaluates the method on few-step image and video generation and releases code, a demo, and a ComfyUI integration.

III. Methodology

VSF encodes positive and negative prompts together, identifies attention values associated with the negative tokens, and reverses their sign before the attention output is aggregated. A scale parameter controls suppression strength, while masks constrain where and how strongly negative guidance is applied.

IV. Main Findings

Experiments on the NegGenBench prompt pairs report stronger negative-prompt adherence than the compared few-step guidance methods while retaining competitive image quality and positive-prompt fidelity. The method is demonstrated with Stable Diffusion 3.5 Turbo, Flux Schnell, and Wan image/video models.

@inproceedings{ICLR2026_86074507,
  author = {Guo, Wenqi and Du, Shan},
  booktitle = {International Conference on Learning Representations},
  editor = {C. Vondrick and B. Hariharan and C. Raffel and L. Pinto and D. Yang and A. Faust},
  pages = {83005--83018},
  title = {VSF: Simple, Efficient, and Effective Negative Guidance in Few-Step Image Generation Models By Value Sign Flip},
  url = {https://proceedings.iclr.cc/paper_files/paper/2026/file/8607450734c14f1e0e1c89d35e1a9218-Paper-Conference.pdf},
  volume = {2026},
  year = {2026}
}