VSF

01 — Overview

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.

02 — Contributions

Key Contributions

• 01

  Introduces value sign flipping as a training-free mechanism for suppressing negative-prompt concepts.

• 02

  Uses attention masking and token handling to localize negative guidance and reduce unintended changes.

• 03

  Evaluates the method on few-step image and video generation and releases code, a demo, and a ComfyUI integration.

03 — Method

Method

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.

04 — Evaluation

Results

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.

05 — Reference

Citation

@InProceedings{Guo_2026_VSF,
  author    = {Guo, Wenqi and Du, Shan},
  title     = {{VSF}: Simple, Efficient, and Effective Negative Guidance in Few-Step Image Generation Models By Value Sign Flip},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026},
  url       = {https://openreview.net/forum?id=W2NINfoVtN}
}