Research map / Image generation / Fast diffusion sampling
Dynamic Negative Guidance of Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Dynamic Negative Guidance (DNG), a state- and time-dependent negative prompting scheme that estimates the forbidden-class posterior by tracking the diffusion Markov chain, evaluated on MNIST/CIFAR10 class removal and Stable Diffusion.
Problem
Conventional negative prompting (NP) uses a constant guidance scale, which is suboptimal or fails to sample the target distribution because the reverse diffusion process is non-stationary and state-dependent.
Approach
Derives a dynamic guidance scale proportional to p(c_bad|x)/(1-p(c_bad|x)) that is large near the forbidden region and vanishes far from it; the posterior is estimated during denoising from the likelihood ratio of unconditional vs. forget-model Gaussian transitions, requiring no additional training.
Key results
- On CIFAR10 single-class removal, DNG outperforms both NP and SLD on all four classes in the high-safety regime (lower KL and FID at equal safety).
- On Stable Diffusion with an unrelated negative prompt, DNG causes minimal CLIP-score change while NP alters images, with DNG performing on par with NP for related negative prompts.
- In 1D Gaussian mixtures, DNG with exact posterior matches CFG, while NP fails to sample the target distribution.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DNG instead of NP or SLD for class removal at high safety | DNG · CIFAR10 class removal | FID | Helped |
| Use DNG instead of NP or SLD for class removal at high safety | DNG · CIFAR10 class removal | KL-divergence | Helped |
| Use DNG instead of NP with a semantically unrelated negative prompt | Stable Diffusion 2 · Text-to-image, unrelated negative prompt | CLIP score | Helped |
| Use DNG instead of NP with a semantically related negative prompt | Stable Diffusion 2 · Text-to-image, related negative prompt | CLIP score | No clear effect |
| Use exact-posterior DNG instead of NP | One-dimensional Gaussian mixture diffusion · One-dimensional Gaussian mixture | target-distribution sampling | Helped |
Surprising findings
- Conventional negative prompting completely fails to sample the target conditional distribution in the 1D Gaussian mixture example.
- NP modifies images even when the negative prompt is semantically unrelated to the positive prompt, suggesting unnecessary diversity loss.
Limitations
- Stable Diffusion experiments are small-scale: only 5 prompts, no FID, and no object classifier verifying removal.
- Hyperparameters (prior, temperature, offset) are prompt-specific and chosen in limited intervals; no general setting found.
- Posterior approximation relies on first-order slow-change assumption and Gaussian transition approximations.
Experimental setup
Datasets: MNIST CIFAR10
Benchmarks: MNIST class removal CIFAR10 class removal
Metrics: safety (percentage of forbidden class generated) KL-divergence to ideal class distribution FID CLIP score
Baselines: Negative Prompting Safe Latent Diffusion
Builds on: Classifier-Free Guidance DDPM Safe Latent Diffusion
Closest papers by meaning
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.91)
- Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Feedback Guidance of Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.89)
- REG: Rectified Gradient Guidance for Conditional Diffusion Models (ICML 2025 · similarity 0.89)
- Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.88)
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.