Safety-Guided Flow (SGF): A Unified Framework for Negative Guidance in Safe Generation

2026-03-17 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Safety-Guided Flow (SGF) is a unified probabilistic framework for safe generation in diffusion and flow models, addressing the need for robust safety mechanisms as these models enter high-stakes domains. It unifies existing heuristic methods like Shielded Diffusion (Kirchhof et al., 2025) and Safe Denoiser (Kim et al., 2025b) under an energy-based negative guidance approach, utilizing a Maximum Mean Discrepancy (MMD) potential. The framework leverages control-barrier function analysis to identify a "critical time window" early in the denoising process where negative guidance must be strong, decaying to zero afterward to ensure both safety and high-quality generation. Experiments confirm that applying guidance in early steps, specifically for windows like [1.0, 0.8] or [1.0, 0.6], significantly reduces attack success rates (ASR) against nudity prompts and mitigates memorization, while preserving diversity and image fidelity.

Key takeaway

For research scientists and engineers developing or deploying generative AI, understanding the temporal dynamics of safety guidance is critical. Your models will achieve superior safety and fidelity by implementing negative guidance strongly in the early stages of the denoising process, rather than uniformly throughout. Over-applying guidance beyond this "critical window" can degrade image quality and stability, so focus your safety interventions strategically to maximize impact and efficiency.

Key insights

Early, strong negative guidance within a critical time window is crucial for safe and high-quality generative model outputs.

Principles

• MMD potential gradients create repulsive vector fields.
• Control-barrier functions justify time-varying guidance strength.
• Early denoising steps set coarse structure, requiring strong initial guidance.

Method

SGF uses an MMD potential gradient to generate repulsive forces against unsafe distributions. Control-barrier analysis determines a critical time window for strong guidance, which then decays to zero, applied in the x0 space.

In practice

• Apply negative guidance in early denoising steps (e.g., [1.0, 0.8]).
• Use RBF kernels for MMD potential in image generation.
• Estimate kernel bandwidth empirically for adaptive guidance.

Topics

• Safety-Guided Flow
• Diffusion Models
• Negative Guidance
• Control Barrier Functions
• Maximum Mean Discrepancy

Code references

• MingyuKim87/SGF
• rtqichen/torchdiffeq
• facebookresearch/flow_matching
• CharlesGong12/RECE
• jaehong31/SAFREE

Best for: Computer Vision Engineer, Research Scientist, AI Researcher, AI Scientist, Deep Learning Engineer

Related on AIssential

• DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising — DiRecT improves diffusion model safety by enforcing constraints only on final trajectories, avoiding intermediate denoising over-constrainment.
• Unified Safe In-context Image Generation in Multimodal Diffusion Transformers via Restricting Unsafe Information Flows — UVR restricts unsafe information flow in DiT multimodal attention during an early "start-up stage" to prevent harmful image generation.
• Robust and Generalizable Safety Steering for Text-to-Image Diffusion Transformers — SafeDIG enables robust and generalizable safety steering for Diffusion Transformers by transferring sparse features across risk domains.
• Greedy Coordinate Diffusion: Effective and Semantically Coherent Adversarial Attacks via Diffusion Guidance — Fine-tuning language models can degrade safety guardrails due to second-order effects, formalized by the Alignment Instability Condition (AIC).
• SafeDiffusion-R1: Online Reward Steering for Safe Diffusion Post-Training — Online reinforcement learning with CLIP embedding steering effectively enhances diffusion model safety without supervised data.
• RiskFlow: Fast and Faithful Safety-Critical Traffic Scenario Generation — RiskFlow efficiently generates realistic safety-critical traffic scenarios by transforming Gaussian action sequences in a single pass with output-space guidance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.