ICLR 2026 Orals

SAFETY-GUIDED FLOW (SGF): A UNIFIED FRAMEWORK FOR NEGATIVE GUIDANCE IN SAFE GENERATION

Mingyu Kim, Young-Heon Kim, Mijung Park

Diffusion & Flow Matching Sat, Apr 25 · 3:15 PM–3:25 PM · 201 A/B Avg rating: 6.50 (4–10)
Author-provided TL;DR

We introduced a unified probabilistic framework for safe generation in diffusion and flow models, using Maximum Mean Discrepancy-based energy potentials.

Abstract

Safety mechanisms for diffusion and flow models have recently been developed along two distinct paths. In robot planning, control barrier functions are employed to guide generative trajectories away from obstacles at every denoising step by explicitly imposing geometric constraints. In parallel, recent data-driven, negative guidance approaches have been shown to suppress harmful content and promote diversity in generated samples. However, they rely on heuristics without clearly stating when safety guidance is actually necessary. In this paper, we first introduce a unified probabilistic framework using a Maximum Mean Discrepancy (MMD) potential for image generation tasks that recasts both Shielded Diffusion and Safe Denoiser as instances of our energy-based negative guidance against unsafe data samples. Furthermore, we leverage control-barrier functions analysis to justify the existence of a critical time window in which negative guidance must be strong; outside of this window, the guidance should decay to zero to ensure safe and high-quality generation. We evaluate our unified framework on several realistic safe generation scenarios, confirming that negative guidance should be applied in the early stages of the denoising process for successful safe generation.

One-sentence summary·Auto-generated by claude-haiku-4-5-20251001(?)

SGF unifies negative guidance in safe generation via MMD potentials and control barrier analysis with time-critical guidance windows.

Contributions·Auto-generated by claude-haiku-4-5-20251001(?)
  • Introduces unified probabilistic framework for safe generation in diffusion and flow models
  • Recasts Shielded Diffusion and Safe Denoiser as instances of energy-based negative guidance
  • Leverages control barrier functions to identify critical time window requiring strong negative guidance
  • Demonstrates adaptive time-critical guidance achieves both safety and fidelity
Methods used·Auto-generated by claude-haiku-4-5-20251001(?)
  • Diffusion models
  • Flow models
  • Maximum mean discrepancy
  • Control barrier functions
Limitations (author-stated)·Auto-generated by claude-haiku-4-5-20251001(?)
  • Proofs assume gradient of MMD guidance aligns with ideal control barrier field near boundary
    from the paper
Future work (author-stated)·Auto-generated by claude-haiku-4-5-20251001(?)
  • Investigate ways to relax assumption by quantifying guidance mismatch
    from the paper

Author keywords

  • Safe generation
  • flow matching
  • control barrier functions

Related orals

Generative Human Geometry Distribution

Introduces distribution-over-distribution model combining geometry distributions with two-stage flow matching for human 3D generation.

Avg rating: 5.50 (2–8) · Xiangjun Tang et al.
Something off? Let us know →