Unified Safe In-context Image Generation in Multimodal Diffusion Transformers via Restricting Unsafe Information Flows

2026-06-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

Unified Visual Safety Regulator (UVR) is a novel, training-free framework designed to prevent the generation of harmful content in Multimodal Diffusion Transformers (MM-DiTs), particularly addressing limitations of existing safety mechanisms in image-to-image (I2I) editing tasks. Grounded in an analysis of attention dynamics within MM-Attn, UVR identifies a task-independent "semantic start-up stage" where unsafe semantics rapidly emerge and can be localized. It then mitigates harmful generation through unified, targeted attention modulation and explicit restriction of unsafe information flow over identified output patches. Experiments on FLUX.1-dev and FLUX.1-Kontext-dev across various concepts, including nudity, IP characters, and inappropriate objects, demonstrate UVR's state-of-the-art safety performance, achieving 91% and 77% erase rates in image synthesis and editing tasks, respectively, while preserving visual quality and fidelity.

Key takeaway

For AI Security Engineers or Machine Learning Engineers implementing safety mechanisms for multimodal diffusion models, UVR offers a robust, training-free solution. Its unified approach for both text-to-image synthesis and image-to-image editing, achieving high erase rates (91% and 77% respectively) with minimal quality degradation, makes it a compelling choice. You should consider integrating UVR for inference-time safety control in FLUX-series or similar DiT architectures, especially for context-insensitive risks like explicit content or intellectual property violations.

Key insights

UVR unifies safety in multimodal DiTs by modulating attention to restrict unsafe information flow at early, localized stages.

Principles

• Unsafe semantics emerge early in DiT generation.
• Attention dynamics reveal task-independent and task-specific stages.
• Targeted attention modulation can block harmful content.

Method

UVR localizes unsafe visual patches using pre-collected "unsafe anchors" and then regulates them via adaptive attention modulation and explicit restriction of harmful information flows, primarily during the semantic start-up stage.

In practice

• Construct unsafe anchors from final diffusion timestep outputs.
• Apply spatial refinement to localization masks.
• Inject Gaussian noise into core unsafe tokens early.

Topics

• Diffusion Transformers
• Multimodal Attention
• Image Generation Safety
• Content Moderation
• Image-to-Image Editing
• Training-Free Methods

Code references

• deng12yx/UVR
• notAI-tech/NudeNet

Best for: Research Scientist, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• 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.
• Visual Self-Fulfilling Alignment: Shaping Safety-Oriented Personas via Threat-Related Images — Implicit visual exposure to threat-related content can shape safety-oriented personas in MLLMs without explicit safety labels.
• Jailbreaking Multimodal Large Language Models using Multi-Clip Video — Video MLLMs are more vulnerable to jailbreaking with multi-clip, diverse, and dynamic video inputs than images.
• Musk’s tactic of blaming users for Grok sex images may be foiled by EU law — EU lawmakers are targeting AI platforms, not just users, to prevent non-consensual sexual image generation.
• Safety-Guided Flow (SGF): A Unified Framework for Negative Guidance in Safe Generation — Early, strong negative guidance within a critical time window is crucial for safe and high-quality generative model outputs.

Open in AIssential →

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