Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·7d ago·5 min read
Keep It CALM: Analyzing the Limits of Global Unsafety in Text-to-Image Generation
TL;DR
CALM (Counterfactual Adaptive Local Modulation) is a new training-free safeguard for text-to-image generation that enhances unsafe content suppression while maintaining benign output quality. It achieves this by applying localized corrections to prompts rather than relying on a broad, global safety signal.
✦ Why It Matters
Implement CALM in your text-to-image generation workflows to improve safety without sacrificing output quality.
Key Takeaways
- Adopt CALM for localized safety corrections in image generation.
- Evaluate the balance between safety and output quality in your models.
- Utilize matched anchors for effective prompt adjustments.
- Explore the implications of localized versus global safety measures.
How It Works
CALM operates by identifying unsafe categories associated with each prompt and applying localized corrections. Instead of removing unsafe signals globally, it focuses on specific tokens that violate safety guidelines, making minimal edits to steer them towards safe representations.
This targeted approach allows for better preservation of benign content while effectively suppressing unsafe elements.