Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 12·5 min read
Towards More General Control of Diffusion Models Using Jeffrey Guidance
TL;DR
Diffusion models often lack explicit control over their output distributions, limiting their flexibility. Jeffrey guidance is a new framework that enhances control by updating marginal distributions towards a target while maintaining the overall structure.
✦ Why It Matters
Engineers can leverage Jeffrey guidance to enhance control and fairness in generative models for diverse applications.
Key Takeaways
- Explore Jeffrey guidance for enhanced control in diffusion models.
- Utilize the framework to target specific embedding distributions.
- Apply the method to improve fairness in image generation.
- Reduce FID scores significantly in your generative tasks.
How It Works
Jeffrey guidance employs a principled approach based on Jeffrey's rule of conditioning, which allows for the adjustment of marginal distributions towards a specified target. This method preserves the conditional structure of the data while minimally altering the joint distribution, enabling more precise control over the output of diffusion models.