Probability-Conserving Flow Guidance
Researchers have developed a new guidance method called Adaptive Manifold Guidance (AdaMaG) for diffusion and flow-based generative models. This technique addresses limitations in existing methods like Classifier-Free Guidance (CFG) by analyzing guidance through the continuity equation. AdaMaG ensures probability conservation and keeps generated samples on the learned manifold, even under strong guidance, by bounding divergence and score-parallel terms. AI
IMPACT AdaMaG enhances realism and reduces hallucinations in image generation, potentially improving the quality and reliability of AI-generated visuals.