Adaptive Manifold Guidance Improves Diffusion Sampling

A new paper on arXiv proposes Adaptive Manifold Guidance (AdaMaG), a plug-and-play method that corrects probability conservation violations in classifier-free guidance for diffusion and flow-based generative models. The authors analyze guidance through the continuity equation, showing that extrapolation-based methods break probability conservation and drive samples off the learned manifold under strong guidance. They prove that the divergence term in guidance blows up near the data manifold, motivating a time-dependent schedule and score-parallel attenuation. AdaMaG bounds both terms at no additional inference cost, offering a principled alternative to heuristic linear combinations of velocities or scores.