Manifold-Optimal Guidance reformulates Classifier-Free Guidance (CFG) as a Riemannian control problem, eliminating the artifacts and saturation typical of high guidance scales.
arXiv · March 13, 2026 · 2603.11509
Why it matters
By recognizing that standard CFG pushes sampling trajectories off the data manifold via Euclidean extrapolation, MOG provides a geometry-aware update that fixes structural collapse and over-saturation. The 'Auto-MOG' feature further eliminates the need for manual guidance scale tuning.
From the abstract
Classifier-Free Guidance (CFG) serves as the de facto control mechanism for conditional diffusion, yet high guidance scales notoriously induce oversaturation, texture artifacts, and structural collapse. We attribute this failure to a geometric mismatch: standard CFG performs Euclidean extrapolation in ambient space, inadvertently driving sampling trajectories off the high-density data manifold. To resolve this, we present Manifold-Optimal Guidance (MOG), a framework that reformulates guidance as