Energy-Guided Flow Matching for Coarse-to-Fine Image Generation
August 6, 2026
EG-FM introduces an energy-guided scheduling that replaces fixed endpoints in flow matching with a heat-kernel-filtered endpoint. This allows the model to explicitly learn a coarse-to-fine generative trajectory from low-frequency structures to high-frequency details.
HOW THIS AFFECTS YOU
●
researcherYou can improve pixel-space generative models by explicitly modeling spectral evolution during training.
●
designerThis approach may lead to higher-fidelity image generation with better structural consistency.