Researchers have introduced Energy-Guided Flow Matching (EG-FM), a novel approach to generative modeling that explicitly models a coarse-to-fine trajectory for image synthesis. The method addresses a key challenge in pixel-space generative models: jointly learning global structure and fine-grained details in high-dimensional space.
Traditional flow matching interpolates noise toward a fixed clean-image endpoint, leaving spectral evolution to be learned implicitly. EG-FM replaces this fixed endpoint with a heat-kernel-filtered endpoint that evolves smoothly from a low-frequency image to a clean image. An image-specific energy-guided scheduling controls the release of high-frequency signals, re-targeting the velocity in flow matching.
The framework requires no adaptation of the backbone or training data, adding negligible cost to both training and inference. In experiments, EG-FM consistently achieved lower FID scores on ImageNet class-conditional generation at 256×256 resolution, reaching an FID of 1.55 at 200 epochs and 1.45 at 600 epochs. At 512×512 resolution, it achieved an FID of 1.58 after only 40 high-resolution adaptation epochs.
EG-FM also transferred successfully to text-to-image generation, scoring 0.85 on GenEval and 83.9 on DPG-Bench. The code is available on GitHub.