Energy-Guided Flow Matching
The paper makes pixel-space flow matching generate images through an explicit coarse-to-fine path.
EG-FM replaces the fixed clean-image endpoint with a heat-kernel-filtered endpoint that gradually releases high-frequency detail. The schedule is guided by each image’s energy, while leaving the model backbone and training data unchanged. The authors report lower FID on ImageNet class-conditional generation at 256x256, including 1.55 after 200 epochs and 1.45 after 600. They also report transfer results for text-to-image generation, with 0.85 on GenEval and 83.9 on DPG-Bench. HF Daily Papers' note
EG-FM replaces the fixed clean-image endpoint with a heat-kernel-filtered endpoint that gradually releases high-frequency detail. The schedule is guided by each image’s energy, while leaving the model backbone and training data unchanged. The authors report lower FID on ImageNet class-conditional generation at 256x256, including 1.55 after 200 epochs and 1.45 after 600. They also report transfer results for text-to-image generation, with 0.85 on GenEval and 83.9 on DPG-Bench. HF Daily Papers' note
score 4