Problem
Research questions and friction points this paper is trying to address.
pixel-space generative models
flow matching
spectral evolution
high-dimensional generation
image generation
Energy-Guided Flow Matching
🤖 AI Summary
This work addresses the limitations of existing pixel-level generative models, which struggle to jointly learn global image structure and fine details due to the implicit handling of spectral evolution in conventional flow matching approaches. To overcome this, the authors propose Energy-Guided Flow Matching (EG-FM), a novel method that explicitly governs the coarse-to-fine generation process—from low-frequency smooth images to high-frequency details—through a dynamic endpoint mechanism based on image energy scheduling, without modifying the backbone architecture or training data. By integrating heat kernel filtering endpoints with an energy-guided strategy, EG-FM achieves state-of-the-art performance with minimal computational overhead: on ImageNet 256×256, it attains FID scores of 1.55 and 1.45 in 200 and 600 epochs, respectively, and reaches FID 1.58 at 512×512 resolution after only 40 fine-tuning epochs, while also achieving GenEval 0.85 and DPG-Bench 83.9 in text-to-image generation.
📝 Abstract
Pixel-space generative models bypass lossy latent compression, yet necessitate joint learning of global structure and fine-grained details in a high-dimensional space. Standard flow matching interpolates noise toward a fixed clean-image endpoint, leaving the spectral evolution to be learned implicitly. In this paper, we introduce Energy-Guided Flow Matching(EG-FM) that explicitly models a coarse-to-fine generative trajectory by moving endpoint. Specifically, EG-FM replaces the fixed endpoint with a heat-kernel-filtered endpoint that evolves smoothly from low-frequency image to clean image.The fraction of high-frequency signal in moving endpoint is released by an image-specific energy-guided scheduling, leading to the re-targeting of velocity in flow matching.Our framework requires no adaptation of the backbone and training data, bringing negligible cost on the training and inference stages. In our experiment, EG-FM consistently achieves lower FID on the ImageNet class-conditional image generation task at $256 \times 256$ with fewer epochs, reaching an FID of 1.55 at 200 epochs and 1.45 at 600 epochs. We continue training the generation task on the setting of $512 \times 512$ resolution, yielding a FID of 1.58 after only 40 high-resolution adaptation epochs.Furthermore, we transfer EG-FM on text-to-image generation and achieve 0.85 on GenEval score and 83.9 on DPG-Bench. Code is available at https://github.com/ysng123/EG-FM.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Energy-Guided Flow Matching
moving endpoint
coarse-to-fine generation
flow matching
image synthesis
🔎 Similar Papers
No similar papers found.
💼 Related Jobs
