Reproduce “FREPix: Frequency-Heterogeneous Flow Matching for Pixel-Space Image Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mingfeng Lin, Jiakun Chen, Liang Han, Liqiang Nie
Researchers from Harbin Institute of Technology (Shenzhen) developed FREPix, a pixel-space image generation framework that employs frequency-heterogeneous flow matching to explicitly model distinct frequency components. This method achieved a FID of 1.91 on ImageNet 256x256 and demonstrated improved efficiency in early training and low-NFE sampling regimes.
Key insights
- Explicitly modeling frequency heterogeneity throughout the generation pipeline (state, paths, architecture, loss) provides a stronger inductive bias for pixel-space image synthesis, leading to improved quality and efficiency.
- Establishing a coarse-to-fine generation trajectory by allowing low-frequency components to progress slightly ahead in their interpolation path effectively guides the model to form structure before refining details.
- Factorizing the network architecture and conditioning the high-frequency module on predicted low-frequency information reduces frequency entanglement and improves generalization by assigning specialized roles.
- Time-dependent frequency weighting in the loss function is crucial for optimizing the distinct learning dynamics of different frequency bands, aligning with their asymmetric recovery difficulty.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.