Reproduce “Improving Progressive Generation with Decomposable Flow Matching” 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
Moayed Haji-Ali, Willi Menapace, Ivan Skorokhodov, Arpit Sahni, Sergey Tulyakov, Vicente Ordonez, Aliaksandr Siarohin
Decomposable Flow Matching (DFM) introduces a progressive generation framework that applies independent flow timesteps across multi-scale representations using a single, shared generative model. This approach yields enhanced visual quality and faster training convergence for high-dimensional image and video synthesis compared to prior methods.
Key insights
- Progressive generation can be effectively achieved with a single, shared model by employing independent flow timesteps for each scale, simplifying architecture and training.
- Prioritizing the generation of coarse-scale structural details during training, by sampling the coarsest stage more frequently, improves overall image quality and detail.
- The framework is agnostic to the choice of multi-scale decomposition (e.g., Laplacian pyramid), providing flexibility while still benefiting from explicitly structured latent representations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.