Reproduce “Decafs: Disentangled Conditional adversarial Flows” 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
Anirudh jain, Sakshi Varshney, Samuel Kaski, Vikas Garg
DECAFS, developed by researchers at Aalto University and Orion Pharma, introduces a conditional generative framework for normalizing flows that utilizes a Lie-group based generator to achieve disentangled latent spaces. The approach enables interpretable and controlled generation across image synthesis and molecular design tasks, demonstrating improved performance in disentanglement metrics and precise manipulation of specific attributes.
Key insights
- Lie algebra can be effectively used to construct an explicitly disentangled latent space within a flow-based generative model.
- Combining normalizing flows with adversarial training and a dedicated predictor network allows for conditional generation without expanding the flow's dimensionality.
- Separating latent factors into disentangled and coupled components provides flexible control, enabling independent manipulation of specific attributes while preserving necessary correlations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.