DECAFS: DISENTANGLED CONDITIONAL ADVERSARIAL FLOWS

Flow-based generative models are powerful due to their invertibility and exact likelihood estimation, but their latent spaces are often highly entangled, making it difficult to perform controlled generation (e.g., changing one specific feature of an image or molecule without affecting others). The authors seek to enable disentangled, controllable generation in these models without violating the strict invertibility constraints that typically limit flow-based architectures.

The authors propose DECAFS (Disentangled Conditional Adversarial Flows). Instead of modifying the flow architecture itself, they introduce a separate, irreversible conditional generator based on Lie groups. This generator is trained to produce a latent space that is closely aligned with the flow's latent space using an adversarial loss (WGAN). The generator is designed to partition the latent space into two components: a 'coupled' block for general features and a 'disentangled' block that uses Lie algebra to isolate specific factors of variation. A predictor network is also used to ensure that the generated latent space maintains the desired conditional structure.

DECAFS demonstrates strong performance in both image and molecular generation tasks. On the MNIST and dSprites datasets, the model achieves state-of-the-art results in disentanglement metrics (such as DCI and MIG) while maintaining competitive generation quality (FID). The authors show that the Lie group-based disentangled block allows for smooth, independent manipulation of generative factors—such as shape, size, or orientation—within the flow's latent space. Furthermore, the model successfully applies these techniques to molecular graph generation, allowing for the controlled adjustment of properties like drug-likeness (QED) and partition coefficients (logP).