Reversible Generative Models Advance on Two Fronts: FFJORD and Glow

Two new research efforts on reversible generative models have been introduced, each taking a distinct approach to the same underlying goal: building generative models whose computations can run in both directions.

The first, FFJORD, stands for "Free-form continuous dynamics for scalable reversible generative models." As its name suggests, it frames reversible modeling in terms of continuous dynamics, aiming to make such models scalable.

The second, Glow, is described as "a reversible generative model which uses invertible 1x1 convolutions." According to its introduction, Glow extends previous work on reversible generative models while simplifying the architecture. The model's reported capabilities include generating realistic high resolution images, supporting efficient sampling, and discovering features that can be extracted from the learned representations.

Background: reversible (or invertible) generative models are a family of models designed so that the mapping between data and a latent representation can be computed exactly in both directions. This property makes it possible both to sample new data and to perform exact likelihood evaluation, which is why they are a popular alternative to other generative approaches.

Together, the two works — one pursuing free-form continuous dynamics for scalability, the other refining the architecture with invertible 1x1 convolutions for realistic image generation and efficient sampling — illustrate continued momentum in research on reversible generative modeling.