Learning Probabilistic Models from Generator Latent Spaces with Hat EBM
This work addresses the challenge of probabilistic modeling with generator networks for researchers in generative AI, offering a novel method that is incremental in building on existing EBM and generator frameworks.
The paper tackles the problem of using generator networks as the basis for Energy-Based Models (EBMs) by proposing the Hat EBM method, which models images as the sum of latent variables passed through a generator and a residual variable, enabling explicit probabilistic modeling without latent inference or Jacobian calculations. Experiments show strong performance on tasks like unconditional ImageNet synthesis at 128x128 resolution, refining existing generators, and incorporating non-probabilistic generators.
This work proposes a method for using any generator network as the foundation of an Energy-Based Model (EBM). Our formulation posits that observed images are the sum of unobserved latent variables passed through the generator network and a residual random variable that spans the gap between the generator output and the image manifold. One can then define an EBM that includes the generator as part of its forward pass, which we call the Hat EBM. The model can be trained without inferring the latent variables of the observed data or calculating the generator Jacobian determinant. This enables explicit probabilistic modeling of the output distribution of any type of generator network. Experiments show strong performance of the proposed method on (1) unconditional ImageNet synthesis at 128x128 resolution, (2) refining the output of existing generators, and (3) learning EBMs that incorporate non-probabilistic generators. Code and pretrained models to reproduce our results are available at https://github.com/point0bar1/hat-ebm.