Soran Ghaderi

TorchEBM: A Composable PyTorch Library for Energy-Based and Transport-Based Generative Models

TorchEBM is a PyTorch library for generative models defined either by a scalar potential or by a transport between densities. Energy-based models, diffusion, flow matching, and Schrödinger bridges are factored into one set of composable primitives: energies and fields, interpolants, couplings, training objectives, samplers, and numerical integrators. Simulation-free objectives such as flow matching, equilibrium matching, and denoising score matching require no sampling in the training loop; sampling-based objectives such as contrastive divergence remain available where a calibrated energy is needed.