A new paper called FFJORD just dropped a genuinely elegant solution to one of the stickiest trade-offs in generative modeling. If you want to train a model by exact likelihood, you typically have to handcuff yourself to invertible architectures with cheap Jacobian determinants — think NICE, Real NVP, Glow. Those models work, but the architecture constraints are a constant headache. You’re always designing around the math instead of the problem.
FFJORD sidesteps the whole mess by defining the transformation as a continuous-time process specified by an ordinary differential equation. Instead of computing a costly determinant, the authors use Hutchinson’s trace estimator to get an unbiased estimate of the log-density. The result is a model that gives you unbiased density estimation and one-pass sampling, all while letting you use whatever neural network architecture you actually want. No more awkward coupling layers or forced invertibility tricks.
The paper demonstrates the approach across three demanding arenas: high-dimensional density estimation, image generation, and variational inference. In each case, FFJORD hits state-of-the-art results among exact likelihood methods that still offer efficient sampling. That last part matters — plenty of models get great likelihoods but require MCMC or other slow sampling procedures. FFJORD doesn’t make you choose.
What’s particularly clever is that this isn’t just another incremental tweak. The continuous-time framing fundamentally changes what’s possible. By moving the computation to a trace estimate, the authors open the door to architectures that were previously off-limits for likelihood-based training. I’d watch for follow-up work that pushes this even further — the architectural freedom here is the real story, not just the benchmark numbers.
