Research map / Probabilistic and scientific modeling / Variational and generative sampling
FFJORD: Free-Form Continuous Dynamics for Scalable Reversible Generative Models
Will Grathwohl, Ricky T. Q. Chen, Jesse Bettencourt, Ilya Sutskever, David Duvenaud
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FFJORD instead of the strongest reported autoregressive baseline | Density estimation · POWER | negative log-likelihood: -0.62 -> -0.46 | Hurt |
| Use FFJORD instead of the strongest reported autoregressive baseline | Density estimation · GAS | negative log-likelihood: -11.96 -> -8.59 | Hurt |
| Use FFJORD instead of the strongest reported autoregressive baseline | Density estimation · HEPMASS | negative log-likelihood: 15.09 -> 14.92 | Improved |
| Use FFJORD instead of the strongest reported autoregressive baseline | Density estimation · MINIBOONE | negative log-likelihood: 8.86 -> 10.43 | Hurt |
| Use FFJORD instead of the strongest reported autoregressive baseline | Density estimation · BSDS300 | negative log-likelihood: -157.73 -> -157.40 | Hurt |
| Use FFJORD multiscale instead of Glow | Image density estimation · MNIST | negative log-likelihood (bits/dim): 1.05* -> 0.99* | Improved |
| Use FFJORD multiscale instead of Glow | Image density estimation · CIFAR10 | negative log-likelihood (bits/dim): 3.35* -> 3.40* | Hurt |
| Use FFJORD instead of the strongest competing normalizing flow | VAE approximate posterior · MNIST | negative ELBO: 83.32 ± .06 -> 82.82 ± .01 | Improved |
| Use FFJORD instead of the strongest competing normalizing flow | VAE approximate posterior · Omniglot | negative ELBO: 99.00 ± .04 -> 98.33 ± .09 | Improved |
| Use FFJORD instead of the strongest competing normalizing flow | VAE approximate posterior · Frey Faces | negative ELBO (bits/dim): 4.40 ± .06 -> 4.39 ± .01 | Improved |
| Use FFJORD instead of the strongest competing normalizing flow | VAE approximate posterior · Caltech Silhouettes | negative ELBO: 104.62 ± .29 -> 104.03 ± .43 | Improved |
| Add the bottleneck trace-estimation trick (ablation) | FFJORD training with Gaussian Hutchinson noise · MNIST | training convergence speed | Improved |
Later papers that compared with this method
20 later accepted papers report FFJORD as a baseline in their results tables; 14 of them report a better number on most of the results they share with it.
- Kernelised Normalising Flows (ICLR 2024 · better on 6 of 8 shared results)
- Building Normalizing Flows with Stochastic Interpolants (ICLR 2023 · better on 4 of 5 shared results)
- Continuously Parameterized Mixture Models (ICML 2023 · better on 0 of 2 shared results)
- Structured Neural Networks for Density Estimation and Causal Inference (NeurIPS 2023 · better on 1 of 1 shared result)
- TO-FLOW: Efficient Continuous Normalizing Flows With Temporal Optimization Adjoint With Moving Speed (CVPR 2022 · better on 6 of 9 shared results)
- Semi-Discrete Normalizing Flows through Differentiable Tessellation (NeurIPS 2022 · better on 2 of 4 shared results)
- Invertible Monotone Operators for Normalizing Flows (NeurIPS 2022 · better on 2 of 2 shared results)
- Accelerating Continuous Normalizing Flow with Trajectory Polynomial Regularization (AAAI 2021 · better on 16 of 22 shared results)
12 more not listed.
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Glow — Glow: Generative Flow with Invertible 1x1 Convolutions (NeurIPS 2018) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Residual Flows for Invertible Generative Modeling (NeurIPS 2019 · similarity 0.89)
- Generative Modeling via Drifting (NeurIPS 2026 · similarity 0.87)
- Statistically Optimal Generative Modeling with Maximum Deviation from the Empirical Distribution (ICML 2024 · similarity 0.87)
- Towards Non-Asymptotic Convergence for Diffusion-Based Generative Models (ICLR 2024 · similarity 0.86)
- Multimarginal Generative Modeling with Stochastic Interpolants (ICLR 2024 · similarity 0.86)
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
