Research map / Probabilistic and scientific modeling / Variational and generative sampling
A Tale of Two Flows: Cooperative Learning of Langevin Flow and Normalizing Flow Toward Energy-Based Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Jointly train the normalizing flow with a Langevin flow that revises its samples (ablation) | CoopFlow on CIFAR-10 · CIFAR-10 | FID: 92.10 -> 21.16; 49.51 -> 21.16 | Helped |
| Increase Langevin steps from 30 to 200 (ablation) | CoopFlow · CIFAR-10, SVHN, CelebA | FID: 21.16 -> 18.89; 18.11 -> 16.97; 6.44 -> 4.90 | Helped |
| Pretrain the normalizing flow before cooperative training (ablation) | CoopFlow · CIFAR-10, SVHN, CelebA | FID: 21.16 -> 15.80; 18.11 -> 15.32; 6.44 -> 4.15 | Helped |
| Use a pretrained normalizing flow and cooperative revision rather than Flow++ alone (ablation) | CoopFlow(Pre) · CIFAR-10 | FID: 92.10 -> 15.80 | Helped |
| Use the 10-step setting instead of 30 Langevin steps (ablation) | CoopFlow(Pre) · CIFAR-10 | FID: 15.80 -> 16.46 | Mixed |
| Vary Langevin step count from the 30-step setting (ablation) | CoopFlow(Pre) · CIFAR-10 | FID: 20: 15.20; 30: 15.80; 40: 16.80; 50: 15.64 | Mixed |
| Set training Langevin step size to 0.01 rather than 0.03 (ablation) | CoopFlow(Pre) · CIFAR-10 | FID: 15.80 -> 15.99 | Helped |
| Set training Langevin step size to 0.02 rather than 0.03 (ablation) | CoopFlow(Pre) · CIFAR-10 | FID: 15.80 -> 16.32 | Hurt |
| Increase training Langevin step size to 0.04, 0.05, or 0.10 (ablation) | CoopFlow(Pre) · CIFAR-10 | FID: 15.80 -> 16.52; 18.17; 19.82 | Hurt |
| Gradually decay Langevin noise during training instead of disabling it (ablation) | CoopFlow(Pre) · CIFAR-10 and SVHN | FID: CIFAR-10: 15.80 -> 14.55; SVHN: 15.32 -> 15.74 | Mixed |
| Use CoopFlow rather than a short-run Langevin flow alone for reconstruction (ablation) | CoopFlow · CIFAR-10 test, 1,000 images | MSE per pixel: 0.1083 -> 0.0254 | Helped |
| Initialize the Langevin revision with the learned normalizing flow rather than noise (ablation) | CoopFlow versus short-run EBM · CIFAR-10 | FID: Short-run EBM: 421.3, 194.88, 117.02, 140.79, 198.09, 54.23; CoopFlow(Pre): 16.46, 15.20, 15.80, 16.80, 15.64, 17.94 | Helped |
Closest papers by meaning
- Markovian Flow Matching: Accelerating MCMC with Continuous Normalizing Flows (NeurIPS 2024 · similarity 0.87)
- Learning Energy-Based Models by Cooperative Diffusion Recovery Likelihood (ICLR 2024 · similarity 0.87)
- Mean Flows for One-step Generative Modeling (NeurIPS 2025 · similarity 0.87)
- Learning normalized image densities via dual score matching (NeurIPS 2025 · similarity 0.87)
- Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling (NeurIPS 2025 · similarity 0.86)
- Normalizing Flows are Capable Generative Models (ICML 2025 · 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.