Research map / Probabilistic and scientific modeling / Variational and generative sampling
A Tale of Two Latent Flows: Learning Latent Space Normalizing Flow with Short-Run Langevin Flow for Approximate Inference
Jianwen Xie, Yaxuan Zhu, Yifei Xu, Dingcheng Li, Ping Li
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 latent normalizing-flow prior with short-run Langevin inference instead of LEBM | LFBM image model · SVHN | FID: 29.44 -> 23.64 | Improved |
| Use latent normalizing-flow prior with short-run Langevin inference instead of LEBM | LFBM image model · CIFAR-10 | MSE: 0.020 -> 0.016 | Improved |
| Use latent normalizing-flow prior with short-run Langevin inference instead of LEBM | LFBM image model · CelebA | FID: 37.87 -> 33.64 | Improved |
| Replace learned inference network with short-run Langevin inference (ablation) | LFBM-VAE · CelebA inpainting | MSE: 0.2074 -> 0.1937 | Improved |
| Replace learned inference network with short-run Langevin inference (ablation) | LFBM-VAE · CelebA inpainting | FID: 0.0562 -> 0.0557 | Improved |
| Use latent normalizing-flow prior instead of EBM prior | Latent prior model · CelebA unsupervised recovery, salt-and-pepper mask 30% | MSE: 0.0256 -> 0.0244 | Improved |
| Use latent normalizing-flow prior instead of Gaussian prior | Latent prior model · CelebA unsupervised recovery, salt-and-pepper mask 30% | FID: 153.01 -> 46.2 | Improved |
| Use latent normalizing-flow prior instead of Gaussian prior | Latent prior model · CelebA unsupervised recovery, single region mask 20 × 20 | MSE: 0.0404 -> 0.0420 | Hurt |
| Increase the depth of the flow-based prior (ablation) | LFBM · CIFAR-10 | FID | Improved |
| Use MCMC inference instead of amortized inference (ablation) | LFBM · CelebAMask-HQ 256 × 256 | FID: 129.65 -> 75.20 | Improved |
| Use MCMC inference instead of amortized inference (ablation) | LFBM · CelebAMask-HQ 256 × 256 | training time: 5.5 -> 20.7 | Hurt |
| Use a normalizing-flow prior rather than an EBM prior | LFBM · CelebA synthesis | synthesis time: 116 -> 105 | Improved |
Closest papers by meaning
- A Tale of Two Flows: Cooperative Learning of Langevin Flow and Normalizing Flow Toward Energy-Based Model (ICLR 2022 · similarity 0.91)
- Markovian Flow Matching: Accelerating MCMC with Continuous Normalizing Flows (NeurIPS 2024 · similarity 0.88)
- Latent Bayesian Optimization via Autoregressive Normalizing Flows (ICLR 2025 · similarity 0.88)
- Normalizing Flows are Capable Generative Models (ICML 2025 · similarity 0.87)
- Annealing Flow Generative Models Towards Sampling High-Dimensional and Multi-Modal Distributions (ICML 2025 · similarity 0.87)
- Normalizing Trajectory Models (NeurIPS 2026 · 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.
