Research map / Probabilistic and scientific modeling / Variational and generative sampling
Generative Flows with Matrix Exponential
Changyi Xiao, Ligang Liu
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 matrix exponential coupling instead of affine coupling (ablation) | MEF · CIFAR10 | bits/dim: 3.336$ \pm $ 0.002 -> 3.324$ \pm $ 0.004 | Improved |
| Use learning rate 0.01 instead of 0.001 (ablation) | MEF with matrix exponential coupling · CIFAR10 | bits/dim: 3.381$ \pm $ 0.008 -> 3.324$ \pm $ 0.004 | Improved |
| Use matrix exponential 1×1 convolution instead of standard 1×1 convolution (ablation) | MEF · CIFAR10 | bits/dim: 3.324$ \pm $ 0.001 -> 3.324$ \pm $ 0.004 | No clear effect |
| Use matrix exponential 1×1 convolution instead of standard 1×1 convolution (ablation) | MEF · CIFAR10 | epoch time: 844.3$ \pm $ 2.6s -> 669.5$ \pm $ 10.1s | Improved |
| Use matrix exponential 1×1 convolution instead of PLU decomposition convolution (ablation) | MEF · CIFAR10 | bits/dim: 3.330$ \pm $ 0.007 -> 3.324$ \pm $ 0.004 | No clear effect |
| Use matrix exponential 1×1 convolution instead of PLU decomposition convolution (ablation) | MEF · CIFAR10 | epoch time: 668.3$ \pm $ 19.6s -> 669.5$ \pm $ 10.1s | No clear effect |
| Use MEF instead of Glow | Generative flow density estimation · CIFAR10 | bits/dim: 3.35 -> 3.32 | Improved |
| Use MEF instead of Flow++ with variational dequantization | Generative flow density estimation · CIFAR10 | bits/dim: 3.08 -> 3.32 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- Glow — Glow: Generative Flow with Invertible 1x1 Convolutions (NeurIPS 2018) · better on 3 of 3 shared results
- Flow++ — Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design (ICML 2019) · better on 0 of 1 shared result
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
- CDFlow: Building Invertible Layers with Circulant and Diagonal Matrices (NeurIPS 2025 · similarity 0.88)
- Expanding Flow Maps (NeurIPS 2026 · similarity 0.86)
- Analysis of Learning a Flow-based Generative Model from Limited Sample Complexity (ICLR 2024 · similarity 0.86)
- Diffusion Generative Flow Samplers: Improving learning signals through partial trajectory optimization (ICLR 2024 · similarity 0.86)
- Annealing Flow Generative Models Towards Sampling High-Dimensional and Multi-Modal Distributions (ICML 2025 · similarity 0.85)
- Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks (ICLR 2025 · similarity 0.85)
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.
