Research map / Probabilistic and scientific modeling / Variational and generative sampling
Densely connected normalizing flows
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 |
|---|---|---|---|
| use DenseFlow-74-10 instead of best prior models on ImageNet | density estimation · ImageNet 32x32 | bits/dim: 3.77 -> 3.63 | Helped |
| use DenseFlow-74-10 instead of best prior models on ImageNet64 | density estimation · ImageNet 64x64 | bits/dim: 3.43 -> 3.35 | Helped |
| add cross-unit coupling (augmentation + preconditioned noise) to standard Glow modules (ablation) | DenseFlow-45-6 · CIFAR-10 | bits/dim: 3.42 -> 3.37 | Helped |
| add cross-unit coupling with intra-module coupling (ablation) | DenseFlow-45-6 · CIFAR-10 | bits/dim: 3.14 -> 3.07 | Helped |
| precondition noise instead of white noise in augmentation (ablation) | DenseFlow-45-6 · CIFAR-10 | bits/dim: 3.40 -> 3.37 | Helped |
| precondition noise instead of white noise (with intra-module coupling) (ablation) | DenseFlow-45-6 · CIFAR-10 | bits/dim: 3.08 -> 3.07 | Helped |
| add latent variable augmentation with white noise (ablation) | standard Glow · CIFAR-10 | bits/dim: 3.42 -> 3.40 | Helped |
| add intra-module coupling fusing dense block and self-attention (ablation) | standard Glow · CIFAR-10 | bits/dim: 3.42 -> 3.14 | Helped |
| use densely connected cross-coupling instead of conditioning on previous representation only (ablation) | DenseFlow-45-6 · CIFAR-10 | bits/dim: 0.01 bpd | Helped |
| add Nystrom self-attention to DenseBlock coupling (ablation) | DenseFlow-45-6 · ImageNet 32x32 | bits/dim: 0.01 bpd | Helped |
| use DenseFlow instead of NVAE/VFlow for training cost | training budget · CIFAR-10 | training GPUs: 1 GPU vs 16 GPUs (VFlow) | Helped |
Closest papers by meaning
- CDFlow: Building Invertible Layers with Circulant and Diagonal Matrices (NeurIPS 2025 · similarity 0.89)
- Representational aspects of depth and conditioning in normalizing flows (ICML 2021 · similarity 0.89)
- Normalizing Flows are Capable Generative Models (ICML 2025 · similarity 0.89)
- On the Universality of Volume-Preserving and Coupling-Based Normalizing Flows (ICML 2024 · similarity 0.88)
- Kernelised Normalising Flows (ICLR 2024 · similarity 0.87)
- Lifting Architectural Constraints of Injective Flows (ICLR 2024 · similarity 0.87)
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.