Research map / Probabilistic and scientific modeling / Variational and generative sampling
Closing the Dequantization Gap: PixelCNN as a Single-Layer Flow
Didrik Nielsen, Ole Winther
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 |
|---|---|---|---|
| Train with ELBO instead of exact likelihood (evaluated exactly) (ablation) | PixelCNN++ flow · CIFAR-10 | bits/dim: 2.924 -> 2.944 | Hurt |
| Train with ELBO instead of exact likelihood (evaluated exactly) (ablation) | PixelCNN (Quad.) flow · CIFAR-10 | bits/dim: 3.090 -> 3.104 | Hurt |
| Train with ELBO instead of exact likelihood (ablation) | PixelCNN (bin cond.) · CIFAR-10 | bits/dim: 3.141 -> 3.141 | No clear effect |
| Remove bin conditioning (uniform dequantization) (ablation) | PixelCNN++ flow · CIFAR-10 | bits/dim: 2.924 -> 3.112 | Hurt |
| Remove bin conditioning (uniform dequantization) (ablation) | PixelCNN flow · CIFAR-10 | bits/dim: 3.141 -> 3.248 | Hurt |
| Evaluate with IWBO(1000) instead of ELBO (ablation) | PixelCNN++ (bin cond., ELBO-trained) · CIFAR-10 | bits/dim: 2.993 -> 2.974 | Improved |
| Evaluate exact likelihood instead of ELBO (ablation) | PixelCNN++ (bin cond., ELBO-trained) · CIFAR-10 | bits/dim: 2.993 -> 2.944 | Improved |
| Stack 2 PixelCNN layers instead of 1 (ablation) | PixelCNN · CIFAR-10 | bits/dim: 3.14 -> 3.07 | Improved |
| Stack 4 PixelCNN layers without dropout (ablation) | PixelCNN · CIFAR-10 | bits/dim: 3.07 -> 3.09 | No clear effect |
| Add dropout 0.2 to 4-layer stack (ablation) | PixelCNN · CIFAR-10 | bits/dim: 3.09 -> 3.02 | Improved |
| Use quadratic splines instead of linear splines (ablation) | PixelCNN · CIFAR-10 | bits/dim: 3.14 -> 3.09 | Improved |
| Combine PixelCNN (Quad.) with PixelCNN++ in a two-layer flow (ablation) | PixelCNN++ · CIFAR-10 | bits/dim: 2.924 -> 2.906 | Improved |
Closest papers by meaning
- Discrete Flow Matching (NeurIPS 2024 · similarity 0.87)
- Asymmetric Flow Models (NeurIPS 2026 · similarity 0.86)
- FlowDiffuser: Advancing Optical Flow Estimation with Diffusion Models (CVPR 2024 · similarity 0.85)
- Normalizing Flows are Capable Generative Models (ICML 2025 · similarity 0.85)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.85)
- CDFlow: Building Invertible Layers with Circulant and Diagonal Matrices (NeurIPS 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.
