Research map / Image generation / Fast diffusion sampling
PixelPyramids: Exact Inference Models From Lossless Image Pyramids
Shweta Mahajan, Stefan Roth
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 modulo-difference fine components in the pyramid instead of a regular image pyramid (ablation) | PixelPyramids · CelebA 256x256 | bits/dim: 0.76 -> 0.61 | Improved |
| add the partial autoregressive structure (Conv-LSTM via squeeze operations) (ablation) | PixelPyramids · CelebA 256x256 | bits/dim: 0.83 -> 0.61 | Improved |
| increase n_squeeze from 0 to 1 (ablation) | PixelPyramids · CelebA 256x256 | bits/dim: 0.83 -> 0.62 | Improved |
| increase n_squeeze from 1 to 2 (ablation) | PixelPyramids · CelebA 256x256 | bits/dim: 0.62 -> 0.61 | Improved |
| increase n_squeeze from 2 to 3 (ablation) | PixelPyramids · CelebA 256x256 | bits/dim: 0.61 -> 0.61 | No clear effect |
| replace Haar wavelet scale-space flow with Paired Pyramid block-autoregressive model | WaveletFlow · CelebA-HQ 1024x1024 (8-bit) | bits/dim: 1.34 -> 0.58 | Improved |
| use PixelPyramids instead of Glow | flow baselines · CelebA-HQ 256x256 (5-bit) | bits/dim: 1.03 -> 0.61 | Improved |
| use PixelPyramids instead of MaCoW | flow baselines · CelebA-HQ 256x256 (5-bit) | bits/dim: 0.67 -> 0.61 | Improved |
| use PixelPyramids instead of WaveletFlow | WaveletFlow · CelebA-HQ 256x256 (5-bit) | bits/dim: 0.94 -> 0.61 | Improved |
| use PixelPyramids instead of fully autoregressive SPN | SPN · ImageNet 128x128 (8-bit) | bits/dim: 3.08 -> 3.40 | Hurt |
| use PixelPyramids instead of ParallelWavelet | ParallelWavelet · ImageNet 128x128 (8-bit) | bits/dim: 3.55 -> 3.40 | Improved |
| use PixelPyramids instead of PixelCNN++ on low-resolution data | PixelCNN++ · CIFAR10 32x32 (8-bit) | bits/dim: 2.92 -> 3.19 | Hurt |
Closest papers by meaning
- Parallel Multiscale Autoregressive Density Estimation (ICML 2017 · similarity 0.88)
- LaMAR: Laplacian Pyramid for Multimodal Adaptive Super Resolution (Student Abstract) (AAAI 2024 · similarity 0.84)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.84)
- NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering (NeurIPS 2025 · similarity 0.84)
- BitDance: Scaling Autoregressive Generative Models with Binary Tokens (NeurIPS 2026 · similarity 0.83)
- HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · similarity 0.83)
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.
