Research map / Image generation / Fast diffusion sampling
Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images
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 |
|---|---|---|---|
| Increase stochastic depth while keeping parameter count fixed (ablation) | Very Deep VAE · ImageNet-32 | Test Loss: 4.30 -> 3.95 | Helped |
| Allocate more of the 48 stochastic layers to higher resolutions (ablation) | Very Deep VAE · ImageNet-32 | Test Loss: 3.98 -> 3.95 | Helped |
| Add residual scaling of the final residual-block layer by 1/√N (ablation) | Very Deep VAE · CIFAR-10, depth 75 | Loss: Diverged -> 2.28 | Helped |
| Add residual scaling of the final residual-block layer by 1/√N (ablation) | Very Deep VAE · CIFAR-10, depth 45 | Loss: 2.31 -> 2.30 | Helped |
| Add residual scaling of the final residual-block layer by 1/√N (ablation) | Very Deep VAE · CIFAR-10, depth 60 | Loss: 2.30 -> 2.29 | Helped |
| Add residual scaling of the final residual-block layer by 1/√N (ablation) | Very Deep VAE · CIFAR-10, depth 15 | Loss: 2.50 -> 2.51 | Hurt |
| Add residual scaling of the final residual-block layer by 1/√N (ablation) | Very Deep VAE · CIFAR-10, depth 30 | Loss: 2.36 -> 2.38 | Hurt |
| Use nearest-neighbor upsampling instead of learned convolutional upsampling (ablation) | Very Deep VAE · ImageNet-32 | Loss | Helped |
| Keep residual connections in the top-down residual blocks instead of removing them (ablation) | Very Deep VAE · ImageNet-32 | Loss | Helped |
| Use gradient skipping above a threshold to avoid destabilizing updates (ablation) | Very Deep VAE · Training across datasets | Training stability | Helped |
Closest papers by meaning
- NVAE: A Deep Hierarchical Variational Autoencoder (NeurIPS 2020 · similarity 0.93)
- PixelVAE: A Latent Variable Model for Natural Images (ICLR 2017 · similarity 0.90)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.89)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.87)
- Scalable Pre-training of Large Autoregressive Image Models (ICML 2024 · similarity 0.87)
- CV-VAE: A Compatible Video VAE for Latent Generative Video Models (NeurIPS 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.