Research map / Probabilistic and scientific modeling / Variational and generative sampling
Diagnosing and Enhancing VAE Models
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 |
|---|---|---|---|
| add a separately trained second VAE to model the first-stage aggregated posterior (ablation) | VAE with learned γ · MNIST | FID: 54.5±1.0 -> 12.6±1.5 | Helped |
| add a separately trained second VAE to model the first-stage aggregated posterior (ablation) | VAE with learned γ · Fashion MNIST | FID: 60.0±1.1 -> 29.3±1.0 | Helped |
| add a separately trained second VAE to model the first-stage aggregated posterior (ablation) | VAE with learned γ · CIFAR-10 | FID: 76.7±0.8 -> 72.9±0.9 | Helped |
| add a separately trained second VAE to model the first-stage aggregated posterior (ablation) | VAE with learned γ · CelebA | FID: 60.5±0.6 -> 44.4±0.7 | Helped |
| use a learned decoder variance rather than fixed γ (ablation) | single-stage VAE · MNIST | FID: 52.0±0.6 -> 54.5±1.0 | No clear effect |
| use a learned decoder variance rather than fixed γ (ablation) | single-stage VAE · Fashion MNIST | FID: 84.6±0.9 -> 60.0±1.1 | Helped |
| use a learned decoder variance rather than fixed γ (ablation) | single-stage VAE · CIFAR-10 | FID: 160.5±1.1 -> 76.7±0.8 | Helped |
| use a learned decoder variance rather than fixed γ (ablation) | single-stage VAE · CelebA | FID: 55.9±0.6 -> 60.5±0.6 | No clear effect |
| train both VAE stages jointly instead of separately (ablation) | 2-Stage VAE · the paper's tested datasets | average performance | No clear effect |
| replace the Gaussian decoder output with a cross-entropy output (ablation) | single-stage VAE · MNIST | FID: 52.0±0.6 -> 16.6±0.4 | Helped |
| replace the Gaussian decoder output with a cross-entropy output (ablation) | single-stage VAE · Fashion MNIST | FID: 84.6±0.9 -> 43.6±0.7 | Helped |
| replace the Gaussian decoder output with a cross-entropy output (ablation) | single-stage VAE · CIFAR-10 | FID: 160.5±1.1 -> 106.0±1.0 | Helped |
Closest papers by meaning
- On Unifying Deep Generative Models (ICLR 2018 · similarity 0.91)
- From Variational to Deterministic Autoencoders (ICLR 2020 · similarity 0.91)
- Entropic GANs meet VAEs: A Statistical Approach to Compute Sample Likelihoods in GANs (ICML 2019 · similarity 0.89)
- Adversarial Variational Bayes: Unifying Variational Autoencoders and Generative Adversarial Networks (ICML 2017 · similarity 0.89)
- Generating Diverse High-Fidelity Images with VQ-VAE-2 (NeurIPS 2019 · similarity 0.89)
- A Primal-Dual link between GANs and Autoencoders (NeurIPS 2019 · similarity 0.89)
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.