Research map / Probabilistic and scientific modeling / Variational and generative sampling
Simple and Effective VAE Training with Calibrated Decoders
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 analytically estimated shared decoder variance instead of a manually tuned β | VAE · SVHN | FID: 27.93 -> 22.25 | Helped |
| Use analytically estimated shared decoder variance instead of gradient-trained shared variance (ablation) | VAE · CIFAR HVAE | -log p: <-5435 -> <-5677 | Helped |
| Use analytically estimated shared decoder variance instead of gradient-trained shared variance (ablation) | VAE · SVHN | FID: 22.25 -> 27.25 | Hurt |
| Use analytically estimated shared decoder variance instead of gradient-trained shared variance (ablation) | HVAE · CelebA | -log p: <-6374 -> <-8446 | Helped |
| Use a per-pixel variance rather than one shared variance (ablation) | VAE · SVHN | FID: 22.25 -> 114.7 | Hurt |
| Use a per-pixel variance rather than one shared variance (ablation) | HVAE · CelebA | FID: 60.7 -> 159.3 | Hurt |
| Use an optimal per-image variance rather than a shared variance (ablation) | VAE · CIFAR | FID: 137.8 -> 89.88 | Helped |
| Use an optimal per-image variance rather than a shared variance (ablation) | VAE · Frey Face | FID: 42.86 -> 56.07 | Hurt |
| Use a learned shared variance instead of fixed unit variance (ablation) | VAE · SVHN | FID: 112.5 -> 22.25 | Helped |
| Use a Bernoulli decoder for color intensities (ablation) | VAE · CIFAR HVAE | FID: 101.4 -> 284.5 | Hurt |
| Use a full categorical decoder instead of a Gaussian σ decoder (ablation) | VAE · CIFAR | FID: 136.7 vs 137.6 | Mixed |
Closest papers by meaning
- From Variational to Deterministic Autoencoders (ICLR 2020 · similarity 0.90)
- Embrace the Gap: VAEs Perform Independent Mechanism Analysis (NeurIPS 2022 · similarity 0.90)
- Adversarial Symmetric Variational Autoencoder (NeurIPS 2017 · similarity 0.90)
- Support is All You Need for Certified VAE Training (ICLR 2025 · similarity 0.89)
- Beyond Vanilla Variational Autoencoders: Detecting Posterior Collapse in Conditional and Hierarchical Variational Autoencoders (ICLR 2024 · similarity 0.89)
- Information-theoretic Generalization Analysis for VQ-VAEs: A Role of Latent Variables (NeurIPS 2025 · 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.