Research map / Probabilistic and scientific modeling / Variational and generative sampling
Pythae: Unifying Generative Autoencoders in Python - A Benchmarking Use Case
Clément Chadebec, Louis J. Vincent, Stephanie Allassonniere
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 |
|---|---|---|---|
| Fit a 10-component GMM on latent codes instead of sampling the N(0,I) prior (ablation) | AE · MNIST ConvNet generation | FID: 26.7 -> 9.3 | Improved |
| Fit a GMM instead of N(0,I) prior sampling (ablation) | beta-VAE · MNIST ConvNet generation | FID: 21.4 -> 9.2 | Improved |
| Fit a GMM instead of N(0,I) prior sampling (ablation) | VAE · MNIST ConvNet generation | FID: 28.5 -> 26.9 | Improved |
| Replace GMM with MAF normalising flow sampler (ablation) | VAE · MNIST ConvNet generation | FID: 26.9 -> 26.8 | No clear effect |
| Use 2-stage VAE sampler instead of N(0,I) prior (ablation) | VAE · MNIST ConvNet generation | FID: 28.5 -> 40.3 | Hurt |
| Use VAMP prior instead of standard Gaussian prior | VAE · MNIST ConvNet generation | FID: 28.5 -> 64.2 | Hurt |
| Use small-beta beta-VAE instead of standard VAE | VAE · MNIST ConvNet reconstruction | MSE (x1e-3): 16.85 -> 5.61 | Improved |
| Add L2 penalty on decoder weights (RAE-L2) to AE | AE · CIFAR10 ConvNet reconstruction | MSE (x1e-3): 2.82 -> 2.25 | Improved |
| Use beta-TC VAE instead of standard VAE | VAE · CIFAR10 ConvNet classification | accuracy: 32.61 -> 45.40 | Improved |
| Use autoencoder instead of VAE | VAE · MNIST ConvNet classification | accuracy: 86.75 -> 93.81 | Improved |
| Use more flexible VAMP prior instead of standard prior | VAE · MNIST ConvNet classification | accuracy: 86.75 -> 92.17 | Improved |
| Use VAEGAN adversarial reconstruction loss instead of MSE loss | VAE · MNIST ConvNet reconstruction | FID: 30.71 -> 5.54 | Improved |
Closest papers by meaning
- Perceptual Generative Autoencoders (ICML 2020 · similarity 0.85)
- From Variational to Deterministic Autoencoders (ICLR 2020 · similarity 0.85)
- A Primal-Dual link between GANs and Autoencoders (NeurIPS 2019 · similarity 0.84)
- Scaling Tractable Probabilistic Circuits: A Systems Perspective (ICML 2024 · similarity 0.83)
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.83)
- JetFormer: An autoregressive generative model of raw images and text (ICLR 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.
