Research map / Probabilistic and scientific modeling / Variational and generative sampling
Deterministic training of generative autoencoders using invertible layers
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 |
|---|---|---|---|
| Replace VAE training with AEF (exact likelihood in expanded ambient space), architecture otherwise identical | Encoder-decoder models with IAF posterior and MAF prior · CelebA 64x64, 64-256 latents, CelebA 32x32, ImageNet 32x32 | BPD and FID: e.g. CelebA 64x64 (256 latents) FID 90.9 -> 55.8 | Helped |
| Use AEF instead of VAE on MNIST with 32 latents | Architecture-matched AEF vs VAE · MNIST 32 latents | BPD and FID: BPD 1.80 -> 1.82; FID 16.2 -> 17.7 | Hurt |
| Use AEF instead of VAE on MNIST with 2 latents | Small conv encoder-decoder · MNIST 2 latents | BPD and FID: BPD 2.40 -> 2.37; FID 57.5 -> 55.8 | Helped |
| Add posterior flow only (no prior flow) to AEF (ablation) | AEF on CelebA 32x32 · CelebA 32x32, 128 latents | FID: FID 64.4 (no flows) -> 108 | Hurt |
| Add prior flow to AEF (ablation) | AEF on CelebA 32x32 · CelebA 32x32, 128 latents | FID: FID 64.4 -> 37.6 (only prior); 35.6 with both flows | Helped |
| Add prior flow to VAE (ablation) | VAE on CelebA 32x32 · CelebA 32x32, 128 latents | FID: FID 77.2 -> 74.1 (only prior); 68.1 with both flows | Helped |
| Add posterior flow only to VAE (ablation) | VAE on CelebA 32x32 · CelebA 32x32, 128 latents | FID: FID 77.2 -> 81.5 | Hurt |
| Remove the core encoder and prior flow from AEF (AEF without flows) vs VAE with both flows (ablation) | AEF vs VAE · CelebA 32x32, 128 latents | BPD and FID: AEF (no flows) 5.90 BPD / 64.4 FID vs VAE (both flows) 6.67 / 68.1 | Helped |
| Use AEF with linear core in expanded space instead of VAE, trained with 2-run / 5-run means | CelebA 32x32 and 64x64 generative modeling · CelebA 32x32 (256 latents) | FID: 63.5 -> 30.9 | Helped |
Closest papers by meaning
- From Variational to Deterministic Autoencoders (ICLR 2020 · similarity 0.91)
- Shape your Space: A Gaussian Mixture Regularization Approach to Deterministic Autoencoders (NeurIPS 2021 · similarity 0.89)
- Decoupling Global and Local Representations via Invertible Generative Flows (ICLR 2021 · similarity 0.89)
- InfoVAE: Balancing Learning and Inference in Variational Autoencoders (AAAI 2019 · similarity 0.88)
- Analysis of Learning a Flow-based Generative Model from Limited Sample Complexity (ICLR 2024 · similarity 0.87)
- Generalization in VAE and Diffusion Models: A Unified Information-Theoretic Analysis (ICLR 2025 · 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.