Research map / Image generation / Autoregressive and latent generation
ControlVAE: Controllable Variational Autoencoder
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 feedback PI control to tune the KL weight and target KL 35 | Conditional VAE dialog generation · Switchboard (SW) | Dis-1 | Helped |
| Use feedback PI control to tune the KL weight and target KL 35 | Conditional VAE dialog generation · Switchboard (SW) | PPL | Helped |
| Use feedback PI control to tune the KL weight and target KL 16 | VAE disentanglement model · 2D Shapes | MIG | No clear effect |
| Dynamically adjust the KL weight rather than keep it fixed | VAE disentanglement model · 2D Shapes | reconstruction error | Helped |
| Use feedback PI control to tune the KL weight and target KL 200 | VAE image generation model · CelebA | FID: 58.71 ± 0.207 -> 55.16 ± 0.187 | Helped |
| Use feedback PI control to tune the KL weight and target KL 200 | VAE image generation model · CelebA | SSIM: 0.675 ± 0.0001 -> 0.687 ± 0.0002 | Helped |
| Set dialog-generation KL target to 25 instead of 35 (ablation) | ControlVAE conditional VAE · Switchboard (SW) | PPL: 8.81 ± 0.05 -> 12.47 ± 0.07 | Hurt |
| Set dialog-generation KL target to 25 instead of 35 (ablation) | ControlVAE conditional VAE · Switchboard (SW) | Dis-1: 6.27K ± 41 -> 6.10K ± 60 | Hurt |
| Increase the disentanglement KL target from 16 to 18 (ablation) | ControlVAE · 2D Shapes | MIG: 0.5628 ± 0.0222 -> 0.5432 ± 0.0281 | Hurt |
| Reduce the image-generation KL target from 200 to 180 (ablation) | ControlVAE · CelebA | FID: 55.16 ± 0.187 -> 57.57 ± 0.236 | Hurt |
| Reduce the image-generation KL target from 200 to 170 (ablation) | ControlVAE · CelebA | SSIM: 0.687 ± 0.0002 -> 0.675 ± 0.0001 | Hurt |
Closest papers by meaning
- Rethinking Controllable Variational Autoencoders (CVPR 2022 · similarity 0.94)
- Variational Autoencoders Pursue PCA Directions (by Accident) (CVPR 2019 · similarity 0.87)
- Property Controllable Variational Autoencoder via Invertible Mutual Dependence (ICLR 2021 · similarity 0.87)
- Guided Variational Autoencoder for Disentanglement Learning (CVPR 2020 · similarity 0.87)
- From Variational to Deterministic Autoencoders (ICLR 2020 · similarity 0.86)
- Beyond Vanilla Variational Autoencoders: Detecting Posterior Collapse in Conditional and Hierarchical Variational Autoencoders (ICLR 2024 · similarity 0.85)
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.