Research map / Probabilistic and scientific modeling / Variational and generative sampling
SQ-VAE: Variational Bayes on Discrete Representation with Self-annealed Stochastic Quantization
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 |
|---|---|---|---|
| Trainable variance (self-annealed stochastic quantization) vs fixed sigma_q^2 quantization | Gaussian SQ-VAE (I) · CelebA 64x64 | MSE (x10^3): 1.09 -> 0.96 | Helped |
| Replace EMA VQ-VAE with SQ-VAE (Gaussian I) under same bit budget | VQ-VAE · CelebA 64x64 | rFID: 18.5 -> 14.8 | Helped |
| Naive categorical decoder replaced by vMF decoder with vMF dequantization (ablation) | NC SQ-VAE · CelebAHQ-Mask | perplexity: 12.6 -> 52.4 | Helped |
| Naive categorical SQ-VAE vs vMF SQ-VAE on CelebAHQ-Mask pixel error (ablation) | NC SQ-VAE · CelebAHQ-Mask | pixel error (%): 6.63 -> 3.51 | Helped |
| Naive categorical SQ-VAE vs vMF SQ-VAE on MNIST MSE (ablation) | NC SQ-VAE · MNIST | MSE (x10^3): 10.89 -> 1.63 | Helped |
| Naive categorical decoder (NC SQ-VAE) vs VQ-VAE on MNIST MSE | NC SQ-VAE · MNIST | MSE (x10^3): 6.24 -> 10.89 | Hurt |
| Naive categorical SQ-VAE on CelebAHQ-Mask vs VQ-VAE pixel error | NC SQ-VAE · CelebAHQ-Mask | perplexity: 46.2 -> 12.6 | Hurt |
| Gaussian SQ-VAE with the most complex variance parameterization (IV) | Gaussian SQ-VAE · CelebA 64x64 | training stability | Hurt |
| Gaussian SQ-VAE (IV) vs VQ-VAE with EMA on speech MSE | Gaussian SQ-VAE (IV) · VCTK and ZeroSpeech 2019 | MSE (dB^2): 29.59 -> 24.68 (VCTK); 34.33 -> 32.32 (ZeroSpeech) | Helped |
| Trainable variance parameterization (III) vs (I) on speech and images (ablation) | Gaussian SQ-VAE (III) · ZeroSpeech 2019 | MSE (dB^2): 33.17 -> 34.35 | Hurt |
| Gaussian SQ-VAE parameterization (III) vs (I) on CelebA rFID (ablation) | Gaussian SQ-VAE (III) · CelebA 64x64 | rFID: 14.8 -> 13.9 | Helped |
| Gaussian SQ-VAE parameterization (II) vs (I) on CelebA generation FID (ablation) | Gaussian SQ-VAE (II) · CelebA 64x64 | FID (generation): 28.2 -> 27.7 | No clear effect |
Closest papers by meaning
- Training-Free Vector Quantization via Gaussian VAEs (ICML 2026 · similarity 0.90)
- Neural Discrete Representation Learning (NeurIPS 2017 · similarity 0.90)
- Variational Bayesian Quantization (ICML 2020 · similarity 0.90)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.90)
- Restructuring Vector Quantization with the Rotation Trick (ICLR 2025 · similarity 0.89)
- Information-theoretic Generalization Analysis for VQ-VAEs: A Role of Latent Variables (NeurIPS 2025 · similarity 0.88)
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.