Research map / Architectures and efficiency / Quantization
DiVeQ: Differentiable Vector Quantization Using the Reparameterization Trick
Mohammad Hassan Vali, Tom Bäckström, Arno Solin
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 |
|---|---|---|---|
| Use DiVeQ instead of STE at high lr/batch size | VQGAN · CelebA-HQ 8-bit FID (lr 2.5e-4, bs 32) | FID: 334 -> 8.44 | Improved |
| Use SF-DiVeQ instead of ST-GS at high lr/batch size | VQGAN · CelebA-HQ 12-bit FID | FID: 155 -> 7.4 | Improved |
| Use SF-DiVeQ instead of DiVeQ | VQGAN · CelebA-HQ FID (lr 2.5e-4, bs 32, 12-bit) | FID: 9.54 -> 7.4 | Improved |
| Use DiVeQ instead of NSVQ | VQGAN · CelebA-HQ 8-bit FID (lr 2.5e-5, bs 8) | FID: 81.5 -> 5.90 | Improved |
| Use DiVeQ/SF-DiVeQ instead of STE at small low-lr setting | VQGAN · CelebA-HQ FID at 10-bit (lr 2.5e-5) | FID: STE 5.28 vs DiVeQ 6.32, SF-DiVeQ 5.57 | Mixed |
| Use SF-DiVeQ instead of STE for speech coding | DAC · VCTK 11-bit LSD | Log spectral distance: 3.20 -> 1.02 | Improved |
| Use SF-DiVeQ instead of EMA | DAC · VCTK 13-bit PESQ | PESQ: 1.67 -> 1.75 | Improved |
| Use DiVeQ instead of EMA | DAC · VCTK 13-bit MFCC distance | MFCC distance: 69.1 vs 72.6 | No clear effect |
| Vary noise variance sigma^2 when <= 1e-2 (ablation) | DiVeQ/SF-DiVeQ · VQ-VAE compression, VQGAN | quality | No clear effect |
| Use proposed codebook replacement instead of NSVQ replacement (ablation) | DiVeQ · CelebA-HQ VQ-VAE reconstruction | reconstruction quality | Improved |
| Random instead of custom init for SF-DiVeQ (ablation) | SF-DiVeQ · VQ-VAE compression | reconstruction | No clear effect |
Closest papers by meaning
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.89)
- Restructuring Vector Quantization with the Rotation Trick (ICLR 2025 · similarity 0.89)
- Learning to Quantize for Training Vector-Quantized Networks (ICML 2025 · similarity 0.87)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.87)
- Enhancing Diversity for Data-free Quantization (CVPR 2025 · similarity 0.86)
- VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025 · similarity 0.86)
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.
