Research map / Architectures and efficiency / Quantization
Differentiable Vector Quantization for Rate-Distortion Optimization of Generative Image Compression
Shiyin Jiang, Wei Long, Minghao Han, Zhenghao Chen, Ce Zhu, Shuhang Gu
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 |
|---|---|---|---|
| add differentiable relaxation of the index distribution for the rate loss (ablation) | RDVQ · DIV2K-val | DISTS: 0.2147 -> 0.1005 | Improved |
| add differentiable relaxation of the index distribution for the rate loss (ablation) | RDVQ · DIV2K-val | FID: 86.93 -> 19.96 | Improved |
| add differentiable relaxation of the index distribution for the rate loss (ablation) | RDVQ · DIV2K-val | LPIPS: 0.5031 -> 0.2321 | Improved |
| add differentiable relaxation of the index distribution for the rate loss (ablation) | RDVQ · DIV2K-val | bpp: 0.0464 -> 0.0247 | Improved |
| use joint RD learning instead of K-means codebook-size rate control (ablation) | RDVQ · DIV2K-val | DISTS: 0.1253 -> 0.1005 | Improved |
| use joint RD learning instead of K-means codebook-size rate control (ablation) | RDVQ · DIV2K-val | FID: 28.08 -> 19.96 | Improved |
| use joint RD learning instead of K-means codebook-size rate control (ablation) | RDVQ · DIV2K-val | LPIPS: 0.2831 -> 0.2321 | Improved |
| jointly optimize autoencoder and entropy model (vs separate AE-Entropy) for prefix-based rate adjustment (ablation) | RDVQ-Adj · Kodak (test-time rate adjustment) | R-D curve / DISTS | Improved |
| use autoregressive completion instead of zero-padding for missing features (ablation) | RDVQ-Adj · Kodak (test-time rate adjustment) | R-D curve / DISTS | Improved |
| use piecewise tau/lambda (tau=0.1 at ultra-low bpp) instead of fixed sharp tau=0.01 (ablation) | RDVQ · DIV2K-val | RD performance | Improved |
| add high-resolution fine-tuning on OpenImage then DF2K after ImageNet (ablation) | RDVQ · DIV2K-val | reconstruction quality | Improved |
Closest papers by meaning
- HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression (NeurIPS 2026 · similarity 0.93)
- Multirate Neural Image Compression with Adaptive Lattice Vector Quantization (CVPR 2025 · similarity 0.87)
- Efficient Generative Modeling with Residual Vector Quantization-Based Tokens (ICML 2025 · similarity 0.87)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.87)
- Learning Optimal Lattice Vector Quantizers for End-to-end Neural Image Compression (NeurIPS 2024 · similarity 0.87)
- VRVVC: Variable-Rate NeRF-Based Volumetric Video Compression (AAAI 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.
