Research map / Architectures and efficiency / Quantization
QARV: Quantization-Aware ResNet VAE for Lossy Image Compression
Zhihao Duan, Ming Lu, Jack Ma, Yuning Huang, Zhan Ma, Fengqing Maggie Zhu
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 |
|---|---|---|---|
| place affine transform after layer norm (position 2, AdaLN) instead of position 0 (ablation) | QARV · Kodak | BD-rate (%): 2.446 -> -0.066 | Improved |
| AdaLN position 2 vs position 1 (ablation) | QARV · Kodak | BD-rate (%): 2.005 -> -0.066 | Improved |
| AdaLN position 2 vs position 4 (ablation) | QARV · Kodak | BD-rate (%): 5.365 -> -0.066 | Improved |
| use Layer Norm instead of Group Norm (ablation) | QARV · Kodak | BD-rate (%): 1.041 -> -0.066 | Improved |
| use Layer Norm instead of Instance Norm (ablation) | QARV · Kodak | BD-rate (%): 10.26 -> -0.066 | Improved |
| add bi-directional inference to latent block (config B vs A), N=2 (ablation) | QARV · Kodak (N=2) | BD-rate (%): 24.96 -> 18.20 | Improved |
| add residual connection to latent block (config C vs B), N=2 (ablation) | QARV · Kodak (N=2) | BD-rate (%): 18.20 -> 14.06 | Improved |
| add bi-directional inference (B vs A), N=3 (ablation) | QARV · Kodak (N=3) | BD-rate (%): 22.51 -> 9.076 | Improved |
| add residual connection (C vs B), N=3 (ablation) | QARV · Kodak (N=3) | BD-rate (%): 9.076 -> 3.761 | Improved |
| increase number of latent variables from N=2 to N=3 with config A (ablation) | Hyperprior-style config A · Kodak | BD-rate (%): 24.96 -> 22.51 | Improved |
| use ConvNeXt instead of NAT residual blocks (ablation) | QARV · Kodak | BD-rate (%): 2.916 -> -0.066 | Improved |
| sample lambda by 1/3-power uniform instead of log-uniform (ablation) | QARV · Kodak | BD-rate (%): -4.28 -> -4.74 | Improved |
Closest papers by meaning
- Hierarchical Quantized Autoencoders (NeurIPS 2020 · similarity 0.92)
- Improving Inference for Neural Image Compression (NeurIPS 2020 · similarity 0.91)
- HyperVQ: Enabling Hyperprior Entropy Modeling for VQ-Based Generative Image Compression (NeurIPS 2026 · similarity 0.88)
- Restructuring Vector Quantization with the Rotation Trick (ICLR 2025 · similarity 0.88)
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.87)
- Information-theoretic Generalization Analysis for VQ-VAEs: A Role of Latent Variables (NeurIPS 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.
