Research map / Architectures and efficiency / Quantization
DynaQuant: Dynamic Mixed-Precision Quantization for Learned Image Compression
Youneng Bao, Yulong Cheng, Yiping Liu, Yichen Yang, Peng Qin, Mu Li, Yongsheng Liang
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 learnable quantization (DPA with DGM) instead of PAMS at 8 bits (ablation) | Cheng2020 · Kodak | PSNR (dB): 36.185 -> 36.649 | Improved |
| remove learnable scale s from DPA (ablation) | DPA · Kodak | PSNR (dB): 36.649 -> 36.185 | Improved |
| add learnable zero-point z to DPA (ablation) | DPA · Kodak | PSNR (dB): 36.323 -> 36.649 | Improved |
| add distance-aware gradient g(x) in place of plain STE (ablation) | DPA · Kodak | PSNR (dB): 36.288 -> 36.649 | Improved |
| add dynamic bit-width selector to DPA (ablation) | Cheng2020 DPA · Kodak | avg bit-width / PSNR: 8 -> 6.42 bits; PSNR 36.649 -> 36.636 | Mixed |
| use DPA quantizer instead of PAMS in the dynamic bit-width model (ablation) | dynamic bit-width selector · Kodak | PSNR (dB): 30.262 -> 36.231 | Improved |
| lower fixed bit-width from 8 to 6 in DPA (ablation) | DPA · Kodak | PSNR (dB): 36.649 -> 35.664 | Hurt |
| use candidate bit set {6,8,10} instead of {4,6,8} (ablation) | dynamic bit-width selector · Kodak | PSNR (dB): 36.432 -> 36.636 | Improved |
| use dynamic bit-width selection instead of fixed 8-bit DPA | Cheng2020 · Kodak/JPEG-AI/CLIC average | BD-Rate loss (%): 1.60 -> 12.18 | Hurt |
| use dynamic bit-width selection instead of fixed 8-bit DPA | Cheng2020 · Kodak/JPEG-AI/CLIC average | speedup: 4.00x -> 5.17x | Improved |
Closest papers by meaning
- One-Shot Model for Mixed-Precision Quantization (CVPR 2023 · similarity 0.88)
- DQT: Dynamic Quantization Training via Dequantization-Free Nested Integer Arithmetic (AAAI 2026 · similarity 0.87)
- When the Teacher Has More Bits: Self-Teacher Latent Distillation for Learned Image Compression (ECCV 2026 · similarity 0.87)
- Retraining-Free Model Quantization via One-Shot Weight-Coupling Learning (CVPR 2024 · similarity 0.87)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.87)
- MetaMix: Meta-State Precision Searcher for Mixed-Precision Activation Quantization (AAAI 2024 · 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.
