Research map / Architectures and efficiency / Quantization
Condition Number Based Low-Bit Quantization for Image Super-Resolution
Kai Liu, Dehui Wang, Zhiteng Li, Zheng Chen, Yong Guo, Linghe Kong
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 condition number minimization of weights before quantization (ablation) | 2DQuant-based PTQ of SwinIR-light · Set5 x2 (2-bit) | PSNR (dB): 37.08 -> 37.15 | Improved |
| use CondiQuant pipeline vs 2DQuant | SwinIR-light 2-bit · Set5 x2 | PSNR (dB): 36.00 -> 37.15 | Improved |
| use CondiQuant vs 2DQuant | SwinIR-light 2-bit · Set5 x4 | PSNR (dB): 29.53 -> 30.64 | Improved |
| quantize only activations vs only weights (ablation) | SwinIR-light · Set5 x2 | PSNR (dB): 32.0117 (weights only) vs 31.3304 (activations only) | Hurt |
| vary gradient step size eta from 1e-4 to 1e-1 (ablation) | CondiQuant · Set5 x2 (2-bit) | PSNR (dB): 37.08 - 37.15 | No clear effect |
| choose target value t as median or midpoint instead of mean (ablation) | CondiQuant · Set5 x2 (2-bit) | PSNR (dB): 37.15 -> 37.12 / 37.11 | No clear effect |
| vary regularization lambda from 0.001 to 0.010 (ablation) | CondiQuant · Set5 x2 (2-bit) | PSNR (dB): 37.15 (0.003) vs 37.08 (0.010) | No clear effect |
| apply proximal condition number reduction | SwinIR-light x2 weights · average condition number | condition number: nearly 3k -> 28.37 | Improved |
| skip CondiQuant on FC1 and FC2 layers (ablation) | CondiQuant · condition number analysis | efficiency | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- 2DQuant — 2DQuant: Low-bit Post-Training Quantization for Image Super-Resolution (NeurIPS 2024) · better on 13 of 13 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- 2DQuant: Low-bit Post-Training Quantization for Image Super-Resolution (NeurIPS 2024 · similarity 0.93)
- QuantSR: Accurate Low-bit Quantization for Efficient Image Super-Resolution (NeurIPS 2023 · similarity 0.93)
- AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR 2024 · similarity 0.87)
- PassionSR: Post-Training Quantization with Adaptive Scale in One-Step Diffusion based Image Super-Resolution (CVPR 2025 · similarity 0.87)
- Thinking in Granularity: Dynamic Quantization for Image Super-Resolution by Intriguing Multi-Granularity Clues (AAAI 2025 · similarity 0.86)
- Binary Quadratic Quantization: Beyond First-Order Quantization for Real-Valued Matrix Compression (NeurIPS 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.
