Research map / Architectures and efficiency / Quantization
2DQuant: Low-bit Post-Training Quantization for Image Super-Resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes 2DQuant, a two-stage post-training quantization method for transformer-based image super-resolution (SwinIR-light) combining MSE-based bound initialization (DOBI) with distillation-based calibration (DQC).
Problem
Low-bit post-training quantization of transformer-based image super-resolution models degrades quality because weight and activation distributions are asymmetric, long-tailed and unlike the CNN models studied previously.
Approach
A coarse-to-fine two-stage PTQ: DOBI searches clipping bounds per quantizer using distribution-specific MSE search (symmetric narrowing for bell-shaped data, fixed-min right-bound search for exponential-like data), then DQC fine-tunes the bounds by distilling output and intermediate features from the FP model with an STE-based gradient.
Key results
- At 2-bit, x2 Set5, 2DQuant reaches 36.00 dB PSNR vs 34.55 dB for DBDC+Pac, the paper's reported gain up to 4.52 dB on Set5 (x2) relative to SOTA.
- At 4-bit x2 Set5, 2DQuant achieves 37.87 dB, only 0.28 dB below the FP SwinIR-light (38.15 dB).
- Reported compression ratios of 3.07x, 3.31x, 3.60x and speedups of 3.99x, 4.47x, 5.08x for 4, 3, 2 bits.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DOBI and DQC together to the quantization pipeline (ablation) | SwinIR-light 4-bit PTQ · Set5 ×2 | PSNR: 34.39 -> 37.87 | Helped |
| Add DOBI initialization before DQC (ablation) | SwinIR-light 4-bit PTQ · Set5 ×2 | PSNR: 37.32 -> 37.87 | Helped |
| Add DQC after DOBI (ablation) | SwinIR-light 4-bit PTQ · Set5 ×2 | PSNR: 37.44 -> 37.87 | Helped |
| Set DQC learning rate to 10^-1 instead of 10^-2 (ablation) | 2DQuant · Set5 ×2 | PSNR: 37.87 -> 37.82 | Hurt |
| Set DQC learning rate to 10^-3 instead of 10^-2 (ablation) | 2DQuant · Set5 ×2 | PSNR: 37.87 -> 37.78 | Hurt |
| Set DQC learning rate to 10^-4 instead of 10^-2 (ablation) | 2DQuant · Set5 ×2 | PSNR: 37.87 -> 37.74 | Hurt |
| Reduce DQC batch size from 32 to 4 (ablation) | 2DQuant · Set5 ×2 | PSNR: 37.87 -> 37.82 | Hurt |
| Reduce DQC batch size from 32 to 8 (ablation) | 2DQuant · Set5 ×2 | PSNR: 37.87 -> 37.83 | Hurt |
| Reduce DQC batch size from 32 to 16 (ablation) | 2DQuant · Set5 ×2 | PSNR: 37.87 -> 37.84 | Hurt |
Surprising findings
- DOBI alone already matches DBDC+Pac, and DQC alone performs worse than DOBI, suggesting training from a poor initialization converges to local optima.
- The quantized 2DQuant model sometimes outperforms the FP model visually (e.g., Urban100 stripe direction), which the authors attribute to FP model errors from aliasing.
Limitations
- DOBI requires weight/activation distributions to be approximately bell-shaped or exponential.
- More MSE search points do not necessarily improve results.
- Requires a calibration set; DOBI and DQC cannot run without it.
Experimental setup
Datasets: DF2K DIV2K Flickr2K Set5
Benchmarks: Set5 Set14 B100 Urban100 Manga109
Metrics: PSNR SSIM
Baselines: MinMax Percentile DBDC+Pac EDSR
Builds on: DBDC+Pac SwinIR Straight-Through Estimator
Closest papers by meaning
- PassionSR: Post-Training Quantization with Adaptive Scale in One-Step Diffusion based Image Super-Resolution (CVPR 2025 · similarity 0.88)
- Binary Quadratic Quantization: Beyond First-Order Quantization for Real-Valued Matrix Compression (NeurIPS 2025 · similarity 0.88)
- AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR 2024 · similarity 0.88)
- Thinking in Granularity: Dynamic Quantization for Image Super-Resolution by Intriguing Multi-Granularity Clues (AAAI 2025 · similarity 0.87)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.87)
- Flexible Residual Binarization for Image Super-Resolution (ICML 2024 · 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.