Research map / Architectures and efficiency / Quantization
Thinking in Granularity: Dynamic Quantization for Image Super-Resolution by Intriguing Multi-Granularity Clues
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 Granular-DQ, a patch-wise, layer-invariant dynamic quantization method for image super-resolution that allocates bit-widths via a granularity-bit controller and refines high-bit patches with an entropy-to-bit mechanism.
Problem
Dynamic quantization of SR models trades off SR accuracy and quantization efficiency; existing layer-and-patch-wise bit adaptation disturbs inter-layer relations and degrades representations.
Approach
Allocate bit-widths per image patch (not per layer) using a Gumbel-Softmax granularity-bit controller over coarse-to-fine multi-granularity features, then refine high-bit patches with entropy thresholds calibrated by EMA (entropy-to-bit mechanism).
Key results
- On Urban100 x4 with EDSR, Granular-DQ reaches FAB 4.97 with PSNR 26.01 dB, versus CABM at FAB 5.80 and 25.95 dB.
- On Urban100 x4 with IDN, Granular-DQ exceeds the full-precision IDN PSNR by about 0.2 dB while using FAB 4.18 versus 32 for full precision.
- With EDSR, BitOPs drop from 527.0T to 73.6T (86.0% reduction) while keeping PSNR close to full precision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GBC to fixed 8-bit quantization (ablation) | EDSR · Urban100 ×4 | PSNR: 26.01 -> 25.97 | Hurt |
| Add E2B fine-grained bit adaptation to GBC (ablation) | EDSR with GBC · Urban100 ×4 | PSNR: 25.97 -> 26.02 | Helped |
| Add ATC threshold calibration to GBC + E2B (ablation) | EDSR with GBC + E2B · Urban100 ×4 | FAB: 5.51 -> 4.97 | Helped |
| Use E2B candidate bits [4,5,6] instead of [4,5,8] (ablation) | EDSR Granular-DQ · Urban100 ×4 | PSNR: 26.01 -> 25.98 | Hurt |
| Use E2B candidate bits [4,7,8] instead of [4,5,8] (ablation) | EDSR Granular-DQ · Urban100 ×4 | FAB: 4.97 -> 5.64 | Hurt |
| Use E2B candidate bits [4,6,7] instead of [4,5,8] (ablation) | EDSR Granular-DQ · Urban100 ×4 | PSNR: 26.01 -> 25.97 | Hurt |
| Use E2B candidate bits [4,6,8] instead of [4,5,8] (ablation) | EDSR Granular-DQ · Urban100 ×4 | FAB: 4.97 -> 5.38 | Hurt |
| Use E2B candidate bits [4,5,7] instead of [4,5,8] (ablation) | EDSR Granular-DQ · Urban100 ×4 | PSNR: 26.01 -> 25.99 | No clear effect |
| Use thresholds [0.4,0.9] instead of [0.5,0.9] (ablation) | EDSR Granular-DQ with ATC · Urban100 ×4 | PSNR: 26.01 -> 25.98 | No clear effect |
| Use one threshold [0.5] and bits [4,8] instead of thresholds [0.5,0.9] and bits [4,5,8] (ablation) | EDSR Granular-DQ with E2B · Urban100 ×4 | PSNR: 26.01 -> 25.97 | Hurt |
| Use three thresholds [0.4,0.6,0.9] and bits [4,5,6,8] (ablation) | EDSR Granular-DQ with E2B · Urban100 ×4 | PSNR: 26.01 -> 25.93 | Hurt |
| Use three thresholds [0.4,0.6,0.9] and bits [4,5,7,8] (ablation) | EDSR Granular-DQ with E2B · Urban100 ×4 | PSNR: 26.01 -> 25.93 | Hurt |
Surprising findings
- Granular-DQ on IDN surpasses its full-precision counterpart by about 0.2 dB PSNR on Urban100 and Test4K.
- Allocating higher bit-widths (e.g., [4,7,8]) incurred the most FAB but gave negligible performance gains.
Limitations
- Mixed-precision solution requires specific hardware and operator support for real compression acceleration.
- Efficacy for accelerating super-resolution of large-size images is modest.
- Attention blocks in SwinIR-light and HAT-S are kept at full precision due to quantization errors; CADyQ/CABM quantization in HAT-S hit gradient explosion.
Experimental setup
Datasets: DIV2K Urban100 Test2K Test4K
Benchmarks: Urban100 Test2K Test4K Set14
Metrics: PSNR SSIM FAB BitOPs Params
Baselines: PAMS CADyQ CABM AdaBM RefQSR
Builds on: CADyQ QuantSR AdaBM Gumbel-Softmax
Closest papers by meaning
- 2DQuant: Low-bit Post-Training Quantization for Image Super-Resolution (NeurIPS 2024 · similarity 0.87)
- AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR 2024 · similarity 0.86)
- Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaptation (ICLR 2025 · similarity 0.86)
- PocketSR: The Super-Resolution Expert in Your Pocket Mobiles (NeurIPS 2025 · similarity 0.85)
- Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025 · similarity 0.85)
- Outlier-aware Slicing for Post-Training Quantization in Vision Transformer (ICML 2024 · similarity 0.85)
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.