Research map / Architectures and efficiency / Quantization
AdaBM: On-the-Fly Adaptive Bit Mapping 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 AdaBM, an on-the-fly adaptive quantization framework for image super-resolution that calibrates two bit-mapping modules (image-wise and layer-wise) with 100 LR calibration images, reducing processing time from hours to seconds.
Problem
Adaptive bit-width quantization of super-resolution networks achieves good accuracy-cost trade-offs but requires hours of quantization-aware training with full LR/HR training pairs, limiting practical use.
Approach
Decouples image-wise and layer-wise bit adaptation (observed to be independent), maps image complexity to an image-wise bit factor and layer sensitivity to a layer-wise bit factor, and calibrates/fine-tunes these mappings and clipping ranges using only a small set of LR calibration images with a feature-reconstruction loss, without QAT on ground-truth HR.
Key results
- On EDSR-4bit Urban100, AdaBM reaches 25.49 dB PSNR at FAB 4.3 vs CADyQ 25.12 dB at FAB 4.9, while calibration takes 71 seconds versus 40 hours for CADyQ.
- Processing time is accelerated by about x2000 relative to existing adaptive quantization methods (CADyQ and CABM) that require 40-70 hours of QAT.
- In static no-QAT comparisons on EDSR 4-bit, AdaBM achieves 25.11 dB on Urban100 at FAB 3.7, versus PTQ4SR at 24.92 dB at FAB 4.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AdaBM instead of CADyQ for 8 / 6MP quantization | EDSR · Urban100 | PSNR: 25.98 -> 25.96 | No clear effect |
| Use AdaBM instead of CADyQ for 4 / 4MP quantization | EDSR · Urban100 | PSNR: 25.12 -> 25.49 | Helped |
| Add image-wise and layer-wise bit mapping, bit-aware clipping, and calibration fine-tuning (ablation) | AdaBM · Urban100, 4-bit EDSR | PSNR: 24.27 -> 25.11 | Helped |
| Add I2B, L2B, and BaC (ablation) | AdaBM with FT · Urban100, 4-bit EDSR | PSNR: 23.51 -> 25.11 | Helped |
| Add L2B and BaC (ablation) | AdaBM with I2B and FT · Urban100, 4-bit EDSR | PSNR: 23.80 -> 25.11 | Helped |
| Add I2B and BaC (ablation) | AdaBM with L2B and FT · Urban100, 4-bit EDSR | PSNR: 24.89 -> 25.11 | Helped |
| Add I2B and L2B (ablation) | AdaBM with BaC and FT · Urban100, 4-bit EDSR | PSNR: 25.12 -> 25.11 | No clear effect |
| Fine-tune the mappings and quantization parameters using calibration images (ablation) | AdaBM with I2B, L2B, and BaC · Urban100, 4-bit EDSR | PSNR: 24.78 -> 25.11 | Helped |
Surprising findings
- Relative layer-wise and image-wise quantization sensitivity orders are largely invariant, enabling the two adaptation policies to be learned separately.
Limitations
- Accuracy is only on-par rather than superior to QAT-based adaptive methods in some settings (e.g., SRResNet 4-bit Urban100 and some Test4K results are mixed).
- Evaluated only on SR networks EDSR, SRResNet, and RDN at scale x4 with fixed bit factors {-1,0,1}.
Experimental setup
Datasets: DIV2K DIV8K Urban100 Set5 Set14 BSD100 Test2K Test4K
Benchmarks: Urban100 Test2K Test4K Set5 Set14 BSD100
Metrics: PSNR SSIM FAB (feature average bit-width) processing time BitOPs model size
Baselines: CADyQ CABM MinMax Percentile MinMax+FT Percentile+FT PTQ4SR
Builds on: CADyQ CABM PTQ4SR OMSE LSQ-style straight-through estimator
Closest papers by meaning
- 2DQuant: Low-bit Post-Training Quantization for Image Super-Resolution (NeurIPS 2024 · similarity 0.88)
- AdaFormer: Efficient Transformer with Adaptive Token Sparsification for Image Super-resolution (AAAI 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)
- TSD-SR: One-Step Diffusion with Target Score Distillation for Real-World Image Super-Resolution (CVPR 2025 · similarity 0.86)
- ResDiff: Combining CNN and Diffusion Model for Image Super-resolution (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.