Exploring Frequency-Inspired Optimization in Transformer for Efficient Single Image Super-Resolution

Ao Li, Le Zhang, Yun Liu and Ce Zhu, "Exploring Frequency-Inspired Optimization in Transformer for Efficient Single Image Super-Resolution", TPAMI, 2025

[arXiv] [Official]

---

Abstract: Transformer-based methods have exhibited remarkable potential in single image super-resolution (SISR) by effectively extracting long-range dependencies. However, most of the current research in this area has prioritized the design of transformer blocks to capture global information, while overlooking the importance of incorporating high-frequency priors, which we believe could be beneficial. In our study, we conducted a series of experiments and found that transformer structures are more adept at capturing low-frequency information, but have limited capacity in constructing high-frequency representations when compared to their convolutional counterparts. Our proposed solution, the cross-refinement adaptive feature modulation transformer (CRAFT), integrates the strengths of both convolutional and transformer structures. It comprises three key components: the high-frequency enhancement residual block (HFERB) for extracting high-frequency information, the shift rectangle window attention block (SRWAB) for capturing global information, and the hybrid fusion block (HFB) for refining the global representation. To tackle the inherent intricacies of transformer structures, we introduce a frequency-guided post-training quantization (PTQ) method aimed at enhancing CRAFT’s efficiency. These strategies incorporate adaptive dual clipping and boundary refinement. To further amplify the versatility of our proposed approach, we extend our PTQ strategy to function as a general quantization method for transformer-based SISR techniques. Our experimental findings showcase CRAFT’s superiority over current state-of-the-art methods, both in full-precision and quantization scenarios. These results underscore the efficacy and universality of our PTQ strategy.

---

Floating-point Visual Results

HR	LR	SwinIR	ESRT	CRAFT (ours)

Qunatization Visual Results (4-bit X4)

Full-Precision	LR	MinMax	PTQ4SR	Ours

Dependencies & Installation

• Python 3.7
• PyTorch 1.10.2
• NVIDIA GPU + CUDA 11.7

git clone https://github.com/AVC2-UESTC/Frequency-Inspired-Optimization-for-EfficientSR.git
conda create -n CRAFT python=3.7
conda activate CRAFT
pip install -r requirements.txt

Training

Floating-point Model Training Procedure

• Download the train datasets and place them in the datasets directory.

• Run the following script to start the training process.

  # Train with 4 GPUs
  # X4 scaling
  bash scripts/dist_train.sh 4 \
    options/train/CRAFT/train_CRAFT_SRx4_scratch.yml 

Quantized Model Training Procedure

• Randomly sample and crop 128x128 sub-images, and place in the datasets/calibration_data directory.
• Run the following scripts to perform quantization:

  # Train with 1 GPU
  # Stage 1: Adaptive Dual Clipping Process
  python PTQ/Adaptive_Dual_Clipping_main.py \
    --output_dir results/ptq \
    --saved_model_path experiments/train_CRAFT_SR_X4/PTQ_models \
    --fp_model_path experiments/pretrained_models/float_models/CRAFT_MODEL_x4.pth \
    --traindir_LR datasets/calibration_data/X4 \
    --benchmarks Set5+Set14+B100 \
    --scale 4 \
    --bits 4
  
  # Stage 2: Boundary Refinement Process
  python PTQ/Boundary_Refinement_main.py \
    --output_dir results/ptq \
    --saved_model_path experiments/train_CRAFT_SR_X4/PTQ_models \
    --fp_model_path experiments/train_CRAFT_SR_X4/float_models/CRAFT_MODEL_x4.pth \
    --ptq_adc_model_path experiments/train_CRAFT_SR_X4/PTQ_models/CRAFT_MODEL_4bit_x4_ADC.pth \
    --traindir_LR datasets/calibration_data/X4 \
    --benchmarks Set5+Set14+B100 \
    --scale 4 \
    --bits 4 \
    --epochs 10 \
    --lr 0.002

Testing

Test images with High-Resolution (HR) Images

• Download the pre-trained models and place them in experiments/pretrained_models/.

  Floating-point models (Google Drive): CRAFT_MODEL_x2, CRAFT_MODEL_x3, and CRAFT_MODEL_x4.

  Floating-point models (BaiduYun): CRAFT_MODEL_x2, CRAFT_MODEL_x3, and CRAFT_MODEL_x4.

  Qunatized models (Google Drive): CRAFT_MODEL_4bit_x4, CRAFT_MODEL_6bit_x4 and CRAFT_MODEL_8bit_x4.

  Qunatized models (BaiduYun): CRAFT_MODEL_4bit_x4, CRAFT_MODEL_6bit_x4 and CRAFT_MODEL_8bit_x4.

• Download test datasets, place them in datasets/benchmark.

• Run the following scripts.

  # Test Set5 (X4) 
  # Floating-point model
  python inference/inference_CRAFT.py --scale 4 --model_path experiments/pretrained_models/float_models/CRAFT_MODEL_x4.pth --folder_lq datasets/benchmark/Set5/LR_bicubic/X4 --input datasets/benchmark/Set5/HR --output results/CRAFT/Set5/X4
  
  # 4-bit quantized model
  python PTQ/PTQ_eval.py --model_path experiments/pretrained_models/PTQ_models/CRAFT_MODEL_4bit_x4.pth --bits 4
  
  # 6-bit quantized model
  python PTQ/PTQ_eval.py --model_path experiments/pretrained_models/PTQ_models/CRAFT_MODEL_6bit_x4.pth --bits 6
  
  # 8-bit quantized model
  python PTQ/PTQ_eval.py --model_path experiments/pretrained_models/PTQ_models/CRAFT_MODEL_8bit_x4.pth --bits 8

• The results will be saved in the results directory.

Results

Floating-point

Model	#Parameters	Set5	Set14	BSD100	Urban100	Manga109
CRAFT-X2	737K	38.23/0.9615	33.92/0.9211	32.33/0.9016	32.86/0.9343	39.39/0.9786
CRAFT-X3	744K	34.71/0.9295	30.61/0.8469	29.24/0.8093	28.77/0.8635	34.29/0.9491
CRAFT-X4	753K	32.52/0.8989	28.85/0.7872	27.72/0.7418	26.56/0.7995	31.18/0.9168

Quantization (X4)

Model	#W/A	Set5	Set14	BSD100	Urban100	Manga109
Full-Precision	32/32	32.52/0.8989	28.85/0.7872	27.72/0.7418	26.56/0.7995	31.18/0.9168
CRAFT-8bits	8/8	32.45/0.8965	28.80/0.7850	27.68/0.7397	26.49/0.7967	31.06/0.9141
CRAFT-6bits	6/6	32.01/0.8820	28.51/0.7717	27.49/0.7272	26.16/0.7808	30.11/0.8922
CRAFT-4bits	4/4	29.46/0.7854	26.92/0.6891	26.32/0.6492	24.50/0.6787	27.03/0.7653

Citation

If you find the code helpful in your research or work, please cite the following paper(s).

@inproceedings{li2023craft,
  title={Feature Modulation Transformer: Cross-Refinement of Global Representation via High-Frequency Prior for Image Super-Resolution},
  author={Li, Ao and Zhang, Le and Liu, Yun and Zhu, Ce},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={12514--12524},
  year={2023}
}

@article{li2025exploring,
  author={Li, Ao and Zhang, Le and Liu, Yun and Zhu, Ce},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Exploring Frequency-Inspired Optimization in Transformer for Efficient Single Image Super-Resolution}, 
  year={2025},
  volume={47},
  number={4},
  pages={3141-3158},
}

Acknowledgements

This code is built on BasicSR, CAT, Restormer, and FQ-ViT.