FiDeSR: High-Fidelity and Detail-Preserving One-Step Diffusion Super-Resolution

⭐ Accepted by CVPR 2026

🔥 News

• CVPR 2026 Accepted
• [2026.03] Paper: CVPR 2026 Paper
• [2026.04] Code and pretrained model are released (training / inference / pretrained models).

📌 Framework

⚙️ Dependencies & Installation

git clone https://github.com/Ar0Kim/FiDeSR.git
cd FiDeSR

conda create -n fidesr python=3.10
conda activate fidesr

pip install -r requirements.txt

⚡ Quick Inference

Step 1: Download the Pretrained Models

Download the following models:

Model	Description	Link
SD 2.1-base	Base diffusion model	Stable Diffusion 2.1-base
RAM	Recognize Anything Model (for tagging)	ram_swin_large_14m.pth
FiDeSR	FiDeSR checkpoint (LoRA + LRRB weights)	fidesr.pkl

Step 2: Prepare the StableSR test datasets

Download StableSR testsets from HuggingFace.

Step 3: Run Inference

python test_fidesr.py \
  --pretrained_model_path preset/models/stable-diffusion-2-1-base \
  --pretrained_path preset/models/fidesr.pkl \
  --process_size 512 \
  --upscale 4 \
  --input_image preset/test_datasets \
  --output_dir experiments/test \
  --hf_scale 0.2 \
  --lf_scale 0.2

🖼️ Results

Trade-off Comparison

FiDeSR achieves the best trade-off between fidelity (PSNR↑, SSIM↑, LPIPS↓) and perceptual quality (MANIQA↑) among existing methods including DiffBIR, PiSA-SR, SeeSR, AddSR, OSEDiff, StableSR, SinSR, and PASD.

Visual Comparison

License

This project is released under the Apache 2.0 license.

Acknowledgments

Our project builds upon PiSA-SR. We sincerely thank the authors for their awesome work.

Citations

@inproceedings{kim2026fidesr,
  title={FiDeSR: High-Fidelity and Detail-Preserving One-Step Diffusion Super-Resolution},
  author={Kim, Aro and Jang, Myeongjin and Moon, Chaewon and Shin, Youngjin and Jeong, Jinwoo and Park, Sang-hyo},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={38270--38280},
  year={2026}
}