You can deploy this model and start earning money today!

XiangZ/
hit-sr

Name: hit-sr
Availability: InStock

HiT-SR

image super-resolution

transformer

efficient transformer

image-to-image

2407.05878

Views

No views yet

HiT-SR: Hierarchical Transformer for Efficient Image Super-Resolution

[Github] | [Paper] | [Supp] | [Video] | [Visual Results]

HiT-SR is a general strategy to improve transformer-based SR methods. We apply our HiT-SR approach to improve SwinIR-Light, SwinIR-NG and SRFormer-Light, corresponding to our HiT-SIR, HiT-SNG, and HiT-SRF. Compared with the original structure, our improved models achieve better SR performance while reducing computational burdens.

🛠️ Setup

Install the dependencies under the working directory:

git clone https://huggingface.co/XiangZ/hit-sr
cd hit-sr
pip install -r requirements.txt

🚀 Usage

For each HiT-SR model, we provide 2x, 3x, 4x upscaling versions:

Repo Name	Model	Training Set	Upscale
`XiangZ/hit-sir-2x`	HiT-SIR	DIV2K	2x
`XiangZ/hit-sir-3x`	HiT-SIR	DIV2K	3x
`XiangZ/hit-sir-4x`	HiT-SIR	DIV2K	4x
`XiangZ/hit-sng-2x`	HiT-SNG	DIV2K	2x
`XiangZ/hit-sng-3x`	HiT-SNG	DIV2K	3x
`XiangZ/hit-sng-4x`	HiT-SNG	DIV2K	4x
`XiangZ/hit-srf-2x`	HiT-SRF	DIV2K	2x
`XiangZ/hit-srf-3x`	HiT-SRF	DIV2K	3x
`XiangZ/hit-srf-4x`	HiT-SRF	DIV2K	4x
`XiangZ/hit-srf-2x-df2k`	HiT-SRF	DF2K	2x
`XiangZ/hit-srf-3x-df2k`	HiT-SRF	DF2K	3x
`XiangZ/hit-srf-4x-df2k`	HiT-SRF	DF2K	4x

To test the model (use hit-srf-4x-df2k as an example):

from hit_sir_arch import HiT_SIR
from hit_sng_arch import HiT_SNG
from hit_srf_arch import HiT_SRF
import cv2
import torch

# detect device
device = torch.device('cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu')

# initialize model (change model and upscale according to your setting)
model = HiT_SRF(upscale=4) 

# load model (change repo_name according to your setting)
repo_name = "XiangZ/hit-srf-4x-df2k"
model = model.from_pretrained(repo_name)
model.to(device)

# test and save results
with torch.no_grad():
    sr_results = model.infer_image("path-to-input-image", device=device)
cv2.imwrite("path-to-output-location", sr_results)

📎 Citation

If you find the code helpful in your research or work, please consider citing the following paper.

@inproceedings{zhang2024hitsr,
    title={HiT-SR: Hierarchical Transformer for Efficient Image Super-Resolution},
    author={Zhang, Xiang and Zhang, Yulun and Yu, Fisher},
    booktitle={ECCV},
    year={2024}
}