|
--- |
|
language: en |
|
tags: |
|
- mohsin-riad |
|
- image-processing |
|
- super-resolution |
|
- upscaling |
|
- real-esrgan |
|
license: apache-2.0 |
|
base_model: xinntao/realesrgan-x4plus |
|
datasets: |
|
- DIV2K |
|
- Flickr2K |
|
library_name: pytorch |
|
pipeline_tag: image-to-image |
|
--- |
|
|
|
# Upscaler-Ultra |
|
|
|
 |
|
|
|
## Model Description |
|
|
|
Upscaler-Ultra is a high-performance image upscaling model built upon RealESRGAN architecture. This model is designed to enhance image resolution while maintaining high quality and preserving fine details. The model specializes in upscaling low-resolution images to higher resolutions with minimal artifacts and maximum clarity, leveraging the proven effectiveness of Real-ESRGAN for practical image restoration tasks. |
|
|
|
### Model Architecture |
|
|
|
This model is based on RealESRGAN (Real-Enhanced Super-Resolution Generative Adversarial Networks), which utilizes: |
|
- Enhanced ESRGAN architecture optimized for real-world image degradation |
|
- Adversarial training with improved discriminator networks |
|
- Perceptual loss functions for better visual quality |
|
- Specialized training techniques for handling complex real-world artifacts |
|
|
|
## Intended Uses & Limitations |
|
|
|
### Intended Uses |
|
|
|
- Image upscaling and enhancement |
|
- Photo restoration and quality improvement |
|
- Digital art enhancement |
|
- Low-resolution image improvement |
|
- Professional photography post-processing |
|
- Real-world image super-resolution tasks |
|
|
|
### Limitations |
|
|
|
- Performance may vary depending on input image quality and degradation type |
|
- Very low-resolution inputs might not achieve optimal results |
|
- Processing time increases with input image size |
|
- May not preserve extremely fine details in heavily compressed images |
|
- Best suited for natural images rather than synthetic graphics |
|
|
|
### Base Model |
|
Built upon [RealESRGAN](https://github.com/xinntao/Real-ESRGAN), specifically the RealESRGAN-x4plus model, with additional fine-tuning and optimizations. |
|
|
|
### API Usage |
|
|
|
The model is available through Replicate API: |
|
```python |
|
import replicate |
|
|
|
output = replicate.run( |
|
"mohsin-riad/upscaler-ultra", |
|
input={"image": "path_to_your_image.jpg"} |
|
) |
|
``` |
|
|
|
Replicate: [mohsin-riad/upscaler-ultra](https://replicate.com/mohsin-riad/upscaler-ultra) |
|
## Citation |
|
|
|
If you use this model in your research, please cite: |
|
|
|
```bibtex |
|
@misc{upscaler-ultra, |
|
author = {Mohsin Riad}, |
|
title = {Upscaler-Ultra: High-Quality Image Upscaling Model Based on RealESRGAN}, |
|
year = {2025}, |
|
publisher = {Hugging Face}, |
|
journal = {Hugging Face Hub}, |
|
howpublished = {\url{https://huggingface.co/mohsin-riad/upscaler-ultra}} |
|
} |
|
``` |
|
|
|
Please also cite the original RealESRGAN work: |
|
|
|
```bibtex |
|
@InProceedings{wang2021realesrgan, |
|
author = {Xintao Wang and Liangbin Xie and Chao Dong and Ying Shan}, |
|
title = {Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data}, |
|
booktitle = {International Conference on Computer Vision Workshops (ICCVW)}, |
|
date = {2021} |
|
} |
|
``` |
|
|
|
## Additional Information |
|
|
|
For questions and feedback, please contact: |
|
- GitHub: [mohsin-riad](http://github.com/mohsin-riad) |
|
- Model Repository: [upscaler-ultra](http://github.com/mohsin-riad/upscaler-ultra) |
|
|
|
### License |
|
|
|
This model is released under the Apache License 2.0. |
|
|
|
### Acknowledgments |
|
|
|
- Special thanks to the RealESRGAN team for the foundational architecture |
|
- Thanks to the open-source community and all contributors who have helped in the development of this model |
|
- Built upon the excellent work of Xintao Wang et al. on Real-ESRGAN |