martin-rizzo/ESRGAN-4x

ESRGAN 4x Models

Overview

This repository contains two powerful 4x super-resolution models based on ESRGAN (Enhanced Super-Resolution Generative Adversarial Network). These models are specifically designed to upscale images to four times their original resolution while maintaining or enhancing visual fidelity.

Model Descriptions

1. Nickelback_70000G.safetensors

   • Architecture: ESRGAN
   • Strengths : Photorealistic upscaling of real-world images.
   • Use Cases : Best suited for scenarios requiring photorealistic output, such as enhancing rofessional photography or increasing resolution in images where realistic content is crucial.

2. foolhardy_Remacri.safetensors

   • Architecture: ESRGAN
   • Strengths : Enhances the quality of comics and anime images by improving detailed texture and edge sharpness.
   • Use Cases : Ideal for users who need high-quality outputs in anime and comic book styles, where fine details are important for narrative or artistic value.

Usage Instructions

1. Installation: Ensure you have PyTorch installed along with any other necessary libraries in your environment.
2. Loading Models: Load the models using a script that supports ESRGAN model inference (such as Real-ESRGAN or custom scripts).
3. Inference: Run your images through one of these models to upscale them by 4x.

Evaluation

These models were part of an evaluation process similar to "Hires.fix":

1. The image was first generated artificially using Z-Image for 8 steps at a resolution of 1088x1600 px.
2. The image was then downsampled by a factor of 2 using bilinear interpolation.
3. One of the ESRGAN models (x4) was applied to upscale the image.
4. The image was downsampled again by a factor of 2 using bilinear interpolation.
5. Finally, 2 steps of Z-Image were applied.

The resulting images had the same resolution as the original but with improved quality. Among these checkpoints, Nickelback_70000G.safetensors and foolhardy_Remacri.safetensors produced the best image quality for photographic and illustrative/comic/anime content respectively, according to my evaluation.

Limitations

Both checkpoints are based on ESRGAN and therefore inherit its limitations:

• They should not be used in applications where accuracy and consistency are critical, such as medical imaging or legal document processing.