Python script for running and inferencing image upscaling models on image folders using optimization techniques like batched tile processing and automatic mixed precision (AMP).

Prerequisites

• Python 3.7+
• NVIDIA GPU that supports CUDA version 12.4 or higher (This Script uses PyTorch with CUDA 12.4).

Installation

1. Create a virtual environment using Python's module or using Anaconda Prompt and activate the environment.

   python -m venv [env_name]
   .\[env_name]\Scripts\activate
   

   or

   conda create --name [env_name]
   conda activate [env_name]
   

2. Update PyPI if necessary:

   pip install --upgrade pip
   

3. Install the package:

   pip install Folderesque --extra-index-url https://download.pytorch.org/whl/cu124
   

Usage

1. Create a new project or workspace.

2. Copy the folder of images that you want to upscale to the workspace.

3. Download the pre-trained model:

   • Place your pretrained model in the project directory. The example for this script uses the RealESRGAN_x4plus_anime_6B.pth model.
   • Download from Real-ESRGAN repository.

4. Create a config.py file with contents matching the one shown below:

   INPUT_PATH = "daskruns"
   OUTPUT_PATH = "testscaling"
   MODEL_PATH = "models\RealESRGAN_x4plus_anime_6B.pth"
   SCALE_FACTOR = 4
   DEVICE = "cuda"
   TILE_SIZE = 400
   THREAD_WORKERS = 4
   BATCH_SIZE = 16
   

5. Ensure that INPUT_PATH and MODEL_PATH points to the correct path of your input folder and the model path respectively.

6. To run the script, copy the path of the config file and run the command:

   Folderesque --conf [conf_path]
   

Output images are saved with "ESRGAN_" prefix in the filenames.

❤ Credits

Immense thanks to:

• Real-ESRGAN authors: Xintao Wang
• BasicSR framework: BasicSR

Troubleshooting

Common Issues:

1. CUDA Out of Memory:

   • Reduce TILE_SIZE.
   • Decrease THREAD_WORKERS.
   • Decrease BATCH_SIZE.

2. Command not recognized:

   • You may want to add the folder path of your environment to the system's path.
   • Alternatively, you can use python -m prefix when running commands.