This project provides a powerful tool for upscaling videos and images to higher resolutions using advanced AI models. It leverages Real-ESRGAN to achieve state-of-the-art results in super-resolution, enhancing the quality and detail of your media.
- High-Quality Upscaling: Utilizes Real-ESRGAN, a cutting-edge super-resolution model, to produce sharp and detailed results.
- Video and Image Support: Upscale both videos (MP4, MKV, AVI, MOV) and images (PNG, JPG, WEBP).
- Flexible Resolution Control: Choose between a specific scaling factor (e.g., 2x, 4x) or a target resolution (e.g., 1080p, 4k).
- Hardware Accelerated: Leverages NVIDIA GPUs for fast processing.
- Audio Preservation: Automatically preserves the original audio track in upscaled videos.
A native macOS app with a user-friendly interface is available on the Mac App Store. Get it here: Upscale & Enhance on the Mac App Store
- Python 3.8+
- NVIDIA GPU with CUDA and cuDNN installed.
- ffmpeg: Required for video processing. You can install it using your system's package manager (e.g.,
sudo apt-get install ffmpegon Debian/Ubuntu).
-
Clone the repository:
git clone https://github.com/your-username/your-repo-name.git cd your-repo-name -
Install the required Python packages:
pip install -r requirements.txt
Note: BasicSR currently requires
numpy<2andtorchvision<0.16, so the requirements file pins compatible versions. For Apple Silicon MLX acceleration, install:pip install -r requirements-mlx.txt
-
Download the pre-trained models: Download the
RealESRGAN_x4plus.pthandRealESRGAN_x2plus.pthmodels and place them in themodelsdirectory. You can find the models in the original Real-ESRGAN repository or other trusted sources. For MLX, placeRealESRGAN_x4plus.npzandRealESRGAN_x2plus.npzin the samemodelsdirectory.
To upscale a single image or video, use the super_resolution.py script with the input_path and output_path arguments.
Example (Upscale to 4k):
python super_resolution.py path/to/your/input.mp4 path/to/your/output.mp4 --target-resolution 4kExample (Upscale by a factor of 2):
python super_resolution.py path/to/your/image.jpg path/to/your/output.jpg --scale 2Example (Use Apple Silicon MLX backend):
python super_resolution.py path/to/your/image.jpg path/to/your/output.jpg --scale 4 --backend mlxIf you hit GPU memory errors on Apple Silicon, add tiling:
python super_resolution.py path/to/your/image.jpg path/to/your/output.jpg --scale 4 --backend mlx --tile 256 --tile-pad 10To lower the process priority (run in the background):
python super_resolution.py path/to/your/image.jpg path/to/your/output.jpg --scale 4 --backgroundYou can also process all files in a directory by providing the input and output directory paths.
python super_resolution.py path/to/your/input_directory path/to/your/output_directory --target-resolution 1080pinput_path: Path to the input video, image, or directory.output_path: Path to the output file or directory.--scale: The upscaling factor (e.g., 1.5, 2.0). Mutually exclusive with--target-resolution.--target-resolution: The target resolution (e.g., 1080p, 2k, 4k, 5k, 8k). Mutually exclusive with--scale.--backend: Inference backend (auto,pytorch,mlx).autouses MLX when available.--tile: Tile size for large images to reduce memory usage.--tile-pad: Tile padding to reduce seams.--pre-pad: Pre-pad input to reduce border artifacts.--background: Lower process priority to reduce system impact.
The convert_x4_model.py script can be used to convert a x4 Real-ESRGAN model to x2 and x1 models.
python convert_x4_model.pyThis will create RealESRGAN_x2plus.pth and RealESRGAN_x1plus.pth from RealESRGAN_x4plus.pth in the models directory.
This project is licensed under the MIT License. See the LICENSE file for details.
This project is built upon the excellent work of the following open-source projects:
- Real-ESRGAN: https://github.com/xinntao/Real-ESRGAN
- BasicSR: https://github.com/xinntao/BasicSR