span-ncnn-py

Custom models:

• ClearRealityV1 (4X) by Kim2091
• SPANkendata (4X) by terrainer
• 2x-span-anime-pretrain (2X) by Kim2091
• 2xHFA2kSpan (2x) by Phhofm Python Binding for span-ncnn-py with PyBind11

Span go brrr for upscaling images.

Current building status matrix

System	Status	CPU (32bit)	CPU (64bit)	GPU (32bit)	GPU (64bit)
Linux (Clang)		—	✅	—	✅
Linux (GCC)		—	✅	—	✅
Windows		—	✅	—	✅
MacOS		—	✅	—	✅
MacOS (ARM)		—	✅	—	✅

Usage

Python >= 3.6 (>= 3.9 in MacOS arm)

To use this package, simply install it via pip:

pip install span-ncnn-py

For Linux user:

apt install -y libomp5 libvulkan-dev

Then, import the span class from the package:

from span_ncnn_py import span

To initialize the model:

span = Span(gpuid: int = 0, tta_mode: bool = False, tilesize: int = 0, model: int = 0)
# model can be 0, 1, 2, 3, 4, 5, 6; 0 for default
# 0: {"param": "spanx2_ch48.param", "bin": "spanx2_ch48.bin", "scale": 2},
# 1: {"param": "spanx2_ch52.param", "bin": "spanx2_ch52.bin", "scale": 2},
# 2: {"param": "spanx4_ch48.param", "bin": "spanx4_ch48.bin", "scale": 4},
# 3: {"param": "spanx4_ch52.param", "bin": "spanx4_ch52.bin", "scale": 4},
# 4: {"param": "2xHFA2kSPAN_27k.param", "bin": "2xHFA2kSPAN_27k.bin", "scale": 2},
# 5: {"param": "4xSPANkendata.param", "bin": "4xSPANkendata.bin", "scale": 4},
# 6: {"param": "ClearReality4x.param", "bin": "ClearReality4x.bin", "scale": 4}

Here, gpuid specifies the GPU device to use, tta_mode enables test-time augmentation, tilesize specifies the tile size for processing (0 or >= 32), and model specifies the num of the pre-trained model to use.

Once the model is initialized, you can use the upscale method to super-resolve your images:

Pillow

from PIL import Image

span = span(gpuid=0)
with Image.open("input.jpg") as image:
    image = span.process_pil(image)
    image.save("output.jpg", quality=95)

opencv-python

import cv2

span = span(gpuid=0)
image = cv2.imdecode(np.fromfile("input.jpg", dtype=np.uint8), cv2.IMREAD_COLOR)
image = span.process_cv2(image)
cv2.imencode(".jpg", image)[1].tofile("output_cv2.jpg")

ffmpeg

import subprocess as sp

# your ffmpeg parameters
command_out = [FFMPEG_BIN, ........]
command_in = [FFMPEG_BIN, ........]
pipe_out = sp.Popen(command_out, stdout=sp.PIPE, bufsize=10 ** 8)
pipe_in = sp.Popen(command_in, stdin=sp.PIPE)
realesrgan = Realesrgan(gpuid=0)
while True:
    raw_image = pipe_out.stdout.read(src_width * src_height * 3)
    if not raw_image:
        break
    raw_image = realesrgan.process_bytes(raw_image, src_width, src_height, 3)
    pipe_in.stdin.write(raw_image)

Build

here

The project just only been tested in Ubuntu 18+ and Debian 9+ environments on Linux, so if the project does not work on your system, please try building it.

References

The following references were used in the development of this project:

xinntao/Real-ESRGAN-ncnn-vulkan - This project was the main inspiration for our work. It provided the core implementation of the Real-ESRGAN algorithm using the ncnn and Vulkan libraries.

Real-ESRGAN - Real-ESRGAN is an AI super resolution model, aims at developing Practical Algorithms for General Image/Video Restoration.

media2x/realsr-ncnn-vulkan-python - This project was used as a reference for implementing the wrapper. Special thanks to the original author for sharing the code.

ncnn - ncnn is a high-performance neural network inference framework developed by Tencent AI Lab.

License

This project is licensed under the BSD 3-Clause - see the LICENSE file for details.