Could you add support for calling Python ?

[juntaosun]

This is a very good project, except for the vision-cli and C++ calling methods.
Could you add support for calling Python?

Because vision-cli processes data via the command line, it's suitable for single image input or output.
For example, if I want to use Real-ESRGAN to process video frames, single image I/O might involve frequent disk reads and writes, which is inefficient.

Each time a command is called, the model needs to be loaded.

vision-cli esrgan -m ESRGAN-4x-foolhardy_Remacri-F16.gguf -i input.png -o output.png

If you could support Python's inference calls to vision, just like the C++ calling method,
that would be much easier to operate. We know that Python is a very simple language, and combined with the vision.cpp inference library,

loading the model once and processing the video input would be fantastic.
I think it would be wonderful. Is this possible?

Python + vision.cpp, ( without using PyTorch)

Or, could you encapsulate the vision.cpp API into a library that Python can call ?

[Acly]

My preferred way to create Python bindings would be to add a C interface to vision.cpp libraries, and then use ctypes from Python.

I can try to add something, but probably don't have time to do a complete package for PyPi at the moment.

[Acly]

Here is an example on how to use the latest version (v0.3.0+) from Python:

First extract a release, or build from source and install.
Then use it from Python like this:

import sys
sys.path.append("/visioncpp_install_path/python")

from PIL import Image
from visioncpp import Device, Model

device = Device.init()
print("Device:", device.name, device.description)

model = Model.load("RealESRGAN-x4plus_anime-6B-F16.gguf", device)
print("Model:", model.arch)

image = Image.open("image.jpg")
result = model.compute(image)
result.save("result.png")
print("Complete!")

[juntaosun]

That's amazing !