OIDN-python

Python Binding of Intel Open Image Denoise Version 0.4 (Based on OIDN 2.1.0, NOT FINISHED)

Install using pip.(support macos_aarch64, win_amd64, linux_x64).

pip install oidn

Features(version 0.4)

C API wrapper

• Wrap original OIDN C APIs using ctypes. functions are stripped oidn prefix, macros are stripped OIDN_ prefix. For example oidnNewDevice -> oidn.NewDevice, OIDN_FORMAT_FLOAT3 -> oidn.FORMAT_FLOAT3.

• Discard buffer APIs, use numpy array as buffers.

Object-Oriented Interface

OOP style interface will be finished in version 1.0

Example denoising image

Denoise image rendered by a monte carlo ray tracer. code

from pathlib import Path
import sys
import numpy as np
from PIL import Image

here = Path(__file__).parent.absolute()
sys.path.append(here.parent.parent.absolute().as_posix())

import oidn

img = np.array(Image.open((here / "CornellBoxNoisy.png").as_posix()), dtype=np.float32) / 255.0

result = np.zeros_like(img, dtype=np.float32)

device = oidn.NewDevice()
oidn.CommitDevice(device)

filter = oidn.NewFilter(device, "RT")
oidn.SetSharedFilterImage(
    filter, "color", img, oidn.FORMAT_FLOAT3, img.shape[1], img.shape[0]
)
oidn.SetSharedFilterImage(
    filter, "output", result, oidn.FORMAT_FLOAT3, img.shape[1], img.shape[0]
)
oidn.CommitFilter(filter)
oidn.ExecuteFilter(filter)

result = np.array(np.clip(result * 255, 0, 255), dtype=np.uint8)
resultImage = Image.fromarray(result)
resultImage.save(f"{here}/CornellBoxDenoised.png")

oidn.ReleaseFilter(filter)
oidn.ReleaseDevice(device)

The image in left is before denoised, rendered by a Monte-Carlo PathTracer, spp=10, width=height=1000. The image in right is after denoised.

There's also more pythonic APIs (Not bindings of OIDN C++ API)

from pathlib import Path
import sys
import numpy as np
from PIL import Image

here = Path(__file__).parent.absolute()
sys.path.append(here.parent.parent.absolute().as_posix())

import oidn

with oidn.Device('cpu') as device, oidn.Filter(device, 'RT') as filter:
    input = oidn.Buffer.load(device, Image.open((here / "CornellBoxNoisy.png").as_posix()), div255=True)
    output = oidn.Buffer.create(input.width, input.height, device=device)
    filter.set_image("color", input)
    filter.set_image("output", output)
    filter.execute()
    Image.fromarray( np.array(np.clip(output.to_array() * 255, 0, 255), dtype=np.uint8) ).save(f"{here}/CornellBoxDenoised.png")

Pending Features

• Support for SYCL and HIP devices.
• Asynchronized executing/buffering APIs.

Update Log

• 0.4 : CUDA device is supported now, torch.Tensor could be passed as GPU buffers. Torch is counted into optional dependencies.
• 0.3.1alpha : Update to new oidn version 2.1.0. Only CPU device is supported at the moment.
• 0.3alpha : Warp nearly full APIs of OIDN1.4.3 in oidn.h. (excluding buffer APIs, buffers are substituted numpy array), add function __doc__, to be comprehensively test.
• 0.2.1 : Support win_amd64 and manylinux1_x86_64 platform.
• 0.2 : Wrap basic device and filter APIs, Initial support for macosx_12_0_arm64 platform.

License:

Apache 2.0

API Document

See here.