PyOIDN: Intel Open Image Denoise Python binding

Yet another unofficial Intel Open Image Denoise (OIDN) Python binding -- but more Pythonic. Checkout docs for more details.

Features

• Directly use NumPy arrays as input/output images.
• Support all OIDN filter types.
• Simple and clean API design.
• Lightweight: only depends on NumPy and the OIDN shared library.
• CUDA device support (via PyTorch).

Install

To install the latest release from PyPI:

pip install pyoidn

The default installation only includes lightweight CPU support with NumPy. Additionally, pyoidn provides support via PyTorch which enables CUDA device support:

pip install pyoidn[torch]

Note: this variant doesn't support Intel Macs, use the default installation instead.

Quickstart

Given a noisy image, plus its normal map and albedo map, denoise and save the result.

import numpy as np
from PIL import Image
import pyoidn


def load_image(path: str) -> np.ndarray:
    return np.array(Image.open(path), dtype=np.float32) / 255.0


color = load_image(color_path)
normal = load_image(normal_path)
albedo = load_image(albedo_path)
result = np.zeros_like(color, dtype=np.float32)

device = pyoidn.Device()
device.commit()

flt = pyoidn.Filter(device, "RT")
flt.set_image(pyoidn.OIDN_IMAGE_COLOR, color, pyoidn.OIDN_FORMAT_FLOAT3)
flt.set_image(pyoidn.OIDN_IMAGE_NORMAL, normal, pyoidn.OIDN_FORMAT_FLOAT3)
flt.set_image(pyoidn.OIDN_IMAGE_ALBEDO, albedo, pyoidn.OIDN_FORMAT_FLOAT3)
flt.set_image(pyoidn.OIDN_IMAGE_OUTPUT, result, pyoidn.OIDN_FORMAT_FLOAT3)

flt.commit()
flt.execute()

# Always check errors if something looks off
assert device.get_error() is None

result_u8 = np.array(np.clip(result * 255, 0, 255), dtype=np.uint8)
Image.fromarray(result_u8).save(output_path)

flt.release()
device.release()

The result:

pyoidn also supports RAII-style resource management using context managers:

with pyoidn.Device() as device:
    device.commit()
    with pyoidn.Filter(device, "RT") as flt:
        flt.set_bool("hdr", True)
        # set images and other parameters
        flt.commit()
        flt.execute()

If you have installed the PyTorch support, you can also use CUDA devices:

torch_device = torch.device("cuda:0")
color_t = torch.from_numpy(color).to(device=torch_device, dtype=torch.float32).contiguous()
normal_t = torch.from_numpy(normal).to(device=torch_device, dtype=torch.float32).contiguous()
albedo_t = torch.from_numpy(albedo).to(device=torch_device, dtype=torch.float32).contiguous()
result_t = torch.zeros_like(color_t, dtype=torch.float32, device=torch_device)
with pyoidn.Device(torch_device) as device:
    device.commit()
    with pyoidn.Filter(device, "RT") as flt:
        # set images and other parameters
        flt.set_image(pyoidn.OIDN_IMAGE_COLOR, color_t, pyoidn.OIDN_FORMAT_FLOAT3)
        flt.set_image(pyoidn.OIDN_IMAGE_NORMAL, normal_t, pyoidn.OIDN_FORMAT_FLOAT3)
        flt.set_image(pyoidn.OIDN_IMAGE_ALBEDO, albedo_t, pyoidn.OIDN_FORMAT_FLOAT3)
        flt.set_image(pyoidn.OIDN_IMAGE_OUTPUT, result_t, pyoidn.OIDN_FORMAT_FLOAT3)

        flt.commit()
        flt.execute()

Notes

• Error handling: use device.get_error() after creating/committing/executing.
• Async example: see tests/test.py.

Documentation

Build locally:

pip install -r requirements-docs.txt
mkdocs serve

Roadmap

• CPU device support
• NumPy array support
• All filter types support
• OIDN buffer support
• RAII-style resource management
• More device types
  • CUDA (torch)
  • SYCL
  • Metal
• Documents
• Examples

License

This project is licensed under the MIT License. See LICENSE.

This project includes:

• Intel Open Image Denoise (Apache License 2.0)