Nuke Denoiser

A Nuke node plugin for denoising CG renders using Intel Open Image Denoise (OIDN).

Bundles a one-click install script.

Forked from mateuszwojt/NukeCGDenoiser with significant Windows stability fixes and quality-of-life improvements.

Requirements

• Windows 10/11 x64 (this fork is Windows-only)
• Nuke 16 (tested on 16.0v8; other recent versions likely work)
• CUDA drivers for Nvidia GPUs (Optional, defaults to CPU denoising)

---

Install

Make sure Nuke is closed before proceeding with the install.

Option A — Straightforward (recommended)

Copy paste this command in a Windows Command Prompt (cmd.exe)

cd Downloads
git clone https://github.com/JTCHE/NukeCGDenoiser
cd NukeCGDenoiser
install.bat

Option B — Double-click

1. Clone this repository
2. Double-click install.bat
3. Restart Nuke
4. Delete this folder

Option C — Python

python install.py

What it does

The installer:

• Copies the plugin folder to ~/.nuke/nuke-denoiser/
• Adds a pluginAddPath entry to ~/.nuke/init.py
• Cleans up any old denoiser entries from previous installs
• Is safe to run multiple times (idempotent)

Uninstall

python uninstall.py

Removes the init.py entry and optionally deletes the plugin folder.

---

Usage

1. In Nuke, Tab-search Denoiser (or go to Nodes > MW > Denoiser)
2. Connect your render and choose a workflow:

Workflow A — Multi-layer EXR (recommended)

Place the Denoiser directly on a multi-layer EXR Read node. Use the layer picker knobs to select which passes to use:

• Beauty layer — the layer to denoise (default: rgb). The denoised result is written to the rgb channels; the original layer is left untouched.
• Albedo layer — albedo or diffuse pass for auxiliary-guided denoising. Set to none to skip.
• Normal layer — world-space normals for auxiliary-guided denoising. Set to none to skip. Must be accompanied by Albedo layer.

All other layers pass through unchanged, so the node works like a Shuffle + Denoise in one step.

Workflow B — Separate inputs

Connect separate images to each input:

Input	Content	Required
0 - Beauty	Beauty pass (RGB)	✅
1 - Albedo	Albedo AOV (RGB)	❌
2 - Normal	Normal AOV (RGB)	❌

When albedo or normal inputs are connected, the corresponding layer picker knobs are disabled — the node reads RGB directly from the connected input. This is useful for split EXRs, PNGs, or renders that output each AOV as a separate file.

Alpha denoising

Enable the alpha checkbox (A) in the Beauty layer knob to also denoise the alpha channel. Alpha is denoised in a separate pass since OIDN processes 3 channels at a time.

Properties

Knob	Description
Device Type	CPU (default) or CUDA
Quality	Balanced (default) or High
HDR	Enable for high-dynamic-range input (recommended)
Enable thread affinity	Pins OIDN threads to hardware threads for performance
Memory limit (MB)	Cap OIDN memory usage; 0 = no limit
Number of runs	Feed the image through the filter N times
Beauty / Albedo / Normal layer	Select which layer from the input EXR to use for each role. Albedo and Normal are disabled when their inputs are connected
Prefilter Auxiliary Passes	Tells OIDN to prefilter noisy albedo/normal before using them for guidance. Leave unchecked if your auxiliary passes are already clean

---

Building from source

Requires Visual Studio 2022 and CMake 3.10+. MinGW/GCC cannot be used — Nuke ships MSVC import libraries (.lib files) and GCC cannot link against them.

cd nuke-denoiser
cmake -B build -G "Visual Studio 17 2022" -DOIDN_ROOT="c:/bin/oidn-2.1.0"
cmake --build build --config Release
copy build\lib\Release\Denoiser.dll denoiser.dll

The output denoiser.dll should be placed at the repo root (or in your ~/.nuke/nuke-denoiser/ install).

OIDN path: Replace c:/bin/oidn-2.1.0 with wherever you extracted the OIDN 2.1.0 release. Do not use a newer OIDN version — see the TBB compatibility note above.

Nuke path: The CMake config auto-detects Nuke from standard install locations. If detection fails, set -DNUKE_ROOT="C:/Program Files/Nuke16.0v8".

---

What's different from the original

The original plugin could freeze or crash the host machine when used with animation sequences or multiple concurrent frames. This fork fixes that and adds a proper Windows distribution.

Stability fixes (C++):

• Global OIDN device shared across all node instances — eliminates per-instance construction/destruction races during multi-threaded Nuke renders
• CRITICAL_SECTION instead of std::mutex for thread synchronization — std::mutex global constructors don't reliably run when a DLL is loaded via LoadLibrary inside Nuke, causing segfaults on lock
• DllMain initializes the critical section on DLL_PROCESS_ATTACH, guaranteed to run before any plugin callbacks
• OIDN DLL search path resolved relative to the plugin DLL at load time (GetModuleFileNameW) — no hardcoded install paths

Distribution:

• OIDN 2.1.0 CPU runtime DLLs bundled in oidn/bin/ — no separate OIDN installation needed
• CUDA device DLL (OpenImageDenoise_device_cuda.dll) not bundled due to size; drop it in oidn/bin/ to enable GPU denoising
• install.py / install.bat — one-click installer that copies the plugin and patches ~/.nuke/init.py
• uninstall.py — clean removal

Why OIDN 2.1.0 and not newer: OIDN 2.4.x ships with tbb12.dll v2022.3, which conflicts with Nuke's own tbb.dll v2020.3 when both are in the same process, causing immediate crashes. OIDN 2.1.0's tbb12.dll (v2021.10) is compatible with Nuke's TBB version.

---

Known limitations

• CUDA denoising is experimental. Requires an NVIDIA GPU and driver 522.06+. Drop OpenImageDenoise_device_cuda.dll from the OIDN 2.1.0 release into oidn/bin/ to enable it. HIP and SYCL are not supported.
• Windows only. The original plugin supports Linux and macOS; this fork's distribution system is Windows-specific. The C++ source itself is cross-platform.
• Inputs must match resolution. If beauty, albedo, and normal are different sizes, reformat them to match before connecting.

---

Acknowledgements

Based on NukeCGDenoiser by Mateusz Wojt, licensed under the original project's license (see LICENSE).

Uses Intel Open Image Denoise, © Intel Corporation.