RenderKit/oidn

What it solves

Intel Open Image Denoise (OIDN) addresses the problem of Monte Carlo noise in images produced by stochastic ray tracing methods like path tracing. By filtering out this noise, the library allows developers to significantly reduce the number of samples per pixel (spp) required to achieve a high-quality image, thereby drastically cutting rendering times.

How it works

OIDN uses a collection of efficient deep learning-based denoising filters. These filters are trained to handle a wide range of sample counts, from 1 spp (preview) to nearly fully converged images (final-frame).

To preserve detail, the filters can operate using only the noisy color (beauty) buffer, or they can utilize auxiliary feature buffers such as albedo and normal maps. The library provides a C/C++ API for integration and includes a training toolkit that allows users to train their own models to optimize for specific renderers, scenes, or content types.

Who it’s for

Rendering application developers and artists who use ray tracing based rendering solutions and want to achieve high-quality, noise-free images faster.

Highlights

• Broad Hardware Support: Supports a wide variety of CPUs (Intel 64, ARM64) and GPUs from multiple vendors (Intel Xe, NVIDIA Turing/Ampere/Ada/Hopper/Blackwell, AMD RDNA 2/3/3.5/4, Apple Silicon).
• Deep Learning Based: Employs pre-trained AI models to remove noise from ray-traced images.
• Flexible API: Provides a simple C/C++ API for easy integration into existing rendering pipelines.
• Customizable: Includes a training toolkit for creating custom-optimized denoising models.
• High Performance: Leverages modern instruction sets (SSE4, AVX, AMX, NEON, XMX) for efficient execution.