Intel Open Image Denoise 3: Fast Temporal Denoising & Supersampling in Cycles — BCON26

Intel Open Image Denoise (OIDN) — what it is

• Intel OIDN (Open Image Denoise) is an AI-based denoiser for ray-traced content creation.
• It can be used for:
  • Final-frame rendering
  • Real-time/viewport preview rendering
  • Denoising for light baking (lightmaps)
• It runs locally (no subscription), is open source (Apache 2.0), and is cross-platform:
  • Windows, Linux, macOS
• Hardware support:
  • CPU: x86_64 (Intel/AMD) and ARM64
  • GPU: multiple vendor backends (Intel / NVIDIA / AMD / Apple)

Integration & API / compute architecture

• Provides a C API with a C++ wrapper, enabling integration into applications like Blender.
• Uses backend abstractions per GPU vendor:
  • Sycl (Intel), CUDA (NVIDIA), HIP (AMD), Metal (Apple)
• Unified shared memory by default (buffers share pointers between CPU/GPU).
• Zero-copy interoperability via Direct3D/Vulkan sharing mechanisms:
  • Avoids copying memory back to CPU before denoising, enabling real-time performance.
• Neural network precision choices:
  • FP16 for image feature/tensor processing (quality for HDR)
  • FP32 mandatory for depth images (precision required)

Blender workflow and how OIDN is used (as presented)

OIDN 2.x is already integrated in Blender 5.2. The talk describes three common usage paths:

1. Viewport denoising (preview)

   • Preview with fewer samples; denoising fills in plausible results quickly.

2. Final render denoising

   • Render with moderate sample counts, denoise afterward to cut render time.

3. Compositor node workflow

   • A render layer outputs auxiliary buffers like albedo and normals.
   • These feed into an OIDN denoiser node, producing a denoised color output.

What’s new in OIDN 3.0

OIDN 3.0 adds several major capabilities:

1. New network architecture using “kernel prediction”

   • Replaces/changes the previous end-to-end approach to predict filter kernel weights.
   • Improves quality and reduces artifacts such as:
     • Ringing
     • Color shifting
   • Better detail preservation compared to OIDN 2.x.

2. Temporal stability denoising (for viewport and animation)

   • Reduces wobbling / frame-to-frame instability when navigating or rendering sequences.
   • Demonstrated improvements vs spatial-only denoising (OIDN 2.5 style).

3. Built-in upscaling: 1x to 3x

   • Enables denoising while scaling resolution up as part of the model.
   • Uses auxiliary buffers (e.g., albedo/normals) for higher-quality upscaling than generic bicubic-type methods.

4. Composable AOV denoising (Arbitrary Output Value)

   • Kernel prediction makes it feasible to denoise multiple AOVs with the same learned kernel.
   • Includes joint denoising of alpha with minimal extra cost.

Temporal denoising inputs (OIDN 3.0 specifics)

For temporal denoising, the model uses additional guiding data beyond color + normals:

• Albedo split into:
  • Diffuse component
  • Specular component
• Roughness
• Depth
• Motion vectors described as 2.5D motion vectors:
  • X/Y motion vectors (where a pixel maps from last frame)
  • Plus linear depth delta (depth change between frames)
• Memory across frames (temporal “history”), effectively increasing effective sample count.

The talk contrasts:

• Spatial model: denoise each frame “from scratch” → causes wobble
• Temporal model: uses prior frame information → reduces wobble/ghosting

Kernel prediction: how it works (high level)

• Prior approach: an end-to-end network that downsamples/upsamples the noisy image (unit architecture), which could produce artifacts.
• New approach: predict the weights of an image-space filter kernel.

Benefits

• Better image quality (less ringing/color shift)
• Same kernel can be applied to denoise other AOVs (enabling fast composable AOV denoising).

Performance-related design detail

• Feature extraction in INT8 (8-bit integers) for speed with little quality loss.
• Filtering still in FP16 for quality where it matters.

Other advantages highlighted

• Consistent denoising experience across:

  • Viewport (few-port) temporal denoising
  • Final frame denoising using closely related models and training data; differences are in the motion-vector inputs (forward/backward motion vectors for final frames, backward-only reasoning for interactive viewport movement).

• Custom retraining

  • OIDN is open-source and includes a PyTorch-based training toolkit.
  • Blender collaboration provides tools/scripts to export training AOVs in the expected format.

Roadmap / expected releases (as stated)

• Alpha OIDN 3.0: Oct–Nov
  • Adds: viewport spatial filter with kernel prediction
  • Lightmap filter on GPUs
  • CPU: final-frame spatial + composable AOV support
• Beta: December (with bug fixes)
• Final release: Q1 2027
  • Enables: viewport temporal upscaling filters on all GPUs, CPU compatible CPUs
• OIDN 3.1: Q2 2027
  • Adds temporal filter to final frame using backward & forward motion vectors

Demo results / qualitative observations (examples)

• Tested in Blender demos on an ultra-portable laptop with Intel Arc iGPU and a high-resolution display.
• Examples noted:
  • Nishita Sky: animation looks much better with OIDN 3.0 than without, while noisy sampling can’t keep up.
  • Italian flat indoor scene: strong improvement over noisy baseline; temporal benefits when new parts enter view.
  • Light change scenario: OIDN 3.0 adapts to changed lighting without the typical ghosting issues the speaker associated with other temporal denoisers.

Main speakers / sources

• Christoph Noaza (speaker/host guiding the presentation)
• Atal Afra (lead developer at Intel; acknowledged during intro)
• Product/source: Intel Open Image Denoise (OIDN) and its OIDN 3.0 technical updates as presented in the talk.