Overcoming Blurring And Artifacts From Blender’S Denoiser
What is denoising and why it can cause problems
Denoising is a process in 3D rendering that attempts to reduce unwanted visual noise and grain in the final image. It works by analyzing pixel colors and patterns to determine areas of noise versus actual scene detail. The Blender denoiser uses algorithms to selectively blur noise while preserving edges and textures.
While denoising can greatly improve renders, it can also introduce new issues if not used properly:
- Blurring of fine details and textures
- Smearing and artifacts around edges
- Blotchy patterns and inconsistencies
These problems typically arise from the denoiser misinterpreting noise for actual scene elements, or having difficulty distinguishing complex textures and geometry.
How the Blender denoiser works
The Blender denoiser is a hybrid OpenImageDenoise plugin that uses two algorithms working together:
- Bayesian – Analyzes individual pixels based on the surrounding pixels to determine if they represent noise or detail.
- Dilation/Erosion – Compares geometric differences between adjacent pixels to preserve edges and surface boundaries.
The combined output merges the detail preservation of the Bayesian step with the edge awareness of the dilation/erosion step to produce the final denoised image.
Key settings that control performance:
- Iterations – How many times to re-apply the algorithms.
- Feature Strength – Overall filtering intensity.
- Radius – Size of the pixel area sampled around each point.
Increasing these values can improve noise reduction but also heightens the chances of damaging detail.
Common artifacts and blurring issues
When pushed too far, the information in an image can get over-simplified, leading to these common denoising problems:
- Blurring of textures, particulates, foliage, hair, skin pores and blemishes.
- Blocky pixelation in subtle color gradients and shadows areas.
- Flattened depth and exaggerated edges around objects and selections.
- Watercolor-like smearing of fine lines and intricate details.
- Removal of intentionally placed noise like film grain.
Identifying the specific cause of unwanted denoising artifacts allows the best approach to resolve them…
Optimizing Your Render for Denoising
Sample counts and noise levels
The key to effective denoising is finding the optimal balance of render samples and noise levels before applying any filtering…
Too little noise from ultra-high sampling (5000+ samples) leaves little for the denoiser algorithms to grab onto, often resulting in over-blurring, lost detail, and an artificial appearance.
Excessive noise (under 250 samples) can confuse the underlying pixel analysis, causing misplaced artifacts, blotchiness, and uneven texture retention.
Ideally, sample somewhere between 250-1200 samples initially, just enough to establish the key forms, edges, and lighting continuity without fully clearing noise. This focused noise gives the denoiser better discernment between detail and remaining grain to selectively smooth.
Lighting and materials
Proper lighting and shading techniques also dramatically assist the denoiser in accurately separating noise from essential detail, including…
- Soft area lighting rather than harsh direct and localized lighting.
- Smooth color gradations on materials over high frequency detail.
- Avoiding extremely dark areas and bright hot spots.
- Simpler flat and moderately bumped shaders first.
- Lower contrast textures.
Renders exhibiting these characteristics tend to denoise cleaner with less chance of artifacts or excessive feature blurring.
Compositing considerations
Plan ahead when using render layers and holdout masks in compositing. Limit complexity until basic lighting and shading is established…
- Initially render beauty elements separately then recombine.
- Denoise individual components before blending.
- Add more renders later once workflow is verified.
This isolates potential issues to specific layers and reduces masking problems around edges that can confuse the denoiser.
Fixing Blurring
Adjusting denoise feature strength
Excessive blurring is perhaps the most common unwanted side effect of over-enthusiastic denoising. Often first impulse is to simply keep lowering the feature strength…but this can worsen artifacts by allowing more residual noise through…
Instead, first check render sample counts and noise levels meet suggested guidelines. Then selectively paint lower feature strength directly only on the affected surfaces. Typically between 50-80% strength retains details clearly while eliminating most noise on common materials.
Using multiple denoise passes
Rather than one high strength pass, stack multiple gentle small-radius denoise nodes at lower feature strength…
5-15 passes at 10-30% strength adds up to enough filtering to eliminate noise yet with subtle enough individual impact to avoid strong blurring side effects.
Combine with limiting each pass to only necessary shader and lighting components to maximize control.
Sharpening filters
As a last resort, localized sharpening via blur or high-pass filter nodes can selectively reintroduce certain details if too aggressively blurred…
- Plug image into bottom socket of blur node to retrieve high frequencies.
- Layer high-pass result back over denoised beauty for extra crispness.
Use mask controls to carefully limit sharpening just to affected regions to avoid increasing noise elsewhere.
Reducing Artifacts
Masking technique
Surface detail inconsistencies and blotchy patches often emerge along texture edges and normal map seams. These confuse the automatic perceptual criteria in the denoiser.
Painting mask layers to isolate these tricky regions allows selective reduction of feature strength or radius just where needed to retain detail clarity.
Tuning feature influence settings
In addition to overall feature strength, adjusting the advanced influence sliders biases the denoiser priorities towards preserving certain image qualities:
- Color – Retains richness over smoothing noise
- Normal – Favors surface bump detail
- Albedo – Values texture and color accuracy
For example, increasing normal influence across the image reduces erroneous blending of edges and gradients, at expense of some residual grain.
Patching problem areas
For small isolated artifact spots not improved by above steps, a semi-manual approach…
- Render region just around flaw.
- Denoise smaller area individually.
- Copy corrected section back into full image.
Careful masking and sample count matching avoids visible seams from patched insert.
Example Scenes and Node Setups
Simple scene using basic node workflow
Even a simple scene with few unique textures can develop issues when denoising that need selective attention…
Adding the denoiser quickly reduced overall noise, but some blotchiness formed on the ottoman seat at normal and albedo seams. Lowering the feature influence on normals in this area cleared up the artifacts while retaining pattern and color detail.
Complex scene demonstrating advanced techniques
More ambitious scenes require careful isolation and treatment of difficult sub-elements…
The cloth scratches and tattoo symbols started blurring excessively despite attempts to simply lower strength across the board. Ultimately a separate denoise node just for the clothing layer at 5% strength prevented loss of finesse while allowing heavier filtering on skin and hair.
Further Reading on Denoising
Resources for deeper understanding of algorithms
For readers interested in the academic research and mathematics underlying Blender’s state of the art denoising approaches…
- Non-local means filtering – Original Bayesian analysis algorithm.
- Mathematical morphology – Framework behind erosion/dilation method.
These seminal works pioneered the core techniques now adapted and expanded upon for application to 3D rendering.
New denoising features on the horizon
The Blender development roadmap promises exciting improvements to denoising such as:
- Intel Open Image Denoise integration – Fast GPU acceleration.
- Optix AI-acceleration
- Native compositor nodes
- Enhanced texture sampling
Check the latest Blender 3.4 release notes and development blog to keep updated on the field of denoising – an area of ongoing research with new advances continually enhancing Blender’s capabilities.
