Denoising is a post-processing step that removes the last traces of grain from a render, so you can stop at fewer samples and still deliver a clean image. It works best on a render that is already close to clean. Used earlier or pushed too hard, it smooths away brickwork, fabric, foliage and the fine edges that make an architectural image believable.

This guide covers what causes noise, when to denoise instead of adding samples, which render passes help, how to set it up, and how to avoid the artefacts that give a denoised image away. It also covers animation, where the risks are different.

Visual Trick – CGIs, Animations, Digital Twins
Present Clean, Realistic Visuals
Visual Trick creates high-end CGI and architectural visualisations that help property teams communicate projects clearly across planning and marketing.

Explore Visual Trick

TL;DR

  • Get the render reasonably clean first, then denoise to finish. Denoising is for the last few percent of noise, not for rescuing a badly lit scene.
  • Feed the denoiser guide data (albedo and normal passes) so it keeps textures and edges.
  • Lower the sample count in steps and check the image at 100% after each change.
  • In animation, over-aggressive denoising can cause flickering splotches, so settle your noise level before the denoiser touches the frames.
  • Fix the cause of noise (weak lighting, bright glossy highlights, fireflies) rather than hiding it.

Table of Contents

What noise is and why archviz scenes produce it

Path-traced renderers work by firing rays at random and averaging what they find. Noise appears when those random rays miss important light contributors, such as bright glossy highlights. According to the Blender manual, more samples raise the chance of finding them, and denoising is the cleanup for the final bit of noise that remains.

Architectural scenes are a tough case for this. Interiors are often lit by light that has bounced through a window or around a corner, so many rays never reach a light source. Glass, polished stone, chrome and water create bright, narrow highlights that are easy to miss. Volumetric effects such as haze or sunbeams add more randomness.

The symptoms differ, and it helps to name them:

  • Fine grain across shadows and indirect light, the classic low-sample look.
  • Fireflies, isolated bright pixels from rare, high-energy light paths.
  • Blotchy noise in reflections and caustics.
  • Grain in depth of field and motion blur, where low ray counts show up as texture. Maxon’s Redshift documentation lists depth of field, motion blur, global illumination and area lighting as typical places.

Grain and fireflies call for different fixes, which is why “just add samples” is rarely the best first move.

How a denoiser actually works

A denoiser reconstructs clean pixels from a noisy, low-sample image. AMD’s GPUOpen write-up explains that denoisers often rely on auxiliary buffers such as albedo, normal, roughness and depth, and that noisy diffuse and specular signals can be denoised separately and recombined to protect fine detail.

Two families exist, and the distinction matters for how you work:

  • Offline denoisers aim at production-quality reconstruction from higher sample counts. This is what you want for final stills.
  • Real-time denoisers are built to cope with very low samples under tight time budgets. They suit live viewports and walkthroughs, and they trade some accuracy for speed.

Most modern offline denoisers are machine-learning based. Blender’s compositor Denoise node, for example, uses Intel Open Image Denoise and accepts optional albedo and normal passes to preserve detail.

When to denoise and when to add samples

The short rule: raise samples and fix lighting until the image is almost clean, then denoise. OTOY’s workflow guidance says to get the render reasonably clean without denoising first, then enable denoising and reduce the maximum samples step by step, checking that the image holds together.

Workflow for sampling and denoising renders

Maxon gives the same advice in different words: tune the settings to an acceptable noise level, then use denoising for the last few percent, especially for animation. It also notes that denoising can be faster than substantially increasing samples.

The saving can be large. Disney Animation reported that its production denoiser cut render times by a factor of 2 to 4 compared with earlier denoisers, and ILM reported worst-case reductions of 2 to 3 compared with not denoising. Feature-film pipelines are not archviz, but the principle carries over: a good denoiser buys back hours.

The same logic appears in consumer tools. Live Home 3D’s documentation for Radeon ProRender says higher sample counts improve quality but lengthen render time, and suggests switching the denoiser on instead of endlessly raising samples once the remaining noise is acceptable.

A quick decision guide

Situation Better move
Heavy grain across the whole frame Raise samples or improve lighting first, then denoise
Almost clean, a little grain in shadows Denoise and lower the sample count in steps
A few bright speckles (fireflies) Clamp or blur the source of the energy, then denoise
Noise in volumetric haze or sunbeams Enable volume denoising where your renderer offers it
Final animation frames Reach a low noise level first, then denoise lightly

Fix the cause before you denoise

Denoisers are forgiving, but they work from what you give them. Cleaner input gives cleaner output, and these habits reduce the load:

  1. Light the scene properly. OTOY advises against compensating for weak lighting with excessive exposure. If exposure has to sit well above 1, raise the emitter or HDR environment power and bring exposure back down.
  2. Tame fireflies. The same guidance recommends GI Clamp to limit high-energy rays, Caustic Blur to smooth caustic and reflection noise, and Hotpixel Removal as a post-process. Its examples suggest GI Clamp values between 1 and 10, Caustic Blur of 0.3 and Hotpixel Removal of 0.5. Treat those as starting points for your own scene.
  3. Soften the worst offenders. Blender’s manual suggests blurring bright glossy spots or limiting global illumination, while admitting that these shortcuts can change the exact result. For a planning or marketing image where materials must read accurately, make these changes deliberately and compare against the original.
  4. Use adaptive sampling. OTOY recommends it so noisy areas receive more samples rather than every pixel getting the same effort.
  5. Test small. Live Home 3D’s guidance favours iterative low-resolution tests before the final render, which is cheap insurance.

Setting up denoising step by step

The menus differ between engines, but the workflow is the same. Here are the two best documented examples.

In Blender Cycles

  1. Enable the Denoising Data pass in the view layer settings.
  2. In the compositor, connect the Noisy Image, Denoising Albedo and Denoising Normal outputs to a Denoise node, as the manual recommends. These guides help preserve edges, textures and fine detail.
  3. Choose a prefilter. According to the same page, None keeps the most detail but expects clean guide passes, Fast is quicker but blurrier, and Accurate prefilters noisy guides and generally gives more detail at extra processing cost.
  4. Remember the sampling guidance: denoising generally improves when more passes are available. Albedo is the recommended minimum, and Albedo plus Normal gives the denoiser more scene information.

In Octane

Enable Camera > Imager > Enable Denoising. If your scene contains volumetrics, also enable Denoise Volumes. The Spectral AI Denoiser page describes the adjustable completion timing, minimum samples, interval and Blend controls. Blend runs from 0 to 1: zero is fully denoised, one is the original image, and values between mix the two. A Blend value slightly above zero is a simple way to bring back a touch of natural texture.

If you are tuning render quality inside a BIM-led workflow, a Revit rendering settings checklist is a handy companion for keeping the earlier settings consistent before the denoiser comes into play.

Artefacts to watch for

Denoising is lossy. HOOPS Luminate’s documentation states that small defects can appear and source detail can be lost. It also explains the radius and strength trade-off: a large radius and high strength remove more noise but also fine detail, while a small radius suits images that already have little noise.

Maxon lists the typical failures: residual grain, oversmoothed textures or small geometry, brush-stroke-like patterns, and flickering splotches in animation.

In architectural images, these show up in specific places:

  • Timber, brick and render textures turning waxy or painterly.
  • Fine foliage and grass merging into smears.
  • Thin mullions, railings and cables losing definition or breaking up.
  • Text and signage softening, which matters on commercial and healthcare schemes.
  • Soft shadow gradients gaining blotches that read as stains on walls and ceilings.

The check is simple: view the denoised result at 100%, flip between it and the noisy version, and look at the materials you care most about. If any of them have lost character, lower the strength, raise samples slightly or blend some of the original back.

Denoising architectural animation

Stills only need to look right once. Animation must look right across hundreds of frames, and a denoiser that treats each frame independently produces slightly different guesses every time. The result is shimmer or flickering splotches, as Maxon warns.

Three practical rules follow:

  • Bring each frame to a low, consistent noise level first, then denoise to finish, following Maxon’s last-few-percent advice.
  • Use guide passes, so surface detail stays anchored from frame to frame.
  • Prefer denoisers that look across neighbouring frames. The Disney Animation and ILM publication describes temporal denoising over a seven-frame region, using motion vectors, diffuse and specular separation, albedo, normals and variance estimates to improve stability.

Before committing to a full sequence, render a short test section with a camera move and some foliage, and watch it at full speed. Flicker that is invisible on a single frame is obvious in motion.

Our perspective

At Visual Trick, every image is produced in-house by one experienced operator, with no outsourcing. That shapes how we think about denoising: it is a finishing tool, and the decisions that matter are made long before the render starts.

Our images go into planning applications, investor pitches, brochures and listings. In those contexts, a waxy brick or a smeared hedge is not a minor flaw. It changes how a scheme is read. So we settle lighting, materials and sampling first, and we treat the denoiser as the last pass that gets the file ready for delivery.

Because the same person handles the brief, the scene and the final output, we can match the workflow to the purpose of the image. A hero exterior for a planning submission is handled differently from a quick interior for a marketing campaign, and an animation is handled differently again. Predictable deadlines come from that control, not from rushing the last render.

If you want clean, realistic and planning-ready visuals without worrying about samples, passes and prefilters, that is the job we do.

Paride

Need clean, planning-ready visuals?

If you would rather commission the finished image than tune the pipeline, we produce custom CGI for developers, architects, planning consultants, estate agents and designers across the UK and worldwide. Our architectural visualisation services cover exterior and interior CGI, photomontage, verified views and animation, all produced in-house.

Visual Trick - CGIs, Animations, Digital Twins

To discuss a project, start from our services page. If your scheme is heading for planning, our CGI for planning applications page explains what we prepare. If you also work on post-production, see our guide to Photoshop post-production with AI.

FAQ

Does denoising make renders look fake?

It can if it is pushed too hard. Strong settings smooth fine texture, so materials look waxy. Keep the render close to clean before denoising, use albedo and normal guides, and check at 100% to keep the result natural.

Should I use denoising or more samples for a final still?

Use both in sequence. Get the image reasonably clean first, then enable the denoiser and lower the maximum samples gradually while checking image integrity, as OTOY’s guidance suggests. That is usually faster than raising samples alone.

Which passes help a denoiser most?

Albedo is the recommended minimum in Blender’s Cycles documentation, and adding Normal gives the denoiser more scene information. These guides help keep edges, textures and fine detail intact.

Can I denoise an animation the same way as a still?

Not quite. Frames denoised independently can flicker, so reach a low noise level first and denoise only the last few percent. Test a short sequence at full speed before rendering the whole animation.

Sources