Learn
What Is Denoising?
Every digital image contains some degree of noise. It might be subtle enough to ignore or severe enough to ruin a shot. Denoising is the process of identifying and reducing that noise, whether in a still photograph or a moving image, while preserving the underlying detail that makes the image worth keeping.
The challenge is deceptively simple to describe and remarkably difficult to execute well. Noise and fine detail can look almost identical to an algorithm. A speck of chrominance noise and a tiny highlight on a textured surface both register as small, localised variations in colour and brightness. The entire discipline of denoising revolves around teaching software to tell the difference.
What noise is
Noise in a digital image manifests as random variations in brightness and colour that were not part of the original scene. It is the visual equivalent of static on a radio signal: information that the sensor added during capture rather than something the lens delivered.
In practice, noise appears as grain, speckle, or colour flecking across the image. It tends to be most visible in the darker areas of the frame, in large uniform surfaces like walls or skies, and in the shadow regions where the sensor was straining to register light. At low levels it can resemble the organic grain of photochemical film. At high levels it obscures detail entirely, turning shadow areas into a churning mess of random pixels.
Understanding what noise looks like is the first step. Understanding what causes it is what allows you to minimise it at the source and treat it more effectively in post.
What causes noise
Noise is an inherent property of digital imaging. Every sensor produces it to some degree, because the process of converting photons into electrical signals is imperfect. Several factors determine how much noise ends up in the final image.
The most common cause is insufficient light reaching the sensor. When a camera's ISO or gain setting is raised to compensate for a dark environment, the sensor amplifies the signal, but it amplifies the noise floor along with it. This is why footage shot at ISO 6400 in a dimly lit room looks grainier than footage of the same scene properly lit and shot at the camera's base ISO.
Small sensor sizes contribute as well. The individual photosites on a smaller sensor are physically smaller and collect fewer photons per unit of time, producing a weaker signal relative to the noise floor. This is partly why a full-frame cinema camera handles low light better than a smartphone, though computational photography has narrowed that gap considerably.
Long exposures introduce noise because the sensor heats up over time, and thermal energy generates spurious signals. High ambient temperatures have a similar effect. Aggressive compression, particularly at low bitrates, can also introduce artefacts that behave like noise in subsequent processing steps, a consideration when working with heavily compressed video files.
Types of noise
Not all noise is the same, and the distinction matters for how you treat it.
Luminance noise appears as variations in brightness across the image. It looks like traditional film grain: a random pattern of lighter and darker pixels overlaid on the image. Of the noise types, luminance noise is generally the least objectionable. In moderate amounts it can even add a pleasing texture that audiences associate with the cinematic look of photochemical film.
Chrominance noise, sometimes called chroma noise, manifests as random colour variations. Instead of lighter and darker specks, you see coloured speckles, often in magenta, green, or blue, scattered across the image. Chroma noise is almost always more distracting than luminance noise because it looks unnatural. The human eye is quite tolerant of brightness variation but immediately notices colour where it should not be. Understanding how cameras handle colour information, including how chroma subsampling reduces colour resolution, provides useful context for why chroma noise can be particularly problematic.
Fixed pattern noise is a consistent, repeatable pattern produced by defects or inconsistencies in the sensor itself. Unlike the random noise described above, fixed pattern noise appears in the same locations on every frame. It is most visible in long exposures and can sometimes be calibrated out by the camera's firmware or subtracted in post using a dark-frame reference.
How denoising works
At its core, denoising is pattern recognition. The algorithm analyses the image, attempts to identify which variations are noise and which are legitimate detail, and then reduces or removes the noise while leaving the detail intact.
The fundamental difficulty is that noise and detail share many of the same characteristics. Both are high-frequency information, meaning they represent rapid changes in pixel values over short distances. A strand of hair, the weave of a fabric, and a cluster of noise pixels can all produce similar patterns. The sophistication of a denoising algorithm lies in how well it can distinguish between these.
Modern denoising approaches fall into three broad categories, each with distinct strengths.
Spatial denoising analyses each frame independently. It examines groups of neighbouring pixels, identifies patterns consistent with noise, and smooths them while attempting to preserve edges and detail. Spatial methods are computationally simpler and work on any source material, including still images. Their limitation is that they have only one frame's worth of information to work with, so they must be more conservative to avoid erasing detail.
Temporal denoising takes advantage of the fact that video is a sequence of frames. By comparing the same region across multiple frames, it can identify noise because noise changes randomly from frame to frame while the actual scene content remains relatively consistent. Temporal denoising is often more effective than spatial alone because it has more data to work with, but it can produce artefacts when there is motion in the frame, since the algorithm may misinterpret movement as noise. Understanding frame rates is relevant here, as the relationship between temporal sampling and motion affects how well these algorithms perform.
AI and machine-learning denoising represents the current state of the art. These systems use neural networks trained on large datasets of clean and noisy image pairs. The network learns to separate noise from detail in ways that are more nuanced than handcrafted algorithms can achieve. AI denoising tools can often recover detail that traditional methods would destroy, and they handle complex textures with more sophistication. The trade-off is computational cost: these methods typically require a capable GPU and more processing time.
In practice, the best results often come from combining approaches. DaVinci Resolve, for example, offers both spatial and temporal noise reduction that can be used together, allowing the temporal pass to handle the bulk of the noise while the spatial pass cleans up what remains.
Where denoising fits in the pipeline
Denoising is typically applied during colour grading or as a dedicated step before final output. Its placement in the post-production pipeline matters more than many editors realise.
The general principle is that denoising should happen before sharpening. Sharpening works by enhancing edges and fine detail, which means it will also enhance any noise present in the image. If you sharpen first and denoise second, you force the denoiser to work harder against noise that has been artificially amplified. If you denoise first, you give the sharpening pass cleaner material to work with and get better results from both operations.
When working with LOG or RAW footage, noise reduction is sometimes applied before the colour transform, sometimes after, and sometimes split across both stages. The specifics depend on the footage, the grading software, and the severity of the noise. Some colourists prefer to denoise the linear or LOG image because the noise characteristics are more uniform before a contrast curve is applied. Others prefer to denoise after the grade because they want to see the noise in its final context.
For editors and colourists managing complex timelines, having well-organised project files makes it easier to apply denoising methodically across a sequence. Keeping project documentation clear about which clips needed noise reduction and what settings were used is a small discipline that prevents problems during review.
The detail-vs-noise trade-off
Every denoising operation is a compromise. Removing noise inevitably removes some detail along with it. The question is never whether detail will be lost, but how much and whether the loss is acceptable.
Aggressive denoising produces that characteristic "waxy" or "plasticky" look that audiences and professionals find objectionable. Skin loses its texture and looks like it has been smeared with vaseline. Fabrics lose their weave. Hair becomes a smooth, featureless mass. The image may be technically noise-free, but it looks processed and unnatural.
The goal is to find the threshold where noise is reduced enough to stop being distracting without crossing into visible detail loss. This threshold is subjective, context-dependent, and often different for different parts of the same frame. A sky can tolerate aggressive denoising because it has minimal detail to lose. A face needs a lighter touch because any loss of skin texture is immediately noticeable.
This is why secondary denoising, applying different strength settings to different areas of the image, is a common technique. Power windows and luminance qualifiers let a colourist apply heavy noise reduction to shadow areas while leaving midtones and highlights largely untouched. The review and approval stage of post-production is where these decisions get scrutinised: what looks acceptable on a grading monitor may reveal problems when viewed at full resolution on a larger screen.
Popular denoising tools
DaVinci Resolve includes spatial and temporal noise reduction in its colour page, available even in the free version. The controls are comprehensive, with separate adjustments for luminance and chrominance noise and the ability to dial in the balance between noise reduction and detail preservation. For many projects, Resolve's built-in tools are sufficient.
Neat Video is a widely used plugin that works within most major non-linear editing systems. Its strength is its noise profiling capability: you select a flat, featureless area of the frame, and Neat Video analyses the noise characteristics, then builds a custom profile for that clip. This targeted approach often yields cleaner results than a one-size-fits-all algorithm.
Adobe Premiere Pro and After Effects include denoising tools that range from basic noise reduction effects to more sophisticated options in the Lumetri panel. Topaz DeNoise AI uses machine learning to handle severe noise in still images and video, often recovering more detail than traditional methods. DxO PhotoLab's DeepPRIME technology is notable in the stills world for its ability to produce remarkably clean images from very high-ISO captures.
Shooting to minimise noise
The most effective noise reduction happens before the camera rolls. Exposing correctly, or slightly over-exposing where highlight headroom allows (a technique sometimes called "exposing to the right"), puts more signal into the sensor and pushes noise down relative to the image data.
Using the camera's native or base ISO keeps the sensor operating at its most efficient. Most cinema cameras have a base ISO where the signal-to-noise ratio is optimal, and every stop above that introduces more noise. Some cameras have a dual native ISO architecture, offering a second low-noise setting at a higher sensitivity.
Lighting the scene adequately is the single most impactful thing a videographer can do to reduce noise. An extra light, a reflector, or even a practical lamp in the background can bring exposure up enough to keep ISO settings reasonable. Choosing cameras with larger sensors and better low-light performance is another option, though it comes with its own trade-offs in size, weight, and cost.
When to leave noise alone
Not all noise needs to be removed. Some degree of grain adds texture to an image and can contribute a cinematic quality that audiences associate with photochemical film. Many colourists deliberately add grain back into digitally shot footage as part of the final colour grade to avoid the sterile, overly clean look that some digital cameras produce.
The aesthetic choice between a perfectly clean image and one with a controlled amount of grain is subjective and depends on the project. A corporate product video typically wants to look as clean and polished as possible. A narrative film or a music video might benefit from the texture and warmth that a light grain structure provides.
The key is intentionality. Leaving noise in a shot because it serves the look is a creative decision. Leaving it in because you did not notice it or could not be bothered to address it is a quality control failure. Annotating specific frames or timecodes during review can help teams communicate clearly about which shots need treatment and which are intentionally textured.
For content creators working with footage from a variety of sources, including smartphone cameras, action cameras, and lower-end DSLRs, noise management becomes a routine part of the post-production process. Having a reliable cloud workspace to organise and search through footage makes it easier to identify which clips in a project need attention before the grade begins.
Frequently asked questions
What is the difference between noise and grain?
In common usage, "grain" tends to refer to the organic, textured pattern of photochemical film, while "noise" refers to the unwanted artefacts produced by digital sensors. In practice, the terms are often used interchangeably. Digital luminance noise can closely resemble film grain, and many denoising tools use the word "grain" in their controls.
Does denoising reduce resolution?
Denoising does not change the pixel count of an image, so the nominal resolution remains the same. However, aggressive denoising can reduce the perceived resolution by smoothing out fine detail. The image may have the same number of pixels but contain less information per pixel than it did before processing.
Should I denoise before or after colour grading?
There is no single correct answer, and professional colourists take different approaches. A common workflow is to apply noise reduction as part of the grading process, after the base correction but before or alongside the creative grade. The important rule is to denoise before sharpening, whichever stage that falls in.
Can denoising fix severely noisy footage?
Modern AI denoising tools can produce remarkable improvements on footage that would have been considered unusable a decade ago. However, there are limits. If the noise is severe enough to have completely obscured the underlying detail, no algorithm can reconstruct what was never recorded. The result may be cleaner but will lack fine detail.
Is temporal denoising always better than spatial?
Temporal denoising is often more effective because it uses information from multiple frames, but it is not universally better. It can introduce artefacts around moving objects, and it is only applicable to video, not still images. Many workflows use both temporal and spatial denoising together, with each handling different aspects of the noise.
Does shooting in RAW give better denoising results?
Generally, yes. RAW files preserve the full data from the sensor without compression artefacts, giving denoising algorithms cleaner data to work with. Compressed formats may introduce quantisation artefacts that complicate the denoising process. This is one of several reasons professional productions favour RAW or high-quality codec workflows.
What is the best denoising software?
It depends on the context. For video work within a grading workflow, DaVinci Resolve's built-in tools and Neat Video are both widely respected. For AI-powered denoising, Topaz is a popular choice. For stills, DxO DeepPRIME is highly regarded. The best tool is the one that fits your pipeline and produces acceptable results on your specific footage.
How much noise reduction should I apply?
Enough to reduce the noise to the point where it is no longer distracting, and no more. This varies by delivery format: footage destined for a cinema screen at 4K needs more careful treatment than content intended for social media at 1080p, where compression will mask much of the noise anyway.
Does sensor size affect noise?
Yes. Larger sensors have larger photosites that collect more light, producing a stronger signal relative to the noise floor. This is why full-frame and large-format cinema cameras generally produce cleaner images at high ISO settings than cameras with smaller sensors. However, sensor technology, processing algorithms, and pixel density also play significant roles.
Can I add grain back after denoising?
Yes, and it is a common practice. Many colourists denoise footage to remove the character of digital noise, then add a controlled film grain emulation on top. This gives the image texture without the harshness of digital noise. Most grading software includes grain generators, and dedicated grain overlays are available as well.