In plain words
Take a photo at a dimly lit dinner and it often looks grainy, full of tiny random coloured specks. That grain is called noise. It happens because little light reached the camera sensor, and the sensor's own electronics add small errors.
Old noise filters smooth the picture by blending each pixel with its neighbours. That hides the specks but also blurs the edges, hair and text. An AI denoiser has learned from many pairs of grainy and clean photos what real texture looks like and what random speckle looks like, so it can remove one and keep the other.
It runs quietly in many phone cameras and photo editors. The same idea, learning to remove noise, is also at the heart of image generators built on diffusion models.
Everyday example
Hearing a friend on a noisy railway platform
On a crowded platform, your friend's voice arrives mixed with announcements and chatter. Your brain knows what her voice sounds like, so it pulls her words out of the din. An AI denoiser does the same with pixels: it knows what edges, skin and fabric look like, so it keeps those and drops the speckle.
Where the example breaks: If the din drowns a word completely, you guess it from context. The denoiser does too: in very dark photos it paints in plausible texture rather than recovering what was really there, which can look waxy or invent detail.
How it actually works
Noise in photos is mostly random. Because it is random, it follows patterns a model can learn, which is what makes learned denoising work.
Where the noise comes from
Light arrives as tiny packets called photons, and their number varies by chance from pixel to pixel. This shot noise is worst when light is scarce. The sensor's electronics add read noise when turning each pixel's charge into a number.
Capture more light when you can
Phones often take a burst of frames and merge them. Random noise differs in each frame, so averaging cancels much of it. Google's Night Sight for Pixel phones, described in 2018, works this way.
Learn from grainy and clean pairs
Train a convolutional network on pairs: a noisy photo and a clean version of the same scene. Adobe's Denoise, released in 2023, was trained on millions of pairs of noisy and clean image patches taken from real raw photos.
Predict the noise, then subtract it
A 2016 network called DnCNN learned to output the noise itself rather than the clean picture. Subtract that from the input and the clean image is left. One model handled many noise levels.
No clean photos needed
In 2018, Noise2Noise showed a network can learn from noisy photos alone, by training it to turn one noisy shot of a scene into another noisy shot of the same scene. The random parts cancel out on average, leaving the real image.
- Deep-learning milestone
- DnCNN, 2016, from Kai Zhang, Wangmeng Zuo and colleagues, which predicts the noise and subtracts it
- Learning without clean images
- Noise2Noise, 2018, also used for medical scans and computer graphics
- Link to image generation
- Diffusion models learn to remove noise step by step. Run from pure noise, that skill becomes a way to create images
- Where you meet it
- Phone night modes, photo editors such as Lightroom, medical scans, cleaning up computer-rendered images
Denoised images can look too smooth. Adobe suggests adding a little grain back, and no denoiser replaces getting more light when you take the photo.
Questions people ask
What is noise in a photo?
Noise is random grain or coloured speckle that shows up mostly in dark areas and dim scenes. It comes from the natural randomness of light hitting the sensor and from the sensor's electronics, not from dust or a dirty lens.
Is AI denoising better than a normal noise filter?
Usually, yes. A normal filter blurs specks and detail together. An AI denoiser has learned the difference, so it keeps edges and textures sharper. The risk is that, in very noisy shots, it fills in texture that looks real but was never captured.
Why do night photos from phones look smooth or painted?
Strong denoising removes grain but can also wipe out fine texture like skin pores, leaves or fabric. What remains looks waxy or watercolour-like.
What does denoising have to do with AI image generators?
A lot. Diffusion models are trained to remove a little noise at a time. To create an image, they start from pure random noise and denoise it many times, guided by your prompt, until a picture appears.
Was this clear?
1 is confusing, 5 is crystal clear. One tap, no account needed.
Spotted something wrong, or know a better example? Suggest a change
Further reading
- Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising (Zhang et al.)arXiv, 2016
- Noise2Noise: Learning Image Restoration without Clean Data (Lehtinen et al.)arXiv, 2018
- Night Sight: Seeing in the Dark on Pixel PhonesGoogle Research, 2018
- Denoise demystifiedAdobe Blog, 2023
- Denoising Diffusion Probabilistic Models (Ho, Jain, Abbeel)arXiv, 2020