Denoise Image: How to Remove Noise from Photos

Every photographer has been there — a great shot ruined by grain. Shoot in dim light, crank the ISO, use a phone's tiny sensor, or push a long exposure too far, and you get noise: that speckled, gritty texture that makes photos look like they were printed on sandpaper. This guide covers exactly how to denoise an image, from one-click desktop tools to command-line automation and OpenCV code — pick the method that fits your workflow.

What Causes Image Noise

Noise enters photos at the sensor level. Understanding the source tells you how aggressively you need to treat it.

High ISO sensitivity. ISO amplifies the electrical signal from your sensor. At ISO 200, you are amplifying a clean signal. At ISO 6400, you are amplifying a weak signal and every imperfection in the electronics — thermal variation, electron scatter, manufacturing tolerances — gets amplified with it. The result is random brightness and color fluctuations scattered across the frame.

Small sensors. Smartphone sensors pack millions of pixels into a sensor the size of your thumbnail. Each photosite (the individual light collector behind each pixel) is tiny, which means it captures fewer photons before the shutter closes. Fewer photons means worse signal-to-noise ratio, which means more noise — especially in shadows.

Low light. When light is scarce, your camera compensates by raising ISO, lengthening the exposure, or both. Either path introduces noise.

Long exposures. Leave the shutter open for 30 seconds on a star trail shot and the sensor heats up slightly. Heat causes thermal noise — a pattern of brighter pixels scattered across the dark areas of the image. This is why cameras have long exposure noise reduction modes (they take a "dark frame" with the shutter closed and subtract it).

JPEG compression. A well-exposed shot can pick up compression artifacts from aggressive JPEG encoding. This is not true image noise, but it looks similar and responds to some of the same treatments.

Two Types of Image Noise

Not all noise is the same, and some tools handle each type differently.

Luminance noise looks like film grain — irregular variations in brightness. Pixels that should be a uniform mid-grey vary between light grey and dark grey. It has a relatively organic appearance and is often tolerable in small amounts. Aggressive smoothing to remove it tends to smear fine detail.

Chrominance noise (color noise) looks worse — random colored speckles, usually red, green, and blue dots scattered through shadow areas. A dark background that should be neutral grey sprouts colored flecks. Chrominance noise is harder to ignore and easier to remove without damaging detail, because removing random color shifts does not destroy edge information.

Most tools let you adjust luminance and chrominance reduction independently. Start with chrominance (remove the colored speckling first), then apply conservative luminance reduction, then sharpen selectively to recover the edge detail you lost.

Desktop Methods

Photoshop 2024 — Camera Raw AI Denoise

Photoshop 2024's Camera Raw filter uses a neural network trained on millions of images to separate noise from detail. It is the fastest path to clean results without manual tuning.

1. Open your image in Photoshop. For RAW files, Camera Raw opens automatically. For JPEGs and TIFFs, go to Filter → Camera Raw Filter.
2. In Camera Raw, click the Detail panel (the magnifying glass icon in the right panel, or press Ctrl+Alt+3 / Cmd+Option+3).
3. Click Denoise. Camera Raw analyzes the image and applies AI-based noise reduction.
4. Use the Amount slider to control strength. For ISO 3200, start at 50 and judge from there. For ISO 12800+, you may push to 70-80.
5. Adjust Luminance Detail and Color Detail sliders to recover edge sharpness if needed.
6. Click OK to apply.

The AI Denoise model runs locally — no cloud processing. It generates a full-resolution DNG from your RAW data, so you keep maximum detail. Processing a 24MP file takes 5-20 seconds depending on your GPU.

Limit: Camera Raw AI Denoise requires a Photoshop subscription (Adobe Creative Cloud, from $22.99/month for Photography plan). License: proprietary.

Lightroom Classic 13.x — AI Denoise

Lightroom Classic's Denoise tool (introduced in 2023, refined through 13.x) works on RAW files in the Develop module.

1. Open a RAW file in Lightroom Classic's Develop module.
2. Go to Detail panel → click Denoise.
3. A dialog shows a before/after preview. Set Amount (0-100). The recommended starting point for most ISO levels appears automatically.
4. Click Enhance. Lightroom creates a new DNG file (it does not modify the original).

For non-RAW files, use the traditional Detail sliders: Luminance (0-100) for grain, Color (0-100) for colored speckles. These are the sliders that have been in Lightroom for a decade — effective but not as powerful as the AI model.

Limit: Lightroom Classic requires an Adobe subscription. License: proprietary.

GIMP 2.10.36 — Non-Local Means

GIMP is free and open source (GPLv3+). Its built-in denoise filter uses Non-Local Means (NLM) — an algorithm that reduces noise by averaging each pixel against similar patches elsewhere in the image, rather than just blurring neighboring pixels. The result preserves edges better than a simple Gaussian blur.

1. Open your image in GIMP 2.10.36.
2. Go to Filters → Enhance → Non-Local Means.
3. Adjust Spatial Patch Radius (default: 3) — larger values catch more noise but are slower.
4. Adjust Patch Noise Sensitivity — higher values apply stronger smoothing.
5. Click OK.

For color noise specifically, first convert the image to Lab color space (Image → Mode → Lab), apply a Gaussian Blur of 1-2 pixels to only the a and b channels (which carry color information), then convert back to RGB. This reduces color speckles without touching luminance detail.

GIMP 2.10.36 is free to download at gimp.org. License: GPLv3+.

DxO PureRAW 4

DxO PureRAW 4 is a standalone preprocessing tool that applies DxO's DeepPRIME XD2 algorithm — arguably the strongest AI noise reduction available for RAW files. It processes RAW files before they reach Lightroom or Photoshop, outputting a clean DNG.

The workflow: drag RAW files onto PureRAW, select DeepPRIME XD2, click Process. The output DNG opens in your normal editor with the noise already removed. Processing a 24MP RAW file takes 30-60 seconds on a modern CPU with GPU acceleration.

DxO PureRAW 4 costs $139 (one-time purchase). License: proprietary. Free trial: 31 days.

Online Tools

imglarger.com — Browser-based AI upscaler that includes a denoise mode. Upload, select Denoise, download. Works on JPEGs and PNGs. Free tier: 5 images/day at up to 5MP. Paid plans from $9.99/month. Files are uploaded to their servers — check their privacy policy if this matters for your workflow.

Let's Enhance — AI enhancement platform with denoise as part of its processing pipeline. Free tier: 10 credits/month. Paid plans from $9/month. Also uploads files to their servers.

Both tools are reasonable for occasional use when you do not have desktop software available. Neither matches Photoshop or PureRAW for fine control.

Command Line: ImageMagick 7.1

ImageMagick 7.1 includes -enhance (reduces noise by averaging neighboring pixels) and the -noise operator (adds or reduces specific noise types). For batch processing on a server or in a build pipeline, this is the practical choice.

Reduce noise with -enhance (applies a 5x5 kernel averaging pass — modest but fast):

# ImageMagick 7.1.1-15
magick noisy-photo.jpg -enhance denoised-photo.jpg

Run it twice for stronger results on high-ISO shots:

# ImageMagick 7.1.1-15
magick noisy-photo.jpg -enhance -enhance denoised-photo.jpg

Remove Gaussian noise with -noise:

# ImageMagick 7.1.1-15
# Reduce noise using a 3-pixel radius median filter (good for impulse/salt-and-pepper noise)
magick noisy-photo.jpg -noise 3 denoised-photo.jpg

Batch process a directory:

# ImageMagick 7.1.1-15
for f in ./raw-photos/*.jpg; do
  magick "$f" -enhance -enhance "./denoised/$(basename "$f")"
done

-enhance is a linear filter — it works well on mild noise but will soften fine detail at high noise levels. For aggressive noise (ISO 6400+), the desktop AI tools above produce better results. For mild noise cleanup in batch pipelines, ImageMagick is fast and scriptable. License: ImageMagick License (permissive, Apache 2.0-compatible).

Programming: OpenCV 4.9

OpenCV 4.9's fastNlMeansDenoisingColored implements the Non-Local Means algorithm for color images. It is the standard choice when you need programmatic denoise in a Python pipeline.

# OpenCV 4.9.0 — pip install opencv-python==4.9.0.80
import cv2

# Load image
img = cv2.imread("noisy-photo.jpg")

# fastNlMeansDenoisingColored parameters:
#   h               — luminance filter strength (10 = default, 3-7 for mild noise, 10-15 for heavy)
#   hColor          — color channel filter strength (10 = default, typically match h)
#   templateWindowSize — patch size in pixels (must be odd, 7 = default)
#   searchWindowSize   — search window size in pixels (must be odd, 21 = default)
denoised = cv2.fastNlMeansDenoisingColored(
    img,
    None,
    h=10,
    hColor=10,
    templateWindowSize=7,
    searchWindowSize=21
)

cv2.imwrite("denoised-photo.jpg", denoised)

Tuning the parameters:

• h=3 to h=7: mild noise (ISO 400-800 on a decent sensor). Preserves fine texture.
• h=10 to h=15: moderate noise (ISO 1600-3200). Default range.
• h=15 to h=25: heavy noise (ISO 6400+). Expect some detail loss — the signal-to-noise ratio is just bad at this point.

For grayscale images, use cv2.fastNlMeansDenoising instead (same parameters minus hColor).

Processing time scales with searchWindowSize. The default (21) is a reasonable balance. If you are processing thousands of images and latency matters, try searchWindowSize=15 — the quality drop is small, the speed gain is significant.

License: OpenCV is Apache 2.0. The Python package (opencv-python) is also Apache 2.0.

Tool Comparison

The right tool depends on your context:

• Best quality on RAW files: DxO PureRAW 4 (DeepPRIME XD2 outperforms Camera Raw in blind tests, though the gap has narrowed).
• Best integrated workflow: Lightroom Classic (denoise stays inside your existing edit).
• Best free desktop option: GIMP with Non-Local Means — honest results, no subscription.
• Best for batch automation: ImageMagick 7.1 for mild noise; OpenCV 4.9 for programmable control.
• Best online (no software install): imglarger.com for quick one-offs.

Compress After Denoising

Denoising changes the image's statistical properties in a way that affects compression. Before denoising, noise creates random pixel variation that is hard to compress. After denoising, large areas of similar color appear — which compressors encode very efficiently.

Run your denoised image through Pixotter's image compressor after processing. You will often get 20-40% smaller files compared to compressing the noisy original — the compressor has less random data to encode. The key is to compress after denoising, not before. Compressing first (especially with JPEG) introduces blocking artifacts that the denoising algorithm then has to work around, producing worse results than either step alone.

If you also need to resize your image for web delivery, do the operations in this order: denoise → resize → compress. Resizing after denoising prevents the resizing algorithm from blurring already-softened pixels twice. Compressing last means the compressor sees the smallest, cleanest image.

For format questions — whether to output to WebP, AVIF, or JPEG after your cleanup — see our guide on lossy vs lossless compression for the trade-offs. If your goal is to sharpen the image after denoising to recover edge crispness, do the sharpening step between resize and compress. Sharpening a noisy image amplifies the noise; sharpening a denoised image only enhances real edges.

If you need to convert the file format — say, from a denoised TIFF to WebP for web delivery — Pixotter's format converter handles that in the browser with no upload required.

Frequently Asked Questions

Does denoising reduce image quality?

All noise reduction involves a trade-off: you are removing random variation, but some of that variation carries real detail (fine hair, fabric texture, film grain used intentionally). AI-based tools like Camera Raw Denoise and DxO PureRAW 4 are the best at distinguishing noise from detail. Traditional filters (Gaussian blur, even GIMP's NLM at high settings) are less precise. Apply the minimum strength that achieves acceptable results — do not push denoise sliders to maximum by default.

What is the difference between denoise and sharpen?

They are opposite operations that work on the same kind of information. Denoising removes random high-frequency variation (the noise). Sharpening amplifies high-frequency differences (the edges). That is why you should always denoise before sharpening — if you sharpen first, you amplify the noise, then the denoising algorithm has to remove both the original noise and the sharpening artifacts. Denoise first, then apply selective sharpening to recover edge crispness. See our guide to sharpening images.

Can I denoise a JPEG, or do I need a RAW file?

You can denoise a JPEG, but results are typically worse than with RAW. JPEG compression introduces its own artifacts (the 8×8 block pattern characteristic of DCT compression) that denoising algorithms can partially smear into the noise pattern. Starting from RAW gives the denoising algorithm clean data and full bit depth to work with. If JPEG is all you have, use a tool that handles JPEG artifacts specifically — Camera Raw's AI Denoise handles JPEG inputs and attempts to separate compression artifacts from true noise.

Will denoising help with motion blur?

No. Motion blur and noise are different problems. Noise is random high-frequency variation; motion blur is directional smearing caused by movement during exposure. Denoising smooths random variation — it will not restore sharp edges that were blurred by motion. Deconvolution algorithms can partially address motion blur, but they require knowing the blur direction and length, and the results are rarely clean.

How much does image noise affect compression file size?

Significantly. Noise is, by definition, random pixel variation. Compression algorithms work by finding patterns and encoding them efficiently — random variation has no patterns, so it compresses poorly. A noisy ISO 6400 shot can be twice the file size of the same scene shot at ISO 200, at the same JPEG quality setting. Denoising before compressing consistently produces smaller files. The exact savings depend on noise level and compression format: WebP and AVIF have better entropy coding than JPEG and tend to benefit more from pre-denoise cleanup.

Is AI denoising always better than manual sliders?

For most users and most images, yes. AI denoise models are trained on paired noisy/clean images and learn to preserve structures that manual sliders cannot distinguish from noise — fine hair, eyelashes, fabric weave, distant foliage. Manual sliders apply uniform mathematical operations that do not understand image content. The exception: AI models can occasionally introduce subtle hallucinations on high-noise images (small pattern repetitions, softened fine text). When output accuracy matters more than aesthetic quality — forensic photography, technical documentation — compare the AI result against manual NLM reduction before committing to the AI output.