AI Noise Reduction

How to Remove Noise from a Photo

What Causes Image Noise?

Image noise is the random variation of brightness and color that makes photos look grainy or speckled. It is an inherent part of digital photography, but some conditions make it far worse:

• High ISO sensitivity — when your camera amplifies the sensor signal to capture more light, it also amplifies electronic noise. ISO 1600 and above produces visible grain on most cameras, and phone cameras can show noise even at ISO 400.
• Small camera sensors — smartphone and compact camera sensors are physically small, so each pixel receives less light. This makes phone photos inherently noisier than those from full-frame cameras, especially indoors.
• Low-light conditions — dimly lit rooms, evening scenes, and indoor shots without flash force the camera to use higher ISO and longer exposures, both of which increase noise.
• Long exposures — sensor heat during long exposures (several seconds or more) generates thermal noise that appears as hot pixels and colored speckles, particularly in night sky and astrophotography.
• Digital zoom and heavy cropping — zooming digitally or cropping a small area of a photo enlarges existing noise along with the image, making grain much more visible in the final result.
• Aggressive editing — pushing shadows, increasing exposure in post-processing, or heavy HDR tonemapping amplifies noise that was hidden in dark areas of the original image.

Types of Image Noise

Not all noise looks the same. Understanding the type of noise in your photo helps you know what to expect from AI denoising:

AI Denoising vs Traditional Filters

Traditional noise reduction techniques have been available in photo editors for decades, but they all share a fundamental problem: they cannot distinguish between noise and real image detail. AI denoising solves this.

Gaussian blur is the simplest approach — it averages neighboring pixels to smooth out grain. The problem is that it smooths everything equally, destroying sharp edges, fine textures, and small details along with the noise. The result looks plastic and unnaturally soft.

Median filter replaces each pixel with the median value of its neighbors. It preserves edges better than Gaussian blur but produces a characteristic blocky, posterized look. Fine gradients and subtle textures are lost, and the output looks like a watercolor painting rather than a photograph.

Bilateral filter is smarter — it only averages pixels that are similar in brightness, which preserves edges. However, it still struggles with complex textures and tends to over-smooth areas with fine detail like hair, fur, and fabric weave.

AI denoising (Real-ESRGAN) takes a fundamentally different approach. Instead of applying mathematical operations to pixel neighborhoods, it uses a neural network trained on hundreds of thousands of noisy-and-clean image pairs. The model learned what real image detail looks like versus noise, and it can:

• Remove noise selectively — grain and speckles are cleaned while edges, textures, and fine details are preserved or even reconstructed.
• Reconstruct lost detail — the AI predicts what detail should exist beneath the noise and generates it, producing a result that often looks better than simple noise removal.
• Handle different noise types — luminance grain, chroma speckles, banding, and compression artifacts are all recognized and treated appropriately.
• Preserve natural look — the output retains the natural appearance of the photo without the plastic or watercolor artifacts that traditional filters produce.

Best Settings for Noise Reduction

Our noise reduction tool defaults to the settings that produce the best denoising results for most photos:

Quality mode is strongly recommended for noise reduction. The Quality model uses the full Real-ESRGAN architecture with more processing layers, which is critical for distinguishing fine image detail from noise. The Fast model works for quick previews but tends to over-smooth subtle textures that the Quality model preserves.

2x scale is the ideal choice for denoising. It provides the AI with enough processing room to clean up noise while only moderately increasing the image dimensions. If your original photo is already at the resolution you need, the 2x increase is a minor trade-off for significantly cleaner output — you can always resize it back down after denoising.

When to consider 4x scale: use it only if you need both noise reduction and a larger image. For example, if you have a noisy 1000×1000 phone photo and want a clean 4000×4000 version for printing. The denoising quality is the same at both scales, but 4x produces a much larger file.

For extremely noisy images (ISO 6400+), the Quality model at 2x is the best combination. If the result still shows some residual noise, you can run the denoised output through the tool a second time for additional cleaning.