How AI Image Upscaling Works (and Why It Beats Normal Resizing)

Make a small photo bigger in a normal image editor and it turns soft and blurry. Run the same photo through an AI upscaler and edges stay crisp, hair looks like hair and text stays readable. The difference comes from a completely different approach to the problem. This article explains both approaches in plain language, so you know what is really happening and when to trust the result.

The problem: inventing pixels that do not exist

Doubling an image's width and height means four times as many pixels. A 1000 × 1000 image has one million pixels; at 2x it has four million. Three out of every four pixels in the new image have to come from somewhere. Every upscaling method is really a strategy for filling them in.

Traditional resizing: interpolation

Image editors traditionally use interpolation, which calculates each new pixel from the pixels around it.

• Nearest neighbour copies the closest existing pixel. It is fast and keeps hard edges, but produces a blocky, pixelated look. It is mainly used for pixel art.
• Bilinear takes a weighted average of the four nearest pixels. Smooth, but noticeably blurry.
• Bicubic uses the sixteen nearest pixels and a curve to blend them. Slightly sharper than bilinear and the default in many editors, but still soft at large factors.
• Lanczos uses a wider neighbourhood and is a little sharper again, at the cost of faint ringing near edges.

All of these methods have the same limitation: they can only blend what is already there. A sharp edge between two pixels becomes a smooth gradient spread across several pixels, which our eyes read as blur. No amount of averaging creates new detail.

AI upscaling: super-resolution

AI upscalers use a technique called super-resolution. Instead of blending neighbours with a fixed formula, a neural network predicts what the larger image most likely looked like.

The network learns this from examples. During training, researchers take a large collection of high-resolution photos, shrink each one, and show the network the small version. The network tries to reconstruct the original, compares its attempt with the real thing, and adjusts itself. After millions of examples it has learned how real-world textures, edges and shapes typically look at higher resolution: that a blurry diagonal is probably a crisp edge, that a soft brown patch with a certain pattern is probably fur, that small dark curves on a light background are probably letters.

ESRGAN: the model behind many upscalers

ESRGAN (Enhanced Super-Resolution Generative Adversarial Network) was published in 2018 and became one of the most widely used upscaling models. It is the model family our AI image upscaler uses. Its key idea is to train two networks against each other:

• A generator that produces the upscaled image.
• A discriminator that tries to tell the generator's output apart from real high-resolution photos.

As the generator learns to fool the discriminator, it stops producing the safe, blurry average that simpler models fall back on and starts producing realistic texture. That is why ESRGAN results look sharp and detailed instead of merely smooth.

What AI upscaling does well

• Edges and outlines: building edges, horizon lines, product outlines.
• Natural textures: grass, foliage, fabric, hair, brick.
• Cleaning up compression: combined with denoising, it removes many JPG blocks and halos while enlarging.
• Small sources: thumbnails, images saved from chats, old phone photos and screenshots gain the most.

What it cannot do

It is important to be honest about the limits. The network predicts plausible detail; it does not recover the true detail that was lost.

• It cannot fix focus or motion blur. If the detail was never captured, there is nothing to rebuild.
• It may guess wrong on very small features. A face only a few pixels wide may come out generic, and small text can be reconstructed into wrong letters.
• It should not be used as evidence. Upscaled images are not suitable for reading licence plates in security footage or for forensic purposes, because the added detail is invented.
• Larger factors increase the guessing. At 8x, 63 out of every 64 pixels are predicted. Results are best when the source is clean and the factor is modest.

Why ClearPix runs the AI on your device

Most online upscalers upload your photo to a server with powerful graphics cards. ClearPix instead downloads a compact version of the model (about 5 MB) and runs it in your browser using your device's graphics hardware through WebGL. That means your photos stay private, there is no queue and no daily limit.

The trade-off is memory. Phones and older laptops have less of it, so the AI pass runs on the image at up to 768 pixels on the longest side (384 on phones), and any extra enlargement beyond that is done with standard resizing. In practice this makes the tool most effective exactly where AI upscaling helps most: small and low-quality images. Read more about the privacy side in why on-device image tools are safer.

Getting the best result

1. Start from the cleanest, largest original you have.
2. Pick the smallest factor that reaches the size you need.
3. Use denoise for compressed or grainy images, but keep it low for clean ones.
4. Compare before and after at 100% zoom, not just as a thumbnail.

If you are upscaling for print, our guide to how many pixels you need to print will tell you which factor you actually need.