Upscaling makes an image bigger. That is the only thing it reliably does. Everything else you have been told about it depends entirely on which of two very different technologies is doing the work, and the difference between them matters far more than any quality slider.
One kind of upscaler adds pixels by calculation and invents nothing. The other adds pixels by informed guesswork and invents a great deal. Both are legitimate. Both are sold under the same word. If you do not know which one you are using, you do not know whether the detail you are looking at was ever in front of the camera.
The information is not hidden, it is absent
Start with what a digital photo is. A sensor samples the light falling on a grid of photosites and records a number for each one. A 1000 by 750 image contains 750,000 samples. That is the complete record. Anything finer than the spacing between those samples was never measured.
When you enlarge that image to 2000 by 1500, you need 3,000,000 pixels and you have 750,000 measurements. The other 2,250,000 have to come from somewhere, and there is no archive to fetch them from. This is not a limitation of the software. It is a limitation of what was recorded.
The television trope of an analyst saying "enhance" and a licence plate resolving out of four grey blocks is fiction in a specific way that is worth naming: it implies the detail is present but obscured, waiting to be brought out. In a downsampled digital image the detail is not obscured. It is gone, in the same way that the sound of a conversation is gone from a photograph of the room.
Traditional interpolation: honest arithmetic
Interpolation fills in new pixels by looking at the neighbouring ones and calculating a value in between. The methods differ in how many neighbours they consult and what curve they fit through them.
- Nearest neighbour copies the closest existing pixel. Enlarging by 4x simply turns each pixel into a 4 by 4 block. The result is blocky, and that is sometimes exactly right: for pixel art, screenshots of low-resolution UI and anything where hard edges are the content, nearest neighbour is the only method that does not smear them.
- Bilinear takes a weighted average of the four surrounding pixels. Fast, smooth, and noticeably soft. It removes blockiness by blurring, which is a trade rather than an improvement.
- Bicubic looks at a sixteen pixel neighbourhood and fits a cubic curve through it. Better edge definition than bilinear, mild overshoot around high-contrast boundaries, and the default in most photo editors for good reason.
- Lanczos uses a windowed sinc function over a still wider neighbourhood. It is the sharpest of the classical methods and the one most likely to produce visible ringing, the faint light and dark halo that runs parallel to a hard edge.
What unites all four is that the output is a deterministic function of the input. Feed the same image in twice and you get the same result. Nothing appears that was not derivable from the pixels you supplied. If a face was six pixels wide in the original, interpolation gives you a larger, smoother six-pixel-wide face, with no eyes, because there were no eyes recorded.
This is the boring, trustworthy option, and it is what the image upscaler here does: 2x, 3x or 4x through the browser's own high-quality smooth scaling, with no model and no invention. Bigger and smoother, not sharper, and nothing in the output that was not in the input.
AI upscaling: plausible detail that was never there
An AI upscaler works from a different premise. It has been trained on very large numbers of image pairs, each a high-resolution photograph and a downscaled copy of it, and it has learned the statistical relationship between them. Given a low-resolution input it produces the high-resolution image that is most consistent with what it learned.
The results can be startling. Skin gets pores. Bricks get mortar lines. Foliage gets individual leaves. Hair gets strands. None of that was measured by the sensor. It is the model's answer to the question "given how this looks small, what did it probably look like large", and the answer is drawn from other photographs of other faces, other bricks and other trees.
For a great many purposes this is genuinely useful and the honest description is still "a plausible reconstruction". The model is not recovering your image. It is generating a new image that is compatible with your image at low resolution. Run the same input through twice with different settings and you can get different pores in different places, and both will look convincing, because looking convincing is precisely what the model was optimised for.
The failure mode is not that it looks bad. The failure mode is that it looks great and is wrong.
Where invented detail becomes a real problem
Text. This is the clearest case, because you can check the answer. Small text in a photo is a handful of pixels per character, and the model has learned what letterforms look like, so it will produce crisp, confident, well-formed characters. They are frequently the wrong characters. A serial number, a licence plate, a street sign, a price tag, a handwritten note: the output will be sharp and legible and you have no way to tell from the image whether it matches what was photographed. Text is the one place where a blurry original is more truthful than a clean enlargement.
Faces. Face upscaling is where the technology is most impressive and most dangerous. Models are heavily trained on faces, so they will confidently supply eye detail, skin texture and the exact geometry that reads as a specific person. Identity lives in small differences, and small differences are what was lost. An enlarged face is a face the model considers likely, not the face that was there. Using one to identify someone is using a guess as evidence.
Anything forensic or documentary. Insurance photos, accident scenes, medical images, security footage, journalism, condition reports on goods you are selling. In all of these the value of the image is that it records what was in front of the lens. Enhancement breaks that contract, quietly, in a way that is invisible in the output file. This is why the widely discussed episode of a phone maker's long zoom producing detailed photographs of the moon, later shown to be adding detail the camera had not resolved, generated so much argument: the pictures were beautiful and they were not photographs of what the sensor saw.
Product photography, with a caveat. Selling something means representing it accurately. An upscaler that invents fabric weave, wood grain or stitching is showing a customer a texture your product may not have, and a return over a misleading photo costs more than reshooting it. This is a smaller problem when you are enlarging an already sharp photo of a product than when you are rescuing a bad one.
When upscaling is entirely fine
Most upscaling is harmless, and the anxiety above applies to a minority of cases.
Small enlargements. Going up by 1.5x or 2x from a decent original is a modest ask. Interpolation handles it with a slight softening that is often invisible at normal viewing size, and the gap the software has to fill is small enough that even an AI model has little room to invent much.
Anything viewed at a distance. Print resolution rules of thumb exist because the eye resolves less detail further away. Around 300 PPI is the standard for something held in the hand. A poster on a wall viewed from a couple of metres is fine at roughly half that, and large-format signage viewed from across a road is fine at a small fraction of it. An image too small for an A4 print can be entirely adequate as a poster, without any enhancement at all.
Backgrounds and decorative use. A hero image behind a heading, a blurred backdrop, a texture. Nobody is inspecting it, so nobody is misled by it.
Illustration, graphics and logos. These are usually better solved by finding the vector original, but where you cannot, upscaling flat-color artwork is far more forgiving than upscaling a photograph, because there is less real detail to fake.
Meeting a hard minimum. Marketplaces, print services and stock libraries often reject files below a stated pixel size. Enlarging to clear a threshold when the image already looks fine at its native size is a formality, not a deception.
Problems that look like resolution and are not
A large share of images people try to upscale are not suffering from a lack of pixels at all, and enlarging them makes the actual fault bigger.
JPEG compression damage shows as 8 by 8 pixel blocking and colored fringing around edges, particularly in images that have been saved and re-saved through several rounds. Upscaling enlarges the blocks. The fix is a better source file, not more pixels.
Motion blur and focus miss are optical failures. The information was never resolved by the lens, so there is nothing for any amount of interpolation to sharpen, and an AI model will simply invent whatever it thinks a sharp version would look like.
Wrong display size is the most common of all: a 400 pixel image stretched by CSS to fill an 800 pixel container looks terrible, and the file is not the problem. If you are not sure which of these you are looking at, why images look blurry or pixelated works through how to tell them apart.
Before reaching for an upscaler, find the original. The version in your camera roll, in the email attachment, on the original site, in the client's shared folder. Recovering the source solves the problem completely, and it takes less time than most people spend on the alternatives.
How to decide, in practice
- Can you get the original at full size? If yes, stop here. Nothing below beats this.
- Does the image need to be truthful? Evidence, documentation, identification, product listings, anything someone will make a decision from. If yes, use interpolation only, or leave it alone.
- How far are you enlarging? Up to about 2x, interpolation is usually enough. Beyond 4x you are asking for mostly invented pixels regardless of method.
- How will it be seen? Work out the actual viewing distance and required PPI before assuming you need more pixels. Posters and signage need far fewer than people expect.
- Is the fault really resolution? Check for compression blocking, blur and CSS stretching first.
- If you do use an AI upscaler, keep the original and know which file is which. Six months later the enhanced version will look like a photograph, because that is what it is designed to look like.
One practical note on privacy. Most AI upscalers run on someone else's servers, which means uploading your file to a third party, and that is a genuine consideration for client work, images of people, medical or legal material, and anything under an NDA. Interpolation needs no server at all, which is why plain enlargement runs locally in your browser here and nothing leaves your machine.
The honest summary
Interpolation gives you a larger version of the image you have. It will look softer, because you are spreading the same information over more pixels, and there is no arrangement of arithmetic that avoids this. Within a modest enlargement that softening is invisible in normal use, and the result carries no claims that the original did not make.
AI upscaling gives you a larger image that looks sharper than your original ever was, containing detail nobody photographed. That is a fair trade for a wallpaper, a print viewed from across a room, or a background texture. It is not a fair trade when the picture is supposed to be a record of something.
Both are worth having. The mistake is not choosing one over the other. The mistake is not knowing which one you used.
