So is AI upscaling real, or is it a trick?

Both words are wrong. A trained model really does produce results that look better than anything arithmetic can manage on a soft source, so it is not a trick. But it does that by predicting plausible detail rather than recovering recorded detail, so it is not recovery either. The correct sentence is that it substitutes, convincingly, and whether that is what you want depends on whether the picture is decoration or evidence.

If nothing is recovered, why does the enlarged file look better than the small one?

Often because the comparison is unfair. A 500-pixel image shown on a 1,500-pixel screen is already being enlarged, by the browser, with the cheapest filter it has. Doing the same job deliberately with a better filter produces a better-looking picture at the same information content. That is a genuine improvement in presentation and no improvement at all in what was recorded.

How big can I safely go?

A rule that holds up in practice: double it and trust it, quadruple it and check it at 1:1, eight times it and only for something nobody will stand close to. The variable is not the factor, it is how much genuine detail the source had per pixel to begin with — a sharp 600-pixel photo survives 4× better than a soft 1,200-pixel one.

What enlarging can and cannot get back

A pixel is a measurement. Between two measurements there is no third measurement — there is only whatever the world happened to contain, and nothing in the file knows what that was.

The part that is arithmetic

Every enlarger starts from the same position. You have a grid of measurements and you want a denser grid. The new positions between the old ones were never sampled, so a value has to be produced for each of them, and the only material available is the values nearby.

Weighting those nearby values is what a kernel does, and the differences between kernels are real but narrow. Lanczos weights seven neighbours along a curve that rings slightly; Mitchell weights four along a curve that does not; an edge-guided method decides which neighbours to trust before it weights them. All three produce a smooth blend of things already known, and it follows directly that none of them can put a frequency into the result that was not in the input.

If the thread in a fabric was never resolved by the camera, no arrangement of weights resolves it now. What changes is how gracefully the absence is described.

That is not a criticism. Describing an absence gracefully is most of what a print needs. A poster at two metres asks for smooth tone and clean boundaries, not for thread count.

The part that is a guess

The other family of enlargers does something categorically different. A network trained on millions of photographs has learned what a blurred eyelash usually looks like sharp, what a soft brick edge usually resolves into, what a two-pixel letterform usually turns out to be. Given a small picture it produces the detail those patterns imply.

This works startlingly well and it is a substitution, not a recovery. The distinction matters most exactly where people care most:

Text comes back confidently wrong.A model has seen far more sharp letters than your blurred ones, so it will give you a crisp glyph. There is no mechanism that makes it the right glyph, and nothing in the output distinguishes the two cases.
Faces come back as a plausible person.Below about sixty pixels a face has no individual features left to enlarge. What a model produces is a face consistent with the blur — the correct skin tone, a sensible eye shape, a nose that fits. It is not the person, and the sharper it looks the more convincing the error.
Nothing in the file marks which pixels were invented.The output is an ordinary image. A viewer cannot tell the computed regions from the recorded ones, which is why a reconstruction should never be the copy you keep or the copy you send on as the original.

This build ships only the arithmetic path, so nothing you produce here is invented in that sense. That is a deliberate limitation and it is also, for some of the jobs people bring here, the safer answer.

Worked through, factor by factor

A 600 × 400 photograph, taken carefully, as it goes up
Factor	Result	What you can still believe
2×	1,200 × 800	Nearly everything. Edges are edges, texture is texture.
4×	2,400 × 1,600	Shapes and tone. Fine texture is now a smooth description of texture.
8×	4,800 × 3,200	Composition and colour. Anything smaller than a fingernail is inference.

The same table for a source that was already soft

Halve every row. A picture that was out of focus, shot in poor light or saved three times by a chat app has less recorded detail per pixel than its dimensions suggest, and the enlargement inherits that rather than the pixel count. This is why two 800-pixel files can behave completely differently at the same factor.

A test you can run in ten seconds

Open your picture, enlarge it, press 1:1 and hold the compare button. You are now looking at your result and at the same source stretched by the browser's own cheap filter, in the same rectangle, at the same magnification.

If the two look different, the method is earning its keep. If they look the same, the source has run out of detail and no tool of either family will change that — the useful move is a bigger original, not a bigger factor.

Questions people arrive with

So is AI upscaling real, or is it a trick?: Both words are wrong. A trained model really does produce results that look better than anything arithmetic can manage on a soft source, so it is not a trick. But it does that by predicting plausible detail rather than recovering recorded detail, so it is not recovery either. The correct sentence is that it substitutes, convincingly, and whether that is what you want depends on whether the picture is decoration or evidence.
If nothing is recovered, why does the enlarged file look better than the small one?: Often because the comparison is unfair. A 500-pixel image shown on a 1,500-pixel screen is already being enlarged, by the browser, with the cheapest filter it has. Doing the same job deliberately with a better filter produces a better-looking picture at the same information content. That is a genuine improvement in presentation and no improvement at all in what was recorded.
How big can I safely go?: A rule that holds up in practice: double it and trust it, quadruple it and check it at 1:1, eight times it and only for something nobody will stand close to. The variable is not the factor, it is how much genuine detail the source had per pixel to begin with — a sharp 600-pixel photo survives 4× better than a soft 1,200-pixel one.