How to Upscale AI Images Without Losing Detail

Generators usually return something around one megapixel. That is fine on a screen and nowhere near enough for print. Enlarging it is straightforward, but there is an important distinction between two very different things people both call upscaling.

Summary

There are two kinds of upscaling and they answer different questions. Pick deliberately.

Resampling makes the same picture bigger. Honest, predictable, slightly soft.
Generative upscaling invents detail that was never there. Looks better, and is not true.
For a generated image, 2x is clean, 3x is acceptable, 4x is the practical ceiling.
Always scale first, sharpen once, then stop.
Start from the original PNG, never a re-compressed JPG.

The Two Kinds of Upscaling

This distinction matters more than any setting.

	Resampling	Generative
What it does	Works out new pixels from the ones around them	Predicts what detail probably belongs there
Result	Slightly soft, faithful	Sharp, partly invented
Adds detail?	No	Yes, and it was never in your file
Safe for documents and evidence	Yes	No
Best for	Anything where accuracy matters	Decorative art and illustration

Careful: A generative upscaler will happily invent eyelashes, brick texture and text that never existed. For a fantasy illustration that is a feature. For a photo of a document, a face, or anything someone might rely on, it is a problem.

How Far You Can Realistically Go

These figures assume you started from a clean 1024 pixel generation, not a screenshot of one.

Realistic quality at each scale factor, resampling a 1024 pixel source

The rule underneath those numbers: if the detail is not in the original pixels, no resampler can recover it. It can only produce a larger, smoother version of what you already had.

Doing It in the Browser

The image upscaler does the whole job in one step. It uses Lanczos resampling, which is the best of the classic algorithms, then runs a sharpening pass so the result does not look like a soft blob.

Drop the original PNG in. Not a screenshot, not a re-saved JPG.
Pick 2x to start. You can always come back and try 3x.
Leave sharpening on Medium. Use Strong only for line art and screenshots.
Choose PNG if the image has transparency, JPG if not and file size matters.

Doing It in GIMP

More control, and the same underlying idea.

Open the image and go to Image, Scale Image.
Set the width. The height follows while the chain icon is linked.
Change Interpolation to NoHalo. This is the important step and most people skip it.
Press Scale.
Run Filters, Enhance, Sharpen (Unsharp Mask). Radius around 2, amount modest, threshold 0.
Export.

Interpolation	Use it for
None	Pixel art only. Keeps hard squares.
Linear	Fast, and worse than the others. Skip it.
Cubic	The old default. Acceptable, slightly soft.
NoHalo	Best for enlargement. Avoids the bright ringing around edges.
LoHalo	Best for reduction. Use when scaling down instead.

Tip: NoHalo for making things bigger, LoHalo for making things smaller. The names describe the artefact each one avoids, and remembering which is which is most of the benefit.

Do You Actually Need to Upscale?

A lot of upscaling is unnecessary. Work out what the destination requires before you enlarge anything.

Destination	Pixels needed	Is 1024px enough?
Blog image	1200 wide	Almost
Full width web header	1920 wide	No, 2x it
Instagram post	1080 square	Yes
YouTube thumbnail	1280 x 720	Yes
6 x 4 inch print at 300 DPI	1800 x 1200	No, 2x it
A4 poster at 300 DPI	2480 x 3508	No, 3x and accept the softness
A3 poster at 150 DPI	1754 x 2480	No, 2x it

The DPI and print size calculator works this out for any size, and the resolution calculator goes the other way if you know the print size you want.

Worth knowing: Printing at 150 DPI instead of 300 halves the pixels you need, and for anything viewed from more than about a metre away the difference is genuinely hard to see. A poster does not need 300 DPI.

The Order That Matters

Repair and grade at the original size, then scale, then sharpen once at the end

Two mistakes account for nearly every disappointing upscale:

Sharpening before scaling. The added definition is thrown away by the resampler, and any halo it created gets enlarged.
Sharpening more than once. Each pass builds on the last one's halos. Two moderate passes look worse than one.

Retouching also belongs before scaling. Repairs made at native size and then enlarged blend far better than repairs painted onto an already smoothed image. That is why the full editing workflow puts upscaling second to last.

When Upscaling Will Not Help

The problem	Why upscaling fails	What to do instead
The image is out of focus	Blur is missing information, not missing size	See fixing blurry AI images
Heavy JPG blocks	The blocks get enlarged too	Regenerate and save the original PNG
Broken hands	Now they are bigger and more obvious	See fixing hands and faces
Garbled text	Larger gibberish is still gibberish	Remove it and add a real text layer

Enlarge an image now

Lanczos resampling with a sharpening pass, up to four times. Free, no watermark, no account.

Open the tool

Common Questions

What is the maximum I can safely enlarge?

Two times is the sweet spot and looks essentially clean. Three times is soft up close but fine at arm's length. Four times is visibly smoothed and useful for print size rather than detail.

Should I use a generative upscaler instead?

It depends on what the image is for. A generative upscaler invents plausible detail, which looks better and is not true. That is fine for decorative art and wrong for anything documentary.

Why did my upscale come out blocky?

You enlarged an already compressed JPG. The compression blocks get enlarged along with everything else. Start from the original PNG the generator gave you.

Do I sharpen before or after?

After. Always. Sharpening first then scaling throws the added definition away, and scaling first then sharpening once is what the whole workflow depends on.

Upscaling AI Images Without Losing Detail