Upscale an image for print, free and in your browser
Choose the print size and dpi, and see whether your picture already has enough pixels. If it falls short, an AI model enlarges it for you, tile by tile, and saves a PNG at exactly that size. Free, nothing is uploaded, and it tells you where the model guesses.
Faster with the free Chrome extension: right-click any picture on a website and send it straight to the print upscaler. Add to Chrome · what it does
Extension installed. Right-click any picture on a website and choose Makerfern → Open this picture in… to bring it here.
Enough pixels for the size you want to print
A print looks sharp when it has about 300 pixels for every inch. A picture from the web, an AI image of 1,024 px or a crop from a phone photo often has a third of that, and printing it larger just stretches the pixels. This tool does the sum for the size and dpi you choose, says whether the picture already has what it needs, and if it does not, enlarges it with a model that has learned what edges, fur and foliage usually look like at a higher resolution.
What it does
- The sum first. Set the print width or height in inches, centimetres or millimetres and the dpi. It tells you the pixels the print needs, the dpi you have now, and how much the picture must grow. If it already has enough, nothing is enlarged.
- Enlarges four times, tile by tile. The picture is cut into 128 px tiles with an overlap, so seams do not show, and each tile is enlarged by the model and then resampled straight to the print size. A huge canvas is never built, which is why large prints work on modest devices.
- Looks at a detail before the long run. Tap the picture, choose a spot, and see it at 100% as plain resizing next to the model's result, with a slider. It takes a few seconds and shows you where the model helps and where it invents texture.
- Logos and lettering go the other way. Flat art gets a high-quality Lanczos resample instead of the model, which draws halos round hard edges. The page tells you when it has done that.
- A PNG that prints at the right size, with the dpi stored in the file and transparent parts kept.
- Honest about time. You see an estimate before you start, and a run that would take many minutes on a device without WebGPU is warned about or limited.
What it does not do
- It does not restore detail. It guesses it, and a guess can be wrong, especially on faces, small text and fine patterns.
- It does not rescue a blurry or badly out-of-focus picture, and it was not tested on text under about eight pixels tall.
- It does not convert to CMYK or embed a colour profile.
- It does not make a small image look like a large original. Past four times, the extra enlargement is a plain resample.
Where to look before you print
- Open the PNG at 100%, not fitted to the screen, and look at the parts people look at: faces, eyes, lettering, edges.
- Look for texture that was never in your picture: smooth skin turned to fur, brickwork that was not there, repeating patterns.
- Print at 100% ("actual size"), never "fit to page", so the file keeps its dpi.
- If the result is not good enough, a smaller print, a lower dpi or a bigger original will beat any enlarging.
Enlarging for a halftone or a DTF transfer? Enlarge first, then open the file in the halftone generator or the DTF-ready PNG tool, which work at the pixels you give them.
Questions
Does it restore detail that is missing from my picture?
No. The model looks at the pixels you have and guesses what fine detail would plausibly be there. Edges come out cleaner and JPEG blocks and speckle are smoothed away, which is a real improvement over stretching the pixels, but the extra detail is a guess. On a face, fur, small text or a fine pattern the guess can be wrong: a different letter, texture that was never in the photo. Look at the result at 100% before you print, and for anything where the exact detail matters (a portrait, a product label), start from a bigger original.
How many pixels does a print need?
Pixels divided by inches is dots per inch, so the pixels a print needs are its size in inches times the dpi. At 300 dpi, an 8 × 10 in print needs 2,400 × 3,000 px and a 12 in wide print needs 3,600 px across. A 1,024 px picture is 3.4 in wide at 300 dpi as it is, and 13.7 in wide once it has been enlarged four times to 4,096 px. The page does this sum for the size and dpi you choose, and tells you whether the picture already has enough pixels, in which case it does not enlarge at all.
How much can it enlarge?
The model enlarges four times, tile by tile. If the print needs less than that, each tile is enlarged four times and then reduced to the print size. If it needs more than four times, the model does its four and a plain resample covers the rest, so a very small picture still looks soft at a large size. Past eight times the model is guessing most of the picture, and the page warns you. The longest side of a print is limited to 30,000 px and its area to 120 million pixels (36 million on a phone); a bigger print is made at a lower dpi, at the same printed size, and the page says so.
Why was my logo enlarged without the model?
Flat art, such as a logo, lettering or clipart, is made of a few flat colours with hard edges. The model tends to draw halos and darkened edges round those, and can redraw small letters as different letters. So the page detects flat art and uses a high-quality Lanczos resample instead, which keeps the shapes exact; it says when it has done so. A resample cannot add sharpness, though: for a logo that has to be printed large, tracing it into a vector with the converter removes the size limit altogether. You can choose "AI enlarge" yourself to compare the two.
How long does it take?
It depends on the browser. With WebGPU (recent Chrome and Edge on a computer, and some phones) a 1,024 × 768 px picture took about 6 seconds in our test on a laptop. Without it the model runs on the processor, and the same picture took about 4 minutes. Phones were not timed, so the page can only estimate for them, and says so; it also takes smaller pictures on a phone without WebGPU, so a run never lasts many minutes without warning. The page shows an estimate before you start, corrects it from the speed it measures, and you can cancel at any time. The first run downloads the model, about 11 MB with its runtime, once.
Is my picture uploaded?
No. The picture is read, enlarged and written into the PNG in your browser, and saved from memory on your own device. The only download is the model itself, which your browser keeps for next time. Close the tab and nothing is left.
What do I get?
A PNG at exactly the print size you chose, with the dpi stored inside the file so print software places it at the right size. Transparent parts stay transparent. A picture with no see-through parts is saved as a plain RGB PNG; one with transparency is saved as RGBA (RGB plus a transparency channel). Either way the file carries no colour profile and is not converted to CMYK, so ask your print shop whether it wants one. The PNG is written in bands, so a large print does not need one enormous canvas.
Which dpi should I choose?
300 dpi is the usual target for something held in the hand, and most DTF and sticker shops ask for it. A poster seen from a few steps back is fine at 150 to 200, and a banner read from across a room needs much less. Choosing a lower dpi means fewer pixels to guess, so a faster and more faithful result. These are rules of thumb, not a manufacturer's specification: your print shop's own figure wins.
Can it fix a blurry or out-of-focus photo?
Not reliably. It sharpens edges and removes compression noise, but a blurred picture has lost its detail, and the model has nothing to base a guess on. It also was not tested on text smaller than about eight pixels tall, so tiny lettering may come out wrong.
What about several images at once?
One image at a time is free. With Pro you can drop up to 20 images, choose one print size and dpi, and get one zip with every PNG and a report that lists, for each image, its size, dpi, enlargement and whether the model or the plain resample was used. Each image keeps its own shape, so only one side is fixed. To halftone them next, open the halftone generator with the enlarged files.
Whose model is it?
Real-ESRGAN, the realesr-general-x4v3 model by Xintao Wang and colleagues, released under the BSD 3-Clause licence (licence and notice). The model file is converted, not changed: its weights are the released ones. Other tools use the same family of models; this one runs it in your browser, tied to a print size and dpi.