Upscale an image without losing quality: what is really possible
You can upscale an image without losing quality, but only up to a point, and the point depends far more on the state of your original than on the size you want. This article explains what an AI upscaler actually does, why ×2 to ×4 is the sensible range, which order of steps gives the cleanest result and how to check it before you download.
Why enlarging a small image always used to look soft
For decades, every request to enlarge a small image went through the same arithmetic. The software took the pixels that already existed and spread them over a bigger grid. Methods called nearest neighbour, bilinear and bicubic differ only in how they fill the gaps: copying the closest pixel, averaging the two nearest, or fitting a smooth curve through sixteen of them.
None of them adds detail. If a strand of hair was two pixels wide in a 640 × 480 photo, after a bicubic ×4 it is eight pixels wide, but it is still one blurry stroke, now four times larger. The detail per centimetre stays exactly where it was; the image simply takes up more space. That is why a 'bigger' file printed at A3 looked mushy, why edges grew soft halos and why text turned into grey fog.
So 'upscale without losing quality' sounded like a scam, and for classic interpolation it largely was. You did not lose information, but you did not gain any either, and at viewing size the softness reads as a loss. The useful mental model: interpolation changes the pixel count, not the resolution of the picture. Two images can both be 5120 × 3840 px and carry completely different amounts of real detail.
This distinction matters later, because some files that look 'high resolution' on paper were already enlarged this way by a phone, an editing app or a website. Upscaling them again starts from a soft base, not a sharp one.
What an AI upscaler actually does when you increase the resolution of a photo
A trained upscaling model works differently. It has seen millions of pairs of low and high resolution images and has learned what sharp edges, skin, foliage, brick, fabric and text usually look like at a finer scale. When you ask it to increase the resolution of a photo, it does not stretch the pixels: it predicts, for each area, the most plausible high resolution version that would have produced the small one you gave it.
The result can look startling. Edges come out crisp instead of haloed, grass gets individual blades, a wall gets texture. But it is worth being blunt about what is happening, because the marketing usually is not: this is reconstruction, not recovery. The model is not finding hidden detail inside your file. Information that was never captured, or that compression threw away, does not exist anywhere in the image. The model invents detail that is consistent with what is there.
In practice that means three things:
- Textures are plausible, not exact. The bark on an enlarged tree looks like bark, but it is not the bark of that particular tree. For a landscape print, that is exactly what you want.
- Unreadable text stays unreadable. If a number plate or a street sign is a smudge of six pixels, no upscaler can tell you what it said. It may produce something letter-shaped, and that is worse than useless.
- Faces are the risky zone. We are very sensitive to small changes in a face, so an invented eyelid or tooth shape is noticed immediately. That is why faces get their own step later.
A good model is also conservative: it sharpens what it is confident about and leaves ambiguous areas smooth. General chatbots that 'enhance' a photo behave differently, often redrawing the whole scene at a modest size; we covered why in enhancing a photo with an AI chatbot.
How far can you upscale a photo? The sensible limit
Here is the rule we would give a friend. From ×2 to ×4, almost any decent photo holds up. Beyond that, it depends entirely on the condition of the original, and the megapixel count on the label is a poor predictor.
A clean 2 MP photo (say 1600 × 1200 px, well exposed, saved once as a high quality JPEG or PNG) will often give a better ×4 than an 8 MP photo that has been sent through three messaging apps. The first has little detail but clean edges for the model to build on. The second has more pixels, but many of them are compression blocks, smeared colour and noise, and the model faithfully enlarges those too: blocks become crisp blocks.
Some practical thresholds, based on what the file looks like at 100 % before you start:
- Sharp at 100 %, no visible blocks: ×4 is safe. ×8 or more is reasonable for prints viewed from a distance, such as posters and canvases.
- Slightly soft, light JPEG artefacts: ×2 to ×4, after cleaning (next section).
- Heavy blocks, colour banding, a screenshot of a screenshot: ×2 at most, and expect a clean but smooth result rather than a detailed one. Our article on the pixelated WhatsApp photo goes into what can and cannot come back.
- Tiny thumbnails under about 300 px: the model has very little to work with. You will get a usable small print or screen image, not a sharp poster.
For print, the target is not 'as big as possible' but 'enough pixels for the viewing distance'. A photo seen at 30 cm wants around 300 ppi; a poster seen from 2 metres is fine at around 120 ppi. That alone often means ×2 is enough where people reach for ×8. The guide on how to upscale an image for large-format printing has the full table.
One honest warning, because the usual advice gets it backwards: do not sharpen before you upscale. A sharpening filter adds thin bright and dark outlines along edges. The model takes those halos as real edges and enlarges them into clear double lines. If your phone or editor already applied heavy sharpening, you will see this; there is no setting that fully removes it afterwards.
How to improve the resolution of a photo: the order that changes the result
Most disappointing upscales are not caused by the upscaler but by the order of operations. Every defect in the input is enlarged along with the subject, so the goal is to hand the model the cleanest possible base. This is the sequence we use:
- Clean noise and compression first. Grain from a night shot or a high ISO setting, and the blocks from heavy JPEG compression, look like texture to an upscaler. Run the photo through a remove noise from a photo step at moderate intensity before anything else. Use the strength slider sparingly: too much and skin turns waxy, and the upscaler will enlarge the wax.
- Sharpen only if the photo is actually blurred. Camera shake or a missed focus point is a different problem from low resolution. If the subject is smeared in one direction, or the whole frame is soft, use unblur an image before upscaling. If the photo is merely small but sharp, skip this step; deblurring a sharp image only adds artefacts.
- Upscale. Now the model works from clean edges. Choose the factor you need for the final size, not the largest available.
- Faces last, at medium fidelity. Face reconstruction works best on a face that is already clean and enlarged. In Airpic you can tick the face option in the upscaler, or, for old portraits where likeness matters most, use the restore old photos tool, which has a fidelity slider. Around the middle of that slider (roughly 40-60 %) the result stays closer to the real person; at full strength, on a very degraded face, the model tends to produce a generic, too-perfect face.
Why faces last and not first: a face restored on a tiny, noisy crop and then enlarged loses the benefit, and a face restored before denoising mixes invented detail with grain. Doing it at the end also lets you judge it on the final size, which is the only size that matters.
If the faces in the original are very small (a group photo where each face is a few dozen pixels), keep expectations modest. The tool tells you when faces are too small to reconstruct reliably, and that is a signal to stay at ×2 rather than push further.
Step by step: upscale an image in Airpic and check it at 100 %
Here is the exact process with the Airpic upscaler, including what to choose and why. You do not need an account, and processing plus viewing the result is free.
- Load the photo. Drag it in, pick it from your files or paste it with Ctrl+V. JPG, PNG (including transparency and 16-bit), WebP, TIFF, HEIC and AVIF are accepted. The EXIF orientation is applied and the ICC colour profile is kept, so a portrait from an iPhone arrives the right way up with its colours intact. A card shows the input size and the output size you are about to get.
- Choose the scale. ×2 and ×4 run as a single direct pass. ×8, ×16 and Max use a tiled mode that can reach up to 600 megapixels. Start with the smallest factor that reaches your target size: if you know the print size, set it under the usage option for printing in large format and the tool recommends a factor for your viewing distance.
- Pick the style, if you go beyond ×4. Under Advanced settings, the sharpness style keeps edges crisp and suits photos of buildings, products and landscapes; the painterly fill gives softer, more continuous texture and suits skin, skies and photos that were never very sharp. If in doubt, try both on the same file: processing is free.
- Decide on faces. Tick the face enhancement option only if there are people and their faces look soft after the first pass. On a clean, sharp portrait it is often better left off.
- Choose the output format. Leave it on automatic unless you have a reason: an image with transparency stays PNG, a JPEG input comes out as a high quality JPEG, and very large PNG results switch to JPEG so the file stays manageable. For the web, convert afterwards with the compress images tool, which is completely free.
- Process and wait a few seconds. In our tests on 19 September 2026, a 320 × 240 px photo upscaled ×4 to 1280 × 960 px in 5 seconds. On 3 September 2026, photos of 0.3 to 1 megapixel took between 6 and 9 seconds at ×4, and about 17 seconds with face enhancement switched on. The first job after a quiet period can take longer while a GPU in the EU starts up.
- Compare at 100 % before downloading. Drag the divider across the before and after view, then press View at 100 %. This is the step most people skip and the one that matters: at screen size everything looks good. At 100 %, check eyes, hair, text and fine repeating patterns. Look for double edges (a sign of earlier sharpening), crisp compression blocks (clean the original first) or a face that no longer looks like the person (lower the strength or use the restorer).
- Download or repeat. If something is off, use Repeat with other settings, change one thing and process again. When it looks right, download. You get one full download free each month; after that, credits that never expire. Your file is deleted one hour after download, one day at most, and it is never used for training.
Ready to try it on your own photo? Open the upscale an image without losing quality tool, run ×2 first, and judge it at 100 % before you go any bigger.
FAQ
how to upscale an image without losing quality
Use an AI upscaler rather than plain resizing, clean noise and compression first, and stay between ×2 and ×4 unless the original is sharp at 100 %. Then compare the result at 100 % before downloading: that is where any loss shows.
can you increase the resolution of a photo after it was taken
You can increase the pixel count and get sharper edges and plausible texture, but the model reconstructs detail rather than recovering it. Anything that was never captured, such as a blurry number plate, cannot be brought back.
how big can I enlarge a small image
×2 to ×4 works for almost any decent photo. Beyond that it depends on how clean the original is: a sharp 2 MP photo can go further than a heavily compressed 8 MP one. For prints, size for the viewing distance, not the maximum factor.
does upscaling make faces look different
It can. Face reconstruction at full strength on a very degraded face tends to produce a generic face. Restore faces as the last step and keep the fidelity around the middle so the person still looks like themselves.
We measured it
- A 320 × 240 px photo upscaled ×4 to 1280 × 960 px in 5 seconds. (19-09-2026)
- With 0.3 to 1 megapixel photos, ×4 took 6 to 9 seconds; with face enhancement on, about 17 seconds. (03-09-2026)
Measured on airpic.app with the GPU already running. If nobody has used it for a while, the first job takes 40 to 90 seconds longer to start.
Upscale your image and compare it
Keep reading
PNG vs JPG: which is better, and when to use WebP
PNG vs JPG: which is better for photos, screenshots and logos, when to use WebP or AVIF, and which conversions lose quality. Pick a format by use, not habit.
How to compress images for WordPress
Compress images for WordPress: resize, pick JPEG, WebP or AVIF, set quality and strip metadata so pages load fast and photos look the same.