AI image upscale

Upscale images 2× or 4× without the usual blur

Give small, soft images more usable resolution with Real-ESRGAN. Add face enhancement for portraits, then compare and download the result.

Processed on our secure servers
Drop an image to upscale
JPG · PNG · WebP · HEIC — Pro AI tool
or press Ctrl+V
PNG · JPG · WEBP · .HEIC · .HEIF · … · max 25.0 MB

The complete guide to AI image upscaling

What AI upscaling actually does

When you enlarge an image the ordinary way, the software has an impossible job. Doubling a 500 × 500 photo to 1,000 × 1,000 means filling in three new pixels for every original one, and a conventional resampler — bicubic, Lanczos, or the plain "scale up" in any editor — can only average the neighbours it already has. Averaging produces a smooth, plausible-looking blur. That is why every enlarged photo looks soft: nothing new was added, the existing information was just spread more thinly.

A model like Real-ESRGAN, which powers this tool, approaches the problem from the other direction. It has been trained on millions of pairs of images — a high-resolution original and a degraded small version of the same picture — until it learned what kinds of detail typically go missing when an image shrinks. Given a new small image, it predicts the detail that should be there: the sharp boundary at the edge of a roof, the directional texture of hair, the weave in a fabric.

That distinction matters for setting expectations. Upscaling is reconstruction, not retrieval. The model is making an extremely well-informed guess about detail that no longer exists in your file. On most photographs the guess is convincing enough to be indistinguishable from a real high-resolution capture. On unusual subjects — dense small text, an unfamiliar logo, a face at very low resolution — it can guess wrong, and the result will look confident but slightly invented.

When 2× beats 4×

The instinct is that more is better, and for upscaling that instinct is usually wrong. 2× doubles each dimension and quadruples the pixel count, which is already a substantial jump: a 1,000 × 1,500 image becomes 2,000 × 3,000, comfortably enough for a full-page print or a retina-density hero image. It is also the setting where the model is doing the least guessing, so results look the most photographic.

4× is the right call when the output genuinely needs the pixels. Reasonable cases: a crop from a wider shot that has to stand alone, a product image customers will zoom into, a scan destined for large-format print, or an archival photo you want to work with in an editor afterwards. In those situations the extra reconstruction earns its place.

Where 4× disappoints is on images that were already adequate. Upscaling a 3,000-pixel-wide photo to 12,000 pixels does not make it look better on a screen that will display it at 1,200 — it makes a large file with slightly synthetic-looking texture. It also costs twice the credits and takes noticeably longer. A useful rule: pick the scale from the size you actually need at the other end, not from the size that is available.

If you need more than 4×, do it in a single pass at the largest scale rather than running the tool twice. Each pass treats the previous output as ground truth, so a second run confidently sharpens the first run's inventions. Artifacts compound quickly.

Face enhancement, and when to leave it off

Face enhancement is a separate restoration pass that runs only over regions detected as faces. General-purpose upscaling models tend to under-serve faces: we are extraordinarily sensitive to facial detail, so a face that is merely as good as the rest of the image still reads as wrong. A dedicated face model, applied on top, rebuilds eyes, lashes, lips and skin texture at a level of specificity the general model does not attempt.

On the right source this is the single most dramatic thing the tool does. A face that occupies 60 pixels in an old group photo or a distant snapshot can come back with recognisable structure. If you are restoring family photographs or salvaging a portrait from a low-resolution scan, turn it on.

It has two failure modes worth knowing. On faces that are already sharp, the pass has nothing to recover and instead smooths skin texture, producing the slightly waxy look people associate with over-processed portraits. And on extremely low-resolution faces — where there genuinely is not enough information to determine what someone looked like — it will still produce a confident, detailed face, which may not match the person. That is a property of every face-restoration model, not a bug in this one, but it means you should not treat an enhanced face as evidence of anything.

Face enhancement adds 1 credit on top of the upscale cost. Because it is optional and independently priced, the honest default is off: turn it on when the faces are the reason you are here.

The 64-megapixel ceiling, and how to work within it

Upscaled output is capped at 64 megapixels — about 8,000 × 8,000 pixels, or a 6,000 × 10,000 rectangle. The limit exists because upscaling is GPU work whose memory cost scales with the output, not the input. An uncapped 4× run on a large photo can exhaust a GPU mid-job, which fails slowly and wastes your credits along with our compute. A firm ceiling turns an unpredictable failure into a clear message before anything runs.

The consequence worth internalising is that the limit applies to the output, so it constrains your input by the square of the scale. At 2×, anything up to roughly 16 megapixels is fine — larger than most phone cameras produce. At 4×, the effective input limit drops to about 4 megapixels, which a modern phone photo will exceed comfortably. That surprises people: the 4× option is most useful precisely on the small images that fit within it.

If you hit the ceiling, you have three good options. Run 2× instead, which is often all the image needed. Crop to the region you actually care about and upscale that — usually the right answer for a detail pulled from a wider scene. Or accept that your source is already large enough that upscaling adds little. Input files are separately capped at 25 MB.

What upscaling cannot fix

Being clear about the limits is more useful than overselling the model. Upscaling cannot recover detail that was never captured. An out-of-focus photo has no fine structure to reconstruct, and the result will be a larger, cleaner rendition of something still out of focus. The same is true of motion blur: the model will sharpen the blur's edges rather than undo it.

Heavy compression is the other common disappointment. JPEG artifacts — the blocky 8 × 8 squares, the colour banding in skies, the ringing halo around high-contrast edges — are structured patterns in the pixel data, and a model asked to enlarge that data may enlarge the artifacts along with everything else. Mild compression noise is cleaned up well; a screenshot of a heavily re-shared social post is not. Always start from the largest, least-processed copy you can find. An original camera export beats a messaging-app download by a wide margin, and that difference matters more than any setting on this page.

Finally, upscaling does not reliably reconstruct text or logos. Both are cases where a nearly-correct guess is obviously wrong to a reader — a rebuilt letterform that is subtly the wrong shape looks worse than the blurry original. If your goal is to read small text in an image, the OCR tool is a better path than making the picture bigger.

How your image is handled

Upscaling is one of the OpusImg tools that does not run in your browser. The models are far too large to ship to a web page, so your image is uploaded over HTTPS, processed on our GPUs, and returned to you as a PNG. That is a real difference from tools like compress, resize, crop and blur, which never send your file anywhere, and it is why every AI tool on the site is labelled as server-processed rather than local.

Practically: the file is uploaded, held while the job runs, and the result is delivered to you. It is not used to train models, and it is not shared. If the distinction matters for a particular image — a confidential document, a client's unpublished work — the local tools remain available and genuinely never transmit anything, and that is a decision worth making per file rather than once.

Costs are shown before a run, not after. 2 credits for 2×, 4 for 4×, plus 1 if face enhancement is on. Failed jobs are not charged. Pro includes 500 credits a month and Business 2,000 per seat, which is a lot of upscaling — the credit system exists to keep GPU work sustainable, not to meter you closely.

Keep going

Frequently asked questions

What does AI upscaling do?

AI upscaling adds pixels while reconstructing edges and texture, so an image can be enlarged more cleanly than with a standard resize. A conventional resize can only average the pixels it already has, which is why enlarged photos look soft. A model trained on millions of image pairs instead predicts what the missing detail should look like — sharpening edges, rebuilding texture in hair and fabric, and cleaning up the mush that ordinary interpolation leaves behind.

Can AI upscaling fix a blurry photo?

It depends on why the photo is blurry. If it is small but sharp — a low-resolution image that was never bigger — upscaling works very well, because the detail is present and simply needs more pixels to live in. If the photo is out of focus or motion-blurred, upscaling cannot invent detail the camera never captured; it will produce a larger, cleaner-looking version of a blurry image. Soft-but-sharp is the sweet spot.

How much can you upscale an image without losing quality?

2× is nearly always safe and looks natural. 4× is the practical ceiling for most photographs: beyond that, the model is inventing more than it is reconstructing, and results start to look painted rather than photographic. If you need more than 4×, upscale in one 4× pass rather than running the tool twice — repeated passes compound the model's guesses and amplify its artifacts.

Should I choose 2× or 4×?

Choose 2× for everyday web images, social posts, and anything that just needs to look crisp on a high-density screen. Choose 4× when you need the largest practical output — print work, a detailed crop from a larger scene, or a product shot that has to hold up when zoomed. 4× takes longer and costs more credits, so it is not the automatic better choice; on an already-adequate image it mostly adds file size.

Is AI upscaling better than Photoshop?

For enlarging a photograph, yes — a trained model reconstructs plausible texture, while classic bicubic or Preserve Details resampling can only interpolate between existing pixels. Photoshop remains better for everything that is not enlargement: precise retouching, compositing, and color work under your direct control. The two are complementary, and many people upscale here and then finish in an editor.

What does face enhancement do?

Face enhancement runs a second restoration pass over detected faces, recovering facial detail in soft or low-resolution portraits. It is genuinely transformative on small group photos and old scans. It can also work against you: on faces that are already sharp it tends to over-smooth skin, and on extremely low-resolution faces it may reconstruct features that do not quite match the person. Leave it off unless the faces actually need help.

How many credits does upscale use?

2× upscale uses 2 credits, 4× upscale uses 4 credits, and optional face enhancement adds 1 credit — so the most an ordinary run can cost is 5. The exact cost is shown before you start, and nothing is charged if a job fails. Pro includes 500 credits per month and Business includes 2,000 per seat.

Is there an output-size limit?

Yes. To keep processing reliable, an upscaled result cannot exceed 64 megapixels — roughly 8,000 × 8,000 pixels. Because the limit applies to the output, it constrains the input by the square of the scale: at 4×, an image larger than about 4 megapixels will exceed it. If you hit the ceiling, either run 2× instead or crop to the region you actually need. Input files are capped at 25 MB.

Will upscaling fix a heavily compressed JPEG?

Partly. The model cleans up mild compression noise, but heavy JPEG artifacts — blocky 8×8 squares, colour banding, halos around edges — are a pattern in the pixels, and upscaling can enlarge them along with everything else. You will usually get a better result from the largest, least-compressed copy you can find: an original export beats a screenshot of a re-shared post every time.

What resolution do I need for printing?

Print is usually specified at 300 DPI, so multiply the printed size in inches by 300. A 4×6 photo needs about 1,200 × 1,800 pixels, and an 8×10 needs 2,400 × 3,000. Large-format work viewed from a distance — posters and banners — is fine at 150 DPI or less. Upscaling a 600 × 900 snapshot 2× lands you at print resolution for a 4×6, which is exactly the kind of job this tool is for.