AI Image Upscaler — Enlarge Photos Without Losing Quality

Every image on your screen is a grid of coloured squares, and when you ask a program to make that grid bigger it has to fill in squares that never existed. For thirty years the answer was interpolation: look at the neighbouring pixels and pick a colour somewhere in between. Nearest-neighbour simply repeats each pixel, which produces the blocky staircase look of a badly scaled icon. Bilinear averages the four closest pixels, which removes the blocks but smears everything. Bicubic — the default in most editors — looks at sixteen neighbours and fits a smooth curve through them, and Lanczos uses an even wider window with a sharper curve. Each is a little better than the last, and none of them can add a single detail that was not already in the original.

That is the fundamental problem. A photo saved at 800 by 600 pixels contains 480 000 samples of the scene. Enlarge it four times and you need 7 680 000. Interpolation spreads the same 480 000 samples over a canvas sixteen times larger and guesses the rest by averaging, so edges soften, textures dissolve into gradients, and anything fine — hair, fabric weave, lettering, the grain of wood — becomes a blur. The result is not wrong, exactly; it is just an accurate picture of how little information was there. Sharpening afterwards only exaggerates the edges that survived and adds halos around them.

An AI image upscaler approaches the problem from the opposite direction. Instead of asking what colour sits between two pixels, it asks what the scene probably looked like at higher resolution — and answers using what it has learned from millions of paired low-resolution and high-resolution photographs. This guide explains how that works, when it is worth using, how to choose between enlarging by two and by four, what happens to prints, faces and old scans, and, just as importantly, what a super-resolution model cannot do.

Super-resolution is the technical name for the task of producing a high-resolution image from a low-resolution one. The models that do it well today are convolutional neural networks trained on a very simple game: take a sharp photograph, degrade it — shrink it, blur it, add noise and compression artifacts — then ask the network to reconstruct the original from the degraded version. The network's output is compared with the true sharp image and its millions of internal weights are nudged to reduce the difference. Repeat this on a vast and varied set of photographs and the network learns something remarkable: what edges, textures, skin, foliage, brick, fur and text tend to look like at fine scale, and how each of them appears after being shrunk and compressed.

Real-ESRGAN, the family of models behind PixGenia's upscaler, took this approach further by training on degradations that mimic what real photos actually suffer — not just clean downscaling but the whole chain of blur, sensor noise, JPEG compression and resizing that a picture goes through on its way from a camera to a chat app. That is why it copes well with images from the wild: a low-resolution product shot pulled from an old website, a screenshot of a photo, a picture that has been through three rounds of messaging compression. The model has seen those exact kinds of damage during training and knows what a clean version usually looks like.

The key word is plausible. When the model turns a smudge of pixels into the texture of a knitted jumper, it is not recovering the real jumper — it is drawing a knit that is statistically consistent with the low-resolution evidence. For textures, edges, foliage and most surfaces this is exactly what you want, and the difference from bicubic enlargement is dramatic: crisp edges instead of soft ones, texture instead of gradient, and no halos. Where it matters is on content whose precise details carry meaning: a face you know, a serial number, small print. We come back to those cases below, because an honest guide to upscaling has to draw that line clearly.

The scale factor multiplies both width and height, so the pixel count grows by the square. Enlarging by two turns 800 × 600 into 1600 × 1200 — four times as many pixels. Enlarging by four turns the same image into 3200 × 2400 — sixteen times as many, or 7.7 megapixels from a 0.5 megapixel original. A 1200 × 800 photo becomes 2400 × 1600 at ×2 and 4800 × 3200 at ×4, which is 15 megapixels and already larger than what many phones capture. A 2000 × 1500 image at ×2 becomes 4000 × 3000, a 12-megapixel file that will be several tens of megabytes as a PNG.

Bigger is not automatically better. The model has to invent a larger share of the output at ×4 than at ×2, so the risk of over-smooth or slightly artificial texture grows with the factor. If the original is reasonably clean and you need a moderate increase — a 1000-pixel image that has to become 2000 for a full-width web banner, a phone photo you want to crop tighter — ×2 gives a result that is almost indistinguishable from a native capture. Reserve ×4 for genuinely small sources: thumbnails, old web images, video frames, low-resolution logos and product pictures, anything under roughly a thousand pixels on its long side that has to become print-sized or fill a modern screen.

There is a practical ceiling too. PixGenia's tool accepts images up to 1 megapixel for ×4 (for example 1200 × 800) and up to 4 megapixels for ×2 (for example 2000 × 2000), which keeps the output within 16 megapixels and the file at a size a browser can still encode and download. If your source is already 12 megapixels, you probably do not need to upscale it at all: it prints at A3 as it is, and what looks like low resolution is usually softness, noise or compression that a photo enhancer handles better than an enlarger.

• 640 × 480 (old web image) → ×4 → 2560 × 1920, enough for a sharp full-screen display or a 20 × 15 cm print
• 800 × 600 → ×4 → 3200 × 2400, roughly 7.7 megapixels, prints at 27 × 20 cm at 300 dpi
• 1024 × 768 → ×2 → 2048 × 1536, a clean upgrade for a slide or a web hero image
• 1200 × 800 → ×4 → 4800 × 3200, an A3 print at 300 dpi with room for cropping
• 1920 × 1080 (video frame or screenshot) → ×2 → 3840 × 2160, native 4K
• 2400 × 1600 → ×2 → 4800 × 3200, a 15-megapixel file for large-format printing

Print is the most common reason to enlarge a photo, and the arithmetic is simple once you know the target. Printers and labs generally ask for 300 pixels per inch (often written dpi) for photographic quality viewed at arm's length; 200 is still very good, and 150 is acceptable for posters that will be seen from a metre or two away. To find the pixel count you need, multiply the print's dimensions in inches by the resolution. A 10 × 15 cm print (4 × 6 inches) at 300 dpi needs 1200 × 1800 pixels. An A4 sheet is 21 × 29.7 cm, or 8.27 × 11.69 inches, so it needs about 2480 × 3508 pixels at 300 dpi. A3 is twice that area: 29.7 × 42 cm, about 3508 × 4961 pixels.

Put those figures next to a typical low-resolution source and the case for AI upscaling is obvious. An 800 × 600 photo printed on A4 would be running at about 70 dpi, visibly pixelated. Enlarged by four to 3200 × 2400 it comfortably exceeds the 300 dpi requirement for A4 and sits at roughly 200 dpi on A3 — good enough for a wall print. A 1500 × 1000 image from a website, enlarged by two to 3000 × 2000, prints at A4 at close to 300 dpi and at A3 at around 180 dpi. The important thing is that the AI result actually holds up at that size: interpolated to the same dimensions, the same image would be soft over every edge, and a print reveals softness far more cruelly than a screen does.

Two practical notes for print work. First, download the PNG rather than a re-saved JPEG: the print shop will do its own conversion, and you want to hand over every pixel intact. Second, upscale before you crop and retouch, not after, so that any further editing happens at the final resolution. If the picture also needs colour correction or sharpening, run the upscaler first, then open the result in the photo enhancer, where the manual sliders will work on the enlarged file.

Screens have quietly become as demanding as paper. A modern phone displays around 400 pixels per inch, laptops routinely exceed 200, and every platform recompresses what you upload. An image that is smaller than the slot it lands in is stretched by the browser with — you guessed it — bilinear or bicubic interpolation, and then squeezed through JPEG again. Starting from a source that is at least as large as the display size, and ideally twice as large for high-density screens, is the single biggest thing you can do for how a picture looks online.

The recommended sizes speak for themselves. Instagram feed posts are stored at 1080 pixels wide; stories and reels at 1080 × 1920. LinkedIn and Facebook cover images run around 1600 to 1200 pixels across. YouTube thumbnails are 1280 × 720 and channel art is 2560 × 1440. A website hero image on a large screen needs 1920 pixels or more, and 2560 to 3840 if you want it to look crisp on a 4K monitor. Marketplaces such as Amazon and eBay ask for a longest side of at least 1000 to 1600 pixels so that shoppers can zoom, and several of them reject or downrank listings whose images are smaller. Any product photo, logo or brand asset that predates these requirements is a candidate for a clean ×2 or ×4 enlargement.

E-commerce is where the plausible-detail principle deserves a moment of thought. Upscaling a product photo sharpens fabric, stitching, wood grain and packaging convincingly, and the zoom view will look far better than a stretched original. But if the exact appearance of a fine detail is part of what you are selling — a printed pattern, an engraved marking, a label with small text — check the enlarged version closely against the original before publishing, because the model's job is to render a plausible surface, not to certify an exact one.

A scan of a family print, a frame saved from a camcorder tape, a picture emailed in 2006 at the size that made sense then: these are the images people most want to enlarge, and they are also the hardest, because they arrive with damage layered on top of low resolution. Scanned prints carry the grain of the paper and often a slight softness from the scanner optics. Old digital photos carry heavy JPEG compression, which shows up as eight-by-eight-pixel blocks in smooth areas and ringing around edges. Low-light photos from early phones carry coloured noise. Interpolation enlarges all of that faithfully, so the blocks become bigger blocks and the noise becomes bigger noise.

This is precisely the territory Real-ESRGAN was designed for. Because the model was trained on images degraded with realistic blur, noise and compression, it has learned to treat those artifacts as damage rather than as content: it smooths the block boundaries, suppresses the coloured speckle, and rebuilds the edges underneath. On a compressed 600-pixel snapshot the effect is often more striking than the enlargement itself — the picture looks like a cleaner version of what it always was. For a scan, the model will usually soften paper grain and remove scanner noise, giving a smoother image that prints well; if you like the grain, you can add a little back in a photo editor afterwards.

Manage your expectations in proportion to the damage. Mild compression and moderate noise clean up beautifully. A picture that has been re-saved a dozen times until the blocks are the size of a fingernail, or a scan of a print that was itself soft, will improve but will not become a modern photo; the model has less to work with and will produce smoother, less detailed surfaces. For heavily damaged images, a workable approach is to upscale by ×2 rather than ×4 — a smaller factor gives the model less to invent — and then run the result through a photo enhancer to restore contrast and colour that the years took away.

Faces are special because we are extraordinarily sensitive to them. A texture that is almost right on a brick wall is invisible; a texture that is almost right on a cheekbone or an eyelid is immediately unsettling. General-purpose super-resolution does a reasonable job on faces that are already a fair size in the frame — a portrait, a group of three — but on small faces, the kind that occupy fifty pixels in a crowd or a class photo, it can produce eyes and mouths that are sharp yet slightly wrong.

The optional face enhancement runs a second model that specialises in facial restoration. It detects each face, reconstructs eyes, eyebrows, nose, mouth and skin texture using what it has learned about how faces are built, and blends the result back into the upscaled image. On old, blurry or low-resolution portraits the improvement can be remarkable: features that were a suggestion become legible, and the person looks like a person rather than a smudge. This option always runs on the server, because the face model is far too large for a browser, and the tool says so when you enable it.

It also comes with honest limits that you should know before relying on it. The face model reconstructs a plausible face that is consistent with the low-resolution evidence; on a very small or very blurry face, the evidence underdetermines the result, and the reconstruction can drift toward a generic face — the same person in general outline, but with subtly different eyes or a slightly changed expression. Fine lines, freckles, scars and asymmetries that make someone recognisable can be smoothed away. It is a superb tool for restoring a portrait to share with family; it is not a forensic instrument, and no upscaler should be treated as one. For photos where identity must be exact, upscale without face enhancement, compare the two results, and trust your eyes.

Super-resolution is not magic, and the cases where it fails are predictable once you understand that it enlarges what it can see. Motion blur is the first. A photo blurred because the subject or the camera moved has no sharp edges anywhere; the information is not merely small, it is smeared across pixels in a direction the model cannot undo. The upscaler will give you a larger, cleaner version of the same blur, with sharper edges at the boundaries of the smear. That can look better, but it is not deblurring. The same is true of focus blur: an out-of-focus background stays out of focus, which is usually what you want, and an out-of-focus subject stays soft.

Tiny text is the second. Letters that are only a few pixels tall are ambiguous at the pixel level: a lowercase e, c and o can be the same three-pixel blob. The model will render crisp, plausible letterforms, and they may not be the letters that were actually there. Upscale a screenshot of a document and the headings will look excellent, but never read a phone number, a licence plate, a serial number or a price off an upscaled image without checking the original. For the same reason, a low-resolution logo will come out clean and sharp, but check the fine typography and any small symbols against the real thing.

The third limit is the one that runs through this whole guide: invented detail. Every texture the model adds is a well-informed guess. In the overwhelming majority of photos — landscapes, products, buildings, pets, portraits at a reasonable size — those guesses are indistinguishable from reality and make the picture far better. In the small minority where a precise detail matters, the model has no way of knowing which detail that is. Finally, an upscaler cannot fix framing, exposure or colour, and it cannot add pixels outside the frame; for those, use a photo editor or a generative fill tool.

• Motion blur and focus blur are enlarged, not removed
• Text under about ten pixels tall may be rendered as plausible but incorrect letters
• Faces smaller than a few dozen pixels can drift toward a generic face, especially with face enhancement
• Fine identifying marks — serial numbers, engraved codes, licence plates — must be verified against the original
• Exposure, white balance, colour casts and framing are untouched: use an editor before or after
• Images larger than the input limits (1 megapixel for ×4, 4 for ×2) must be reduced or cropped first

PixGenia's upscaler takes a browser-first approach. For images up to roughly a megapixel — most web images, screenshots, old photos and thumbnails — the super-resolution model itself is downloaded to your device (a few megabytes, cached after the first use) and runs there, on your graphics hardware through WebGPU where the browser supports it, or on the processor otherwise. Your image never leaves your computer or phone in that case; you can switch the network off after the model has loaded and the enlargement still completes. The picture is cut into overlapping tiles, each tile goes through the network, and the tiles are blended back together so that no seams are visible.

Larger images, the face-enhancement option, and devices that cannot run the model locally use a server instead. In that case the tool tells you clearly — a line reading "large image: processed on our servers" appears before you start — and the image is uploaded over an encrypted connection, enlarged by a bigger version of the same model family, returned to your browser, and deleted. Nothing is stored: not the original, not the result, not a thumbnail. The upscaled file exists only in your browser until you download it, and reloading the page discards it.

Server processing is metered with a daily allowance per visitor so that the tool can stay free for everyone, and the allowance is only ever checked when you actually need the server; you are never shown a limit for work your browser does itself. If the server declines an image because the daily allowance is spent and the image is small enough, the tool simply falls back to processing it locally. No account is required for either path, and the download carries no watermark and no reduced-size version.

The tool accepts PNG, JPEG and WebP files up to 50 megabytes. Orientation metadata from phone cameras is respected, so a portrait shot comes out upright. Transparency in PNG and WebP sources is preserved through the enlargement: the alpha channel is upscaled alongside the colour, which makes the tool useful for logos and cut-out product images destined for a larger layout. Animated GIFs and video are not supported; save a single frame as PNG first.

Output is always PNG at the full enlarged size. PNG is lossless, so every pixel the model produced reaches your disk unchanged, and it is the right format to hand to a printer, a designer or another editor. The trade-off is size: a 3200 × 2400 photographic PNG is typically between 10 and 20 megabytes, and an 8000 × 6000 one can be well over 50. The tool displays the output dimensions and file size before you download, and the file is named after your original with the factor appended — photo-x4.png — so that versions do not get mixed up. Camera metadata is written back into the PNG when the source had it.

If you need a smaller file for the web or for email, open the downloaded PNG in the photo editor on the same site and export it as JPEG at quality 85 to 92 or as WebP; both will be a fraction of the size with no visible loss at normal viewing distance. Do not convert to JPEG before upscaling, though: the model handles existing compression well, but there is no reason to add more.

The whole process takes under a minute for a typical web image. The list below walks through it in order; the compare view at the end is worth a few seconds even when the result looks right at a glance.

• Open the upscaler and drop your image on the page, click to browse for it, or paste it from the clipboard with Ctrl+V or Cmd+V. PNG, JPEG and WebP up to 50 MB are accepted.
• Read the size line: it shows the original dimensions and, live, the dimensions you will get. Choose ×2 for a moderate, very safe enlargement or ×4 for small sources that need to become large.
• Decide on face enhancement. Leave it off for landscapes, products and documents; turn it on for old or blurry portraits, knowing that it always uses the server.
• Press Enlarge. A small image loads the model into your browser on first use — a few seconds — then shows tile-by-tile progress with a time estimate. A large image shows the server notice and a single progress bar. Cancel at any time.
• Use the compare view. Hold the button to flash back to the original, drag the split line across the picture, or switch to side by side. The view opens at 100 % so you are judging real pixels; zoom to 200 % or 400 % on eyes, text and fine texture.
• If the result looks over-smooth or a detail seems wrong, try the other factor, or toggle face enhancement off and compare again. Each run is independent.
• Download the PNG. It is the full enlarged size, named with the factor, with no watermark. For a JPEG or WebP version, open it in the photo editor and export from there.

Desktop editors have all added AI enlargement in recent years. Photoshop's Super Resolution and Lightroom's Enhance feature perform a ×2 enlargement using Adobe's own models, work on raw files as well as JPEGs, and integrate neatly with the rest of the editing workflow; they require the applications to be installed and, for most users, a subscription to the suite. Topaz Gigapixel is a dedicated desktop upscaler that offers factors up to ×6 and beyond, several models tuned for different kinds of content, batch processing and very fine control, and it is the tool professionals reach for when hundreds of images need enlarging with adjustable settings. It is a paid application that runs on a capable computer.

Free online upscalers occupy the other end of the spectrum and vary enormously. Many of them run the same open-source Real-ESRGAN family that PixGenia uses, so image quality is often similar; the differences lie in what happens around the model. Common restrictions include a low resolution cap on free output, a watermark, a mandatory account, an upload of every image to the provider's servers regardless of size, and a queue at busy times. Some are excellent, and the sensible thing is to try the same image on two of them and compare at 100 %.

PixGenia's position is straightforward: professional-grade super-resolution with the two factors that cover almost every real need, free, without an account, with the full-size PNG and no watermark, and with small images processed on your own device so that they are never uploaded. What it does not offer is the depth of a desktop tool — no batch processing, no raw support, no choice of models or per-image sharpness settings, no factors above ×4. If you enlarge a handful of images a week and care about privacy and simplicity, it is the right tool; if you enlarge hundreds and need fine control, a desktop application earns its place.

"Can I really enlarge an image without losing quality?" You cannot recover detail that was never captured, but you can enlarge without the softness and blockiness that interpolation causes, and the model adds plausible fine detail that makes the result look like a native higher-resolution photo. For most images the practical answer is yes. "Which is better, ×2 or ×4?" Use ×2 when the source is decent and you need a moderate increase; use ×4 for small sources that must become large. When in doubt, try both and compare. "Does it work on screenshots and logos?" Yes, and transparency is kept; check small text against the original.

"Is my photo uploaded?" Small images are processed entirely in your browser and never leave your device. Large images and the face-enhancement option use a server, the tool tells you beforehand, and the image is deleted after processing. "Is there a watermark or a size limit on the download?" No watermark, and the download is always the full enlarged size. "Why does the first run take longer?" The model is downloaded to your browser once and cached afterwards. "Can I enlarge a very blurry photo?" It will be larger and cleaner, but blur is not removed; the tool works best on images that are small rather than blurry. "What is the maximum size?" 1 megapixel of input for ×4 and 4 megapixels for ×2, with a 50 MB file limit.

Enlarging an image well comes down to three decisions: whether the picture is small or merely damaged, which factor gives you the pixels you need without asking the model to invent more than it has to, and whether the details that matter — a face, a line of text, a marking — are ones you must verify against the original. Get those right and AI super-resolution turns an old web image into a print, a thumbnail into a banner, and a compressed snapshot into a picture that looks like it was taken yesterday.

PixGenia's image upscaler puts that within reach of anyone: Real-ESRGAN super-resolution at ×2 or ×4, an optional face-restoration pass for portraits, a professional compare view at real pixel size, and a full-resolution PNG download — free, without an account, without a watermark, and with small images processed in your own browser so they never leave your device. Drop in the picture that has always been too small and see what it was hiding.