You have a photo that's 600 pixels wide and you need 3,000 to print it or use it as a banner. You scale it up, and the result looks like a mosaic of blurry squares. That's normal: traditional software only knows how to stretch the pixels that are already there. AI upscalers do something else: they recreate the missing detail based on what they've learned from millions of images. Skin texture, the weave of a fabric, the grain of a brick — the model guesses what should be there and draws it in.
Below you'll find three methods tested on the same images: Skilzy's AI Lab, free upscalers online or offline, and the feature already built into your phone and computer. For each one: the real scaling factor, the processing time, the price, and the cases where things go off the rails. None of them require technical skills, and two of them cost nothing.
Why an Enlarged Image Goes Blurry (and What AI Changes)
A digital image contains a fixed number of pixels: when you enlarge it, the software has to invent the missing ones, and traditional software invents them badly.
An 800 x 600 pixel photo contains 480,000 pixels. Displayed at 1600 x 1200, it needs 1,920,000 — four times as many. Those extra 1,440,000 pixels don't exist anywhere: they're manufactured.
A traditional tool (Preview on Mac, Paint, your phone's resize function) uses interpolation: it averages the neighboring pixels. The result is smeared edges, flat patches of color, and that staircase effect on diagonals.
An AI upscaler works differently. The model was trained on millions of pairs of “low-resolution image / high-resolution image.” It learned what an eyelash, a strand of wool, or a tile grout line looks like up close. Faced with your photo, it doesn't compute an average: it proposes a plausible reconstruction.
Hold on to that word: plausible. It doesn't mean accurate. On a face, the model might draw in skin texture that was never there. On blurry text, it might turn an “R” into a “P.” You never recover lost information — you get a sharp image that looks a lot like the original. For a banner, a YouTube thumbnail, or an ad, that's more than enough. For a document meant to serve as evidence, it's worthless.
One useful benchmark before you pick a tool: printing requires roughly 300 pixels per inch. A 4 x 6 inch print therefore needs about 1180 x 1770 pixels, and a letter-size page around 2480 x 3510 pixels. Compare that number to your current photo: often a simple 2x is enough, and the smaller the factor, the cleaner the result.
Method 1: Enlarge Your Images in Skilzy's AI Lab
Skilzy's AI Lab brings upscaling and image generation models together in a single interface, with credits included in the subscription.
The problem when you use upscalers one by one: every good model lives behind its own subscription. You end up with one account to generate, a second to upscale, a third to cut out backgrounds, and three monthly charges for a handful of images a week. The AI Lab built into Skilzy avoids that pile-up: you use the real tools straight from the platform, credits are included, and there's no other subscription to sign up for.
In practice, the process takes five steps:
- You drop your image into the AI Lab.
- You pick the scaling factor (2x or 4x in the vast majority of cases).
- You specify the image type: portrait, product photo, illustration. A face isn't processed the same way as a packshot.
- You run the render and compare before / after at 100% zoom, not as a thumbnail.
- You export in the format you need, JPEG or PNG depending on the use.
Where Skilzy really stands out is in everything around that click. The platform is a school: more than 15 programs cover image creation, video (UGC, faceless channels, avatars, voiceovers), music, automation with n8n, prompt writing, and AI-assisted coding. You don't just learn how to enlarge a photo — you learn when to do it within a production pipeline. If your goal is to produce sharp ad visuals that work in both print and display, the AI advertising creation program covers exactly that sequence: generation, retouching, upscaling, and adapting to the right formats.
Two certifications are state-recognized and eligible for funding: RS7439 (AI-assisted content marketing) and RS6792 (AI and sales). A B2B offer also exists for training organizations.
Its limits, honestly: the AI Lab runs on credits, so it isn't the tool for processing a catalog of 800 photos overnight. For raw, repetitive volume, go to method 2.
Method 2: Free AI Upscalers, Online or on Your Computer
Upscayl is the most solid free option: it's open source, it runs directly on your computer, and it never sends your images over the internet.
Upscayl is available on Windows, macOS, and Linux. It ships with several models from the Real-ESRGAN family, offers 2x and 4x (more if you chain passes), and processes entire folders in batches. It's free, with no account, no watermark, and no limit on the number of images. Its drawback: it relies on your graphics card. On a recent machine, expect a few seconds per image; on a 2018 office laptop, you can hit two to three minutes per file.
Web services, on the other hand, ask for nothing but a browser:
- upscale.media: 2x and 4x, very fast, account required to export at full resolution. The free tier caps output size.
- Bigjpg: specialized in illustrations and drawings (anime, manga, flat color areas). Free with a queue, 4x maximum on the free tier.
- Krea: its enhancement module is the most creative of the bunch. It doesn't just enlarge, it reinvents detail. Excellent on an illustration, risky on a family photo.
- Magnific: the benchmark for extreme upscaling with added detail, but the tool is paid and clearly aimed at professional use.
Two precautions with online services. First, your images go to a server: avoid ID photos, client documents, or visuals under a confidentiality agreement, and read the terms of service before uploading anything. Second, free offers shift every quarter: quotas, watermarks, and maximum resolutions change without notice, so always check the current pricing page.
Worth noting: if your source image came from a generator, you're usually better off regenerating it directly at high resolution rather than enlarging it afterward. Our roundup of the best tools for generating images with AI for free breaks down each one's native resolutions.
Method 3: The Feature Already Built Into Your Phone and Computer
Your phone and computer already know how to improve a blurry photo, but in most cases they clean it up without actually enlarging it.
It's a distinction that saves a lot of wasted time. Here's what each native option actually does:
- iPhone (Photos app): Edit, then Enhance, adjusts contrast, exposure, and perceived sharpness. No extra pixels are created. Useful on a dull photo, useless on a photo that's simply too small.
- Google Photos: the Sharpen feature corrects motion blur and noise. It's offered on Pixel devices and to Google One subscribers depending on the device. Again, that's denoising, not upscaling.
- Samsung Gallery: the Remaster option slightly increases definition and restores some detail. The gain is real but modest, on the order of a cautious 2x.
- Preview on macOS, Photos on Windows: pure resizing, no AI. You'll get exactly the blur described at the start of this article.
- Photoshop and Camera Raw: Adobe's Super Resolution doubles the linear dimensions — four times as many pixels — with a very faithful result. In Photoshop, Image then Image Size also offers Preserve Details 2.0 resampling, effective up to 2x or 3x. It's the safest method when you want zero invention, but it requires a Creative Cloud subscription.
When to favor native tools: it's instant, free, offline, and your images never leave the device. For an Instagram post or a photo to send in a message, it's often enough. If what you actually need is retouching rather than resolution, the guide to changing a photo's background from your phone covers the other half of the subject.
The Comparison: The Same Images Run Through Four Tools
On the same set of four images upscaled 4x, no tool wins across the board: the ranking shifts depending on whether it's a face, a product, an illustration, or text.
Test protocol: a 500 x 500 pixel portrait from an old profile, a 900 x 600 product packshot, an 800 x 800 illustration, and a 600 x 400 screenshot containing text. 4x target for everything, examined at 100% zoom.
| Tool | Time per image | Price | Strength | Weakness |
|---|---|---|---|---|
| Skilzy AI Lab | 15 to 30 s | Monthly subscription | Portraits and product photos, pairs with generation | Runs on credits, poorly suited to very high volume |
| Upscayl (local) | 10 s to 3 min depending on machine | Free | Batch processing, no images sent online | Depends on your graphics card, sometimes plastic-looking skin |
| upscale.media | 5 to 15 s | Free with limits | Speed, zero installation | Capped output resolution, upload required |
| Photoshop Super Resolution | 5 to 20 s | Adobe subscription | Maximum fidelity, no invention | Factor limited to 2x linear |
| Phone's native feature | Instant | Free | Dull or slightly blurry photos | Adds almost no pixels |
The takeaways from the test, beyond the table. On faces, the AI models are all comparable as long as you stay at 2x; from 4x on, they smooth out pores and harden eyelashes, which produces that slightly waxy look. On the packshot, they all do well: manufactured objects have crisp edges that help the model. On the illustration, the most creative tools come out ahead. And on text, every single one failed at least partially, including the best: small characters were redrawn with errors.
A practical consequence for logos: don't enlarge a pixelated logo, remake it. A vector file stays sharp at any size, and our article on creating a logo with AI explains how to go from a generated concept to a clean file.
The Five Mistakes That Ruin an Upscale
The final quality depends as much on what you do before and after the upscale as on the model you choose.
- Starting from a JPEG that's been recompressed ten times. A screenshot of a screenshot sent over WhatsApp arrives already covered in artifacts. The AI mistakes them for detail and enlarges them faithfully. Always hunt down the original file before processing the one you received.
- Going for 8x in one shot. Two 2x passes almost always give a cleaner result than a brutal 4x, and a direct 8x stays very risky except on simple flat areas.
- Cropping afterward. Crop first, then enlarge: you don't waste computation on areas you're going to throw away, and you gain sharpness on the part you keep.
- Approving it at thumbnail size. An image that looks sharp in a 200-pixel preview can be a disaster at full size. Check at 100%, and prioritize the sensitive areas: eyes, teeth, hands, text, logos.
- Forgetting file size. A 4,000-pixel-wide visual easily weighs 6 MB. For the web, re-export to WebP at around 200 KB, otherwise you're trading a pixelated image for a page that takes five seconds to load.
Key Takeaways
Start by calculating the resolution you actually need: a lot of problems are solved with a simple 2x. If you have a lot of files and a decent computer, install Upscayl — it's free and unlimited. And if your photo is only dull or slightly blurry, try your phone's built-in feature first: it's instant and often enough.
One last habit to build: for anything containing text or brand identity, regenerate rather than enlarge.
