Image Tools·4 min read

How AI Upscaling Works, and Where It Honestly Fails

Everyone has one: a photo that only exists at 600 pixels wide, an old scan of a family print, a logo the designer exported tiny and never sent again. Making those bigger without turning them to mush is what super-resolution models are for, and once you know what the model is doing you can tell when it will save you and when it will disappoint. This article covers what the ESRGAN model inside the AI Image Upscaler actually predicts, the honest limits, and how to get the most from 2x and 4x.

Open the free AI Image Upscaler All insights

The situations that call for AI upscaling

Small originals, big requirements.

The classic case is the irreplaceable small image. A print got scanned at low resolution years ago and the print itself is gone. A client's only copy of their logo is a 512 pixel export. The one photo of a discontinued product is the web thumbnail. In each case you need more pixels than exist, and no amount of asking will produce a bigger original.

There's also the routine case: enlarging decent images for print, banners, or high-density screens, where a 2x pass simply keeps things crisp at the new size instead of soft.

How AI upscaling works: prediction, not enlargement

The model draws what was probably there.

Classic resizing, bicubic or bilinear, computes each new pixel as an average of its neighbors. Averages are always softer than their inputs, which is why stretched images look smeared. An ESRGAN network takes a different route: it was trained on enormous numbers of paired small and large images, so it learned what edges, textures, and gradients look like at higher resolution. When it upscales, it predicts the detail that plausibly belongs there.

In this tool the model loads through UpscalerJS with TensorFlow.js, fetched once from a CDN and cached, then runs on your own GPU through WebGL. There are two ESRGAN variants: Standard, which is smaller and faster, and High detail, which can recover more texture but needs more graphics memory. If the AI can't run at all, for example with WebGL disabled, the tool falls back to a non-AI enlarge-and-sharpen pass and tells you so on the page.

The word predicts matters. The output detail is plausible, not recovered. That's the source of both the impressive results and every limitation that follows.

A worked example with real numbers

Inputs, caps, and outputs.

Input: avatar.png at 640x480, run at 2x with PNG output. Output: avatar-2x.png at 1280x960, with edges the model re-drew rather than stretched. The same file at 4x becomes 2560x1920. Pick JPG or WebP instead and the extension follows, with a quality slider for either.

The caps work like this: output is limited to 4096 pixels on the longest side, so 2x accepts up to 2048 pixels and 4x up to 1024. Feed a 1200 pixel wide photo to 4x and it first scales down to 1024, then upscales to 4096. For 8x, the 4x and 2x models run one after the other, so inputs are reduced to 512 pixels. Files are capped at 20 MB each, Exact output width can resize the result up to 8192 pixels, and batches come back as upscaled-images.zip.

The upscaling limits nobody prints on the marketing page

What super-resolution cannot give back.

A model can't restore information the camera never captured. A license plate that's an 8 pixel smudge stays unreadable; the network just draws a sharper smudge, or worse, invents characters that were never there. The same goes for distant faces and tiny text. This is why upscaled images are worthless as evidence: the added detail is a statistical guess wearing a convincing costume.

Damage gets amplified too. JPEG block artifacts, noise, and motion blur are all patterns the model happily enhances along with real detail, and 4x on a heavily compressed source often produces waxy, plastic-looking textures. Knowing these upscaling limits up front saves you from expecting a rescue that isn't coming.

When upscaling beats rescanning, and when it doesn't

Real pixels always beat predicted ones.

If the physical original still exists, rescan it. A flatbed at 600 dpi captures genuine information, and genuine beats predicted every single time. The model is the fallback for when the print is lost, the negative is gone, or the source was born digital, like a screenshot or a web export, and no better version ever existed.

The two also combine well: rescan at the highest optical resolution your scanner offers, then apply a 2x pass to clean up and enlarge further. Used that way, the model adds to a good scan rather than replacing it.

Mistakes that make upscaled photos worse

The usual self-inflicted wounds.

These four patterns account for most disappointing results.

Feeding a heavily compressed JPEG to 4x. The model magnifies the compression artifacts along with the subject. Use the cleanest copy you can find.
Compressing the output hard right after. You paid processing time for predicted detail, then threw it away. Compress gently, and last.
Defaulting to 4x when 2x looks more natural. Faces in particular can turn artificial at 4x; the smaller factor often wins on realism.
Expecting readable text or plates to emerge. Prediction isn't recovery, and anything that looks recovered was actually invented.

Practical tips for sharper results

Squeeze out the last bit of quality.

Crop before you upscale. Trimming dead space with Crop Image means the pixel budget and the size caps go entirely to your subject. If you're preparing an image for a cutout, upscale before background removal so the segmentation model gets the extra edge detail to work with.

Use the before and after slider on faces and text, where invented detail shows first. Try High detail on a machine with a decent graphics card, and drop back to Standard if it is slow or fails. Save as PNG if you will edit further; JPG or WebP suits a final photo.

One more trick: for a target size between factors, run 4x and then scale down to the exact dimensions. Downscaling a larger prediction acts like supersampling and often lands crisper than a direct 2x.

The upscaler next to the related image tools

Pick the tool that matches the direction.

The Image Resizer is the right choice whenever you're going smaller or just need exact dimensions, since shrinking never requires invented detail. The Image Compressor reduces file size at unchanged dimensions and belongs at the end of any pipeline. And the Background Remover slots in after upscaling, as covered above, so your enlarge-then-cut order stays consistent whenever you touch old product photos.

Try it now

Open AI Image Upscaler

The tool is one click away. No sign up, no upload, no payment.

Open AI Image Upscaler

Write a review