Ask “which AI upscaler is best” in any photo or AI-art thread and you’ll get five confident, contradictory answers. They’re all sort of right, because the honest answer is that it depends on the image, and nobody’s holding still long enough to prove it. The showcase galleries are no help. Every upscaler’s marketing page shows the one photo where it happens to win.
The only way to know for sure is boring. Take your own image, run it through two or three upscalers, and put the results under a slider you can drag. The differences that disappear in a gallery jump straight out the moment you can flip between them in the same frame.
The three you’ll usually be choosing between
There are dozens of models, but in 2026 most real comparisons come down to a few.
Real-ESRGAN is the default free, local one, and for most AI-generated images it’s the one to beat. It adds real detail, handles anime and photoreal about equally well, and it’s fast. Its signature is sharpness: crisp edges, occasionally a little too crisp, with faint ringing along high-contrast lines. The 4x-UltraSharp variant leans into that for AI art, and the anime 6B model is the pick for line work.
SwinIR is the transformer-based option. It scores best on the raw quality metrics and holds photographic detail well, but it’s slower, and plenty of people still think Real-ESRGAN looks better to the eye because it invents nicer texture. Its output just has a different character, which is exactly the sort of thing you only catch in a direct comparison.
Topaz Gigapixel is the commercial route: a polished desktop app, no models to pick, with face and texture models tuned for photos. It reads smoother where Real-ESRGAN reads sharper. Worth knowing before you get attached, Topaz is subscription-only now, around $29 a month for the Personal plan (cheaper if you pay for the year), with no one-time Gigapixel purchase for new buyers anymore.
What to look for when you slide between them
Upscaling isn’t magic, and it isn’t lossless. The model is guessing at detail that was never in the file. So when you drag between two upscales, you’re really checking three things.
The first thing you notice is the style, sharp against smooth, and it’s mostly taste right up until it isn’t. A sharp model makes textures and text pop. Skin and gradients usually come out better through a smoother one. Slide the same face between Real-ESRGAN and Topaz and you’ll know which you prefer pretty fast.
Next come the artifacts. Drag the handle onto a high-contrast edge, a roofline against sky or dark hair against a pale wall, and watch for ringing: a faint bright or dark halo hugging the edge. GAN-based upscalers like Real-ESRGAN are the usual culprits, and it’s easy to miss on a still until you can flip the same edge back and forth.
The one that matters most is hallucination. An upscaler will happily invent pores, hair strands, and fabric weave the source never had. Sometimes that’s the detail you were after. Sometimes it’s a face that isn’t quite the same person anymore. Line the upscale up against the original and stare at the smooth areas. If new “detail” is turning up in mid-tones where there was nothing to recover, the model is making it up.
Here’s the same shot upscaled two ways, one softer and one sharper. Drag between them and start asking which “detail” is real and which is just the sharpen dial cranked up: ::imgi{id=shuVRa7}
A rough starting point, then verify on your own image
If you mostly finish AI art, start with Real-ESRGAN 4x-UltraSharp. It’s free and its sharpness suits illustration. Anime and line art, the 6B model. For photographs where faces and skin carry the shot, Topaz Gigapixel’s photo models are hard to beat, and that softer look is usually what you want on a portrait. For wringing the most measured detail out of a photo when you don’t mind the wait, SwinIR.
Treat all of that as a starting point. The whole reason to compare is that the ranking flips depending on the picture. An upscaler that’s magic on anime can smear a landscape, and the one that nails skin can wreck fine text.
Running the comparison without fooling yourself
Feed every upscaler the exact same source file. Match the final dimensions too, since a 4x and a 2x upscale at different output sizes aren’t a fair fight. Then drop the results into imgi and drag. Two models is a straight before/after; for three or more, the multi-compare mode holds several outputs behind one link and lets you swap which pair you’re viewing, which is the sane way to rank a whole batch.
Whichever upscaler wins on the picture in front of you is the right one, and that’s a two-minute test you can run before you commit to a download or a subscription.