AI Upscalers for Outdoor Boulder
Enter the hub password to continue.
That's not right. Try again.
Which AI upscaler to use, and what it can't fix
Researched 28 July 2026. Every price and setting below came from the vendor's own page or model card, fetched that day and linked. Nothing is from memory, because this stuff changes monthly.
The job: turn phone photos of prairie, foothills, rock and alpine lakes into full-width page banners (about 2200 by 600 pixels, under 300KB) on a site whose credibility depends on not showing terrain that isn't really there. The thing to beat: the free tool already on your laptop, which tripled the detail on Rabbit Mountain but still smeared the grass and made the rock look waxy.
The short answer
Use Topaz Gigapixel through Replicate, at 5 cents an image. It's the one that can't invent scenery even by accident, which matters more here than anywhere else. Before you spend anything, though, try the free model in the last section: at the file sizes this site uses, it may well be indistinguishable, and it costs nothing to find out.
And the honest part: upscaling buys you about one quality step, not two. It cannot turn a 640-pixel photo into a good full-width banner. Nothing can. The real fix is the original file.
1. Why this one, in plain terms
Upscalers split into two camps, and almost no review mentions it because most of them test faces rather than landscape.
- Faithful ones enlarge and sharpen what's actually in the photo.
- Generative ones invent plausible new detail. On a portrait that's flattering. On a photo of Royal Arch it can add a rock ledge that has never existed, which on this site is the same kind of error as publishing a wrong trail distance, just with a picture around it.
Topaz Gigapixel wins because it is faithful by construction, not by a setting you have to remember. Topaz sorts its models into three families, and only one of them is the non-inventing kind. The endpoint we'd use has no prompt box, no creativity slider and no resemblance dial. There is nothing to leave switched on. I checked this against the live model page rather than taking the docs' word for it.
Asked directly what to use for "640x427 soft, already-compressed landscape photos, enlarge 3.5x, do not
invent terrain detail," Topaz's own documentation answers Low Resolution 2, which is
described as recovering detail "without modifying the source characteristics."
Why through Replicate rather than Topaz directly: Topaz's own API wants $50 a month for 500 credits. Replicate bills per image with no subscription, and every one of your banners lands in the cheapest tier because they're only 1.3 megapixels. For 40 photos that's the difference between $50 and $2.
Runner-up: Magnific "Precision"
Magnific (owned by Freepik) is the only vendor that has drawn the faithful-versus-inventing line explicitly in its product, with an endpoint literally titled "High-Fidelity Super-Resolution (No Hallucinations)". Three catches: it needs a paid plan on top of per-image cost, it gives you no control over output size, and its defaults add a little synthetic film grain, which is invented texture even if it's harmless. Their own technical docs also contradict their marketing page on whether it can add elements, which looks like a copy-paste error rather than a lie, but it means you'd verify by eye. Worth testing only if Topaz disappoints.
2. The full comparison
Cost is for your actual job: roughly 1.3 megapixel output. Every column scrolls into view on a phone, nothing is hidden.
| Tool | Invents detail? | Cost per photo | Commercial use | Notes |
|---|---|---|---|---|
| Topaz Gigapixel via Replicate (the pick) | No No prompt or creativity controls exist | $0.05 | Fine. Their restriction is on reselling the service, not on publishing your own photos | Best available evidence for photo fidelity, though see section 4 on how thin that evidence is |
| 4xNomosWebPhoto (free, on your laptop) | No | Free | Yes, with a credit line | Trained specifically on photos degraded the way yours are: phone noise, lens softness, repeated compression. Try this first |
| Bria | No | $0.04 | Strongest of anyone here: trained only on licensed stock, with legal indemnification | Almost no landscape evidence either way |
| Recraft crisp upscale | No Sharpens rather than rebuilds | $0.006 | Yes on paid plans | Cheapest paid option. Less lifting power |
| Magnific Precision (runner-up) | Mostly no Adds slight synthetic grain by default | About €0.10 to €0.20, plus a paid plan | Included on paid plans | No control over output size |
| Real-ESRGAN (what we ran tonight) | A little at the texture level | Free | Yes | 2021 design. Known to smear grass and wax rock, which is exactly what we saw |
| Magnific Creative, SUPIR, Clarity, Topaz Wonder / Bloom / Redefine / Super Focus 2 | Yes, by design | varies | varies | Don't use any of these on this site. Turning the dial down still leaves invention on |
| Google Vertex Imagen | Leans faithful | not confirmed | Yes | Watermarks by default, and the watermark is invisible, so you would not catch it by looking |
3. What it costs to actually decide
A bake-off on your three hardest photos (one grass-heavy, one rock-heavy, one water and sky) across four different models: 57 cents. Doing all 40 remaining photos afterwards: two dollars.
So money isn't the deciding factor here at all. The only real question is whether the result clears your eye, which is a look-at-it decision rather than a spreadsheet one.
4. The honest caveat, which is the important bit
Almost every upscaler comparison online tests faces, text and product shots. That isn't an accident: a reviewer can tell at a glance whether a face came out right, and nobody can do that with grass. So treat every ranking above, including mine, as a hypothesis to test on your photos rather than a proven result.
The photos that actually need enlarging are nine page banners between 1,024 and 1,269 pixels wide: Walker Ranch, Gross Reservoir, Eldorado, Gregory Canyon, Nederland, South Boulder Creek, Wonderland Lake, Anne U. White, and the camping photo. Taking those to 2,240 is roughly doubling, not tripling. That is well inside what a good upscaler does convincingly, and it is the difference between "this probably won't work" and "this should look fine."
The arithmetic that still bites is for anything genuinely tiny. Rabbit Mountain is the real example: its usable detail is about 620 pixels wide and it needs to fill 2,240. That is the case where an upscaler buys one step and no more.
And compression eats much of what you paid for. Your good banners currently run at a file size that's perfectly sensible for the web and also low enough that a lot of the fine detail an expensive upscaler adds gets thrown straight back out when the image is saved. Which is precisely why the free model deserves the first try.
So can an upscaler rescue a 640-pixel photo into a good banner?
No. It can rescue it into an acceptable one, which is a different thing. Expect less smearing in grass, less plastic on rock, more honest-looking texture. Expect it to still read as a bit soft at the top of a page on a good screen, and the people most likely to notice are the ones who actually hike these trails.
The ranked fix, in order: (1) find the original full-size file, or reshoot, since these are trailheads twenty minutes from your house and your phone shoots forty times the pixels; (2) if a place can't be reshot, use the small photo somewhere it's displayed small and looks great, and give that page a different banner; (3) upscale only what's left. Upscaling is step three, not step one.
5. Try the free one first
The model on your laptop is a 2021 design. The people who care most about photographic upscaling have moved on to newer ones, and 4xNomosWebPhoto is trained specifically on photographs degraded the way yours are: real camera noise, lens blur, and the repeated compression that happens when a photo gets saved for the web more than once. That's an unusually close match to your actual problem, it runs on your laptop with no graphics card needed, and it costs nothing. There are ten published before-and-after examples you can flip through yourself.
Licence is free for commercial use as long as the author gets a credit line somewhere on the site, which one line in a colophon covers.
Sources
Vendor documentation and model cards, all fetched 28 July 2026: Topaz API and pricing · Topaz model selection · Low Resolution 2 · Replicate Topaz · Bria · Recraft · Magnific Precision · Magnific pricing · fal.ai Topaz · Stability pricing · Vertex Imagen · Picsart · Let's Enhance · OpenModelDB · Philip Hofmann's models · spandrel.
Comparison blogs and forum threads, treated as weak evidence because they test faces rather than landscape: rangy.ai 2026 roundup · Upscayl discussion on photo models.
Private · For outdoorboulder.com · ← Hub