Upscale a Product Photo — Free, Fast & Private
Listing photography has a hard floor and a soft ceiling, and most sellers hit the floor by accident. Every marketplace publishes a minimum size for a main product image, and most switch on a zoom or magnifier once the file is comfortably above it. The floors sit broadly between about a thousand and two thousand pixels on the longest side, but the exact figure differs by platform and changes without much announcement — so the number to work to is the one in your own seller help today, not a figure repeated on a blog. Below the floor the listing is rejected or the image is silently downgraded. Above it, the zoom viewer turns your photograph into the closest thing an online shopper gets to handling the product, which is why an image that merely clears the minimum and an image that rewards magnification are two different assets.
Reshooting is better than enlarging, and enlarging is much better than shipping an undersized file. If the product is still on the shelf and a phone is to hand, a fresh photograph in daylight against a plain wall beats any reconstruction, because it collects detail rather than predicting it. Enlargement earns its place when reshooting is not available: the product has already shipped, the shot came from a supplier, the only surviving copy is a crop from an old catalogue. In those cases the model on this page rebuilds the picture at 2x or 4x with reconstructed texture instead of the averaging blur an ordinary resize produces — the difference shows most on the surfaces buyers look at, which are fabric, grain, machined metal and moulded plastic. The on-device upscaler is the tool doing that work.
The order that keeps the cut-out clean. Frame first: crop to the composition the listing needs in the on-device cropper, since pixels reconstructed in a margin you are about to trim are pixels wasted. Cut out second, while the boundary between product and backdrop is still original data — the background remover has a real edge to follow there, and white background product photo covers the plain-backdrop convention most marketplaces expect for a main image. Enlarge third, so the model spends its whole effort on the product rather than on a backdrop already discarded. Compress last, in the image compressor — a 4x PNG can run to twenty megabytes, which is right for an archive and disastrous for a product page that has to load on a phone. Category-specific compression targets live on the Etsy and Amazon listing pages.
What the free tier covers, and what Pro adds. The upscaler runs realesr-general-x4v3 inside your browser: 4.9 MB, checksum-verified, executed through WebGPU or WebAssembly in 512px tiles. Free accepts inputs up to 1080px on the long edge, produces up to 4320px at 4x, and stamps a small "minipx.com" watermark in one corner — fine for testing a workflow, not fine for a live listing. Free also includes 2 AI runs shared across the three on-device AI tools before a cooldown starts at 60 minutes and lengthens by 30 each further time. Pro removes the watermark, makes on-device runs unlimited, and lifts the caps to 2048px in at 4x and 4096px at 2x, both ceilinged at 8192px of output — which matters for a seller processing a batch, because unlimited on-device runs are what make a catalogue practical at all. The engine itself is documented at the AI image upscaler hub.
Cloud HD, and where to spend the ten HD runs. Pro accounts see a switch beside the run button, off until you turn it on. Switched on, that one photo goes to our GPU, where a JPEG pre-clean pass runs before HAT-L rebuilds it at 4x — the pre-clean is the part that matters for e-commerce, because supplier images and re-downloaded catalogue crops have usually been re-encoded several times. Intake is 25 MB and 4 megapixels, output reaches up to 16× the pixels of the input — a 4 MP photo comes back at about 64 MP, and a full-size source takes four to five minutes. Each run spends one of the 10 HD runs in your month, inside a wider allowance of 150 runs on Pro Monthly or 200 on Pro Annual, resetting on the first of the month with no rollover; a failed run is not charged. The sensible split is obvious once you see the numbers: hero images and anything that will be magnified get an HD run, and the rest of the catalogue goes through the unlimited on-device path.
Where reconstruction is convincing, and where it will embarrass you. Texture is the model strength: knitwear, leather, wood grain, brushed aluminium, ceramic glaze all come back with believable structure. Geometry is safe too — straight edges stay straight, and a product silhouette does not wander. The failure case is text. Packaging copy, ingredient lists, model numbers and serial plates that had already dissolved will be rebuilt into letterforms that look sharp and say the wrong thing, and a customer reading them is being misled by your listing. Photograph any label you need legible as its own close-up shot. The same caution covers anything a buyer might rely on: dimensions printed on a box, a certification mark, a warranty number.
Where the images go. The enlargement runs on your own hardware — the model downloads to your browser once, caches, and works with no connection afterwards, so a batch of listing images never leaves your machine on the free path. The one exception is a Cloud HD run you switch on yourself as a Pro subscriber: that single photo is uploaded encrypted, processed in the United States on Cloudflare R2 storage and a Modal GPU, deleted when the run finishes, and in all cases within 24 hours, and is never used to train anything. Every other step in this workflow — cropping, background removal, conversion and compression — has no upload path at all. If the finished images are also going out as social posts, upscale an image for social media covers the different sizing pressures there, and fix a pixelated image is the page for a source that arrived visibly blocky.