How to upscale an image without losing quality

Enlarging a photo is not the same as inventing a new one. If you already have a file and you need it bigger for a print, a crop, or a stubborn CMS that wants 2000 pixels on a side, the job is to add plausible high-frequency detail without rewriting the subject. That is what Lumenfix image upscaler is for. This guide explains the honest limits, the 2×, 4×, and Ultra HD credit paths, and how to avoid the usual quality traps.

Start with the file you have

Open the original, not a screenshot of the original. A phone export, a camera JPEG, or a PNG from an editor is a real still. A 400-pixel social download of that still is a different, worse input. Super-resolution models cannot recover texture that was already discarded by a previous resize. They can only invent a texture that looks like the neighborhood of each pixel. The closer the source is to the sensor capture, the less the model has to guess.

Correct orientation before you care about scale. Lumenfix fixes JPEG EXIF rotation in the browser and shrinks any side longer than 4096 pixels so you do not pay to upload a 40-megapixel file the model will immediately downsample. If your master is a 24-megapixel raw, export a high-quality JPEG first. This product does not ingest camera raw.

What “without losing quality” actually means

People use that phrase to mean three different things.

First: the file should not look softer than the original at the same viewing size. A naive upsample (the kind a word processor does) is blurry. A Real-ESRGAN 2× pass is usually sharper at the new pixel grid. That is the Fast 2× path on /image-upscaler.

Second: edges and type should not ring or melt. Cheap upscalers put halos around lettering and hair. If your subject is a screenshot or a slide, do not start on the upscaler. Use unblur text when the letters already exist as pixels, and do not expect a generative model to “fix” a word. Lumenfix will not rewrite the sentence.

Third: identity should not drift. Face-aware stacks can rebuild a mouth that was eight pixels wide into a different mouth. The homepage photo enhancer uses GFPGAN plus a modest 2× pass for that reason. If you only need more pixels and the face is already sharp, stay on the upscaler and skip the face restore.

You will not get a medium-format capture out of a 640-pixel guest-list photo. “Without losing quality” here means “without an obvious extra generation step,” not “lossless in the information-theory sense.”

2×, 4×, and Ultra HD credits

Fast 2× can draw on the free daily allowance: one photo without an account, or 10 credits a day once you sign in; free results are JPGs up to 2048 px with a small watermark. Everything else costs credits, and each button shows its price before you run it. You do not need an account first: you can buy a credit pack at checkout, and the account is created from your payment email. Results paid with credits have no watermark and are not downsized.

Pro is a subscription that adds 400 credits a month. 4× does not need the subscription; any credits work. Ultra HD still costs its own credits. That split is cost, not a dark pattern. Topaz High Fidelity calls are much more expensive than the standard models.

Every button reads its price from the server quota payload. On the Standard SeedVR model, 2K is 2 credits, 4K is 4, 2× is 6, and 4× and 8K are 8 each; Fast 2× is 4. If the balance is too low, the page sends you to credit purchase. You do not need to sign in first: checkout creates your account. Credits and Pro do not replace each other.

A practical order: try 2×. If the file is for a poster and the 2× still looks soft at arm’s length, use 4×. Reach for Ultra HD only when you are printing large and you have already decided the composition is final.

Workflow that usually wins

Crop first in an editor you trust. Upscaling a huge empty wall wastes model time and can smear the subject.
If the file is an old portrait, run the photo enhancer before the upscaler. Faces benefit from GFPGAN more than from a raw 4×.
If you need more canvas rather than more pixels, stop. That is image extender, not this page.
If a caption or a sticker is in the way, remove it before you enlarge. Otherwise you upscale the junk.
Compare on a before/after slider at 100%. Zoom the parts that matter: eyes, type, fabric. If those failed, do not keep the file because the background looks “AI sharp.”

What the models will not do

They will not turn a 120-pixel face into a fashion plate. They will not undo a heavy beauty filter. They will not reconstruct clipped highlights. They will not make a screenshot of a UI look like a vector file — that is PNG to SVG in the browser, a different job.

They will not strip a photographer’s watermark. Lumenfix does not ship a watermark remover. If you painted your own caption, the remove tools are for that. If the mark belongs to someone else, you need permission, not a model.

Cost and storage

Every paid model call writes a provider_call row. Outputs live under a tools/ prefix and expire. Download anything you care about the same day.

If you want the longer version of “is this even a camera photo,” run the AI image detector first. It never uploads the file. An already-generated still can still be upscaled; you should just know what you are enlarging.

A short checklist

Use the original file.
Prefer 2× unless you are printing.
Do not upscale text-heavy screenshots on this page.
Do not expect Pro to include Topaz.
Keep the master. The result is a derivative, not a raw.

If you are evaluating the page for a team, look at three artifacts and ignore the rest of the marketing chrome. First, the credit balance: it should drop by the price on the button after each job. Second, the watermark: no result should carry one. Third, the credit buttons: their labels come from tools.quota, so a future price change should not require a frontend deploy of a magic number.

A note on viewing conditions. Judge the result at the size you will actually use. A 4× file that looks crunchy at 400% on a retina monitor can still be the right file for a foam board across a room. Conversely, a 2× file that looks fine in the slider can fall apart on a 27-inch display at 100%. The slider is a diagnostic, not the delivery surface.

Color management is boring and it matters. The upscaler does not promise a calibrated print profile. If you are sending the file to a lab, convert to the lab’s requested space in an editor after you download. Do not expect the Worker to attach an ICC profile that matches your monitor.

Finally, keep a short written note of which button you used. Six months later you will not remember whether a hero image was a cheap 2× or an 8-credit 8K. The signed-in history list helps for a week. After that, your own folder names are the record.

That is how you enlarge a still without pretending the model found detail that was never there. The upscaler is a tool for files you already have. Use it that way and the quality loss you were afraid of is mostly the quality you already lost before you arrived.