Upscale an Image 4x
Upscale an image four times in width and four times in height — sixteen times the pixels — in a single native pass. Nothing is rescaled before or after, so the output is entirely the model's own work rather than a resized version of it. What that model produces is reconstruction and not recovery, and the difference is what decides whether the result is usable in print or only on screen.
Use this without the search next time. Prathom Workbench puts Prathom's tools in your toolbar.
Add to Chrome — freeDrop an image here, or click to browse
Up to 25 MB. Uploaded temporarily, deleted after processing.
What it does
- Native four times factor, single model pass
- No pre-scaling or post-scaling of the result
- Sixteen times the pixel count of the source
- Reconstructed texture rather than recovered detail
How to use AI Upscale 4x
- 1
Start from the best source you have
A model reconstructs from what survived. An already-compressed or already-enlarged file gives it artifacts to enlarge rather than detail to work from.
- 2
Run the pass
The image goes through the super-resolution model once at its native four times factor, with nothing rescaled before or afterwards.
- 3
Check faces and text before anything else
These are the two places where invented detail is obvious. Everything else tolerates a plausible guess; a face and a word do not.
How it works
One model, one pass, nothing else.
RealESRGAN takes the image and outputs it at four times the width and four times the height. That factor is fixed in the model's architecture — it is not a parameter that was chosen here, it is what the network was built to do. This page uses it directly, which is why the output is exactly four times and why there is no interpolation step before or after to soften the result.
Sixteen times the pixel count is worth stating plainly, because it is what the file size and the processing time both follow. A one megapixel photo comes back as sixteen megapixels.
Reconstruction, not recovery
The mental model that causes disappointment is the one from television: enhance, and the detail appears. That is not what is happening.
The model was trained by taking sharp images, degrading them, and learning to undo that degradation. What it acquired is a very strong sense of what real photographic texture looks like at higher resolution — how skin pores are distributed, how fabric weave repeats, how a leaf edge behaves. When it enlarges your image it applies that knowledge to fill the space between your pixels.
For most of a photograph this is genuinely excellent, because most of a photograph is texture that follows rules. Bark, cloth, stone, hair in bulk, foliage — these all upscale convincingly, and the invented detail is indistinguishable from the real thing because it is drawn from the same distribution.
Where the guessing shows
Two subjects break the pattern, and they are the two that matter most.
Faces are the first. A face is not generic texture; it is a specific arrangement that a viewer knows intimately. When the model reconstructs an eye from a handful of pixels it produces an eye, and often a slightly different eye than the one in the room. At small source sizes this can shift someone's appearance in ways that are hard to name but easy to feel.
Text is the second and it is worse, because text has a correct answer. The model knows that letterforms should be crisp, so it makes them crisp. It does not know which letter it is looking at when only four pixels remain, so it produces the most probable one. The output reads cleanly and says the wrong thing, which is strictly more dangerous than an unreadable blur.
Choosing this over 2x or 8x
The three upscale pages differ in what happens around the same model pass.
At 2x, the four times output is halved afterwards. Downscaling averages neighboring pixels, which suppresses the model's occasional mistakes and produces a conservative, clean result at a modest size.
At 8x, the four times output is doubled with lanczos interpolation. No new detail is created in that final step — it is ordinary resampling — so an 8x result carries the same amount of real information as a 4x one, spread across four times the pixels.
This page is the middle case and the honest one: every pixel came from the model, nothing was thrown away, nothing was padded out.
Publication gate
This page ships once the workflow has been run against a texture-heavy source, a portrait small enough to stress facial reconstruction, and an image containing small text, with all three inspected at full size against their originals.
Examples
Small product photo for print
Product surfaces and hard edges upscale well. The model has clear structure to extend and no faces or lettering that must be exactly right.
Cropped photo with a distant sign
Text is reconstructed into confident letter shapes that are frequently the wrong letters. Crisp and legible is not the same as correct.
Frequently asked questions
Does upscaling recover detail that was lost?
No. It generates detail that is plausible given what remains. A super-resolution model is trained on pairs of images, a sharp original and a degraded copy, and learns what degraded regions usually looked like before. When it enlarges your photo it applies that learned expectation. The result often matches reality closely, because most textures behave predictably, but nothing is being read back out of the file.
Why does small text come back looking sharp but wrong?
Because the model has learned that letter-like shapes should resolve into crisp letters, and it will produce crisp letters whether or not it can tell which ones. At small sizes the difference between similar characters is a handful of pixels, and once those are gone the model fills the gap with whatever is most typical. This is the single most important thing to check, since the output looks more confident than the input did.
How is this different from the 2x and 8x versions?
All three use the same model at the same four times factor. The 2x page halves the result afterwards, and the 8x page doubles it with ordinary interpolation. This page does neither, so it is the only one of the three where every output pixel comes from the model. If you want the most detail per pixel, this is the setting.
Is there a size limit on what I can upload?
In practice yes, set by GPU memory rather than by a stated number. Sixteen times the pixels has to fit in video memory during the pass, so a large source can be refused where a smaller one succeeds. If a job is rejected for capacity, reducing the source dimensions and running again is usually the quickest route.
Is the AI Upscale 4x tool free, and do I need an account?
It is free and there is nothing to register for. No account is created, no address is collected, and no reduced-quality preview is withheld for a plan that does not exist. The constraints that do apply come from the workflow itself rather than from billing: 25 MB in, a shared GPU queue, and a thirty-minute window to collect the output.