"Enhance" is a television idea. Real upscaling is more interesting than the fiction and considerably more limited, and knowing where the line is saves a lot of disappointment.
The short version: a super-resolution model can make a small image genuinely look better at a larger size, because it has learned what real detail looks like and can synthesise something plausible. It cannot recover information that the camera never recorded, because that information does not exist anywhere in the file.
Why it is better than stretching
A conventional enlargement interpolates: each new pixel is a weighted average of its neighbours. Averaging is a blur, so a stretched image is a bigger, softer version of the original. Edges that were one pixel wide become four pixels of gradient.
A super-resolution model does something different. Trained on millions of pairs of small and large versions of the same image, it has learned the statistics of real detail — what an edge looks like at higher resolution, how texture behaves, what a fine line becomes when you can see it properly. Given a small image, it produces a large one consistent with those statistics.
The result genuinely looks sharper, not merely bigger. Edges are crisp rather than smeared, texture has structure instead of mush, and compression artefacts are often cleaned up as a side effect because the model has learned they are not what real images look like.
What it cannot do, stated plainly
It cannot read unreadable text. If a licence plate or a document is too small to resolve, the model will produce sharp, confident, plausible characters — and there is no reason to believe they are the right ones. This is the most important limitation to understand, because the output looks authoritative.
It cannot fix motion blur or a missed focus. Those are not resolution problems; the information was destroyed at capture and no amount of resolution recovers it. An out-of-focus photo upscaled 4× is an out-of-focus photo at four times the size.
And it invents. Everything it adds is plausible rather than true. For photographs and illustrations that is exactly what you want. For anything where the detail is evidence — a document, a plate, a face you need to identify — it is the wrong tool, and a tool that lets you believe otherwise is doing you harm.
Filum labels this honestly: it enlarges and sharpens, it does not restore or de-blur, and it says so rather than promising enhancement.
Four times each side, and what that means
The scale is exactly 4× per side, so a 500×500 image becomes 2000×2000 — four times the width, four times the height, sixteen times the pixel count. That last number is why very large sources are handled differently: sixteen times the pixels of an already-large photo is more than a browser can process in a reasonable time, so an oversized source is first scaled down to a workable size and then upscaled 4× from there. When that happens Filum tells you the exact size that was processed rather than quietly changing what you asked for.
Why it takes a few seconds, and why that is the point
Online upscalers are faster. They run on server GPUs built for exactly this, and they are very good at it. They also require your image to be sent to them first, and most of them charge — upscaling is one of the more reliably monetised image operations on the web.
Filum runs the model in your browser instead. The model and its runtime are fetched once from this site, about 18 MB, and cached; the work then happens on your own processor. It takes a few seconds rather than being instant. In return the image never leaves your computer, there is no account, and there is no per-image charge.
The model is realesr-general-x4v3 from the Real-ESRGAN project, released under the permissive BSD-3-Clause licence and self-hosted here rather than pulled from a third-party network — so no one else learns what you are upscaling either.