Skip to content
Server-sideDeleted in 30 minutes

Enlarging Is Arithmetic. Upscaling Is a Guess

Two operations share the word upscale. One redistributes the pixels you have across a larger grid, which is arithmetic — reliable, instant, and incapable of adding anything. The other synthesises detail that was never captured, using a model that has learned what detail of that kind usually looks like, which is a guess that happens to be very good. The first can be run across four hundred files without looking. The second produces four hundred confident inventions, and somebody has to look at all of them.

Use this without the search next time. Prathom Workbench puts Prathom's tools in your toolbar.

Add to Chrome — free

What it does

  • Enlargement by resampling, nothing invented
  • One rule applied across the folder
  • Runs in the browser, nothing uploaded
  • Deterministic, identical every run

How to use Batch Upscaler

  1. 1

    Check that enlarging is what you need

    If the goal is meeting a minimum dimension, resampling is the right tool. If the goal is recovering detail, no amount of enlargement will produce it and the model-based page is the honest option.

  2. 2

    Untick "Never enlarge", then set the target

    The tool refuses to enlarge by default, so a first run at a larger width returns the original size. Untick it, then set a width, a percentage, or a fit box to apply to every file.

  3. 3

    View one result at full size before accepting the batch

    Enlargement makes existing softness larger. If the result is unusable at full size, the source resolution is the problem and the batch will not fix it.

How it works

This page uses the browser-side resizer, configured for enlargement. Each file is decoded on your device, drawn into a larger canvas with the browser's smooth interpolation, and encoded back out into a zip.

Interpolation estimates the color of each new pixel from the ones around it. It is a weighted average, so the result contains no information that was not already present — the same picture, occupying more pixels, slightly softer because the same detail is now spread further.

Nothing samples randomly, so the same folder always produces the same output, and a failure on one file does not stop the batch.

What a model adds, and what that costs

The model-based upscalers in this catalog do something different in kind.

They have learned, from an enormous number of image pairs, what a given kind of detail looks like at higher resolution — how a brick edge sharpens, how hair resolves into strands, how a leaf gains a vein. Given a small image, they produce a large one containing detail of that kind.

That detail is plausible rather than true. On foliage nobody can check it, so plausible is indistinguishable from correct and the result is excellent. On a face, a serial number, a product label, or a logo, the correct answer exists and the model does not know it, so the invention is wrong in ways that matter.

Which is why they are single-file tools with a queue in front of them. The constraint is not a technical limitation to be worked around; it is proportionate to an operation whose output has to be looked at.

Where enlargement genuinely helps

Being specific, because the honest use cases are narrow and real.

Meeting a minimum. A specification demands a thousand pixels and your assets are nine hundred. Resampling clears it, adds nothing false, and takes seconds.

Filling a known space. An image has to occupy a fixed area in a layout and being slightly soft is preferable to being tiled or letterboxed.

Matching a set. A batch where one file is smaller than the rest can be brought up so the set is uniform, which the store catalog page cares about.

In all three the goal is a dimension rather than detail, and that is the distinction this whole page turns on.

What to do instead when you need detail

If the real problem is that the images are too small to be useful, the order of preference is short.

Find the original. A larger version usually exists — in a camera folder, in an email, from the photographer, or in whatever system produced the file that got downsized.

Rephotograph or rescan. For anything physical, this is faster than it sounds and produces real information rather than a good guess.

Then, if neither is possible, use the model-based upscaler on the files that matter, one at a time, and check each result against what the image is for.

Running four hundred files through an inventing model unattended and shipping the output is the one approach that reliably produces a problem nobody notices until a customer does.

What has been tested

The resizer this page uses is already in production, and the enlargement claim was checked directly rather than assumed.

A 1200×1800 JPEG with "Never enlarge" unticked and a width of 2000 came back 2000×3000 and 14 percent larger as a file — more pixels holding the same information, which is exactly what this page says enlargement is and is not.

Worth knowing before you start: "Never enlarge" is ticked by default, so a first run at a larger width returns the original size and looks like nothing happened. That is the tool refusing to soften your image without being asked.

Still to check on a later pass: a batch of small heavily compressed thumbnails, and a mixed-resolution batch brought to a common size.

Examples

Assets just under a required minimum

assets/*.jpg - 900 pixels wide, spec asks for 1000
resized-images.zip - 1000 pixels wide

The intended case. A small enlargement to clear a threshold costs almost nothing and adds no risk, because nothing is invented.

Old thumbnails you want restored

thumbs/*.jpg - 200 pixels, heavily compressed
larger, and just as soft

The wrong expectation. Resampling spreads two hundred pixels of information across a thousand. What you want here is the model-based upscaler, one file at a time, with each result checked.

Frequently asked questions

Why is there no model in a batch upscaler?

Because model upscaling invents plausible detail, and unattended invention across a folder is exactly what nobody should ship. The catalog's model-based upscalers run one file at a time on a GPU queue, which is a constraint that matches what the operation deserves.

Is enlarging pointless, then?

No, it is useful whenever a specification demands a dimension you narrowly miss, or when a file has to fill a known space. It just cannot make a soft photograph sharp, and any tool promising that from a resample is describing something else.

How much can I enlarge before it looks bad?

Perhaps twenty to fifty percent before the softness becomes obvious, depending on how sharp the source was and how closely it will be viewed. Doubling is visible on any screen. Quadrupling looks like exactly what it is. A sharp original tolerates more enlargement than a soft one, because there is more real edge detail to spread.

What about the GPU upscalers here?

They use a superresolution model and genuinely add detail, at two, four, or eight times. They also invent, which matters on faces, text, and products — the logo upscaler page covers what that costs when the correct answer is knowable.