How the numbers are produced

The rules the engine computes from

The relationship between letter height and reading distance is not a matter of opinion and has not been for decades. Sign practice, drawing on research from the United States Sign Council and the Pennsylvania Transportation Institute, works from two figures:

• Best impact. One inch of capital letter height for every ten feet of viewing distance. That is 8.33 mm of cap height per metre, or an angle of 28.6 minutes of arc at the eye. Above this, a line reads as somebody walks past without their having decided to read it.
• The legibility index. A six-inch letter is readable at a hundred and fifty feet by a person with 20/20 vision. That is 3.33 mm per metre, or 11.5 minutes of arc. Below this, letters stop being letters.

Between the two, a line is readable to somebody who stops and looks at it, and mostly lost on somebody who does not. That band is where a great many garment prints end up.

Why your line gets the height it does

Cap height on a garment is not chosen freely: it is whatever fits. The chest of an adult medium gives about two hundred and eighty millimetres of usable width, the back about three hundred and forty, a sleeve about ninety. The engine sets the longest line as large as that width allows, using an average upper-case advance of 0.62 of the cap height, and sets any subordinate line — a trailing year, a bracketed tag — at sixty per cent of it, which is the conventional treatment and also the reason such a line is the first thing to disappear.

Cap height is converted to a point size by dividing by 0.72, the cap-to-em ratio of the grotesque this engine assumes, so that the figure can be handed to a printer who thinks in points.

Which words survive

Each line is checked against the two thresholds at each distance. Anything below the legibility index has stopped being words. For a line in the band between the thresholds, the engine applies a glance budget: a person walking past has roughly a second and a half, reading research puts fixations at about four a second, and useful letter identification at around eight characters per fixation when type is comfortably above threshold and much less when it is not. The engine therefore returns the leading words that fit that budget and names the rest as lost.

That last calculation is the crudest thing this service does and we would rather say so than dress it up. It is a reasonable model of a glance, not a measurement of one. The thresholds either side of it are solid; the division of a marginal line into “these words” and “not these” is an estimate.

Why the pictures are resampled and not blurred

A Gaussian blur is the obvious way to make type look distant and it is the wrong one, because it softens a hairline and a heavy stroke in the same proportion. What actually happens as you back away is that the eye stops resolving detail finer than about one minute of arc, and a stroke narrower than that is not softened — it is averaged into the cloth and gone, while a heavier stroke beside it survives the same treatment intact. That difference is precisely the thing worth seeing, and a blur destroys it.

So each view is produced like this:

1. The line is drawn at full resolution, at its real proportions inside the print width.
2. The engine computes how many minutes of arc the print subtends at the chosen distance. That is how many independent samples a 20/20 eye can take across it.
3. The full-resolution render is box-filtered down to exactly that many samples. A box filter is an area average, which is what an integrating receptor does.
4. The result is reconstructed back up to display size by bilinear interpolation.
5. The mark Simulated at typical daylight is drawn into the pixels of the finished image.

The views are magnified: they show what your eye resolves at that distance, at a size you can inspect on a screen. That is a deliberate choice, because a true-scale rendering of a five-metre view would be a few pixels tall and would tell you nothing.

The mark on every image

Every generated view carries Simulated at typical daylight burned into the image data. It is not an overlay that can be hidden with a browser tool, and it survives a screenshot, because an image produced this way looks photographic and is not a photograph. On the free plan a watermark is also applied; it is positioned so that it never covers the simulation mark, and it never covers the list below.

What this service cannot know

This list is shown in full on every result, on every plan, and is never abbreviated or covered.

• The light it will be read in. These figures assume open daylight. A club room, a warehouse or a floodlit touchline all move the threshold, and low light moves it a long way.
• The angle it will be read from. A chest print seen from forty-five degrees is narrower than the same print seen head-on, and the arithmetic here is head-on.
• Ink spread in the fabric. Ink wicks along the yarn before it cures. On a knitted jersey a stroke can finish visibly wider than it was on the screen, which closes up the counters of small letters and costs you more than the height does. On a woven fabric the same print spreads much less. This is the single largest source of error in anything below about fifteen millimetres of cap height.
• The typeface you will actually print in. This engine assumes an upper-case grotesque of ordinary width and stroke weight. A condensed face is narrower and fits more, a light weight loses its strokes sooner, and a face with fine hairlines loses them first of all.
• Whether the reader is looking. Every figure here describes a person who has already decided to read the garment. Most people have not.
• The wearer's posture and the drape of the cloth. A print on a relaxed jersey is not flat, and a curved surface loses height at the edges.

Where a language model is used, and where it is not

No number on this site comes from a model. Cap height, point equivalent, angle at the eye, sample counts, thresholds, which words survive — all of it is computed in your browser by code you can read the behaviour of on this page.

A language model is used only to phrase an answer to a question you type in your own words, and it is given the computed result rather than the other way round. It cannot change a figure, and it is not asked to.

What this service has nothing to do with

We only measure how far your words can be read. What they say is entirely yours.

Glancerange never evaluates content, stance or subject matter. It never writes or edits copy, never suggests different wording, never shortens a line for effect, and never recommends a typeface on the basis of what the line says. Stroke weight and width are treated as measurements, because they are, and never as personality.

The acceptable use policy still prohibits unlawful material, hate and harassment, and use of somebody else’s trademark or copyrighted text. That is a platform floor rather than a judgement about lawful opinions, and the two are not in tension.

Do not use this for

Medical, legal, financial, safety, employment, credit, housing or judicial decisions. Safety-critical signage is governed by standards this service does not implement.

Your text and model training

The lines you type are not used to train any model. See privacy for how long anything is kept and how to have it deleted.