Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Johannes Neumeier


Johannes Neumeier
Last Active
  • Re: I trained a neural network to kern a font (mostly)

    Without retraining the network some quick tests seem to mostly point you towards "potential" kerning pairs. Aside from the p% value I'd much more like to know how confident the script is about the amount it suggest a pair should be kerned - is it, based on the training data, a heavily scattered or homogenous pair. I see the end goal of the p% for automating kerning classes, though. Based on how often a character shares kerns with high confidence with other characters those could be suggested as groups.

    I did only run the script on two fonts and manually went in to compare only a few of the kerning finds. Some observations:
    • Characters that I'd often myself consider to have mirrored kerning, so like A V T O o, seem not to be recognised as such (mirrored kern values for also the same symmetrical glyph on either side). I can't surmise how the learning material would not reflect that also, but maybe other factors have more weight in the output, e.g. kerns that are mirrored or near equal should probably emphasise that in evaluation to skew the result to symmetry, where such is identifiable from the training material
    • Some kerns that I would think are common and are also non-negligent in value from my own manual kerning seem to have been completely disregarded, e.g. F-o A-T (first image)
    • Identifying classes seems to work reasonably well, meaning it finds similar shapes to receive the same kerning

    A-T kern missing completely (the serif?), while T-A does have an okayish value, A-V-A a tad more lose than my manual:

    A-V-A-T-A all with kerns, but A-V-A and A-T-A disturbingly unsymmetrical:

    Example of visually quite obvious and also not uncommon missing F-o kern, that wasn't in the table at all:

    Good example of appropriate identification of kern classes, even though the actual kern value is small:

    A c -10 p= 83 %A d -10 p= 95 %A e -10 p= 87 %...A g -10 p= 92 %A o -10 p= 95 %A q -10 p= 95 %

    Mixed results of kern class identification:

    F n -25 p= 99 %F p -25 p= 99 %...F r -25 p= 81 % < Good: Less certain, but left stem identified correctly as same group with n and p...F v -25 p= 55 %F w -25 p= 69 %F x -65 p= 61 % < How come different class?F y -25 p= 83 % < ~25% more certainty than for F-v and F-w?F z -20 p= 40 %

    Very interesting work! I'm curious how this develops.
  • Re: Is it ok to call a "typeface design" the UI of a font software program?

    The objective is to differentiate "font software" from an "image of lettering", clearly and quickly with minimal jargon. 
    Perhaps I am of a simple mind, but is it not, in fact, quite obvious to your average Joe or Josefine that the one is the font and the other is the stuff you make with it?

    I would go so far as to say that the idea that fonts are software is foreign to most people to begin with, and inventing a remedy for this self made problem might be solved simply by calling things what they are in the users' mind.
  • Re: Launch Timing

    Also, this is interesting (singular rather than plural on the book search)
    You can combine those hits by appending wildcard * or _INF.

    Also, comparing the decrease in "font" and simultaneous increase in "typeface" to their combined total shows that it might be the terminology shifting from the DTP age "font" to the more distinguished "typeface" term, without an actual decline in the volume of the results.

    While it does inform to the purchase timing, all in all this says little about the industry as a whole.
  • Re: 2017 Font Purchasing Habits Survey Results (worth the read!)

    49% of people say they have read an entire EULA before

  • Re: Using literature in a type specimen

    Use only an excerpt and it's Fair Use.
    Fair Use and using something for commercial gain are not quite the same, though, are they. It's not just the length of what you quote, but the motivation, that matters most for fair use - like research, education, commentary, news, etc.

    Or by way of example, it wouldn't be Fair Use for L'Oreal to advertise with quote's from George Martin's Song of Fire and Ice, describing the Lannister's fair hair, even if they only use just one phrase.