Re: face quantization

[Didier Verna <didier@lrde.epita.fr>]

Reiner Steib <reinersteib+gmane@imap.cc> wrote:

> The algorithm in `gnus-face-from-file' itself hasn't changed.
> But the default of `gnus-convert-image-to-face-command' is
>   "convert -scale 48x48! %s -colors %d png:-"
> instead of
>   "djpeg %s | ppmnorm | pnmscale -width 48 -height 48 | ppmquant %d | pnmtopng".
>
> It's quite possible that the convert differs from the later.  Could
> you investigate?

        After some testing, it appears that convert does a better job (in the
sense that the result is closer to the original image), although I tend to
prefer the former solution because the results are brighter and a bit
"artificial", kinda special-effect :-)

        I also noticed that given my original jpg faces that are already at a
size of 48x48, the average quantization I get is 4 to 6 which is quite poor.
Maybe the size limit of 726 has something to do with it. As it changed in the
last few months (I cound't find a trace of this in the ChangeLog) ?


        Finally, it might be worth considering that after quantization to 4 or
6 colors, the resulting PNG images are *bigger* than my original JPEG ones.
So why not using JPG in the headers instead ?

-- 
Didier Verna, didier@lrde.epita.fr, http://www.lrde.epita.fr/~didier

EPITA / LRDE, 14-16 rue Voltaire   Tel.+33 (1) 44 08 01 85
94276 Le Kremlin-Bicêtre, France   Fax.+33 (1) 53 14 59 22   didier@xemacs.org