tarticle-tgtimes-ascii-art-rendering.mw - tgtimes - The Gopher Times
HTML git clone git://bitreich.org/tgtimes git://enlrupgkhuxnvlhsf6lc3fziv5h2hhfrinws65d7roiv6bfj7d652fid.onion/tgtimes
DIR Log
DIR Files
DIR Refs
DIR Tags
---
tarticle-tgtimes-ascii-art-rendering.mw (3596B)
---
1 .SH tgtimes
2 Synthetic ASCII Art
3 .2C 80v
4 .
5 .PP
6 Sometimes, an entirely new way to approach problems is found.
7 The media is often fond of covering these topics, be it quantum computing, blockchains, machine learning...
8 With these big toys come people experimenting with it.
9 .
10 .PP
11 This time we are reviewing the combo of Machine Learning and ASCII art.
12 .
13 .PP
14 I was expecting to present cute attempts at drawing images with computer-made text, but this is nothing of the sort.
15 Be prepared to see Science at the service of Art.
16 .
17 .IP "Generated Typewriter Art"
18 This research paper (no less!) shows that it is possible to write software for placing characters,
19 later typed during 6 hours by a human operator (for this example).
20 It is unsettling to see details much smaller than the characters themself be drawn on paper,
21 along with shades of grey of various intensities.
22 .
23 .FS
24 https://graphicsinterface.org/wp-content/uploads/gi2021-13.pdf
25 .FE
26 .
27 .IP "Generated ASCII Art in 2010"
28 This is possibly the state of the art of 2010 technology.
29 It was announced in the yearly conference SIGGRAPH hence presented to an audience full of computer graphics engineers.
30 The work of three researchers from Hong Kong, Xuemiao Xu, Linling Zhang and Tien-Tsin Wong, shows results of surprising accuracy.
31 The story does not tell whether there ever was a job offer "looking for ASCII artists for a scientific experiment" posted on the job board of the Chinese University of Hong Kong.
32 While the paper contains the complete math used, it also illustrates and explains methods to acheive this level of accuracy.
33 And no, it is not exactly machine learning, but hand-crafted strategies, combined statistics and other data massaging.
34 After all, it was published five years before things like Tensor Flow were introduced...
35 .
36 .FS
37 http://www.cse.cuhk.edu.hk/~ttwong/papers/asciiart/asciiart.html
38 .FE
39 .
40 .IP "Generated ASCII Art in 2017"
41 Is seven years enough time to improve upon that previous acheivement?
42 Quoting the previous paper as well as others in its own work,
43 Osamu Akiyama of the Osaka Faculty of Medicine kept the ball rolling.
44 This throws the big guns of machine learning to reach higher skies.
45 Its input data were Japaneses BBS such as 5chan (2chan) or Shitaraba,
46 which extends the ASCII set to all of unicode, notably the CJK set.
47 If the result of the paper are not enough to convince you,
48 the "Bad Apple" often used as a video demo in the Asian market have been converted in its entirety.
49 Something out of reach if doing every frame by hand.
50 The TensorFlow and Python code used is released publicly,
51 and an online demo is offered for the curious.
52 .
53 .FS
54 https://nips2017creativity.github.io/doc/ASCII_Art_Synthesis.pdf
55 https://nips2017creativity.github.io/
56 https://yewtu.be/watch?v=8GulN69Cgbg
57 https://www.vice.com/en/article/zmymwx/machine-learning-ascii-art-neural-net
58 https://github.com/OsciiArt/DeepAA
59 .FE
60 .
61 .PP
62 Is it so futile? Not so sure.
63 After all, representing anything with a computer is a matter of making a reality fit onto something terribly awkward and unnatural: a display.
64 The pixels, the square elements praised for providing a grid to throw data at, are promising, but themself have their quirks to be worked around.
65 For instance, sub-pixel geometry uses the same techniques as those presented by these papers for improving the realism of images beyond what a single pixel can offer.
66 It is, for ASCII art like for anything else, a matter of representing something, real or fictious, through a medium of some kind.
67 ASCII art has the ability to fit an image somewhere where there could only be text.