Trained a stylegan on Arabic Calligraphy (1/3)

3:29 PM · Jan 6, 2021

27
103
17
795
with mosaic frames (3/3)
1
2
1
59
Replying to @zaidalyafeai
wow. do they make any sense?
2
0
0
12
No they don't. The size of the dataset is only ~ 500 images. It would interesting to collect a larger dataset to see if they can capture language as well.
1
0
0
34
Fascinating! Would love to learn more about data / training, do you have a colab?
1
0
0
5
Replying to @zaidalyafeai
The edges look great! I'd be interested in a vector representation of the forms. Any components readable or pronounceable at all?
1
0
0
1
No, only on the letter level.
1
0
0
1
Replying to @zaidalyafeai
Fascinating! Specially given the difficulty of the domain. It seems the network learned the global form and some local graphemic structures but didn't learn how to connect graphemes into words. How much data do think the model needs to make the produced text intelligible?
2
0
0
5
Maybe there a linking to an nlp model should be made for that.
0
0
0
2