update : 12/06/2026
i was working with the mnist dataset and the tensorflow method to create this , and my model trained well ( 0.91 accurancy 0->1 ) , but it never worked during testing , so i gave up the mnist set

INSTEAD… i tried to make something new with the help of codex and power of json !
introducing the NEW and IMPROVED visual digits !

( github+FULL source code for PROPER info ( you won’t regret if ur interested )

in a nutshell tho :
now you can train it yourself , it currently has 10 samples / digit but you can now train it on your own system and even show it to me through github if your training dataset is superior !
it might be weaker than the CNN pathway but its simpler in working so yeah !

https://github.com/bhuvan-rai/visualdigits