Post B6UVpmiya74UPZxcRc by wolf480pl@mstdn.io
 (DIR) More posts by wolf480pl@mstdn.io
 (DIR) Post #B6UVJJtgE2XnjqUF9c by unfa@mastodon.social
       0 likes, 0 repeats
       
       The difference between talking to a machine learning model to write down my words and talking to a machine learning model to draw me a picture is that without the machine learning model I can still type all the words myself exactly the same. But can you draw that picture without the machine learning model? Can you make that music track you have asked a computer to synthesize for you yourself? I can. All I ask for is that the photographer does not claim to be a painter. Prompters are not artists.
       
 (DIR) Post #B6UVJK9HI31MWDmhSy by wolf480pl@mstdn.io
       0 likes, 0 repeats
       
       @unfaAnother difference between talking to a machine learning model to write down your words and talking to a machine learning model to draw you a picture is that in the former case, you're not asking the model to add new informationText in a computer is very rigid. Sure, you can change the font, or color individual letters, but the model isn't trying to do that, and certainly it is not its task to maintain adequate spacing between letters, draw all the serifs, and make the text look nice1/
       
 (DIR) Post #B6UVpmiya74UPZxcRc by wolf480pl@mstdn.io
       0 likes, 0 repeats
       
       @unfaAll the information you want the model to put in the text was already present in your speech, and the rest is standardized.Whereas if you as a model to draw a picture, it is adding details that you didn't describe, and often didn't think of. It has to, because unlike in a dream, in a computer picture everything has to have a defined shape and color.A human drawer would make hundreds of implicit decisions about these things. An ML model just randomly picks them from a distribution.2/
       
 (DIR) Post #B6UW3nD7ZQw1yQjwwK by wolf480pl@mstdn.io
       0 likes, 0 repeats
       
       @unfaIf you had to generate warning signs using a standardized vocabulary of symbols and placement rules, and needed an ML model to convert your natural language description of those signs to pictures, that'd be more analogous to speech-to-text.
       
 (DIR) Post #B6UWM5VQHvNTva1ZuC by wolf480pl@mstdn.io
       0 likes, 0 repeats
       
       @unfaSo the question remains: what to do when you really don't care about the details in a picture, and just want "a car with a spiky bumper" drawn in a way that is immediately recognizable and does not communicate anything beyond "a car with a spiky number", but there is not a standardized set of graphic symbols to express that.
       
 (DIR) Post #B6UdyJzBbDf1WJ7l6O by wolf480pl@mstdn.io
       0 likes, 0 repeats
       
       @unfabtw. thanks.You helped me remember why I love programming