I did a few image generations on Bing Chat or whatever it is called today which uses Dall-E3 as the image generator.
I was looking in particular for an Atari Joystick image in black and white with a red button. It generated some really cool images. Sharing my favorite one from the selection.
And that is where I was, wonder if I can just do it on my own computer and it might actually do a better job on a different model. Not the case.
The models I have tried don't seem to comprehend what Atari is. I guess that is where the whole training element then comes in. You have the ability to train the model on your own server to be able to do what you would like. Perhaps that is the next frontier in this experimentation.