I can spend all day playing with Dall-E Mini, honestly, it blows my mind.
And i've seen images generated by Dall-E 2, they're even more mindblowerer
i played with it for a while a few days ago.. it was fun for a while but then you kind of learn its limits…. it seems to be able to put characters in to scenes well, but it doesn’t seem to be very good at making them do any thing besides just being in the scene.. like “wario wacking off luigi” might just show wario and luigi in a scene but you can’t make them punch each other really or any thing weird too much..
oh yeah maybe pro-tip: after you click Run the first time, if it gives you the “network too busy” error you can just keep pushing enter over and over to keep trying instead of having to click “OK” and “Run” again over and over..
but i’m also afraid of like……… are we even going to get any thing more advanced than this? i’m waiting for when they decide “you know what, people are abusing this AI too much! we will not release any thing else after this!”
> are we even going to get any thing more advanced than this
have you seen the images generated by the people who have access to the beta of Dall-E 2?
People in several technology forums have noted that the developers AND google (part of the developing team) are kinda refusing to open the source of the second version of this technology, partially because of that, fears of being abused.
@berkberkman @FF777 i think for these things (neural networks) they use arrays of GPUs., You could run a local instance of Dall-E Mini with a single strong off-the-shelf GPU, a guy in 4chan says that it's running one and it has an output of an image every 10 seconds. The stronger the GPU the faster.
@hideki @berkberkman
the only thing i’m wondering is, where does it get all of its database from? how does it know what hatsune miku looks like? would you have to store a huge database on your computer i guess? and how humongous would that be?
@FF777 @berkberkman What i know is that you program the AI model, and then they fed the model million of random images each one with a caption and the model little by little learns to identify what it sees.
By the time the training ends, the AI built a knowledge database. So if you download a pre-trained model it already has knowledge on how everything looks. But you could also download a "blank" model and train it yourself with the images you feed it.
So TLDR no, you don't need to download the "source" database, just the knwoledge database that the AI formed during it's training. How big is that database i don't relly know.
@FF777 @berkberkman so i checked the repos and i think the pre-trained model for dall-e weights around 2gb
@hideki @berkberkman
nice