FamilyDev
patreon
What does SD know about love?
🕑 Added 2024-07-05 18:43:54 +0000 UTCHow smart is SD? How does the neural network understand some tokens? That's what we're going to find out today.
The first interesting prompt without any negatives is an empty prompt. What will the neural network draw if it isn't prompted in any way? Let's give her something to create what she wants.
We get a cloudy silhouette resembling a woman with long hair. What can we learn from this? We can learn that in the database of this checkpoint, there are many images of women, and there are so many of them that we get a woman 99 times out of 100. The neural network didn't create anything, but it drew what was most represented in its database.
Let's try the same thing on another checkpoint.
We got a room with a window. This checkpoint has seen lots of different rooms. So, this model is more suitable for this task.
And finally, let's try our third checkpoint.
A woman again, but not a silhouette, but almost a photograph. Yes, you got it right. Checkpoint is clearly focused on realism and women. This way, you can check what particular checkpoint is focusing on.
The next one is the AI token. I wonder how the neural network will draw itself.
In 10 out of 10 cases, it's an Asian woman with a robotic body and neon signs on the background. You can imagine what this token was trained on.
The future. How does the neural network see it? Is it pessimistic or optimistic?
It seems that the neural network was trained for the future on posters for a film about space. The short summary is that such philosophical tokens will not lead to some kind of discovery. You will just find out how and on what neural network was trained on.
Everyone uses a huge number of tokens for their art, and the same number of negatives. But what do these negatives affect? Let's try to figure it out.
The first negative that we will write in the prompt is "poor quality".
We get a very blurry picture of a girl.
And now let's try to write a "blurry" prompt.
We get the same blurred picture, but now of nature. It turns out that these two negatives have almost the same effect on the result, so there is no point in writing them together. Check what you write in negative and do not copy 100, 500 negative, the point is to remove what you don't want, and not all bad things for the sake of all good.
Masterpiece. I've seen this prompt from other people millions of times, and wondered how it affected the image? Is it true that one prompt can make a masterpiece out of any prompt?
As a result, neural network recognizes this token as a girl more often naked on the picture on the wall. Yes, everything is so ordinary and simple, neural networks don't know how to think or how to create a masterpiece from your prompt.
Finally, love. It's a complicated feeling that even people can't explain. What does AI know about it? Will it understand our human mind? Can it love and subtly portray love in one image?
A heart. More often, with the text 'LOVE'. That's all. For a neural network, love is a painted heart with text.
Result. A neural network is a deterministic system that allows you to create incredibly beautiful things if you know how to work with it. The neural network doesn't know how to think, interpret your words; it can't give you something new, because its basis is limited and fine-tuned.
