It is not a folder of pictures
Explain, to someone who has never used one, what a model is holding, and why "it just copies things" is wrong in a way that matters at the desk.
If you remember one thing
The model kept the patterns and threw the pictures away. That is why it can make something new, and why it cannot make your shot.
Almost everyone starts with the same mental picture: a gigantic shoebox of images, and typing a prompt is rummaging through it. That picture is wrong, and it will cost you hours.
A filing cabinet gives you back exactly what you put in it, which is why nobody is impressed by one.
A cook who has made ten thousand dinners has no photographs of any of them, and can still make you something new.
Here is the number that kills it. The models you use every day were trained on billions of images. The file you are using is a few gigabytes. You cannot put a billion pictures in a few gigabytes. There is nowhere for them to be.
So what did survive? Relationships. That skin has soft edges and metal has hard ones. That light falls off. That faces have two eyes above one nose. That the word "dusk" tends to arrive with long shadows and orange skies. It kept the rules and lost the receipts.
It is closer to an artist who has looked at everything and remembers none of it specifically.
What people assume
A search engine with a paintbrush. You describe a picture, it finds the nearest one and hands it over slightly changed.
What it is
A machine that learned how pictures tend to go, and then builds one that follows those tendencies. It has no particular picture in mind, including yours.
This also explains the thing that frustrates people most. Ask for something common and it is superb, because it saw a great deal of it. Ask for something specific to you and it will confidently invent a plausible neighbour, because a plausible neighbour is genuinely the best it has.
You should now be able to say
- Why a few gigabytes cannot contain the training set
- What it kept instead of the pictures
- Why your show's asset is never in there, no matter how you word it
- Why common subjects come out better than rare ones
How to check yourself
Ask for something genuinely obscure and something very common with the same amount of description. Note which one needed you to supply a reference picture. That gap is the lesson, and you can feel it in one minute.
How this goes wrong
People spend an hour rewriting a prompt for a thing the model has never seen. No wording fixes a gap in what it was shown. Reference pictures fix it.
Do this
Generate the same subject twice: once as a generic version, once as a specific real object you know well. Write one sentence on what the specific one got wrong.
What counts as done
- Two generations from the same prompt structure
- One sentence naming a specific thing the model invented
- A guess at whether a reference picture would have fixed it
What you are given
- docFoundations worksheet
Check yourself
0 of 3 answered- Question 1: not answered
- Question 2: not answered
- Question 3: not answered
Answer before you scroll back. Getting one wrong here is worth more than getting it right by looking, and only your first answer counts towards points.
1Why can a model a few gigabytes in size not be storing its training images?
2Your production designed a prop last month. What is the most reliable way to get it into a generation?
3Why are common subjects usually rendered better than rare ones?
If you want to go further
Optional, and none of it is required to finish the course. Every link was opened and checked, and none of them is trying to sell you a different course.
