TwiddleLab / Artificial Intelligence / Generative AI: Making New Things
Project: Design a Critter Generator
What you will learn
- A generator learns the cloud of its training data; directions in its code map let you steer one feature at a time.
- Quality costs something: bigger codes or more steps give sharper results, but take more model or more time.
- Check outputs for near copies of training data before you use them.
- Honest provenance, like a signed Content Credentials label, tells people how a picture was made.
How the lesson goes
Design a critter species that has never existed, and prove where it came from.
Start by choosing your generator.
Train your generator on the critter collection.
Press Train. Watch the loss and the samples.
Find the fur direction in your code map.
Drag the yellow tip to turn the arrow. A good fur arrow adds fur and leaves the ears alone.
Make a family of 5 critters along your fur arrow.
They start spiky. Choose the step size so the last one is very fluffy, and nobody leaves the cloud.
Your studio, from data to signed picture.
Every idea of this course has a place in it.
Set the quality knob: sharpness 90 with the smallest setting.
Your generator's knob: code size for an autoencoder, steps for diffusion.
More code numbers or more denoising steps keep more detail, but cost model size or time. The smallest setting that reaches the goal is the best deal.
A customer orders a fluffy, long-eared, orange critter. Design it.
Drag the point on the code map and tap a colour.
Filter the candidates: keep originals, reject near copies.
Drag each candidate into Keep or Reject.
A candidate 90% or more alike to a training critter is close to a copy of someone else's picture. Rejecting near copies keeps the species original.
Life: give your critter an honest Content Credentials label.
Drag the true lines onto the label, the rest to Leave off.
A provenance label is only useful if it is true: the picture was generated by AI, with your studio's tool, by you, on this date.
Why label AI pictures with signed credentials?
Signed credentials record who made a picture, with which tool, and that it is unchanged since signing. They do not stop copying, and they only help when tools and platforms keep them attached.
Sign and publish your species card.
Your critter, its label and a signature.
Previous lesson: Can We Detect Generated Content? · Part of the course: Generative AI: Making New Things
Lessons are interactive and need JavaScript. Turn it on to play them.
