Model Type: Generative Adversarial Network (GAN)
What it does: Learns to generate new images in the style of your dataset — from scratch.
Media: Images

FastGAN is a type of AI model that can invent new images after learning from a set of real ones. You show it a bunch of examples — and it learns to “hallucinate” new ones that look similar, but aren’t exact copies. Imagine training it on 100 of your own portraits, and it generates endless new faces in that style. Or you feed it abstract patterns, and it starts making its own.

How it basically works

FastGAN is made of two parts:

• a Generator that tries to create fake images
• a Discriminator that tries to detect which images are real and which are fake

They train by fighting each other (hence: adversarial). The generator improves by fooling the discriminator — and the result gets more convincing over time.

Workings

1. Pick a visual theme

What kind of image world do you want to train the model on?
Examples:

• A set of your own drawings
• Photos of hands, stones, shadows, objects, patterns
• Aesthetic category: “DIY sci-fi interfaces” or “found typologies”

2. Collect your dataset

• Aim for 50–300 images
• Images should be consistent in format (e.g., size, background, framing)
• Resize all images to square (e.g. 128×128 or 256×256)

3. Train the model

• This is more technical — we’ll help set it up using Google Colab
• Training may take 1–2 hours on a GPU
• You’ll see samples improve over time

4. Generate new images

• Once trained, the generator can produce new images endlessly
• These can be downloaded, curated, or used in other projects

Resources