# VAE GAN Diffusion course: Module 4 — Generative AI & LLMs module: Module-4-Generative-AI-LLMs type: pdf source_url: https://personal-learn.armco.dev/files/Module-4-Generative-AI-LLMs/General/21_June/VAE_GAN_Diffusion.pdf pages: 7 --- [page 1] FROM NOISE TO IMAGINATION How Machines Learn to Imagine A visual journey through Autoencoders, VAEs, GANs and Diffusion Models AE VAE GAN DIFFUSION Mindgraph AI · Generative AI Foundations 01 [page 2] MODEL 01 Autoencoder — Squeeze, then Rebuild Think of it as a sketch artist who looks at a face, draws a tiny rough thumbnail from memory, then redraws the full face from that thumbnail. Original Image Full detail, large size Encoder squeezes detail down Code a tiny number summary Decoder rebuilds it back Reconstructed Close copy, slightly blurry Used for compressing images, removing noise, and spotting unusual / faulty data. Mindgraph AI · Generative AI Foundations 02 [page 3] MODEL 02 Variational Autoencoder — Compress into a Cloud Same sketch artist — but now the tiny thumbnail isn't a fixed point, it's a fuzzy cloud of possibilities. Pick any point in the cloud and you get a new, believable face. Input Face Encoder Latent Cloud a dash of randomness Decoder New Faces Every draw is slightly different The randomness lets a VAE generate brand-new, never-seen examples — not just copies. Mindgraph AI · Generative AI Foundations 03 [page 4] MODEL 03 GAN — Forger vs. Detective Two networks battle each other. The Forger paints fakes. The Detective tries to catch them. Both improve, round after round, until the fakes are undetectable. The Forger (Generator) Takes random noise and paints an image, trying to fool the Detective. The Detective (Discriminator) Looks at images and guesses: real photo, or Forger's fake? VS They train together, again and again — the Forger learns from every catch, the Detective learns from every fake — until the images look completely real. Mindgraph AI · Generative AI Foundations 04 [page 5] MODEL 04 Diffusion — Sculpting an Image Out of Pure Noise Start with TV static. Step by step, the model removes a little noise at a time, guided by a text prompt — until a clear picture emerges. Pure Noise Step 1 Step 2 Step 3 Clear Image "a golden retriever puppy in the snow" This step-by-step denoising is what powers most modern text-to-image AI tools today. Mindgraph AI · Generative AI Foundations 05 [page 6] THROUGH THE YEARS The Timeline of Generative AI 2013 VAE Variational Autoencoder introduced 2014 GAN Generative Adversarial Networks born 2015 Diffusion First diffusion-based generative model 2020 DDPM Diffusion models become practical 2021 DALL·E Text-to-image goes mainstream 2022 Stable Diffusion Open, fast diffusion image models Mindgraph AI · Generative AI Foundations 06 [page 7] SPOT THE PATTERN Which Tool Uses Which Model? Diffusion Models DALL·E 3 Midjourney Stable Diffusion Nano Banana Flux Sora (video) GANs StyleGAN faces Deepfake video tools Art generators (early) Image super-resolution VAEs / AEs Image compression Anomaly / fraud detection Latent space in Stable Diffusion's core Most image tools you use today (Nano Banana, DALL·E, Flux) are diffusion models — GANs and VAEs paved the way. Mindgraph AI · Generative AI Foundations 07