Generative Adversarial Networks (GANs) and Stable Diffusion
summary
Walkthrough builds a DCGAN on MNIST using Keras with Conv2DTranspose, Adam and binary cross-entropy, training for 50 epochs until noise converges into legible digits; discriminator and generator losses are computed and updated separately. Then Stable Diffusion via keras_cv.models.StableDiffusion runs text_to_image (default 512×512, 50 denoising steps), supports inpainting and textual inversion, and uses mixed precision and XLA to cut runtime from ~9s to under five seconds while preserving diverse outputs and reproducible seeds.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 0:28What are GANs
- 1:10Generator vs Discriminator
- 2:49Neural Networks
- 4:50Deep convolutional Gans
- 5:30Convolutional computation
- 6:50Tensorflow Keras
- 7:30Transposed Convolution
- 9:41Training Loop
- 11:10Generator Loss
- 11:37Code Walkthrough
- 19:57Stable Diffusion
- 27:10Stable Diffusion with KaraCV
- 28:47Performance Results
- 30:15Code
- 30:38Code Example
- 39:05Questions
- 41:02Resources
related talks
01h 21mGenerative AI with Diffusion Models
40mOverview of Generative Adversarial Networks (GANs)
43mHitchhiker’s Guide to Generative Adversarial Networks (GANs)
37mHarnessing the power of Generative Adversarial Networks (GANS) for supervised learning
01h 17mDeep generative models for image...
01h 00mTutorial on Deep Learning IV
53mGenerative AI: Overview and use cases
46mHow to Make Stable Diffusion Generate Not Just Anime and Cats, but You