Generative Computer Vision
Generative and Discriminative Modeling
Representation Learning and Multimodal Alignment
Diffusion and Flow-Based Generation
What You Will Learn
Core Concepts: What makes a model generative, latent spaces, and the three objectivesbehind modern image synthesis — likelihood, adversarial, and denoising.
Representation Backbones: In-depth study of the Transformer and the Vision Transformer (ViT), followed by the encoders that give generative models their semantic handles: CLIPfor language-image alignment and DINO for self-supervised visual features.
Classical Generative Families: Variational Autoencoders (VAE) and discrete tokenization, and Generative Adversarial Networks (GANs).
Diffusion Models: The core formulation (DDPM) and fast sampling (DDIM), classifier-freeguidance, latent diffusion, and the techniques that make generation controllable —structural conditioning and text-guided image editing.
Frontier Topics: Diffusion Transformers (DiT) and flow matching, the architecture andobjective behind the current generation of image and video models.
Hands-On, Project-Based Learning, gain experience with essential tools like Python, PyTorch, and Hugging Face Diffusers.
Coding Intensive: You must be a "coding person." You will be required to read and write code for exercises, exams, and projects.