But how do AI images and videos actually work? | Guest video by Welch Labs

Summary by Krater

This video explains the mathematical and physical intuition behind text-to-video diffusion models like Wan2.1 and Stable Diffusion, connecting high-dimensional diffusion processes with Langevin dynamics, random walks, and score-based generative modeling.

From the video

Answers: How do AI text-to-video diffusion models work?

• Diffusion models in machine learning
• Brownian motion in random walks
• CLIP model representation space
• Denoising Diffusion Probabilistic Models
• Score-based generative modeling
• Classifier-free guidance in diffusion

What it concludes

• Diffusion models shape pure noise into realistic images by progressively removing noise in reverse time steps.
• The learning objective of diffusion models is mathematically equivalent to learning to predict the total noise added across the forward steps.
• Classifier-free guidance conditions generative models by combining unconditional and conditional vector fields to steer image generation toward specific text prompts.