A Very Short Introduction to Diffusion Models

Abstraction: Forward noise addition and learned reverse denoising for image generation

Key points:

• Diffusion models work via two processes: forward diffusion adds Gaussian noise with variance βₜ over T Markov steps until the image becomes isotropic Gaussian; reverse diffusion trains a neural network p_θ to recover the original from noise
• The reparameterization trick allows sampling xₜ at any arbitrary timestep directly, avoiding costly sequential computation through all intermediate steps
• Reverse process parameterizes mean μ_θ(xₜ, t) and covariance ∑_θ(xₜ, t) at each timestep; for small enough βₜ the reverse distribution is Gaussian
• Key models: DDPM (Ho et al. 2020), Latent Diffusion Models (LDM), Stable Diffusion, DALL-E 3, Imagen by Google
• Outperforms GANs (training instability, mode collapse) and VAEs (surrogate loss issues) for high-resolution, diverse image generation
• Loosely based on non-equilibrium thermodynamics; operates in a latent continuous feature space similar to VAEs

Connections: Stable Diffusion · Dall E · Openai · Google · Diffusion Models · Generative AI · Image Generation

Source: https://kailashahirwar.medium.com/a-very-short-introduction-to-diffusion-models-a84235e4e9ae