Short answer
Diffusion models generate data by learning to reverse a gradual noising process—currently the dominant approach for high-quality image synthesis.
Why it matters
Understanding diffusion clarifies why generation takes iterative steps, how guidance works, and why edits/inpainting are natural extensions.
Diagram
Article
Math-lite
Focus on the iterative refinement idea and conditioning signals.
Examples
Text-to-image sampling
A UI shows progress as noise becomes a landscape over many steps. Users tweak guidance scale to balance prompt adherence versus diversity.
Inpainting a signboard
An editor masks a storefront sign; diffusion regenerates only that region conditioned on surrounding pixels and a new text prompt.
FAQ
How is diffusion different from a GAN?
GANs train a generator against a discriminator in an adversarial game. Diffusion models train a denoiser with a likelihood-related objective and generate via iterative refinement. Diffusion systems are often easier to train stably at high resolution, at the cost of heavier sampling.
What is latent diffusion?
Latent diffusion performs the denoising process in a compressed latent space learned by an autoencoder, then decodes back to pixels. This reduces compute compared with pixel-space diffusion.
Why does the same prompt look different each time?
Sampling starts from random noise (a seed). Different seeds explore different regions of the learned distribution unless randomness is fixed.
Sources
-
Denoising Diffusion Probabilistic Models
Ho et al.Key paper on DDPM-style diffusion generative models.
-
High-Resolution Image Synthesis with Latent Diffusion Models
Rombach et al.Influential latent diffusion work underlying many open image systems.
-
Google AI Blog
GoogleAccessible updates on generative media research and products.