Begin with latent space
VAE maps data into continuous latent variables and supplies the intuition behind later latent generation.
Diffusion model paper path
Understand diffusion models from DDPM to Stable Diffusion
Diffusion models define a gradual noising process and learn to reverse it through denoising. DDPM showed that this route could produce high-quality images with stable training. Latent Diffusion moved the process from pixels into a compressed latent space, making high-resolution text-to-image generation much more practical. VAE and GAN provide the useful contrast for understanding what changed.
A four-paper generative-model path
Compare GAN and DDPM
GAN trains a generator against a discriminator; DDPM learns iterative denoising, usually trading slower sampling for more stable training.
Arrive at Stable Diffusion
Latent Diffusion performs diffusion over compressed representations and adds conditioning, providing the research foundation for Stable Diffusion.
Four papers around the diffusion-model breakthrough
Reading chronologically separates latent representations, adversarial training, pixel-space diffusion, and latent-space diffusion.
- 01
Auto-Encoding Variational Bayes
It made probabilistic latent-variable models trainable with ordinary backpropagation.
- 02
Generative Adversarial Networks
It reframed generation as a game between a generator and a discriminator.
- 03
Denoising Diffusion Probabilistic Models
It turned iterative denoising into a stable route to high-quality generation.
- 04
High-Resolution Image Synthesis with Latent Diffusion Models
It moved diffusion into a compressed latent space, making high-resolution text-to-image generation practical.
FAQ
Diffusion model paper FAQ
Which paper introduced the basis of Stable Diffusion?
High-Resolution Image Synthesis with Latent Diffusion Models is the main research foundation. It performs diffusion in a compressed latent space.
Why add noise before denoising?
The forward process turns a complex data distribution into simple noise. The learned reverse process removes predicted noise step by step to generate a sample.
How are diffusion models different from GANs?
GANs learn through a generator–discriminator game and sample quickly but can be unstable. Diffusion models are often more stable and cover the distribution better, but require multiple sampling steps.