← AI Terminology

Diffusion Model

A diffusion model is a generative model that learns to reverse a gradual noise-addition process — it is trained to incrementally denoise data starting from pure Gaussian noise, ultimately generating realistic samples (images, audio, video) from nothing.

DALL-E 2, Stable Diffusion, Sora, and most modern image generators are diffusion models.
Why It Matters in AI
Diffusion models produce higher-quality, more diverse samples than GANs and are more stable to train — there's no adversarial game. They powered the text-to-image generation wave (2022–present), producing photorealistic and artistic images from text prompts at commercial scale. They've since expanded to video (Sora), audio (AudioDiff), protein design, and drug discovery.
Key Points
Aspect Description
LDM Latent Diffusion Models (Stable Diffusion) — diffusion in compressed latent space for speed
DDPM Denoising Diffusion Probabilistic Models (Ho et al., 2020) — foundational paper
vs GAN More stable training, better diversity; slower sampling (requires many denoising steps)
Conditioning Text, image, or other signals guide the denoising — enables text-to-image generation
Forward process Gradually adds Gaussian noise to data over T steps until it's pure noise — a fixed Markov chain
Reverse process Learned neural network (U-Net or Transformer) predicts and removes noise step by step
Simple Analogy
Start with a photograph, gradually bury it under random static until it's pure noise — then learn to reverse that process. A diffusion model learns to remove snow from a snowy image step by step. Text conditioning is the instruction that guides which image to unbury from the noise.
Common Usage Examples
  • Stable Diffusion: pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1")
  • DALL-E 3 via API: openai.images.generate(model="dall-e-3", prompt="...")
  • Sora (OpenAI): video diffusion model generating high-quality 60-second clips from text
  • AudioDiff / MusicGen: diffusion-based audio generation from text descriptions
  • ControlNet: adds spatial conditioning (edge maps, depth maps) to Stable Diffusion generation
Summary
In short: Diffusion models learn to generate by learning to denoise — starting from pure noise and iteratively sculpting a realistic image, video, or audio output.