Definition
Diffusion models train a network to denoise data that was progressively corrupted with noise. At generation time, they start from pure noise and iteratively denoise into a sample.
Stable Diffusion, DALL·E-style systems, and many audio generators use diffusion or related score-based methods.
In simple terms
Imagine a photo slowly dissolving into static. A diffusion model learns to rewind that film — step by step turning static back into a coherent image.
Where you see it
- Text-to-image tools like Stable Diffusion.
- Image editing and inpainting.
- Emerging video and audio generators.
How it works
1.Forward noise
Gradually add noise to training examples.
2.Learn to denoise
Network predicts noise or clean signal at each step.
3.Sample
Iterate denoising from random noise to a finished sample.
Why it matters
- Diffusion set the quality bar for controllable generative media.
Often confused
Diffusion models are LLMs.
They are a generative paradigm often used for continuous data (images); LLMs usually model discrete tokens.