Diffusion models are a class of generative AI that create images by learning to reverse a gradual noising process. During training, the model sees images progressively corrupted with Gaussian noise over many steps until they become pure static. It then learns to predict and remove the noise at each step. At generation time, the model starts from random noise and iteratively refines it into a clean, coherent image.

Stable Diffusion (Stability AI) and DALL-E (OpenAI) made this approach mainstream for text-to-image generation. They work in a compressed latent space rather than pixel space, making generation faster and more memory-efficient. ControlNet adds spatial conditioning — you can guide generation with edge maps, depth maps, pose skeletons, or segmentation masks, giving precise control over the output composition. Imagen and SDXL pushed image quality further with larger models and better text encoders.

For computer vision practitioners, diffusion models matter beyond art generation. They're used for synthetic training data generation (creating labeled images for rare scenarios), data augmentation (generating variations of existing training samples), super-resolution (enhancing low-resolution images), image inpainting (filling masked regions), and domain adaptation (translating images between visual styles while preserving content).