Diffusion Model
Diffusion Model
An AI technique that generates new data by gradually removing noise
In Simple Terms
A diffusion model is an AI system that learns to reconstruct clean data from noise. During training, the AI takes a clean image and gradually adds random noise — like TV static — until the original is completely unrecognizable. It then practices the reverse: starting from that noisy state and removing noise step by step to restore the clean image. By mastering this denoising process, the model can generate brand-new images from pure random noise that match a given condition, making it a core technology behind high-quality AI image generation services.
Behind the Name
The English name for this technology is "Diffusion Model." Diffusion refers to the physical phenomenon where ink dropped into a glass of water gradually spreads until it blends throughout. The model is named after this idea — because it applies the same concept to data, gradually spreading noise through it just as ink disperses in water.
Take a Closer Look!
A diffusion model is an AI system that learns to reconstruct clean data by first corrupting it with noise, then reversing that process. It plays a central role in image generation and is one of the key technologies powering AI image generators.
The model operates in two phases: a forward process that adds noise, and a reverse process that removes it.
During training, the AI is shown a clean image, and random noise is added to it little by little — like static gradually overtaking a picture — until the original content is completely unrecognizable.
The AI is then trained to run the process in reverse: starting from the fully noisy state, it learns to remove noise step by step and restore the original clean image.
By learning the rules that guide noisy data back toward natural-looking data, the model becomes capable of generating new images that match a given condition — starting from nothing but random noise.
Compared to earlier image generation techniques, diffusion models produce remarkably high-quality, photorealistic results with far greater expressive range.
Many of the widely-used high-quality image generation services today are built on this technology.