Diffusion Model
Model Difusi
The architecture behind most image and video generators.
A diffusion model generates by starting with random noise and repeatedly removing it, guided by the prompt, until the noise resolves into an image. Training teaches the model to predict what noise was added to a real image, so it learns the reverse operation.
It works well for images because images are high-dimensional and a pixel-level process captures fine detail. The cost is speed: the denoising happens in many steps, and video multiplies that by every frame.
Text-to-image models also inherit the biases of their training data, more visibly than language models do, because the visual details of a concept are what the model learned. A model asked for a doctor produces a specific kind of doctor, and the model has no way to represent the alternative.
Related terms
Generative AI
AI that produces new content: text, images, audio, code, video.
Multimodal
A model that handles more than one kind of input or output: text, image, audio, video.
Dataset
A collection of examples used to train, evaluate, or ground a system.
IP
Who owns what the model produced, and what it was trained on.
Bias
A systematic skew in what a system produces, inherited from the data it learned on.
Want this working inside your team?
It starts with one conversation about your workflow.