Diffusion Model
A diffusion model is a class of generative models. Training models a gradual process that adds noise to data and learns its reversal. During generation, the model typically starts with noise and develops a sample through multiple steps. Diffusion models are best known for images but can also be applied to audio, video, and other data types.
Text prompts, reference images, or other conditions can guide generation. Results are stochastic: the same prompt may produce different outputs depending on the initial seed and settings. Quality, speed, and control depend on the model, data, sampler, and compute budget. Generated content is not automatically factual, original, unbiased, or free of rights concerns and requires review before use.
- Generative model class using iterative denoising
- Widely used for image generation
- Also applicable to audio, video, and other data
- Outputs depend on seed and settings
- Review rights, facts, and bias before use
Frequently asked questions
A diffusion model is a generative model that learns to reverse a gradual noising process and use it to generate new data.
No. They are well known for image generators but can also be applied to audio, video, and other data representations.
