Generative modeling creates new text, images, audio, or video that resembles patterns learned from real-world data. Unlike systems that classify, predict, or retrieve, it produces novel outputs from learned relationships. section 1.0
Generative systems appear through several connected approaches and applications:
- Transformer architectures apply attention mechanisms, originally developed for natural-language processing, to tasks including image processing. section 1.0
- Diffusion models generate images by adding random noise and progressively refining it according to patterns learned during training. section 1.0 section 1.5.1
- Text-to-image systems translate natural-language prompts into detailed images and can edit existing photographs by removing objects, changing backgrounds, or cropping. section 1.0
A prompt becomes an image
DALL-E 2, Imagen, and Stable Diffusion demonstrate text-to-image generation: a written description becomes a visually detailed output. Related uses include design visualization, photo editing, medical diagrams, and generated essays or code. section 1.0
Generative modeling also raises questions about originality and copyright. Outputs may combine learned relationships into new concepts, yet some systems can produce elements highly similar to artworks in their training data. section 1.5.1
