Producing pictures from a description. Quality reached commercial usability quickly, and the difficult questions turned out to be legal and social rather than technical: what the training data was, whose styles are being reproduced, and what happens to the people whose work it replaces.
Technical view
Dominated by latent diffusion with text conditioning through cross-attention, increasingly with transformer backbones and flow-matching objectives. Control mechanisms include guidance scale, negative prompts, image-to-image initialisation, inpainting and structural conditioning. Persistent weaknesses are text rendering, precise counting and compositional relations between multiple objects. Commercial deployment centres on licensing of training data and provenance marking of outputs.
Related
Part of AI Terminology Index, section: Modalities and Applications.