Text-to-image generation is the capability of producing an original picture from a written prompt: type “a watercolor fox reading a newspaper in a cafe” and the system creates an image matching that description. It is the capability readers name, distinct from the underlying method. Where the diffusion-models entry explains how today’s tools generate images, this entry covers the category, the user-facing ability that products like DALL-E, Midjourney, and Stable Diffusion all deliver, regardless of the technique inside.
Two papers anchor the modern category. OpenAI’s 2021 “Zero-Shot Text-to-Image Generation” (Ramesh et al., arXiv 2102.12092), the work behind the first DALL-E, showed that with enough data and scale a single model could generate plausible images for arbitrary text prompts; its abstract reports the approach “is competitive with previous domain-specific models when evaluated in a zero-shot fashion.” The 2021 “High-Resolution Image Synthesis with Latent Diffusion Models” (Rombach et al., arXiv 2112.10752), the basis of Stable Diffusion, then made high-quality generation efficient enough to run on ordinary hardware, which is what turned text-to-image from a lab demo into a mass-market capability through 2022.
The mechanics matter less to most users than the shift it represents: producing professional-looking visual media became a matter of describing what you want in words. The same capability has since extended into text-to-video (systems like Sora) and other media, making “describe it and the model renders it” a general pattern rather than an image-only trick.
Why business readers should care: text-to-image collapsed the cost and time of producing custom visuals for marketing, design, and prototyping, while raising hard questions the technology does not resolve. Training data and copyright are contested, the same tools enable convincing fakes and misuse, and outputs can reproduce biases present in their training images. Treat it as a fast, cheap drafting and ideation capability whose provenance, licensing, and authenticity still require human judgment.