DALL-E is OpenAI’s original text-to-image model, introduced in January 2021 as a 12-billion-parameter variant of GPT-3 trained to generate images from text captions. It worked by treating image generation as a sequence prediction problem: a discrete variational autoencoder converted images into tokens, and an autoregressive transformer learned to predict those tokens from a text prompt, with a CLIP model used afterward to rank the resulting images. It was trained on around 250 million text-image pairs scraped from the internet and could produce plausible images for combinations of concepts it had never seen paired together, like an armchair shaped like an avocado. DALL-E was never released as a public product; OpenAI kept it as a research preview and later replaced it with the diffusion-based DALL-E 2, which produced sharper images at a fraction of the parameter count.