DALL·E

DALL·E, presented in January 2021, was the model that put text-to-image generation on the map. OpenAI describes it plainly: a 12-billion-parameter version of GPT-3, trained on a dataset of text–image pairs to generate pictures from written descriptions. The architectural point mattered as much as the pictures — this was a language model, the same kind of network that writes text, made to emit an image instead. The capabilities OpenAI demonstrated are the ones every later generator is still judged on: controlling attributes of an object, drawing several objects in the right relation to one another, showing perspective and three-dimensionality, rendering an object's internal structure, inferring details the prompt never stated, combining unrelated concepts, and applying transformations to an existing image. The company also showed zero-shot visual reasoning and geographic and temporal knowledge — the model could be asked for a phone from a given decade. What DALL·E never had was customers. It was a research demonstration: no API, no product, no price list. The public only saw the examples OpenAI chose to publish, at a resolution its successor would multiply by four. The mass audience arrived a year later with DALL·E 2, and image generation reached the wider public only through that model and ChatGPT. The profile is kept as a historical entry. The model has no availability of any kind today; the current capability lives in the GPT Image 2 profile.

▸Videos

No videos yet.

▸Reviews