DALL-E generates images directly from text captions
A 12-billion-parameter version of GPT-3 learns to draw what it's told
Quick facts
- Institution
- OpenAI
- Parameters
- 12 billion
- Method
- Autoregressive Transformer over text and image tokens
What happened
OpenAI announced DALL-E, a 12-billion-parameter version of the GPT-3 Transformer architecture trained to generate images from text descriptions rather than text from text. Aditya Ramesh and co-authors described the underlying method in 'Zero-Shot Text-to-Image Generation': the model treated both text and image patches as tokens in a single stream, using a discrete variational autoencoder to compress images into a manageable vocabulary of visual tokens the Transformer could then predict autoregressively, one token at a time, conditioned on the caption. DALL-E could generate plausible images for captions describing combinations that had likely never co-occurred in its training data, such as anthropomorphized objects or animals rendered in specific artistic styles, and it could also apply requested edits or transformations to existing images.
Why it matters
DALL-E demonstrated that the same Transformer approach behind large language models could be extended to generate images directly from natural-language descriptions, helping trigger the wave of text-to-image systems, including Stable Diffusion the following year, that made image generation broadly accessible.
How we know
OpenAI's own announcement describes the model's scale and GPT-3 lineage directly; the companion technical paper on arXiv documents the token-based architecture in full detail.
Sources
- OpenAI. DALL-E: Creating Images from Text · Primary source (author-declared)openai.com · Cited as a "primary" source (no stronger domain match).
- Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, et al., OpenAI. Zero-Shot Text-to-Image Generation · Primary source (author-declared)arxiv.org · Cited as a "primary" source (no stronger domain match). · Link is live and its text matches the event's key terms (Jul 2026)
See something wrong? . Corrections with a source get fixed fastest.
Part of a timelineArtificial Intelligence30 events · From a wartime theory of neurons to machines that write, paint, and fold proteinsView all →