OpenAI introduces the 12B text-to-image model DALL·E
OpenAI introduces DALL·E, a 12B autoregressive Transformer that models text and discrete image tokens as one sequence.
What happened
Event details
OpenAI presented DALL·E as a 12B text-conditional image model. It joins up to 256 text tokens with 1,024 discrete image tokens and models the 1,280-token stream autoregressively with a 64-layer decoder-only Transformer. OpenAI's examples used CLIP to rerank 512 candidates and display the top 32 for each prompt. This was a research disclosure, not a product launch: no model weights, external demo or API were released. OpenAI also documented important weaknesses. Attribute binding deteriorated quickly as more objects were added, and semantically equivalent prompt rewrites could produce sharply different results.
Assessment
Why it matters
DALL·E extended large-scale autoregressive sequence modeling to text-conditioned image generation and helped start the modern wave of generative visual models, while its first release remained non-public and visibly brittle.
Availability
Access notes
OpenAI published research details and selected examples, but did not release weights, an external demo or an API.