OpenAI announces DALL·E 3 and its synthetic-caption training method
OpenAI announces DALL·E 3; the accompanying paper attributes much of its stronger prompt following to recaptioning training images with detailed synthetic descriptions.
What happened
Event details
OpenAI announced DALL·E 3 and said it would work with ChatGPT, which could turn a user request into a more detailed image description. The accompanying paper placed the central intervention in the training data: OpenAI trained a dedicated image captioner, used it to produce detailed synthetic descriptions, and trained the image generator on those recaptioned examples. The disclosed experimental stack used a T5-XXL text encoder and a text-conditioned U-Net latent-diffusion model. DALL·E 3 also used a convolutional U-Net diffusion decoder distilled to two denoising steps. OpenAI reported DrawBench short- and long-prompt correctness of 70.4% and 81.0%, compared with 49.0% and 52.4% for DALL·E 2. These are results under the paper's evaluation setup, not a guarantee for every real-world prompt. September 20 was the announcement, not a same-day general release. OpenAI made DALL·E 3 available to ChatGPT Plus and Enterprise users on October 19. It did not disclose the model's parameter count, training-set size or text context length.
Assessment
Why it matters
DALL·E 3 used recaptioned training data to improve complex prompt following and made language-model-assisted prompt rewriting a mainstream text-to-image interaction.
Availability
Access notes
September 20, 2023 was the announcement date; initial ChatGPT Plus and Enterprise availability began on October 19, 2023.