OpenAI unveiled DALL-E, a 12B-parameter autoregressive Transformer that modeled up to 256 BPE text tokens followed by 1024 discrete image tokens from a dVAE, demonstrating zero-shot compositional text-to-image generation.
DALL-E & GPT Image
Also known as: DALL-E · DALL-E 2 · DALL-E 3 · GPT Image · gpt-image-1
Plain English
OpenAI's pioneering text-to-image model lineage—beginning with the original DALL-E in 2021 and evolving through DALL-E 2 and DALL-E 3 into native multimodal image generation inside GPT-4o (gpt-image-1).
Technical Overview
The lineage traces three distinct architectural eras: DALL-E 1 modeled discrete dVAE visual tokens autoregressively with a 12B Transformer; DALL-E 2 (unCLIP) paired a CLIP latent prior with cascaded diffusion decoders; DALL-E 3 introduced dense synthetic recaptioning; and GPT-4o Image Generation (gpt-image-1) unified visual synthesis directly inside an omni-modal multimodal backbone.
Timeline (4)
2021
Release2022
PaperOpenAI introduced DALL-E 2 (unCLIP), replacing autoregressive token generation with a prior that mapped text captions to CLIP image embeddings and a cascaded diffusion decoder that synthesized 1024x1024 images and semantic variations.
2023
ReleaseOpenAI announced DALL-E 3 with native ChatGPT prompt expansion, showing that re-captioning training images with detailed synthetic descriptions fixed most spatial binding and attribute-mixing failures in earlier text-to-image diffusion models.
2025
ReleaseOpenAI rolled out native GPT-4o image generation in ChatGPT (and in the API on April 23, 2025 as
gpt-image-1), superseding standalone DALL-E 3 by uniting world knowledge, multi-turn conversational editing, and accurate multi-line text rendering inside an omni-modal multimodal model.
See also (10)
Open full graphDALL-E & GPT Image