DALL-E

The first DALL-E research studies text-to-image generation through autoregressive modeling of text and discrete image tokens. A discrete variational autoencoder supplies an image-token representation, allowing the transformer to model a joint sequence rather than directly predicting each raw pixel value. This architecture should not be described as the same as every later DALL-E model. The DALL-E 2 research, Hierarchical Text-Conditional Image Generation with CLIP Latents, uses a prior that produces CLIP image embeddings from text and a decoder that generates an image conditioned on that representation. The studied decoder uses diffusion, while the prior and decoder have different functions. DALL-E 3’s captioning report investigates training a bespoke image captioner, recaptioning the training images with descriptive synthetic text, and measuring resulting prompt-following behavior. That report states that it does not cover the model’s full implementation details, so a high-level lineage should not invent undisclosed architecture from a name alone. The historical launch integration also used ChatGPT to refine an idea into a detailed generation prompt. Runtime prompt refinement is distinct from replacing training captions. Historical editing descriptions include DALL-E 2’s selected-region image changes; these should not be generalized to every version or API operation. Current availability is another separate fact. OpenAI’s API deprecation record lists May 12, 2026 as the removal date for dall-e-2 and dall-e-3 snapshots. A surviving launch article, research paper, or old model page does not establish that those API models remain callable today. Their research can still be studied as historical systems.