DALL-E 2 (branded DALL·E 2) is a text-to-image generation model announced by OpenAI in April 2022 as the successor to the original DALL-E. It produced 1024x1024 images, four times the resolution of its predecessor, and its photorealistic outputs and editing features made it one of the most visible AI systems of 2022, ahead of the chatbot wave started by ChatGPT later that year.12
Architecture
DALL-E 2 is described in the paper "Hierarchical Text-Conditional Image Generation with CLIP Latents," which calls the approach unCLIP.1 Instead of the autoregressive Transformer over image tokens used by the first DALL-E (a 12-billion-parameter derivative of GPT-3), DALL-E 2 works in two stages. A prior network maps a text caption to a corresponding CLIP image embedding, and a diffusion decoder then generates an image conditioned on that embedding; additional diffusion upsamplers raise the output to 1024x1024. The decoder is developer-reported at roughly 3.5 billion parameters. Building on CLIP's joint text-image space gave the model strong semantic control and enabled "variations" of an input image, as well as inpainting and later outpainting for editing and extending pictures.
Release and access
OpenAI announced DALL-E 2 on April 6, 2022, as a limited research preview with a waitlist, citing risks such as misuse for disinformation and the generation of harmful content.2 Access widened through a paid-credit beta in July 2022, the waitlist was removed in September 2022, and an image-generation API entered public beta in November 2022.3 Deployment mitigations included training-data filtering to reduce explicit content, prompt and output filters, and initial restrictions on generating photorealistic faces of real people.
Reception and legacy
DALL-E 2 arrived months before Imagen from Google and Stable Diffusion from Stability AI, and together these systems set off the 2022 boom in generative image tools that later produced open-weight successors such as FLUX.1 Its diffusion-based design marked the field's shift away from autoregressive image generation, a direction that also informed OpenAI's later video model Sora. OpenAI superseded it with DALL-E 3 in 2023, which was integrated into ChatGPT alongside models such as GPT-4o, and DALL-E 2 was subsequently wound down as an offering.