DALL-E 2
By OpenAI
DALL-E 2 is a text-to-image generation model from OpenAI that creates original images from natural-language prompts using a diffusion process guided by CLIP-derived embeddings, and it was the version that first made OpenAI's image…
Definition
DALL-E 2 is a text-to-image generation model from OpenAI that creates original images from natural-language prompts using a diffusion process guided by CLIP-derived embeddings, and it was the version that first made OpenAI's image generation broadly available through a standalone web app and API. It introduced mask-based inpainting and outpainting, letting users edit regions of an existing image or extend it beyond its original frame using a new prompt.
Overview
DALL-E 2 is a text-to-image generation model from OpenAI that solves the problem of producing an original image from a written description, succeeding the original DALL-E with substantially higher visual fidelity and more accurate adherence to prompt details such as object relationships and style cues. It was also OpenAI's first image model to reach a wide audience through a dedicated web interface and a public API rather than remaining a research-only system. Under the hood, DALL-E 2 combines a CLIP text encoder with a diffusion-based image decoder: a prior network maps the CLIP text embedding to a corresponding CLIP image embedding, and a diffusion model then generates an image conditioned on that image embedding, upsampling through additional diffusion stages to reach final resolution. This two-stage, embedding-mediated design is what let it separate "what CLIP thinks the image should represent" from the literal pixel synthesis, an architecture-level detail not shared by every later diffusion model. Within the text-to-image space, DALL-E 2 sits between the original token-based DALL-E and later, more capable systems such as DALL-E 3 and Midjourney's newer versions; its CLIP-guided diffusion approach differs from Stable Diffusion's latent diffusion trained on a fixed autoencoder and from Muse's masked-token prediction. It was notable at release for introducing inpainting and outpainting features that let users edit specific regions of an existing image using a mask and a new prompt. In practice, DALL-E 2 has been used for illustration, concept art, marketing imagery, and rapid visual prototyping, accessed through OpenAI's own interface, its API for third-party integrations, and for a period through Microsoft Bing's image tools before newer models superseded it. Its editing features made it a common choice for iterative image refinement rather than one-shot generation alone. Its limitations included difficulty rendering legible text within images, inconsistent handling of compositional prompts involving multiple objects with specific spatial relationships, and occasional artifacts in hands, faces, or fine detail typical of diffusion models of its era. As OpenAI released DALL-E 3 with improved prompt following from richer training captions, DALL-E 2 became the choice mainly for lower-cost generation or workflows already built around its specific API and editing endpoints rather than for state-of-the-art fidelity. Its inpainting and outpainting endpoints also remained useful in their own right for a time, since not every successor model exposed an equivalent mask-based editing interface through a simple API call, making DALL-E 2 a practical fallback for developers whose tooling depended specifically on that editing behavior even after newer models surpassed it on raw image quality.
Key Features
- Generates images from text prompts using a CLIP-guided diffusion pipeline
- Supports in-painting to edit specific regions of an image
- Supports outpainting to extend an image beyond its original frame
- Produces multiple image variations from a single reference image
- Available through both a standalone app and the OpenAI API
- Applies content filters restricting sensitive and photorealistic misuse
- Improved photorealism and detail over the original DALL-E model