DALL-E 1, stylised as DALL·E, is the first version of a text-to-image model developed by OpenAI. Announced in January 2021, it uses deep learning methodologies to generate digital images from natural language descriptions, known as prompts. The name is a portmanteau of the Pixar robot character WALL-E and the Spanish surrealist artist Salvador DalÃ.
DALL-E 1 was a milestone in generative artificial intelligence, demonstrating that a Transformer-based model could produce coherent and diverse images from textual descriptions. It was succeeded by DALL-E 2 in 2022 and DALL-E 3 in 2023, each improving on realism, resolution, and prompt understanding.
History and background
OpenAI revealed DALL-E 1 in a blog post on 5 January 2021. It used a modified version of GPT-3, a large language model with 175 billion parameters, to generate images. The model was developed alongside CLIP (Contrastive Language-Image Pre-training), a separate model trained on 400 million image-text pairs scraped from the Internet. CLIP's role was to rank DALL-E's outputs by predicting which caption from a list of 32,768 randomly selected captions was most appropriate for an image, helping filter the best results.
The release of DALL-E 1 generated significant public interest and discussion about the capabilities and implications of generative AI. It was not immediately made widely available; access was initially restricted for research preview due to ethical and safety concerns. The successor DALL-E 2 entered beta in July 2022 and was opened to everyone in September 2022.
Technology
DALL-E 1 had three components: a discrete variational autoencoder (VAE), an autoregressive decoder-only Transformer model with 12 billion parameters, and a CLIP pair of image and text encoders. The discrete VAE converted images into sequences of tokens and back, as the Transformer did not process image data directly.
The input to the Transformer was a sequence of tokenised image captions followed by tokenised image patches. Captions were in English, tokenised by byte pair encoding with a vocabulary size of 16,384, and could be up to 256 tokens long. Each image was a 256×256 RGB image divided into 32×32 patches of 4×4 pixels. Each patch was converted by the discrete VAE to a token from a vocabulary of 8,192.
This architecture was similar to that of GPT-3 but adapted for image generation. The autoregressive model predicted image tokens sequentially, conditioned on the text caption. CLIP, based on contrastive learning, was used to rank and select the best image from a larger set generated by the model.
Capabilities
DALL-E 1 could generate imagery in multiple styles, including photorealistic images, paintings, and emoji. It could "manipulate and rearrange" objects in its images and correctly place design elements in novel compositions without explicit instruction. For example, when asked to draw a daikon radish blowing its nose, sipping a latte, or riding a unicycle, it often drew the handkerchief, hands, and feet in plausible locations.
The model showed an ability to "fill in the blanks," inferring appropriate details without specific prompts, such as adding Christmas imagery to prompts commonly associated with the celebration and placing shadows appropriately. It exhibited a broad understanding of visual and design trends.
DALL-E 1 could produce images for a wide variety of arbitrary descriptions from various viewpoints, with only rare failures. Mark Riedl, an associate professor at the Georgia Tech School of Interactive Computing, found that it could blend concepts, a key element of human creativity. Its visual reasoning ability was sufficient to solve Raven's Matrices, visual tests often used to measure human intelligence.
Impact and legacy
DALL-E 1 helped popularise text-to-image generation and spurred further research in generative AI. It was followed by DALL-E 2, which used a diffusion model with 3.5 billion parameters, and DALL-E 3, which was integrated into ChatGPT in October 2023. Microsoft implemented DALL-E 3 in Bing's Image Creator and Designer app, and Copilot runs on it. In March 2025, DALL-E 3 was replaced in ChatGPT by GPT Image's native image-generation capabilities.
OpenAI began adding watermarks to DALL-E generated images in February 2024, using metadata in the C2PA (Coalition for Content Provenance and Authenticity) standard. DALL-E 1's release marked a significant step in the development of artificial intelligence and machine learning, influencing subsequent models and applications.