DALL·E is an early example of a neural network that turns ordinary written prompts into pictures. Give it a sentence, or a sentence plus a reference image or partial image, and it will produce an original image that matches the description — often with surprising creativity.
How it works
The original DALL·E was built by OpenAI as a text-to-image version of a transformer language model. It used a large autoregressive architecture — a 12-billion-parameter model related to the family of GPT models — and was trained on many pairs of text and images. The model treats text and image information as a single stream of tokens (the implementation worked with long token sequences) and learns to predict the next token, which allows it to generate coherent image outputs from text inputs.
Because it was trained on paired text and images, DALL·E could do more than generate an image from a caption. It could accept both text and an image prompt together, and it supported a simple form of inpainting: the model could regenerate a rectangular region of an input image (the academic implementation had restrictions such as regions that extended to the bottom-right corner) so the result would match the combined text-and-image prompt.
The name is a playful mash-up of Salvador Dalí and the Pixar robot WALL·E, which hints at the blend of imagination and automation behind the idea.
What it did well — and where it struggled
The striking thing about DALL·E was its creativity. Given whimsical prompts like “an armchair in the shape of an avocado,” it produced a set of distinct, photorealistic and stylised images that matched the idea. Prompts that combined unexpected concepts — animals wearing clothes, furniture shaped like food, or objects with impossible materials — often produced charming and sometimes uncanny results.
That said, the model had clear limitations. It handled single objects and simple spatial descriptions reasonably well, but its performance declined as prompts demanded precise placement, multiple distinct objects, or complex relationships between them. Longer, more detailed instructions could confuse the model, leading to odd arrangements or errors in counting and orientation. These were not flaws in imagination so much as challenges in reliably mapping precise, structured instructions into pixels.
There were also broader questions: how such tools affect creative work, how they should be used and credited, and how training data should be handled. Those concerns were part of the conversation around DALL·E as much as the technical results.
Practical note
DALL·E is best thought of as a demonstration of what text-to-image models can do — creative, sometimes uncanny image generation from plain language. If you want to experiment with this kind of tool, the useful takeaways are the same now as then: write clear, concise prompts; try a few variations; and be aware that the model may struggle with precise multi-object scenes. Where to try such systems depends on current offerings from research labs and commercial services.
