DALL·E, also written DALL-E, is a family of text-to-image models developed by OpenAI that generate original images from natural-language descriptions and can edit existing images through inpainting and outpainting. In generative AI, it is one of the systems that made prompt-driven image creation a mainstream capability, alongside Stable Diffusion and Midjourney.
The name combines the painter Salvador Dalí and the robot WALL·E. Three main generations appeared between 2021 and 2023, and each used a different architecture, so the name refers to a product lineage rather than a single technique. Later, OpenAI moved default image generation in ChatGPT to natively multimodal GPT models, while DALL·E remained a reference point for the field.
How DALL·E generates images from text
Every DALL·E version solves the same task: map a text prompt to a set of pixels that matches its content, style and composition. The approach changed substantially between generations, moving from autoregressive token prediction to diffusion.
The table below compares the three generations by architecture and key capability.
| Version | Released | Core architecture | Notable capability |
|---|---|---|---|
| DALL·E | 2021 | Discrete VAE image tokens predicted by a 12-billion-parameter transformer | First broad demonstration of compositional text-to-image |
| DALL·E 2 | 2022 | CLIP embeddings with a diffusion prior and diffusion decoder | Higher realism, variations, inpainting and outpainting |
| DALL·E 3 | 2023 | Diffusion model trained on detailed synthetic captions | Much stronger prompt following and legible short text |
The first model compressed each image into a grid of discrete image tokens with a variational autoencoder and trained a transformer to predict those tokens after the text tokens. DALL·E 2 instead generated a CLIP image embedding from the text and then decoded it into pixels with a diffusion model, an approach described as unCLIP.
DALL·E 3 improved mainly through better training captions. A captioning model rewrote the descriptions of training images into long, precise texts, so the generator learned to respect details such as object counts, positions and colors. At generation time, a language model expands the user’s short request into a detailed internal prompt.
Capabilities and parameters
Through the API and the chat interface, DALL·E exposes a small set of controls that shape the output. Users cannot adjust sampling steps or guidance scale directly, unlike open models where these settings are visible.
| Option | Values | Effect |
|---|---|---|
| Image size in DALL·E 3 | 1024×1024, 1792×1024, 1024×1792 | Square, landscape or portrait framing |
| Image size in DALL·E 2 | 256×256, 512×512, 1024×1024 | Lower cost at smaller sizes |
| Quality | Standard or HD | Finer detail and consistency at higher cost |
| Style | Vivid or natural | Dramatic versus more photographic rendering |
| Edit with mask | DALL·E 2 editing endpoint | Inpainting of a selected region |
Prompt rewriting is an important behavior of DALL·E 3. The submitted prompt is usually revised before generation, and the API returns the revised version. This improves results for brief requests but means that the exact wording is not fully under user control.
DALL·E compared with other image generators
DALL·E differs from its main competitors less in raw image quality than in access model, control and integration. The comparison below reflects the general profile of each system rather than any single release.
| System | Access | Typical strength |
|---|---|---|
| DALL·E | Hosted service through chat and API | Prompt following and conversational refinement |
| Stable Diffusion | Open weights, local or hosted | Deep customization, ControlNet and fine-tuning |
| Midjourney | Hosted service | Distinctive aesthetic and artistic styling |
| Native multimodal GPT models | Hosted, inside the chat model | Image generation with shared conversational context |
A practical consequence is that DALL·E suits teams that want reliable results without managing models, while open-weight diffusion models suit teams that need fine control over composition, custom styles or on-premises deployment.
Writing effective prompts for DALL·E
Because DALL·E 3 was trained on long, descriptive captions, it responds best to prompts that describe the scene the way a caption would: subject, setting, composition, lighting and visual medium. Lists of disconnected keywords, which suit some other generators, tend to give less predictable results.
- Name the medium first, such as watercolor illustration, studio photograph or flat vector icon.
- Specify composition: viewpoint, framing and where the main subject sits in the image.
- Describe lighting and color palette explicitly instead of relying on mood words alone.
- Refine conversationally: ask for one change at a time so the rest of the image stays stable.
- State exclusions plainly, for example no text in the image, when unwanted elements appear.
Iterating in small steps is usually more efficient than rewriting the entire prompt, because each new request is interpreted against the previous description in a chat session.
Safety measures and content provenance
DALL·E operates under usage policies enforced by classifiers on both prompts and outputs. Requests for sexual content, graphic violence or images of real public figures are refused, and the system declines to imitate the style of living artists by name.
Images generated by DALL·E 3 carry C2PA provenance metadata that records their AI origin. This metadata helps platforms label generated media, but it can be removed by screenshots or re-encoding, so it supports transparency rather than guaranteeing detection.
Limitations and common mistakes
Despite strong prompt adherence, DALL·E shares the typical weaknesses of diffusion-based generators. Knowing them helps users write better prompts and check results before use.
| Problem | Cause | Remedy |
|---|---|---|
| Garbled long text | Letters are learned visually, not spelled | Keep text short or add it in an editor |
| Wrong object counts | Weak precise counting | State counts explicitly and verify the result |
| Inconsistent characters across images | Each image is sampled independently | Describe features in detail every time |
| Unexpected details | Prompt rewriting adds content | Check the revised prompt and request fewer additions |
In plain words
DALL·E is a tool that draws pictures from a written description. You type something like a red bicycle leaning against a blue wall at sunset, and it creates a new image that matches.
It learned this by studying a very large number of images paired with descriptions. It does not copy one stored picture; it builds a new image from learned patterns, much as an illustrator draws from experience.
A real-world example
A small educational publisher needs 24 chapter illustrations for a children’s science book within two weeks and a limited budget. The designer uses DALL·E 3 in landscape format at 1792×1024, describing each scene in detail, for example a cross-section of a volcano with labeled layers drawn in a soft watercolor style.
The designer generates several candidates per chapter, selects the best one and adds all labels in layout software, because short generated text is unreliable. Characters are described in the same words every time to keep them recognizable. The team finishes the illustrations in about four working days, and a science editor checks every image for factual accuracy before print.