Prompt DALL-E

Writing prompts for DALL·E

DALL·E is OpenAI's text-to-image family. DALL·E 1 (2021) was a demo only; DALL·E 2 (April 2022) was the first public release with inpainting and outpainting; DALL·E 3 (October 2023) became integrated with ChatGPT Plus and Microsoft Bing/Copilot. The big jump in v3 is not just quality — it is that DALL·E now inherits GPT-4's language understanding, so it follows complex instructions and renders text inside images at a level that Midjourney and Stable Diffusion still struggle to match.

When you use DALL·E 3 inside ChatGPT, your message goes through a hidden step that rewrites your short prompt into a detailed paragraph. Writing "a cat in a hat" can become 600 characters describing lighting, mood and composition before reaching the model. This is great for casual users but a problem if you want exact control — in which case you should ask ChatGPT to "use my prompt verbatim".

Versions, syntax and parameters

Inside ChatGPT, just describe the scene naturally; expanded prompts cap around ~1000 characters. Via direct API, model dall-e-3 accepts: size (1024x1024, 1792x1024, 1024x1792), quality (standard or hd) and style (vivid for hyperreal/dramatic or natural for muted realism). Pricing runs $0.04-0.12 per image via API; ChatGPT Plus and Bing Image Creator are included in their respective subscriptions, with Bing being a free way to access DALL·E 3.

Strengths and weaknesses

Strengths: best-in-class text rendering inside images (signs, logos, posters), photorealism, and instruction adherence inherited from GPT-4. Great for marketing imagery, social posts, presentation slides, mockups and blog covers. Weaknesses: the safety filter is stricter than Midjourney's or Stable Diffusion's — it refuses public figures' faces, branded logos, violence, sexual content and many copyrighted characters. Outputs are a single image per request rather than the 4-image grid Midjourney returns.

Compared with Midjourney and Stable Diffusion

vs Midjourney: DALL·E 3 wins on literal instruction following and on text inside images; Midjourney wins on artistic, painterly aesthetics. vs Stable Diffusion: DALL·E is closed-weight, paid and runs in the cloud only; SD is open-source, free and runs locally on your GPU. DALL·E gives you zero fine-tuning knobs; SD gives you LoRAs, ControlNets and full sampler control.

Editing and inpainting

DALL·E 2 historically exposed inpainting (replace a masked region) and outpainting (extend the canvas), but these were deprecated. DALL·E 3 does not publicly expose inpainting — if you want to change part of an image, you either describe the new version and regenerate, or send it back to ChatGPT asking for an edit (which still regenerates from scratch).

FAQ

Can I sell images generated by DALL·E commercially? Yes. OpenAI's Terms grant you full ownership and commercial-use rights to images you generate, whether through the API, ChatGPT or Bing Image Creator.

Is a longer, more detailed prompt better? Not necessarily. DALL·E 3 already enhances your prompt internally — overdoing detail can compete with that enhancement and produce worse results. Be specific about what matters and let the model handle the rest.

Can I ask DALL·E to edit a specific part of the image? Not really — the public surface does not expose inpainting. Each "edit" generates a brand new image, which means consistency between iterations is hard to keep.

Why does it refuse my prompt? Likely a filter trigger: real people's faces, copyrighted characters, brand logos, violence or sexual content. Rephrase to a generic version (e.g. "a businessman in a blue suit" instead of a named CEO) and try again.