What Is DALL-E? How It Worked and What Replaced It in 2026

DALL-E was OpenAI’s text-to-image model, designed to turn natural-language prompts into original images. The model helped push generative image tools into the mainstream, but it has since become a legacy product, superseded by newer GPT-based image generation.

The shift correlates with how quickly small businesses are adopting AI. In a 2025 Shopify survey, 75% of store owners said they used AI tools.* 

Anyone arriving at DALL-E today may be seeking one of two things: an explanation of what it was, or the newer tools that do image-generation now.

What is DALL-E?

DALL-E was OpenAI’s generative AI model for creating images from natural-language written prompts. As opposed to retrieving an image, the program generated a new one based on instructions about the subject, the setting, or the composition.

DALL-E sat within the broader generative AI category alongside tools that generated text, code, audio, and video. It’s an investment that has captivated small business owners. The International Data Corporation reported in June 2026 a shift toward generative AI as the top forward-looking technology priority, overtaking traditional AI and process automation.

For the businesses that adopted it, the appeal came down to a handful of practical reasons:

• Fast drafts and iteration. DALL-E could turn a written prompt into a visual direction, and generate another version when the prompt changed. 
• More room to experiment. You could explore visual directions before committing to a finished production. Gartner found 77% of marketing organizations using generative AI had adopted it for creative development tasks.
• Lower technical barrier. You could describe what you wanted in everyday language instead of constructing the image manually in design software. No design experience required.

How did DALL-E work?

The original DALL-E, released in 2021, worked as an autoregressive transformer, the same supporting architecture family as early GPT models. OpenAI described it as a decoder-only transformer that treated text and images as one sequence of tokens, much like how a large language model works. Tokens are used by AI models to read, process, and generate information. 

The later DALL-E versions changed the latent approach. DALL-E 2 used diffusion, starting from noise and progressively refining it into an image that matched the prompt. This new model first converted the text into a CLIP-based (Contrastive Language–Image Pre-training) representation, and used a diffusion decoder to turn that representation into an image.

DALL-E 3 also used a latent diffusion model, but improved prompt following by training on much more detailed image captions. OpenAI said this helped the model understand the objects, relationships, and details specified in a prompt.

DALL-E 1 vs. DALL-E 2 vs. DALL-E 3

DALL-E’s image quality and control improved with each generation, but the biggest enhancement was the way in which the models followed prompts:

DALL-E 3 was the last model in the family. OpenAI has since deprecated and removed it from the API, recommending GPT-Image-2 for current image generation and editing. The company retired the official DALL-E GPT in ChatGPT on August 30, 2026.

For anyone still searching for DALL-E today, ChatGPT 2.0 Images is the latest successor as of September 2026.

How do you create AI-generated images after DALL-E’s retirement?

You can still do the things people used DALL-E for, like generate an image from a prompt or edit an existing image, but the tools now live elsewhere.

• ChatGPT. ChatGPT Images 2.0 is OpenAI’s current image-generation experience. You can create from a prompt, upload an existing image for editing, change the aspect ratio, and ask for details such as text or a transparent background.
• Your Shopify admin. If you’re on Shopify, you can generate images without leaving Shopify. Sidekick accepts natural-language prompts, can work from uploaded reference images, and lets you refine an image with follow-up instructions before saving it to your store.
• Microsoft Copilot. Copilot lets you generate images from text prompts or upload an existing image and describe how it should change.
• OpenAI API. If you’re a developer, you can use GPT-Image-2 for programmatic image generation and editing.

How to create your first image in ChatGPT Images 2.0

Here’s how to create an image in ChatGPT Images:

1. Open Images in ChatGPT

Select Images from the sidebar. 

You can enter your instructions as typed text or a voice prompt, or attach an existing image or other file to give ChatGPT more context for what you want to create. OpenAI also lets you generate an image from an ordinary ChatGPT conversation rather than opening Images first.

2. Describe what you want

Start with the image itself, and add any reference material that matters. In the example below, a thrift store’s Instagram page is uploaded as visual context, along with instructions to create a fall-collection promotional poster and keep the existing logo prominent.

3. Generate and review the first draft

ChatGPT creates the image from those instructions and automatically saves generated images under Library so you can return to them later.

4. Edit the first draft

Open the generated image and use the Describe edits field to request changes in plain language. 

For more targeted adjustments, the image editor also includes controls to remove the background, erase part of the image, or resize it without generating a completely new version.

5. Share or download the finished image

Open the share menu to copy a link or send the image directly to supported platforms such as X, LinkedIn, or Reddit. Select Download to save the final image file for use elsewhere.

How businesses use AI image generators like DALL-E

Here are some business use cases for AI image generation:

Content creation and design

The 2025 Shopify survey of store owners found the most common AI use case was creating content, at 69%.* AI image generators let you create early concepts, promotional graphics, backgrounds, and other visual assets.

A growing number of retailers expect to handle more of that work themselves. According to Deloitte’s 2026 Retail Industry Global Outlook report, 94% of retail executives surveyed expect to bring more marketing activities in-house as AI tools enable work such as content generation and creative automation. 

If you’re on Shopify, you can do this directly in Sidekick, which generates images from natural-language prompts, accepts reference photos, and supports follow-up instructions for refining a result.

Children’s apparel brand Brave Little Ones, for example, uses Sidekick to quickly turn creative ideas into storefront changes. The brand takes advantage of Shopify’s built-in AI to modify the site ahead of big sales. 

“Sidekick allows us to come up with a concept and then immediately create a section and throw it on the website and see if it worked or not,” says Cofounder Jon Ezell.

Product prototyping

Before paying for a sample, you can use AI image generators to see what an idea might look like. DALL-E could turn a written product concept into visual mockups, making it easier to compare shapes, colors, finishes, or packaging directions during the early stages of product prototyping.

That workflow has continued with newer generative AI tools.

For example, The New Black AI, found in the Shopify App Store, lets brands generate garment concepts, preview them on virtual models, and produce technical packs before pushing finished visuals to Shopify.

Creative storytelling

AI image generators can also help you put products into a story instead of simply showing them against a blank background. A product photo can become a holiday scene or a visual concept built around a specific mood or moment.

Sustainable activewear brand Reprise, for example, has used Shopify Magic to give product photos a festive treatment during the holiday season.

“When you cater to those different holidays, people pay attention because it’s top of mind,” says Mary.

Concept art

AI image generators can help turn a loose visual direction into something concrete. They can be used to sketch out moodboards, environments, packaging directions, and campaign concepts that could then be handed to designers or creative teams for development.

That exploratory workflow also shows up in Shopify tools. Tinker, Shopify’s free AI creative app, brings together more than 100 specialized AI tools for creating images, videos, logos, and product photography. 

“Tinker’s image just always comes out the best. It always takes my feedback. I normally get the picture I want in the first, if not second, try,” says Lena, cofounder of jewelry brand Loire.

Fashion design

AI image generators can help fashion designers turn an idea into something they can respond to. The description of a silhouette, fabric, color, or detailing can generate visual variations before a designer commits to a sample, which can help narrow down which directions would be worthwhile to explore further.

The value is in the speed of exploration. A designer might test the same dress in different fabrics, or compare colorways without sketching every variation manually.

The limitations of AI image generators like DALL-E

For retailers, AI image generation comes with a few trade-offs:

• The generated images still need fact-checking. AI can alter product colors, proportions, packaging, or other details while producing something that looks perfectly plausible. OpenAI’s current terms require you to evaluate outputs for accuracy before sharing them.
• The bias can show up in the people it generates. A 2025 Scientific Reports study generated 320,000 images across 32 professions with Stable Diffusion XL. Men made up at least 90% of images for 20 professions, while white people were the most frequently generated group in 24 professions. For businesses creating lifestyle imagery, representation needs a human check, too.
• The customers may not like your synthetic imagery. A 2025 International Journal of Information Management study found consumers were more likely to avoid services advertised with AI-generated rather than real images; respondents also perceived the AI imagery as less trustworthy. 
• The use of commercial images and copyright laws needs separate evaluation. You own your output, to the extent permitted by law, under OpenAI’s Terms of Use, but OpenAI also notes that output may not be unique. Shopify, too, doesn’t claim ownership of images created with its media-generation tools and permits retailers to use them for business purposes inside or outside Shopify. Shopify-generated images carry an invisible watermark that doesn’t restrict commercial use.

As for copyright protection, the US Copyright Office concluded in 2025 that AI-generated material can be protected where there’s sufficient human authorship, but prompts alone generally aren’t enough.

Note: This information is for general educational purposes; copyright, advertising, privacy, and AI-disclosure rules vary by jurisdiction and use case. Read more about how to use AI for small businesses.

The top tips for using technology like DALL-E

Many of the prompting habits that worked well with DALL-E are still relevant to newer image generators, like ChatGPT Images and Shopify Sidekick.

Provide clear and detailed descriptions

Include details such as the subject, setting, lighting, camera angle, colors, or materials when they’re relevant.

For ecommerce imagery, it also helps to specify where the image will appear. A homepage hero image may need negative space for copy, while a product image may need a plain background and an unobstructed view of the item.

Example prompt: “Generate a hero banner image for my summer sale featuring a beach scene with sunglasses and towels, in a bright and vibrant style.”

Experiment with different prompts and styles

If you try different ways of describing the image you want, you may get a variety of results. This can help you learn what details matter to the AI image generator you use and refine your prompts over time. Be intentful, and get creative if something doesn’t feel right. Test variables such as the camera angle, lighting, background, or visual treatment and compare the results.

Create different iterations of an image

If an AI image generator gets the general idea right but misses a detail, one way to fix it might be to keep the good parts while changing the instruction. You could, for instance, ask for more realistic lighting and texture if an image appears too cartoonish. A wider crop, a different camera angle, or more space around the subject might solve a composition problem. 

Current image tools make editing easy because you can use conversational prompts or select specific parts of an image.

Curate and filter the output

Check the result for composition, product accuracy, awkward text, inconsistent branding, and unwanted details the model may have added. For ecommerce imagery, compare generated products against the real item before anything reaches a product page or ad.

Provide context and feedback

Tell the model what the image is for, who will see it, and which elements need to stay fixed. A request for an ecommerce website, for example, might specify the page layout and where headline copy will sit. An email newsletter, on the other hand, may need a different crop altogether.

When revising an image, be specific about what worked and what didn’t.

Use AI where the image can afford to change

The more a customer needs an image to document the real product, the less room there is for AI to improvise.

That makes generative tools a natural fit for backgrounds, campaign concepts, scene changes, and early creative exploration. Shopify’s own media-generation tools are built around jobs like replacing backgrounds, changing lighting, and extending an existing image.

For product-defining details, keep a closer eye on the source image. Even OpenAI’s newer image models, which improved considerably on DALL-E, still have limitations. Currently, OpenAI says its editing tool is imperfect, and an edit can extend beyond the area selected.

*Based on a 2025 survey of 500 Shopify merchants conducted in English across Australia, Canada, the United Kingdom, Ireland, New Zealand, and the United States. Respondents were established merchants with two or more years on the platform. Results reflect the experiences of this specific sample and may not be representative of all merchants.