# How do Midjourney, DALL·E 2 and Stable Diffusion compare?

Date: 2023-09-10 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/midjourney-dalle-2-stable-diffusion-compare/

**Answer:** Midjourney gives the most polished, stylized images with the least effort, through a paid Discord subscription. DALL·E 2 is the simplest to use in a web page and sells credits, while Stable Diffusion, now in its SDXL 1.0 version, is an open model you can run for free on your own computer, with by far the most control and the steepest learning curve.

All three tools do the same basic job: you describe a picture in words, and a model builds it, usually by starting from random noise and refining it step by step until it matches your description. Where they differ is in how you reach them, what they cost, how much control they give you, and what you are allowed to do with the results. Those differences matter more to a beginner than small gaps in image quality.

## The three tools in brief

**Midjourney**, a San Francisco company led by David Holz, is used through Discord, the chat app. You join its server, type `/imagine` followed by your prompt, and the bot replies with a grid of four images; you then choose which ones to upscale or vary. Version 5 arrived in March, followed by 5.1, which applies more of its own styling, and 5.2, with better image quality. Midjourney withdrew its free trial, citing heavy demand and misuse, so using it means a subscription, currently between $10 and $60 a month depending on the plan.

**DALL·E 2** is OpenAI's image model, used through a plain web page: type a prompt, click generate, get a set of images. You can also edit an image or ask for variations of one. OpenAI sells credits: $15 buys 115, and each credit runs one prompt, which returns several images. OpenAI gives users full rights to use the images they create commercially, including selling and printing them.

**Stable Diffusion** is an open model from Stability AI, which means the trained model files can be downloaded and run by anyone. Its newest version, SDXL 1.0, came out on July 26. It makes images at 1,024 by 1,024 pixels, handles hands and lettering better than earlier versions, and needs fewer "magic words" in the prompt to look good. Stability says it runs on consumer graphics cards with 8 GB of video memory. You can also use it through Stability's own paid web app, DreamStudio, its API or its Discord server. Its license, CreativeML OpenRAIL++-M, is worth reading before any commercial project.

## One prompt, three tools

The fairest way to compare is to send the same prompt to each. Try this one:

```prompt
A cozy bookshop on a rainy evening, warm light from the windows, a cat sleeping on a stack of books, watercolor illustration
```

Here is what to look for in each set of results, rather than which one is "best":

- **Style.** Did the tool add a look of its own, or stay close to "watercolor"? Midjourney's 5.1 model is known for adding more of its own styling, so compare it with the others on this point.
- **Prompt details.** Check whether each tool kept all four elements: the bookshop, the rain, the warm light and the sleeping cat. Missing details tell you how closely each one reads a prompt.
- **Text and hands.** Look at any sign on the shop. Lettering is still a weak point for most image models, and SDXL's announcement names it as one area it improves.

Run the prompt more than once in each tool. One image says little; four or eight show you the range.

## Side by side

| | Midjourney | DALL·E 2 | Stable Diffusion (SDXL 1.0) |
|---|---|---|---|
| **Where you use it** | Discord bot | Web page | Your own computer, or web apps and an API |
| **Cost** | Subscription, $10 to $60 a month | Credits: $15 for 115 prompts | Free to run yourself; paid in hosted apps |
| **Ease for a beginner** | Medium: learning Discord commands | Easiest | Hardest if you install it yourself |
| **Control** | Pick and upscale from each grid of four | Edits and variations | The most: open settings and custom training |
| **Licence** | Set by Midjourney's terms; read them for your plan | Full commercial rights to your images | CreativeML OpenRAIL++-M; read it first |

> **Try it:** Write one prompt with four concrete elements (a subject, a setting, a light source and a style). Run it in two of the tools, four images each, and count how many of the four elements appear in each image. Then change only the style word and run it again.

## Which should a beginner start with?

If you want results with the least setup, start with DALL·E 2. It runs in a normal web page, the buttons are self-explanatory, and pay-as-you-go credits suit occasional use. It is also the easiest place to learn the core skill all three share: describing subject, setting, light and style precisely.

If you want striking images for posters, mood boards or social posts, and you are comfortable in a chat app, Midjourney is the most rewarding next step. Expect to spend an evening learning its commands and parameters, and decide on a plan before you begin, since there is no free trial.

Choose Stable Diffusion when you want control or privacy: you can run it offline, change every setting, and even train it on your own style. The trade-off is time. Installing it and learning its settings takes real effort, and you need a capable graphics card. You can try it in DreamStudio first, before installing anything.

Whichever you pick, two habits apply everywhere: save your best prompts with the settings you used, and read the tool's terms before you use an image in paid work. Legal questions about AI images, including lawsuits from artists over the images these models learned from, are still open.

## Sources

- DALL·E Now Available in Beta: https://openai.com/blog/dall-e-now-available-in-beta/
- Announcing SDXL 1.0: https://stability.ai/news/stable-diffusion-sdxl-1-announcement
- Midjourney (Wikipedia, revision of 8 September 2023): https://en.wikipedia.org/w/index.php?title=Midjourney&oldid=1174518917