How do Flux, Midjourney and DALL·E 3 compare for realistic images?

For most of the past year, the choice for realistic AI images came down to Midjourney or DALL·E 3. On August 1, a new company, Black Forest Labs, entered with Flux.1, and its team is not new to the field: it previously created Latent Diffusion and several Stable Diffusion models. Now there are three serious options, each with a different personality.

"Realistic" usually means three things: believable people (hands, skin, faces), physically sensible scenes (light, shadows, reflections), and fine detail that holds up when you zoom in. Those are the criteria to judge by, whatever the tool.

The three tools side by side

Flux.1 comes in three versions. [pro] is the top model, offered through the company's API and partners. [dev] has open weights, meaning anyone can download the model files, but its licence is for non-commercial use. [schnell], German for "fast," is the quickest and is released under the permissive Apache 2.0 licence. All three are built on a 12-billion-parameter transformer, the same broad architecture used in language models. Black Forest Labs says [pro] and [dev] beat Midjourney v6.0 and DALL·E 3 on image quality and prompt following in its own tests; note that those tests predate Midjourney V6.1 and are not independent.

Midjourney V6.1, released on July 30, focuses on exactly the problems that make AI photos look fake: more coherent arms, hands and bodies, fewer pixel artifacts, better skin texture, and more precise small details such as eyes and distant faces. It is also about 25% faster for standard jobs.

DALL·E 3 is the oldest of the three but has an advantage for beginners: it is built into ChatGPT, so you can describe what you want in ordinary language and ask for changes in conversation.

A worked same-prompt test

A good test prompt names the subject, the setting, the light, the camera and one tricky detail. Here is one:

Candid photograph of an elderly fisherman mending a net on a wooden dock at sunrise,
weathered hands in sharp focus, his breath visible in the cold air,
soft golden side light, shallow depth of field, 50mm lens, natural colours,
a small sign on the boat behind him reading "MARIE-ÈVE"

Use it as written in Flux.1 and Midjourney (add --ar 3:2 in Midjourney for a photo shape). In ChatGPT, paste it and add "Use this prompt exactly; don't rewrite it," since ChatGPT otherwise rephrases prompts for DALL·E 3. Generate several images in each tool, then score them. The scores below are illustrative:

Score on the full-size image, not the thumbnail; most flaws in realistic images hide in fingers, teeth, ears and background text.

Which one should you choose?

Choose Midjourney if you want the most consistently beautiful, photographic results and are happy to learn its parameters. Choose Flux.1 if you value control and openness: [schnell] and [dev] can run on your own hardware or a hosting service, and [pro] is available by API for apps. Choose DALL·E 3 if you already pay for ChatGPT and want to work by conversation, or need images that follow a long, specific description. For anything you publish, check the licence of the exact version you used, especially Flux.1 [dev].

Which one handles text in images best?

Text has long been a weak point of image models, which tend to produce letter-like shapes rather than words. All three makers now claim progress. Midjourney says V6.1 draws words more accurately when you put them in quotation marks in the prompt. Black Forest Labs lists typography among the areas where it says Flux.1 [pro] and [dev] beat Midjourney v6.0 and DALL·E 3. OpenAI emphasizes that DALL·E 3 follows the wording of a prompt closely, which helps when the text is short.

In practice, short text works far better than long text in every tool: a name on a boat, a two-word shop sign, a single headline. Put the exact words in quotation marks, keep them short, and mention where they appear. Accented letters, such as the È in the example, are a good stress test.

Flux.1 is only days old, so its makers' claims still need wider testing, and results vary from prompt to prompt in all three tools. Generate several images, zoom in on every letter, and plan to fix small errors in an image editor. For a poster or logo where the words must be perfect, adding the text afterward in a design tool is still the most reliable method.