#14 Midjourney - Is the powerful AI tool beneficial for type designer?

At the beginning of November 2022, the AI tool Midjourney received a major new update. Version 4 (--v 4) is now available and I have tested the AI tool for its usability in the area of type design.

output created with midjourney

Version 4 can be accessed via the Midjourney Discord bot. There is no access to the open code, so no parameters can be set. The bot can only be used to generate additional versions of an image. Within one minute, the model generates four images, which in the later upscaling process reach a maximum image size of 2048 × 2048 pixels. This is the highest known resolution at this stage. In experiment #14 I test whether this AI tool is helpful for type designers.

Many of the generated images match the prompts. It is noticeable that Midjourney often renders the images as illustrations. It is therefore clear that this AI tool has its own playful, mystical style, which is almost certainly based on the training data. Only the prompt "A word on white background" was unable to generate a matching image. The probability of a match is 63%.

Since the experiments with this AI tool are proving promising, I have further refined the prompts to investigate whether more precise properties of letters can also be realized. My goal was to generate letters with properties like serifs or different font styles. I was able to achieve the following results with Midjourney. Letters with serifs can be generated reliably by the AI tool, but not distinguished from sans serif, or slab serif. It is likely that the model has received many letters with serifs as training data.

Another approach to generate serif and sans serif fonts can be to use the names of well-known antiqua or grotesque fonts. Even if the generated images do not look like the Bodoni, it is still recognisable that the model knows at least one difference between the two typefaces and reproduces it in its style.

With the addition of "italic" in the prompt, it can certainly be interpreted that the generated letters are shown in an italic style. No significant differences are discernible in the font styles hairline, thin, light and bold. Even the most accurate of the text-to-image models I tested cannot reliably convert type design specific terms. In further experiments with Midjourney, I found that this AI tool is particularly good with textures and produces outstanding results in the area of lettering and 3D lookalikes.