Midjourney may have just solved a big problem with AI image generators

AI image generators continue to make new advances almost by the day. Last week, it was the launch of Ideogram with its improved text rendering. Now it seems Midjourney may have just solved one of the biggest problems holding back text-to-image AI: consistency across images.

Specifically, Midjourney has added a tag that's intended to create more consistent characters. Since diffusion-based AI image generators create new images every time, even when using the same prompt, it can be very difficult to create two (or more) images in which a character looks like the same person. This has made it hard to use AI for storyboards or anything that requires that consistency, but the long-requested new feature tells the AI when to try to respect the appearance of the character in a reference image.

Midjourney's solution is a new character reference tag: -cref. You generate an image in Midjourney (the tag seems to works best with Midjourney images), copy the URL, create a new prompt and then add -cref as a parameter along with the URL of the original image. While it doesn't appear to be perfect, it does seem to generate much more similar looking facial features, body shape and clothing for characters across a series of images, based on early examples that users have been sharing.

The approach could eventually solve similar problems for other subjects in AI image generators. I predict that Midjourney could aim to further develop the feature to recognise more than just human characters. See our round up of AI art tutorials to learn more about image generators.