Categories
Uncategorized

Be weird to diversify Midjourney images

Generate more diverse and visually interesting images with Midjourney’s ‘weird’ parameter. By breaking away from typical prompts, you’ll be able to generate more unique and interesting results.

We know that the data used to train foundational AI models perpetuates bias, and in Midjourney that most clearly manifests1 in results involving people; unless explicitly instructed otherwise, the people in generated images tend to look beautiful.

To show you what I mean here’s a prompt I wrote, with the resulting generated images below it:

candid photograph, close-up, a young Korean woman with bleached blonde hair, bright natural light, happy, joyful --style raw

The results look great, but clearly fall squarely within conventional standards of beauty. Not everybody looks this beautiful (myself included), and we should reflect that.

Midjourney 5.2 introduced a new parameter, called (somewhat unfortunately) --weird. Here’s how it’s described:

Explore unconventional aesthetics… this parameter introduces quirky and offbeat qualities to your generated images, resulting in unique and unexpected outcomes.

Midjourney documentation

By adding the parameter at the end of a prompt, and experimenting with its value (from 0 to 3000), you can get results of people which are more characterful and, in my opinion, more interesting. Here’s the same prompt I used above, but with the --weird parameter added:

candid photograph, close-up, a young Korean woman with bleached blonde hair, bright natural light, happy, joyful --style raw --weird 1000

This results in images of people who are much more unique and varied; a little more goofy, a little more dorky, a little more human (for want of a better word).

You should run a few tests with that value; sometimes you need to ramp it up to 1000, as above, but sometimes a lower value is sufficient.

Here’s a second example, without the parameter:

photographic portrait, close-up, a young woman with tan skin and freckles, bright natural light, happy, joyful --style raw

Note that I didn’t specify ethnicity in this prompt and it defaulted to caucasian, which is another illustration of the bias in the data (and something to be aware of when writing prompts). But run the same prompt again with a --weird value of 100:

photographic portrait, close-up, a young woman with tan skin and freckles, bright natural light, happy, joyful --style raw --weird 100

While the --weird results are still beautiful, they’re less typical, and more ethnically diverse.

Here’s one more example, this time with a male subject:

candid photograph, close-up, a mid-30s Indian man with dark beard, bright natural light, happy, joyful --style raw

Handsome — and very similar — dudes. But when we add a --weird value of 100:

candid photograph, close-up, a mid-30s Indian man with dark beard, bright natural light, happy, joyful --style raw --weird 100

The people generated here are far more varied and much more like ‘average’ (sorry) people.

Until foundation models are made less biased, and results made naturally more diverse, it falls to the prompter to be conscious of what they produce. For now, --weird is a useful technique to make that happen.

  1. It was suggested to me that this may not be bias in the dataset, but an intentional decision by Midjourney to train the model to favour beauty. I accept this could be correct, but I think that it shouldn’t be; because it’s much easier to add ‘beauty’ in the prompt than it is to exclude it. ↩︎