We’re excited to announce our newest feature. Personalization! It’s the ability to design a personal style unique to you on niji.

You can do this by rating images.

Below is a comparison of a character produced by a personal style (left) vs. niji’s default style (right):

Step 1: Rate at least 200 image pairs on https://nijijourney.com/teach-niji

Step 2: Use the --p flag to see your personalized result

for example:

Step 3: Use your shortcode

When you run this command, you can get a “shortcode” at the end of the generation.

You can then take this and share this to keep recreating the same style. When you make more ratings, the “shortcode” changes, allowing you to access different versions of your own style.

You can also use this shortcode to share your personalized style with friends: using --p with a shortcode allows you to run someone else’s personal style on your account.

for example:

To see your personal style, you’ll need at least 200 ratings. More ratings will result in a style that better matches your tastes! You can check how many ratings you’ve done by checking the top of https://nijijourney.com/teach-niji for the count.

If you don’t like either picture in the set, MAKE SURE TO SKIP IT. 200 ratings of pictures you really like > 1000 ratings of pictures you don’t like.

When skipping and rating, try to consider whether this is what you like for that category. So even if the category is something you don’t like, consider whether this could be the best thing in that category. For example, your style will lack the ability to draw a good man, if you only rate pictures of women, per the rules above.

If you want a personalization that only draws very particular things, feel free to ignore this final instruction!

This feature has been long in the making because it has been hard to know whether we’ve made something “personalized.” Niji itself is already the desired personal style of the researchers.

Here at niji, we are obsessed with describing aesthetics. When we tune niji, we look for proper “visual grammar” when deciding whether to push the model in a certain direction.

(What is visual grammar: Lesson 1: Fundamentals of Measurement and Abstraction: The Theory of (How to Draw) Everything)

However, just like verbal language, the non-grammatical version is often more charming. And so, to capture that charm factor which exists outside of the right and wrong of “academic” art techniques, we devised an experiment to come up with a more objective way to test charm.

This produced a lot of what I felt to be dubious images!

To my surprise, their owners never fail to pick them every time, and I now suspect that aesthetics is something truly unique to everyone.

During very early development, we hosted an event where we invited 10 people to produce around 10k ratings each. We were surprised when three of the styles that came back were identical.

When we followed up with the participants, the organizer of the event admitted that two people did not show up, so he filled out the quiz himself three times. It turns out, at 10k ratings, the personal style constructed is as unique as a fingerprint: we accidentally identified a person with it.

Today, we’ve reduced this number drastically to craft your personal style. Although more ratings is always better, we estimate that 200 is the minimum necessary to get started with making your style.

One thing we wonder a lot about is the entanglement between style and subject matter.

As humans, we have selective attention: we pay close attention to the things that we like, and discard the things that we don’t like. In other words, paying attention is how we show affection.

If you care about something, you stylize it. If you don’t care about it, you don’t.

I’ve come to think of personal style as kind of like a special nickname that you make for the things that you are fond of: it is inextricably tied to subject matter. We discovered this on accident when we were studying certain personal styles which produced extremely handsome men. On closer inspection, we realized that these handsome styles had one thing in common: they were all made by women

What’s fascinating is that these women came from all sorts of occupations, but their personal styles produced handsome, well-drawn men, regardless of their artistic ability.

This measurement of “how much you care” can be quite drastic for concepts that require interpretation. In the above experiment, since men can usually hold coffee, most personal styles tend to draw something reasonable when asked to draw “a man holding a coffee.” However given a prompt like “a cat in pajamas,” there is a drastic difference in output, since cats do not usually wear pajamas. We can measure how much somebody likes cats, based on how this prompt is interpreted through personal style.

There is a certain appeal of this type of system: it is a hope that a third-party arbiter could tell us something hidden about ourselves. It has the same sort of mystique as horoscopes or meyers briggs tests.

(We’re surprisingly into this type of thing: Kindling the Spark of Inspiration)

I’ve always thought of niji as a way to explore the catalogue of human art history: it is a tool for refining visual ideas through a summarizer of human art data. Now, with personalization, it can also be used to explore not just the outer world, but also the inner self, to surface up various facets about your persona that you may not have thought about.

Of course, like all other types of horoscopes, take this with a grain of salt.

Now that we’ve started rendering out what used to be only in our brains, what’s going to happen to humans? It is evident to me, after the personalization experiment, that we are all way more different than I thought. Are these personal aesthetics going to converge, because they can now be shared and communicated in the open? Or are they going to diverge faster, because we are now sharing the weirder variants?

This field is nascent. We could be headed towards a future where each of us has a personal aesthetic presence to enrich the visual information that we receive. Or we could be going towards the future where each advertiser sells me new socks with the mathematically calculated best anime girl that meets my preferences perfectly.

It is certainly an exciting time to be alive and working on this problem.

As always, thank you for your support on this journey.

Related Articles