日本語: 新しい niji のスタイル?!

There are two ways to construct a road: you can either set a predetermined paved path, or you can put down a lawn, observe the so called “desire paths” that people naturally wear down through the grass and pave those over into a proper road.

Now that niji 5 has been released for about over a month, we’ve had some time to observe usage statistics and prompting patterns. Just like the case with physical roads and lawns, people often wandered off the optimal usages that we’ve carefully prepared.

As model makers, we were curious what we could do with this data. As it turns out, the model we trained from the difference between “what we tuned niji to do” vs. “what people wanted to use niji for” was so interesting that we decided to release it as its own separate style.

We’ve served a full 4-course meal before with the style event, so we hope that you’ve saved some room for dessert.

prompt: miku enjoying a lavish feast [model sample, results subject to change]

prompt: miku enjoying a lavish feast [model sample, results subject to change]

The hidden vector of model quality

There is an interesting thing about model quality which cannot be demonstrated in pure pictorial output: it is the model’s willingness to draw what people ask for.

prompt: A very nice kakapo [model sample, results subject to change]

prompt: A very nice kakapo [model sample, results subject to change]

Consider this example: if we tune a model that draws very nice kakapos, but nobody ever asks for any kakapos, then it is about as effective as if we have never tuned a model that draws kakapos at all. The opposite is true: small details that we didn’t exactly balance feel like they’re a big part of the model when the users eagerly prompt for them.

Untitled