Each Niji model drives a bit differently. Today we’ll be sharing some tips for handling Niji 7!

Precision Prompting

Compared to our previous offerings, this model will draw more of exactly what you want, nothing more, nothing less.

For example, given the same prompt:

The output from Niji 6 is a more robust picture: It has lighting, a background, and a more interesting crop.

In contrast, the output from Niji 7 has less accompanying elements, but it is more precise: the details that I asked for, “lime green hair, blue eyes, hair in a single bun”, stand out because the model doesn’t put other things in there.

It takes some getting used to, but you’ll see it really shine when the prompts are more demanding. For example, on this longer prompt:

The image quality is very comparable at a glance! (Every one of our models is good! 😉) Where you’ll see Niji 7’s strength is in the precision prompt following: how the hair stays in a single bun, how the belt stays white while the frills stay blue, how the floor is always made out of cobblestone.

As a general rule of thumb, when working with Niji 7:

FAQs

We’ve been observing community questions since launch. Below is a compilation of the most frequently asked.

I want a background

The default style doesn’t come with a background.

If you want a background but you don’t have a good clue of what you want, you can use a generic word like “indoors” or “outdoors”

a girl with short lime green hair, blue eyes, hair in a single bun, outdoors

I want a fullbody picture

Specifing what sort of shoes that your character is wearing will turn the picture fullbody.

I want less details

Some keywords like “minimal,” “simple,” or “flat shading” work well here.

I want less flat

A variety of approaches work here, depending on what you are going for:

You could use words that specify some strong lighting condition, like “golden hour”

You could also use words that specify a more rendered style, like “painted”

You could also use camera words that specify more depth, like “depth of field”

I want to emulate the look of an older model

You can use sref to do this!

To do this, you can input the same prompt into an older model, to generate an image.

Here you can see I have used Niji 5 expressive, which has limited prompt following, to create the image on the left.

Then I can feed the result through sref and the same prompt to combine niji 7 precision prompt following with the look of the older model!

I want to see more differences in the options

Turn the “chaos” parameter up. (The slider is also called “Variety” in the web ui)

I want to see less anime

You can see less anime results by flipping the style setting from “default” to “raw”

Designing a Prompting System

In this blog, we talk a lot about designing aesthetics, but the way that a model responds to words is just as important of a challenge, when making a model.

First-hour users (”cat girl maid”) want to see something exciting. 1000th-hour users (”girl with green hair, blue eyes, cat ears, hair in a single bun, wearing a maid costume, sitting on a couch by a window, night time”) want to see something consistent. And this is many times at odds with one another.

Cinematic One-Shot

Traditionally we have leaned more towards making a good one-shot recommendation. Which is to say, no matter what you ask for, we prioritize giving you something that looks cinematic and mind-blowing.

But this presents a problem: if the simple prompt and the well-thought out prompt produces the same result, then how do you get differing levels of excitement?

Do we want differing levels of excitement?

To illustrate why we might want some pictures to be less exciting than others, here’s a snapshot of the clips used to make the niji 7 launch video.

There’s a lot of girl faces for sure, but to make those shots stand out, we need a lot of things like backgrounds, props, closeups, long shots which are not girl-focused.

And so, to produce this video, we could not only just make the peak result good (as we have traditionally done). We also had a very strange task of designing a system that would get you reliably unexciting shots when you asked for it.

In practice. It’s a question full of trade-offs.

On the one hand, if it’s possible to get unexciting shots, then new users will struggle because some percentage of beginner prompts will return unexciting results.

On the other hand, the experienced user will have access to wider range of results.

After much deliberation, we leaned a little bit closer on the experienced user side for this release. You’ll find that niji 7 prompting is overall more difficult, but more rewarding once you master it.

As the audience for generative AI matures, I think we see a greater number of people who want to see the efforts of their knowledge reflected in the prompts that they type, and niji 7 truly honors that desire.

Personally, I think it’s never been easier to make an anime with our models, so I hope to see you get out there and do it.

Whether your favorite model is Niji 7, Niji 6, or even Niji 5 Expressive, the beauty of having a garage of many cars is that each is fit for a different purpose. I want you to have the power of taking your racecar to the track, your camper van into the wilds, and your hatchback to pick up the groceries.

Whatever model of ours you choose to drive, you can be assured it was constructed with care and devotion. If one day you should lose a marble that which rolls underneath your driver’s seat, you’ll find that that the underside of the seat is beautiful when you look down to retrieve it.

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

관련 기사