Tutorial: Intro to Image Generation

If you prefer to watch instead of reading, we have a video version of this same tutorial available here.

Hello there! Today, we'll teach you how to create your own NovelAI Diffusion masterpieces.

This tutorial will cover the absolute basics to create images with NovelAI Diffusion Anime, our state of the art image generation model. As you may guess from the name, it focuses on anime-inspired art!

In this case, we're already logged in, subscribed, and ready to go. First off, let's hop into the Image Generator from the main dashboard; click that Image Generator banner to jump right in!

This is your creative studio; to begin we will start with a simple idea and type it into the prompt box.

We'll go with no humans, flower field, sunset to create a basic background image.

Since we want an image of a landscape with no characters on it, the no humans tag at the start is very important here.

Press the 'Generate' button, and the AI will start creating your image!

Look! A beautiful picture! Our very first masterpiece!

Now, to explain prompting, we must first talk about tags. They're one of the most unique aspects of NovelAI!

The AI has been trained in English, but in tag form, rather than natural language, as in normal sentences you'd use to describe your vision with.

Think of tags as the elements that make up all images.

These tags give us a lot of control over the image composition, and there are a lot of them. You can get super detailed with tags, and with enough of them you can get very consistent results across different images.

Take note of the little circle on the side of each tag. These circles indicate the AI's knowledge of any given tag. The brighter, the more knowledgeable.

Don't worry if this sounds complicated, the AI will suggest tags to you as you type, making it all a little easier. Look at the suggestions pop up as we type!

By the way, tags should be separated with a comma, and a space. This makes it easier to read and keep track of your prompt too, especially since some prompts can get pretty long.

You can also check how much space you have left with the Context Limit bar. It fills up as your prompt grows in length! Look at it go!

Let's generate a girl next!

First off, let's get rid of the tag no humans and add the tag 1girl right at the start of the prompt. Things closer to the front of your prompt will have a bigger impact on the overall image.

Let's also add some details. For hair tags, we're going with messy hair, brown hair and finally the tag green eyes for the eyes. We'll use the tag from side to prompt for a specific camera angle, and the tag school uniform for the outfit.

Our prompt now is: 1girl, flower field, sunset, messy hair, brown hair, green eyes, from side, school uniform

Let's press the generate button and see what we get.

Voila, she's looking just as we've described in the prompt!

To show how we can keep editing this image, we're going to click on the seed button below the image, now we can make adjustments to the prompt without changing the entire image.

Speaking of which, you can get a more granular control over what the AI will focus on by strengthening or weakening tags.

To strengthen a tag, all you have to do is surround it in {curly brackets}. The more brackets you use, the stronger the effect will be, but too many can have some strange effects. Also, make sure you use the same amount on both sides unless you're at the very end of your prompt.

We can also do the opposite and tell the AI to weaken a tag by surrounding it in [square brackets]. This works well for things like make-up that can sometimes turn out a little too extreme...

Let's use the messy hair tag as an example. Here you can clearly see how the hair progressively gets messier when we go from a weakened [[[[[messy hair]]]]] tag, to a normal messy hair tag, and then to a strengthened {{{{{messy hair}}}}} tag.

Now, there's also times where the AI adds things to your image which you didn't ask for. For that we have Undesired Content.

Here, you can just write down anything you don't want the AI to generate. It's that simple!

And yes, you can also use {curly brackets} and [square brackets] here as well.

As an example, let's add petals into Undesired Content and watch how all those flying petals disappear from the image.

We can also switch to a more horizontal resolution down in the Image Settings section. Depending on what you're trying to generate you might find that some aspect ratios work better than others.

Here you can also choose how many images you want to generate at once.

Keep in mind that generating more images increases the cost in Anlas—the NovelAI currency used for image generation. Users subscribed to the Opus tier can generate as many images as they want without spending any Anlas, as long as they stay within the "normal" image dimensions, and only generate one image at once.

All the images we've been generating so far appear in the History on the right. If you ever generate horrors beyond belief, you can also remove them from your History, to avoid nightmares.

Once you're satisfied with your results, you can click the 'Download' button below any image to save it to your computer.

If you've generated up a storm you can also grab all of your generated images at once by clicking on 'Download ZIP'!

For user privacy reasons, NovelAI doesn't save any of your generated images. This means that if you close or refresh the page, all your images will be lost. You will get a warning pop-up, should you accidentally try to leave. So don't forget to save any of the images you like!

Another good thing about saving your images is that you can go back and recover the prompt and settings used to create the original image simply by dragging and dropping it on the site!

Do keep in mind that, by default, the seed value of the original image also gets imported. While this can be useful in some cases, if you want a different image with the same prompt, then all you have to do is click on the seed value to clear it.

Once you're comfortable with the process of generating images, you can start playing around with more advanced settings like steps, prompt guidance, sampler, and prompt guidance rescale. The default values are already optimized for a good experience in most use cases, but experimentation can always lead you to great results.

Don't worry if you feel like you messed something up while experimenting with these settings. You can always press the reset button to return them all back to the default values.

If you'd like to learn more about how these settings work, we highly recommend checking the rest of our official NovelAI Image Generation documentation, or taking a look at one of our more in-depth tutorials.