How to Get Started with NovelAI Website Image Generation

NovelAI has 2 different parts:

• Am AI story writer, which we have reviewed in-depth here
• An AI anime image generator, also known as NAI Diffusion

Since its release, the anime AI generator has blown up in popularity, with commentators joking that users have abandoned the story generator for the anime image generator.

There’s truth to the memes: in the NovelAI Discord, we noticed the anime-image channels were much busier than the ones dedicated to story writing.

While you can try the AI story writer for free, you’ll have to pay to use the AI anime art generator.

That’s why we’re going to help you take an in-depth look at NovelAI’s Anime capabilities, before deciding if you want to purchase a subscription.

We’re going to show you the results it can get, and compare it to alternatives like Niji Journey so you can decide it if NovelAI is worth it.

First Impressions

We’re using Opus, the $25/month plan since we wanted to get an idea of NovelAI’s full capabilities.

There are 3 main settings:

• NAI Diffusion Anime (Curate): Good baseline quality and predictable subject matter. Recommended for streaming.
• NAI Diffusion Anime (Full): The expanded training set allows for a wider variety of generations.
• NAI Diffusion Furry (Beta): Beta model for furry and other non-human content.

I didn’t want to spend much time learning the settings, I just wanted to generate images. So I started typing:

There’s a cool autocomplete function: these are the actual tags the model has been trained on, with the tags at the top being the ones the model has the most knowledge of.

That’s to say: you don’t have to stay stick with the tags, but if you do, it’s guaranteed the model will have some knowledge of your subject.

The tags are also great to browse when you need some extra inspiration or ideas.

The actual generation is fast, taking less than 5 seconds for a 512 x 768 image. You get 3 options after you generate an image:

• Edit image: Draw on top the image
• Variations (cost 18 Anlas): Generate 3 new variations that are similar to the base image
• Enhance: Re-generates the image at a higher resolution

Edit Image

I thought this feature was going to re-generate the image with whatever I drew on top of it, making a beautiful new artwork.

Unfortunately however, it is only a standard editor with limited brush options.

Variations

You can generate 3 new variations if you like the base image, but want to see other versions.

You can repeat this process infinitely: select one of the variations you make, and create new variations based off of that one until you get something you like. Keep an eye on your Anlas though.

Variations cost 16-18 Anlas to generate.

Enhance

Enhance passes a selected images through NovelAI a second time, and you can tweak the prompt that goes with it. This is ideal if you like your image but want to change some aspect of it.

You have a couple of options to set:

• Strength: How much the image will be changed. Higher strength = more change.
• Noise: How much detail the new image should have. Higher noise = more detail.
• Magnitude: A combined stat for Strength & Noise
• Upscale amount: How much to increase the resolution of your enhanced image

Enhancing costs quite a bit of Anlas: 20 Anlas when you turn up Strength to the maximum

I found that the “new text prompt” part of enhancing did not work very well. This is what happens when you enhance:

You can see that higher strength creates more variation, and that higher noise must be paired with higher strength to work, or you will get unintended noise effects.

It seems like enhancing is an ideal feature for when you want to create new images that are similar to the original.

Users have reported getting higher quality images by not using NovelAI’s upscaler, but saving the image at 1x (original size) and using a different upscaler.

Waifu2x is a commonly recommended anime image upscaler. (Many websites with the same name, but the Japanese website is the only official one. You can change the language to English: https://waifu2x.udp.jp/index.html)

OK, but how do I generate beautiful images?

Here’s some tips on making the most of your generations, and doing popular things in NovelAI:

Improving the quality of your images

If you’ve spent any time in the Discord you’ll see that everyone puts “masterpiece” in their prompts. This is because it works. It makes the model pull from higher quality images with better art and better compositions.

We recommend you create a string of tags that can generate a style you love consistently.

Here’s an example, feel free to try it:

{{masterpiece}}, {{ultra detailed}}, {ultra quality}, {dramatic shadows}, {cinematic lighting}, intricate expression, 

You can see in this image, the image is nice without any additional descriptors, only the subject. But when you add the descriptors it really pops!

Curly braces such as: {masterpiece} make the tags stronger. You can add multiple curly braces, to push the strength of tags up and down relative to each other.

Square braces such as: [bad anatomy] make the tags weaker. You can add multiple square brackets.

Here are some common descriptors to add to your images to beautify them in different ways:

• masterpiece
• cinematic lighting
• cinematic angle
• dramatic lighting
• dramatic angle
• ultra detailed
• intricate details
• beautiful
• intricate expression

Conclusion

We recommend diving in and experimenting to make the most of NovelAI!

Many people however are concerned about burning through their Anlas by experimenting. This is especially true if they’re on the lower plans and only have ~200 generations/month.