Using NovelAI for the first time

Today, I registered an account with NovelAI to try generating images. From what little research I’ve done, there are a few tricks to creating a good looking image.

NovelAI

NovelAI is a service that can generate images from text. Such technology is also known as Text2Img.

Tips for NovelAI

It seems that there is a kind of canned statement that can be used to improve the quality of NovelAI’s output.

For example,

• masterpiece
• best quality, high quality
• highly detailed
• illustration
• beautiful detailed eyes, beautiful detailed face
• 4K, 8K
• oil painting
• realistic

However, too much prompting like this may lead away from the output that was originally desired. For example, if you want to output an animated picture and you add realistic, you will not get an animated picture.

The quality can be further improved by specifying in detail the elements you do not want output. Specifically, I specified the following text as a negative prompt.

{{{multiple views}}}, upside down, tentacle hair, sign, hat, man, cave,
light blue, blurry, picture frame, image frame, bad face, pointy ear,
animal,mask, cross, man, name plate, lowres, bad anatomy, bad hands, text,
error, missing fingers, extra digit, fewer digits, cropped, worst quality,
low quality, normal quality, jpeg artifacts, signature, watermark, username,
blurry, bad feet,

NovelAI has several parameters. In general, the larger the scale parameter, the more important NovelAI seems to consider prompts. However, as far as I could tell, a scale parameter of around 9~10 gave me just the right picture; if I set the scale to a value above 20, the drawing would break up.

Character Drawing

After several trials and errors, I found a prompt that generated a nice character. Here I generated a character named Katsushika Hokusai from FGO.

To output this image, I have specified japanese Brush Painting. Also, even though it is a watercolor painting, I felt that specifying 3D improved the quality.

Also, when specifying a color at the prompt, the closer the color was to the beginning, the more important it seemed. For example, when specifying red hair, blue eyes, the eye color often seemed to be red. However, this is a sensory thing and may not be the case in reality.

And although this is currently an unavoidable problem when using deep learning models, we could not generate characters with a small number of data. Characters from famous games and cartoons can be generated, but those that are not are very difficult to generate. It is necessary to specify the hair style and clothing in detail.

Personally, I feel that NovelAI would not eliminate jobs for existing illustrators and cartoonists. I follow and support several illustrators through pixivfanbox and other channels, and I don’t think their appeal has been diminished by NovelAI. In fact, I often visit Comitia, an event for original stuff. If you have heard of Comic Market, it is the so-called event for fan art. Comitia, on the flip side, is an event for original works. I don’t feel that software like NovelAI has made the original works offered at Comitia any less appealing. I will continue to look for interesting works at Comitia.

By the way, I came across a paper [1] the other day. In this paper, They found some cases where stable diffusion models blatantly copy from the training data. It is important to consider that, at least as of today, the diffution model may produce output that looks as if training data were copied. I think we also need to be careful when using image generation tools that they may have these characteristics.