Let's understand three things that AI struggles with

Let's understand three things that AI struggles with

There are three things that AI clearly struggles with when generating AI illustrations.

Three things that AI struggles with.
  • AI cannot accurately depict drawings of hands.
  • AI cannot accurately depict drawings of genitalia.
  • AI cannot accurately depict drawings of sex.

In this article, I will explain these three characteristics and consider countermeasures.

Contents

AI cannot accurately depict drawings of hands

When generating illustrations with AI, one quickly notices the unnaturalness of the hands.
There are many cases where if the hands look unnatural, the entire illustration appears unnatural, making it a fatal weakness in AI illustration generation.

AI has difficulty accurately depicting hands
AI has difficulty accurately depicting hands

AI is not good at depicting hands, so it is also not skilled at making characters hold objects such as swords or weapons.

There is TEXTUAL INVERSION which specializes in correcting the hands used for Negative Prompt, but it does not lead to a dramatic improvement.

An example of TEXTUAL INVERSION that specializes in accurately describing hands

The easiest and most convenient solution

The easiest and simplest measure is to... not depict hands. While this may be a passive solution, I think it may not be necessary to spend time forcing this solution until AI illustration technology makes further progress.

How specifically can we avoid drawing hands?
Enter the following words as prompts for avoiding hand drawing.

Prompt examples for avoiding hand drawing
  • arms behind back both arms behind one's back
  • arms behind head both arms behind one's head
An example with the prompt  "arms behind back"
An example with the prompt "arms behind back"
An example with the prompt  "arms behind head"
An example with the prompt "arms behind head"

These two words are very useful and can effectively avoid the difficulty of drawing by hand.
They can also be used in scenes where characters are lying down, so please try applying them in various situations.

AI cannot accurately depict drawings of genitalia

Unless using an AI model specialized in that field, it is impossible to accurately depict genitalia through AI. Japanese artwork that could be used as training data is almost always censored with mosaic covering almost 100% of the area.

In Japan, it is not permissible to publish uncensored images due to legal issues. Therefore, it is possible to conceal them by applying a mosaic.

Even if it's a little strange, I can cover it up with a mosaic
Even if it's a little strange, I can cover it up with a mosaic

To solve it, rely on LoRA

However, there are cases where uncensored content is preferred overseas, so in those cases, I think the only option is to introduce a specialized LoRA.

You can search for penis or vagina on civitai (https://civitai.com/).

Even with the use of specialized LoRA, challenges remain.
When multiple LoRAs are used together, they can interfere with each other and decrease the quality.
In other words, it is difficult to use a LoRA specialized for character expression and a LoRA specialized for genital depiction at the same time.

AI cannot accurately depict drawings of sex

AI, not only in terms of sex, but also in depicting scenes involving multiple people engaging in physical contact, tends to struggle.

AI has difficulty depicting scenes involving multiple people interacting physically
AI has difficulty depicting scenes involving multiple people interacting physically

This causes various problems such as the fusion of the body or the increase of legs and arms.

The method to solve this is also dependent on LoRA

To solve this task, the best approach would be to use LoRA that specializes in sex positions.
For example, you can try searching for doggystyle, cowgirl, or yuri on civitai (https://civitai.com/).

There are still challenges even when using specialized LoRA.
When multiple LoRAs are used together, they can interfere with each other and decrease the quality.
In other words, it is difficult to use a LoRA specialized for character expression and a LoRA specialized in sex position at the same time.

I would like to introduce my recommended tools, LoRA and Textual Inversion, that I use below.

The textual inversion that I recommend is...

Even when used in conjunction with LoRA, which specializes in character expression, Textual Inversion has relatively little interference.

Sex positionURL
Missionary
Bound
Wrists
Corneo's POV Bound Wrists Missionary Embedding
https://civitai.com/models/4725/corneos-pov-bound-wrists-missionary-embedding
Cowgirl
position
Corneo's Cowgirl position embedding for anime
https://civitai.com/models/4218/corneos-cowgirl-position-embedding-for-anime
Paizuri
POV
Corneo's POV Paizuri Embedding
https://civitai.com/models/4475/corneos-pov-paizuri-embedding

The LoRA that I recommend is...

Weight adjustment is essential when using LoRA, which is specialized in character expression, at the same time, as they interfere with each other.

Sex positionURL
Doggy
Style
doggystyle v2
https://civitai.com/models/37907/doggystyle
Cowgirl
position
POV Squatting Cowgirl LoRA [1 MB]
https://civitai.com/models/8877/pov-squatting-cowgirl-lora-1-mb
Paizuri
POV
POV paizuri v2
https://civitai.com/models/36836/pov-paizuri
Paizuri
straddling
straddling paizuri
https://civitai.com/models/55299/straddling-paizuri

There is also an aspect that depends on the idea

If I consider other possible measures to avoid this, there are methods such as Tentacle sex (Tentacles sex can be expressed sexually without LoRA), which I often use. It can be expressed without a partner, so a post-coital description might also be good.

This image was generated using only prompts without using LoRA
This image was generated using only prompts without using LoRA

It depends on the ideas around here.
See this article for prompts.

Contents