Official Danbooru, the learning source, releases a statement regarding the automatic illustration generation AI 'NovelAI'
Image generation AI such as Stable Diffusion and Midjourney performs learning on datasets consisting of large amounts of images. NovelAI, which has become a hot topic for its ability to generate high-precision two-dimensional illustrations, is said to include many images from the overseas two-dimensional image site Danbooru in its learned dataset. The official management team of Danbooru has released a statement regarding NovelAI.
Contact | Danbooru
https://danbooru.donmai.us/contact
Danbooru is a site where you can post and share 2D illustrations, and it is especially popular because it has many tags that indicate the content depicted in the illustrations and has excellent searchability. However, the problem is that many of the two-dimensional illustrations uploaded to Danbooru are reprints of works posted on SNS such as Pixiv and Twitter.
The management of Danbooru says, ``To request deletion of an image, please email [email protected]. Attach the URL of the relevant image on Danbooru and proof that you are the rightful owner of the image. We would appreciate it if you could do so.''He added that he would take action against unauthorized reprints.
Furthermore, Danbooru released the following statement regarding NovelAI, which uses Danbooru's images as a dataset.
First of all, we would like to make it clear that we are not affiliated with NovelAI and do not endorse or condone what they do.
The most efficient method for artists who do not want their drawings used by NovelAI is to contact NovelAI directly, delete their drawings from the company's learning data, and remove them when generating AI art. Requesting that users exclude their name or handle from the 'prompts' they enter.
At this point, NovelAI's training data is already complete, and due to the structure of the AI model, even if you delete a picture from Danbooru, the information learned from that picture will not disappear from NovelAI's hands. Please note that. please.
Even if a picture has never been posted on Danbooru, there is a high possibility that it will be collected by NovelAI etc. as soon as it is posted on Pixiv or Twitter.
In other words, we hope that you can consider that the pictures that NovelAI can collect from Danbooru can also be collected directly from Pixiv and Twitter.
In the first place, most AI art generation services such as NovelAI are based on Stable Diffusion, a software provided by Stability.ai, to create their own AI models.
Stable Diffusion itself was created by collecting 5 billion pictures from various sources on the Internet as training data.
This includes not only sites that artists often use, such as Twitter, Pixiv, ArtStation, and DeviantArt, but also Reddit, Pinterest, Facebook, Google Image Search, and countless others.
If the picture has been imported into Stable Diffusion, deleting it from Danbooru now will probably have little effect.
In other words, what artists should understand is that the AI model of the AI art generation service is learning not only from Danbooru but also from the Internet as a whole.
As long as the pictures are publicly available somewhere on the Internet, there is always a risk that they will be used as learning data.
There is no way to prevent it other than contacting the developer of the AI model.
Danbooru management introduces a site where you can check whether your work is being used for training image generation AI.
``Have I Been Trained?'' that allows you to search whether your work has been used without permission by image generation AI - GIGAZINE
In addition, it is possible to notify the AI developer by declaring that you do not want your picture to be used for AI learning and using the following service.
Spawning
https://spawning.ai/
There are many posts on Twitter about what kind of illustrations can be generated with NovelAI.
#novelAI This is amazing... I can do anything now... pic.twitter.com/eRO2LxHKwe
— Polygon Z (@ZporigonZ) October 4, 2022
It was amazing when I loaded VRChat photos into NovelAI lol pic.twitter.com/S1VmdJWBrW
— Nemumu Shito???? Robobo Shito???? In operation (@nemumusito) October 4, 2022
#NovelAIDiffusion #novelAI
— Kamipallet (@kaimipallet2) October 4, 2022
I had a photo of a figure made into an illustration using AI.
After adjusting the parameters to make the most of the original picture, it looks a lot closer pic.twitter.com/yOEgrSlSYl
#novelAI
— cross (@CROSS_VoidLove) October 3, 2022
Apparently this is what happens when you throw Irasutoya at this guy . This guy might actually be worse than I thought.pic.twitter.com/aULQiWx9ka
Consideration of fixing the texture of the picture with black and blonde yuri prompts. Do not use names of existing artists.
— Episode 852 (@8co28) October 5, 2022
Yuri output itself is stable.
1 is the coloring that NovelAI is good at, 2 is anime coloring with deep shadows, and 3 is watercolor-like, giving it a slightly nostalgic look.
I thought there wasn't a wide range of designs, but it depends on the text. #NovelAI #NovelAIDiffusion pic.twitter.com/dRepgc83CK
#NovelAI #NovelAIDiffusion
— Episode 852 (@8co28) October 4, 2022
Three-sided drawing of the character (sometimes diagonally)
Although it feels a little strange, it is quite high quality
3D modeling seems to be going well in the VTuber world and VRC world pic.twitter.com/nL3xasLiXY
Cute brown female brat in summer.
— MABO (@G03MIAq1YTrRVOR) October 5, 2022
Evil smile is useful for the female brat's expression. After that, I am thankful that when I put Armpit in, it naturally moves to show my armpits. #NovelAIDiffusion #novelAI pic.twitter.com/4P2HNS16lN
It's amazing how just adding words like rain, wind, and butterflies creates a fantastical feeling... #NovelAIDiffusion #novelAI pic.twitter.com/PYex6P3c7z
— MABO (@G03MIAq1YTrRVOR) October 4, 2022
[Translated by DeepL] In the warehouse of the school gymnasium, a bloody girl in a sailor suit is sticking out a kitchen knife. There was no light in his eyes, and he looked like he was about to attack. #novelAI pic.twitter.com/ePaBcUx1J1
— Aldehyde @ “Kimi wa Koushi” Volume 1 will be released on September 9th! (@alde_hyde) October 4, 2022
I tried to generate the popular one from the illustration (first one) that I drew on the site! #NovelAIDiffusion #NovelAI pic.twitter.com/pewfqJKN4C
— picco (@picco_xxx) October 3, 2022
#NovelAI #NovelAIDiffusion
— Episode 852 (@8co28) October 4, 2022
They also provide endless collections of croquis poses like the one in the first one.
2.3. It is possible to output full-color pictures of the same character in the same costume in various angles and poses. It is simply too powerful as a reference material for studying pictures (there are some inconveniences, but it is within the acceptable level) pic.twitter.com/1zkTwgpMsh
Wow, the pose specification in the 3D model is perfect #novelAI pic.twitter.com/Q2y7a5aGi1
— Max (@minux302) October 3, 2022
#NovelAIDiffusion #novelAI has been released 4
— Episode 852 (@8co28) October 3, 2022
Well, anyway, it's dangerous, but what's more dangerous than anything is this... Her habits and tastes... are found out, that's it... Ha~ A little girl is too strong This is... pic.twitter.com/Q02rchIYq6
I tried using NovelAI and it was really amazing...
— MABO (@G03MIAq1YTrRVOR) October 3, 2022
It was so cute that I made only little girls.
#NovelAI pic.twitter.com/LWcUcrt1t8
Isn't it a monster to be able to generate images so accurately with instructions that are only crudely translated using deepL? ? ? ? ? #NovelAIDiffusion #novelAI pic.twitter.com/2tNaHDiF8t
— Monochrome @ Reverberate aiming for birds (@MoNoKuRo1030) October 3, 2022
#I moved the image output with NovelAI with TalkingHeadAnime3 I think this is probably the quickest way to start becoming a 2D VTuber pic.twitter.com/V8DvQChkRs
— MAX@VRC (@akagi_max) October 3, 2022
I see, AI-kun #NovelAI #NovelAIDiffusion pic.twitter.com/RZVxVZZxhl
— Episode 852 (@8co28) October 3, 2022
The two end with a happy kiss...
— Denden (@den2_nova) October 3, 2022
It's amazing that you can create a kissing picture that would easily fail without any discomfort. Spells are ALT. #NovelAIAIDiffusion #NovelAI pic.twitter.com/JdqG160Szz
How to use NovelAI……I understand it.
— hoshimi12 (@hoshimi12) October 3, 2022
Left: Output picture Right: Uploaded picture pic.twitter.com/Ni2cQH6UHQ
I made a light novel-like picture, so I made it look like a cover #NovelAI #NovelAIDiffusion #aiart pic.twitter.com/YQYBwC9BWl
— Nikimonaka (@nkmonaka) October 4, 2022
There are many other models that use Danbooru images as datasets. There are various types of Stable Diffusion derived models, and the following article summarizes which models use what datasets for learning.
Summary of various specialized model data that can be used with image generation AI 'Stable Diffusion' - GIGAZINE
Additionally, NovelAI was originally an AI that automatically generated novels, and illustration generation is just one of its functions. The image generation part is based on Stable Diffusion, but it implements a unique algorithm. However, unlike Stable Diffusion, which was developed as open source and is also available to the public, NovelAI's source is not publicly available.
We decided to release #NovelAIDiffusion models as standalone releases:
— NovelAI (@novelaiofficial) October 3, 2022
NovelAI Diffusion Anime (Curated)
Good baseline quality and predictable subject matter.
NovelAI Diffusion Anime (Full)
The expanded training set allows for a wider variety of generations.
Related Posts:
in Software, Web Service, Manga, Anime, Art, , Posted by log1i_yk