astronglitch.bsky.social
Insomniac Illustrator 🔞 | EN/PT-BR
Links: astronglitch.carrd.co
Comms: vgen.co/AstronGlitch
Stream: twitch.tv/AstronGlitch
Discord: AstronGlitch
[email protected]
395 posts
357 followers
1,773 following
Regular Contributor
Active Commenter
comment in response to
post
While I do believe we have enough information to make a lot of conclusions, my main goal here is simply to spread the information I found, not necessarily to make people agree with me.
Ideally, I agree they should be more clear with what they do.
comment in response to
post
About the YouTube, the paper explains how they got the videos, based on keyword searching in the entirety of YouTube, meaning there's no way they could've filtered for only copyright free content as there's no way to check that for random videos on YouTube outside of manually.
comment in response to
post
Ideally having the training data disclosed would be good, but what we can guarantee so far is that for sure they did it on top of a pre-trained model, meaning there's stolen data in there.
comment in response to
post
I'm glad it's helpful
comment in response to
post
Outside of this, as I said before, I don't think is good to support a company that openly says that they wanna save money by not paying artists and creatives.
comment in response to
post
about the creation of a dataset to improve 3D head generation made fully by scrapping data from YouTube videos, which are publicly available but not necessarily copyright free.
comment in response to
post
guarantee they aren't gonna run into any issues (i.e. they're careful not to be sued as well)
- Are they using copyrighted material on their training data?
While I can't confirm which data they used on what, as that information is not disclosed, one of the papers I linked (funded by them) is (+)
comment in response to
post
I can see what you mean and I can help clarify some of these points.
- Are they aware of the unethical problems with SD?
Yes, they mentioned in their YouTube video that Midjourney and SD are being sued for copyright infringement and when they use those in the initial stages, they check after to (+)
comment in response to
post
Besides that, KRAFTON also funded the creation of other stolen dataset and funds more research on GenAI because they don't want to pay the creative workers, which is also something I can't get behind.
comment in response to
post
It's a good question and I can understand where it comes from but the way I see it they're aware that they're using something made with stollen data, they don't need to use it in any way if they don't want, so it's still also their fault.
comment in response to
post
That's fair, I tried to be as informative as possible but in the end, I know some people are just not gonna care about the problems. All we can really do is make sure the information is out there.
comment in response to
post
If I just said that, it would be exactly the same as all other posts we've seen, a statement without reference that would lead to the same misinformation I'm trying to solve.
This isn't written as my opinion or a simple statement but as a reference to whoever wants to understand what is happening.
comment in response to
post
The docs version if someone prefers to read it like this:
docs.google.com/document/d/1...
comment in response to
post
that way you don’t have to wonder where I got each information from.
- While I heavily despise GenAI, I made this thread the same way I’d make a scientific paper, as objectively as I could and backed up with evidence. (23/23)
comment in response to
post
Some disclaimers:
- I am an artist and (ex?)scientist and have some previous knowledge on how AI works despite never working directly with it.
- You might note some of the source links are repeated, that’s intentional so each part has its own references attached to it, (22/23)
comment in response to
post
Most likely, any game created by KRAFTON (the company behind inZOI) will have significant usage of GenAI, given the company heavily funds research in that field in order to save money, manpower and time in their game development. (21/23)
comment in response to
post
TLDR
On all their papers and websites, there’s no mention of creation of a proprietary model, only working on top of pre-trained models or directly with them, namely Stable Diffusion and Midjourney. (20/23)
comment in response to
post
These numbers are basically impossible even for Disney-sized companies to create fully on their proprietary data, let alone a small company or individual.
github.com/CompVis/stab...
laion.ai/blog/laion-5b/
proceedings.neurips.cc/paper_files/...
(19/23)
comment in response to
post
which includes 5.85 billion image-text pairs. Even after that, fine-tuning usually requires between 1k and 27k extra images in order to give the resulting images a specific style or direction. (18/23)
comment in response to
post
- Can someone actually create a GenAI solely on their own data?
Currently, I highly doubt it.
While lots of the big GenAI models didn’t fully disclose their training datasets, Stable Diffusion did and for its first version they used the LAION-5B dataset, (17/23)
comment in response to
post
model trained from the base models, that is, they are trained with additional data for generating images of particular styles or objects, on top of the pre-existing model, already trained on billions of images.
x.com/PlayinZOI/st...
stable-diffusion-art.com/beginners-gu...
(16/23)
comment in response to
post
with an entirely “company-owned and copyright issue-free” dataset creating what is called a custom model.
While for the general public (that’s not familiar with GenAI terminology), a custom model may sound like something totally made by them, it really is a (15/23)
comment in response to
post
- So KRAFTON lied in their tweet?
Kind of, what I assume they mean is they fine-tuned a previously existing (pre-trained) model, most likely Stable Diffusion v1.5 from the papers I’ve read but I can’t confirm that, (14/23)
comment in response to
post
arxiv.org/abs/2501.05803
arxiv.org/abs/2406.14272
proceedings.neurips.cc/paper_files/...
(13/23)
comment in response to
post
There’s no mention of contact with the people shown in these videos, or the video creators, regarding the usage of their image.
In a third paper, they discuss a method of fine-tuning pre-trained GenAI models with less data than usual. (12/23)
comment in response to
post
it “improves the pre-trained T2I model by up to 20%”.
In another paper, they present the MultiTalk dataset, comprising over 420 hours of talking videos in 20 languages scrapped entirely from YouTube, which they used to enhance 3D talking head generation. (11/23)
comment in response to
post
- Is KRAFTON GenAI ethically sourced (i.e. built with proprietary data)?
Again, not really.
In the most recent paper I found supported by them (from Feb 2025), they propose something called DAS, which in their own words is “a method for aligning diffusion models“ and (10/23)
comment in response to
post
but you’re most likely using the result of those as well since they said more than 90% of all employees are currently using AI on their work.
youtu.be/Hb-AIDMmCRI?...
ddaily.co.kr/page/view/20...
(9/23)
comment in response to
post
arts, voices and translations quickly and more efficiently, while cutting costs of paying more humans to do it, so, even if you don’t personally use the generative functions, we can’t tell what exactly was GenAI used for in the development stages, (8/23)
comment in response to
post
- Can I just play the game avoiding the GenAI features?
Not really. While you can avoid generating anything yourself, KRAFTON said themselves they use GenAI for their advertisement and game production, such as to generate game ideas, (7/23)
comment in response to
post
Some of the reasons mentioned by their team are saving money and time on voice acting and 3D modelling, as well as not having to wait for a new hire to learn about game plot before writing a dialogue since the AI would write it for them.
ddaily.co.kr/page/view/20...
(6/23)
comment in response to
post
- Why does KRAFTON use GenAI?
In their own words, to make game development faster and cut costs, as well as to innovate game production. (5/23)
comment in response to
post
The company behind the game (KRAFTON) has an AI-focused division and funds research in the development of generative AI in Korea.
krafton.com/en/more-expe...
youtu.be/Hb-AIDMmCRI?...
(4/23)
comment in response to
post
- First of all, does inZOI use GenAI?
Simply put, yes, in multiple ways. You can generate 3D assets and textures in game based on pictures and prompts and they’ve used multiple AI generated pieces in their advertising. (3/23)
comment in response to
post
I’m gonna do this based on what I’ve seen people confused about and the misinformation I’ve seen running around. It’s a pretty long thread but I tried to be somewhat concise while explaining everything properly.
I’ll try to include all of my sources. (2/23)