- 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)
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)
- 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)
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)
Comments
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)
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)
https://x.com/PlayinZOI/status/1833030670219903067
https://stable-diffusion-art.com/beginners-guide/#What_are_custom_models
(16/23)
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)
https://github.com/CompVis/stable-diffusion/blob/main/Stable_Diffusion_v1_Model_Card.md#training
https://laion.ai/blog/laion-5b/
https://proceedings.neurips.cc/paper_files/paper/2023/hash/a5755ccd0efeca8852ae0a1193f319f6-Abstract-Conference.html
(19/23)