HOME | DD

Description

No offense against AIarts critics in general. There are many real issues. But I really dislike and reject the request to ban it or/and call it "stealing", which is just a misleading term that spreads big misinformation about the mode of action to the public.

For example: Many people think, AI is duplicating things from images from its training dataset that fit to the prompt and mash them together. But this is far away from the truth. Latent Diffusion models connect semantic terms and sentences to their trained parameters. These parameters came from their training process, where they try to duplicate several image many times till another AI is confirming, it's good enough(Generative Adversial Network). They don't use the images from their training dataset anymore, when they create images from prompts. This isn't even possible, because they juse use a memory space in size of 2-4 gigabytes(Thats why SD can work on your own PC with enough Vram). This is not enough space to contain all 4 billion images from the Laiyon 4B dataset.

To anyone who is saying, that this have to call "stealing" i ask this question:

If its then also stealing, when a human try to duplicate existing images to improve his own skill and create with this skill new images after that? Are these new images stolen works?

_____________________________________________________________________________________________________________________________

-Used characters generated with Midjourney by UtauRadon.

-Edited with Stable Diffusion and Gimb.

Description

No offense against AIarts critics in general. There are many real issues. But I really dislike and reject the request to ban it or/and call it "stealing", which is just a misleading term that spreads big misinformation about the mode of action to the public.

For example: Many people think, AI is duplicating things from images from its training dataset that fit to the prompt and mash them together. But this is far away from the truth. Latent Diffusion models connect semantic terms and sentences to their trained parameters. These parameters came from their training process, where they try to duplicate several image many times till another AI is confirming, it's good enough(Generative Adversial Network). They don't use the images from their training dataset anymore, when they create images from prompts. This isn't even possible, because they juse use a memory space in size of 2-4 gigabytes(Thats why SD can work on your own PC with enough Vram). This is not enough space to contain all 4 billion images from the Laiyon 4B dataset.

To anyone who is saying, that this have to call "stealing" i ask this question:

If its then also stealing, when a human try to duplicate existing images to improve his own skill and create with this skill new images after that? Are these new images stolen works?

_____________________________________________________________________________________________________________________________

-Used characters generated with Midjourney by UtauRadon.

-Edited with Stable Diffusion and Gimb.