Raw LLM Responses

You can not poison AI datasets. Thats not how things work. As someone who trains AI. The issue is that it hurts real fans and supporters who are getting supbar quality art, that they even pay for. While it doesn't do anything to us who train AI. And the increased complexity in the images just bloats file sizes. AI datasets are noisy and sub par already. They are trained on the principle of "more data no matter the quality, then fix it in post". There are a few ways to prepare the data set. The easy way is to just scale it down and auto crop. This removes allot of the noise already, since the patterns are at pixel level and like 3/4 pixels gets removed/merged. Then there are several tools to clean images, waifu2x with real-esgran is free. The way AI trains is literally by adding noise to images. Then comparing the before and after, it learns how to make a noisy image less noisy. This is also why you can clean noisy images and remove the noise. Even if the image is noisy you can still compare it to a even noisier image and get useful training data from it. And you average it across multiple images, most if not all that are not poisoned. So when it comes to scrapers if 10 images out of a 100 million are poisoned, it doesn't really matter, since half the dataset is basically garbage to begin with. JPG compression and subpar lighting when using phone camera is way more detrimental, but we can still make the quality you can see despite that. For us who hand pick our images, we can make sure its only the best quality and can just circumvent any poisoning.