Raw LLM Responses

Yeah, about that... Glaze and Nightshade mainly rely on attacking the Variational Autoencoder (VAE) by shifting concepts in the Latent Space during training. Most AI guys just train LoRAs (Kind of like small neuralnets trained on specific style/concept) which usually doesn't involve any VAE modifications at all. So, trained images without any modifications would still look like the trained concept or style, but with some noise artifacts here and there. But if we go even further with this concept, just setting Min-SNR-Gamma (Minimum Signal-to-Noise Ratio Gamma) to 1 will make Diffusion models ignore certain features like noise and focus on more pronounced features. But even if you train a model from scratch, there's some research techniques, generative or not, that can uncloak artworks. From stuff like IMPRESS or GLEAN to just simple upscaling and Gaussian noise. I think one of the biggest mistakes Glaze team made is not obfuscating their code, which means that anyone can look through it and see how they've done their adversarial noise. Because the noise is just an algorithm and not a generative model that dynamically applies it, people can reverse it if they really want it. I would be happy if there actually were some reliable ways to protect your art, but current projects just offer some false sense of security. Unfortunately, for now, best defence from AI is just not posting your stuff publicly.