Raw LLM Responses

Whatever that AI bro done is besides the point. NightShade has already been defeated. It is dead and useless as of Wednesday. A group of researchers just put out LightShed which has a near 100% success rate at detecting nightshade and almost as good at undoing it. I would strongly recommend any artist trying to protect their work use Glaze instead of whatever ends up replacing nightshade. Glaze can definitely be defeated too, but it is a lot safer option for legal reasons. What makes Glaze a fundamentally better idea is that developing a tool to break Glaze means you are aiming to pirate content. Institutions are not a lot less likely to green light research into cracking DRM, and companies will think twice about internally developing a tool that's very existence can loose them a lawsuit. This is kind of besides the point, but at 2:50, you seem to have a misunderstanding of what a LoRA is. In simple terms, LoRA is a way to fine tune a model on a machine that normally wouldn't be capable of training a model that size. If a LoRA works, then a full fine tuning or training would definitely work. In more technical terms, instead of training the weights W, LoRA uses W`=W+AB where A and B are learned vectors instead of the large weights matrix. The end result is a retrained model using a less capable but cheaper training method. Nightshade should protect against a LoRA(or at least one with no counter measures) because a fine tuning is retraining with a high bias for the samples you are tuning for, so nightshade is extra affective against LoRAs because they are weighted extremely strongly compared to the rest of the dataset.