Protección del artista de glaseado y belladona

Both tools come from the SAND Lab at the University of Chicago, led by Ben Zhao. Glaze was released in 2023 and Nightshade in January 2024, and both are free. Glaze also has a web version, WebGlaze, for artists without powerful computers. Glaze addresses style mimicry. Someone can fine-tune an image model on a few dozen pieces by one artist, then produce unlimited work in that style. Glaze adds a perturbation, a pattern of small pixel changes, that makes the artwork appear to a model's image encoder as though it were in a different style. A model fine-tuned on glazed images learns the wrong style, while a person sees little change. Nightshade is offensive rather than defensive. It targets training on scraped data. A shaded image of a dog is changed so a model perceives features of something else, such as a cat. Because each concept appears in relatively few training images compared with the whole dataset, the researchers argued that a modest number of poisoned samples could make a model respond to "dog" with distorted results. The aim is to raise the cost of scraping without permission. How well do they work? The perturbations are tuned against specific model components, and that is their weakness. A 2024 study by researchers at ETH Zurich and Google DeepMind found that simple methods, such as noisy upscaling, could remove Glaze-style protections enough for mimicry to succeed. Later work, including LightShed in 2025, trained detectors to spot and strip poisoning perturbations. The Glaze team has disputed some findings and updated its tools. The key misconception is that protection is permanent. An image published today cannot be updated later, while attackers and models keep changing. The tools can also leave visible artifacts, especially at high intensity.