Drawing Marie Antoinette ✦ Ohuhu Alcohol Marker drawing 

I’m not sure how this is relevant to the video, but according to the website they _claim_ that it’s still effective. However, this claim is somewhat questionable as Glaze and Nightshade are designed for architectures based in U-Net. I’m not going to explain it too much in-depth since it’s a very deep subject, but basically it helps to better map annotated training data in lower-dimensional latent space to allow for more efficient utilization of annotated samples, allowing for annotated segmentation to be utilized in localized areas and making it so that fewer images/training data can be used-basically, it allows for faster and better generalization, allowing an AI model to learn concepts more quickly and with less data. Ultimately, all modern AI image generators-so-called “diffusion” models-generate images from random visual noise (like the static on an old TV), rather than being a “collage” like is commonly thought. They do this by first turning training images into full random noise, learning the attributes and patterns from said images, and then reversing that process using new, random noise to create a novel image. Why the nature of U-Net can create a problem for Glaze and Nightshade is that it’s already dated. In SD3 for example, U-Net is completely replaced with (MM-)DiT, or “diffusion transformers.” Basically, it is more efficient than a U-Net at creating better images, as DiT helps the text better match the images in training and creates a better “flow” between elements. Glaze and Nightshade, under their current coding paradigm, primarily attacks U-Net. While it is effective against training on the most common types of model architectures (SD 1.5-based ones especially, as a lot of people can run it and it’s easier to train; for SDXL-based models, Glaze and Nightshade have also proven effective), when a new base model is being created, often there are new technologies employed, like (MM-)DiT. Therefore, Glaze and Nightshade always end up being one step behind. At this point, Meta and other companies training on AI aren’t going to be using UNet; Glaze and Nightshade, at least right now, are mostly going to protect you from those who are inexperienced (most LoRA trainers) are who can only train/know how to train on semi-outdated model architectures. Advanced model trainers, like Meta on Instagram, are much more knowledgeable about cleaning their dataset and filtering out Glaze and Nightshade images, such as running scripts or using SUPIR to clean up the noisy data and slightly upscale it (which would be effective against a screenshot taken, say, with another device or of the screen itself). Further, Glaze and Nightshaded images can actually be beneficial to model trainers, and can be used to intentionally _improve_ model quality-even if they prove to be effective against the model. This is something known as _Negative Reinforcement_ or _Negative Conditioning_. This is basically when you train a model on images you _don’t_ want, and then insert it as a negative embedding or something of the sort to get improved quality outputs than the model would otherwise create. To sum it all up, for the vast majority of extant models (older models like SD 1.5, SDXL, and especially small~medium-sized LoRA models) and against the vast majority of model trainers (who don’t have huge teams or huge resources), who are still using somewhat outdated model architectures due to computational requirements or just their pre-existing knowledge, Glaze and Nightshade will be effective. However, for most new models, and against almost all corporate or “base” models, Glaze and Nightshade are very unlikely to be effective. That effectiveness goes even down with closed models like Firefly, Midjourney, or DALL-E 3, whose closed architectures make it very difficult for the Glaze and Nightshade researchers to find which areas to attack. Further, such images can be trained via _Negative Reinforcement_ to create better AI-generated images.

Yeah! I’m working on two huge projects, one contract work and one being my book manuscript, and I don’t realistically have time to upload videos- and writing scripts, recording, and editing takes a really long time. I’m trying my best to keep at least a once a week upload while I’m working on these projects but when i have time to actually edit again, I’ll bring back voiceovers in my videos.