Raw LLM Responses

So, this seems to specifically try to confuse image classification models, such that descriptions generated might end up being "blueberry muffin" instead of "dog". Image classification models trained on nightshaded imagery will become more robust and learn things closer to "what actually constitutes a muffin / dog". Wrongly classified images could end up hurting models trained on them if it is a significant portion of total artwork which ends up with wrong descriptions, but if it is a very small percentage, it will just represent a small bump in the optimizer momentum which does not meaningfully change model weights. This will over time lead to image classifiers which perfectly understand the concepts and don't just learn very basic features to figure out what something is. After that point you'll need to fully destroy your artworks for them to be classified wrong. I suppose one could try seeing whether they can confuse the better image classifiers using this software at this point whatsoever (Llama 3.2 90b or ChatGPT for example). If it understands the images just as well as when they aren't nightshaded, it is already fully useless. What you do accomplish is primarily that some image search engines may not find your artwork anymore. Especially if it actually works, your images would become undetectable under the correct keywords. Your nightshaded bullshit could also make any image search in existence worse, when I search for dogs and find tons of images of muffins instead. I'mma apply nightshade to some images real quick and have pixtral describe them with and without nightshade. Nightshade intensity maximal (this mangled the output image visually, btw) Result: Generated descriptions are almost identical despite the image being visually mangled by nightshade at maximum strength. Occasionally, the description gets slightly worse. Try it yourself if you don't believe me. If you give both images to a large model at once for comparison it'll be like "what do you mean. This is the same image, but one has slightly more compression artifacts" In conclusion: Nightshade accomplishes nothing. If you look closely, you can definitely tell the visual artifacts in your images. You're just making your art worse for absolutely no benefit whatsoever and much like those spamming their generated imagery everywhere, you primarily would damage the internet as a whole if this worked in the first place.