Trancor How to spot Generative AI (even if it has all 10 fingers and toes)

🎯 Загружено автоматически через бота: 🚫 Оригинал видео: 📺 Данное видео принадлежит каналу (@TrancorWD). Оно представлено в нашем сообществе исключительно в информационных, научных, образовательных или культурных целях. Наше сообщество не утверждает никаких прав на данное видео. Пожалуйста, поддержите автора, посетив его оригинальный канал. ✉️ Если у вас есть претензии к авторским правам на данное видео, пожалуйста, свяжитесь с нами по почте support@, и мы немедленно удалим его. 📃 Оригинальное описание: I’ve seen quite a few “how to spot ai“ videos that were simply talking about errors in the construction of the image. Too many fingers, weird eyes, weird faces, etc... But AIs are getting better; And I’m not hearing anyone talk about the errors baked into the training data itself. Update : Given the comments, I’ve started to make an ai detector python program that can run from command-line, or basically, automated image checking of the issues I talk about in the video. But also with a window you can launch and view image statistics on. I’ll update with a GitHub link when the basic version is up and running. But, given this is youtube, we know how that will go. What makes an AI image stand out? -The *Blue* channel on real Digital Photos will *ALWAYS* be more noisy than Red and Green channels. -Consider the hardware used to take the “photo“ or make the “digital painting“ -JPEG compression is image wide, chunks of slightly off tint/hue colors, with purple and green tendencies. -JPEG compression should be more visible around curves in an image, but in 2x2 into 4x4 pixel blocks. -If it doesn’t have *BOTH* JPEG issues, then its AI. -So, non-square JPEG artifacts in PNGs with just random “JPEG Green & Purple“ tinted pixels; -Or odd noise patterns mixed throughout the image (per object noise patterns) -Red, Green, & Blue channels all have the same noise means its a digital image, not specifically AI, but if its a “photo“ with uniform channels, it’s AI generated. -Even “digitally painted“ AI art styles will add inconsistent noise to every object in the scene, which an artist’s general Style AND Approach/How to use the program wouldn’t do; when using photoshop, procreate, gimp, etc. -Lines crossing 64x64 pixel boundaries tend to not stay congruent. (12:04, as I mention it, look at the tree-ground line between the horse’s legs on the right of the image) -If you spot stochastic sampling (like pixel gaps between similar colors randomly in the render), I’d put my money on cgi render, not ai Note / Mess-ups - 5:00- Took the photo’s last year, I just suck at keepin’ my damn hands still. 6:00- I’ve been on sets and productions, but I’m a technical artist. I’m what comes after film or rendering or need to create pipeline tools to work with film or camera signals for work. Mostly writing code to work with images/video. 7:00- The sharpening “appears“ like JPEG compression, but the lack of the chunking I was talking about is the indication of “style“ over “file compression“ over “ai blatantly using random noise all over the place“. There is visible JPEG artifacts in the tan/beige area, but again, looks like “attempted style,“ not “JPEG compression“ 7:15- There is potential for straight line without jpeg compression, but you’d notice a blocky edge where the curved part meets the straight line. So, the area above what I have selected. 7:55- Random JPEG noise but no distinct squares. 9:18- *Horizontal lines I made the music in Aiva, and hold the license. Its an ai music generator you can customize pretty deeply.