Photoshop Detector AI Is Useless

Key Takeaways

you see a site of instagram has become a platform that demonstrates how social media can really create a false perception of reality and in some aspects it has brought harm to a lot of people especially kids and so there are places such as our slash instagram reality that are born to compare the instagram post versus the reality but on the flip side people that unrealistically photoshop or use filters mostly just want to present the best image of themselves to others but the consequence is that it creates a never-ending cycle of toxicity moreover when people unrealistically portray themselves online it sometimes backfires for instance celebrities can be criticized a lot and for some odd reasons like not being truthful to their fans because they edited their photos what is this false advertising and it even became trendy to expose celebrities that photoshop their photos for a good period of time and this is where the ai paper that i've been wanting to revisit comes in fal detector short for detecting photoshopped faces by scripting photoshop was published around september 2019 and it went viral some may say it became a tool for witch hunt but at the very least it has the ability to expose people who have photoshopped their faces not that it is always accurate to be more precise a fail detector was trained to detect facial warping it is also one of the most common ways to edit faces in photoshop where you literally just pull their faces and this ai has two functions one is determining the likelihood of a face being photoshopped and another one is evaluating the region that may be photoshopped and also producing an unwarped face based on the suspicion which is really cool so in order to train this ai the author used a data set of images and for every image in the data set it is passed through an algorithm to randomly work the faces which means if the starting data set has hundred thousand images after randomly warping each of the faces 10 times the data set will now have 1 million training data to use but would i say this ai is trained to perfection not really well my argument is not because this ai's architecture is not precise enough or that the training data set is not a correct representation of the real world photoshopping it is more of how taking photos itself can vary based on the machine specs you choose to use and many other factors that i am not mentioning in this video i think most of you already know how there is an adjustable thing on a camera called focal length and the smaller the focal length the better someone looks so let's say someone uses a different focal length to take their photos it seems that my results show the smaller the focal length the more likely the image would be predicted a higher percentage of being photoshopped well of course these images were never edited in photoshop it is the focal length that changes how the same face can look differently and still be suspected of being photoshopped and have regions of jaws being predicted as photoshopped well then which one of these faces would you say is the most accurate representation of the person they all are but just presented at a different focal lengths okay i think the problem lies in the fact that to evaluate something so delicate like photoshopped faces it is definitely a much more difficult task than most people would expect it to be it is not just looking around the pixels and guessing if they were stretched or not there is more than one factor that can affect how a phase can look like in a 2d environment so to create an ai or a tool that can fully detect what is actually being photoshopped or if it is just how it looks like being taken on the given camera or machine is actually kind of hard and don't get easily fooled by a simple heat map it may look like it's showing the region that's like oh it is photoshopped heavily or lightly well that is not necessarily the entire truth but most importantly you also need to refer to the value of how likely this image is being photoshopped too like this image is being predicted a probability of 99.27 percent that is photoshopped but look how lightly the heat map was drawn so don't just go by looking at the visualization like the other day i saw an old video demonstrating fal detector on a phone app and it creates a heat map for the results of the detection a lot of people may just jump to the conclusion oh these are the regions that are heavily being photoshopped well no we have been given absolutely zero context it might as well be a detector of the region where the jaw exists if you don't believe me let me run the same exact image that video has my low resolution of the same image produces a probability of 0.22 of the image being photoshopped and the 4k image that i found online produces a probability of 24.84 of being photoshopped while the heat map region is totally off from what the video shows what i am trying to say here is that don't interpret the results immediately in a way that your mind wants to believe like oh kim kardashian definitely photoshopped herself and this heat map just shows how she is shamelessly doing so don't do that because first the result might not be accurate and second the result might not provide a full view like the probability of this image being photoshopped is not even being towed at the end of the day just be a bit more perceptive take everything with a grain of salt and don't jump to conclusions this ai is not particularly good or bad or toxic it is just a rather interesting research someone has decided to put their time into so the rest of us can enjoy and discuss or even learn about the knowledge the author has gathered through this research paper that is shared with all of us for free if you want to play around with this ai i'll link my installation tutorial down in the description and this video is sponsored by infinite red infinite consulting handles your mobile web and ai needs if you are looking for someone to build your app visit with the link down in the description if you have any questions about the installation or want to share your results or thoughts you can head over to my discord channel i am quite active there and big shout out to mark schwinn and many other patreons that support my work through patreon and lastly follow me on twitter if you haven't and i'll see you all in the next one