Image Completion AI - Predict Pixels Just Like Text Predictions [Image-GPT]

Key Takeaways

predictive text system where it can guess accurately what you want to say when you are texting someone by suggesting a few words on top of your keyboard for you to easily finish your sentence for some of them it strained upon a large amount of text beforehand to be able to learn what comes next after the zap you typed a white predictive text well what if you put the same logic where you predict or autocompletes the next pixel for an image just like when you keep tapping the predictive text where it creates a lone and confusing paragraph but instead produce an image so you may already heard of GPT 2 or even GPT 3 which are language models where you can produce an extremely convincing text which completes whatever you input into the AI model or generate a paragraph out of nowhere that is coherent so openly I made this language model where it dumped all the text on the internet into that model and have a super large amount of parameters which is like 175 billions brute force a near-perfect text generator that can basically do anything I think open iya has completed one of the most important steps towards perfecting data science which is basically brute forcing first like amass in order for mathematicians to come up with a mathematical formula first L required them to brute force and explore the area to understand how things work in that math room ok so back to the topic of outer completing or predictive pixels open there I recently published a paper where they use the exact same logic with gp3 text learning but on images instead the concepts is fairly straightforward and the AI just basically looks at all the previous pixels and add a new one based on all the previous ones then it takes in the newly generated one two and generates the next next pixel and which just continues until you generate a whole image for the input right now apparently they only provide 32 times 32 image size so that's all we can admire for right now so how exactly can it predict a crop image generating natural views seems to be the best fun for this AI if the crop is exactly where the birds are standing on the output has some really interesting results here's one where it even generated reflections - which is crazy the same thing goes to grassland it generates all kinds of things that could potentially be in the grassland Mount Rushmore has a body or leg sometimes which is hilarious I also trimmed another image of a bird and it logically generates back the birds tail and the branch which is amazing removing the grounds of a castle also made some pretty interesting generations and with the pixel art of a castle the AI gets really creative with it and has all kinds of cool castle designs for simple logos like Photoshop it sometimes can generate back the letter P and s but sometimes it just goes funny the same thing happens to Google's logo there are a few ones that are very close to spelling out Google or just a lot of doodles which is pretty funny I tried on various other logos too like Firefox one did really well generating back but Apple just kind of deford so how much can the present pixels affect the generation I use the socially awesome awkward penguin as an example if you cut at the half-way without the presence of the blue pixels it just completes the whole image in red which is pretty logical but if you leave a line of blue pixels the following generations will mostly be blue - for some more fun examples I tested on some meme formats it does not exactly produce a proper meme but the AI seems to know what exactly the objective might be like this is a human billboard no I mean penguin billboard yeah asked for the pictures of animals or human faces it runs really accurately like the Shiba right here you can see or even the guy in the butterfly meme is now wearing a suit what it's struggling is generating saitama's body yes this is amazing uh it has some problems generating from half of the phase to make it look not deformed or probably it's my problem a guy on Twitter called Michael phrasing hope I pronounced that right had some more fascinating results than me he was able to produce something longer which I have no idea to do for example the ocean one right here I'm guessing this is a boat and it is a really cool one in my opinion it's like an extending view into the deep sea that was generated by GPT image but the best result he had is on the human faces according to him it takes around like one hour to complete each of these results in the end knock on the lights is probably the best example of boot force out there opening yeah literally is testing the limit of brute forcing and this is going so well I love it if you want to play with this AI I'll link a collab down in the description which is made by Alfredo and join my disk or if you have any questions or join just for fun follow me on Twitter if you haven't and now see you all next time [Music]