AI That Could Soon Replace Vector Artists [DALL-E]

Key Takeaways

who knew pikachu can be such a chat playing a grand piano or even riding a motorbike this is not until a couple days ago open ai decided to once again blow our mind away with its newest ai models just five days into 2021. we are off to a great start and in the ai part of course its name is inspired by the artist salvador dali and pixar's wall-e or wally is a trained neural network that can create images by associating concepts within the given text descriptions this has only become possible because of the success of gpt3 language model which can complete any text by literally learning nearly everything the internet has to offer which is absolutely mental so how insane is this dolly ai model well let's look at the results together right away openai did not release their trained model so we would be looking at the official demo instead as of why they did not make the model available i will elaborate on that later on at the very top we can see the text prompt that is used to generate the images below these are already being generated previously by open ai so we are just loading images here but surprisingly there are a ton of demos and you can go play around yourself on their blog but as amazing as it looks it is still cherry picked by another ai that they made called clip which is also released at the same time and it ranks the similarity between the given text and the image so what we are seeing now is already being ranked and the more samples that are being taken the better quality of the images become and it also shows how well dolly and clip work together as a dual especially as a sorting function but generally as we can see dolly is spectacular at generating images that show traits of style transfer and just basically anything that is changing the surface of something this also ranges from art styles textures and colors it is also good at generating a typical object when mentioned in the text prompt things that often appear in the data set show high consistency at generating the overall shape or structure of it like the stop sign however when it is about object symmetry like faces it starts to struggle as faces can be symmetrical and asymmetrical at times to further nick pick the false to make me look smarter you can tell that for some image generation it is bad at generating the exact precise description if the text prompt gives ideas that are abstract or unseen before the ai seems like it would fail to process every single detail and only succeed in just a few ones when the number starts to get too big it struggles at generating the right amount of it or the shape starts to deform and you can clearly tell that it is not a pentagon and if a word could have multiple meanings we can see different objects being brought forward and shown to us interestingly they also provided demos for objects that would contain concepts like reflections to test out how well the ai would understand it and as expected it only occasionally gets it right same goes to a lot of concepts like playing the piano would require you to sit outside not inside pikachu novel view generation is hard to be consistent but also can be proven successful in some cases and you can also tell it to generate illustrations too but it is hard to stay consistent generating symmetrical faces with asymmetrical features there is image completion alongside the text prompt too so generating images based on a specific part of the image can hugely affect the outcome like the mannequin here so they all look really consistent or to use the image to let the ai to refer to and it really is sad that i can't play with the trained model myself it just looks so much fun to test it but if it's actually released there are all kinds of let's say potential issues that could be a little problematic for instance just by looking at the quality of these illustrations i feel like it might be a big threat to a lot of small businesses like icon or cartoon illustrations distributing companies let alone photoshop artists or just artists in general who might lose some of their smaller deals there's just so much money that could be involved in this and could definitely slam pretty hard on the freelancer markets on the bright side this does give other people easier access to other creative aspects such as generating infinite amounts of the same texture just with slight variations maybe reusing assets in games would not be a thing in the future anymore because it is just too convenient to get something slightly different and just training gpt 3 can cost up to 5 million usd according to the internet so no one would probably invest this much money into something that is not entirely perfect yet also there are similar concerns whereby people were able to extract the training data from the gpd2 model there is a paper about this problem which showed how gpt2 was capable of extracting personal information about someone when a correct prompt is given note that dolly has the same concern but it still has the possibility of being able to extract personal ids or credit cards are unintentionally being included in the data set too there's just too many malicious things that you can do with this who knows if this dolly can photoshop or even deep fake just a text prompt away not even mentioning those not safe for work stuff and how well it can possibly do well no more with the speculation on what dolly can do but if it's me i just generate some dumb like a bunch of pout faces with a variation of my mako and they did release the code for clip so if you do want to try it out i'll link its blog and the collab down in the description the least block from openai would also be down there too and this video is sponsored by infinite red infinite consulting handles your mobile web and ai needs if you are looking for someone to bear your app visit with the link down in the description thank you guys for watching hope you enjoyed it openai is just amazing at blowing our mind away and this is just a wild start for the year 2021. shout out to mark schwinn and many other patreons that support my work through patreon and i'll see you guys in the next one