Is Midjourney art directable?

i started experimenting with mid journey a couple of weeks ago it was a lot of fun seeing the images the ai would come up with it wouldn't always make the best decisions and the images weren't always the best but overall it's exciting to see this new field developing right before our eyes so with this video i wanted to explore mid journey a little bit more to see how much of an input i can have in the final image in my previous test i relied heavily on the random decisions of the eai that's of course part of the charm of this type of image creation but in order for ai tools to become more useful to artists we have to have some way of steering things to a specific direction so with that in mind i started experimenting with prompts and prompt settings in order to try and get closer to what i had in mind did i succeed let's find out [Music] the first thing i wanted to try out was the image weight command it was suggested in the comments of the previous video and after reading the manual i figured out how to use the correct syntax so what does this command do we can point the ai to a specific image on the internet and tell it to use that as a visual inspiration but as the manual says we're not really building on top of that image it's just used as a general reference i decided to use one of my own 3d renders as the bass i want to see how the ai would interpret this robot in a cyberpunk manner so my text prompt reflected that but even though the image weight was set quite high in order to influence the outcome as much as possible the results i got back were not what i had in mind the main reason i was disappointed by the results has to do with the fact that the ai images don't really have any similarities to the base image i do see the use of a similar color palette but other than that there's not much else for example i would have liked to see the same three quarter angle or some more unique characteristics of the form like the multiple wires on the robot instead we get a more bland and uninteresting image first off it's a frontal view which is not the most exciting angle the ai did try to copy some other elements but it's nothing to write home about for example the ai robots do have similar blocky features as the original robot there's also an attempt to use screws in some parts of the head but yeah overall the ai images aren't really close to the original to see if i could influence that a little bit better i tried an even stronger weight before the image weight was set to around 70 percent and now it's all the way up to 90. and in order to steer the overall style a little bit closer to what i had in mind i also used hajima surayama as the style to copy sarayama is extremely well known for his shiny robot characters and there's a ton of images of his work on the internet so i knew that the ai would not have a problem finding references unfortunately the results weren't that great these are definitely not sorayama robots and on top of that they don't even look close to my image i wasn't sure why that was happening and initially i thought that i might have confused the ai on one hand i was using the image weight command and was asking to use my image as a reference and at the same time i was asking the ai to copy the style of sarayama so it might have been a user error as a last-ditch effort i decided to adjust the prompt a little bit more and i only focused on delivering a robot character in the style of surah yama unfortunately vai had other plans what i got back was more of a depressed surah yama the robots were suddenly not shiny they weren't vibrant and they were way too human-like these look more like a soviet artist's interpretation of a robot rather than a surah yama robot just as a reminder this is suryama super shiny surfaces dynamic forms and characters that have anthropomorphic features but are clearly robotic and this is the ai result there's quite a bit of difference there so yeah my first attempt at controlling the ai wasn't going as planned just to make sure that my prompt wasn't throwing things off i took out the word realistic just to make sure that this word wasn't the reason i was getting more human features in there but that word was certainly not the issue the results i got back were even worse than the ones before but i guess that's to be expected the more descriptive we are the more guidance the ai has the less information it has the more stuff it has to come up with but there's always this fine line we need to tow if there's a ton of descriptors the ai might get confused if there's too few a lot of things will be left to chance after this huge fail i decided to just experiment with the ai in a more loose way just to see where things would take me i used more basic and vague descriptors in my prompts and the results were certainly much better i think there's two reasons for that first off i think by keeping things loose and vague the ai has more room to play but most importantly i think it also has to do with a user's expectations in my previous attempt i had a very specific image in mind and i was trying to get as close to that as possible but by coming into it without a specific outcome in mind you're more willing to accept whatever image the ai will come up with the manual says that if you're looking for something specific you should try to be as descriptive as possible but i believe that for me at least the loser approach works much better and as the team says in the manual using strong feelings or mystical sounding themes can work great for example using prompts like sense of awe or the will to endure and that's what i was finding as well whatever the case may be though these ai assisted images are still a great starting point for an artwork whenever i get an image back from the ai i can see the potential it's immediately obvious to me how the image could be improved by getting rid of some bits modeling some others a little bit better adjusting the lighting and composition the skeleton of a great image is there it just needs some refinement and extra love and that's the amazing thing if you think about it we essentially have a storyboard artist at our fingertips that can spit out images in a really short amount of time are they work of arts on their own no but they can be after some heavy editing i still though wanted to explore a little bit more of the different styles the ai could output just to see how well mid journey could imitate things and to do that i used artists i'm already familiar with gregory croutson jeff coons simon salahag and a couple of others the results were surprising i first started with gregory kreutzen i have a whole video about this artist but in case you're not familiar with his work he's a photographer shooting very cinematic and dream like pictures his images are mostly shot in suburbia and require a ton of pre and post work it's not the easiest style to copy but i was hopeful that the ai could manage to output something interesting and i must say i was quite impressed when i got back the results from the first prompt of course it's nowhere near a final state but what i'm interested in is a first rough sketch something to get the creative flow going and this one has all the elements of a good sketch the overall mood is spot on the suburban setting is correct and the color palette is in the right range everything feels to the point since i was successful with this one i decided to play a little bit more with prompts using gregory christian as the base style and the results were equally good as we've seen with the previous video the ai has some issues with the human form and you can barely tell that the man is supposed to be floating in midair but still as a sketch it's perfect it conveys the style and mood pretty well i then tried another favorite artist of mine simon stalinhack i'm sure most of you know him already but in case you're not he's a painter slash illustrator usually drawing jaw dropping futuristic wastelands suburbia is also heavily featured in his work and his images are as cinematic as scrutin's even though he's using traditional media to do that the ai results were quite good disregarding how some things look for example the design of the robot i really like a lot of the elements in the picture the composition is quite strong the color palette is very reminiscent of stellenhex work and the overall mood is spot on the ai keeps doing some of the weird things i've seen to do before but it doesn't matter there's a good base there for example the robots are made out of houses even though i didn't really ask for that or what i assume is street lights are just floating in space but yeah these things can easily be redefined when working on the final artwork so with that first attempt being a success i had to try some more prompts with stalinhack as the style base [Music] the usual airs are there like things being in the wrong place but the ai images feel like a nice starting point for an artist-driven artwork [Music] i then tried several different artists with differing levels of success copying david hockney was kinda semi-successful the color palette is kinda there and the human forms while not really hockney-like we feel like a very distant cousin of his so vai gets a pass for this one trying to imitate jeff koons on the other hand was a huge failure mostly because the ai couldn't imitate the shiny look of coon's famous dog sculptures and also because the ai doesn't seem to know how a dinosaur looks like it's a common occurrence with mid-journey there are things that it just cannot figure out how to draw in order to help things i decided to be more specific so instead of dinosaur i used the word t-rex but things didn't really improve the t-rex looked more like a child's interpretation of a t-rex imitating geiger though was a bit more successful the forms still don't look like dinosaurs and the images don't really use the shapes geiger used in his work but it's an interesting set of images to look at as i've mentioned multiple times before they're good enough as a starting point finally i try to see how well the ai would do with william eccleston he's a photographer known for his slice of life banal type of images he has a very distinct style in his colors and his compositions and he's one of my favorite photographers i used two of his images as a bass but i did not use the image weight command i just tried to describe the images and see what the ai would come up with the framing on both of ecclestone's images is very creative and i'm sure it's something we won't see from the ai majorence framing is not as adventurous so here's how mid journey interpreted the images even though they look nothing like eggleston's photographs they do have some nice properties i think if i was a bit more descriptive the results might have been better but i don't think we would ever get as close to the real photos either way it's interesting to see how ai deals with the subject matter descriptors like hard shadows and sunlight really help with the style of the image so always keep that at the back of your head when you're creating yours these are really nice descriptors to use of course if it fits what you're going for it's also hilarious to see the ai struggling with simple things like how a bicycle looks like or seeing the nondescript contents of the plate on the dinner table i'm sure if i redo these images in a couple of months the results will probably be 10 times better from what i understand the ai is still being trained so probably we will soon get to see some jaw dropping results [Music] even though i didn't manage to control the output of the ai i feel that i did get some better understanding of how things work but my opinion hasn't changed since the last video i still think that it's a really awesome tool for sketching out ideas i'm planning to create some artwork based on these ai designs so keep an eye for that video hopefully it's gonna be sooner rather than later but what do you think of the results were you pleasantly surprised or were horrified by the images let me know in the comments below take care and i'll see you on the next one [Music]

Channel: Dimitris Katsafouros

Views: 28,063

Keywords: AI, neural network

Id: 6vVnGuL1o58

Length: 13min 19sec (799 seconds)

Published: Fri Jun 17 2022