Scaling Midjourney AI art for PRINT with Topaz Labs Gigapixel

this is the 30 second version mid journey is a machine learning ai system specifically designed to turn words into images and it does a pretty good job but they are small meanwhile topaz gigapixel is a machine learning ai system specifically designed to scare images up and it's pretty good at it and we can take those two differently specialized ais and make them fight no no we can make them work together to make print size images [Music] i'm a big fan of print i really believe that if you're a digital artist or designer or a photographer the one of the best things you can do is print your work and stick it on the wall and then just let it sit with you you get a different connection to the work when you see it physically under different conditions throughout the day and if you're in the process of developing some artwork you start to spot things that you that kind of bug you in ways that you don't necessarily see when they're on the screen i think it must be the same as like when you print out documents you spot typos in the printouts that you can't spot on the screen as well i wanted to see if i could print the results of a mid journey ai system which is like the dahly one as well that other people might be more familiar with the idea of asking an ai for images of something and then having those up as prints around around in my studio kind of tickled my fancy i have another video about mid journey linked up there somewhere but a quick recap is it's a system that's been trained on a lot of images for concepts like a tree and work by lots of different artists a tree can give us something like this and a tree by jeff koons is something like this it doesn't understand what a tree actually is but it knows what images that have been tagged or titled with a tree what they look like and how it does that is roughly like this let's run another tree and we can see as it iterates over the images until it gets to the end results roughly what it's doing is making lots of very tiny images just like a few pixels across filter just random colors until it's got a few hundreds if not thousands of them and then it scores each of those tiny images for which one looks most like all the images of a tree it has and it picks the winner so if it's produced an image full of lots of red pixels that will get rejected but if there's one that's kind of slightly dark around the edge lighter in the middle and then with a little blob in the center that's the most tree-like at that tiny small scale it then doubles it and then fills in all the missing pixels with some more more random crap basically and then asks the same question which of these slightly bigger ones looks most like a tree does the same with uh jeff coons a tree by jack coons which these tiny images most looks like a tree and also looks most like an image by jeff coons and then it scales it up and scales it up scales up always scoring towards the result until you end up with something that looks both like a tree and an artwork by jeff coons so let's do it again with something more complex like an ancient landscape filled with a field of flowers and snow topped mountains with fractal pink fluffy clouds and spaceships so it has to score against landscape field of flowers snow top mountains fractals pink fluffy clouds and spaceships um so it'll take one of these small tiny pixel groups and it will score highly against that and then scale it up and scale it up what i'm getting at with all of these things is it's quite computationally expensive to keep generating these hundreds if not thousands of images fling their random pixels and then scoring them so they they keep things smaller when you type into the prompt you get quite small results because it wants us to be the deciding factor for which one is good before it commits to then making the bigger version of it when you pick one it'll then go through the same process of that but it's focusing all its computational energy on just that one for the next couple of sizes computers aren't really fast or cheap enough to do that at scale yet you can use all that processing power to serve lots of people small images or a couple of artists doing bigger images but even with that every time you double the width and the height of something it quadruples the pixels and it still has to generate multiple versions of it so it gets so like exponentially more expensive every time you jump up a scale so that leaves us with small images that look good on the screen or on the phone but not really for print my image is about 2048 pixels by 1024 and when we're printing we're looking about 300 pixels per inch the minimum you can get away with is about 150 so with a 2000 ish pixel uh image to work with we can print around 7 inches which is about that that's not really war size i prefer it to be about 20 inches by 10 inches maybe even 30 inches but that's starting to push the limit practicality the print i have here which is on archival paper with archival ink which means it's good for about 200 years not fading the really great thing about printing in archival quality is the blacks are really good the details are amazing the colors are really sharp but that and also means that if there are pixels in it you're definitely gonna see the pixels so solution one is just send the image to the printer and let the printer software scale it up it's not great two is use a package like photoshop or another general purpose image package to scale it up or three use specialized image scanning software which is what topaz gigapixel ai is just as mid-journey is an ai designed to generate images from words and nothing else gigapixel is an ai designed to scale images up and nothing else it's been trained on hundreds of thousands of photos the type of things photographers take photos of so like landscapes and mountains portraits and buildings and architecture the typical photographic subjects but less so abstract art concepts so when it scales things up it doesn't have to generate like a hundred or thousand variations and score them all on the big computer to fill in the pixels um instead what it's doing is it's going how do i make these mountain pixels look more mountainy it has one job to do and it doesn't have to be on a big powerful cloud computer system it can actually run that on my laptop so let's give it a whirl with our pinky spaceshipy mountainy landscape i've loaded it into the app and we can have a look at the preview or move around and have a look at different bits of it so you can see we've got fields here that have what look like flowers in the foreground and um it's quite good because just these few pixels on the left it's decided to turn into this because it's been trained on landscapes it's kind of going oh this is a landscape bulletin there with clouds up here again it's looking at these and it's probably going well these quite cloudy i'll make clouds the spaceships up here not quite so much mid journey's going oh this is kind of a spaceship over here we've turned it into something but it's definitely better than the one on the left or scoot around a little bit over here back to the landscape again so let's let it do its thing and it's filled in this grass and flowers it's quite amazing considering that this over here on the left is what it it had to work with let's have a look at a little bit more back onto typical landscape it's quite good and then up here as well there's one last thing you manage to turn this set of pixels into this over here so this is the one from mid journey loads up into photoshop so 2048 pixels across and now i'm going to scale it up image image scale and we're going to use 600 and we're going to use the automatic uh way of doing it there's a number of different choices let's just let photoshop do its thing so that's what it looks like scaled up now i'm gonna paste the topaz one straight over the top so this is the topaz version you can already see it's so much better if i go down to something like fields over here turn topaz off you can see over here all this field detail that photoshop's tightened on this little circle down here turn topaz back on it knows what fields look like it doesn't really know what this like sticky up area is but it's giving it a go if i just move across over to this side we can see again just a little bit the mountain here just quick toggle on off normal scaling topaz scaling just go up to this bit here steering photoshop is difficult off on let's just zoom in one more time this is 200 now that's photoshop that's topaz so even though it hasn't turned it magically into a spaceship it certainly made it high enough quality that i can print that out and be quite happy that it's going to be okay which is what i did with this image right here if i grab my camera you'll be able to see the difference between the topaz scaled one and then the just the blown up and i've brightened the colors up a little bit on this one [Music] so hopefully you can see the difference between the two of them topaz gigapixel ai goes up to 600 percent which takes our original image to 12 000 pixels wide just over and at 300 dpi that's about 40 inches or a meter in sensible measurements now i'm not sponsored by topaz labs i pay for the software with my own money or i guess rather me as an artist because being an artist is kind of my job and that involves this sort of thing so business suspense anyway it boils down to this at the moment pick your ai that's best suited to the purpose and then chain them together i can see a situation where say i wanted the digital frame on the wall where the processing get the image up to a certain size before then the second specialized scaling ai takes over to take it the last couple of steps to make it bigger and then bigger again and that's it hopefully you found this useful if you have images from one of these machine learning generative art systems so they're trying to figure out how you can generate bigger files to print them i think this is a good way of doing it there's a trial version you could probably use that if you've enjoyed this video hit the like button subscribe leave a comment all that good stuff because it really does help we ticked over a thousand subscribers the other day which is huge for me um and you're all awesome i will not see you next week because i have a thing on i'm taking a holiday but i'll be back the week after that see you next time [Music] [Applause] [Music] you

Channel: Dan Catt

Views: 62,314

Keywords: gigapixel, gigapixel ai, how to use midjourney, how to use midjourney ai, mid journey, midjourney, midjourney ai, midjourney commands, midjourney tips, midjourney tutorial, topaz gigapixel, topaz gigapixel ai, ai, ai art, concept art, dall-e 2, digital art, openai, videographer, vlog, vlogger, vlogging, youtuber

Id: JQFPbB91Cdc

Length: 12min 5sec (725 seconds)

Published: Thu Jun 23 2022