How to stop getting deformed Midjourney images...

have you ever opened mid journey and tried to generate something like this a full body photo of a man wearing a bright green tux standing on the red carpet surrounded by a crowd sharp detail and then you randomly get corrupted nonsense as a result mid journey is supposed to be better than this right where's the problem here if you've made enough AI art or watched a few of my tutorials the next thing you probably do is go through a checklist in your head it seems like a good prompt it's clear visually descriptive it's an existing style it's not focusing on text or fingers or anything else that the AI is bad at and there's plenty of training data for images just like it based on everything that matters this should go great right wrong mid journey is going to struggle with prompts like this one every single time and understanding why is the key to making better art in mid Journey also you're in luck because they just released a feature that allow you to fix this problem using a single button I'm telling you this is so important just look at this when you generate that prompt I just showed you and zoom in on the details our subject's face looks crumpled up the suit looks cheesy instead of elegant the crowd is deformed and the entire image just feels very AI in all four variations this isn't a problem with the prompt this isn't because there aren't enough similar images in the training data so what went wrong how do we stop getting results like that and instead create photorealistic art like this and wait a second hang on is that the same prompt how'd I do that we're getting ahead of ourselves let's break it down the problem here is that mid Journey literally got overwhelmed and tried to do too much at once think about it for a diffusion model some images require less compute than others the reason this image sucks is because mid Journey assigns the same amount of compute to every gen generation approximately one GPU minute on their super computer cluster obviously that's a problem a close-up portrait of a subject's face with a blurred background is a lot simpler to accurately render through diffusion than a fullbody professionally designed outfit with 80 High Fidelity faces behind it so when it tries to do that using just one GPU minute it falls on its face and gives you these pathetic results to get around this deficiency we need to pick what mid journey is working on when the goal is to make mid Journey's GPU Focus first on the most important part of the image then slowly and carefully fill in the details and we're in luck because mid Journey just quietly added the tool for the job reframe if you don't know reframe is the combination of pan Zoom change aspect ratio and remix mode so here on the alpha site it gives you a chance to visually see what of those actions you're performing how it'll change the composition of the resulting image and lets you adjust the prompt while you do it so coming back to our red carpet example instead of trying to do all of this at once we should instead start closeing on our subject and work our way out using reframe doing that means we're spending our compute on the parts of the image that people will actually pay attention to Let's watch this in action we start with a prompt to generate the close-up of our character instead of being on the red carpet we put him on a dark background with lights and a distant crowd for our first image this way like for the original our subject will have the right content in the background right behind him the difference is it can spend its full compute on getting this part right and here obviously mid journey is powerful enough to do a great job at that now we click reframe there are two ways to reframe an image either by changing the aspect ratio or zooming out by combining these you can put the starting image anywhere in the composition of your frame here we'll start by setting the aspect ratio to 2x3 and choose start to put our guy's face on the top click here to edit the prompt and now our prompt from earlier can come out to shine the result pans down and shows more of the crowd more of our character and a bit of that classic red carpet and looking at these faces see how much better they already are compared to that original obviously this is still a hard image and as we work our way down and out you can see we start to hit the limits of what mid journey is able to add to a single image the faces start to get bad not demons spawn bad but not good either and now we know why the more we ask for Mid Journey has to work harder and tries to get more done but eventually it just has too much to pay attention to and the edges of our image still go crazy and after all that still you've created a mid-journey image with a coherency that was otherwise impossible so with this method not only do you have full control over your composition you're finally able to see what you're doing to combine pan and zoom to put your subject exactly where you want them but you can also use this feature to alligate Mid Journey with the resources to complete the parts of your request that actually matter to you starting with the most complex and most important parts of the image and working your way out doing this will give you a huge leg up in all kinds of generations this is super powerful for consistent character generation start close up where the consistency from your CF is the best and of the highest quality then slowly reframe the image step by step to put that character exactly where you want them this method is Miles better than trying to one shot the same scene and same character I mean talk about deform look at that H but apply this to your architecture your generated paintings or sculptures or carvings and you're going to see that you're able to get some mighty fine artworks that were otherwise impossible this was kind of an intense topic for YouTube so I really appreciate you taking the Deep dive with me GPU resource management for your AI art is a really technical concept if you appreciated it please go ahead and click that like button it really is the best way to tell YouTube that I did a good job explaining cool stuff I'll link to this video where I talk about consistent characters in this one I created a GPT that writes every prompt you'll need to get great characters every time anyway thank you so much for watching and I'll see you in the next one

Channel: Glibatree

Views: 3,764

Keywords: AI Story Telling, AI Art, Midjourney, ChatGPT

Id: biwSQNVa9w4

Length: 6min 45sec (405 seconds)

Published: Mon Jun 17 2024