How to improve 3D people in your renders using AI (in Stable Diffusion) - Tutorial

[Music] hi everyone my name is dear I'm from the digital bunch and today we're here to show you how we use stable diffusion in our projects now recently we've shared a few reals of us doing little tests with stable diffusion and and artificial intelligence and it's actually had an amazing feedback and a lot of you were asking to um for us to do this tutorial so here we are and we've been internally experimenting with this these tools because we think it's really important to have our foot in the door and really see how it's evolving um it's definitely had a few great results and we've definitely had a few mixed results and you know not everything is definitely production ready for now but when you think about it a year ago or two years ago nobody ever even thought that AI would be here to help artists because this was considered to be the creative industry that robots and automation is never going to touch so here we are um a lot of these assumptions are obviously different to um you know what we were thinking and um we just need to stay on top of it really and and see how it goes so without further Ado let's dive into a stable diffusion tutorial and let us know what you think in the comments below let's start from the beginning stable diffusion is an incredibly fast growing open-source software project it's a deep learning text to image model released in December 2022 before you start you need to install stable diffusion and we're not going to show it here because you can find very clear instructions on how to do so on their website we're working with automatic 1111 which is the web interface for stable diffusion this interface can be a bit confusing at first but we'll walk you through it and you'll see how it can give users a lot of features and options to play around with we have already got a desktop shortcut to the script that opens the platform you will find it in the files that you will install look for the name web UI user in this script you will find the url to copy and paste into your browser you always use the same URL so you can even bookmark this tab on your browser and keep it accessible let's start with Opening Our file in Photoshop when we need to work around an issue that stable diffusion still has it doesn't process large images yet we only asked it to process a small part of the image so we need to crop the part we're most interested in and save it out as a separate file when we open stable diffusion we go to the imageo image Tab and here we select the in paint tab now drag and drop the cropped image to this area and select the elements you want to edit with a brush [Music] now you can pick the model you want to work with each model is trained to do something different so if we're changing people in our image we will want to select a data model that is specialized in faces and people the one we pick here is realistic vision and we will leave a link to this model in the description below for example another model that we have experimented with a bit is photon and we found that it works really well with realistic vegetation and environments okay now we can proceed to typing your prompt the way this works is that there's a positive and a negative prompt try to keep these as simple as possible and easy to understand usually it's good to Define what the element you need to change is and add adjectives like photorealistic high quality 8K describe the lighting natural lighting if this is an inter exterior shot and you can add the lens like 50 mm lens this is very similar to the way prompts work in mid journey and you can play around with it to check what results you get now moving to the negative prompts this is where you can type in the results you don't want here it's worth adding adjectives like anime cartoon 3D render drawing artifacts ugly deformed Etc now moving down we get to a lot of various settings and features a few important points here is um to check the only masked options since we want AI to generate stuff only in the areas we selected in this section we want to input the value of 768 pixels because we found that this is the optimal resolution this model works with anything larger won't be as good quality-wise and anything smaller would will not have enough detail for our final image we set the batch size to four so that stable diffusion generates four different images that we can choose from we can obviously add more but it will take longer to process and usually one of the four Images works pretty well now the most important den noising strength we usually set it to 35 this is a value that tells us how different the newly generated image will be from our original image the higher the value the more different the output we found that at one we get some pretty weird results like these all we want to do here is to make this model more realistic not change it completely so we're pretty happy with a value between 25 to 45 then we click generate and let the magic happen it usually takes about 1 minute on our 4070 TI card and this is computed locally sometimes we like to select the face and the body separately for better results then we click this folder right here to open the folder in which these images are saved these options are pretty similar but you can pick the one you like here since we selected the whole person the face itself isn't the most photo realistic and you can still see some artifacts on it you can then paste it back into your visualization and here's your final result some before and after previews stable diffusion is really good for tweaking clothes and even though you might still get some weird results sometimes it does look more realistic than 3D models we also tested fixing people that were already generated by stable diffusion and it's not as stupid as it sounds sometimes you really can improve the results even further of course this is just an AI tool so it does hallucinate sometimes especially if you bump the D noise value you can get some pretty weird creepy or funny results that's it for today now we really hope that you enjoyed this and that you found it very useful now we would love to see how you guys are using it and the outcomes that you've achieved because some of our stuff is kind of cringy and weird um obviously this is super exciting so we really look forward to hearing your comments and let us know what else you'd like us to show you or what else you'd like us to test um we love doing R&D so honestly this is going to be amazing and see you hopefully next time

Channel: The Digital Bunch

Views: 5,442

Keywords:

Id: cxvd_5Tot7Y

Length: 7min 26sec (446 seconds)

Published: Wed Feb 07 2024