SECRET FREE Stable Diffusion 2.1 TRICK! As Good As MIDJOURNEY!

Key Takeaways

hello humans when we scan your overlord and today I suppose I owe you a little apology because it seems like I was maybe a little too harsh on the stable diffusion 2.1 model now you might ask wow really you like the 2.1 model now after everything bad that you said about it well did you suddenly become a photorealism lover oh no no no no absolutely not it's just that well there is actually a very easy and very simple way to make the 2.1 model way way better and generate images that are very close to what mid Journey can do and all of that for absolutely free so let's take for example a simple prompt photo of a beautiful woman looking at the camera centered medium shot sitting in the background etc etc using the 2.1 model if I click on generate you get this kind of image not bad not terrible a little boring certainly not the kind of images that I personally would like to generate really just nothing exceptional but if I add this argument events to The Prompt are by mid-journey CGI underscore animation and I click on generate suddenly the same image becomes way way better we're talking about a completely different style here and all of that just by adding two words so what is this is this dark magic is this something that I did did I use my special AI brain to generate this new image well no actually in this time I didn't really do anything special because all of that is thanks to small little files called textual inversion embeddings because yes ever since the stable diffusion 2.0 model came out the community found out that texture inversion embeddings worked much better than with the old 1.4 or the 1.5 model now for those of you who don't know what extreme version buildings are they're basically very small files like usually a few kilobytes that contain Trend data of a small part of a neural network while that means is that it's kind of like dream move when you can change stable diffusion with your own images but at the end you don't get like a huge ckpt file you instead get a DOT PT or bin file that you can use in your prompt but to get these files you either download them from others from the other people of the community or you train them yourself and before the 2.0 model came out these embeddings weren't that good I mean they weren't bad but they were not as good as training an entire model with dream booth and embeddings trained on the 1.4 or the 1.5 do not work with the 2.0 models in vice versa and now that the 2.0 was trained in a different way compared to the 1.4 or 1.5 these embeddings trained on those models seem to work even better than before and one of the great things with the Bendix is the fact that you can use multiple embeddings at the same time meaning that you can mix and match them him for pretty much infinite combinations for example let me add some negative prompts to begin with and generate a new image this is the kind of image that I get if I simply add another parameter to The Prompt let's say I use Viking Punk 63 which is another embedding and then click on generate this again gives me a completely different image and all I'm doing is just adding in a bending on top of another and you can keep going as much as you want and this is what makes these embeddings so powerful it is the Infinite combination of all the different styles put together and even more since this is a simple keyword you can also select the keyword and then either decrease its strength to have a completely different result or add another one on top of all the previous embeddings used I mean look at the images that we generated right here this kind of images would be impossible to make with the 2.1 alone and with the quality of the details and outputs I would even say that this could be even better than the 1.5 I mean this is the kind of image that you would generate using the current mid Journey but you're doing this for absolutely free now then where do you actually find these embeddings and how can you use them well I would say to be able to easily find and take some inversion embedding the website that I recommend is called civitai.com which is a website where you can find plenty of models trained by the community and on the right side we have here a filter where you can filter by the model types and you can filter by checkpoint external version IP network aesthetic gradients or Laura and if you only want to see the textual inversion models all you have to do is just check this box right here and you will see a bunch of models available for free for you to download but make sure that this embedding was actually trained on the 2.0 models because otherwise they will not work so oftentimes you have to click on a model and read if this was indeed trained with the stable division 2.0 as a base or you can also use tags like sd2 700g 68 that basically lists all the models trained with that specific model and the second side is of course hangingface.com but unfortunately on this side it is actually very difficult to find the right text to inversion models because not everybody is using the right tags so correctly filtering them all is absolutely impossible but I do have two username for you two people who are actually very good at creating those sticks to inversion embeddings the first one is Shadow X Shinigami who is well known for creating his mid Journey paper card model that is actually absolutely fantastic and you should definitely try it out and the second user is called conflict X who created the CG animation model that is really fantastic if you want to generate these new style images they are really really good and also the Viking Punk model which just looks absolutely fantastic now at the end of the video we will come back to these models because I will actually give you my personal list of all my favorite models that I tried this past few days and that I think you should definitely download and try it yourself because they are really really good so then how do you download and use them well first you're gonna go into files and versions if you're on a hiking face then you can click on vikingpong.pt and then click on download and if you see a PNG file you can also click on it and download it too and there's also a way for you to share an embedding file so once you've downloaded these files you're gonna press Ctrl X to cut them go into your stable diffusion web UI folder into embeddings and you're gonna paste those files right here and then you can launch stable diffusion and the next time that you'll learn stable the fusion you will see here a text telling you all the texture inversion embeddings that were loaded with stable diffusion meaning that all of these embeddings are ready to be used right now and to use them it's actually super simple all you have to do is just remember that there is a keyword that you need to use usually it is the name of the model if it's not the name of the model the owner will usually say what keyword you need to use for example in our case for Viking Punk the key word is Viking Punk so if you want to use that in painting all you have to do is just go into your prompt and type Viking Punk and well that's it you can just click the generate button to generate a brand new image with this new embedding and look at the Quality I mean come on now this is really really super cool this is some high quality stuff and again as I said really the best thing with this is that you can mix and match them as much as you want that's what makes these embeddings so powerful and I can tell you they were never that powerful before with the previous models so yes in a way the 2.1 model is actually way better than the 1.4 or the 1.5 at least on that front now I know what you're gonna say okay well that's great and all but how how can I actually train my own embeddings well it is actually fairly simple and you can do it directly from stable diffusion all you have to do is just click on the train Tab and here you have an option to create a new embedding now keep in mind that you still need a beefy graphics card to be able to do this I think you need around 12 gigabytes of vram to be able to do this and it also takes the same amount of time as if you wanted to train a stable diffusion model using dreambooth now in this video I'm not gonna go into too much details on how to train a texture inversion embedding I will leave that for a future video because this is definitely way more complicated than using dreambooth because there are way more options and it's definitely less precise than dream Booth but at least they will show you the basic principle and the recommended parameters this way you can start training right now but again I'm not gonna go into too much details today I will probably leave that for a future video so here you're gonna have to choose a name for your in Beijing so let's say for example you want to train a mid journey in Bedding you would put a name here that you would remember that is not only easy to use but also that stable diffusion does not know so do not put something super generic like man woman painting or anything like that so in my case what I usually do is that I start by calling them EMB EMB for embedding underscore and then the name of the style that I want to train so in my case it will be MJ for mid-journey for the initialization text you're gonna leave it at default because this way the keyword will always be the name of the embedding which means that you can change it as much as you want so if you decide one day to change the name of the embedding you can simply rename it and then use the new name in The Prompt for the numbers of vector per token here I recommend 8 which is usually what most people use and then you're gonna click here to create a new embedding and then you're gonna click on the pre-process images tab now the pre-process images tab will basically crop your images into the width and height that you choose right here and also will create captions for each images which will make the final training way more precise so here you're gonna input your Source directory The Source directory is when your simple images are located the destination directory is when you want your final cropped images to be or the width and the height do not forget that if you are training on the special 768 version you need to put them both at 768 then you're gonna ignore this you're gonna check here create flipped copies that will basically create a flipped copy of your image which will basically double the amount of images that you have then you're gonna click here on autofocal crop point which will basically automatically crop your images and only living the most important subject at the center so you don't have to do it yourself it will be done automatically and then you're gonna click here to use bleep for caption which will basically scan every single image and create a text file that will basically describe what this image look like and then you're gonna click on pre-process in in the end you should have something like this when you will have a bunch of images both flipped and for each image you will have a text file that will basically describe what this image looked like as you can see right here after this is done you're gonna click on the train tab here you're gonna select your embedding so in our case it was EMB underscore MJ for the embedding learning rate I saw a lot of people debating which one to use most people use the default 0.005 but I would actually advise you to use only 0 0.4 which will basically make the training a little bit less aggressive but it will still be strong enough so that you don't waste too much time for the batch size you're going to leave it at one for the gradient accumulation steps here you want to basically put half of your total images so if you have for example 200 images on the gradient accumulation steps you want to put like a hundred if you have 110 images you're gonna put here 55 life and so on and so on so the data set directory is where your final cropped images are so do not mix with the initial sample images that you got you need to copy the URL from this folder the one with all the text files this is the folder URL that you need to copy and then you're gonna paste it right here here you're gonna leave everything by default again for the width and height you want to select 768 and for the max steps here you're gonna put 150 and here you're gonna save an image to log directory every 50 steps and save a copy of the embedding to log directory every 50 steps then you're gonna click on read parameters from text to image tab when making previews which is why it's very important that you go back to text to image in here that you input the name of your embedding so in my case it was EMB underscore MJ and these are the parameters that it will use to create a preview image every 50 steps here for the drawable tags you're gonna put it at 0.1 and for the latent sampling method you're going to choose deterministic and finally after all of that you can simply click on train embedding and it will start the training now it will take a while it could take more than an hour depending on your graphics card but in the end when everything is done when the training is complete if you're going to text your inversion folder into today's date the name of your embedding then embeddings folder you will see here A bunch of dot pte files and these are your finished embedding style then you can then copy and then paste them into the embeddings folder and as you can see here in my previous attempt I've actually saved every 100 steps so this is basically a checkpoint for every 100 steps of training and if you want to see which one is the best in your prompt all you have to do is just input the name of your embedding then Dash and then the number of your checkpoint so let's say I want to see the 800 checkpoint so we simply write Dash 800 and then click on generate just forgot to put Viking Punk here but this is is basically the final result so it's really up to you to try it out and see which one works the best for you and here's my personal selection of the best embeddings that I've used this past few days starting with number one the mid journey in Bedding so of course all the links for these embeddings will be in the description down below and the meat journey in Bedding is absolutely fantastic it gives absolutely beautiful results and to be able to use this you need to use the keyword art by mid-journey and for example if we try it out in my prompt and we generate a new image this is the kind of results that you get so yeah uh looking really good definitely an embedding that you absolutely need to download the second embedding is called anthro and basically allows you to create this anthropomorphic animals that are just super super cute and it works really really well now this embedding unfortunately works better with the 2.0 rather than the 2.1 but it also works pretty well with the 2.1 but if you really want to train out train on first with the 2.0 and as you can see you have right here plenty of prompt examples that you can use to generate these kinds of images and it's really super fun the next one is called remix and this one is a little strange because the kind of results varies from image to image when you can actually get pretty interesting results just use the by remix keyword in your prompt and see what kind of images you can actually get as you can see in my case it works very very well so yeah just try it out yourself it is actually very interesting the next one is called CGI animation and this is really good if you want to create Disney type images so the CGI animation embedding will not transform your character into Disney it will only enhance the effect you can get pretty similar results without using the embedding but using the embedding in combination with those words basically make it way more powerful and more precise and it actually works very very well with animals so yeah again China yourself and this is actually pretty cool the next one is called noling case it is basically a model that allows you to basically generate an image inside of a case or inside of a translucent box as you can see here the kind of images that you get with this embedding model are absolutely fantastic just do not forget to rename the files to whatever trigger word you want to use otherwise using them like that is kind of difficult to remember the next one is already a model that we saw earlier in the video It's called Viking Punk it basically creates this kind of Viking cyberpunk theme mixed together to create this absolutely fantastic images and this is probably one of my favorite models too because of how versatile it actually is so this is definitely a model that you want to download and the last one is called v-ray render which basically create this kind of 3d effect on your images almost like these came from a super recent video game I mean look at those graphics and the keyword that you need to use to make it work is simply called v-ray render and this is the kind of images that you can get in combination with the mid Journey embedding absolutely fantastic so yeah there we have it folks now we can actually say that stable diffusion 2.1 is actually useful for something and I'm gonna be very honest I was not expecting that because I was actually having a lot of fun these past few days trying out all these embedics in combination with each other it is actually very very fun and the images that you get are absolutely fantastic so yeah definitely something that you need to train out right now and it will only get better from this point and there you go thank you so much for mp3 supporters for supporting my videos you guys are absolutely awesome congratulations also to this week's AR challenge winner CNC gig for his fantastic fantasy cake submission absolutely gorgeous and if you do want to participate where air challenge that we do every week on Discord you can click the link in description down below to join my Discord server and maybe YouTube can be featured in the next video that being said thank you so much for watching don't forget subscribe and smash the like button for the YouTube algorithm and I'll see you guys next time bye