Playground AI tutorial Prompt Engineering 101

let's start from the beginning if I were to enter this prompt men in a suit you'll often get varying results and quite honestly unpleasing results for that matter and in some cases deformities double hits double torsos and as a result of this most people think man this program sucks meanwhile on the home page of playground AI there are these beautiful images that so many of the members have made so what gives how did they get to these final results to start off I think there are some basic things that you need to understand about how a stable diffusion works even with this simple prompt I'm going to show you something that will make a big difference on the outcome of the results that we previously got you see that the coherency of the image is slightly better the quality can use some more work but more on that in a second don't worry about the cropped images we'll also address this very soon but notice there are no double heads some deformities in the faces and hands and that's mostly because the AI needs more information but you will see that there is a wide variance the types of images we're getting to understand what's happening here there are some things that you need to consider stable diffusion was trained on a massive database with dimensions of 512 by 512 once you stray from those original aspect ratios the likelihood of you getting deformities double heads or unpleasing results will increase if we go back to the prompt that we entered man in a suit is a very general term so in turn you are going to get a variation of results so the first thing to consider When developing a prompt is to be as specific as you can and be very descriptive building from a simple prompt is the best way to have the most control over the image you get utilizing adjectives to describe nouns is a great way to approach prompting also knowing the fact that these data sets are images from stock sites or Google or similar places many of these images have tags so ask yourself the image that I want what would it be tagged with by building your prompt from scratch you will start to develop templates for yourself to reuse and tweak and create even more amazing images now how we can do this is very simple building your prompt from scratch utilizing seeds and negative prompts to prove it to you I'm going to take this not so flattering image and transform it to an amazing portrait the first thing you want to do is utilize the same seat basically a seed is a random number that is generated by stable diffusion utilizing the seed will keep certain characteristics of that image somewhat consistent the next thing we want to do is identify things in your image that you don't want toggle on exclude from image this is known as negative prompts reviewing the image there are various things here we can put into the negative prompt a cropped head it looks more like an artistic painting rather than photorealistic there are too many buttons and it's an all gray suit perhaps we want to change the color let's enter all those things that we don't want to see in the image now let's generate a new image and now we see the image is slowly taking shape the head is still cropped off but don't worry about that we will address it later and you can also fix cropped images in canvas but at least now it's looking a bit more photorealistic let's enter some negative prompts regarding the hands it's not terrible there is some potential there but it's also not perfect for now we will enter these words into the negative prompt now let's start to shape the image often what I like to do is ask myself questions we have a man in a suit what kind of suit what color suit let's give him a blue suit but what kind of suit is it is it a plaid blue suit a leather suit let's give him a silk suit let's also give him a purple tie man is a very general statement who is this man what does he look like let's create a handsome man as we look at the generated image you see that we have a man in a silk suit the light purple tie we can further emphasize that later on the hands are looking much better still need some work and we still have the cropped hit don't worry about that for now what are some other elements that you think we need in this image well let's put him in some sort of environment for now let's keep it very general and put in nature background at Sunset previously we had no background now we have a nice sunset with a nature background notice we haven't added any more negative prompts yet I discovered along the way too many negative prompts can also negatively impact your image so it's always best just to put in what's required at this point we have what I call a foundational prompt we have a subject in an environment the next step is to add what's known as modifiers whether they be Artistic Styles certain details of the image and once again thinking adjectives to describe the noun so let's focus on the man once again perhaps we want to give this man some ethnicity since I'm Filipino we're going to use Filipino and I do have some Spanish blood in our bloodline let's further enhance the background to add some mountains and what is your subject doing are they reading a newspaper having a coffee in the cafe for this example let's do something simple we'll simply put waving hello one of the advantages of using a seed versus image to image with a seed you can still tweak the person's pose every change you make will change the image but you are not committed to one pose as you are with image to image but we see here now this subject is waving hello we've got some mountains in the background and now it's really starting to take shape now let's talk about the fine details some very common words used in prompts for details are high details intricate details or even a word like ornate is a fancy way of saying give me fancy details in the beginning of the prompt we're going to enter highly detailed photo and to give it more style we're going to indicate fashion photography using a type of Photography like Fashion sports Wildlife well often call up images that have more of a professional polished look indicating camera models will also tend to inherit characteristics of those cameras remember think tags however if you notice from the result of this image the hands are getting a bit more deformed and even the face now while people will automatically think oh I gotta put something in the negative prompt like long fingers which we should do this may be a case of the order of the words prompt order in stable diffusion 1.5 is a thing it could be that this fashion photography prompt is conflicting with the general prompt let's remove it and put it at the end instead and now we see we don't have the long fingers but we have more fingers so even that slight change could make a difference in the output of your your image back to the negative prompt here let's enter long fingers we already have too many fingers but at this point I wouldn't worry too much about having too many fingers because this image is going to continue to change moving things around in the prompt can result in a better output I ended up putting fashion photography and the camera information last highly detailed photo we moved after describing the attire and the result was back to four fingers in a thumb a little bit of a better image the face is a bit distorted but that's okay don't worry I promise you it will fix itself so the point of this is yes order matters whatever is priority in your image typically you want to put it at the front of your prompt however if you start getting unpleasing results move around some of these words before where you move on now that I have a good foundation this is where I would change the model or even add a filter let's select realistic vision and you'll clearly see because we worked on the raw stable diffusion version Now using the model we have a much cleaner image more details the hands could use some work but we'll get to that in a second basically a transformed image with RPG we see a boost in the contrast and quality of the image almost like a high dynamic range image the hands look slightly better but still could use some work and with rev animated we have more of a much better hands the thumb could use some work but a much more artistic and stylish image some details are a bit off but we can change that and I'm going to show you how one of the tips that I can give you is that rev animated does half body full body shots very well and it tends to give you very pleasing hands before I decide on the final look and polish of the image this is where I would make some adjustments to the seat let's change the five to a six and we have a much better rendition of the previous image he's got a bit of a Stumpy pinky but that's okay the suit is a lot cleaner it's got some great details let's change the six now to a seven once again we get a different variation of the same image personally I like the way the 66 looked and now we're going to change the front of the seed let's try an eight now we're getting somewhere the fingers look a bit more acceptable except this one's a little too far from the others but the whole point of this is that playing with these individual numbers in the seed May fix existing problems or give you different variations all I can tell you is to experiment with each seed number increase them decrease them as you see with this image that I just did again the hands look much more acceptable yes the style is changing quite a bit but it's a much more pleasing image I'll do one more here we'll do eight and change the three to a four now this may look like he's only showing three fingers it's just the pinky is positioned in front of the other finger so technically it's okay it's just a weird looking finger now let's assume that I want this image to be my final image now I would bring this into image to image and experiment with other filters let's run this through realistic Vision we see the image under image to image with an image strength of 30 which is not a bad place to start judging by the results here we see that the hands are so warped now there's a few things we can do here we could add negative prompts but first I want to try increasing the image strength to 60 that way it doesn't sway from the original image all too much and there you go the hands are a lot better probably this one is better than that one and if we open up the image we see here now it's got more of a photo realistic finish I could further experiment bring this into image to image and try RPG for example with the same image strength and we see similar to realistic Vision RPG also produces a very realistic image as well so it's not always about the right prompts and the right filters to use using the right negative prompts adjusting your seeds accordingly will give you much better results than spamming multiple images at a time in practically gambling with AI so to speak once you learn to develop a good basis of a foundation for images you want to create this will open up so many possibilities and much more flexibility to get the results that you want until the next video my friends the this is playground AI

Channel: Playground AI

Views: 78,157

Keywords: Playground AI tutorial Prompt Engineering 101, playground ai, playground ai tutorial, prompting basics, how to prompt, stable diffusion, stable diffusion tutorial, prompting 101, prompt engineering, prompt crafting, ai art, ai art community

Id: TMdH8uP0NXE

Length: 14min 34sec (874 seconds)

Published: Sun Jun 04 2023