Stop STRUGGLING with AI Art Prompts | Basics to Advanced masterclass

in this video I'll share with you some secrets and Advanced Techniques that will help your images get to the next level so let's go this will be the first of three videos that will help you get from an idea to a final beautiful image and first we need an idea for this you can actually go to civit AI not only there are beautiful images in there but they have an actual prompt to see how they were made in this case I like this cut so let's make a cut inspired image at the beginning I really like to create four variations at a time to better see what the model understands battery size refers to how many images will be generated for batch and batch count refers to how many patches there will be every time you click generate so a batch count of 4 with a battery size of 1 would generate one image four times and a batch count of one with about the size of 4 would generate 4 images at once okay stable Fusion make me a cat driving a supercar in a cyberpunk city this is honestly not as bad as I expected but it's also not one I want so we gotta tackle the basics first formatting let's see how the original poster typed its prompt for this click this little icon right here and then you can either read The Prompt or straight up copy the generation data you can paste the generation data into the positive prompt then click this blue icon right here and automatic 1111 will replace everything to match the input data notice that the prompt is separated by commas and that's because stable diffusion struggles with natural language understanding making words like to the right with that which less meaningful we can also see that there are words that don't really describe what's going on in the image and rather its overall quality I call this enhancers and they definitely work some better than others but I recommend using some of them now let's throw this in the trash and type our prompt again I will use PNG info for this you can find your outputs by clicking this folder icon right here now let's select the first image with it and drag it right here this will give us all the generation data but you can just click send to text to image when prompting I tend to follow a very specific order keeping in mind that the words at the beginning of the prompt weighed more and therefore are more important than the word at the end I'll put what type of image I want at the very beginning whether it is a photo an illustration a painting at that Ben I'll add the main subject so in this case I want a raw photo of Aika next I'll type the action driving followed by the place or environment a futuristic city and finally the style in this case cyberpunk this is what my main prompt will look like and then we'll add some enhancers for this I use a preferred template create shortcuts by using this dialogue you can type your template into the positive negative prompt or both and then click save and choose a name to import just select it in the drop down menu and click the style icon lastly since the enhancers took some weight of our main subject the cat we will select and use control app to emphasize its importance if you want a certain word to gain or lose priority select it and use the don't turn up or down arrow to automatically type this parenthesis and the number I don't really recommend going above 1.5 of course over time I'll change the prompt as I see fit in order to get the best image I pretty much only think about the negative prompt in case there are very specific things I don't want the damage that are some words that could lead into a misunderstanding of the main prompt else I just put some basic template negative prompt the negative prompt can be a very good tool though but it is really really hard to control it in benefit of a specific result I am always looking at Civic AIS prompts and checking for good Styles and answers and negative prompt options but if you want a large list of possible Styles and other stuff check these websites I'll leave in the description something you should remember is that even if a word exists if it hasn't been used enough during the training of the model it won't recognize it so you may find a really cool style that you love but it is possible that stable division just doesn't know it don't worry though the next episode of this masterclass will solve just that you will see that when we import it from PNG this number here changed this is the seat of the image in other words the image ID and this is a very powerful tool that allows us to see what every single word on our prompt actually does knowing the images ID means that you can generate the same image every time while also creating slide variations of it for example using our initial prompt let's change cat to dog and generate again as you can see the image is pretty much the same but with a dog instead seats are crucial for understanding how stable diffusion interprets your prom as you can see how one image changes instead of having to deduce it by looking at multiple new images with this in mind let's compare the images that our original prompt made with the reformatted prompt Generations I would dare to say that this is much better even though the cut has completely disappeared but before fixing this I want to change the aspect ratio to better fit a cinematic shot the aspect ratio has a massive effect on the image even with the same Brandon seat the image is pretty much completely different and this one even created more than one image per generation to better control this I'd suggest looking into the recommendation each model gives you depending on the sizes of the images that it was trained on it will be more precise using those aspect ratios and sizes also think about what you create and what's the usual format you would see it in for example selfies are usually taking on a 9x16 aspect ratio while landscape photography usually uses 16x9 and now it's time to iterate this means basically clicking generate and changing little words on the prompt until you find an image that you actually kinda see fit then you can block it by copying this seat clicking this icon right here and just iterate over and over again but with the same image once I find a prompt that I think can work well I'll also go over and play with the CFG scale technical speaking the CFG scale I have no idea what it does but I like to call it the creativity scale the higher the number number the more literally it will take the problem and the harder it will try to follow it the lower the number the more freedom stable division will take while generating after iterating for a while I ended up with this image and these parameters as you can see I have changed the samples and the sampling method each sampling method will process the image in a very different way the sampling steps are the times each image is processed even though that doesn't necessarily mean it's the same image sometimes if you change the sampling steps by a lot it will generate a completely different image in the image viewer you can actually see how the image is being processed almost in real time but how can you know which sampling method is the best and combined with how many sample steps oh wait there's also CFG skill now too so yeah there is no best combination that you can use always either but don't worry there's a way to know what's the best combination you're looking for right now and it is using scripts down here in the script option select the XYZ plot this scripts allow you to create a matrix of generations with pretty much every combination you need this is super helpful and now we will use it to see the next first I want to test the best CFG scale so I'll type my code current one eight now comma five comma seven comma ten you can test more if you want to but then I'll also change the number of sampling steps current then 10 then 30 and then 45 and finally the best three Samplers error a DPM 2m Karas and uni PC what the script will do is create generations of the same prompt the same seed the same everything but first with 8 then with 5 and then win 7 and 10. this will create a matrix using every possible combination between these parameters I really like the generations created by Euler at 25 steps and CFG scale of 5. so let's go with this for now hey I hope you're enjoying the video as a way to say thanks for watching this Vlog let me teach you a really Advanced and cool technique this is called or at least I call it this prompt blending very few people know that you can actually change the prompt while the image is still generating here is how first you type what you want to add it in between these square brackets and then you choose one of three options option one switching steps for this you separate the concepts with a vertical bar this will process the prompt switching the word every sampling step in this case sampling Step 1 would be forest and step 2 would be City and then three would be Forest again and so on this is the result keep in mind that the first word will always have a bigger impact than the next by the way you can actually put more than three words in here I don't really use this one much but here comes the good part option 2 switch here you can actually choose when you want the switch to happen by writing the first concept a double dot the second concept a double dot again and now these steps when you want the concepts to switch this is extremely useful as it gives us a really high control over the image it is a very accurate way to create Blends between Concepts and it can be used for as much stuff as your imagination can figure out this case I'll compare to creating a skeleton woman with and without this option without it it manages to create a cool looking female skeleton but it does some weird things in the breast and more importantly it's really hard to control how much skeleton I want versus how much woman I want and that sounds weird okay but next I use switch now not only the skeleton is way more normal but it also portrays femininity in other aspects like the pose and I have a ton of control over the final result you can either type the samples by number or by percentage by typing 0. the percentage number and remember that the first word always has more weight so switching at 50 sampling steps will not give your result health and health but instead more of a 70 30. and now on to option 3. I not and removed this is a way to either take words out of the prompt or put them in at certain specified sampling steps this is really useful for compositional purposes in this example I have the word River in the prompt because I want the composition to create a Ribery shape in the middle of the image Ribery does that exist I don't think that exists okay well never mind but I don't really want the river per se this way I can make the word influence the composition but not the final result and adding words mid generation actually works in Reverse I usually use it for words that have a strong concept of living okay what is concept bleeding first concept bleeding is when a concept or word has some implied or unexpected effects on your image even if the world itself has no such implications for example in the first image we generated adding the word green can change the whole composition of the results even with the same seat and prompt even though it doesn't really make sense to humans to come at this we can use this option for example let's try adding green at the third step and amazingly enough not only did the composition not change this time but also the color applied way better concept breathing is actually really important and can be used in your favor as well in this case back to our image after trying the prompt with different seeds I can see that it isn't actually as consistent as I'd like it to be to make it more consistent I tried prompt blending in this case I'm gonna try and use one man driving for the image to First generate the human form and then swap it by a card driving and it helped a little but it still created a bunch of images that were not inside the car nor subject Focus for this I took advantage of the concept bleeding that comes with the prompt perfect face as in order to create a perfect face the AI needs to see the face and have a portrait style image most likely when it was trained perfect face was given only two prompts that showed a face which makes sense and therefore creates a bleeding without actually specifying anything I got way more consistent results the consistency improved a lot now it's not perfect but adjusting the prompt a little we could actually use it to generate good images directly on the next video I will take this image and change it so my cat is the one actually driving the car we'll learn about models loras and other super useful stuff so make sure to watch it as well if you have some cool prompting techniques make sure to comment them in the comment section below and have fun see you

Channel: Not4Talent

Views: 40,709

Keywords: stable diffusion, sd, stable difusion, promting, prompt, tutorial, guide, stable diffusion prompt guide, stable diffusion tutorial, ai art, ai, crate ai art, beautifull ai art, create good prompts, talk to ai, talk to stable diffusion

Id: 9H0oOexrupY

Length: 12min 13sec (733 seconds)

Published: Mon May 01 2023