Explaining 6 More Prompting Techniques In 7 Minutes – Stable Diffusion (Automatic1111)

prompting can be a challenge even if you know the basics and I wanted to explore more prompting techniques so you have more options for bringing your ideas to life drop a like so the algorithm recommends this video and let me give it to you bite-sized so you've probably heard of the mysterious break keyword an operator which many struggle to understand well when you type prompts it fills this number on the top right hand corner of the prompt box called tokens the brake operator when used in all capital letters fills the current token limit with padding characters to create a new chunk these new chunks are created every 75 tokens and these chunks are processed to generate images and while that isn't useful on its own there's a practical application to using break which you may find helpful so let's explore this break can be used to help mitigate the effects of color bleeding wear colors in your image aren't located where they're specified in your prompts and the effectiveness varies from SharePoint to checkpoint as some do a good job at managing colors without this trick I also want to stress the importance of ensuring that you're using the correct prompting style for your images for better accuracy and note that the placement of break might be different on other checkpoints but the concept is still the same now if we were to generate a portrait with different colors in our prompt you can see how they aren't exactly in the location specified we wanted the dress blue the eye is green and the afro orange and we got this which isn't too far off but has some artifacts to improve this we can make some adjustments to our prompts by using the break keyword between our prompts where we specify color I'll also put the prompts which use color in Brackets with waiting and as shown in these images we're getting a far better set of results where the color is much better placed than before if you notice that particular color is weak then increase the waiting for that prompt to draw it out further as previously mentioned the placement of the break keyword may be different on other checkpoints but the theory is the same and once you find an image you like overall final adjustments can be made using impainting now something which isn't really explained all that much is the difference between tagging and writing when prompting as certain checkpoints prefer one over the other and while both will work they operate differently tagging means using predefined tags from websites like Danbury within your prompts which tells stable diffusion that your drawing references from this website's collection of images written is prompting by actually describing what it is you want in short phrases separated with commas and drawing from the HTML image tags for billions of images online on which stable diffusion was trained on now here are the benefits and drawbacks for tagging the result you get is entirely dependent on how many images are available for that particular tag and how the tags are formatted on the website so for example about where to type in black afro as a hairstyle style it struggles to figure out what that is because the prompt black afro isn't a tag on Danbury which results in this wavy hairstyle but it does recognize the tags black hair and afro separately so using the tags separately in my prompt will yield better results but they won't be perfect as there's only 2.3 000 images on the website that use the tag afro compared to the 474 000 that use the tag braid and using this will be instantly recognized by stable diffusion but turning to written I'll use my favorite checkpoint absolute reality with a more grammatically correct prompt including the black afro the results are what I expected but we have the benefit of using whatever words we want outside of Danbury tagging such as specifying a full afro for a better result now we still will get this part of the fact and I think that's due to the wide number of afros who scarcely defined naming conventions but I found that adding the term men's 1970s 4 after awake helps as most Google search results seem to associate that term with a puffy Afro as opposed to braids dreads or other types or styling but these additional terms and phrases would have been difficult to draw out if we were limited to tax as there definitely aren't enough images that cover these more Niche Styles now you can get different types of camera shots in your images depending on how you describe both the image and the type of shot you want I'm using XYZ plot to test a variety of camera shots and this is the result where we can see how the prompts and waiting impact the type of image we get looking at the results I had to add in some weighting to make it work but the different angles help the images look more distinct there were other kinds of prompts you can use in which some will work and some will not but hopefully these seven act as a good starting point for whatever you need stable diffusion is also capable of generating different visual styles by specifying a style before the term such as art style Within your problems this could be a flat Manga style the painted impressionism or even a realistic style which borders on 3D you will notice that manga and 2D give similar results as would 3D and realistic because they are pretty similar in terms of what they actually produce so it's worth using tools like XYZ plot or plot Matrix to remove redundant prompts and find ones that give you the results you actually want lastly some checkpoints handle style changes better than others so if you struggle to get the output you want consider using a different checkpoint of adjusting the waiting now I wanted to move on to clip skip which isn't really a prompting tip but can improve the results of your prompts clip skip represents the layers of the clip model when generating images and the clip model is the text to image generation model that takes text and produces images as a result now why should you care well for each clip your image goes through the resulting picture will be more than more legible to your prompts more accurate and it's some cases more broad by setting your clip skip to a value such as two or three you will get a less legible result which will be more accurate to your prompts as it doesn't overthink what you're trying to describe for example clip skip 2 gives us a dark-skinned woman with the nephro while sex gives us a woman in a forest without the styling and one will give us a less accurate result aiming more for the female with the style the setting can be found on the settings stable diffusion and I'd suggest using a value of 2 but you can go up to a value of 12 although I'd suggests the highest being free for optimal results and then we have the word and which I never see being used and likely because it complicates things further but for the sake of understanding it let's explore it further the operator is and in all capital letters and this will combine different prompts into one before and after where and is used and may be useful for combining different concepts and art styles into one before making adjustments through normal prompting where commas are used for example if we use our previous prompt by adding Gal Gadot and Amazonian whatever we can see how it tries to merge the two together with the rest of our prompts whereas using a normal end has a weaker impact with the Tiara being the only clothing of an Amazonian whatever in the image but hopefully this video is as useful as the previous one and if you liked it then be sure to hit the like button and consider supporting with the links in the description this is bite size genius and I hope you enjoyed

Channel: Bitesized Genius

Views: 11,080

Keywords: Stable Diffusion, stable diffusion prompt guide, stable diffusion controlnet, stable diffusion prompts, BitesizedGenius, Automatic1111, stable diffusion lora, stable diffusion extensions, stable diffusion embeddings, stable diffusion checkpoints, stable diffusion anime, stable diffusion scripts, stable diffusion video, stable diffusion img2img, Stable diffusion realistic, stable diffusion models, stable diffusion install, stable diffusion tutorial install

Id: UeR0yZOYS0Y

Length: 7min 29sec (449 seconds)

Published: Wed Aug 16 2023