This is fantastic. I work for an AI image gen site, and I'm always trying to explain the principles of prompting and what ( ) does vs [ ], or how to structure a prompt from most to least important. You concisely break down so many things here, thank you!
Break is perfect for describing individual people or items in an image. Never heard of it before watching this, but it was exactly what I needed to have two people of different races and dress styles conversing in the same frame!
BREAK helps you keep certain sections of your prompt cohesive. Since every chunk is 75 tokens, if you were to run into a situation where one of your token segments was abruptly ended (such as "wearing a white/shirt and blue jeans") it can lead to unwanted results. If you know you're close to your token limit before starting another section (like before the clothing portion), you can put BREAK into the prompt and ensure everything sticks together in the next chunk. The AI also seems to put more emphasis on things when processed alone. Splitting colours into separate segments can ensure they don't bleed into each other or colour incorrect items.
Great video. My SD setup is quite complex and I'm making some startlingly photorealistic stuff, but even I learned a couple tips here. Thanks for the upload.
It was a good video. Whenever I search for SD guide, all the videos that came up, only showed how to install it and a little overview. This video gave me a good idea on prompting. Good work 👍
I use what you call "Alternating Words" all the time. It is one of my favorite tricks for creating realistic looking faces by using a few different real people, but by alternating through 3 or 4 people using this technique in general it won't look like any one of them.
I give that if I wanted to create consistent faces I sold use a name, but then I found you can just plug in a number instead of a babe, so something like this "a tall woman with brown hair named 5437". Would give me the same face in all the images from that generation
Thanks so much for making this, faaaaaar too many people in the space seem to assume everyone can code with Python like a god so none of this has to be explained, yet those like me are sitting there with NO FUCKING IDEA how any of this shit works! lol So again, thanks, just knowing these basics makes a gigantic difference.
Personally, for my characters, I wrote a Python script that uses ChatGPT to generate a bunch of prompts based on a template, and then write the resulting prompts out to a file which I then run through SD using the Prompts from File
Awesome breakdown, I went through and tested each one as I watch this video and I have leveled up for sure and know a bit more how SD advanced prompting works, hope to see more good stuff!
Thank you for another very informative and helpful video! It's much appreciated! Regarding the BREAK, it actually looks like a slight bump in quality WITH the break. Nice!
Fantastic tutorial man! I've been, on and off, toying around with Auto1111 since early December. I've generated thousands if cool images, but... the output consistency is all over the place. Some great, but with lots of crap in between. I've mostly been winging it because I haven't seen anyone cover the features together well, anywhere, and there are just so many variables that can completely disrupt your output. It's time consuming. Also, many tutorials assume you already know much of this info and are hence obscure and incomplete lessons. E.G., I was pushing buttons, changing prompts, and observing the destruction... lol 😆 🤣 So that's lot's of "oh, hmmm, that's a neat, um, FAIL... f...uggggly." So thanks again! Great content, liked 👍 👌, subbed ❤, saved.
BadDream and UnrealisticDream is a negative embeddings not just simple negative prompt, I think you don't explain it (or I'm missing it?). without proper installing those two negative embeddings in \stable-diffusion-webui\embeddings\ those words don't have a real meaning and don't processed properly. those two files are not LoRA either, it's embeddings, that could be downloaded in civitai /models/72437 (I can't put real URL because it always deleted). from the video, on result log below generated images, I saw those two words are not processed as embeddings as it really intended. if those negative embeddings are used properly, there's a log like this: medium shot of a woman with full lips, golden eyes and a white crop top, long black hair, snowy weather, octane render Negative prompt: BadDream, (UnrealisticDream:1.2), (NSFW) Used embeddings: BadDream [48d0], UnrealisticDream [5f55]
Yes, this is something I missed and I plan to make a video explaining embeddings, LoRa's etc soon, as it's not as simple as copying the generation data. That should then give some context for future videos, without explaining it each time.
Thank you for explaining this so well! I am reinstalling SDXL because my first install was flawed somehow…this makes experimenting with other interfaces like Comfy easier! Such amazing software!
Thanks for this amazing SD prompt tutorial. Around the 07:00 mark you talk about prompt editing and I got a bit confused. Are we able to save images from the sampling steps? If so, how. What is the benefit of using prompt editing? Thank you!
So prompt editing is just deciding at which sampling step a prompt should be used, via either a number or a percentage of the total sampling steps. It's a bit like swapping prompts from one to the next at a stage you specify.The benefit is would be getting a unique image that blends one prompt with another or delaying the start of a prompt until later in the image generation. You will get the resulting image when the generation is complete. Hopefully that helps?
@@BitesizedGenius this does help. I also found a medium article in the topic that helped me understand it better. I realize now it influences the final image, but the step in which it’s introduced will determine how much influence. Thanks again for the great tutorials 👏🏾👏🏾👏🏾
Is there a way to use prompt editing with loras? When I try something like [,keyword1:,keyword2:0.7] it doesn't seem to work. Do you have to do some trick to escape the ':' in the lora? Is it even possible (with an extension or addon?) Thanks so much for this awesome guide! I come back to it again and again! @@BitesizedGenius
Is there a way to control angle and perspective. Like the low angles that imply the viewer of the picture would be small and looking up at a huge object or person that towers over them?
It's a tactic, make the adverts so offensive you're compelled to pay the subscription. I pay for TH-cam only to get rid of the onslaught to my senses from the ads. There should be a ban on embedded ads in the videos for VPNs or whatever! There are also different versions of TH-cam that block all the ads and are free but they don't suit my needs.
The formatting of prompt if very helpful. I tend to rattle whenever I am generating a prompt because I'm overwhelmed with all the settings the A1111 have.
@@BitesizedGenius I've also watched thosoe and added to my playlist for SD. I'm eager to grasp the prompt techniques so I won't have any troubles when I have an idea.
The thing that was frustrating the most for me about generating prompts with how long it took to do that for each prompt. And a lot of the times I wanted to generate multiple prompts with different settings in each one. I know there's extensions in stable diffusion that kind of help with that, but I developed a prompt generator for chat GPT that structures my prompts and I can ask it whatever I want. You should check it out sometime.
This is a really good tutorial, great! I have a problem, the AI keeps creating a duplicated character, like a "twin". I can't resolve it by writing the negative prompt. Please help!
Hard to say based on that information. What checkpoint are you using and do you get this issue when copying the generation data from the checkpoints example image?
Hi! Does anyone know how to fix an error when trying to use x/y/z plot with prompt S/R. It doesn't give me the different images, instead it reads the following error: RuntimeError: Prompt S/R did not find Batman in prompt or negative prompt. Help would be greatly appreciated. Thank you!
Hey, that means that your prompt in XYZ cannot be found inthe prompt box, so check the wording as I think its case sensitive and ensure your commas are in the right place to signify the seperation of prompts
Hello Sir, can you help with my query about Hand and fingers deform ho when creating artwork like extra legs extra fingers. sometime the went missing or not corrects as per human anatomy. I tried negative prompts, but still issue remains.
Token Limit - 'isn't that important' - How? Why? If 75 is the limit what happens to the rest? How are they used? Randomly selected etc from what's left or what happens to them? I've seen people say that 75 is the limit, but on the other hand many example prompts from places like civitai etc are a mile long...especially negative prompts. I'd be interested to know how those long prompts are used by SD.
With stable diffusion there really is no token limit. It reaches that limit and then just resubmits more information. The problem is the longer the prompt the less likely it is to incorporate the details later on in the prompt without emphasis.
Hey, i'll do a seperate video breaking it down in future and showing how it can be useful, but information is scarce, so I just gave the example most commonly found online.
Thanks as always! I have an off-topic question, is there any way to make StableDiffusion not show people but only clothes? I put no human, no girl, etc. in the negative prompt and it still shows people.
So that will likely depend on your checkpoint and most are trained for people. You can either try and find a checkpoint geared towards cloths, train your own or do people and clothes but remove the person through controlnet, photo editing etc.
HI, I am new to SD and I tried those prompts yesterday and my results were nowhere as good as yours even though I used everything you used in your prompt. Can you help? Thank you.
"Octane render' is a really bad prompt for realistic characters as Octane is a 3d rendering program so ( great for volumetrics) but makes figures look 3d and plastic skinned. Just thought I'd point that out. I see it in loads of prompts and don't get why.
@@snickle1980 Yes you see that and "Unreal Engine" less these days. I like the odd ones like "An Arafed image of..." (saw it in a comfyui workflow) so googled it) and used it when stuck.
So confused as to why the StabilityAI team doesn't just explain this stuff. Instead of people having to theorize and endlessly test to find out how to best use their tools.
How to save prompts? i have found myself is a button on down right of skip, and was very hard to find, no one of the beginners tutorial talk about it's incredible the bad UI of Stable difussion like wtf the save button is always on upper left corner!? wtf happen with the AI and all technology LMAO so stupid UI sorry, and like to much and tanks the models are so easy to change on the upper left!? so ituitive to use, like i was changue the model notting happened the Stable difussion even it do not work at all lol
Hey, so i've done a video on the TXT2IMG and IMG2IMG UI, which may help and i've got a video on UI settings on the way. But you can save prompts using the save style button under the generate button. Check my video on where files are saved for useful locations 😁
SD Web UI could definitely benefit from some serious usability improvements, such as saving and loading of prompt settings more than what is currently supported by "Save style". However, Web UI is so amazing that no one is complaining. I plan to look into Comfy UI to see if that solves some of these usability issues.
I don't know why it took so long to find a clear tutorial and explanation online for this topic, ty for secret sauce
shut up
This is fantastic. I work for an AI image gen site, and I'm always trying to explain the principles of prompting and what ( ) does vs [ ], or how to structure a prompt from most to least important. You concisely break down so many things here, thank you!
Break is perfect for describing individual people or items in an image. Never heard of it before watching this, but it was exactly what I needed to have two people of different races and dress styles conversing in the same frame!
Ah good idea... It can also be good for weakening the impact of prompts by placing some padding so it's lower on the weighting!
hey, thank you bro
Thank you for explaining this. Game changer.
This is an excellent and comprehensive guide to Stable Diffusion. Well thought out and informative, thank you for this!
BREAK helps you keep certain sections of your prompt cohesive. Since every chunk is 75 tokens, if you were to run into a situation where one of your token segments was abruptly ended (such as "wearing a white/shirt and blue jeans") it can lead to unwanted results. If you know you're close to your token limit before starting another section (like before the clothing portion), you can put BREAK into the prompt and ensure everything sticks together in the next chunk.
The AI also seems to put more emphasis on things when processed alone. Splitting colours into separate segments can ensure they don't bleed into each other or colour incorrect items.
Thanks, i'll be revisiting this keyword soon in an upcoming video so this reaffirms my research!
Omg best explaination I ever see about prompting of stable diffusion, thanks you for being clear and fast.
lovely - to the point - clear and concise - thank you
Legend - Finally someone who just gives clear concise explanations on how everything works! Subbed
Underrated youtube channel. You deserve way more attention
Thanks Azula!
I have no perfect pipeline before I show this video. Thank you so much @Bitesized Genius
the BEAT explanation of SD that’s out there. Feels like I found an oasis. This vid needs to get more views :(
Great video. My SD setup is quite complex and I'm making some startlingly photorealistic stuff, but even I learned a couple tips here. Thanks for the upload.
I just got into Stable Diffusion and the whole week I was generating images without knowing these techniques. Thank you.
Well, I also have the same journey as you. Generated more than dozen of these for the last week without even knowing what to write for the scene 😂
The format of your videos is so incredible, and the value contained within this video in particular is amazing. Thanks much!
Thanks, let me know if I missed anything as i'm thinking off additional prompting tips!
i never found a video that helped me more than this one keep up the good work
It was a good video. Whenever I search for SD guide, all the videos that came up, only showed how to install it and a little overview. This video gave me a good idea on prompting. Good work 👍
This is fantastic. Thanks for the tutorial.
7:55 Try Put BREAK after each concept keywords and see what happens.
(8:29) Those lines are vertical, not horizontal. Their proper name is "pipes".
One of the more useful introduction to SD prompts I've watched. How do you only have 900 subs?
Thanks, much appreciated!
Excellent video. Very detailed explanations. Thanks for your work.
I use what you call "Alternating Words" all the time. It is one of my favorite tricks for creating realistic looking faces by using a few different real people, but by alternating through 3 or 4 people using this technique in general it won't look like any one of them.
It's also useful for getting some consistency with how characters look, regardless of whether you use a different seed or not!
I give that if I wanted to create consistent faces I sold use a name, but then I found you can just plug in a number instead of a babe, so something like this "a tall woman with brown hair named 5437". Would give me the same face in all the images from that generation
excellent video! thank you so much! learned a whole bunch
Thanks so much for making this, faaaaaar too many people in the space seem to assume everyone can code with Python like a god so none of this has to be explained, yet those like me are sitting there with NO FUCKING IDEA how any of this shit works! lol So again, thanks, just knowing these basics makes a gigantic difference.
Personally, for my characters, I wrote a Python script that uses ChatGPT to generate a bunch of prompts based on a template, and then write the resulting prompts out to a file which I then run through SD using the Prompts from File
Awesome breakdown, I went through and tested each one as I watch this video and I have leveled up for sure and know a bit more how SD advanced prompting works, hope to see more good stuff!
Thanks for the tutorial!
Thank you. Very useful. You got a subscriber and a like.
Very useful ! Thank’s
best guide I've seen yet.
Thanks!
Very informative video. Thanks
Thank you sooo much for making this! 💎✨
How do you do the image with comparison? Like 4 gens or batches with different settings and labels?
Thank you for another very informative and helpful video! It's much appreciated! Regarding the BREAK, it actually looks like a slight bump in quality WITH the break. Nice!
4:57 You may have missed colons in the middle two generations which is why they look so similar, but I think the point still stands fine.
Fantastic tutorial man!
I've been, on and off, toying around with Auto1111 since early December. I've generated thousands if cool images, but... the output consistency is all over the place. Some great, but with lots of crap in between.
I've mostly been winging it because I haven't seen anyone cover the features together well, anywhere, and there are just so many variables that can completely disrupt your output. It's time consuming.
Also, many tutorials assume you already know much of this info and are hence obscure and incomplete lessons.
E.G., I was pushing buttons, changing prompts, and observing the destruction... lol 😆 🤣 So that's lot's of "oh, hmmm, that's a neat, um, FAIL... f...uggggly."
So thanks again! Great content, liked 👍 👌, subbed ❤, saved.
This is really great stuff.
Subscribed! This information has been tremendously useful for my channel.
Loving the content, keep it up!
I'm trying, but the images I see are not that nice and clear. Do I need a plug-in or is there a different setting? I didn't know..
Hi , thanks for the video !
4:30 Does the ":" character not matter?
You use it for the value 1.8 but not for 1.2 and 1.3
It matters, I just didn't notice it was missing 😂
BadDream and UnrealisticDream is a negative embeddings not just simple negative prompt, I think you don't explain it (or I'm missing it?). without proper installing those two negative embeddings in \stable-diffusion-webui\embeddings\ those words don't have a real meaning and don't processed properly. those two files are not LoRA either, it's embeddings, that could be downloaded in civitai /models/72437 (I can't put real URL because it always deleted). from the video, on result log below generated images, I saw those two words are not processed as embeddings as it really intended. if those negative embeddings are used properly, there's a log like this:
medium shot of a woman with full lips, golden eyes and a white crop top, long black hair, snowy weather, octane render
Negative prompt: BadDream, (UnrealisticDream:1.2), (NSFW)
Used embeddings: BadDream [48d0], UnrealisticDream [5f55]
Yes, this is something I missed and I plan to make a video explaining embeddings, LoRa's etc soon, as it's not as simple as copying the generation data. That should then give some context for future videos, without explaining it each time.
Thank you for explaining this so well! I am reinstalling SDXL because my first install was flawed somehow…this makes experimenting with other interfaces like Comfy easier! Such amazing software!
wow i was confused with circular and square brackets and numbers..thanks for clarifying
Hi Bitesized Genius - Regarding the prompt from file/text - how do you generate negative prompts here? Really stuck would love your help thanks!
Thanks for this amazing SD prompt tutorial. Around the 07:00 mark you talk about prompt editing and I got a bit confused. Are we able to save images from the sampling steps? If so, how. What is the benefit of using prompt editing? Thank you!
So prompt editing is just deciding at which sampling step a prompt should be used, via either a number or a percentage of the total sampling steps. It's a bit like swapping prompts from one to the next at a stage you specify.The benefit is would be getting a unique image that blends one prompt with another or delaying the start of a prompt until later in the image generation. You will get the resulting image when the generation is complete. Hopefully that helps?
@@BitesizedGenius this does help. I also found a medium article in the topic that helped me understand it better. I realize now it influences the final image, but the step in which it’s introduced will determine how much influence. Thanks again for the great tutorials 👏🏾👏🏾👏🏾
Is there a way to use prompt editing with loras? When I try something like [,keyword1:,keyword2:0.7] it doesn't seem to work. Do you have to do some trick to escape the ':' in the lora? Is it even possible (with an extension or addon?) Thanks so much for this awesome guide! I come back to it again and again! @@BitesizedGenius
veritical is now horizontal. :/
Is there a way to control angle and perspective. Like the low angles that imply the viewer of the picture would be small and looking up at a huge object or person that towers over them?
use prompts like from below, from above, birdseye view, fisheye lens etc and add weighting if required
@@BitesizedGenius Thanks dude
What's correct punctuation between prompts: comma's or line break?
commas but either will work, line breaks are more for personal formatting so you can view things easier
@@BitesizedGenius in my most recent tests line breaks yields incredible consistency.
TH-cam is really falling off, a 60 second unskippable ad...
There's a new pipe for that.
It's a tactic, make the adverts so offensive you're compelled to pay the subscription. I pay for TH-cam only to get rid of the onslaught to my senses from the ads.
There should be a ban on embedded ads in the videos for VPNs or whatever!
There are also different versions of TH-cam that block all the ads and are free but they don't suit my needs.
@@patrickfitzgerald927 search the word in my post, but without a space. You'll thank me.
Awesome video dude!
Very nice, thanks for your tips 🙂
The formatting of prompt if very helpful. I tend to rattle whenever I am generating a prompt because I'm overwhelmed with all the settings the A1111 have.
Thanks, hopefully this video makes you less rattled. I also did a video on TXT2IMG and IMG2IMG where the settings are explained.
@@BitesizedGenius I've also watched thosoe and added to my playlist for SD. I'm eager to grasp the prompt techniques so I won't have any troubles when I have an idea.
The thing that was frustrating the most for me about generating prompts with how long it took to do that for each prompt. And a lot of the times I wanted to generate multiple prompts with different settings in each one. I know there's extensions in stable diffusion that kind of help with that, but I developed a prompt generator for chat GPT that structures my prompts and I can ask it whatever I want. You should check it out sometime.
This is a really good tutorial, great! I have a problem, the AI keeps creating a duplicated character, like a "twin". I can't resolve it by writing the negative prompt.
Please help!
Hard to say based on that information. What checkpoint are you using and do you get this issue when copying the generation data from the checkpoints example image?
This was very helpful! Thank you :)
How to check image used to learn the checkpoint?
Images from my video are available on Patreon
Great video and channel! You got a new subscriber!
Would love to see how to create characters based on training the model with images of myself.
Hi! Does anyone know how to fix an error when trying to use x/y/z plot with prompt S/R. It doesn't give me the different images, instead it reads the following error: RuntimeError: Prompt S/R did not find Batman in prompt or negative prompt. Help would be greatly appreciated. Thank you!
Hey, that means that your prompt in XYZ cannot be found inthe prompt box, so check the wording as I think its case sensitive and ensure your commas are in the right place to signify the seperation of prompts
Hello Sir, can you help with my query about Hand and fingers deform ho when creating artwork like extra legs extra fingers. sometime the went missing or not corrects as per human anatomy. I tried negative prompts, but still issue remains.
Hey, i've got a video about this on my list so keep an eye out!
Token Limit - 'isn't that important' - How? Why? If 75 is the limit what happens to the rest? How are they used? Randomly selected etc from what's left or what happens to them? I've seen people say that 75 is the limit, but on the other hand many example prompts from places like civitai etc are a mile long...especially negative prompts. I'd be interested to know how those long prompts are used by SD.
With stable diffusion there really is no token limit. It reaches that limit and then just resubmits more information. The problem is the longer the prompt the less likely it is to incorporate the details later on in the prompt without emphasis.
Hey, i'll do a seperate video breaking it down in future and showing how it can be useful, but information is scarce, so I just gave the example most commonly found online.
Nope. I don’t have enough brains cells for this. I’m going back to midjourney 💀💀💀
😂
Where i can download "Lora add details"?
civitai detail tweaker
Thanks as always! I have an off-topic question, is there any way to make StableDiffusion not show people but only clothes? I put no human, no girl, etc. in the negative prompt and it still shows people.
So that will likely depend on your checkpoint and most are trained for people. You can either try and find a checkpoint geared towards cloths, train your own or do people and clothes but remove the person through controlnet, photo editing etc.
HI, I am new to SD and I tried those prompts yesterday and my results were nowhere as good as yours even though I used everything you used in your prompt. Can you help? Thank you.
Are you using a good checkpoint and not the default on installed?
@@BitesizedGenius i figured that out and ima try it. Thanks! I am a total newb
@@BitesizedGenius checkpoints are what "models" are called right?
@@TheMagista88 No worries, check out my other videos, some gems in there for beginners :)
@@BitesizedGenius def sybbed. Thanks ;)
Nice work
SD works with easy subjects such as women's faces. But when you apply it to something that doesn't already exist it really struggles.
Like for Manbearpig reference ;)
Very nice video, how do i remove the background bokeh from human subject
try putting bokeh into negative prompt? or writing sharp in positive
@BitesizedGenius Didn't work but thanks, I use Fooocus btw. I put bokeh,depth of field in negative and small aperture I'm positive still no dice
In the prompt weighting section you didn’t put colons so it didn’t accurately interpret it exactly the way you intended
what's this software tho?
th-cam.com/video/WJUsETmv_r8/w-d-xo.html
@@BitesizedGenius thanks :D
Dude.. man-bear-pig!! LOL :D
i just understood prompt from file/text XD
It's self explanatory but it also doesn't explain much🤣
@@BitesizedGenius 🤣
bad dream is a negative embedding and you need to have that file before it will even do anything.
Thanks!
Thank you
"Octane render' is a really bad prompt for realistic characters as Octane is a 3d rendering program so ( great for volumetrics) but makes figures look 3d and plastic skinned. Just thought I'd point that out. I see it in loads of prompts and don't get why.
In the early days of stae diffusion, "octane render" was a hugely popular prompt.
Old habits die hard. Most people arent that creative.
@@snickle1980 Yes you see that and "Unreal Engine" less these days. I like the odd ones like "An Arafed image of..." (saw it in a comfyui workflow) so googled it) and used it when stuck.
So confused as to why the StabilityAI team doesn't just explain this stuff. Instead of people having to theorize and endlessly test to find out how to best use their tools.
This is sub-worthy content for sure.
Thanks
You should get a pop filter on your microphone.
I have one on my microphone, but i'm clearly doing something wrong with my audio set up, so i'll look into it 😊
vertical line
Mine still look like nightmares, lol
"Yee Hee!"
How to save prompts? i have found myself is a button on down right of skip, and was very hard to find, no one of the beginners tutorial talk about it's incredible the bad UI of Stable difussion like wtf the save button is always on upper left corner!? wtf happen with the AI and all technology LMAO so stupid UI sorry, and like to much and tanks the models are so easy to change on the upper left!? so ituitive to use, like i was changue the model notting happened the Stable difussion even it do not work at all lol
Hey, so i've done a video on the TXT2IMG and IMG2IMG UI, which may help and i've got a video on UI settings on the way.
But you can save prompts using the save style button under the generate button. Check my video on where files are saved for useful locations
😁
SD Web UI could definitely benefit from some serious usability improvements, such as saving and loading of prompt settings more than what is currently supported by "Save style". However, Web UI is so amazing that no one is complaining. I plan to look into Comfy UI to see if that solves some of these usability issues.
No matter the prompt or setting for me the photo comes out to pure pixelated garbage
Why you speak in a question?
A lot of words but I feel like nothing has been explained lol.
ai photo movie
th-cam.com/video/uP04emczDi8/w-d-xo.html