Prompting HACKS no-one talks about

[@vanhoe0]

I'm going to explain the BREAK keyword from my understanding, please correct me if I'm wrong as I'm not a programmer. So the original Stable Diffusion is built on a 75 token limit, which means that the things that you wrote after the 75 token will be ignored. Being open source, the A1111 community found a way to increase the limit by modifying it to divide the prompts into chunks of 75 tokens. (1chunk=75tokens). But there's a problem, the chunks does not divide the prompts as we intended. As an example: "woman eating blue car" Assuming the prompt above will be counted as 100 tokens (1 word = 25 token), which is more than 75 tokens. So SD will try to divide the prompt into chunks to fit into 75 token. It might led the prompt to be split into chunk and chunk. Which will probably generate an output of a woman eating a blue thing with a car on background. We don't want the car on the background, we want to see the action of eating a car. So we use the BREAK keyword to help the AI to better understand our prompt. In this case we use BREAK to try and separate the prompt in a much more understandable context for the AI: "woman BREAK eating blue car" Now the chunks will be and , which generates the output closer to what we intended. Hooray! As a side note, the words/tokens on the start of the prompt weigh higher than the end of the prompt. So if you write a long prompt and wants the AI to prioritize some of the sentences at the end of the prompt, you can BREAK them up. Example: masterpiece, highres, highly detailed, best quality, original. by greg rutkowski If you're having problem summoning rutkowski's style on your output, you can try adding BREAK before "by greg rutkowski" to put more weight to it.

[@Not4Talent_AI]

Pinning the comment cuz it is really well explained, thank you!!

[@interestedinstuff]

Excellent. I didn't know about the SD training input limit. Explains why some other SD programs limit the input string. FYI Tokens for words are not per character. They are roughly per syllable. So broken is two tokens 'bro' and 'ken'. The LLM's tokenise so that words like Disable and Disaster aren't completely different. The LLM learns that 'dis' appears in certain types of words. While 'able' appears also in many words. Lots of words are concatenated from smaller words, as you've prob noticed.

[@mrhoangpro]

I thought BREAK is in prompt regional, but it turns out Prompt Regional using BREAK

[@phadrus]

This was incredibly helpful. Thank you.

[@maitotechlab9035]

Wow, adding a few breaks into my prompts made the output be way closer to what is expected, and i use an online service.

[@generalawareness101]

Concept bleeding is my bane of existence with AI art.

[@Not4Talent_AI]

hahhahahha

[@clearandsweet]

Bruh I learned like 20 different things watching this video I swear your channel has the most useful and practical information for SD on KZbin. Thank you so much for all the videos.

[@Not4Talent_AI]

Really happy that you learned and liked it! Thank you so much for the super positive comment!

[@not-a-lot-of-options]

Took me a few minutes to understand when I first learned of the [] stuff, and I found this description helpful to understand how to use it. [Initial Prompt:Final Prompt:When to Change] [dog:cat:0.5] halfway through stop generating dog and start generating cat You can also leave out the initial or final prompt to start/stop generating a certain prompt after a certain number of steps [dog::0.5] dog is ignored AFTER half way through generation [:dog:0.5] dog is ignored BEFORE half way through generation

[@Not4Talent_AI]

Yes!! I explain it in another video and disregarded that in this one... maybe I should had explained it here as well xD. Thanks for the easy to follow explanation. I'll pin the comment for now if anyone need the info

[@not-a-lot-of-options]

@@Not4Talent_AI All good! I also realised you can nest them. [[elephant:dog:5]:cat:10] Generate elephant, at step 5 switch to dog, at step 10 switch to cat. I can't think of any practical use for this, and that example prompt is generating some nightmarish results lol, but it's interesting nonetheless.

[@Not4Talent_AI]

Damn, you just reminded me of something I wanted to try for this video and completely forgot XD glad you did! Thanks

[@lefourbe5596]

@@not-a-lot-of-options when i was trying to generate more variety of woman for classification on a certain style. i remembered your tips. [woman|1girl] or [woman:1girl:0.2] can give a good balance of young adult with anime style as output and when i input a style lora even later like [[woman:1girl:5]:"LoraStyleKeyword":10] it will help acomplish a good pull of diverse stylized character and poses regardless of style constraints. it's quite hard to get good looking classification images for training. this help the model not sacrificing quality for variety.

[@leluch1616]

Gracias por todos tus consejos bro, mejore un 3000% y me ayudo a entender cosas complejas!!!! super recomendado sigue asi saludos desde Ecuador!

[@Not4Talent_AI]

Muchas gracias!! Me alegra oir eso ajjaja saludos desde españa!

[@WallyMahar]

Your eyebrows and stubbl are god-like. Oh yeah, great prompting stuff.

[@Not4Talent_AI]

😂😂😂😂😂😂 thanks!!! (For both ahhaha)

[@Kalafina2CSUG]

Best channel for taking AI art to the next level!

[@Not4Talent_AI]

Glad you liked it!! Thanks!!!

[@codenamewolf1358]

You really blew me away with how concise, thorough and useful this information was. Subscribed and hoping to see more up to date SD content!

[@Not4Talent_AI]

Thank you!!! I hope to do so! I'll try my best for sure :3

[@hutukaTB]

Yo man this channel is the gold mine of SD knowledge, hope it will blow up soon man!

[@Not4Talent_AI]

It did! Thank you so much for the support!!

[@phadrus]

Most channels scratch the surface and talk about the same basic things, it’s useful, but these videos are in a class of their own. Deep dives while still simple and easy to understand. Nice work.

[@Not4Talent_AI]

Thank you so much!! glad they help

[@testales]

Well done, I didn't expect there are still quite a few things I didn't know about prompt editing!

[@Not4Talent_AI]

haahha thanks! Good thing about AI is that there is a lot of room for experimenting xD

[@johiny]

I like more pausing videos but all the tips were great, i will test them tonight!

[@Not4Talent_AI]

I'll keep it in mind!! tyty!

[@suma-xx4jf]

Found your channel yesterday, even tho you only have 6 videos until now, they helped me a lot. happy to be your 1k subscriber, keep the good work :)

[@Not4Talent_AI]

Thank youu!!

[@boltvanderhuge8711]

Another good trick relating to prompt editing is delaying negative prompt refinements, like "distorted face" or "missing fingers", to keep them from influencing the composition or content

[@memoryhero]

The depth of control you put on offer in this video rivals whole series' from other channels.

[@Not4Talent_AI]

Thank you so much!!!!

[@MilanKnolTwee]

This is the best ai art generating channel!! Thank you for this info 🎉

[@Not4Talent_AI]

Hahahhahah tyty!!!

[@MugiwaraRuffy]

Bro, this BREAK prompt is a total game changer. Awesome tip!

[@Not4Talent_AI]

Ty!!! Dont get your hopes too high tho, it still messes up depending on the prompt😂

[@MugiwaraRuffy]

@@Not4Talent_AI Well, I still got some better results with a couple of BREAKS then without, using the exact same prompts. So yeah... :D

[@neosora9912]

Found your channel yesterday, instant subscribe, keep on the good content my friend👌

[@Not4Talent_AI]

thank you so much! will try for sure

[@PyrelordPazuzu]

Ayyy we finally get to see that ice cream prompt! I have used AND before but only for composable lora (which I see someone explained on a comment here already) so seeing a different application of that is definitely interesting enough to experiment on.

[@Not4Talent_AI]

hahahhaha yes! It was a test to see if anyone would say "bla bla AND bla bla". I have to try Latent couple still, will start testing today or tomorrow probably

[@unknownuser3000]

Great video I liked the style. Subbed and looking forward to future tips.

[@Not4Talent_AI]

Thank you so much! Will try to keep providing the best tips I can😁

[@thomasnurse792]

Huh, these are actually great tips, I wasn't aware of any of them. Thanks!

[@Not4Talent_AI]

Happy to hear that! Ty!

[@MikeRenouf]

5:00 wait a second... you're saying SD sometimes includes prompt elements from earlier image generation tasks, even if they're not in a current prompt?!? That's WILD! 😲

[@Not4Talent_AI]

yep, mainly when using Xformers and it gets noticeable if you havent restarted SD in a while. It's recommended to close and re-open the cmd every now and then

[@justinharubra]

Great guide! Loving the new pace, easier to understand, good examples, coherent explaining. Thank you for taking your time to teach us all this useful info! Would also like to add a shortcut, if you press and hold ”ctrl” while pressing up and down buttons, you can control the weight of the prompt that the cursor is on. Can also take multiple words if they are selected. This turns "amazing" ----> into "(amazing:1.3)" - with just "ctrl" + up button x3. And also "best quality" ------> into "(best quality:1.4) with "ctrl" + up button x4.

[@Not4Talent_AI]

Thanks!!! Yes, I actually explained it in another video, maybe should have done it here too. Already have a pinned comment, but useful info for sure. Ty!

[@nio804]

Have you tried the "neutral prompt" extension? It adds two variants of the AND keyword that are less likely to result in a mess. The regional prompting extension is also fun. In prompt mode, it can create a mask from the area of the image affected by a part your prompt (eg. long hair) and then apply another prompt to *only* that area, so you can "replace" long hair with blue hair and the effect is localized to whatever part of the image was "long hair"

[@Not4Talent_AI]

I still have to try them fully,but I plan on doing it! Didnt know abou the neutral prompt extension though, sounds really interesting. Thank you!!!

[@Shabazza84]

Excellent insights. Thx bud. The keyboard blening thing to control concepts in and out is just such a valuable thing to understand. And you really nailed it with cmpact and good explanations!

[@Not4Talent_AI]

thank you so much for the kind words!!

[@zizyip6203]

You can use hire fix prompting and steps to do the bleeding or blending to. Just set the set for the results you want.

[@thevoid6756]

I am gonna come back to this video a lot!

[@Not4Talent_AI]

hahahahah hope it helps!!

[@ianwilliams7740]

I thought I was going mad with prompt-bleeding, after generating tens of images .. they started definitely driting but being influenced by what came before.. i will try and do a restart a bit more often

[@Not4Talent_AI]

Yep! If you have the --xformers command in args it is more common (i think). Even tho its still worth to have it😂

[@Cadmeus]

Thank you once again for sharing your knowledge, your videos are some of the best out there. Your tips on prompt bleeding are very useful, especially when you start to notice tokens inside a word affecting the image (e.g. "asian" in "caucasian" giving the wrong ethnicity)

[@Not4Talent_AI]

Glad it was useful!

[@lanaya9tf]

Appreciate your efforts to explain everything. Thanks a bunch

[@Not4Talent_AI]

Thank you for the support!

[@johnkrumpotick8659]

Lots of great tips, I like the high speed overview too.

[@Not4Talent_AI]

Thanks! Glad it helped

[@interestedinstuff]

Superb video again. The prompt step parameter I had no idea about. Gracias por todos.

[@Not4Talent_AI]

hahahaha de nada, gracias a ti! xD

[@RealPapershaper]

I've been using SD since December and I've never known about these advanced techniques, thank you so much man.

[@Not4Talent_AI]

Happe to hear that! Thank you

[@HermitagePrisoner]

Should be an official SD tutorial 😎Many thanks!

[@Not4Talent_AI]

Haaha tyy!!

[@MariusBLid]

Very cool! Did not know this 😀

[@Not4Talent_AI]

Hope it helped then!! ty!

[@luke11pl1]

I never really comment on youtube videos but i gotta say this channel is a hidden gem! you deserve more views and subs!

[@Not4Talent_AI]

Thank you!!! I'm really shocked at the amounts of support it's been receiving. Really happy with it!

[@luke11pl1]

@@Not4Talent_AI keep doing the great job you've been ! I am kinda new to the whole thing but never seen any info that you give here on other channels😁 really appreaciate

[@Not4Talent_AI]

@@luke11pl1 Glad it helped !!! tyty!

[@Shabazza84]

This knowledge also saved me from going ballistic whenever I try to make a portait format image. Normally, as soon as you go like 9:16 -ish format, the AI will spit out characters that have a selfie pose, no matter what you do. With the hand that usually holds the phone right up your face. Because selfies have this format and there is a ton of selfies on the net the AI saw during training. And if you then don't know why that is happening, you start ripping your hair out over it. (I did). And the "heterochromia" hack is a life save to this day. (the whole concept described here tbh)

[@Not4Talent_AI]

ahahhaahha happy to hear!! tyty

[@fredmcveigh9877]

You are fantastic .I watch others that do videos about Stable diffusion and they are very , very good .But You my friend are so into the prompting that I am amazed .Even your last tip about dropping the picture into the positive prompt was incredible and you said it like it was no big deal .Wow ,thanks a lot .

[@Not4Talent_AI]

Makes me really happy to read comments like this. Thanks a lot!! Glad that it is helping :3

[@saqlainrazee4607]

Very nice and informative video. I learn about keyword blending in your previous video and learn how this work. Now I am using chatgpt to create prompts and its give amazing results. Thank you.

[@Not4Talent_AI]

Smart idea! happy to help! Thank you

[@jason-sk9oi]

Could you please post what you prompt ChatGPT with?

[@saqlainrazee4607]

@@jason-sk9oi which one stable diffusion prompts or chatgpt prompts?

[@jason-sk9oi]

@@saqlainrazee4607 ChatGPT to create the SD prompt. Thank you 😊

[@kenjin5758]

Hi, here is one I use (with variations depending on the model and what I want): Work like a charm: Suppose I use an AI that outputs an image when I enter a prompt. For example, here is an example of a prompt that produces a beautiful image of a girl standing in a gentle breeze. 「masterpiece, best quality, ultra-detailed, illustration, 1girl, solo, sitting, an extremcly delicate and beautiful, beautiful detailed sky, eautiful detailed eye, side blunt bangs, hairsbetweeneyes, close to viewer, breeze,Fying splashes,Fying petals, wind」 The prompt requires the following elements. (1) Words to describe the method of depiction, such as watercolor or oil painting (2) Words to describe external features, ornaments and belongings (also specify colors and patterns) (3) Words to describe age, hair color, hairstyle, hair length, eye color, eye shape, facial expression, breast size, and clothing (4) Words to describe background details (5) Words to direct the pose from head to toe can also be added to the prompt. Be sure to include "masterpiece, best quality, ultra-detailed, illustration, 1girl, solo," at the beginning of the prompt. Please use plain text in your response, no "" or line breaks. Prompts may use danbooru tags. Based on these, please tell me a 150-element prompt that can generate illustrations for the following themes. - - More experimental: Suppose I use an AI that outputs an image when I enter a prompt. For example, here is an example of a prompt that produces a beautiful image of a girl standing in a gentle breeze. 「masterpiece, best quality, ultra-detailed, illustration, 1girl, solo, sitting, an extremly delicate and beautiful, beautiful detailed sky, eautiful detailed eye, side blunt bangs, hairsbetweeneyes, close to viewer, breeze,Fying splashes,Fying petals, wind」 The prompt requires the following elements: (1) Words to describe the method of depiction, such as watercolor or oil painting (2) Words to describe external features, ornaments and belongings (also specify colors and patterns) (3) Words to describe age, hair color, hairstyle, hair length, eye color, eye shape, facial expression, breast size, and clothing (4) Words to describe background details (5) Words to direct the pose from head to toe can also be added to the prompt. Order works best with: content type > description > style > composition - Content type: What type of artwork you want to achieve? Is it a photograph, drawing, sketch, 3D render..? - Description: define the subject, subject attributes, environment/scene. The more descriptive you are with the use of adjectives, the better the output. - Style: we’ve seen the most common ones above, but there are also “sub-categories” - lightning, detail… - Composition: it refers to aspect ratio, camera view and resolution. Be sure to include "masterpiece, best quality, ultra-detailed, illustration" at the beginning of the prompt. Prompts should use danbooru tags but are not constrained to. Based on these, please tell me a 150 to 250 elements prompt that can generate illustrations for the following themes: - -

[@aliciaandrade1148]

the last tip blew my mind, great videos

[@Not4Talent_AI]

hahhahah glad it helped. Thanks!

[@TyronLolXD]

Your channel is so good, I watched all of your videos in full length and the value is insane, and super straight to the point with some jokes here and there. Interesting stuff here bro. Great job! You have a new subscriber.

[@Not4Talent_AI]

Thank you so much!!!! Glad you like it😄

[@UNKNOWN-kg7nx]

thanks, super helpful

[@Not4Talent_AI]

Happy to hear that! Thanks!

[@jessepacoma4219]

Whoa these are so useful. Thank you very much for taking the time to make these videos!

[@Not4Talent_AI]

ty for watching!!

[@LouisGedo]

4:23 Very clever!

[@Not4Talent_AI]

ty! haahhaha

[@350zsean]

Great vids!

[@Not4Talent_AI]

Thank you!

[@matthewhutchinson1247]

Yup you earned a sub! Great video

[@Not4Talent_AI]

thank you so much!!

[@mikeulator]

Thanks for this, GREAT tips. Do you mind sharing the prompts you used to generate the examples starting 5:45 ?

[@Not4Talent_AI]

sure! bath: RAW photo of a [top view photo of a bowl of soup : (aerial perspective:1.2) of bath, woman bathing, full body, covered with a towel : 5], Best quality details, realistic, High definition, High quality texture, epic lighting, Cinematic film still, 8k , soft lighting, anime style town: RAW photo of a [top view photo of a bowl of soup : top view of a cute village, snow, videogame, low poly, muted colors: 6], Best quality details, realistic, High definition, High quality texture, epic lighting, Cinematic film still, 8k , soft lighting, anime style ship: /this one I tried replicating and it was created because of the bleeding on past prompts, so it will probably not work but/ RAW photo of [vector art, illustration, logo, a bowl of soup, (simple icon:1.2) : landscape, pirate island, pirate ship : 3] , masterpiece, Best quality details, realistic, High definition, High quality texture, epic lighting, Cinematic film still, 8k uhd, dslr, soft lighting FlowerPot: RAW photo of a [closeup photo of a flower pot, bouquet of flowers, product photography, dark background, standing : fireworks, night sky, firework show making shapes, flower-like firework, (explosion:0.7) of color : 9], Best quality details, realistic, High definition, High quality texture, epic lighting, Cinematic film still, 8k , soft lighting, anime style ice cream:RAW photo of [ food photography of an ice-cream cone, strawberry, soft lighting, mint and chocolate, tasty : minecraft, snowy mountains, spiral path, overcast, sunlight:8] , masterpiece, Best quality details, realistic, High definition, High quality texture, epic lighting, Cinematic film still, 8k uhd, dslr, soft lighting donnut stuff: [Food advertisement for an blue donut, food focus, food photography, solo, : Photo of a woman, going through a portal, picking through, half body, climbing up from another dimension, cy-fi, beautiful futuristic outfit, looking at viewer, space, glow, portal emitting light: 11], Best quality details, realistic, High definition, High quality texture, epic lighting, Cinematic film still, 8k , soft lighting, anime style

[@mikeulator]

@@Not4Talent_AI Thank you so much! I tried the same thing not long before finding your video but I did not have the same kind of luck with blending. Could have just been my subjects or trying to be too complicated. I should have examples to compare and learn from now, I appreciate it!

[@Not4Talent_AI]

@@mikeulator nice! It is really weird. in my experience if the first prompt doesnt give super predicable results, it is very hard to know what will happen😂

[@twistedbydsign99]

Its interesting to me that the prompt bleeding bug ended up being a feature on mid journey when people wanted to create consistent characters. They would generate lots of the same character, then try to set the scene with them

[@Not4Talent_AI]

Really? 😂😂😂😂 i had no idea

[@TheDanielsSk]

I'm surprised at how great your videos and you're at explaining stuff, no need to go faster, take your time! Waiting for any video you put out Você ainda vai longe no KZbin tenho certeza

[@Not4Talent_AI]

Had to translate that part hahaha but thank you so much!! Appreciate the support :3

[@thomassynths]

Dude your pacing is great, dont promise to go faster! High quality information. Also in case it's not clear, prompt editing is supposed to be fully recursive (barring bugs). I also had tried prompt editing on loras. I will take a look at the code to see if it is easy to support it.

[@Not4Talent_AI]

Ohhh someone that knows code and can check it??? Legend. It would be great! Thank you for your feedback and the comment!

[@CienWill]

This is FK gold, awsome content man, subscribed.

[@Not4Talent_AI]

Thank you so much!!!

[@artemgrauberger8775]

good videos man, keep it up!

[@Not4Talent_AI]

thank you so much!!

[@luke.perkin.inventor]

These are really good tips!

[@Not4Talent_AI]

Thank you! happy to hear that :3

[@prettyawesomeperson2188]

You should get into the habit of uploading your example photos, and results from your prompt editing, so we can copy and paste the settings.

[@Not4Talent_AI]

Do the metadata get uploaded? It is a really good idea if so ngl. The prompt editting prompts are further down in the comments. Someone asked for them before, if you are interested!! But thanks, didnt think of that idea!

[@JerryAyres687]

Nice tips! Thanks!

[@Not4Talent_AI]

Ty for the kind comment!

[@kyleaguas5218]

Tip #5 actually has a pretty cool use for it in the "Latent Couple' extension. Basically it splits your image into segments and using "AND" to segregate each prompt for e.g. prompt: a castle background AND a woman posing for a photo in a model shoot style AND a man wearing a suit While the prompt above is bs at the moment, Latent Couple takes the initial prompt, a castle background, and generates a background. It then generates the 2 other prompts at their specified segments. This is a very half-ass explanation but hope you look into it!

[@Not4Talent_AI]

OOOOOOOOOHH that makes sense, I'll try it!! Ty!!

[@lefourbe5596]

I Saw aitrepreneur talking about this extension. It's really cool but my slug brain wasn't able to make it work. It should work with differents Loras too (with another extension : composable Lora) wich should be great for character interaction. Yep that's deep !

[@Happ1ness]

6:50 There's an extension called lora-block-weight, which allows you to have more control over LoRA's. It's new to me, so I can't give many advices. If you want, you can look for it, there are some videos on Japanese, showing the usage of it.

[@Not4Talent_AI]

Oh, I'll check it out then, thanks!

[@el33tkrew]

Would it be possible if you could give more examples? Thank you

[@Not4Talent_AI]

Sure! I'll try. Tyty!

[@Jfreek5050]

Hoo boy I must really be behind with SD because when you brought up that kewyword that impacted the steps [city:3] I was like *is this common knowledge why did he not elaborate on this*

[@Not4Talent_AI]

My bad! Thought I explained it better. It isn't common knowledge, but it can be deduced. When you think about it, the word "city" is usually portrayed with exterior images, this means that in training, the AI will most likely create a tendency to make "city" prompts into exterior photos. And to combat that we do as in the video. All words have stuff like that, I just chose city as an example, but it can happen with other stuff.

[@CasperLind1]

thx this makes sense and i would never have thourght about it

[@Not4Talent_AI]

hope it helps !

[@GosuNoKami]

Can you please make a video on unprompted? Also on control net mixed with other extensions?

[@Not4Talent_AI]

Sure!! I'll put it on my list! The next video is about controlnet + latent couple. Then other extensions like openpose editor. I will make a video on wildcards, unprompted and this type of deals on the future too. Ty for the suggestions!

[@GosuNoKami]

@@Not4Talent_AI awesome! Thanks

[@cyrille8693]

It"s funny that I'm watching your video today and earlier I was watching a video on Nerdy Rodent yt channel called "A1111 Stable Diffusion Extensions You Might Have Missed!" and AND seems related to an SD extention called neutral prompt which.... obv exploit the bleedings prompts you're talking ab in your video. I can remind that sad sunday afternoon when I had to struggle with a character running into a city under alien attack and he couldn't help to wear a short and a top with a bib number like "My boy, you cannot run if you don't have a number!". Don't ask me anything ab Neutral Prompt btw, I'm not sure what AND AND_PERP AND_SALT really implies. And thanks for your video, btw, very inspiring when it comes to exploit bleeding keywords.

[@Not4Talent_AI]

Hahahaha thanks!!! Ill look into that extension, cuz I heard about it but never actually tried. And yeah, bleeding is soooo annoyng 😂

[@Rasukix]

AND is supposed to be used with latent couple to split the prompts per item etc, I think that's right at least

[@Not4Talent_AI]

It is used in latent couple, but it wasnt made just for that. But you are right in that, Latent couple uses AND, probably through the fact that and creates mutiple images through 1 prompt. Then it tells it where to mix each image based on the mask input, I guess😂😂

[@Rasukix]

@@Not4Talent_AI it may even be getting confused based on the same word being used for these 2 separate entities

[@Not4Talent_AI]

@@Rasukix Could be😂

[@luxecutor]

Thanks!

[@Not4Talent_AI]

oh wow, thank you soo so much!!!

[@stickmaker10111]

If I were using diffusers instead of the webui, is AND/BREAK a feature of the text encoder/clip?

[@Not4Talent_AI]

ngl I have no idea (I dont think so tho)

[@justinwhite2725]

BREAK is similar to the :: operator in midjourney. AND makes auto1111 consider the image to be both things. So it's trying to render it as each of those things at the same time... With interesting results.

[@Not4Talent_AI]

Had no idea abou :: for mj. I think someone asked about it before... would have to check when I can

[@S4SA93]

Mate you just made another pinaccle video on all of this stuff. You should probably make/sell a pdf book with all your knowledge it's so nice, in-depth but still crisp and compressed, really amazing

[@Not4Talent_AI]

Thank you!!! Yes, srry! Tried making the short but it was too much stuff in one video. I decided I'll do that for a complete full video. The next one (I literally started thinking how to approach it today hahaha. I'll try to have multiple characters in different positions and using everything we've learned so far. see how it goes xD)

[@S4SA93]

@@Not4Talent_AI You're the best ;) I can't wait!

[@Not4Talent_AI]

@@S4SA93 ty!! Hope to deliver XD

[@renealbrechtsen9743]

Could "AND" be used for more consistent characters in a picture ? For example if one character has a green body and orange hair and the other had a blue body with pink hair ? Or would "BREAK" be better in that case ?

[@Not4Talent_AI]

Break would be better. Even though i'd highly recommend to just use regional prompter. And will just mix both characters into one probably

[@renealbrechtsen9743]

@@Not4Talent_AI how can I do regional prompting ?

[@Not4Talent_AI]

It's an extension. If you are on a1111, go to "extensions > available > -look for "regional"- and install an extension called "regional prompter". Then apply and restart the UI. You can also install it with a link, using this one: github.com/hako-mikan/sd-webui-regional-prompter.git @@renealbrechtsen9743

[@renealbrechtsen9743]

@@Not4Talent_AI I'm on comfyui.

[@Not4Talent_AI]

Oh, then no idea 😂😂 i guess there will be something similar to it tho

[@kenjin5758]

Hello, thanks again for the tips ! I like the format like last time. Taking your time and focusing on quality will help you in the long run. From what I understand by reading articles, most AI that use text as input functions with token as vectors (see in webui the numbers close to your prompts input). Vectors like we learned in math. The difference is that if we learned operation like 2 or 3 dimensional vectors, the AI can operate on N dimensional vectors, far more complex. Concept consist then on finding the location of our vector in this N dimensions space, then selecting near vectors that will mathematically generate images that looks like what we want. So, the further your two concepts (or N concepts) are, the more the "average" will be inaccurate and blending will be difficult. I'm therefore extrapolating that the keywords BREAK / AND and prompt editing generally tend to give the human an indirect way to construct more coherent vectors / group of vectors instead of letting the AI decide. Could be way too much extrapolation so feel free to correct me if you know better, I'm not an expert. See scholar.harvard.edu/files/binxuw/files/stable_diffusion_a_tutorial.pdf: "Understanding prompts - Word as vectors, CLIP" part.

[@Not4Talent_AI]

I have no idea XD What you said makes sense tbh, could be! I'd love to talk with someone who makes AI at some point, to talk about stuff like this on the channel. And ty for the feedback!

[@johiny]

great resource friend thx

[@kenjin5758]

​@Not4Talent yes, that could be awesome !

[@lefourbe5596]

Clear explanation ! Good work !

[@rebootjac]

6:52 Are you sure about that? I have had good results with simply highlighting the name of any of the three in a prompt: TI, model, or Lora reference, holding down CTRL and using either the up or down arrow on the keyboard to add emphasis or de-emphasize that part in the prompt. SD will then wrap that highlight with parentheses and tack on a number. Should look like for example this: (EasyNegative:0.5) EXCELLENT video btw, didn't know most of this already.

[@Not4Talent_AI]

I'm not 100% sure, but weight variations do work for sure, so yes, this: " (:1.3) would absolutely work. What I tried was: "[ : : 10 ]" Which didn't give me good results. or even ""

[@rebootjac]

@@Not4Talent_AI Interesting! Your syntax doesn't work for me either but when I use [ :: : #] that does produce different results when I change the step amount AND lock the seed. Same thing if I put the colon after the first bracket as well.

[@Not4Talent_AI]

@@rebootjac it did for me too, but the results chenged ina very negligible way. Have to experiment more though. Ty for sharing your results!!

[@GooseAlarm]

I knew prompt bleeding was a thing. I didn't know it had a name though. It’s like the AI remembers what was just prompt and keeps recreating it.

[@Not4Talent_AI]

Tbh I made up a name cuz didnt find anything online. Some people call it bias. I think bleeding sounds way cooler tho. Unless you mean the bleeding that goes from the previous image to the next. In that case bleeding is the actual name

[@AI_Rhythms2]

"she is the blade of miquela btw" AND (xdxd) man I love your videos :)

[@Not4Talent_AI]

Hshahahh tyty!!!

[@interestedinstuff]

Point two about mixing two prompts. How do you get the second prompt to start at a certain point. I've watched the video a few times and I've missed the bit that says, put prompt 1 here, then set this value, then put prompt 2 here.

[@Not4Talent_AI]

in prompt editing there is the [word1:word2:changeStep] method you basically use the same but like this [prompt1:prompt2:changeStep]

[@losing_interest_in_everything]

Do you think we can use OR and WITH in our prompt like we use AND or BREAK? For me it would make sense because they are also keywords in Python

[@Not4Talent_AI]

Are they!!??? WTF I'll try it asap (edit: I've tried them and there doesn't seem to be a difference between WITH/OR and with/or. Which there is on BREAK/AND vs break/and)

[@losing_interest_in_everything]

@@Not4Talent_AI It could have been interesting if it was ^^

[@Not4Talent_AI]

@@losing_interest_in_everything Yeah, got excited for a second hahaahah

[@nubeslocas]

at 5:24 I don't understand how to write prompts using these technique about mixing different prompts. Is it using [promptA:promptB:5] ?

[@Not4Talent_AI]

yes! It is exactly that. Can be later than 5, depending on total sampling steps. But you got that right

[@nubeslocas]

@@Not4Talent_AI thanks a lot for your videos and responses!

[@Not4Talent_AI]

@@nubeslocas ty for watching!! I'll always try to respond, if you have any doubt just ask :3

[@feilox]

Hey for concept bleeding XXX:3 and XXX:8, is that out of 15 steps or 25 or 1-10? I need to know the ratio.

[@Not4Talent_AI]

Steps: 30, Sampler: DPM++ 2M Karras

[@pastuh]

I like to use AND, it's like the same prompt but iterates over all these keywords again. And I never describe long sentences, because some models are trained differently.. some only keywords, some only actions.. BREAK just proves, because all tokens is counted.. Same as GPT.. Tokens is chunk of word or whole word..

[@sherpya]

it’s mindblowing nobody else talking about prompt editing and the reason of bleeding, that also affects midjourney, but unfortunately no prompt editing 😅

[@Not4Talent_AI]

Yeah, tbh don't really know how to combat it in midjourney. But at the same time, midjourney is pretty good at understanding prompts

[@Ekkivok]

i have a question, do you a tip can create two characters based on LORA model on the same image ? is there a prompt for that ? thanks for your answer :)

[@Not4Talent_AI]

Hi! It is pretty pretty hard id you want to get them both in just one image. There is the option of adding both loras and tryingto use BREAK and lowering the loras weight a lot. But it wouldnt work that good. Latent couple can get you halfway there, but even still it usually just implodes when using more than 1 Lora. I actually acomplish something like what you want in this video though. Even if it was through some inpainting kzbin.info/www/bejne/l3PMeIx_rNSBlc0

[@Ekkivok]

@@Not4Talent_AI thanks for your fast replie !! i will check that now ^^

[@Aaisn]

Isnt BREAK the same as emphasis weight like 0.9 or lower or higher?

[@Not4Talent_AI]

No, break is a separation/reset of tokens. (Tokes at the beggining of the promt have more importance. Break is like having 2 begginings, pretty much)

[@Aaisn]

@@Not4Talent_AI damn nice information

[@LouisGedo]

5:15 Wow.... that skeletal hoodie person image is very cool........ if you achieved it, would you be able to give the recipe for doing it? P.S. Oh, 5:22 might be the prompt....... I'll continue watching!

[@Not4Talent_AI]

ofc!! I made it in this video kzbin.info/www/bejne/b3mToIKbrdeoprs this is the metadata for it: RAW photo of a [sexy woman : skeleton: 10], inextreme cold, Best quality details, realistic, High definition,High quality texture, epic lighting, Cinematic film still, 8k , soft lighting, anime style Negative prompt: EasyNegative, closed_eyes, lowres, letterbox, deformed, out of frame, ugly, morbid, mutation, malformed limbs, missing arms, miss finglegs, watermark, loli, cat ears, Steps: 40, Sampler: Euler a, CFG scale: 7, Seed: 2864224920, Size: 512x768, Model hash: 272365759d, Model: ivory_v12, Regional Prompter: "mode:Horizontal,divide ratio : 1,1, Use base : False, Base ratio : 0.2, Use common : False, Use N-common : False" (you can copy this in your positive prompt and then hit the blue icon under generate)

[@lefourbe5596]

This is what i like on small Channel. WE are mostly advanced user who search for UP to date perfection and here a newcomer in the YT game is able to provide every tips needed on time ! I hope you don't overwork yourself and get tired so soon. A community provide great motivation. but don't let it be a weight too heavy to lift.

[@LouisGedo]

@@Not4Talent_AI Thank you! I've not used SD for months now since paying for an annual subscription to MJ. I did that because at that time, MJ images were so superior in quality in general to what I could achieve with SD, that it didn't make sense for me to divide my time and energy between the 2 platforms. But lately I'm seeing some remarkable progress with certain SD forks and add ins that I'm getting tempted to dedicate some time to seeing how much SD has improved over the past 7 months or so.

[@Not4Talent_AI]

@@lefourbe5596 Thanks!!! It is a hard spot to handle in terms of work, but I'm here to stay! Hope to retain the quality as much as I can, or even, increase it over time. Ty for the support and understanding

[@Not4Talent_AI]

@@LouisGedo Yes, SD improves sooo fast xD it's also super good for control. You have the best of both worlds though hahaha

[@paperbea]

I'm going to have to beat the BREAK thing into my mind because I can almost never get the things I want the colours I want...

[@Not4Talent_AI]

It works pretty nice for simple things. It will still be really hard to get the colors right in more complex scenes though. Like I show with the sofa example

[@losing_interest_in_everything]

Can you show us how to create highly detailed 4K images? I'm having a hard time doing that despite the fact that I own a powerful PC 😭😭😭

[@Not4Talent_AI]

I have a tutorial on inpainting, a commenter left pretty insightful information. that comment is pinned in the video. Also there is some info in that vid's description, and ofc the video itself. In any case, I'll eventually do a video on how to upscale, but I also need to get the hang of it hahahah

[@losing_interest_in_everything]

@@Not4Talent_AI I'll check this video. If you can show us how to upscale it could be nice. Thank for the reply

[@fernando749845]

Is your beard glued on hair by hair?

[@fernando749845]

Excellent video by the way!

[@Not4Talent_AI]

hahaahhahah what? XDDDDD Don't think so, no hahhaah

[@Not4Talent_AI]

@@fernando749845 thank you so much! Glad you liked it! I'll make a beard review next xD

[@Framehacker]

For my understanding BREAK simply means that things shouldnt be mixed together and AND means that it should. That would explain your generation with ice cream AND a red dress.

[@Not4Talent_AI]

from a technical standpoint it doesnt do that. But from a practical that is a great explanation xD Ty!

[@WiseOwl_1408]

Spaz content

[@Not4Talent_AI]

ngl, I dont even know wtf that means XD

[@jimhowkings1974]

BREAK works for me . also it'd be amazing if the images were less nsfw :)

[@Not4Talent_AI]

Wasn't aware it was nsfw. I'll keep it in mind!

[@jimhowkings1974]

@@Not4Talent_AI they weren't 100% but safer better I also have a suggestion Creating male characters with many models is hard Can you do a research on the difference between models when it comes to creating male characters

[@Not4Talent_AI]

@@jimhowkings1974 Sure! Even though it is mainly based on training images. There are way more female images than male, so it makes it hard to create men. Even more so with anime models, as otakus are horny af I'll try to make generations of male more often (even though I will have a tendency to create women still xD) Btw there are models trained on man to counterbalance that disparity. Don't know any specific name atm, but they exist

[@lefourbe5596]

Can't really agree on the suggestion with all respect due. I know it's a good intention but... The there is many styles and diverse examples here along with the slightly lewd art. Really high quality work already. There is MANY Trial and error in the making. Generating NSFW (even if it's not) in the making make the process just both more enjoyable and faster. don't take that away from him. Also most of the SD community is used to it, i would not suggest someone starting to go "SFW". I don't want him to be held back like this. Well if you plan to share that in middle school art class or kids... I Can see why it's problematic indeed. i would like to make video too. However i struggle with my energy and my accent and i suck at editing. i made only 2 video for private discord and it was Time consuming and tiring. And they wasn't even good (but provide good help). Now imagine i get a someone just suggesting me to spend more effort and time to combat the output that most of us usually want to see and/or made us start ? That would hurt my motivation. Even if like him i Ask for feedback. No harm done but i though i have to defend my new favorite SD youtuber. Kind regards ;)

[@Not4Talent_AI]

​@@lefourbe5596 hahahahah No problem! I take all feedback equally, at the end of the day, it is feedback not a forcing statement. I didn't think it was nsfw cuz, as you say, most of us are used to it. I consider it good feedback to have in mind, that not everyone is as used to this stuff as the SD community that goes on civitAI xD Appreciate the feeling! tyty

[@dsl3226]

great pace to your tuts and decent content, i rarely need to speed it up compared to other channels but for the love of god please stop with the pointless spergy text editing and fullscreen webcam, i literally have to scroll away from the video constantly to be able to chug thru without building up a strong urge to punch the screen lmao

[@Not4Talent_AI]

😂😂 ty! I'm toning down the text on recent videos. Mainly on parts that don't really need it for anything. Just on the intro and to point out important stuff. About the camera, I will keep using a full camera mode for parts where I'm not specifically talking about the tutorial itself. Even though I'm trying to buy a camera that lets me put it further away so it isnt a full face screamer five nights at freddy's style. Atm I cant do much about it tho, so it is what it is. Trying to improve it, but cant atm

[@jeremiasgiglio7295]

I NEED TO SUPPORT THIS, give me your Patreon please I need more for my team.

[@Not4Talent_AI]

hahahahahha So glad you liked it!!! Don't really hava patreon though (even though it is in my banner xD) Dont have enough time to create extra content. Srry for that, but thank you so much!

[@mrhoangpro]

I usually rewatch the video because It is too fast 😢

[@Not4Talent_AI]

Yeah, I still need practice on how to pace my videos, srry!

[@mrhoangpro]

I don't know why IF I dont turn on regional prompter, the BREAK doesn't work. It still gives random results. do you have telegram group or discord to discuss

[@Not4Talent_AI]

I have a discord myself, not a discord group yet. but you can ask me: Not4Talent#9861 Also, break is not a magic solution. As I say in the video, it just helps a little in specific situations. but it still can fk up complex prompts most of the time

[@mrhoangpro]

@@Not4Talent_AI i did send discord request

[@Francisco-Danconia]

This is next level bro!

[@Not4Talent_AI]

thank you!!