Mastering Negative Prompts in Stable Diffusion

[@SeaHorseOo]

no christmass is my favorite. There's no bias that is an experiment. It imply there might be some bias that are hardcoded by intelligent design... 🙂

[@NorthTonawanda1]

I'd like to know what all those complicated sounding camera settings are all about without having to learn to be a professional photographer. A handy visual guide showing the effects of various aperture settings and different cameras, lighting techniques, vocabulary, etc. These are the things people really want to learn. I'm sure lots of people will start putting out these sorts of guides.

[@MakingPhoto]

Good idea, Andrew! I have a cheat sheet with the camera terms that MJ recognizes. But you've given me a few ideas.... Here's the camera style sheet. bit.ly/photography-mj

[@irenarupicpietri504]

@@MakingPhoto Thanks for the sheet.

[@stunspot]

Two things: 1) there's a couple of prompt generator sites with good pictures for all their keywords. 2) ASK CHATGTP. Seriously, I do this ALL. the. Time. "I'm making an ai generated picture that's a photograph of Cybernetic Cinderella fleeing the disco leaving her mirrored manolo blanacs behind. What would be the appropriate camera and settings to capture that?" ... "Uh huh. How about film stock? Focal length?". Even it doesn't always get it right, it's a great way to really quickly learn the basics. Like, I know my lenses now. Yay!

[@photoshopeandotv]

Thanks again for your videos! 😎

[@MakingPhoto]

Thank you, Photoshopeando! Keep creating!

[@richardglady3009]

Thank you for the overview and specific negative prompts. Thanks for explaining how to create these prompts. Thanks for explaining “remix.” In the end, as my brother (the photographer) pointed out…AI is not the same as taking a photo. Using Tensor Art, I created a perfect character on the first attempt…just 8 toed. Now I know how to (hopefully) get rid of spare toes. Thanks.

[@MakingPhoto]

No problem! Negative prompting is a bit hit-or-miss. TensorArt doesn't have inpainting, but that would also be a tool to help you fix a small problem.

[@irenarupicpietri504]

Thank you so much for this beautiful video explaining lot of things i already made wrong. I can finaly make it better and spare bunch of tokens now.

[@MakingPhoto]

Glad it was helpful! Keep creating, Irena!

[@cynthiairion5203]

How about the weight of the negative prompts? I never know how much is ideal.

[@IdentityCrisis1581]

Every time i use negative prompts, it just thinks its a positive prompt and doubles down on making more of what i tell it not to do. And i have used them on multiple apps and websites. They all don't seem to know the difference. An example is that if I want long hair. And telling it long hair doesn't work, giving it a short hair negative prompt. All it does is short hair and completely ignores the long hair positive prompt. And it seems like the more negative prompt words you use. The more confused it seems to get and doesn't know what you do and don't want. So it uses it all as a positive prompt and applies the details randomly as if it's cherry picking the prompt. I also noticed that all these programs want to default to a particular detail and it's almost impossible to make it stop. An example is that it seems to want to make all women as some perceived perfect 25-30 age range with perfect skin and a super model body. If you are trying to make a more mature woman who has an average figure and looks like any natural woman over 50 that you might meet on the street. It has difficulty with that. It always wants to default to that perfect young supermodel and just ignores your prompts both positive and negative. Another issue is that it will try to default to Asian women or anime girls even when that is miles away from what your prompt describes. I have had them make whole batches of young cartoon anime girls despite specifying mature female and realistic style and Caucasian, or African American. There is definitely some bias in how these LLMs are being trained. I understand that most people only want to make anime, and have an Asian woman fetish. So it's trained to produce the most common results. But that makes it almost impossible to get results different from that. Most of them don't even have loras of white women. They have a white man Lora. But no white woman Lora. And you can't use the white man Lora because then it won't make females, even if you specifically tell it to. And it's all because of the programers training the AI to be bias towards what they would want to make with it, and assume that everyone else wants the same thing. I was trying to make a realistic robot one time and it put some Asian girls face on it. I did not put anything in my prompt about women, females, anime, Asians or Asian women. But it just couldn't help shoving that into the image. And not a one time thing. It wanted to do it all the time, and using a negative prompt only made it double down and make an Asian girl instead of a robot. Now I don't have anything personal against Asians. But if that's not what I'm describing, then I don't want to see it. If I'm making a female ninja character. I expect her to be Asian. But when I'm describing a robot like you would see in an 80's sci-fi movie. I don't want to see a human face on it of any kind, Asian or otherwise. This bias towards everything Asian is out of control. The programmers don't seem to understand that not everyone is trying to make pictures of anime or young Asian girls. They can't accept that. It's almost as if they are offended by anything else so they force the AI to make that regardless of what's in your prompt. It's just plain annoying.

[@rasmadrak]

Try making a person holding a hat.... it's impossible. They're all wearing it on top of their head, just holding it simultaneously. :D

[@IdentityCrisis1581]

@@rasmadrak yeah, I noticed that it has trouble with things like that too. And you have to be careful describing what colors things are when you describe clothes. Like if you tell it that someone is wearing a green jacket and blue pants. It starts making random stuff green and blue, like it will give them green hair and blue lips and the jacket and pants won't be green or blue

[@Majd.19]

Hello, i hope you're doing well, I'm just wondering how could i make midjourney generat cyborg art that has featureless face, because when i include this word " featureless face " it not responds

[@MakingPhoto]

Interesting question! The AI has a difficult time with featureless faces. Faces have, by definition, certain features. I tried a few different prompts, but all had features.

[@MakingPhoto]

See my video on the Midjourney v5 prompt parameters. I made your featureless cyborg!

[@shockadelic]

[twins] helps reduce duplicate faces

[@MakingPhoto]

Great tip!

[@3RDEYELOVE]

do Negative Prompts work in MJ?

[@MakingPhoto]

Yes, I've done a separate video on this. Type --no after the prompt and list the elements you don't want to see.