📝｜prompting-help

[thorn crown]

regarding the main question: while its probably possible with img2img/controlnet color tools, you may find it quicker and easier to learn to use regional prompting. either with a plugin or one of the fancier controlnet things i haven't tried yet. check out Tiled Diffusion regional prompting if you're interested in that path, it's pretty nice.

[tall haven]

here ya go

[fiery pecan]

Thx, i try it once i get off work.

[pine stratus]

Hey guys, I've noticed that I am able to get consistency in multiple images when there is only one character (the same character) present in all of them.

The first image has the prompt "A boy wearing a red shirt and blue jeans walking down the street".

The second image has the prompt "A boy wearing a red shirt and blue jeans eating a pizza outside a pizzeria".

The third image has the prompt "A boy wearing a red shirt and blue jeans watching a movie on his iPad while sitting on the couch at his home".

The fourth image has the prompt "A boy wearing a red shirt and blue jeans playing basketball outside".

It is evident that the same boy is in all of the first four images.

However, if I introduce a new character, consistency gets thrown out the window.

The fifth image has the prompt "A boy wearing a red shirt and blue jeans watching TV on the couch with his sister who is wearing a light blue dress".

Note that the boy is wearing different clothes this time. It seems like his sister his wearing his clothes and he is low-key wearing his sister's clothes.

The sixth image has the prompt "A boy wearing a red shirt and blue jeans eating a cake at the dining table with his sister who is wearing a light blue dress". You can't even see the boy's sister in this image. Maybe he is sitting on top of her since you can see some long hair behind him.

All six images have the same seed*.

All things considered, characters are only consistent when they are the only character in the image. If I introduce a new character (a family or friend for instance), stable diffusion gets confused on who is who and gets characters mixed up while generating images. How can I fix this and keep consistency among characters across multiple images even when new characters are introduced?

[main oasis]

anyone know a way using wildcards extension or dynamic prompting to include the same random word from a list multiple times?

for example, if I had a list of colors and used wildcards extension to choose from the list twice:
__color__ apples in a __color__ basket
I will get results for any combination of apple colors and basket colors, but I want only results where the color of the apples and the color of the basket matches

[ivory galleon]

Guys, im having trouble creating humanized versions of inanimate objects. Does anyone have any good strong prompts to ensure some stuff like full body portraits with proper anatomy or half body portraits with a non distorted face? Maybe some tips on prompting?

[vital remnant]

Does anyone find that cutting and copying a prompt seems to affect the outcome differently than writing it out yourself?

As an example, these two images are using the exact same prompts/weights/settings/seed as each other, but the left image is when I re-wrote the prompts manually, and the right is when I just cut and copied them

Almost like the lora, vae, or checkpoint isn't properly applying to a prompt added as one paste

The outcome is reproducible, at least for me anyway

Also I now hate hands

[fiery pecan]

What website are you using if you dont mine me asking?

you can try to use controlnet. I am guessing for this situation HED might be the best to create the outline for the ai to refer to.

[azure otter]

hey guys, I'm new to this whole stable diffusion thing, I've ready and watched multiple videos, websites and I cant seem to get much more than this from my first renders in 512... when I move to imgtoimg and upscale things dont seem to get much better, can anyone suggest a fix/jump in a call and give some feedback

[silver valley]

Hey, it seems like you used a lora with a to high strenght

Lower the lora strenght to 0.5 and try again

[azure otter]

thank you will give it ago

[cinder sky]

If you got the lora from Civitai, taking a look at the desc may give you the good number for the strength

Usually should be between 0.5 to 1.3 for extreme

[fiery pecan]

Like other have said. You need to reduce the lora strength. I open up your png for metadata.

[fiery pecan]

I can jump into a call and chat. if you have question about AI image in general.

[azure otter]

I've jumped in Diffusers 3 if you don't mind

[fiery pecan]

give me like 5 or 10 min

[azure otter]

no worries at all

[marsh vine]

so ik pinned is a list of negative prompts you'd probably want, but is there any list for positive prompts? I just need some ideas because I don't know how to stretch a few words into 30 prompts like I've seen some do

[coarse herald]

Is it possible to do [concept1:concept2:0.5] in comfyui ? Like u can in A1111

[hollow crane]

https://namemc.com/skin/e7b5d88a838c7e52 I want this to anime

[fluid wing]

Im trying to create a similar style for vector art but I struggle and it makes it too realistic, what steps should I do to get there

[merry crane]

Does anyone know how can I achieve results like this?

[urban tangle]

It's the studio ghibli art style. There are a number of LORAs available if you're using a UI, like this one - https://civitai.com/models/54233/ghiblibackground

Otherwise prompting for studio ghibli style or environment. "Cel shading" may help.

[clever yew]

Hey does anyone know how to make a face looking forward to look in its side?

[thorn crown]

if your model is trained on danbooru style tags, you can try "looking away" and "profile" in positive prompt, and "looking at viewer" in the negative one. for models that use the traditional prompting, just "profile" in positive should be enough. (it should be enough for both, really, just mentioning other options)

[clever yew]

Ooh I will try that

[atomic flume]

Any tips on how I can make this a higher resoulion ?

I'm trying to go for more details

[fiery pecan]

Bruh, the image is already 4400x8800px big. How high you want to go? 😆

[atomic flume]

but when I look at it 100% zoomed in it looks a little out of focus

[fiery pecan]

@atomic flume What model you use and how do you upscale it to this size from the original image size?

[atomic flume]

I first used retile in image 2 image control net + ultimate SD upscale 2x upscale

then I dropped it into extra

[fiery pecan]

I think you may need to upscale it again using ultimate sd upscale once you did it the first time. Also look at other upscaler. You can find comparison online. Some maybe more suited to make image more clearer. It gonna take quite awhile. But i think it may work.

[pastel juniper]

Hello 🙂 I’m looking for help with prompts, models, or Lora to make studio ghibli inspired scenes. For some reason when I type out what I want it to do, it’s only doing like half of the prompt. Or it tries to do it all but it looks unfinished and broken.

Examples of what I mean.

I was able to get some good ones. I’ll post them in a sec. But I don’t have that clean line or smooth finish.

[candid dome]

Any tips on making vector art style with SDXL? Stuff like silhouettes, contrasting colors etc I remember but there are vector art keywords I've forgotten

[tender remnant]

@marsh vine

borbolrta

[minor hatch]

Hey guys n gals, Im wondering if there is a positive or negative prompt that could help me:
I'm trying to generate a QR code blended into an image, but the subject (a cookie) is just small and centered or fragmented, i don't understand how can i make it for example take the whole QR code, if I write big it's not that big

current:
humongous chocolate cookie, chocolate chips,broken in half, liquid chocolate cream, 8k, ultrarealistic, detailed, crispy cookie texture

negative:
ugly, disfigured, low quality, blurry, nsfw, small

[marsh vine]

thanks

[tender remnant]

/imagine,butterfly

[iron loom]

is there something like a "random" pose lora or prompt ? just so it gives me some ideas what pose a character i want to make without constantly teaking till i get 1 pose and then repeat the teaking for next few batches and so on

[normal dome]

how do i generate landscape or 16:9 photos without uncut in web app ?

[north saffron]

I'm new to prompting and trying to generate an abstract art work. I'm using a prompt like: "abstract artwork with pattern and geometrical shapes, flat, 4d, grayscale"

The generated images look fine but I'd like 3 things:

1. I would like to avoid shadow. (already put shadow in the negative prompt but it still generates shadow).
2. Sometimes it generates an art with the room surrounding it. I already put "room, sofa" in the negative prompt, but it still generates the room sometimes.
3. I'd like to have infinite pattern instead of having a centered piece in the image.

How would you improve the prompt?

[north saffron]

Some more examples where the surrounding room is created. How do I remove them?

[marsh vine]

put () around the negative prompt?

stack () until it works

oh also dont use grammer

minimalist, abstract, artwork, simple, pattern, geometrical shapes, paste, beige, lightweight.....

((shadow)), (((room))), sofa

[kindred garnet]

is there any style for this images? there is a huge fake wave on facebook, just trying to replicate it. im interested how they did that

its called watermelon art there

and many people just think its real xD

but im not able to replicate such an image

[clever yew]

BINGO!!!

[clever yew]

Thank you!! It worked!

[thorn crown]

any time

[brazen valley]

how does sdxl styles work? its just modifying the prompt?

[sour light]

Hello everyone, I rencently found this image, do you think it's an AI-generated picture ?

If so, do you have any ideas on how to reproduce this style? It's got a sort of old video game cover vibe.

I wasn't able to find anything about the creator / model used

[sour light]

yeah there's an album but nothing more 😦

[obtuse torrent]

some D&D checkpoint + vintage 70 video game cover + prompts ...
Use PNG info or other interrogating Clips@sour light

[sour light]

alright thanks for your help

[fallen smelt]

complete noob how do you change aspect ratio with button resize?

[quiet loom]

Troubleshooting: New to creating AI art. It seems like most faces and bodies are always distorted. Anyway to clean this up?

Link from breath of the wild, jumping off a building into an ocean filled with sharks, cartoon,
Negative prompt: multiple limbs, anatomically incorrect,
Steps: 20, Sampler: Euler a, CFG scale: 3, Seed: 3111928799, Size: 512x512, Model hash: 7dd0e6760f, Model: arcane-diffusion-v3, Version: v1.4.1

[rich crow]

try playing around with the cfg scale. also check the civitai of that model to see if it has any recommended prompts to include. i was going to suggest hi-res fix but it should be making good pictures at 512x512

[quiet loom]

thanks ill try it out

[silver valley]

The smaller the face on the image the more it gets disorted.
Close up Portraits give the best face Quality.
For large images you would need to Inpaint the face in img2img

You can try an other resolution like 512x768, higher cfg scale maybe 7-8
And 30 steps for example

[arctic tapir]

umm... can somebody help? what is that purple tint on the neck and collar?

[arctic tapir]

i was doing img2img

[silver valley]

You need a vae for that model

Its for color correction and fixes purple spots

[arctic tapir]

oh my god

you're a live saver

I did heard of VAE before jumping into SD but never thought would be this useful

[fiery pecan]

it will open a whole new world for you! have fun!

[arctic tapir]

i really want to make my own LORA out of my OC... is the process complex?

[lavish trench]

👋 Hello guys

I'm trying to make a map and mix 2 different styles, something like near-Futuristic center of the town and a Post-Apocalyptic suburbs.

but dang is hard, maybe in-painting could be a solution but still problematic.

The idea is to replace all of this area with something different that looks poor and degraded.

That's the prompt used 🤔
A video game perspective map from top to bottom, in the center a small near futuristic city with a few white buildings surrounded by lush parks with trees and fountains, with wide white sidewalks and narrow streets, at the edge of the image there are low buildings built with scraps and waste in a post-apocalyptic style, extremely hyperdetailed, perspective view top to bottom

[rich crow]

I think outpainting would be perfect for this. You would just change the prompt to generate the post apocalyptic buildings.

[lavish trench]

Indeed I see no other solution ¯_(ツ)_/¯

Maybe controlNet? 🤔

[rich crow]

It's the best solution imo, I see no reason to use other ones

[lavish trench]

Because I'm lazy AF 😅

[rich crow]

You could even get fancy with it and use https://github.com/ljleb/prompt-fusion-extension
For a few cycle of outpainting to sort of blend the destruction too

Nah outpainting is ez, if anything it enables you to be lazy

[lavish trench]

Wow this extension is very nice... intersting really! 😄

[rich crow]

Ya very cool

[lavish trench]

Inpaint do the job.. ~ mmmmmeh. 🥴

Oh... look what I've found!!
https://youtu.be/yDYhIuS8hJ4

[rich crow]

Just double checking, you're using inpainting and not outpainting?

Outpainting should be easiest

[fiery pecan]

what does \(style\) do? i saw some image uses it.

[arctic tapir]

⬆️, also, what is :d that people use?

[lavish trench]

:d the emoticon 😋

[lavish trench]

Now I just need to upscale it 😁

[arctic tapir]

😋

[arctic tapir]

bro, how do you upscale? img2img or extras

[lavish trench]

https://www.youtube.com/watch?v=A6dQPMy_tHY

[obtuse torrent]

because the parentheses are used as an emphasis multiplier in SD...
() = * 1.1
...
When you need to introduce a "real" () because it was used in the name of a tag or model... that notation is used... so that what is in parentheses is not highlighted and is considered as a single thing with them ..(Google T. sorry)

It is related to the old booru tags... it doesn't work the same in all models or Mix because there are also models that are not based much on the tag system

[arctic tapir]

hm? so lets say i wanna make aqua from konosuba, should i do aqua \(konosuba\) ?

if i dont use LORA

[obtuse torrent]

IDK, it probably doesn't have a relevant effect...

The example you give... will have an effect if the model knows "aqua" and/or "konosuba"... from before... and it won't have it if it doesn't know them... that will happen regardless of how you write konosuba

that way of including a parenthesis as a sign with no special meaning... only makes sense for disambiguations done in Booru system, like for example:

ink (object)
ink (medium)

or if someone trains some model, and uses () when creating the Trigger Words, example:
character_name_(outfit_01)
character_name_(outfit_02)

anyway "aqua" case is special because apart from the possibility that there are other characters called "aqua" in other series,
"aqua" is commonly considered a range of colors (shades of blue to shades of green)

[crude cradle]

Sorry, I'm new, so I input a checkpoint, but it didn't seem to do anything. Have I misunderstood what checkpoint/models are supposed to do? I was going for this:
https://prompthero.com/ai-models/avogado6-diffusion-download

Is there a solid tutorial on how exactly I can guide my model/train it to get closer to that style presented? Super impressed by what people have created on here btw!

[arctic tapir]

will this work? the amount of prompt? iirc, 77 is the max

[whole elbow]

is it possible to include the seed in a dream command?

[arctic tapir]

dream command?

[whole elbow]

yeah, you know, when you tell the bot to dream

[arctic tapir]

hmm, you mean where you put the prompt or text, yes?

[whole elbow]

yeah

slash dream, in the bot channels 🙂

[arctic tapir]

ahh, i thought you were referring to the SD webui, idk abt that, sry!

[whole elbow]

@delicate kernel created something I'd like to develop a bit 🙂

no worries, thank you 🙂

[silver valley]

Thats not a checkpoint. Thats a Lora. Its used together with a checkpoint.

[silver valley]

77 isnt the max. In the webui its nearly unlimited

But dont use to much

[low pendant]

how do I set the command "/style"?
and also "negative words"?
I only manage to write the prompt, and I don't understand how to type the other commands

[arctic tapir]

so what with people claiming about the 77 max?

[silver valley]

Its only for the default SD. And maybe some other webuis

[crude cradle]

@silver valley Gotcha! Are there good places to learn all these general concepts for stable diffusion? Went down that rabbit hole and now understand Lora's a bit more from https://aituts.com/stable-diffusion-lora/

So do I manually create the Lora folder in my models folder in stable diffusion? That was the only thing I was confused by

Also, why doesn't everyone basically just use inpainting/outpainting + lora's? Or are most people?

[silver valley]

The lora folder should be in models. You dont have to create it

[crude cradle]

Ahhh found it you right

What checkpoint should I use for lora models?

[arctic tapir]

anything, if you're not sure, try AnyLORA checkpoint

[silver valley]

Loras can work with any models. Some better than others

[crude cradle]

Whoa! Finally created something decent! Thanks a ton!

Fascinating!

it's a generic anime girl, but def going to try stuff later this week when not tired, super helpful/so glad I ended somewhere with at least a result lol ^

(I have no coding experience etc. so this entire process has been just super cool to learn as a whole)

[arctic tapir]

Same here!, i think i just advance my technique and found some way to make my image looks good

[modern bridge]

Hi, what's the prompt to have this hairstyle pls

[arctic tapir]

probably used a LORA

[modern bridge]

It's also probably midjourney so there is no lora

[rich crow]

Probably did use a lora but I wonder if we could still prompt for it in an attempt to hit it

Long bangs or hair to the side mabey, similar prompts like that

[grave dew]

Hi guys, I have 2 questions concerning prompts :

• Does the danbooru tags work with every chekpoints ? Or its better to only anime ones ?
• I see people writing multiple words using spaces and others with underscore, their is a better one ? (ex : brown eyes, brown_eyes)

Thanks

[quiet zodiac]

I'm trying to make a comedy/tragedy mask across a variety of faces. I can't seem to get the second emotion to convey to the other side, however. Does anyone have any tips on getting the frown to the second side of the face?

a split personality revealing a smile and a frown across one 36 y.o. face

[hardy dew]

try using the cyclical sampler __@color__ apples in a __@color__ basket - https://github.com/adieyal/sd-dynamic-prompts/blob/main/docs/SYNTAX.md#cyclical-sampler

[obtuse torrent]

its better in anime ones ... or mix with a percentage of booru models in them high enough for understand them too (the tags)

Normally no, there are differences but in general they are not more exaggerated than those that are commonly produced by other random factors.
separators become important when:

1. The prompts are longer than 75 tokens... so the way they are distributed will affect their relative importance in the composition.
2. When the separator is an intrinsic part of the tag (IE the "on_one_knee" tag)
   (However, it is no longer so important because internally the AI ​​corrects misspellings, omitted letters, etc.)

[barren copper]

need help in generating stuff something like this... really new in SD so idk

[grave dew]

Ok thanks !

[knotty ginkgo]

try:
a colorful wave of smoke on a black background, a 3D render by Paul Feeley, behance, color field, black background, behance hd, quantum wavetracing
or
a colorful wave of colored smoke on a black background, digital art, inspired by Aleksander Kobzdej, shutterstock, red blue purple black fade, 4 k hd wallpaper illustration

[obtuse torrent]

the exact style must be supported by one of Jinx's Loras or _Lycos (Arcane version)...

[restive sparrow]

anyone would know how to get this hair style?

[obtuse torrent]

@restive sparrow
at which checkpoint model, do you need to use ?

[restive sparrow]

this

[obtuse torrent]

mmm.. add bangs to negative ...
() emphasizes ...
It's hard to make that hairstyle work on older anime models, but that's a mix.. it will depend on how photographic it is.. and a descriptive line perhaps? "straight hair, middle part, uncovered forehead" ..@restive sparrow

[quiet zodiac]

Some hair styles have exact names, it may be worth googling that. For instance, a quick search reveals "sleek and straight"

[obtuse torrent]

without a doubt you are right, however, what we are looking for is not the name of the hairstyle... but the words or phrase that that specific SD model recognizes and triggers the generation of that hairstyle... I have no way of verifying what works and what doesn't in Cetus-Mix...

but yes, "sleek and straight" is recognized too in photographic models...

[frail sentinel]

I'd like to do animation, where I can upload images with a prompt then use the output in EbSynth. What's the best way to get a faithful reproduction of a figure in an uploaded image (in terms the figure's pose)? Also what's the best place to generate images? Here/DreamStudio/somewhere else?

[oblique hound]

Hi guys. I made 2 forest pictures. I rly like colors of the second picture and environement of the first. Can i do something to use colors of the second picture and put it into the first? (Tried var seed but it doesnt work). Except inpainting

[fiery pecan]

Do you use any lighting or artist keyword on both of the images?

[oblique hound]

I used wildcart on the first (no artist tagging) and just prompts on the second with lora "forest" and more details. I didn't even set up (pink colors). I used prompts "sunray"

[fiery pecan]

You can add additional keyword of what you want from the second image to the first one then render in img2img. You experiment with the denoising strength

[oblique hound]

Is it in xyz plot ?

[fiery pecan]

Xyz plot? Why do you use that?

[oblique hound]

I didn't, I didn't use add keyword in total. Newbie in that option

But I got the way to try. Thanks !

[fiery pecan]

Basically you just describe in keyword what you want.

You want pink tree. Add "pink tree*. It is best you look up photography terms. That would help you alot on how to describe things.

Maybe "strong shadow", "sunray". And other stuff. Cant think much. At work atm lol

For these low light shot. Try "Epi noise offset" it help make low light render

@oblique hound dm me the render if you manage to make it work. I am curious how turnout!

[oblique hound]

Yeah I'll luk

[oblique hound]

i got this after denoising 0,83 and var seed denoising 0,74

but i guess this is easier to render than fully green forest

[oblique hound]

this is vibrant and i like this even more. I used base (2nd pic), -1 seed but 0,85 denoise + var seed of third pic on 0,88 den str. Prompt: detailed, masterpiece, 8k resolution, deep pink forest, fairycore, sunrays, pixies, lush trees, a river, lora:forest:1 lora:more_details:0.75, and neg: BadDream, FastNegativeV2, UnrealisticDream, green colors

[rigid radish]

Anyone else having decent results with the CutOff extension for color bleed/breaks? I try to avoid relying on too many extensions and would rather master a good prompt instead, but it's been too difficult to keep colors in check (yellow eyes makes other things yellow, etc), that CutOff has been critical. Unless any of you have some better tips for colors in prompts? 🙂
I've tried:

• adding colors to every token, like gray room, red dress, blonde hair, etc, with decent success, but it's tedious and limiting
• trying that BREAK prompt with abnormal success (the colored objects looked weird)
• colorizing in Photoshop and img2img, not a fan at all
• color grid in ControlNet but couldn't really figure it out
• and lastly CutOff, which is working well

[solemn root]

I'm getting some strong doubling on a prompt that I like. What's the best prompting tactics to get a single object or person? I'm help would be greatly appreciated 🙂

[silver valley]

Hey, dont go to high on native resolution.
Try solo as tag or for anime models
1girl, 1boy

[solemn root]

@silver valley I want a wide 16:9 and already tried 1man lol. Any other key words possibly

[wheat escarp]

YetAnotherWildcardCollection

This is a compilation of around a dozen wildcard collections pruned for any errors and duplicates and organized for ease of use. I will be updating the project gradually.

https://github.com/LulzRose/YetAnotherWildcardCollection/

Requirements:

SD-Dynamic-Prompts Extension - https://github.com/adieyal/sd-dynamic-prompts

How to use:

Download the wildcards or folders you would like and place them in your wildcards directory in /path/to/stable-diffusion-webui/extensions/sd-dynamic-prompts/wildcards.

[rich crow]

wont work as well if youre using a 1.5 model since theyre trained on 512x512. the model will start drawing an extra person after those first 512 pixels essentially

[solemn root]

Yes but, are there any tricky words to help reduce the doubling on stuff bigger than 512

[obtuse torrent]

Thanks for your effort, I hope that many people can get some benefit and progress thanks to it. You should contact an administrator to evaluate setting the reference in the appropriate channel.

[obtuse torrent]

tag "solo" and emphasize it with ()... if small replicas of your character are duplicated maybe (duplicate" and "jpeg artifacts" or "clone" in negative can be of help...
A technique that can sometimes help is to start with a detailed description of the landscape, and then describe the character.
tag "landscape/seascape/cityscape/etc" " might be helpful.. tag "scenary" is also useful but it has a habit of trying to exclude characters from the scene...
start (48 x 9):(48 x 16) --> 432:768 native + Hires.fix (upscale x 2 ..default) .. for a first evaluation
then you can continue scaling with other tools in Img2Img or Extras

[stuck kindle]

i want a picture of two bubbles with bubbles faces, bubbles hands, bubbles legs, holding hands on pure clouds but the AI didn't achieve to do the faces hand and legs how can i do better ? this is my prompt : two bubbles WITH bubbles faces, bubbles hands, bubbles legs, holding hands, pure clouds, 8k, high resolution, Ultra High quality, hyper realistic

[fiery pecan]

You can try use blending prompt for it. may or may not work. how it work is you input the first keyword then the second. It will render the first keyword first UP TO A POINT and continue to finish it with the second prompt. for example [hands|bubbles:0.2]. 0.2 is how far the first prompt which in this case is the hands to be rendered before continuing rendereing it using the second keyword.

mess around with the weight see which one work better,

also be sure to experiment with other model. they may work better.. or worse. you need to find that out yourself. if you can find lora related to the kind of visual you want to do would also be very helpful.

[stable valley]

hi there, sorry if this has been asked before. i want to use 2 text inversions in my prompt to make a merged face. what syntax should i use?
(xxxx:yyyy:1)?

[quiet zodiac]

@solemn root Try using one of the SD2.0 models on civitai. They are trained higher than 512x512. I've been able to natively go as high as 768x1024 without doubling. Prompting really isn't going to matter.

@stable valley I was asking about something simliar the other day. Trying to get comedy/tragdey expressed on one face.

[dire shuttle]

Hello, I am trying to make an image of a 6 years old girl but it always appear the result with a face of a 6 years old girl but with body of a teenager or also with big breasts. I already included in the positive and negative prompts sentences like those but still appears with body of a teenager or adult female. Could you please help me with this?. Thanks. Prompts used for example (1.- Positive: No breast, no breasts, not adult body, etc. 2.- Negative: Breasts, young body, not teenager body, not adult body, etc. )

[tired vigil]

How to use this pic to get the accurate picture??

[quiet zodiac]

@dire shuttle Negatives in the positive prompt don't work well. Don't tell SD what you don't want, tell it what you do want. Also be careful of double negatives in your negative prompt. doesn't "not adult body" in the negative prompt exclude children types? Drop the "not/no" in your negative prompt.
Try being explicit like...
image of young girl eating her breakfast while wearing a costume
If you're specifying "wonder woman costume" that could be a pitfall. The term woman might be artificially aging your concept.

[dire shuttle]

Thank you. I will do. Also this appears when I use cyborg insta onde woman in the prompt. This in previous ocasions

[quiet zodiac]

@tired vigil The first thing you' want to do is find the best image and then frankenstein the head on to a costume from that time period for a better quality image. Here's my PNG attempt at that and a .JPG derived by placing an arbitrary background behind him.
Here's a 1st draft result using this prompt and an image guide a 50% strength.
a male author from the 1890s with a bushy moustache
You might want to specify bald as you craft your prompt.

[tired vigil]

I will try

[stuck kindle]

Thanks, I didn’t know we could do that, I want to learn more about how to do prompts, is there documentation for stable diffusion’s bot ?

[fiery pecan]

Welcome, Though i dont have a set documentation on how to prompt. alot of it is through looking ta other people prompt and look at various reddit thread on prompting.

[stuck kindle]

Okay, maybe yt videos? Or just links please

[quiet zodiac]

@tired vigil Reverting to model Realistic Vision 1.4 seems to work well with this concept. I've added a slightly blurry outdoor background to the image reference portrait.
a bald 64 y.o. male author from with a bushy moustache
I don't think it's possible to get an exact resemblance. But you could generate something that was close and upsizeable.

[oblique hound]

Guys, how to make spherization in sd? like this

[warm mountain]

Good keyword might be "photography sphere" as that's the name of the object in that photo used for that effect

Lens sphere, photography sphere, glass ball.. try a few out

[oblique hound]

thanks, i will

[warm mountain]

And post the results in an appropriate channel if you want, I'd love to see

[oblique hound]

what channel btw?

[oblique hound]

doesnt work 😦

[warm mountain]

¯_(ツ)_/¯

Try interrogating CLIP on that image or similar ones to see what it thinks it is

[livid jetty]

Any suggestions for techniques to compose a scene with multiple subjects with their own descriptors? For example, I'm looking to generate an image of two engraved stones, with one depicting a tree and the other depicting a mountain.

I'm using sd-webui. I've tried:

• Using prompts that describe multiple subjects. But SD seems to "blend" the concepts no matter how I phrase it (often both stones end up with trees and mountains). SD/CLIP doesn't seem to have a strong sense of the prompt's grammatical structure, even when I'm explicit about left/right placement or use words like "beside" or "next to."
• Using the AND keyword (https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#composable-diffusion) - doesn't seem to make much difference, but maybe I'm not phrasing things right?
• Inpainting seems to work decently, e.g. generate two stones with one design, then use inpainting to replace one of the stones' designs with the other desired design. But that's fairly labor intensive (lol at myself calling any of this "labor intensive," these tools are all amazingly powerful)

[quiet zodiac]

@livid jetty Think of it as a multi-step visual effect. Render one stone with a tree, and then another with a mountain. Composite them together in a size that is the exact size of your final SD output. In this case, I chose 768x576. Using the composite image as a guide, submit your prompt.
two engraved stones, with one depicting a tree and the other depicting a mountain

If you ease up on the image guide strength, you can get more random results.

[livid jetty]

Let me see if I'm following you correctly. Your suggestion is to generate the two images separately, combine them into a single image outside of SD (like with Photoshop or something), then run the composite through img2img in SD to smooth out the background, fix lighting, etc.?

[quiet zodiac]

Yep.

[livid jetty]

Gotcha. That seems like a good approach. If I want the items to match (like if the stones were supposed to be a matched pair) then I can probably use inpainting to produce the second stone, maybe add some variation, then compose them

[quiet zodiac]

There may be other solutions as well. I struggle to mix dual concepts in pure prompting. Image guides are my first goto when I get stuck.

[livid jetty]

Yeah there are clearly lots of approaches to this! Thanks for your advice, it seems like it would produce reliable results.

[pastel juniper]

Can anyone recommend artists or styles to get that traditional 2d anime style of art. So far my stuff is looking too 3d lol. I tried studio ghibli and idk if it’s the model I have but I’m not getting anything close

[obtuse torrent]

@pastel juniperI'm really too old, that this phrase lost its meaning for me ..."traditional style" in anime..., the most obvious is to change to an 2D anime-oriented model or incorporate the help of Lora Lycos oriented to a specific style ...do a search in Civitai for it
Regarding prompts... specifically, you can add 3D, CGI, realistic, 2.5D to the negative prompt...
Make an appointment to a specific period of the animation or study:...IE '2D anime in 80s anime style,
referencing styles of color usage (don't put them together if you don't know what they do):
"anime coloring"
"cel shading"
"flat color"
"gradients",
"high contrast",
"limited palette",
in the Anime channel they can recommend models... and maybe they will help you with prompts specific to the styles they use https://discord.com/channels/1002292111942635562/1091193032273043516

[pastel juniper]

Thank you sooooo much I’m about to try these!

[feral kite]

Hello everyone,
I am with SD1.5 now and trying to get as realistic close up shots of people as possible with I2I. Unfortunately my prompts don´t work that great.
Is anyone here, who would be interested in collaborating with me to improve them?
This would be awesome, thanks.

[quiet zodiac]

Post your problematic prompt to get feedback.

[fiery pecan]

When you said you using SD 1.5. You meant you using the one that came with A1111?

[fiery pecan]

TBH, i dont really have a link that i can post cuz like i said. i pretty much look at other people prompt, look at reddit and occasionaly stumble on some nice prompt advice. Experimenting with prompt is how i expand my knowledge. I am sorry for unable to give a definative answer.

[stable valley]

try this https://civitai.com/models/65214/age-slider

[tight robin]

Can anyone make it realistic

[dire shuttle]

Thank you Eva, I will do

[runic oar]

Hey guys What are the Positive and negative "L" and "R" prompts that I've seen in some SDXL workflows? What's the best way to use them?

[mild lodge]

Any recommendations for having people lay on a mattress istead of inside a mattress?

[quiet zodiac]

Try sleeping, resting. Lay is what chickens do to eggs, lie is what people do in beds. Try lie, perhaps.

[ornate pier]

Hi there, I'm learning how to generate realistic looking photos of people, but for some reason every so often it starts to generate images with nudity in them.

I haven't generated anything like that intentionally so I 'm not sure why it happens, and even when adding NSFW, nudity etc. into the negative prompts and emphasising them using parentheses' it still does it albeit less frequently.

is there any way that I can prevent it from generating anything NSFW at all?

[obtuse torrent]

If people are dressed, briefly describe what they are wearing...
change to other models, there are some that are NSFW oriented, and it will always appear

[mental shadow]

I'm sorry if this has been asked already...

when using the SDXL bots, you can have a separate "Style".
Is there a special syntax for that when writing prompts? Or does the bot literally just add for example "photography style" to the prompt? or ...?

[obtuse torrent]

@ornate pier this is 100% joke, but also about 98.5% true...

[fiery pecan]

yeah like AntarEx said. Switch to different model. use model base on SD 2.1 to have lesser nsfw image unless the creator said it has been trained with nsfw images.

[obtuse torrent]

@merry crane you like the colors or the style

[merry crane]

The style. That anime cartoonish look

[obtuse torrent]

@merry crane
which do you think is more similar?

none? .. I was seeing the color ..

[merry crane]

Third row last 2 look similar.

[obtuse torrent]

Model: darkSushiMixMix_colorful, .. for colors : ((orange_theme, plain colors, limited palette, warm colors, 2D anime art, stylized in the 80s anime style))@merry crane

[merry crane]

https://civitai.com/models/24779 Is this the one?

[obtuse torrent]

Model: kakigori_V2, anf this one is new but whit the old plain color styleModel: kakigori_V2,

[obtuse torrent]

is same author .. other version whit 2.5D ... look for colorfull ver for plain colors ...

@merry crane

[merry crane]

Thankyou so much. Really Appreciate it

[obtuse torrent]

@merry crane ok i am erasing examples posts now ... HF ..bye

[merry crane]

Sure

[pine slate]

anyone could help me with prompting in nmkd gui, im new to the software

i can speak english and chinese

greatly appreciated

[fiery pecan]

I havent used NMKD in a long while but it should be pretty much the same as A1111. the formatting used in A1111 should work in NMKD.

[fiery pecan]

What are you looking to create?

[pine slate]

Beijing, (Four people: 1.5)+ all facing backwards.There is a food deliveryman, a construction worker with a helmet, a real estate agent in a suit, and a courier holding a package.

i tried to post the negative prompts in the tab

but the images are still shit

[fiery pecan]

what model are you using?

[pine slate]

v1.5

[fiery pecan]

the base one is ass. go to Civitai.com and download some model there. There is a shit ton you can choose form. To make it easy. just choose some of the popular one.

[pine slate]

1sec

[fiery pecan]

dont write it like a sentence. write it like you adding a tag or a keyword.

Beijing, (Four people:1.5), shot from the back. Looking away, backward facing, food deliveryman, Construction worker with helmet, real estate agent in a suit, courier holding a package

you can try something like above see if its any better.

[pine slate]

1s

is there a best seed

[pine slate]

a bit better

but still shit

[fiery pecan]

its the model. you need to change model also there isnt a best seed. I mean there probably is a seed that produce the exact thing that you want but that thing is randomised so dont think too much about it.

[fiery pecan]

still using SD 1.5?

[pine slate]

yeah

im downloading those good ones

@fiery pecan whats a VAE?

[fiery pecan]

It essentially make the image better with brighter colour, better face shape or more detail in the image and more.

[rich crow]

most models don't do well with generating pics of the back of people's head. without a lora, at best you can prompt for "back of head" and pray you get a good seed.

[stone igloo]

FAQ: Why are my images blurry?

In order to ensure a safe experience, the DreamStudio website has a NSFW classifier that will detect and blur any potential NSFW images. While in most cases the classifier will appropriately identify NSFW images, there may be occasional false positives due to the nature of how these systems work. We will continue to work on and improve the classifier to make false positives less and less likely! You are not charged for any images that are blurred

[signal cradle]

I'm trying to get huge, disneylike, squarish eyes, but realistic, and in a realistic model.

however, no matter what I do, the model gives me small eyes

[pine stratus]

I made a reddit post describing what I need help with. https://www.reddit.com/r/StableDiffusion/comments/15aoymp/generating_more_than_one_character_consistently/ TL;DR - how to get consistent characters across multiple images with regional prompter. How should I craft prompts? Is regional prompter even the way to go or is there a better way?

Feel free to reply here or in the reddit post.

Also not sure if this belongs here or #🤝｜tech-support if you want me to move this let me know and i will

[fiery pecan]

Looks like you need to swap to another model that has those feature or use a lora. Dont ask me which model/lora to use. You need to browse civitai for it.

[wary stone]

So I dont know how to make the eyes look good. Everytime their blurried or ugly and i dont know what to do. I tried to use loras and inpaint but to no avail

[tribal hawk]

any tips on making clean faces? sometimes I get weird colors on the face like

[wary stone]

try restore faces or Hires, fix or another model/lora

[tribal hawk]

I was guessing it's because the model because other models don't have this issue as often as this one, thx

[silver valley]

Hires fix or Inpaint

Or adetailer extension which is Inpainting at last step

[wary stone]

Hires Fix does work sometimes but the results are sometime meh, very meh

[wary stone]

I will try that

[silver valley]

Can you Provide an example with the Hires fix settings you use?

Mostly its just a setting that can change a lot

[silver valley]

Okay, try 15 hires steps and for esrgan based upscalers your denois should be below 0.5

[wary stone]

OK

[silver valley]

Also upscale by 1.2 can be to less to get it highres

[wary stone]

Uh yes, but if i upscale to much it shows the out of RAM error message

[silver valley]

Yea thats why you need xformers

Then you can upscale by 2

[wary stone]

Ah, ok. Then i will add them

[rough raptor]

Anyone know how to achieve a girl with two different pets? I'm promoting "Beautiful redhead wearing glasses in a jungle beside a black dog and a black cat"

Prompting*

[rustic citrus]

Hello, I'm usibg comfyUI and I wonder How should I connect prompts to the CLIPTextEncodeSDXL . Following Bot style, I have this "formula" : "Style" {My prompt} , "Style spec". Positive: cinematic photo {My prompt} . 35mm photograph, film, bokeh, professional, 4k, highly detailed
Negative: drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly, as an example. My question is, wich part should I connect in text_g and text_l ? Text_g should only be "my prompt" and text_l Style+ Style spec ? Same for negative prompt, should it be my negative prompt_l "my negative prompt " and negative prompt_g the style negative prompt ? It seem's to be easier in the refiner as the CLIPTextEncodeSDXLRefiner has only one positive and negative, I connected in each the sum of positive or negative . Am I right ?

[wary stone]

Thx, works perfectly and the results are top notch

[silver valley]

Perfect, np

[wary stone]

btw, is it noral that the consloe says progress 100% and the UI say 97%?

[silver valley]

It should say that only a few seconds before its done

[wary stone]

Well... The console spits out an error

[silver valley]

Can you show it in #🤝｜tech-support ?

[thorn oxide]

Hello, I have some issue while prompting stable diffusion, I would like to decrease the contrast of my image, so I try to add "low contrast" in my prompt, but it doesn't have any notable effect. And when I try like this (low contrast:1.5) It seems doing the opposite effect. Is there an other way to make understand SD that I want an image with lower contrast?

[steep jungle]

I am having some trouble generating a prompt that would use art style (eg. impressionism) and generate some futuristic content. So basicly mixing old style with new content. It seems that style drives content and thats quite understandable given how diffusions work. Negative prompting can help with this some, but do anyone of you know some advanced tricks to enforce the style without affecting the content that much?

[soft sand]

How do I get cute stubby hands/feet instead of lame ass fingers lol

[quiet zodiac]

@rough raptor SD can't do dual concepts yet. You can image guide it, however. For instance take a picture of a friends standing in a jungle with a dog at her side while holding a cat. Or frankenstien/photoshop something close. Then submit that image as a guide, and SD wil "improvise" a new version of that image, based on the submitted prompt.

@steep jungle I have had some good luck with that, but noticed some models don't honor artistic styles very well. Here's a Rembrandt in a space station setting, using the civitai model Deliberate_v2

I start each prompt like this:
in the style of Rembrandt, ...

[tawny plover]

hey guys
i wonder that
how to creat stirrup legwear steady
it always just creat pantyhose

[quiet zodiac]

@thorn oxide Try adding "well lit" to your prompt.

[tawny plover]

[quiet zodiac]

Do you mean beeches and jodhpurs , or are you talking about chaps?

[tawny plover]

wait a sec

sock like this

toeless and heel is uncovered

it suggested that the prompt is stirrup legwear,but it comes out wrong

i'm really confused

this one is more precisely

[steep jungle]

Im not a big fan of using artists names, coz it drives the content so much.

[thorn oxide]

@quiet zodiac Thank you I'll try

[tawny plover]

help plz

[rough raptor]

Thank you @quiet zodiac

Can you give imagine inspiration on this Discord?

[quiet zodiac]

I just posted a similar dual concept reference a couple of days ago.
#📝｜prompting-help message

[soft sand]

How do I get hands like this? I've tried negative prompting fingers and toes and yet it still gives me fingers and toes

[obtuse torrent]

Although it's not the best, one way to do it is to create your image normally (visual composition), then send it to Img2Img, and regenerate it (changing the prompts to reference the specific art style) you can play around with the D.S. So that it changes a little or a lot from the initial generation, if you use Upscalers, try to make them consistent with the style because in some cases they completely erase it.

[steep jungle]

first, thanks for your response!

1. I use photoshop as upscler (sorry misclicked enter)
2. i2i is working in some cases, but often the image where you start need to be "clean" and it kinda translates easily to the the other image. I am not sure what you mean by playing around in D.S. could you please elaborate? I am using quite many generative ai systems atm, but i use the 0.6 the "thing that controls how much you change the pic" in automatic1111. (sry, cant check the name of the thing fast atm) .

@obtuse torrent misclicked enter, thanks for your response!

[obtuse torrent]

just start using "white_thighhighs" + "toeless_legwear ... and then add the following if needed. with emphasis until skin is visible, bare_toes, bare_heels ((all in positive)

[astral rampart]

hey guys, new here... and I need help already... trying to generate "bottom part" of the hamburger bun... and it seems impossible... i've tried literally gazillion prompts, even wrote everything what goes into a burger into negative hoping i'll get just the bun.. but no luck 😄

basically this... (stock photo) ... even in img2img no luck to get it done, no way....

[obtuse torrent]

Yes Denoising strength =DS .. "thing that controls how much you change the pic" ...If you make a "futuristic" pic .... with which you are satisfied ... but you want to see how other styles would look on it, you use Img2Img (you take it as an example or guide for the generation) ... now you need get away from it enough to change the style... but not so far as to ruin your original Pic... that's where you can "play" by adjusting the DS..

[hollow epoch]

Hi guys, I just started using Stable Diffusion week ago. Now I'm trying to make a background that fits the products, but I've met the same problem too. I'm totally rookie on whatever coding or model training stuff. Here are the setting I used. Any expert here mind sharing your thought, experience or solutions?

Here are the image and mask.

[jovial cradle]

how to use image2image in discord?

[silver valley]

Not supported here

[finite karma]

Is there any way that I can use XYZ plotting to test different artstyle Loras over one image?

Like testing different checkpoints results of one image but with artstyle Loras instead

[astral rampart]

i believe that outpainting is what you need

[astral rampart]

for example... this is just pure outpainting

[astral rampart]

with a better model, you get better images.

[sullen rover]

what model is that? @astral rampart

[digital jewel]

Im trying to create a tapestry, any ideas on what models would be best for this? This was made with SDXL which I love but I want to use controlnet so I need something compatbile with controlnet.

[tulip ingot]

do you guys to suggest any specific templates for generic positive/negative prompts? ( more specifically I'm diving further into SDXL )

[serene flower]

Does anyone know what the style setting in dreamstudio does? For e.g. when selecting realism. Is it a positive/negative prompt? And did they ever show what the prompt/negative prompt is when that style is selected?

I suppose this is somewhat related to pocket's question.

[fathom niche]

hello there, I just stared using stable diffusion and I want to make some awesome art and my quesiton is how can I make something like this image:

[alpine vector]

https://youtu.be/W7zvb5wiYAI

Boban, check this video. There you find your answer for this spiderman picture. Cheers.

[strange meteor]

AMA - I developed an LLM, what questions should I ask to test it out?

[barren spire]

How can I make adetailer make the same face twice? It's a 2 panel comic type of thing

[brittle dawn]

Hi there, can anyone help me making a prompt to get smth like these images? I was struggling so much to achieve this look, but all my attempts are a failure 😭

[charred root]

Hey friends, what's the best way to control the camera position relative to the subject?

[quiet zodiac]

You can't really control the camera that much. You can request shot styles, framings and lens types. So, think like a photographer.

lst_shot_framings = ["(ECU)","(CU)","(MFS)","(FS)","(MCU)","(MS)","(LS)","(ELS)"]
lst_shot_types = ["extreme long shot","long shot","medium shot","close-up shot","extreme close-up shot","glamour shot","reverse",
    "low angle shot", "high angle shot","Silhouette shot", "wide shot","overhead shot","side-view", "centered-shot", "over-the-shoulder",
    "back view","selfie", "first-person view", "view from airplane","aerial view","satellite-view","landscape shot", "panoramic shot",
    "tilt-shift","product shot","long exposure","ultra wide shot"
    ]
lst_fstop_type = ["f/2.8","f/4","f/5.6","f/11","f/16","f/22"]
lst_lens_type = ["12mm lens","14mm lens","24mm lens","35mm lens","50mm lens","85mm lens","100mm lens",
    "wide-angle lens","fish-eye lens","telephoto lens","macro lens","16–35mm Zoom lens","70–200mm telephoto lens",
    "macro",
    ]
lst_shot_modifiers = ["sharp focus", "vibrant high contrast","centered in frame", "facing camera","person in foreground"
        ]

But I generally start a prompt with "facing forward" or "facing camera" to make sure the subject looks at the camera.

I often end my prompt with "camera rule of thirds"

[charred root]

ok i c ty, what about deforum tho? it kinda rotates the camera

[quiet zodiac]

Deforum is a new term to me. From what I can quickly google, it looks like someone wrote an extended prompt token tool to take the result from one frame and submit it back as an image guide for the next frame. I don't believe the pipe symbol | is part of the SD prompt base. Kind of lets you dive into a picture.

[strong rose]

Can I generate stable diffusion Images here from discord?

[verbal granite]

I am currently fighting SD to the death just to make it generate this hairstyle. Does anyone know of a lora for this one?

[next flint]

r u using 1.5 based or sdxl

[verbal granite]

..did you make those terms up

I'm sorry, I'm very new lmao

[next flint]

are you using the new stable diffusion xl model or a different one?

if ur using xl it would say "sd_xl_base_1.0_safetensors" in the checkpoint dropdown

[hard phoenix]

anyone knows if you could fuse characters in stable diffusion? i have been trying to do that for quite sometime resulted in failures

[verbal granite]

it says "sd-v1-4.ckpt" in the checkpoint dropdown but I do have a couple of models that I also tried like Counterfiet and AnythingV5

[next flint]

okay then counterfiet and anythingv5 are 1.5 models and sd-v1-4 is a 1.4

ill send a lora in a sec that should be good for that hairstyle

[verbal granite]

Thank you so much!

[next flint]

https://civitai.com/models/76937?modelVersionId=93108 i think hightop fade would be what you are looking for

[verbal granite]

Thank you very much! I got kind of close a few times but it's a struggle

when you use trigger words for lora do you have to add lora:hi_top_fade_hairstyle or can it just be hi_top_fade_hairstyle?

[dull folio]

does anyone know what i should write if i wanted the ai to make a similar hair?

[next flint]

in the prompt include "lora:hi_top_fade_hairstyle:1,hi_top_fade_hairstyle"

at least i believe that should work

[pine moth]

Hey, are bracketed weights and :1.3 etc style weights still working with SDXL? i've been trying but don't think they are working, what about prompt timing, is this meant to function as like with old SD or is there a new format? thanks

[verbal granite]

Thank you for the help! I had to bail for a while but I'll play with it soon!

[quiet zodiac]

Try googling the style. It seems to be called "undercut", or "sidecut". If you supply your image as a guide, it can get you close.

facing forward, a woman with and undercut hair style wearing a tank top in the park

[narrow ether]

asking for suggestions for img2img to get rid of lines like these:

[steep ivy]

hard to say without knowing your settings

[narrow ether]

if you had this image, and the task, what would you put into the negative prompt? i'm looking for prompt ideas.

[chilly raven]

actually i try to create a picture with batman, who stands behind the dead body of joker ... problem is, the pictures always only shows batman xD

[narrow ether]

google prompt keyword boosting

[tidal spade]

I want to have SD give eyes like this
How would you recommend going about doing so?
(Theyre the "All seeing eyes of god from Blood Blockade Battlefront)

[loud geyser]

Anyone know how to do TEXT in SDXL? Example prompt, I want to make a pretend "password list" @loud geyser if you know

[viral peak]

im currently working on this image and i was wondering how would i go about removing this string of hair?

[narrow ether]

inpaint with a different seed or variation

[crude trench]

How do I use multiple Loras without creative an eldritch monstrosity?

[coral pewter]

is there a good way to get a plain white background

[plucky hamlet]

Hey

[obtuse torrent]

Add white background, transparent background....

[ashen goblet]

hey. I need a little bit of help with keywords for generating male anime characters. Especially adult males are extremely difficult to generate, always looking like young boys or girls.. I'm already having girl, woman and boy on negative prompts and male, man on positive, but it still generates young boys that honestly look like girls, is there anything else I can do?

[weary needle]

have you tried azovya's age slider?

[ashen goblet]

first time hearing of it, so no

[weary needle]

This is essential imo

https://civitai.com/models/65214/age-slider

Combine it with (## year old) in the prompt

[ashen goblet]

sounds as essential as the multi colour extension haha

[weary needle]

link? I've never heard of that lol

[ashen goblet]

https://github.com/hnmr293/sd-webui-cutoff

it doesn't always work for me, but it has made multi colour gens much much more consistent

[weary needle]

Ah, i guess it's to fix color contamination. seriously annoying issue, i hope it works

[ashen goblet]

most of the time it at least drastically improves the odds haha

[stiff nova]

Hi there, what's the easiest way to run a img2img batch (ex. 100 images) and as it goes to the next image to be able to adjust the denoising strength

So like start with 0.0 and finish at 1

[crude trench]

Do you have any other essentials? Just started to use SD

[weary needle]

most models are mixes of each other, so the output is nearly identical

[dawn hawk]

I'd increase denoise

[dawn hawk]

Looks really tough, I'd prolly write not about eyes at all but instead something like glowing wheel or something like that then later on invert it in paint

[dawn hawk]

Remove it in paint net, send it to inpaint

[dawn hawk]

Experiment with each LoRA to see which one is causing issues at full strength, then lower the strength of that LoRA

[dawn hawk]

simple background is also a keyword

[dawn hawk]

What model are you using? Also, I'd definitely add male in there. You can probably also add a beard then remove the beard later on with inpaint

[dawn hawk]

I mostly find success with lowering the strength of the color. If I still face color issues, I typically go to regional prompting. Tbf tho I never tried it out

I know ppl do this for more consistency, but you can also get consistency differently I found.

[dawn hawk]

You may be interested in the scheduler extension forA41

https://github.com/ArtVentureX/sd-webui-agent-scheduler

I'm not sure if it's doable with that tbh

It honestly sounds like

You are best off writing a Python script yourself

[stiff nova]

yeah I tried using the x/y/z plot and was hoping that it would automatically select the image and go to the next one on it's own but it just showed an error

[stiff nova]

I'll try this

[hollow epoch]

Thank you so much!

[astral rampart]

np, if you need any help you can also dm me, I'm also trying to learn things

[quiet zodiac]

@chilly raven You have discovered SD's Achilles heel. It can't handle two concepts very well, so revert back to traditional methods. Produce an image of the dead joker, then produce and image of the batman and place the cut out bat man on top.

[chilly raven]

ah thanks for the info. 🙂

[kindred minnow]

ahhh someone has stolen my name

[quiet zodiac]

We're twinsies!😄

[obtuse torrent]

⚛️ ⚛️65

[oblique hound]

Any ideas how to prompt a spear behind head? Tried to use controlnet and got only 1 success with 50 generations (and still got bad arm anatomy). P.s: I can't disable controlnet cause it can't draw this stance just by using prompts

I wanna draw a spear like this, but in original stance

I forgot a name of lora that adds weapons, like dual wielding, staff, bow , sword etc, maybe it can help?

[solar dust]

So

let's say I want to use a specific character's Lora for imgen but there are multiple options available

is there a way to know which one is the best without manually testing one by one?

[vale nacelle]

what I need to do to use Swarm to create alterations of the image?
I can't connect image from Load Image with KSampler
different nodes

[wheat stag]

Any tips for improving them textures of tree leaves/ background vegetation?

[wheat stag]

Otherwise I find the sdxl 1.0 yielding relatively stunning/ sublime results with little prompting. However, the aforementioned textures I find bad

[rich crow]

without **manually **testing one by one?
automatically test them one by one with the X/Y/Z script 😈

[sinful oar]

if I were trying it, I'd generate something like a softedge in controlnet, and then put it in gimp or photoshop and start hacking on it. then use the modified sketch to guide generation. I dont think you're gonna be able to wordsmith it. Time to see what your artistic skills are

[strange meteor]

how much gpu mem is needed for sdxl base 1.0?

[arctic tapir]

what is this error in img2img?

[harsh notch]

Hello,
Does Stable Diffusion XL offer any sort of grouping for the positive prompt?
I often have the issue, that elements are "bleeding" over when describing multiple elements.
For example, I want to generate an astronaut with a golden helmet and a black visor.
When I only have those two elements, I can get fairly good results (first image).
But as soon as I add "white space suit" to the prompt, I cannot get it to generate a golden helmet anymore.
The gold is also "bleeding" into the white space suit.

I know some other models offer curly braces {} to group prompts.
Is something similar available to Stable Diffusion XL?

[dawn hawk]

Probably don't use openpose but something like lineart or canny. Alternatively, you can draw a stick in the starting picture

[dawn hawk]

6.5 in ComfyUI, don't try in A41

[strange meteor]

How to deal with token length in SDXL which is now restricted to 77?

[strange meteor]

Getting this error while trying to run SDXL

[oblique hound]

I can easily do it with editing but I wanted to make 1tap-gereration 😁

[oblique hound]

Is it possible to make this stance without openpose? I don't think so 🤔

[ashen ridge]

how do I redo the generated image without retyping the prompt?

[quiet zodiac]

@oblique hound A classic use of image guides. Cut out the posed figure, add a spear behind her head and drop in a background. The key is not to use a lot of steps. This final uses only 9 samples steps. The more steps you use, the more likely the spear behind the head will become the spear in front of the head.
facing forward, beautiful asian chick in armor with spear behind her head, long brown hair

[oblique hound]

What did u use in total ?

Photoshop to cut a girl and then put where ? What kind of controlnet option?

I see that on the second picture right hand is better (especially clothes) maybe it's just a game with steps and denoising, but what comes before?

[deep saffron]

many ways to rome, i used inpaint to remove her sword, sent it to inpaint sketch, draw a spear, send it back to inpaint again on a low denoising strength to remove the blur around the spear. mask what you want to keep and inpaint the not masked area

1st inpaint: fill with background
sketck inpaint: a long thin spear behind her head
last inpaint: first i interrogated the clip and used it as base; a woman in a costume holding a long spear in her hands behind her head (masked her face and body)

for more detailed hands you can use inpaint again, or of course control net

[quiet zodiac]

@oblique hound You can do the steps in photoshop, but I use after effects for my masking. After cutting out the figure, I found a spear on line and did the same thing. With the pose sandwiched between two spear layers, I dropped in a background and added a touch of soft blur to help separate the figure from the BG. In my OP, the second image is the guide I used to derive the third image. I used no controlnet or inpainting, just prompting. I guess the photoshop/AE work could be considered "out painting". Here's another image with a different background. No denoising used.
POS:
facing forward, beautiful japanese chick wearing a revealing red dress holding a spear behind her head, long brown hair
NEG:

MODEL:SG161222/Realistic_Vision_V1.4, unipc
STEPS: 40, CFG: 6.5, Guidance: 0.4768, 768x576

[distant rivet]

can links be used in prompts .... like picture link on the internet or a directory pic link?????

[deep saffron]

never tried it, but i don't think so

[harsh notch]

Did I post this into the wrong topic?

[solar dust]

I think the 4xAnime6B upscaler is not working anymore. I'm recently getting very bad results when using it.

Is anyone else having the same problem?

[solar dust]

Yeah, some people told me that an auto update from automatic11 broke 4xAnime6B

so it's not just me.

[oblique hound]

@deep saffron @quiet zodiac thanks ! I'll try to handle that when I arrive home 😁

[prisma grail]

I'm trying to make the apple unblurred but it insists that it will be blured how much I even try, I put blur:1.3 in negatives, it's so annoying

[solar dust]

4xAnime6B scaler is no more

does anyone know a substitute?

[silver valley]

It should be still there

[solar dust]

I mean that it doesn't work anymore

[silver valley]

Oh why that

[solar dust]

an update from automatic11 broke it.

[silver valley]

Hmm i can test it later

[silver valley]

Fatal Anime 50000 or 4xAnimeSharp or Remacri are good upscalers

[solar dust]

Thank you.

[dawn hawk]

Forgot to hit enter
You can use commas to reduce the weights and also separate them, CLIP actually lightly understands that
There's also regional prompting / area conditioning that you can use

[solar dust]

Anime6B is giving really bad results

texture-wise.

[dawn hawk]

Have you tried "focus apple"

Also what other things are you doing

Like is it inpainting or is it upscaling

If it's inpainting what is your denoise

If your denoise is below 0.3 it can result in blurriness

Uhh

When upscaling

So when keeping the image it uh shouldn't result in this much change

It also lightly looks like there's no VAE? Or maybe you are using BlessedVAE idk

[solar dust]

@silver valley Couldn't find any of the 2 upscalers you mentioned in Civitai.

[dawn hawk]

https://upscale.wiki/wiki/Model_Database

[solar dust]

Thanks but that's a HUGE list.

I'm trying to ctrl F those 2 but still can't find them.

Found one of them, finally.

What are these folders btw?

Do I need them?

[prisma grail]

well I want overall focus on everything, I am doing inpainting and I think my vae didn't load after restart which is really annoying, I'm using kl-f8 anime2 as vae

[dawn hawk]

Sharp

Maybe detailed

And then after throw it into some slight img2img upscaling

Upscaling makes things sharper

[silver valley]

Thats maybe because your lora strenght is to high

[prisma grail]

my prompt is 1girl, animal_ears, apple, blue_eyes, blue_hair, blush, carrot, crossed_arms, hair_ornament, hair_ribbon, long_sleeves, looking_at_viewer, maid, maid_headdress, , red_apple, rem_\(re:zero\), ribbon, short_hair, solo, table, x_hair_ornament, (open squiggly mouth:1.3), (squinting:1.2), (dilated pupils:1.3), thought bubble, (white rabbit_ears:1.2), (white ears:1.2), (fake white rabbit ears:1.2), (hungry:1.3)

negative:
black rabbit ears, black fake ears, black ears, bad-hands-5, easynegative, pureerosface_v1, verybadimagenegative_v1.3, ng_deepnegative_v1_75t, (black borders:1.1), bad-image-v2-39000, (blur,blurry,focus, out of focus:1.7), worst quality, low quality

[dawn hawk]

Get rid of the negatives

Uhm

[prisma grail]

ok

[solar dust]

The strengh is 1.

[dawn hawk]

Get rid of the negative embeddings

I mean

Not everything

[prisma grail]

yea lol

[silver valley]

Try 0.6 or 0.8

[dawn hawk]

So bad-hands-5, easynegative, pureerosface, verybadimagenegative, ngdeepnegative, badimage-v2 whatever

[solar dust]

I only use 2 negatives

bad-hands-5 and EasyNegative

that's it.

[prisma grail]

hmm

[dawn hawk]

One of them might be trained on more "bokeh" pictures

[solar dust]

I only add more when I specifically do not want something to appear.

[silver valley]

You only need the .pth file. Then put it into models/esrgan

[dawn hawk]

I typically just use (bad quality, worst quality:1.3)

And then add things as needed

multi-panel, greyscale, monochrome,

And then whatever

Textural embeddings are not great negatives actually since you can't quite control them

Whenever I do use them I use them at 0.4 or 0.6 weight

0.6 also is roughly the weight for overtrained LoRAs for some reason

I'd also then add to the positive prompt "sharp"

Maybe detailed but unsure about it

And change the order

[prisma grail]

it's better not

now

[solar dust]

Change the order?

[prisma grail]

does order have any effect

[solar dust]

Does the order of prompts have any impact on the final product?

[prisma grail]

I have seen people put masterpiece on the start and the end but didn't think much of it

[solar dust]

My go-to at the beginning was (masterpiece, best quality, ultra detailed)

and then the prompts

nowadays I try to add more things

like (detailed skin texture), (8k), (HDR), (sharp focus), (extremely detailed), (intricate), (soft light), (dramatic light), (sharp), (HDR)

[dawn hawk]

rem_\(re:zero\), looking at (red:0.3) apple, table, hungry, sharp, fake (white:0.5) rabbit ears, (blush:0.5), maid, maid headdress, (squinting, dilated pupils:1.2), thought bubble, open squiggly_mouth, short hair, crossed arms, hair ribbon, (blue:0.4) hair, short hair, BREAK, focus apple, sharp, detailed,  ```

Something like that

[dawn hawk]

Yes, if you use A41

The words at the front are weighted more

Idk about other UIs

But in this case

We do not want masterpiece, best quality

because it's likely to cause bokeh

[solar dust]

and (photorealistic:1.5), (hyper realism) for when I want to make real world cosplays.

I use automatic11.

[dawn hawk]

And the other things typically don't really work, aside from extremely detailed, which should not be a boilerplate, and soft light, dramatic light,

[solar dust]

What's boilerplate?

[dawn hawk]

Things you just repeat every time because it's required even tho it's kinda fucking clear

[prisma grail]

can't I put bokeh in negative?

[dawn hawk]

Yeah you can

But

[prisma grail]

what's break for

[dawn hawk]

Don't make the prompt fight itself

It's so that it creates a new batch

If you are using A41

At the start of each batch

[prisma grail]

uhm I'm using stable diffusion webui I have no clue what a41 is

[dawn hawk]

The tokens there are weighted more

So

Everyone unfortunately

Calls their thing

Stable Diffusion Webui

[prisma grail]

yes

[dawn hawk]

So we instead go by the maintaner name usaully

There's automatic1111

ComfyUI

SDNext which used to be Vlad

InvokeAI

EasyDiffusion I think

And other things

Kohya

Whatever

[prisma grail]

yes well then automatic1111

[dawn hawk]

Automatic1111 = A41

[prisma grail]

I see

well in that csae

[dawn hawk]

Try my prompt

See what happens

[prisma grail]

is "BREAK" a keyword?

lol

[dawn hawk]

Yes

It makes it start a new batch for the tokens

[prisma grail]

oh

[dawn hawk]

The start of each batch is weighted more

It's whatever I should eventually get to that in my guide

But I'm learning so much shit about inpainting it's unreal

[solar dust]

Okay but does automatic11's prompt order matter?

[dawn hawk]

Yes

Put boilerplate at the end it's mostly useless anyway

Everyone used to have best quality, masterpiece at the start

But that's a bad idea

You want the most complex stuff at the start

So that it can affect the composition of the image more

[prisma grail]

thanks for the tip

[dawn hawk]

Yw

[prisma grail]

how do I add particle effects like shiver?

[quasi bone]

Hi, I would like to learn more about mashing various styles for generating images. For example Barbie movie style applied to Harry Potter scenes. How would I go about to improving the prompt?

Positive: Harry Potter as Ken in a Barbie World, flying on a broom around Hogwarts.
Supporting: Pink car, pink wand, movie style, cinematic, round glasses, Daniel Radcliffe, Ryan Gosling

Result:

[prisma grail]

what's supporting?

are you talking about negative prompt?

I would first figure out what you exactly want, for example get two referneces

[quasi bone]

Supporting goes into text_l node, not sure what it does but it probably reinforces text_g prompt in some way

[prisma grail]

I see, well.

you have multiple ways of doing that

though hmm

I have never used comfy ui so I don't really know the capabilties

[quasi bone]

I'm starting to understand that there's much more to it than just finding magic words that work everywhere... Thanks for tips though, two references kind of make sense 🤔 If I'd try to describe parts of each reference, it might be the result I'm seeking. Just gotta learn what triggers to use and maybe try incorporating loras to achieve final results

[prisma grail]

well you could look into loras and textural inversions

though I don't know if it works in comfyUI, I only use automatic1111

[solar dust]

Is it true that the latest NVidia update caused SD to go slower?

[sly fjord]

@solar dust I'm using drivers from February I believe because about a month ago with an update (RTX 2060) SD got ridiculously slow. What took previously about 15 seconds to generate took suddenly more than 5 minutes. Downgrading the drivers helped instantly.

[solar dust]

Tsc

And how do I revert this?

I'm using RTX 3060.

And the speed difference when compared to my old 1660 Super is almost nothing.

[sly fjord]

If it's really just the drivers causing this, you can try older drivers. I would recommend using the Studio Drivers, they are more stable.

You can download them off of https://www.nvidia.com/de-de/geforce/drivers/ and install them using Nvidia Geforce Experience

@solar dust But i suggest you first have a look if xformers is in use, and if you don't have any arguments like --lowvram or so, there are several things that could slow it down

[solar dust]

These are my arguments.

[sly fjord]

Here are all the arguments for A1111, --no-half is also performance argument and could slow down the process, but I'm not entirely sure, I'm not using it 🤔https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Command-Line-Arguments-and-Settings

[sly fjord]

I also have a question for anyone who may help. I'm trying to find a way how to prompt SDXL to replace part of an object with something different. Is there a way to use something like "instead" in prompt? I know I could go for inpainting or regional prompter (once it's working with SDXL), but I'm searching for a way to do this within the prompt itself. An example would be "a photo of hamburger, cigarettes instead of meat". No matter what I do (give more strength to keywords, use negative "meat") it renders normal burger with meat in it.

[dawn hawk]

You'd prolly specify it like "Barbie World style" or somehing like that

[dawn hawk]

Have you tried it with "shiver lines"? If that doesn't work I'll take a crack at it later

[dawn hawk]

Don't think it works like that. It tries to have everything in the positive prompt inside the picture.

[prisma grail]

hmm

[dawn hawk]

Inpainting really sounds like your best option

[sly fjord]

That's what i thought. Never mind, hopefully regional prompter or something similar will be available soon

[prisma grail]

it's what I am doing but it's not helping as I don't know what they are really called

[dawn hawk]

No that msg was to someone else

someone wanted cigarettes in their burger for some reason

To you I was like "you should try 'shiver lines'"

[prisma grail]

ohhh

[dawn hawk]

And I'm waiting to see if that works

[prisma grail]

srry

[dawn hawk]

If not

Then I'll do a bit of scouting

[prisma grail]

yea I'm atm taking a break lol

but I tried doing nervous lines/shaking/vibrations but the model I used seem not to understand

[dawn hawk]

https://danbooru.donmai.us/wiki_pages/trembling

The danbooru tag seems to be trembling

What model did you try your stuff on?

So my python is acting weird, it's gonna be a while before I can try it out myself

[prisma grail]

hmm

that is good to know

[dawn hawk]

I think I was running SD the whole time in ComfyUI with Python 3.11

That'd be hilarious if it's true

[prisma grail]

lol

I have anythingv4.5 and divine elegance mix v4

wait wtf is it dumb

hmm...

anythingv4.5 creates the effect at clip 8

alright I got it, divine elegance does do it at clip 2

nice

thanks for the help

[dawn hawk]

What exactly worked

[prisma grail]

tremble

the effect

but it's not always perfect

[dawn hawk]

I mean what did you use

[prisma grail]

(tremble:1.5) with the two models

[dawn hawk]

That seems pretty high

Ima try things out on my end

[prisma grail]

I don't know if trembing is better and I put it high cause I found it hard to get anything

[dawn hawk]

Can you send me a pic of the effect you want

Just so we know we're talking about hte same thing

[prisma grail]

woh cnave you yor......

is sleeping better than sleep?

like if the -ing is prefered or not

[dawn hawk]

sounds like it depends on model, but I think more ppl would tag an image as sleeping than sleep

[silver valley]

Dont use --no-half with an rtx3060 use --no-half-vae

[dawn hawk]

I'm looking into it, I think you need to look for a model that was more trained on manga stuff

Than anime stuff

[prisma grail]

when do you want --no-half?

[prisma grail]

I see

[dawn hawk]

This is w/r/t shivering

[prisma grail]

w r t?

[silver valley]

When you have an gtx1660 or an rx5700xt

[dawn hawk]

with regards to

I'm currently looking if maybe manga LoRAs have picked up the concept

[prisma grail]

alright

[prisma grail]

I haven't thought about looking into loras but that would be beneficial

[dawn hawk]

fluffy + darksun seems to have really good effect for me

0.2 merge ratio

[prisma grail]

what does merge ratio mean

oh

I see

I'll try it out a bit later

[dawn hawk]

I think most of it is from fluffy

[prisma grail]

which model is fluffy?

[dawn hawk]

Can I DM

[prisma grail]

alright

yes

[solar dust]

Can I use --no-half-vae along with --xformers?

[silver valley]

Yes

[tawdry minnow]

i'm trying to try outpainting like this video https://www.youtube.com/watch?v=vpmy_6cyI7c
by using controlnet, only paint+lama / controlnet11 model inpaint, but after all, i only get this on both side of image instead

[prisma grail]

damn... I am also on that video

I don't think mine even downloaded, did you download it manualy?

[granite crescent]

are you supposed to keep the width and height the same? (512 x 512 or 1080 x 1080)
if I try to do something like 1080 x 1600 it will generate two people morphed together instead of one person and the results will be all wonky.

[prisma grail]

weird...

what model are you using?

[granite crescent]

Epicrealism. I can show you the 1080 x 1080 generation and the 1080 x 1600 generation to show you the difference

[prisma grail]

perhaps the model is not good with large sized.. though just a guess, but try a different one and see if it does produce similar results?

[granite crescent]

Same prompt using two different models. It's generating two people even though the prompt says nothing of multiple people.

the first model is SDXL

[prisma grail]

hmmm

that's odd...

I don't know why

[granite crescent]

am i writing something wrong in the prompt?

[prisma grail]

that is confusing...

[granite crescent]

if i set the width and height to 1024 x 1024 the result looks normal with just one person.

but i dont like this aspect ratio for portraits

[obtuse torrent]

@granite crescent try a 2:3 aspect ratio but more centered around the base of the model (1024) IE 768 x 1152

[fierce patio]

Any moderator in here?

[rich crow]

yes for sdxl try using these resolutions

that pic with the double face is the expected result if you stray from the resolutions that the model is trained on

[granite crescent]

Thank you @rich crow and @obtuse torrent 🙏

[tawdry minnow]

i thought use controlnet like in video it will automatic extend the picture for me but it's not, i need to use inpaint instead

[slender pasture]

Are there any discord servers specifically about prompt building? 🤔

[slender pasture]

put multiple people and twins in the negative prompt

[silver valley]

The epic realism model is 1.5 based so its best resolution is all near 512x512.
Try 512x768 and use highres fix to upscale it. Then you dont get duplicates

Sdxl is made for 1024x1024 so you can try 768x1024

[zenith isle]

how could I make wings bigger? Whatever I tried, to get that size

parameters

masterpiece,(bestquality),highlydetailed,ultra-detailed,(church),solo,(1girl),(detailedeyes),(blueeyes),(longblondehair),calm,(golden armor),((winged angel)),((wings, white, shine)),(Angelic), Aura,God Ray,Feathers,(shinehalo),(rosary),(candle),(cross:1.2),
(8k, RAW image, best quality, masterpiece:1.2)1girl,blurred background,blurred foreground,lips,looking at viewer,motion blur,realistic, lora:ShinyAngel:0.5 lora:2dhxWings_v1.4:0.5
Negative prompt: sketches,(worst quality:2),(low quality:2),(normal quality:2),lowres,normal quality,((monochrome)),((grayscale)),skin spots,acnes,skin blemishes,bad anatomy,(long hair:1.4),DeepNegative,(fat:1.2),facing away,looking away,tilted head,Multiple people,lowres,bad anatomy,bad hands,text,error,missing fingers,extra digit,fewer digits,cropped,worstquality,low quality,normal quality,jpegartifacts,signature,watermark,username,blurry,bad feet,cropped,poorly drawn hands,poorly drawn face,mutation,deformed,worst quality,low quality,normal quality,jpeg artifacts,signature,watermark,extra fingers,fewer digits,extra limbs,extra arms,extra legs,malformed limbs,fused fingers,too many fingers,long neck,cross-eyed,mutated hands,polar lowres,bad body,bad proportions,gross proportions,text,error,missing fingers,missing arms,missing legs,extra digit,extra arms,extra leg,extra foot,
Steps: 30, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 3087966093, Face restoration: CodeFormer, Size: 512x768, Model hash: d7e2ac2f4a, Model: majicmixRealistic_betterV2V25, Denoising strength: 0.3, Hires upscale: 1.5, Hires steps: 20, Hires upscaler: SwinIR_4x, Lora hashes: "ShinyAngel: 52d3aa6156ee, 2dhxWings_v1.4: db3ba64ac381", Version: v1.5.1

[silver valley]

Maybe the 2d Wings lora didnt got trained on big wings

Try without the lora and try
Large Angel Wings,

[zenith isle]

thanks, it works

[tired vigil]

I'm curious, can you show what you got ? :D

[junior sonnet]

Hellooo

My prompt is huge coral, in the middle of the city, warm tone, with fish flying around the coral, sharp detail, contrast, smokey

Why its blocked?

[solar dust]

What's the model you used on the right?

Also which prompts did you use?

[potent laurel]

Hello guys, has someone try to generate images from a CCTV view? sdxl does a much better job than any 1.5 but I would ask for some advice. For now, "CCTV footage of [scene]" is what it has worked better

[junior sonnet]

Can someone tell me? Why my prompt blocked?

[rigid radish]

Anyone know a site that has a gallery of comfyui SDXL workflows that's easy to browse and navigate? I'm especially interested in xyz type workflows

[junior sonnet]

Itried to prompt:woman, wearing full crystal armour(weigthing:1.0), translucent, on top of building, shot from below, edge lighting, cool tone, cinematic, colorful style:Anime negative_prompt:skin
But the negative prompt doesnt appear

So it keep showing skin..what to do?

Can anyone help?

[solar dust]

@silver valley

Started happening after I replaced --no-half for --no-half-vae.

[silver valley]

Do you use --xformers ?

Also what settings did you tried?

[silver valley]

Okay then something was set to high

[solar dust]

Also those are my usual settings and I never had that problem with it.

I have 12GB of VRAM.

[silver valley]

Your upscaling to 2k with highres fix

Thats a bit much

[junior sonnet]

How to remove skin? Negative prompt keep removed when i enter my prompt

[silver valley]

Skin ? Only her face has skin

[junior sonnet]

Thigh

[silver valley]

Oh

Full body armour

[junior sonnet]

My prompt already use that..full armour is less than full body?

I added negative prompt nudity and skin, but when i enter my prompt..it only show until style anime

[silver valley]

Hmm maybe nudity gets blocked

Try add nsfw

Or skimpy

As negative

[solar dust]

@silver valley Yeah, something is wrong

even 512x512 is failing.

[silver valley]

If you get an out of vram error you need to restart SD before trying it again

[solar dust]

I already restarted my whole computer

[junior sonnet]

Word skin is removed also

[solar dust]

and the error persists.

[silver valley]

Is your gpu driver updated? Does some other Programms run in the background?

[solar dust]

I did outdate my gpu driver because people said the current one was slower.

[junior sonnet]

May i give you my prompt? And you test it in bot channel? Or maybe checkit?

[solar dust]

It was working just fine for a moment

and then all of a sudden I get that error.

After like, 3 images generated.

[silver valley]

Hmm i thought the driver issues got fixed

[silver valley]

Sure i can try

[solar dust]

I'll update the drivers and see if the problem persists.

[junior sonnet]

Thankyouu

[solar dust]

@silver valley Error was fixed but now it's back on being slow.

[silver valley]

how slow

[solar dust]

Man I wish Nvidia didn't screw up things with the newest update.

Well, what was once something that took less than 3 minutes with those settings is now taking longer than that.

[silver valley]

hmm

[north sun]

[(arms up:1.1)::0.8], what does"[::0.8]" do

[stiff cobalt]

I’m trying to reverse engineer this prompt. Any ideas on what art style this is called? Speed painting comes close. Any others? Thanks!

[granite crescent]

instead of specifying a shirt color like "red", "purple", "magenta" is there a way to really dial in an exact shade?

[rich crow]

ime no, so extensions like regional prompter https://github.com/hako-mikan/sd-webui-regional-prompter help

https://github.com/hnmr293/sd-webui-cutoff is an alternative

[tired vigil]

Can someone tell me how I convey the following to the chat AI generator? "A cross between bobby hill and marilyn manson"

[rich crow]

there's a few tricks you can do for this, you can do [Bobby Hill:Marilyn Manson:.5] and it will first start off with bobby for the first half of generation and marilyn for the second half
you could also try [Bobby Hill|Marilyn Manson] where it will alternate every step: i.e. step 1 it generates bobby hill, step 2 it generates marilyn, step 3 it generates bobby hill, step 4 it generates marilyn, etc etc

[tired vigil]

Sick

[slender pasture]

Have the amplifiers been adjusted for XL 🤔

I find that if I add a amplifier over :1 on most anything my outcome goes all wacko

I can't for the life of me figure out why It keeps disfiguring her last muinet 😐

I've removed all the amplifers, and im going though looking for any keywords that may be repeditive in any way 😐

Finally figured it out-

I had to strip the code of almost all of the amplifies except for things that I specifically wanted control over like her hair gradient and the color theme-

Getting used to the lack of need for excessive tagging is taking some time 🥲

[slender pasture]

🤔

[obtuse torrent]

I'm not sure if that type of image conforms to the rules of the server.. it should be consulted.

[slender pasture]

mine or? - cuz i can delete it?

[obtuse torrent]

i m not moderator so..

[slender pasture]

I think they're refering to mine cuz my OC was in lingerie-

I deleted the image 🤷‍♀️

[obtuse torrent]

I only expose a possible point of view so that you take it into account nothing more...
I was not asking for anything and I do not feel offended, I understand that they are only trying to explain their doubts and/or results

[slender pasture]

Im not offended at all 🙂 I'd rather be on the safe side- ty

[west lark]

hi I am incredibly new to AI, I want to generate this style of images everytime, how is that possible?

does anyone know the prompt?

[slender pasture]

deepbooru it 🤷‍♀️

Looks like

Limited Pallet
Sketch

maybe watercolor

[west lark]

bro I created it, I only wrote frodo and bilbo. Of course its not always gonna be this with a prompt thats so simple. There were three other generation with different style, I only want this style always

I know with stable diffusion you can have control of the style and all

anybody?

[obtuse torrent]

@west lark if you're right, the recommendations may not be effective on all models... do you remember what type of model and the name of the one you used as a base? or was it in a bot?

[west lark]

among four only one is in this style, I am sure there is a prompt to always get this image

[obtuse torrent]

I only have normal SD running with A111 UI, my models are 1.5 base I can only recommend some tags and use Controlnet with Only reference:"D&D style" "1850 year style" "((monochrome, brown theme, sepia))", "storybook illustration" "border" @west lark Apparently you are generating with a bot or application, if so and the application does not show you more details about the models used, it is not what you will achieve, because they apply styles in an "invisible" way and only show results

[solar dust]

What do you mean amplifiers?

Because sometimes SD insists into difiguring my characters as well.

[west lark]

thank you so much. you are the best

I will try this on some base model, I have gtx 1660 and don't know how to prompt

[slender pasture]

i mean things like (portrait shot:1) 👈 that

[west lark]

is there a place where I can try different sd models? to know which one always results in this?

[slender pasture]

Amplifying the keyword with modifiers or whatever

[west lark]

downloding all of them is gonna take long time

[obtuse torrent]

Or I hope you have luck with this..SD can run in different forms, one of them is A1111UI but it is not the only one, in this a lot depends on the OS and the type and amount of VRAM available..
----check the requirements first... if your system is not very robust I can PM you a link to one of the paid sites, which will let you experiment for free at low resolutions and which has interfaces more similar to A1111 to familiarize yourself with the concepts and the large number of tuned models that exist

[radiant maple]

anyone got negative prompts to prevent deformed/ugly eyes? everything else is near perfect but the eyes are always janky

[west lark]

please do. I want consistency. Also I have 6gb vram

my pc takes too long to generate I need something faster, i need an models running on the cloud with API endpoint which I will prompt through an LLM

all autonomously

may I dm you?

[obtuse torrent]

For the specific use you are looking for, I am not the right person because my PC is as limited as yours and I possibly know less about the script language... but if you consult the anime channel in a polite way, they will help you, because many people there have knowledge and experience doing the same ,, https://discord.com/channels/1002292111942635562/1091193032273043516 gl

[solar dust]

It's amazing how even when I use 1 single lora it still can lead to disfigurement/extra limbs.

[west lark]

hi guys how do you ensure a certain action?
like I prompt legolas shooting an arrow at an orc and orcs either are no where to be seen or he is not shooting at them.

Can AI still not handle complex compositions like that?

also if I say two three character names, the model just breaks. and shows generic stuff, it can't even handle multiple characters?

[solar dust]

Multiple different characters have always been a challenge for SD.

[quiet zodiac]

@west lark Try adding to the front of your prompt:
in the style of Aubrey Beardsley
Your reference looks a lot like images from the Art Nouveau Era. Perhaps keyword that, too.

Stable Diffusion can basically only handle one concept at a time. If you pre-compose an image with your three different characters, that can go along way to strengthening the output of your prompt.

[slender pasture]

Yeah I had the same issue, but you know you can change the number at the end of the lora right?

Like <good hands.lora:1> 👈 You can make that a :0.3-:0.5 and it tones it down a bit

takes some tiknkering

2 people + is a fun challenge

[high solar]

Hey all, does anyone know if its possible to save a controlnet image as part of a prompt or style to not have to keep doing the process of upping the image, selecting its processor, etc

[crystal cipher]

I'm trying to make a floppy wide-brimmed hat for a character sketch of mine, but the hats just won't be floppy. prompt:limp, flaccid, droopy, relaxed, flabby, squashed, flimsy, flexible, pliant, supple, wide brimmed cloth floppy hat, on a chair, downturned wide brim negative_prompt:rigid, stiff, sturdy, inflexible, sound, firm, compact, man, woman, person. The best I've come up with is this:

What am I doing wrong? On the bots in sdxl

[west lark]

thank you so much. After spending hours on different prompt sites, I did see this Art Noveeau Era thing but thanks for Aubrey Beardsley part. Also what is pre-composing? I am an absolute noob as an artist, I am a developer.

[hard phoenix]

anyone has prompt that i can use to fuse character its really hard to that with stable diffusion

[fierce patio]

fuse? your conFUSEing me

[hard phoenix]

its really difficult to explain basically in mid journey you could put to characters together and get the fused version of it, was wondering if it would be possible to do that in stable diffusion.

you can look it up

[fierce patio]

show me a picture of it

[hard phoenix]

this

turns to this:

or something of sort

[fierce patio]

dont they call that blend mode in MJ

[hard phoenix]

shiii it might have been what they called it

[fierce patio]

oh you want a promp for that in stable diffusion

[hard phoenix]

yep if there is any

Yessir appreciate it bro

[fierce patio]

heck im going to try it

[hard phoenix]

💀

got this

[gritty drum]

anyone provide sample prompts to add furniture in room

[thin ether]

Does anyone know why my results suddenly got brighter?

I did change two prompts between these two images;
(Light Brown hair:1.2)
got changed to
(Light Golden Brown hair:1.2)

(freckled nose:1.2)
got changed to
(lightly freckled nose:1.2)

Now, in my head since I added golden to the hair and lightly to the freckled nose I start to wonder if that might somehow have influenced the colouring of the image?

Does anyone know if that's the case and if so, how can I prevent it?

NOTE: all other prompts remained entirely unchanged

Hm, so it actually is the word ''Golden'' in the hair colour

[dark rune]

Hey folks. I'm just starting out with Stable Diffusion. I've had good results using DreamStudio but not so much locally with SDWebUI (Automatic1111). On DreamStudio there's a handy styles dropdown that helps get the desired style easily (Photographic, Cinematic, Cartoon, etc). How can one replicate this locally? In the Prompt itself? Some add-on? 🤔 (Sorry if this is the wrong channel for this question.)

[orchid ore]

@dark rune if nobody answered, if you want learn how to prompt, probably interrogate images in img2img can help you.
What is wrong with your images made in A1111? Tags you mentioned are working as well in A1111.
Have you proper model, or can you show your settings?

[quiet zodiac]

@west lark By pre-compose, I mean supply an image as a reference. It can be a photoshop hack, that's what nice about SD, it is forgiving when sampling the reference. Roughly cut out an elf, dwarf, and ranger and place them side by side on a background. Size your pre-composed image to the same size your requesting for output. I generally use 768x576 for a lot of work. Then prompt elf, dwarf, ranger in the wilderness.

[dark rune]

Here is a screenshot of my settings.
COMMANDLINE_ARGS="--skip-torch-cuda-test --upcast-sampling --no-half-vae --medvram --opt-split-attention --listen --use-cpu interrogate"

[silver valley]

Whats your GPU?

[orchid ore]

running on cpu, it must take very long time. Interrogate isnt to Commandline_args imo. it is just thing on img2img second tab.

[dark rune]

I'm on an intel Mac. 32gb Ram, 6 core i7, Graphics -> AMD Radeon Pro 5300M 4 GB

[silver valley]

Ahh okay

[dark rune]

as far as I understand it, I can't use the AMD chip but only cpu

[silver valley]

To get styles you need some prompt words to describe it.
Like photorealistic, photo of...,

[dark rune]

I'll try tinkering with that then. Each image takes like an hour so it will be slow going lol. Thanks for the help, folks!

[orchid ore]

curious if comfyUI on cpu is as well faster. @dark rune you can probably try it. If it saves you few minutes, but dont know.

[quick haven]

I'm trying to use SDXL APIs to outpaint images, although everytime I outpaint I get basically a framed version of my image. What am I doing wrong?
does anyone have similar issues?

[west lark]

Hi, But I am going to be generating it fully autonomously with a prompt generator. I want a composition hack that is text based. Like prompt where I can seperate, background, characters and action they are doing. Sorry, I am mostly familiar with text based models and you can control ouputs by few shot learning etc. I need some help. In which I do composition programatically.

[quiet zodiac]

I don't think SD can do that at this point. It even fails on simple things like.
a syrian man wearing a black suit dancing with an african women in a red dress
SD will mix up the ethinicity and clothing type and clothing colors. You'd be surprised how much variation you can get from one reference image. Using a reference image also allows you to control placement, pose and background.

[rich crow]

Best way we can achieve what you want is with this extension https://github.com/hako-mikan/sd-webui-regional-prompter

[junior sonnet]

so, i got this output..i set the style to low poly and unexpectedly it only turns the armour into the low poly..is it because i use ((wearing full armor)) ? so it only change the armor, not the whole?

[west lark]

bro, that's so awesome, Thank you for sharing.

[lavish trench]

is (shower cabin) correct? 🤔
because I'm not able to have a shower box at all ¯_(ツ)_/¯

[young storm]

Hello, can any of you help me find the style of this image / a prompt / artist name to generate something similar? I only care about the drawing style. Thank you!

[crisp lily]

According to google image search
Yū Itsuki
Japanese illustrator

[young storm]

thank you very much, i did not realize this was an image that exists online- I'll try to prompt artist's name to see if i can get something similar.

[vivid ore]

I am finding difficulty trying to generate the look of original anime characters in the negative prompt. What are the prompts people use to reliably generate male characters to avoid that overly feminine/ neotenous look?

[plain egret]

anyone know what BREAK & AND does in prompts?

[granite crescent]

how do you prompt to separate colors for different clothing items? For example, if i prompt beige blazer and black pants the results will often generate beige blazer and beige pants, or black blazer and beige pants, etc.

[orchid ore]

@granite crescent nothing i think will work 100%
Personaly i did "Full body image. Man wearing (beige-blazer:1.5) with (black-pants:1.5)"
And it somehow works. Also using portrait layout 768x1024
Actually works pretty well.

@plain egret BREAK keyword is to extend token limits. And AND i dont know.

[rich crow]

adding to ^, usage of BREAK has to do with token padding, so anything you type immediately after BREAK has a higher weight

[granite crescent]

thanks i will try this now.

[tired vigil]

Someone help

Idk how to use LoRa properly

I wanna generate an image like this

But idk how to

[plain egret]

add lora:X:Y to prompt

X is the lora name, Y is the weight e.g. 0.8, 1.1

[tired vigil]

Can you show an example

Yo?

[orchid ore]

if you are using A1111 there is icon under generate and it is matter of few clicks. I mean using loras or other features. I think it is middle icon, cant confirm now.

[obtuse torrent]

@tired vigil did you download that from citivai? ..in the presentation page it often explains how to use the model... normally you need to first download it and then move the file to the corresponding directory of A1111
then refresh the UI and use it as explained above

[tranquil shard]

Hi, I'm finalizing an extension for WebUI to catalog and make it easier to work with and research prompts.
You can get it here - https://github.com/AlpacaInTheNight/PromptsBrowser. It have instructions on how to install and use it.

I also created a prompts collection you can download and try with this extension - https://github.com/AlpacaInTheNight/prompts_portrait/

[tired vigil]

i know but why it dont look like the image i want it to be

it lookin like this dawg

[digital bluff]

how to write prompts that displays text on objects in sdxl

[obtuse torrent]

@tranquil shard I sincerely appreciate your effort and the spirit of sharing this with us, I hope it helps a lot of people
I recommend, if you haven't already done it", that you try to contact a moderator so they can evaluate fixing your message on the appropriate channels
(no problem posting this here but it won't be prominently visible for long)

[tranquil shard]

Thanks. Whom from moderators can I contact to pin link to my extension? Someone from community mods?

[warm nexus]

Hey guys, I need some advice. I'm trying to make my automatoc 1111 to produce a very specific concept: hair that are made from metallic plates. I've tried prompting "metallic hair", "plate hair", "metal ribbons", "mechanical hair" and at best it draws hair with metallic sheen, at worst it keeps ignoring that part of the prompt

any ideas on what I can do?

[tired vigil]

hjelp

[warm nexus]

This is the best result so far https://cdn.discordapp.com/attachments/1030156141960904714/1136800531264852079/00060-3195082628.png

And this is roughly what I wanted https://cdn.discordapp.com/attachments/1030156141960904714/1136821016820207667/GbWA9iK.png

[obtuse torrent]

@warm nexus control net tool or regional prompter ....

[warm nexus]

I don't really know how to use that

and i fear that it doesn't understand the concept itself rather than being overriden

since naked prompt with metal hair doesn't work either

[obtuse torrent]

this the template for Cnet ..

[warm nexus]

no, not that, I mean how the hair looks in that model

not posing

you see how hair is made from solid slabs of metal?

rather than individual strands

[digital bluff]

how to write prompts that generate text pls help me

[obtuse torrent]

@warm nexus I have no experience with prompts for cgi....in terms of tags this is just "long hair", "straight hair", "high ponytail" or just "ponytail" .....
what really must be missing is some global descriptors and/or models used, for CGI animation that give a less detailed result... it's difficult if you ask without metadata or prompts to analyze....well
I think you were looking for exactly the opposite, in that case you should add even more detail descriptors for the hair case, perhaps use other upscalers and/or methods that allow adding more details in those processes, such as reimplanting the hair so that it is generated even higher resolutions

[oak plaza]

I saw some people saying that brackets don't work the same in SDXL as the previous versions, i have also seen people using all sorts of variations on BREAK and brackets - Does anyone have any good resources / references or guides on how to properly prompt for different models?

[obtuse torrent]

@oak plaza you check links here how a first step ... https://discord.com/channels/1002292111942635562/1080946152318443610

[tired vigil]

help? pls

[obtuse torrent]

@tired vigil I don't have the model but I guess it's enough that you get 1 generation with matadata so you can see how it was done

[tired vigil]

whats matadata? sorry for my ignorance

[oak plaza]

I had a look through these links, but its mostly just posting to official pages and communities for tools, I'm having a hard time finding clear prompting guides or explanations.

[obtuse torrent]

@oak plaza no one here could help you... in the chat ?https://discord.com/channels/1002292111942635562/1089974139927920741sometimes ... it's a matter of time before the right people read it to you, unfortunately my equipment is very basic so I haven't invested time checking an extension that I can't run well ..

[obtuse torrent]

when someone intends to share the way they achieve an image, that person generates it with all the necessary settings... information attached to the image that indicates how to fill in all the parameters and what models were used to generate it ....= metadata

[tired vigil]

k

can u make that a more simple explaination? sorry for asking too much

[obtuse torrent]

@tired vigil give me the ink where did you get the image you want to imitate... or the model, I suppose you downloaded it from somewhere where you saw the image..

[digital bluff]

......

[obtuse torrent]

It's the same thing... readable text in the image is a feature of that model (which I don't have) only someone who is using it or the documentation itself can help you, sometimes it's a matter of luck to coincide with someone in the chat who have experience

@digital bluff normally you would only need an image with metadata to know how it was done

[tired vigil]

i wanna imitate this kind of style

[obtuse torrent]

@tired vigil no metadata ...for it i asked you for the link ..

[tired vigil]

oh

https://civitai.com/models/20809/ceres-fauna-hololive

[digital bluff]

thanks 4 response

[rustic ravine]

unless you only want text, mostly luck

[digital bluff]

nah text on some object

u see i tried on no style

i got 50% correct results

like text with correct placement

[rustic ravine]

best bet is to generate that object, then use gimp/paint/photoshop to add the text; it is very hard for latent-diffusion models to make text

[digital bluff]

well i am not that pro in adding text which blends too perfect

as the sdxl got an update to generate i thougt it will be natural

if the ai itself generate

[obtuse torrent]

Definitely the new models can do that... I've seen it, the old ones (1.x 2.x) do something similar but the text will always be unreadable, just scribbles

[digital bluff]

i got perfect image

one time

first and last

i didnt download sadly

i wanted to make an ai based webtoon ,keeping in that mind i should also focus on things like character expressions, i have a doubt in same seed if i get a character good if i give same seed to my friend and ask him generate different pose with same character ,is it possible if he use same prompt

with slight changes

[rustic ravine]

p.s. seed doesnt do that, your best bet is to be very descriptive in the prompt

[digital bluff]

i will do but if i wanted to collab

with some one

should i share the seed

as style changes without same seed

[rustic ravine]

(seed changes randomness)

its useful sometimes

but not for different angles

[tired vigil]

https://civitai.com/models/20809/ceres-fauna-hololive

[dawn hawk]

BREAK makes A41 create a new batch at that moment, this is useful if you want to separate certain concepts to make sure they are even less likely to bleed into each other
AND is used to merge two tokens together, e.g. you can merge horse AND tiger, and will get some mish mash in between them. There's different ways to combine them, but for that you need extensions like the neutral prompt extension which allows u to use AND_SALT and AND_PERP

[dawn hawk]

Have you tried (helmet, long metal embellishments:1.2)?

[digital bluff]

is there any dedicated channel for seed discussions

[dawn hawk]

Also, ig download the metadata from here if it exists

[warm nexus]

no, but I specifically disabled helmets so it would stop obstructing the face. I'll try that now

[dawn hawk]

You can add face to the prompt

It's not like all helmets hide the face

[dawn hawk]

Seed doesn't matter that much

One seed will net two very different pictures based on the prompt

If you want to generate text

Just add it in with paint

And then inpaint it

[obtuse torrent]

@tired vigil --metadata ...

[tired vigil]

so what do i do?

[obtuse torrent]

@tired vigil well... there's a lot you need to look at with that metadata... to get a similar image....
at least you must have all those settings the same... the same yesmix 1.5v model or a similar one, the lora at: 0.7 weight, a 4x ultra sharp upscaler, etc. so a large number of parameters that can make the image look very different if not the same ...