📝｜prompting-help

[drowsy chasm]

Hello everyone! My friend and I are working on a project together and we want to keep track of our progress and which prompt produced which image. We tried saving everything into a shared folder, but that didn't work out well because we wanted a better way to view the information. Is there a tool that can help us keep track of the prompts, trials, and other information, something similar to Weights and Biases?

or do you recommend a way that worked out for you

[simple finch]

hi, newbie question, what is this? is it a different UI tool other than Auto1111?

[steep fossil]

Yeah. Apparently we now have Comfy UI and Invoke (which I've not used yet)

Auto needs a good categorization system for add-ons now 🙂

We should also have some security system, or at least user reports for safety ratings

[quiet zodiac]

@drowsy chasm I'm not sure if this helps you, but the Houdini interface for Stable Diffusion, called MLOPs, now contains a field for writing information into the metatags of the JPGs it generates. I typically write my positive/negative prompts, the model, CFG, steps, and image guide strength. It's nice to be able to go back to an image a few weeks later and discover what prompts you used to generate it.

[drowsy chasm]

this sounds amazing! thank you so much will test it out

[next flint]

I see a lot of text to image for qr codes but how do i do image to image ones

[drowsy chasm]

you mean an image to qr code ?

[next flint]

kinda. Im looking for something like controlnet shuffle but it shuffles it into a qr code

[honest cave]

What's the prompt to hide a hand? It keeps appearing if front of the character's face but I can't get it to just.. not be there

[thorn dock]

How important/relevant/effective to use characters like ( ) | to separate semantically distinct things in a prompt? Or, I saw people using BREAK between smaller phrases. Is there a guide or a common knowledge regarding those? Thanks!

[quiet zodiac]

The prarenthesis are used to add weight to a concept contained within. Not sure about the vertical pipe symbol. I've never used that. I believe the BREAK is used on long prompts because there is an initial 75 token limit. The break signals the processor to add another 75 so tokens aren't lost.
In my experience, good sentence structure is a better approach to distinguish separate concepts rather than symbolic tricks. However, Stable Diffusion is not very good at managing more than two concepts. Often, portions of the second concept you introduce will bleed into the first. That's why its hard to render 'some dude in a black suit dancing with some chick in a purple sequin dress'.

[thorn dock]

thanks a lot, that's very helpful 🙏

[orchid ore]

@thorn dock | is i believe from dynamic prompt extension
https://github.com/adieyal/sd-dynamic-prompts#basic-usage
and break could be in prompt as well when you are using Regional prompter, it change all keyword in BREAK

[thin creek]

hello there

i have a problem with regional prompting

sometimes the ai switches around the sides of the screen

and sometimes it only looks at one prompt

with:
a portrait of a woman with black hair and red eyes
ADDCOL a portrait of a woman with blonde hair and blue eyes

i sometimes get

here it got the sides correct

but the eyes are purple

and the last one is literally perfect but switched sides again

uh

what do i do?

its pretty confusing

[coarse karma]

How to avoid this?

[drowsy gazelle]

Click on hires fix

[tired vigil]

how do i give negative prompts

[cinder dagger]

https://github.com/Jeremy-Fuller/Prompts

[crude crystal]

Anyone know how to use Temporal Kit? I followed this tutorial along to 15:30 when he says to click Prepare EBsynth and nothing happens, I don't get the key or frames folder created or filled. https://www.youtube.com/watch?v=rlfhv0gRAF4&ab_channel=enigmatic_e

edit: I manually created those 2 folders and then it worked

[orchid ore]

@thin creek i have solved your hair issues. But red eye i cant will try again. I just checked common prompt and in prompt i have
.
.
Woman portrait ADDCOMM
blonde hair with Green eye ADDCOL
black hair with blue eye

I have not single wrong blonde black side
But trying on photorealism i have not anime.
i found some similar model

[thin creek]

ooooh. i will try putting them at the back of the line

thx a lot

[orchid ore]

do you have regional prompter extension? @thin creek

[thin creek]

yes

[orchid ore]

ok

[thin creek]

is there a way to fade the left image into the right?

or merge 2 different styles in the middle

[thin creek]

i seem to have troubles using 2 different loras for each side. it just applies both to the whole image

[orchid ore]

try apply them in each collumn not before ADDCOMM. What is before ADDCOMM is for whole image. If you put it there then this is the reason.

[thin creek]

i have it like:
stuff i want ADDCOMM
<style1> style1 ADDCOL
<style2> style2

the styles in <> are loras

[orchid ore]

o.k. tried and result is very bad. try move them in ADDCOMM line. I am going to try something as well

[thin creek]

kk

[coarse karma]

Ok. Thanks

[orchid ore]

@thin creek doesnt work. But i have only loras for general things, not for persons.

[thin creek]

oh ok

well that is unfortunate

anyway thx for your help!

[thin creek]

well....it works with inpaint

[orchid ore]

np, pitty it doesnt work, as well probably very depends on model

[thin creek]

its pretty funny

[thin creek]

only weird part is that here i only coloured the bottom left part

yet it changed the eye completely

[orchid ore]

i think it is because it is on half of the image.

How about generate some image without loras, send it into img2img and here use controlnet iP2P with lora?

[hybrid harness]

is this where I would gofor help downloading or no?

[orchid ore]

#🤝｜tech-support @hybrid harness

[thin creek]

iP2P?

[hybrid harness]

Thank you

[thin creek]

never heard of it

[orchid ore]

it is part of controlnet

you have statue of Eva, and you can type turn her into statue of Adam

[thin creek]

i heard control net is pretty big

but idk what it is

[orchid ore]

yes model is pix2pix

controlnet is something that add very much value to SD.

[thin creek]

wow your help with regional promting was amazing!

[orchid ore]

i am not sure....

[thin creek]

look

works perfectly

just how i wanted

idk where the horse comes from

but looks cool too i guess

[orchid ore]

have you use addcomm? probably horse is part of your figure, i mean some word that triggering horse.

[thin creek]

hmm

lora:castlevania_style_offset:1 ADDCOMM
(beautiful moon: 1), dark sky ADDCOL
ADDROW
ADDCOL
a portrait of a beautiful female knight with blonde hair in armor

this is what i got

nice npc for a dnd campaign

[orchid ore]

knight can trigger it i believe, not sure about castlevania

[thin creek]

yeah

im trying sth funny now

[thin creek]

nvm

i got sth cool instead

@orchid ore i cant fix the eyes but i am already happy

[orchid ore]

yes difficult good you are happy, thats the point! 🙂

scarry eyes 🙂

[thin creek]

lol

both she and the horse are on heroin

[orchid ore]

that is controlnet, for sure good to install,
yes 😄

[thin creek]

but the horse is the only one whoi enjoys it

[thin creek]

ill try it out

[teal valve]

this is weird ... why is it related to age..
girl 6-12
female 17-25
whenever i use these tags i get age difrences
girl shouldn't mean kid female...

[orchid ore]

woman weird as well would suppose >18

[quiet zodiac]

@teal valve That's exactly what the word 'girl' means {a female child or adolescent}. A woman is an adult female. I find it incredibly shocking to see so many prompters using the term "1girl."

[orchid ore]

adolescent is older than 12, isnt?

[teal valve]

Yeah but then allot of tags like 1girl, catgirl, doggirl and so on end up backfiring at you

And i gotta go around for example i go with
Mature female, fox ears and big tail.

Kinda anoying that i can't simply use a single prompt for that

[orchid ore]

try use magical "-" in mature-female. Or adult-woman or whatever. It is mighty symbol

[undone rivet]

Hello everyone, I would like to know if does exist any guide, to learn about the "base" for an optimal syntax/structure for a prompt...?

[lyric goblet]

Which model has best or most variants of "horns" and what is the best way to prompt them using inpaint sketch?

[drowsy gazelle]

1girl is a booru tag for anime models if you want a single girl/woman in the picture. it won't work right if you aren't using an anime model

Same with 1boy

[strange galleon]

Can I take one of the generated images and get the AI to modify only parts of it? I need the character to have longer hair and I want to remove the random blue jewel on a necklace it gave him.

[orchid ore]

@drowsy gazelle you know what? it is working with photorealistic non anime models as well. With some probably better and with some probably worse, model i am trying work well 1girl

you can send it in img2img and there choose inpainting @strange galleon

[sullen rover]

anyone know what model this image used? it wasnt listed on thee civit ai gallery ppost

i want to know since it looks like QQ/tencet AI generation style

let me pull up some example images

[orchid ore]

have you tried if it has some metadata?

[sullen rover]

let me see

well, the image didnt have any, but the other images in he same post had used anything v33 and the taiwanese food lora

[drowsy gazelle]

There are a lot of realistic mixes that also have some anime in them. Example: I've made anime stuff with deliberate before.

[orchid ore]

so at least something. It is difficult get exact model. With prompt and seed probably posible, but difficult.

i think that 1 girl Blip or Clip understand as well. Mmnt

i did it with 2.1 but cant find it somewhere, so i dont know 🙂

[teal nacelle]

is there a way to fix a great image that has clipping problems? ie a backpack strap going though the shoulder

[next flint]

So sometimes when i click on the lora i want to use it does something and othertimes it doesnt. What is the difference between calling model keywords and calling a lora model in the form of lora::1(name)

[analog basin]

hiiiii, my name is Florencia, I'm wanting to launch a beta version of a cooking app and I want to know what kind of prompts can work to generate recipe images. If you can give me a hand I would be very grateful

[drowsy gazelle]

I don't really think SD can make recipes like that? I think ChatGPT would be better IMO.

SD can hardly handle a few words without garbling everything

[fringe parrot]

Hi guys, I am trying to re create a character from a book series. I was wondering if anyone can help me with it. The character's name is "Skippy the magnificent"

[cinder saddle]

do you guys have any recommendations for sprite making websites
and sprite as in the pixel sheet

[tidal junco]

Guys, how can I make better full body images without the face lose quality? I'm suffering a bit on that

[cinder olive]

Any prompts for showing less skin? She's a warrior not a pole dancers.

Prompt - Masterpiece, best quality, a beautiful dragon-woman, fantasy, Dungeons and Dragons, detailed red hair, detailed dragon yellow eyes, detailed black dragon wings, Serrated Horns, wearing yellow medieval armour

Negatives - Sexy, Big Breasts, NSFW, lewd, risque, nude, hat, laser, lightsaber, lowres, bad anatomy, bad hands, bad faces, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blind, bad eyes, ugly eyes, dead eyes, blur, helmet

[plucky osprey]

Hi! Im using txt2img.py to generate images, is does not has negative prompt as parameters but I did some research and it looks like if you give the model for example a prompt a man riding a horse - tree, dog, cat:-5.0 the prompt will be a man riding a horse and everything after - will be negative prompt, and the -5.0 will be the negative weight, can someone confirm it?

I'm assuming it mainly from some issues on github and this blog post by stability ai https://stability.ai/blog/stablediffusion2-1-release7-dec-2022

[calm spindle]

Hey guys, can anyone help me with a prompt? How can I force angles like these? I got that by pure luck. I had this in the prompt (front view:1.5), (straight on:1.5) (upper body:1.5) plus a lot of other stuff.

Most of the times, the subject of the picture is just sideways while facing the camera

like this, but facing the camera.

[kindred dagger]

how do i make a prompt for spiderman so they colors of the suit is different then blue and red

[drowsy gazelle]

I wonder if looking at viewer would work

[calm spindle]

That only makes the head turns to the viewer, the body still stays at an angle most of the times

[drowsy gazelle]

Have you tried a controlnet with the human pose?

See the third one

[sand edge]

anyone have any tips on showing 2 characters with extreme size differences? for example, peter pan and tinkerbell

[obtuse torrent]

@calm spindle front view,
from front,
They are fine and should work, however the anime models have numerous biases, one of them having long hair forcing a side view to show it off.
Therefore, forcing the pose with controlnet is a good option.

[calm spindle]

ah, that may be it

Ill search a bit for controlnet

Thank you

Welp

(from front:1.5),

blue sky,

(white clothing:1.5),```

really dont get how these things works tbh

[silver valley]

You cam do that with Tiled Diffusion Extension with Prompt Control

[buoyant charm]

Or Regional Prompter

[obtuse torrent]

use : portrait, medium shot, frontal shot
1 girl, (white clothing:1.5), looking at front
blue sky in background

.... add to negative : (side view:1.5) .... this will slightly increase your chances of getting that pose without controlnet

[silver valley]

try it with:
(wearing green gold spidernet suit)

[sand edge]

@silver valley @buoyant charm badass, these both work great, thanks

[buoyant charm]

RP uses a bit unfortunate terms (base prompt and common prompt) that I always forget how to use properly. It's explained in the readme but I wish they chose something more intuitive...

[wanton girder]

ok some people dont know this but you can fuse two things with | like this cat|fox

what i dont know how weight here works is it like this ? (((cat))) or do we use numbers ?

[drowsy gazelle]

Each pair of parentheses is the same as multiplying it by 1.1. so 3 parentheses= 1.331.
You can write (cat:1.3) to get a similar effect.

I think the numbers are better because it's cleaner and more obvious what number you're going for. With the parentheses, you might screw something up by not closing them right

[obtuse torrent]

if you mean the weight in relation to the general prompt, I use this form ((cat|fox)), I have never tried the other one

[tawdry panther]

How do I remove NSFW filter from automatic 1111

[wanton girder]

so i would write (cat:1.3|fox:1.5) ????

[wanton girder]

you should question this in the automac 1111 discord there are plenty of people who make NSFW content who are experts for your needs 🙂

[tawdry panther]

Discord

[wanton girder]

yes they have one

@tawdry panther

[tawdry panther]

Link

[wanton girder]

ok one moment

you have mail

@tawdry panther

[sand edge]

Ya, it took me a while to figure it all out, but it's very powerful once I got the hang of it

[solar dust]

If I use upscaling, the images take way too long to load, even when using the medvram command

is there any other way to make things faster?

[drowsy gazelle]

Are you generating and upscaling at the same time?

[solar dust]

Yes.

Is it possible to do those things separately?

[drowsy gazelle]

Normally I generate an image and then upscale it afterwards if I like it

Here's a tutorial on upscaling - you can do it multiple times on one picture

https://www.reddit.com/r/StableDiffusion/comments/xkjjf9/upscale_to_huge_sizes_and_add_detail_with_sd/

[feral salmon]

Does anyone know what style is this?
I can't replicate it in sd

[next flint]

do you happen to have the original image file with the metadata?

[feral salmon]

https://www.metadata2go.com/result#j=44323314-7e34-46c4-91a1-dc1426bace60

[next flint]

If you had the original file you could use https://github.com/SupaGruen/StableDiffusion-CheatSheet to check for all the relevant data

[feral salmon]

Tried it: No Stable Diffusion EXIF data detected

[next flint]

then i would suggest image searching for similar styles or trying to contact the artist

[lyric goblet]

Not sure if this is the right channel but why do I get this half shade on the eyeglobe when inpainting and how do I get clearer eyes that is not so hazy? Is it higher res, settings or prompt?

[worn kettle]

i have a dreambooth model on myself that i trained using realisticvision 2.0, 30ish pictures at 1500 steps. i also tried with 2000 steps and got the same results. i'm finding the model only gets my face right with some serious prompting. i've found that "detailed ___ face" and "natural skin" get me close sometimes, but other than that it's a guessing game. does anyone have any tips?

[solar dust]

Yeah, about my previous question

I have slow results because my GPU is a 1660

apparently it operates prompts at 1/5 or 1/6 speed of a 2060 so I'm thinking about replacing my GPU with that

is it worth the investment?

[feral salmon]

invest in cloud gpus, you can get good gpu's (rtx 30 series) for less than 0.2 cents per hour

[solar dust]

???

Cloud gpus?

That's the first time I've heard about such thing.

[feral salmon]

Basically, you rent powerful gpus on the cloud. e.g vast.ai

And you pay only for the duration you use it

do some research, it's worth checking it out

[pine moth]

Hi, can someone explain what the 'AND' function does in a prompt? hard to search for it since and is a commonly used word lol, TIA

[silver valley]

Here is the wiki Information for AND:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#composable-diffusion

[pine moth]

ah thanks, i was on there but didn't spot that, cheers. i still don't quite understand what it is doing, it does something that i like though

[dusty folio]

hey anyone here got some experience with the extension grounded sam's prompts?

talking about the grounded dino section

[spring tiger]

Does anyone have any tips or good resources I can read on how to generate specific scenes? Im finding it hard to wrangle SD into doing it through prompting, and can't find images close enough to it to use controlnet to do it

[stiff ibex]

What promt should i use to get images in this Art Style

[cold gust]

Hey y'all I'm new to Stable, how do I make good prompts? I heard there's some logic behind it but idk

[cold gust]

Sure, thanks man!

[fierce pumice]

Hey - How can I add specific elements to my image like aloe vera leaves using prompts

[calm spindle]

Is there any way to prevent hires.fix from changing the image like this? Left is Highres fix.

I am using DPM++ SDE Karras as Sampling Method

[silver valley]

Sure you need to select an ESRGAN based upscaler like esrgan4x anime6b
Then set the denois to 0.25 for example.
The lower the denois the less the image get changed

[calm spindle]

Thank you!

[calm spindle]

Thank you agian @silver valley, its working flawlessly. Have been able to quickly generate them without hires to get a preview and then use hires when I am ok with it 🙂

I changed to UniPC as sampler, 4x UltraSharp with a denoise of 0.4

[silver valley]

Np, looks awesome!

[neat falcon]

Does SDXL on Clipdrop support negative prompts?

There is no separate field for it, is there perhaps some special syntax for it, or is that simply not a thing on Clipdrop?

[drowsy gazelle]

I don't think they have negative prompts

[neat falcon]

Alright. Thanks for the info!

[warm trout]

it's giving me some very bad results lately, what happened?

"doctor holding patient's hand in bed and hiding medical bill behind his own back"

all generated images are like this, no sight of the medical bill (or the doctor's back)

[quartz python]

How can I "isolate" parts of a prompt? For example if I want a picture with two cars, one red with large wheels and one blue with small wheels, how do I prevent the prompt from bleeding over?

[shut otter]

Hi, can somebody help me with this?:

Type: IMG2IMG
Image: potrait photo of me
Prompt: full body shot, bald

**Expected output: a full body shot of me bald

Actual output: the same potrait photo of me, but bald**

[vital hinge]

If you are using ComfyUI this person demonstrated their workflow to achieve something to this effect:

https://www.reddit.com/r/StableDiffusion/comments/10lzgze/i_figured_out_a_way_to_apply_different_prompts_to/

Found it while googling about your question because I was wondering something similar

[quartz python]

Thanks, I'll look into that

those examples look perfect

[slate mural]

Anyone know of a good model for styles like this? This is from reddit, don't even think it's ai but still the style I'm looking for

[drowsy gazelle]

It's an original illustration

https://www.artstation.com/artwork/VJBGdb

[obtuse torrent]

search in Civitai for Ghibli? study it is very likely that SD recognizes this study/style without a model but in Civitai there are them

[worn dome]

I'm tryng to generate an image of a Throne, but every single throne I get got a king on it

I want just the throne

[tired vigil]

so do we all ^^

[next flint]

other than the glossy hair any suggestions for making this more realistic?

[solid zinc]

How would I go about reverse engineering an image to get prompts that generate similar images?
I know about clip decoders and midjourney's "Describe" feature, but what is the best one to use with stable diffusion done locally or free online?

[silver valley]

Automatic1111 webui has a clip interrogate feature in img2img that can be used for that.

[solid zinc]

I’ve tried using it but it loads forever, and nothing happens- though everything else works perfectly

Longest I’ve let it sit for is 30 minutes- don’t know if it’s supposed to take longer

[silver valley]

You should then check the console when its running

If it runs into an error you can post a Screenshot in #🤝｜tech-support

[solid zinc]

Alright thank you man, I’ll do that next time

https://tenor.com/view/salute-good-bye-chao-see-ya-two-finger-salute-gif-26334064

[silver valley]

Np 🙂

[obtuse torrent]

add "No Humans" add "object concept" to positive .... add humans and people to negative ... try removing Portrait reemplace for full view or full shot .... you can add Throne Focus .... use object concept art or "Gold Throne concept art" ..... check effect of : removing the part that mentions Carl Karssson.....

[worn dome]

What's "object concept"?

[obtuse torrent]

Gold throne? but is a old tag in boru models maybe work ...

@worn dome Now I'm seeing the effect of some words from your original prompt,

conflicts between details and full view ... or excesive attention to "object" word and no to "white background" part .... (Txt2Img).. but you are in Img2img... (image guiance) ..@worn dome

@worn dome ok image whit metadata..

[mighty sage]

anyone know of a good Lora or just general prompting to get something simliar to this? tried many variations of hockey mask covered in goo, SD doesn't really know what a hockey mask is, or just thinks its a surgical mask

[rotund smelt]

Recommended positive/negative for prompting out hands? Found some amazing LORAs but they have some serious problems with fingers.

[obtuse torrent]

Without knowing which are those Lora that you mention, I don't understand you well, has tried with embedding / loras--- that improve the generation of hands? .... https://civitai.com/tag/negative
If you really don't want hands on composition you can try
"hands behind back", "hands in pockets", "hands behind head", maybe "hands out of frame" will work... or "No hands/Handless"

[rotund smelt]

Thank you kindly, I thought I had found all the hand negs but I had not. badhandv4 is helping tremendously. I try to keep details of the art to a minimum as it's adult-oriented and I know this is a SFW server.

[modest rover]

hello, in negative prompts, does putting brackets make a difference in the image?

like for example: (worst quality, low quality:1.4), or worst quality, low quality:1.4,

is there a difference?

[obtuse torrent]

It's just my opinion, I haven't read the theory behind it, but I think that if the model has a certain influence from models based on tag systems, then if it does have an influence, for the quality descriptors you mention it's hard to see why. crossed attention has a lot of influence, but for other tags such as excluding a specific color or a certain clothing, those changes are more clearly appreciated ...

[gusty coyote]

how would i prompt darker lines for a picture like this?

[shy sun]

May be a dumb question but Google hasn't helped. Tried Anything and a few others, but I can't get the effect I want. I want anatomy such as the nose size, eye shape, lips etc to be carried over to the anime style image so it really looks like the person and not an anime approximation. It seems the way I'm doing it, it's just the vague hair style, skin color and eye color of the original image. Any prompts to help this?

[sullen orbit]

'' , '' when writing prompts is necessary ?

[obtuse torrent]

@shy sun....really anime no is real 😄 .. and real no is anime ... You may have a more favorable result if you only apply a style that adds typical anime strokes and colors to a photograph (Img2Img)... but that has been available for a long time in photography programs and smartphones

@sullen orbit ?

[sullen orbit]

@obtuse torrent ''1woman,(dark skin:1.2),earrings,hairband,ponytail,hair ribbon,hair clip,long hair '' for example, these '' , '' are needed to separate the words, or not?

[obtuse torrent]

@sullen orbit normally if is it have less of 75 words (tokens) in** amime models** (tag sytem models ?) "," It is not "necessary" but it has consequences .... for exmaple ... compound words can be tokenized wrong... or when you go past 75 tokens A1111 will not split the prompt the same way it would if there was a "," in it.... resulting in a different relative weight assignment and that can cause some changes in the generation ---
(It's my opinion, I may be wrong, let's see if someone else contributes something)

[robust pike]

can someone help me to create a character like this in automatic1111

[errant pilot]

why is it impossible to simply prompt a servant pouring a drink into a Kings cup

it always bugs out and never works

[next flint]

could always do depth map, blender, or openpose

[drowsy obsidian]

hello, im not sure if this is the right place to ask. I don't need an answer but more of the concept so i know what the theoeretical path looks like.
I've noticed that models really influence the look of characters, specifically im using it for anime, the faces, the eyes, you can really tell what model someone is using from how the style is(at least, thats how it feels being 1 week new to this stuff). I see the model explained as the 'artist' (in really simple terms) so it makes, but what if i wanted to generate something with quite a specific style?
IF we use persona 5 as an example (this image), something super unique, really simple, if i wanted to theoeretically make something similiar, am i correct in thinking that the only possible way to do this would be to train a lora/model/something? which obv would need a lot of references and realistically might not be doable hence the thereoetical, or am i understanding this all wrong? I'm pretty sure from everything im seeing the answer is training but i just wanted to check here first before i dive into it

[next flint]

Well it would be hard to do it without a model but not impossible

do you have an example of the style or character you want to use

[drowsy obsidian]

hmm, if we stick with the example above, i wanna convert an already existing character into that style

my very noob and limited guess is its not possible but maybe im missing something obvious or smth :O

[next flint]

well you can definitely mix and match if thats what ur asking

https://modelverse.cs.cmu.edu/model/sd-dreambooth-library~persona-5-shigenori-style

[azure plinth]

Is there a current list of known artists in SDXL ?

[vocal dawn]

What Promt For Greek Statue Like this ?

[umbral veldt]

Question for you prompting wizards...
When blending a prompt, such as [ cat | dog ] , does it work the same to add more words to that blend, such as [ cat | dog | chicken ] ?

[obtuse torrent]

yes

and if you need 75% cat and 25% dog ---> [cat|cat|cat|dog] ... you got the idea?

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#alternating-words @umbral veldt

[obtuse torrent]

Greek Statue ?

Greek Statue ? , "[god name] god/goddess of [god/goddes ambit description]"

[umbral veldt]

So you couldn't weight words like this to get the same effect? [(cat:0.75)|(dog:0.25)]

[obtuse torrent]

IDK..
test it,
I have never done it ,

@umbral veldt
The way I see it, that way of writing it maintains the same relative weight between cat and dog... but it would make both lose relative weight against the rest of the prompt... test is necessary ..

[north canyon]

Hello everyone, i have a stupid question, every time i put i color inside my prompt, i always see that a lot of other elements inside the producted image will be of that color, there's some way to prevent this?

[obtuse torrent]

You can add extensions like controlnet or regional prompter, which help to subdivide the composition or guide the generation in a better way, but you must familiarize yourself with them to take advantage of them...@north canyon
https://github.com/Mikubill/sd-webui-controlnet.git
https://github.com/hako-mikan/sd-webui-regional-prompter#2d-region-assignment-experimental-function

[drowsy gazelle]

How do you combine two words together if you don't have access to the [dog|cat] format?

Is there another way to do it, like dog AND cat?

[obtuse torrent]

@drowsy gazelle .... the [dog|cat] format? .... is a feature of A1111... webUI... it.. alternate 2 (or more) words in each step of generation ..... IDK about others UIs ...

[drowsy gazelle]

Yeah I'm using another version that doesn't have it

[hoary mantle]

sorry chaps, how do you actually ad negative prompts?

[next flint]

like in a1111?

[keen epoch]

Whenever I try to generate a character in 1960x1080 (landscape) it generates multiple of the same object twice or thrice, sometimes merging the hands together. How do I make it so it only generates one of the character and fills the rest with background?

[ember oracle]

I would i prompt this kind of pose

?

[obtuse torrent]

This is something common when it comes to directly generating this type of resolution.
I recommend generating an initial image at a low resolution and then achieving that with tools like Hires.fix and/or upscalers
For example:
653x360 initial
Hires.fix (x2)

• another upscaler (x1.5)
  or any other similar combination
  maybe you can generate a:
  980 x 540 + Hires.fix (x2)
  ... Add "clone" "cloned" "Duplicate" "jpeg artifacts" .. to negative

[keen epoch]

Thanks. I messed around with some hires.fix settings on automatic1111. What should my hires steps and denoising strengths be set to ideally?

[obtuse torrent]

if you set steps at 0 .....it is normally is fine (repeat the same value that you have for the normal generation) ...D.S. depends a bit on the scaling method, normally 0-51 is possible...
05-30 if your base builds are fine and don't need much change
30-51 if you need to give the AI ​​more freedom to alter what is generated at the beginning....

[keen epoch]

Alright. Thanks

[obtuse torrent]

for anime models? it is tag "upside-down" ... "lying", "on_back", "legs_up" and you need a lucky seed ||to no get a NSFW pose||

[keen epoch]

Denoising strength can only be 0 to 1 in webui. Only setting it to 1 gives me a clear image, 0-0.5 generated a blurry image. I set scaling to the same as normal

[obtuse torrent]

My mistake, when I told you numbers, I was talking percentages... I assumed you knew that the D.S. goes from 0 to 1 (0% to 100%)

@keen epoch

[sudden quiver]

Can someone help me with getting a samurai cowboy witb a oni mask holding a lever action, I'm only able to use this servers bot

[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

[solar dust]

Why is that option not available in img2img?

[blissful surge]

is there a way to img2img this image to both remove the white background and add a landscape as well as make the actual character look different/more epic? im using automatic1111 if that helps

[worn dome]

As far as I'm concerned Stable Difussion doesn't work with Alfa channel, but you can get it into Photoshop and use this tool, then work the selection and get the character over a transparent bg, then generate your bg separately in Stable Diffusion and blend the two images together in Photoshop

I'm trying to generate some images of a royal palace background, I don't need something specifically, but I could do with some prompting help cos I'm not sure what should I write down

I've got here some reference images, but I don't need the bg to be exactly like them

[wicked fossil]

Hello! I'm not really sure if this is a right place to ask, so sorry..

I'm looking for a way to restyle an image. I made a 3D sketch with some character sitting in a car, and I want to restyle it using one of the models I have. I tried to use img2img, but it does not save the structure of the shot. Character sometimes in wrong pose or car is wrong e.t.c. Is there a good tutorial on restyling images. All I saw on YT are for photos and people, but some shots I need are without characters at all..

I'm using local web-UI (Automation111 or something like this, do not remember)..

[orchid ore]

if i understand properly, but i doubt, i would use img2img with about 0,3 denoise strenght. Default is too strong, if remember properly about 0,7 probably. @wicked fossil

[wicked fossil]

Oh, yes, it does help, thank you! Weird parameter name for a beginner user 🙂

[orchid ore]

glad it helps

[haughty island]

Hi guys, I'm new to SD. The faces on my characters look off and I don't know how to fix it, anybody that could help?

What am I doing wrong?

This is kind of the result I was going for:

(face-wise)

[minor remnant]

prompting-help

#📝｜prompting-help

cloudless sky

#cloudless sky

[slow ibex]

What its suppose to look like. what i got.

How do I stop this two tone effect?

[willow mural]

can anyone suggest an extension for a1111 to automatically run a prompt through different checkpoints and/or different loras?

[obtuse torrent]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#xyz-plot Feature in Txt2Img

[silver valley]

Thats already in build in auto1111 called X/Y/Z prompt. At the bottom in txt2img

[obtuse torrent]

What type of prompt should be used depends on the model, when the basses have a presentation page where they recommend the settings, and other things... depending on whether the model understands the concepts or not, phrases like "look wise" "wise gaze" can help you or tags like "tsurime" or emotions/personality traits like "cool", "confident", "self-reliant"..you can also try other models or options hires.fix or scripts for better details

[slow ibex]

Almost there, just need to get rid of the finger feet.

this was after 30 steps

I'm gonna give it some more steps to work on. Like 40.

[slow ibex]

I can see it trying. Gonna try adjusting the prompt to fix the fingers and feet

[vapid crater]

my stable diffusion is behaving really weird and I need help with my prompts.
My prompt was: A magnet shaped like a U, 2d game asset
Negative prmopt: Detailed background
and this is what I got

[tribal talon]

Why the heck does this keep happening!

It gets even worse when I upscale, these random green blurs on certain parts

[silver valley]

Thats mostly if your model is missing a vae

[tribal talon]

But I'm not D:

[vapid crater]

oof is my question getting away now 😦

[silver valley]

Which one did you use?

[tribal talon]

clear VAE

[silver valley]

What model do you use ?

[tribal talon]

It happens on any model I use, but this specific one is azovya RPG artist tools v3

[vapid crater]

I'm mainly using stable diffusion for game assets and it has worked better before

[silver valley]

Try an other vae like 84000-mse

Clear vae is for anime models

[tribal talon]

It does it even with anime models though

[silver valley]

Can you show me an example?

[vapid crater]

here is a better answer:
A magnet shaped like a U, 2d game asset
Negative prompt: detailed background
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 2350699442, Size: 512x512, Model hash: cc6cb27103, Model: model, Version: v1.4.0

[silver valley]

You should try an other model

[vapid crater]

which model?

[silver valley]

Maybe this one:
https://civitai.com/models/47800/game-icon-institutemode

[vapid crater]

okay ty

now I just need to watch a tutorial of how to install it again lul

[silver valley]

You download the model. And put it into the models/stable-diffusion folder

Thats it 🙂

[vapid crater]

oh xD

[silver valley]

Then you can select it in the dropdown

[vapid crater]

just in here?

[tribal talon]

I can't find one of the examples! But I have seen it! Is the VAE thing the only possibility? If so I can troubleshoot from there on my own

[silver valley]

@vapid crater In the stable-diffusion folder

[silver valley]

Yes

[vapid crater]

mine is called stable-diffusion-webui 🤔

[tribal talon]

Alright, ty, I can likely find a fix on my own with that.

[silver valley]

For photorealistic models you need a vae like 84000 mse and for anime for example the kl-f8-anime2 vae

[orchid ore]

@vapid crater fourth from bottom folder

[vapid crater]

OH

[silver valley]

Stable-diffusion-webui/models/stable-diffusion

[vapid crater]

ah the one I already have is there

ty!

what a weird file extension name gameIconInstitute_v30.safetensors

[silver valley]

Yea .Safetensor are the new file extensions for models(checkpoints)

[vapid crater]

ah

[silver valley]

Make sure they are over 1.99gb in size if you Download more. Everything below dont go in this folder

[vapid crater]

what's different between models? Is it the data they have or the algorithm?

yeah that one I just downloaded is 3.97GB

[silver valley]

The models contain different training data. The model you had used before is the official SD 1.5 version.
The Community models are better trained or specificly trained on subjects and are based on the 1.5 model

[vapid crater]

aha

so if I want a model for game assets I should look for a model with lots of images related to game assets in its data

cuz they are trained on a bunch of images right?

[silver valley]

Yes the one i linked is such a game asset model

[vapid crater]

aha

do I need to type game asset in my prompt tho?

sometimes it helps but if the model is already trained for game assets it should be useless right?

[silver valley]

Yea i would try game icon, game asset, 2d, icon, U shaped magnet:1.2,

[vapid crater]

what does :1.2 do?

[silver valley]

It gives the word more attention

[vapid crater]

wow that's very cool

[silver valley]

You can use (word) or word:1.1

[vapid crater]

I also tried to use this as an image prompt but it was too similar even though Denoising strength was very low

wait I meant CFG scale

[silver valley]

Denois has to be high

0.8-1

[vapid crater]

what if I want to have a slight 3D shape on it?

but still from a platformer angle

and smoother curves

[silver valley]

It should generate it more 3d then 2d

I cant try it out. Not at the pc rn, but later

[vapid crater]

wtf I got this

game icon, game asset, 2d, icon, U shaped magnet:1.2,
Negative prompt: detailed background
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1140004488, Size: 512x512, Model hash: cc6cb27103, Model: model, Denoising strength: 1, Version: v1.4.0

with the image prompt

my stable diffusion isn't making any sense

[silver valley]

You didnt switched the model

The model hash is still the same

[vapid crater]

uh I thought i did but apparently not

game icon, game asset, 2d, icon, U shaped magnet:1.2,
Negative prompt: detailed background
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1388204124, Size: 512x512, Model hash: c112297163, Model: gameIconInstitute_v30, Denoising strength: 0.62, Version: v1.4.0

BRUH

wait I copied the wrong info it used this: Negative prompt: detailed background
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 4265811788, Size: 512x512, Model hash: c112297163, Model: gameIconInstitute_v30, Denoising strength: 0.9, Version: v1.4.0

my image prompt seems to have done 0 effect

[silver valley]

Okay what if you try the prompt in txt2img ?

[vapid crater]

game icon, game asset, 2d, icon, U shaped magnet:1.2,
Negative prompt: detailed background
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1206477096, Size: 512x512, Model hash: c112297163, Model: gameIconInstitute_v30, Version: v1.4.0

this is text2img

not even a u shape

not even 2d

[orchid ore]

i made this but with CN. Which means download tons of models.

[vapid crater]

it looks a lot like my image prompt tho

[orchid ore]

it is your image as source

[silver valley]

Seems like the model is more for 3d assets

[vapid crater]

oh oof

[silver valley]

Can you try.
Two pole magnet, red and white. U shaped, 2d, comic style

[vapid crater]

I'm looking to get something like this but I took this straight from google

I thought the AI could achieve that with denoising strength adjustments

why else would you use your image prompt if it gives the same results

[silver valley]

Yea normaly it can do that

I will test it later

[vapid crater]

okay

[orchid ore]

maybe horse shoe would be better than U shape

[vapid crater]

xD

I get your point but that was a funny similarity

[vapid crater]

it seems like it only understands "red white"

Two pole magnet, red and white. U shaped, 2d, comic style
Negative prompt: detailed background
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1858827973, Size: 512x512, Model hash: c112297163, Model: gameIconInstitute_v30, Version: v1.4.0

that was text2img

[silver valley]

hmmm okay

[vapid crater]

sometimes I feel like drawing my own stuff is easier than using AI 😂

If you want some context, this is all the art I've managed to make with AI so far but now I can't even make a fucking magnet! 😡

Is this a joke? I got this by using the prompt "Magnet"

[tribal talon]

Draw a magnet and put it in img2img maybe

[vapid crater]

I did

[tribal talon]

It's thinking fridge magnet then

[vapid crater]

I used this as my prompt

and this is the best result I could get

maybe this model has zero images of U-shaped magnets

[tribal talon]

Make a magnet lora :D

Show it the ways of the magnet

[vapid crater]

what's lora?

[tribal talon]

It's like a model but instead of being an entire thing of art it's taught a specific concept, for example you can teach it how to recreate specific characters, specific clothing, hairstyles, the list goes on

[vapid crater]

you mean in here?

[tribal talon]

No, tbh I've only dipped my toe into loras, it's not made in the UI, you make it elsewhere, I am not the most knowledgeable

[vapid crater]

oh

[tribal talon]

Well into making them. I'm very experienced using them!

[silver valley]

Yea its something that a model may didnt know

Maybe not even the base 1.5

Or maybe with an other word

[vapid crater]

but it seems like I need a different model though, any tips for similar stuff like this? It's 2D but still slightly 3D

[tribal talon]

Maybe some kind of tech lora knows how to make a magnet, but that would be a shot in the dark

[vapid crater]

oof

[silver valley]

Hmm maybe an anime model or toonyou

[vapid crater]

I can at least be proud of being the first to encounter a problem in 2023

cuz you are almost never the first 😂

[tribal talon]

I encountered a similar problem. I can never make one of my characters because she has a veil covering her eyes, and no models have been trained on eye veils

[vapid crater]

oh lol

[vapid crater]

hmm ok

I guess I will just make my own in blender with an almost 2D camera angle and then just ask the AI for textures

cuz I want it to look old and rusty

[vapid crater]

"Something made of Iron shaped like a U"

bruh

[stable moss]

Guys I'm trying to generate a set of photo realistic character images in stable diffusion using Random names in my prompt. I'm planning to make a data set out of this results and do a dreambooth training.
Right now I'm struggling to generate the character in all angles like front,side,3/4 views, and back. I'm not getting similar faces in each angle.
I'm using epic realism for this. Any lead or help would be appreciated.

[orchid ore]

@vapid crater try shape of letter C

[vapid crater]

the fact it can't even understand shape U is surpricing

[slow ibex]

mine are purple on one of my checkpoint models

[slow ibex]

You might need to be more specific on that one

[silver valley]

You then need a vae file for the model

[solar dust]

https://cdn.discordapp.com/attachments/1011743094309396631/1127403236756824175/image.png

Why is that option not available in img2img?

[silver valley]

In img2img you need to use the sd upscale script

Its nearly the same

[tired vigil]

anyone can help me to make better this photo

here

1girl,armlet ,bangs ,bikini ,purple bikini ,blunt bangs,bracelet ,braid ,braided ponytail ,breasts ,earrings ,eyebrows visible through hair ,flower ,flower bracelet ,flower on liquid ,hair flower ,hair ornament ,hairclip ,highres ,jewelry,large breasts ,long hair ,navel ,necklace ,parted lips,partially submerged ,purple flower ,purple hair ,purple nails,side-tie bikini ,sitting ,solo,stomach ,swimsuit ,tassel,tassel earrings,thigh strap ,violet eyes ,water ,wet ,white flower,raiden_shogun,aesthetic,more flowers,(exceptional, best aesthetic, new, newest, best quality, masterpiece, extremely detailed, anime, waifu:1.2)
<lora:raiden shogun_LoRA:1>

1girl,armlet ,bangs ,bikini ,purple bikini ,blunt bangs,bracelet ,braid ,braided ponytail ,breasts ,earrings ,eyebrows visible through hair ,flower ,flower bracelet ,flower on liquid ,hair flower ,hair ornament ,hairclip ,highres ,jewelry,large breasts ,long hair ,navel ,necklace ,parted lips,partially submerged ,purple flower ,purple hair ,purple nails,side-tie bikini ,sitting ,solo,stomach ,swimsuit ,tassel,tassel earrings,thigh strap ,violet eyes ,water ,wet ,white flower,raiden_shogun,aesthetic,more flowers,(exceptional, best aesthetic, new, newest, best quality, masterpiece, extremely detailed, anime, waifu:1.2)
<lora:raiden shogun_LoRA:1>

i am new to this thing and peoples are doing much better things than me

just asked for can i make it better

ohh i copied wrong thing

lowres, ((bad anatomy)), ((bad hands)), missing finger, extra digits, fewer digits, blurry, ((mutated hands and fingers)), (poorly drawn face), ((mutation)), ((deformed face)), (ugly), ((bad proportions)), ((extra limbs)), extra face, (double head), (extra head), ((extra feet)), monster, logo, cropped, worst quality, jpeg, humpbacked, long body, long neck, ((jpeg artifacts)), deleted, old, oldest, ((censored)), ((bad aesthetic)), (mosaic censoring, bar censor, blur censor)

here

yeah

https://huggingface.co/YoungMasterFromSect/Trauter_LoRAs/blob/main/LoRA/Genshin-Impact/Raiden Shogun/raiden shogun_LoRA.safetensors

https://civitai.com/models/4829/raiden-shogun-lora-collection-of-trauters

np np

oh i saw them

and any guide for imprinting

i will

yep

inpainting

its works weird

doesnt fix limbs

ohh

last question

why my pc crashes

i have rx 6600xt

and randomly crashes when using sd

for temps

yep

so i need to wait some time

low res?

which gpu you have

ohh

you can do sd with laptop

i want to buy a laptop but i am scared for sd to run poorly

it is hot?

laptop model?

i am planing to buy more budgets

like asus tuf

or hp victus

people dont reccomend that one usually

ohh i found problem probably

i use 512x512

than upscaling it to 2k

i dunno

from highres fix

selecting animeb6

something like that

it just works

in same res

why am i getting out of memory error

i hate this

8

changed sampling steps and res

experimanting

but nothing changed

oh my

this looks good

in image

difference is so much

do you make it with control net right?

still out of memory error

with 8 gb vram...

i need absoulty an nvidia gpu

same res

1280 720

changed nothing

--xformers doesnt work on amd

yeah

i entered some commands

export COMMANDLINE_ARGS="--lowvram --no-half-vae --opt-sub-quad-attention --opt-split-attention-v1 --autolaunch"

low

okey

thanks anyway

i am going to buy an 3060 laptop because of you

i am going uni next year

so i need a laptop

i am in other country saddly...

some people said 6 gb vram isnt enough

but you use it pretty smooth i guess

i learned my lesson

amd is not for sd

absoultly

it is pain

i have a lot of problam installing on linux

and then

this memory issue

[vapid crater]

I gave up on SD and made my own magnet lol

I thought it could give me more details and more interesting results

but I do need better textures

maybe I should try to use this as an img2img prompt

never heard of that

🤔

[tired vigil]

2070 super or 3060

which one is better for sd

2070 is waaaay cheaper in my country

and i can trade my rx 6600 xt with that

i guess 2070 is way way way better than rx 6600xt

i have some graph about it

if we say

3050 is on par with 2070 on ai

stil better than my gpu

but 3050 is 50 class gpu

2070 is 70

probably 2070 is more powerful at ai

[slow ibex]

It would be awesome to see Apple silcon on a list like this

[tired vigil]

you are right...

intel is that bad

so i need mux

asus tuf has mux switch

in budget class

its pretty neat

yep

i will do that

but i certainly buy an nvidia gpu

stable diffusion is so much fun

[vapid crater]

what are those Stock texts tho?

interesting

oh you used that google image as a prompt xD

I was also hoping I would get an image faster by using SD instead of making my own but it was faster and easier by making my own in blender xD

it looks like a helmet xD

magnet helmet 2023

lol

[tired vigil]

Hi, how do I stop these weird blurry artifacts appearing when I do small-scale inpainting?

(I have whole picture checked, for the record)

[dreamy rover]

the picture is blurry on a lot of spots... do you mean blurry spots where your mask ends? i think this can happen if you have the "Mask blur" option set too high

[dreamy rover]

If the "Mask blur" is not the thing you are looking for and instead your problem is that the newly drawn skin or background is too blurry, maybe try "highly detailed skin texture", "sharp", "natural skin". Common negative prompts to prevent blurry skin: "airbrushed" "cgi" and sometimes "doll"

Does someone have an idea how i might get this pose (not exactly, lookalike is fine)? I´m already struggling to control her fingers a little bit...

[solar dust]

Do generated images have any kind of metadata so I can use it to know which prompts were used to make it?

[obtuse torrent]

This is not always the case...often there are options that allow you to determine what generation data is added to the metadata...it is not uncommon that some authors choose not to add them or only a few of them, also and cases that use an external editing program alter or delete them.

[slow ibex]

I've successfully been able to use a LoRA

[solar dust]

So you're saying that, by default, they're not added?

[obtuse torrent]

I don't remember which is the default setting.. also A1111 has had many updates.. and not all people generate with A1111 ........ if you use A1111 you can see that in settings

[solar dust]

Thanks

btw

is it true that I can use SD without an internet connection?

[potent saddle]

hii!!! i need help, please. I want to make a Spanish-style card but changing card number 12 (which is a king) for that of a gaucho

[native thorn]

I'm using the stability api service and struggling to get results that reflect my prompt. I've perhaps handicapped myself as I'm looking to generate some 300 images with a common style. In this case I've selected 'water colours'. Here is an example prompt: a watercolour depicting a user zooming a picture on a mobile phone

I should note that I'm also forcing the aspect ratio to create a 'swatch'.

[native thorn]

So I'm trying to generate an image of an orphaned child (to represent software that is no longer supported) 'a watercolour depicting a lost child' and I get the reponse - Invalid prompts detected - so I'm guessing more overly zealous censhorship

[dim pelican]

hey
What prompt should i use to create a image from a reference image
Please help

[sly fable]

Hello, I'm new to Stable Diffusion. I used a lot Midjourney and loved it. I saw Stable Diffusion and wanted to try because it's "unlimited to use". Do you know why is it soo bad with things like characters or stuff from pop cultur thing. I tryed to do some Mass Effect and lol characters and it never worked not even close. Even with the image to image. Maybe I'm just using it wrong so I came here to find if you can give me any advise for this

[silver valley]

A prompt that describes the input image

[silver valley]

Hey do you use a webui for SD ?

The base model of SD cant create every character from games or movies. So people started to train their own files to get specific characters. These trained files are called Loras. And they can be loaded together with a model to get a good image of the Character.

[sly fable]

Yes the Automatic11111 thing like this

[sly fable]

Do you know where I can find this ?

[silver valley]

Sure you can find different models and loras on Civitai.com
Make sure to place models(checkpoints) in the models/stable-diffusion folder and Loras in models/loras

[sand smelt]

now I'm wondering how MJ can get so many pop-culture characters (relatively) right. Are they all trained in the model itself or does their process pull up LORAs / TIs in the background, in response to prompt keywords?

[sly fable]

Thank you so much !

[silver valley]

I think they mostly mix stuff in the background to get the stuff you want.

[silver valley]

Np, i would suggest to first try different models. Maybe some of them know more characters. If not then search for loras of the chars.
Here is a guide on how to use Loras:
#🤝｜tech-support message

[arctic tapir]

how can you prevent more than one character appeared

[silver valley]

Using a smaller resolution like 768x512 and then upscale it instead og using higher resolution at the start

[arctic tapir]

hmm, does using a 1:1 also keeps prevent the double character?

like, 1024x1024

[silver valley]

More like 512x768 or 768x1024 as its a Portrait format

[arctic tapir]

ahh 2:3

[wanton girder]

You can fuse two thing together with the | symbol between them like this cat|fox you make this symbol when you press "alt gr" + the button left from the "Y"

[muted dragon]

hey guys, as anyone found a way to replicate those "mini-x" kind of setup and looks almost as you are transposing what you want into a diorama sort of thing? I've seen it often done with videogames like "mini-thelastofus", "mini-cyberpunk" and they look as if they all were isometric in nature

would be really cool to build upon it

[fallen grail]

is there a way to generate a prompt x amount of time before it randomize? i use wildcards and dynamic prompt

[solar dust]

Is it true that my SD is not using the internet when running?

[lament crescent]

A challenging prompt no ai has yet come close to mastering: A bear with six eyes

Since apparently no person has ever drawn this it is impossible for any system to imagine it. If anyone can get SD to comply I'll be very impressed.

[cosmic linden]

I need some help, I'm not that experienced with stable (or other Ai generators) but I'm getting a problem that i not faced before
in preview my arts look kinda fine but when they finish they look kinda weird like that

should I up the sampling steps? I'm using 60 rn

[rich crow]

are you using any loras? alsoi whats your cfg at?

[cosmic linden]

I'm using 2 loras rn, lora:pastelMixStylizedAnime and yaohu

my cgf is at 8

[rich crow]

what are the lora weights at? ive seen this happen with multiple loras in the past, they usually dont work well together at higher weights

id try lowering the lora weights and see if that helps if theyre both at :1 or something

8 cfg is fine btw

[cosmic linden]

I'll try lowering a bit, maybe its a lora conflict

I removed all loras and tryied with 100 steps, got a bit better result, I'll try to insert the loras slowly and get the results that I want, any suggestion in improving the quality of the details in the image? still looks kinda unfocused to me

[obtuse torrent]

Lora style 0.2 - 0.5 (you can increase it more if the style does not show well)
.. character Lora 0.5 - 0.9 (in this range the character should be shown relatively well and leave room for the other Lora)
the Sampling steps depend of the Sampling method you use, but in general there is little benefit to going beyond 50.
to test the image of: 10-20.
to generate with a good seed of: 24-50

[cosmic linden]

thanks for the advice, I lowered the style lora and got some pettry good results, I'll be more mindful of the sampling steps but I'm getting some consistent results with 60-70

[obtuse torrent]

@cosmic linden ok......pastelmix recomendations (uploader)...
Guide
For the settings or parameters, I recommend using these settings.
Sampler: DPM++ 2M Karras
Steps: 20
CFG Scale: 7
Hires. Fix: On
Upscaler: Latent (MUST!)
Hires Steps: 20
Denoising Strength: 0.6 ()
() I prefer using 0.6 since it's the sweet spot of this model. If you can find a better setting for this model, then good for you lol.
Latent upscaler is the best setting for me since it retains or enhances the pastel style. Other upscalers like Lanczos or Anime6B tends to smoothen them out, removing the pastel-like brushwork.
------- HF bye

[cosmic linden]

[obtuse torrent]

@cosmic linden normally on their presentation page the models bring a brief description and recommendations for use, and sometimes known problems are mentioned... it is always good to review them well... and if you have more time read questions asked by other users and the answers of the authors..

[cosmic linden]

I generally read it and look into other generations to have ideias of what people are using, I just messed with the steps but I overlooked the Latent upscaler, I'll look into it

I'm not very versed on all of this so thanks for the patience

[obtuse torrent]

NP, sooner rather than later there will be topics in which you have more experience and you will be advising me... besides, the fun is in experimenting and more than once trying something not recommended will give you better results

[obtuse torrent]

?? https://civitai.com/models/83923/little-room

[prisma raft]

there is something very refreshing about uncharted territory

where everyone can figure something out for the first time

i've learned a lot from noobs experimenting (and really am still myself a noob experimenting) and thats super super cool

[sly fable]

How can I force stable diffusion to use one type of color ?

Like I say red and that came blue 💀

And also, I'm interested to use controle net. If you know how can I do it with 1.5 and the webui

[ivory brook]

hey guys, i want to try and paint rest of the face with inpaint tool, is it possible? if its not, what extension can do it?

i just get stripes and thats it

[silver valley]

for example blonde hair:1.3, green shoes:1.3

[worn dome]

I've got a ship inside the black area and I don't want it to be there. I'm trying to erase with inpaint and I dont get it

Prompt: sea
Negative prompt: ship, blurred, bad quality

Someone could help me?

[silver valley]

what are your img2img settings

[worn dome]

Changing denoising strength I got a better result, but still a bit rought difference in the colour of the sea

[silver valley]

can you give me the image? ill try it

[silver valley]

hmm thats pretty hard. SD always tries to add something

you woule be faster by editing it with a editing tool like Pain.NET

[worn dome]

what's that?

[silver valley]

free image editing tool, not as good as photoshop but its okay

[worn dome]

I've got Photoshop

It´s sometimes helpful with some tricky edits SD is probably not prepared to do yet

[silver valley]

the best i got with 1.5 inpainting model and denois 0.8

[worn dome]

are there models to inpaint?

Is that ControlNet Inpaint?

[silver valley]

for inpaint you should use specific inpainting models

[fast socket]

nyway to use stable diffusion more as a filter than generation? As in, generating art in a certain style and wanting to convert another image to that style as well?

[exotic robin]

use the content aware fill or generative fill in photoshop, those are great for removing small unwanted objects like that ship

[worn dome]

content aware fill is in beta?

[exotic robin]

Yes

[worn dome]

cos I don't have Photoshop Beta

[exotic robin]

Content aware fill is in the regular photoshop, Generative fill is beta

[worn dome]

I had, but the free trial expires

[worn dome]

Perfect!

[snow musk]

im trying to make a character lay down, but all im getting is a pose with her rm crossed, i have a prompt lay on back but it aint working

[upper lotus]

H

[fierce dove]

(Img2Img) Any tips for prompting a 2D cartoon image towards being something more like a 3D CGI/Photorealistic image? I’m close but not quite there yet.

[sly fable]

Ok thx

[cerulean heath]

Hello guys.. playing ewith the deform and somehow, the images get very fast abstract.. and it's not part of my prompt.. smth in the setting.. please any help ❤️

[shell sleet]

what is the best sampling method for anime?

[tired vigil]

I want to create a small cute shibi, with blonde hair and blue eyes, for the backround juste a beautiful color.
But i dont know exactly how to do it with speacial prompt any experts here ?

[exotic robin]

lower the denoise strength to like 0.3 or 0.4

[cerulean heath]

this one?

[obtuse torrent]

there are models (checkpoint) that facilitate this, and models specialized in them (LoRA, Lyco), look for them in civitai, by the way check the prompts they use... there is always a possibility that other models somehow understand the concept and generate it ... https://civitai.com/tag/chibi

[fast socket]

Anyway to use stable diffusion more as a filter than generation? As in, generating art in a certain style and wanting to convert another image to that style as well?

[solar dust]

And what do you call the area beyond full shot?

[wanton girder]

🤷

[obtuse torrent]

@wanton girder
The following terms may not work on all models or be understood in the same way.... if there is some realism and photography in them: Wide open view, Extreme wide open view, panoramic, scenery, etc..

[wanton girder]

👍

@obtuse torrent

[sly fable]

How can I tell my AI ton focus on one thing ? Like I want a guy with a helmet that look like a camera, but never something near a camera came out

[wanton girder]

to fuse two things use the | symbol you can make it with alt gr+ the button left from Y than use it like this helmet|camera should it be more an camera not an helmet use weights () they tell the ai to generate from this word more like this helmet|(camera) , the ai dont understand context but words. enjoy @sly fable

text is hard, has anyone good tips ?

[sly fable]

Oooooh ok thank you !

[wanton girder]

@sly fable ah i forgot , to make weight heavier just add more () to an word like cat,(cat),((cat)) you only need one of them and dont use the same word over and over

no problem

[sly fable]

Ok tysm !

[wanton girder]

and use nagative promts too, so you can avoid mixing with other data that has same vibe but totally different content,

[sly fable]

I will !

[wanton girder]

as an example when you want an real photo, just tipe in negative promt waht you dont want like ,3d,render,oil painting, and so on

@sly fable

[sly fable]

Ok ok ty ! You help me a lot

[wanton girder]

its a little sad that here is not an guide with all the promt tricks and good promt words to get what we want.

[sly fable]

It is !

[rich crow]

thats for you and me to make, stabilityAI just trained the model and let us play with it. I'm sure theres guides out there in the internet if you checked, or you could write your own 😉

[twilit arrow]

Is this the right place to ask for prompt tips?
If so ... how do I keep all characters in my renders from having the same expression?
For example, (one woman smiling) and (three men frowning) can have them all smiling or frowning...

[rich crow]

#1100170312106127410 can do that for you

[twilit arrow]

How? Oh you mean the damn unicorn thing lol

If anyone has any prompting-help advice on how to have multiple characters with different expressions, I'd be way grateful

[dusk harbor]

what would be a good prompt to turn my photo into an 80s anime

[solar dust]

I struggle to put multiple characters with different outfits in an image.

For example, I want the AI to draw a man and a woman sitting and chatting with each other in a cafeteria, both wearing different outfits. How the hell am I supposed to describe that to SD without it mixing things up?

[dusky raven]

Hi! I have a question. I would like to know if it's possible to use stable diffusion in this way: I upload my own sketch (png) and then make it inot a rendered image. For now I found a lot of tutorials regarding using the sketching option within the stable diffusion program but I don't know if my png sketch that I uploaded is really used by the model or if it's just not visible for it. Can anyone help me?

[craggy pulsar]

What is the best input format to make sure my prompt is understood? For example, if i were to use piercing blue eyes as a prompt how do i avoid it being understood as "piercing blue" eyes?

[rich crow]

you could try using your sketch as an image with controlnet. check out the gallery for examples https://github.com/Mikubill/sd-webui-controlnet

[heady moon]

Can some1 explain me what exacily "BREAK" doing in prompt?

[orchid ore]

In Regional prompting are keywords like ADDCOL, ADDROW, ADDCOM turned to word break in picture name.
But in A1111 here is the answer
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#break-keyword

[dusky raven]

thanks, I'll try it!

[delicate drum]

there is a MJ style I like but fail to reproduce. any idea to get those brush strokes in sky?

Van goghs a starry night over a huge cyberpunk city, recursive network graphs within neural networks made of stars and galaxies in space, moebius + pascal campion, incredibly detailed + sharpen + professional lighting, cinematic, awe inspiring --ar 16:9

[kindred garnet]

any tips for sdxl prompting, i dont get good results

not completely bad but compared to others not really good

i struggle with extremly blurry results

[fast socket]

Anyway to use stable diffusion more as a filter than generation? As in, generating art in a certain style and wanting to convert another image to that style as well?

[hidden mulch]

Hi, is there a way to try different models in the bot rooms? Not sure how to point to different models. Any tips?

[wary basin]

Hey there SD community. I had a hopefully somewhat quick question for you I'm guessing one of you can answer without an issue!
I was wondering if you guys know a way to put a generated image from SD, infront of a pure white background. An example is below. I am able to do this right now by first uploading a pure white background into inpaint, mask an area, and then generate an image. Which works well enough, but if I want to use the shape of a specific image and get control net involved, the quality takes a big hit, and then the pure white background disappears. Do any of you know a way I can do this in a much more straightforward fashion? Any help would be great. An example is below. If it helps, I already have rembg installed and can remove backgrounds from generated images. Thanks!

[brazen valley]

how do i do weighted promts and other things

is there a wiki for it

?

im using sdxl bot

[hidden mulch]

Hi there, im curious how you make animals fur change colors in a prompt? I am having alot of trouble.

[sage moss]

how do i make my results less creepy

my prompts are 1girl, portrait, full body, brown eyes, extremely detailed eyes, standing:1.2, oversized (t-shirt), (black shorts), platform boots, monotone background, long (black hair), cute face, style of Masashi Kishimoto and for the negatives ((blurry)), nsfw, out of focus, large breasts:1.1, 1boy, (deformed:1.3), poorly drawn face, poorly drawn eyes, lowres, ugly, mutation, mutated, worst quality, bad quality, twins, 2girls, 3girls, 4girls, multiple people, 2boys, 3boys, 4boys,

i don't like how it looks

i want it more anime like

i'm also using sd webui if that's important

[blissful schooner]

Anyone knows how I can mix multiple faces into a single individual? I dont want to get multiple characters to appear in the image.

[dim pelican]

how can i share a reference image for SD to dream it?

anyone there?

Please DM me the answer in case i dont reply its very much useful for me

[wanton girder]

add blurry as an negative promt and high detail,sharp, to the positive ones ?

[wanton girder]

use the fuse promt, i has explained it 2 days ago #📝｜prompting-help message

[kindred garnet]

tried allready 😅 , i kinda gave up with sdxl for now

[wanton girder]

@kindred garnet can you give me your promt and tell me what you wanted to be generate ,eventually can i find out the problem

[echo bramble]

Hello, what's a smart way to find if the style of an artis (that I might use in my prompt) is actually recognised by Stable Diffusion ? Is there a test you can think of that I could use ? Maybe trying to generate the artist's work (e.g. a vase with sunflowers by Vincent Van Gogh) ?

[silver valley]

Your model needs a vae file for color correction. And you should try a resolution of 512x768

[solar dust]

Looks pretty anime-like to me.

And speaking of VAE

@silver valley Are those good enough?

I don't understand much about VAE, my friends told me to get these.

[silver valley]

You should get the kl-f8-anime2 vae too. Its very good for anime

[solar dust]

Could you please give me the link?

Also does SD switch vaes automatically?

[hard gate]

Can this discord server support img2img input?

[hard elm]

Please see the pinned comments in the bot sections for all information on SDXL. You can also see #📣｜announcements for the latest release information

[young trail]

Hopefully this question is pretty basic, I'm a newer to the AI space so I am still trying to experiment with tools to understand more about SD. That being said does anyone know what I would need to do to make two more variations of this image with the head in a different position? I'm going for King Ghidorah

[young trail]

From what I have been reading that seems like the way to go, I'm not familiar with Controlnet either. Does anyone have any experience doing something similar to this? Controlnet settings that you might use to adjust this image to a different pose?

[rich crow]

I'm sure someone might have a better method but the best I can think of is to controlnet tile that and text2img until you get a different pose. Usually you can use openpose to change poses but that's reserved for humanoid figures I think. Mabey openpose face could help but unsure, could give that a few generations to see if it works.

An alternative I would try if I were to do what you're trying to do is edit the image around and flip/tilt it to whatever pose you want it to be in then img2img that or controlnet tile that.

[young trail]

That's a great idea! that would definitely help with adding some variation to the heads, It's always the simplest solutions haha

[wanton girder]

you need canvas inpainting and control net so you need an amatuer freundli tool that can use free models is bug free and has its own discord to help.

you need InvokeAI first search for their discord.
Its an software like paint but very simple and easy to use in there discord are many people who can explain to you how to do it.

[tight light]

anyone got any suggested prompts for making sure head, torse, legs, feet are all facing the same direction?
i keep having issues, especially with back shots, where the head faces away but the rst of the body still faces the viewer

like this

[obtuse torrent]

there will always be failures, unless it's a build guided in some way... regional prompter or Controlnet multiposes ...

[tight light]

yeah im using controlnet

[tight light]

middle row, second from the right. This happens a lot

always in that spot

and it's always in back shots with ot without CN

so i assume its something to do with back shots

[obtuse torrent]

@tight light that is a product of the way the image is generated without a specific guide you will always have these problems appearing, the same for upside down or unusual poses

try to get a good template for controlnet to guide you, it's hard work with the editor... good luck

[white fable]

any particular reason on why abyssorangemix2hard is generating black images? i thought that was the whole point

[restive sparrow]

i generated this character, is there anyway, that i can consistently generate this same character ( face, hair, ears, look)

[white fable]

copy the seed and up the batch count

[rancid maple]

Please some can help here. I am trying to generate images with size of width:1216 height:832 but i always get images with 1024x1024

[floral fossil]

using which UI?

[rancid maple]

using bot on this discord

[floral fossil]

try to resize so pixel count adds up to 1048576

[slim ivy]

what is the input of negative prompts on the bot? --negative prompt?

[floral fossil]

negative_prompt

[slim ivy]

danke

[floral fossil]

after you write dream, if you click outside of the box you can see

the additional parameters you can add

[rancid maple]

i am putting this prompt but still getting the 1024x1024 image resolution .. Red leaves tree with white trunk , best quality, outdoors, close-up, high detail, 2k style:Photographic width:1216 height:832

[floral fossil]

that doesn't add up to a mega pixel

1216*832 = 1011712 (nvm don't think this matters as long as it's close to a megapixel)

you could try genning & using the resize option on the prompt, might get close to the resolution you want

wait

[floral fossil]

i checked the bot mentions and those are 832x1216

i think the height is bigger though rather than width

for some reason

[rancid maple]

i saw some other prompts using same size and their images match the same resolution but when i try with same prompt it's coming out at 1024x1024

[floral fossil]

your last one is 832x1216

but the wrong dimension is altered

also i think other people probably hit the resize button

and got a random dimension

[rancid maple]

one more thing i am confused some one entered along prompt they did not got token limit error when i tried i get LONG PROMPT message with this strawberry milkshake, bar menu, canon eos r 3, f / 1. 4, iso 2 0 0, 1 / 1 6 0 s, 8 k, raw, unedited, symmetrical balance, beautiful volumetric cinematic lighting, masterpiece, best quality, award-winning, coherent, exquisite cinematography, film still, RAW, shot on 35mm. 108MP, HDR, intricate extremely-realistic textures
Negative: unrealistic, semi-realistic, smooth plastic details, uninteresting, ordinary, ugly, plain, dull, sparse, painting, airbrushed, illustration, drawing, toy, doll, digital art, vector art, clipart, comic, 3d rendered. low-resolution, low-quality, smeary/smudged, overexposed, oversaturated, image grid, watermark

[floral fossil]

your prompt has a lot of commas, /, isolated letters/numbers/etc.

those are all individual tokens

you can have a near paragraph long prompt as long as all the words are tokens by themselves

each comma and space is a token

you could avoid using commas entirely a lot of the time

[rancid maple]

i thought comma was not counted

[floral fossil]

what was the prompt that went through? i think there is a 77 token limit still

[rancid maple]

i just copied some one else prompt and tried the image resolution for me is same 1024x1024 as the other member's image was proper size mentioned in Prompt of this one .Mother's Day poster, , lovely child sending flowers to beautiful mother with big wavy hair, shining big eyes, background in a warm home, perspective, light and dark contrast, warm colors, bright background, style:Pixel Art width:1152 height:896

[floral fossil]

i think commas count as tokens, but there's special attention applied to tokens within a certain range from a comma (at least in webui)

[rancid maple]

i just checked now the original message had LONG PROMPT warning . its correct 77 token limit is there for everyone

this image has same size as mentioned in prompt width:1152 height:896

and when i try with same prompt i get width:1024 height:1024

[floral fossil]

there might be a special interaction if you get the size via pressing the resize button vs manually inputting size

whenever i get a size through resize button it just works

might not be open to manual size testing right now

[shrewd temple]

Hi, why can't I use commands : style and aspect It shows me an option. Not a valid choice.

[floral fossil]

actually i noticed width/height aren't valid options? you have to use aspect i think

[floral fossil]

click style/aspect and select from the list. "landscape" isn't a valid aspect for instance

[shrewd temple]

Well, I didn't notice that, thank you

[rancid maple]

Thanks for the help i was just checking the words in prompt and manually using size now it works perfectly as you mentioned.

[prisma grail]

I'm trying to make objects and the issue is that it produces them at an angle, how is it I could make it only generate images that face the camera head on?

[prisma grail]

I don't know what keywords to use

[uncut flame]

front view

[prisma grail]

I'll try

it's producing better results but I still can't get it to be precise, any tips?

updated my model to v4.5 and it's a lot better

[restive sparrow]

anyway to find out what seed i used? since i generated an image after this one

[obtuse torrent]

parameters

white background

anime girl with light brown hair, her hair covering right eye, brown red eye, fox ears, brown hair, smile, right eye covered up by hair, by hair, bust shot, close up, blue bow tier dress, bright brown hair, hair covering right eye, two red marks on face, hair coveres right eye, red mark left side of cheek, red mark right side of cheek, one eye,
Negative prompt: low quality, worse quality, brown hair, bow tie, white hair, bow tie on hair,
Steps: 40, Sampler: DPM2 a Karras, CFG scale: 8.5, Seed: 845172862, Size: 768x768, Model hash: 876b4c7ba5, Model: cetusMix_Whalefall2, Clip skip: 2, ControlNet 0: "preprocessor: reference_only, model: None, weight: 1, starting/ending: (0, 1), resize mode: Crop and Resize, pixel perfect: False, control mode: Balanced, preprocessor params: (-1, 0.5, -1)", Version: v1.4.1

[white fable]

umm

go there

drop the image u generated

theres the seed

might be obvious but here goes the seed

anyone got any questions feel free to dm me lol

[kindred garnet]

any idea how i can get rid of framed pictures?

always struggling with that

ok this is pretty dark, but it has black bars on left and right

[solar dust]

I can see the bars.

Are they intentional?

[inner lantern]

Maybe I just have no idea what I'm doing, but I always get really disheartened trying to use stable diffusion to rough out a character for a story. Like, I feel helpless trying to fight the AI from just doing whatever it wants. I want a character with blue hair, yellow eyes, with black steer horns sprouted from her head, and it basically ignores half of what I say. The horns? If they ever happen (like 1% of the images) it's the wrong kind. It REFUSES to do blue hair. Even when putting it at the top of negative prompts, it DEFAULTS to blonde hair, and blue eyes.

I just don't understand how I've seen so many AI recreations of well known characters, accurate down to the intricacies of the outfits or the extremely unique hair styles. I try to steer it with the prompts and instead of refining as I tell it what not to do, it shrugs its digital shoulders and goes, "you're wrong, blonde hair is awesome."

[tired vigil]

Two times in a row trying different variations it gives me this instead of images,any idea what i might be doing wrong please?

[hollow tapir]

1. It depends vastly of what prompt (and negative prompt) you're using. I'm guessing you're using stuff like blue hair woman red eyes green shirt with horns. The AI does not understand really well how those words work together, how they fit, what's the context for them, etc. So with a prompt like that you can easily end up with a woman with red hair green eyes and red shirt and horns every word you used will be there.... Just not how you expected it. So you've got to learn how to write prompts correctly, try to use attention/emphasis syntax https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#attentionemphasis, use the BREAK keyword in your prompt, etc...
2. to avoid any kind of color contamination, cf : https://civitai.com/models/18840/no-more-color-contamination-read-description
3. the cfg value you're using can also impact how strongly the AI should follow your prompt. Don't overdo it tho.... anything past 10 will usually yield deep fried results if you're not using some special techniques. (not gonna talk about those for now at least)
4. same for the clip skip value. tldr leave it to 1 unless specifically told so by the model's author/if it's a derivative from "that one" model
5. people usually recreate well known characters using loras

[inner lantern]

"Loras?"

Thanks for the response, I'll definitely be checking that out. Nothing more frustrating than having an extremely capable tool and feeling like you're just playing with rocks.

[hollow tapir]

Lora, yup https://stable-diffusion-art.com/lora/

also cf #🤝｜tech-support message

[native plinth]

Hello, I'm new in this community! I was wondering, why does it come out with the kid? I only want the woman. My prompt is this: mature, motherly, blonde, 80's mom

[rich crow]

because you have motherly as a prompt, and the model probably knows that mothers have children

search google for mother, youll see that almost every picture has a baby in it

[restive sparrow]

ty!!

[restive sparrow]

Ty!!, Sorry im new

[dim pelican]

Hello, I want to share refer image to SD to make a new one, what prompt should i use?

[dim pelican]

Someone help

[blissful schooner]

My image always get messed up once I include Gal Gadot lora. Any idea why?

a 32 yo (gldot:0.3 | IU1:0.6),long hair, dark theme, soothing tones, muted colors, high contrast, (natural skin texture, hyperrealism, soft light, sharp),red background,simple background. lora:FilmVelvia3:0.4 lora:Gal_Gadotv3:1 lora:LORA_iu_v35:1

I am combining her face with another famous celeb. But still, even if I just use Gal Gadot, things are messed up. Either the bg goes awry or the body bones look really bad.

[noble charm]

weird question, but what would be the best model to make photorealistic catboys for a meme?

[static moat]

So uughh, probably a dumb question:
For some reason, I’ve been struggling with img2img. I have the program installed locally, and I am trying to have the AI add one matching thing to the image without changing anything else, or the base style of the image itself. I’ve messed with the settings and the image seems to always come out re-painted, messed up, or the thing I specified just doesn’t get added.
I’m a bit new to this, so I’m pretty sure I just don’t know the correct settings. 😛

[silver valley]

For that you need to Inpaint into the image.
Load the image into Inpaint tab in img2img.
Mask the part where you want something to add. Then prompt for it.
Select Mask only and denois 0.5-0.9 for example

[silver valley]

You need to lower the lora strenght by turning the 1 to 0.5 for example

[polar token]

What kinda prompt to make this kinda style

https://media.discordapp.net/attachments/1101178530865352815/1130442547911729172/Crackhead_aesthetic_professional_detailed_steps-59_seed-0ts-1689588889_idx-0.png

[static moat]

That got the area selected, but nothing is happening when I try to add something. I have moved the denoising slider all over the place, but for whatever reason, it refuses to edit the selection.

Not sure if Im missing an addon or something

[stable valley]

is there a format for a prompt to merge/mix 2 loras or text inversions of differnt faces?

i have tried [xxx:yyyy:1.5] where x y is trigger word

[noble charm]

^^^
basically trying to make a world leader into a catboy as a meme, any model suggestions?

[blissful schooner]

Thank you! That seems to have done it.

[hollow tapir]

also you don't want to use "restore faces" option for anime style output.

[steady pagoda]

hey guys, Im doing a female photo and I would like her to have a lil catholic cross necklace is there a way to emphasis I want a small cross and not like a "50 cent" one ? lol and also how to avoid getting cross on other piece of clothing ? any tips ? (im getting a lot of double cross)

I guess "cross necklace" is like a double use cuz necklace alone works pretty fine .. maybe with inpainting ..

[arctic tapir]

How do you even prompt this? (the single red line)

or you just edit in photoshop?

[stable valley]

try Cruxific instead of cross

[hollow tapir]

ears ?

[arctic tapir]

the red line

hair

[obtuse torrent]

yes PS... or inpaint .. or: .... black_hair, streakked_hair, red_hair + RNGod bless

[white fable]

yeah its w.e hmu if u have any questions

[hexed kernel]

It has nothing to do with what I asked...

[silver valley]

You can try lower the value of the lora. But i would recommend using a ghibli lora instead. Thats trained on the ghibli style

[hexed kernel]

OK thank you

[solar dust]

How did you manage to prompt 2 different people in the same image?

Let alone with different ages and clothing?

[final dome]

any prompt about cat? please

[whole skiff]

Does SDXL understand natural language? I was using 1.5 & 2 mostly with comma separated keywords but no setence.

[lean pond]

does anyone know what dreambooth repo works best with AMD GPU

[zealous depot]

Is it possible to tell @delicate kernel to select one of the 2 images and then make alterations to it?

[restive sparrow]

what would be the prompt to get this pov, you see the front side of the person, just the back?

[tall haven]

if i'm generating an image and i see something in the live preview process, is there a way to stop it right there so i can try to hone in on that image result before it continues getting processed? the end results are sometimes drastically different and i see some live previews that i really love

[tired vigil]

How do i get a higher quality image from A1111? The images generated aren't bad, just when i zoom in they tend to be pretty pixalated.

[tall haven]

what is your height/width set to ?

[tired vigil]

521 width 720 height normally, changes on what i'm making

[tall haven]

that would contribute to low res, but you can use a 2x or 4x upscaler to get larger higher quality images, sometimes lowres negative prompt can help a little too but im far from an expert

[tired vigil]

Don't think my computer could handle much more then what i have, 12gb gpu doesn't seem to be enough. Thanks for the help regardless

[silver valley]

Whats your gpu?

You should be able to easily upscale with 12gb vram

[tired vigil]

6700 XT

Dunno

So far i keep getting the "Not enough Vram" message when i try it

[silver valley]

I have tested this card and can upscale 512x768 by 2

Do you use the recommended Commandline args for amd ? In your Webui-user.bat

[tired vigil]

Yes

--medvram --no-half-vae --opt-sub-quad-attention --opt-split-attention-v1 --autolaunch
git pull

[silver valley]

Looks good. But you also can remove --no-half-vae. Shouldnt be needed for 6000 series

[tired vigil]

Alright

[silver valley]

Then try 512x768.
Highres fix, upscale by 2.
Denois 0.5
Hires steps 10
Upscaler Esrgan4x

[tired vigil]

Just got RuntimeError: Could not allocate tensor with 655360 bytes. There is not enough GPU video memory available!

[silver valley]

@tired vigil Okay then completely restart SD and try upscale by 1.8

Or 1.5 to test.

Then you also need the Tiled Diffusion Extension. Install that and then only enable the Tiled Vae.
That will help to get the upscale done.

Then you should be able to upscale by 2

Ive generated this image with an AMD 6700XT.
You can load it into PNG-info tab to check the Settings used.

[tall haven]

can someone tell me how to prevent multiple bodies from being added? it looks great in the live preview until the last few seconds itll add a conjoined body or an extra body in

[silver valley]

You need to lower the base resolution and then upscale the image

[tall haven]

its currently set to 512x512 and upscaling to 1024x1024

[silver valley]

The 1.5 based models got trained on 512x512 and the 2.1 models on 768x768 resolution

[silver valley]

Ohh okay then show your upscale Settings

Probably the denois is to high

[silver valley]

Okay first disable restore faces. That breaks faces when using together with Highres fix.
Set hires steps to 10 or 15
Then set the denois to 0.55

If you use esrgan4x as upscaler you can use a denois of 0.5 or lower

[tall haven]

that seems to work great except i lost her full body

[silver valley]

Okay, you can also set an other base resolution. For example 512x768 is good for portrait style or full body.

Im off now gn

[tall haven]

thanks for your help, gn!

[obtuse torrent]

you can use settings tab for that...
.....the fewer steps you will get more intermediate images....
consider that this will increase the space when saving each generation

[tall haven]

thank you i will experiment with that

[tall haven]

is there a way to exclude specific seeds ?

[white fable]

umm, never generated something like that but probably something among the lines of purple_mist purple_mist_background

[tall haven]

could also run that through a controlnet process and keep the pose/perspective etc

[restive sparrow]

ty!!

also how would i go from just the right upper part of the face, where only an eye and some hair is visible

[tall haven]

how can i minimize/reduce the amount of this "text" jibberish

[fiery pecan]

Can you share the actual image? so can get the meta data for the prompt.

I would like to do some testing and let you know.

[reef snow]

hi, i have a stupid question, whaty do you type in to get multiple people showing? i tried "2 people" or "1girl 1boy" but doesnt do anything

[fiery pecan]

@reef snow You can try using controlnet with an image of multiple people.

[reef snow]

i'm gonna try it, thanks 🙂

[white fable]

usually when you just want to show something specifically you just say stuff like legs_out_of_frame, or only_legs_in_frame, leg_focus

[scarlet citrus]

so in the current stable diffusion I always had good luck generating larger women with the "curvy" keyword, but now even with () around it, it doesn't seem to want to generate anything but a slender woman? any advice?

[white fable]

So umm you could try highlighting the word that you want, press alt or control i dont remember and press the up arrow key a couple of times to let it know that you really want curvy in the gen

[scarlet citrus]

TY I'll try this!!

[simple finch]

quick question, is this formula correct?
[from:to:When to make the change]

I'm trying to make something like this
a photo of living room with royal furniture style, and a carpeted floor, and a cabinet, white ceiling designed with [a painting of heaven: starry night :20]

but no results

[tropic garnet]

Hello!
I recently started using Stable Diffusion running in a local instance. I've learned most of what I know from the great guides here: www.stable-diffusion-art.com/beginners-guide/
Currently I'm trying to create character illustration/portraits of the party members for my first D&D campaign, using D&Diffusion checkpoint.

My main question currently is** is there a way to "restrain the reach" of adjectives in my prompt to specific part of it**.
For example, I'm trying to create a Dragonborn/Lizardmen with Gold scales, wearing an armour made of black scales, which is understandably confusing for Stable diffusion.
Is it possible to "retrain the reach" of the colour in the prompt?

Another question is the effect and usage of punctuation such as quotation mark and commas. Does they have an effect? My test haven't found much, but I see a lot of prompt using at least commas.
The only think I could find about this subject is the use of exclamation mark to increase the strength of some word.

[obtuse torrent]

when have 0.xx format ---> 0.20 (at 20% of steps)

[obtuse torrent]

curvy may have a conflict with Ethiopian (stereotype)
if the emphasis() doesn't work, you can try changing curvy .. to chest and hips

[simple finch]

you mean [something to replace : replace with : 0.2]?

so just writing 20 doesn't work?

[obtuse torrent]

@simple finch
any value.... "when" refers to a percentage of the steps.....
The change will be made after x% of the steps...
the format to write a percentage is 0. xx where xx is the % .. 0.xx is between 0.01 and 0.99

[simple finch]

Ok got it, Thankss.

[obtuse torrent]

It all depends on how tokenize the model, but in general in short prompts (-75 tokens) the effect is close to 0, in longer prompts the grouping effect that A1111 performs appears... when using the punctuation to separate and group "packages" of at most 75 tokens (packages have the same weight among themselves)... then variations appear since the same token will have different relative weight depending on whether it is grouped in a package of more or less tokens

[scarlet citrus]

ah thank you, that makes sense! I'll definitely play around with it, totally didn't think about that

[tropic garnet]

Thank you for your answer.
So if I understand properly, with the default settings, the punctuation doesn't have much effect?

[obtuse torrent]

only if your prompt is short < 75 ..( there may still be differences due to misunderstood compound words or descriptive phrases that lose meaning) ... I recommend that you do some tests keeping seed and other parameters constant, and thus appreciate the effect of adding or not separators and/or punctuation marks