📝｜prompting-help

[tired vigil]

Characters

Wuxia style

[lusty salmon]

https://media.discordapp.net/attachments/661714761964322828/1213083800343683152/c5f4dcd97b63e54871281817e83c6341.png?ex=65f42f54&is=65e1ba54&hm=244e2726dec004e0678ea817e2502dd210293aaabd1059cedbad10fa6b546a87&=&format=webp&quality=lossless&width=460&height=675
Does anyone know what prompts would make this kind of image?
I just cant replicate it

[ornate osprey]

Hi, I'm just getting back into Stable Diffusion after long absence. I'm trying SD XL Turbo and it doesn't even have negative prompt. Why is that? And what version are you using, that does have negative prompt?

[muted haven]

why it happening and how to prevent it? with prompt "photo of futuristic robot, ball in hand, ruins of city in background" I get robot with ball, but moment I change ball to skull face of robot also changing to skull. It is stable cascade with diffusers

[fathom marsh]

Stable cascade will put skulls everywhere except for women's faces, if there is a skull in your prompt. (It can't separate subjects in general.)

[nocturne osprey]

recreate this aerial view of a building with a swimming pool located on the lower gorund of the building located by the 195 highway of miami city

[sturdy mural]

hi

iwanted to prompt and i got this

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

what does that mean

[muted haven]

thanks

[torn cliff]

Not sure if it counts as prompting or not, but does anyone know of a Lora that has scifi like armor with a smooth glass/digital display faceplate?

[wraith tulip]

how do you put animal ears on the side of the head where human ears would be? Like a human with cat ears on the side of the head.

[warm bough]

m

[sturdy mural]

hey guys, found this beautiful generated image and i wanted to know how i can genrate like this and what the base model could be

[tall onyx]

lose weight girls

[oak swallow]

Do you guys have a negative prompt copy paste?

[quiet raft]

I am not using any thing like good hands, perfect hands etc.. What should I use in the Positive and negative to avoid these?

[hazy radish]

what model would you recommend for generating a pixelart like this?

[wraith tulip]

how do you put animal ears on the side of the head where human ears would be? Like a human with cat ears on the side of the head.

[ebon jackal]

@prisma dagger

[prisma dagger]

im here

[ebon jackal]

i think the issue is that the trees and landscape and the tram have no focal point

[prisma dagger]

seems hires fix was causing it

[prisma dagger]

how would i do that

just write focus on tram?

[ebon jackal]

you could try that. or like

i dunno, say wheels, axels, overhead wires, dirt on tracks

[prisma dagger]

just add that to the prompt or..?

[ebon jackal]

yeah just throw those in

[prisma dagger]

i will try that

oh wow yeah focus on tram helped a lot

[ebon jackal]

better?

[prisma dagger]

this is just with focus on tram added

[ebon jackal]

the more you tell it how nice the focal point is, the more it pays attention to it. your original prompt made the tram just part of the trees sarajevo, the tracks!

yes!

[prisma dagger]

right i see

[ebon jackal]

tell the tram how pretty it looks. not joking, stuff like beautiful lights, cobbled path

[prisma dagger]

the more detail added the better

i dont want cobbled path thats actually wrong lol

[ebon jackal]

oh then put in negative cobblestones

[prisma dagger]

yeah i will

[ebon jackal]

negative prompt cobbles, paving, urban

[prisma dagger]

here is the real one before and after renovation

[ebon jackal]

so you want the first one?

but modern tram?

[prisma dagger]

tbh im just experimenting

[prisma dagger]

more like a Tatra K2 tram

the ones we use RN

before the new ones arrive nest year

[ebon jackal]

oh then tell it a tatra k2 tram, be specific

[prisma dagger]

well i try specific a lot of the time but it doesnt know

i mean it doesnt hurt to try

[ebon jackal]

if you showed me a photo of it, it probably has seen that photo

[prisma dagger]

well it cant generate a MiG-29 the most widely photographed fighter jet in the world

[ebon jackal]

has a mig ever sat down beside you cuddling a cat for a selfie?

no? then its only photographed from complex angles so it hasnt a clue 😄

like if SD was trained on this photo and similar from the same side of the train, it wouldnt know what the other side looks like no more than you know the dark side of the moon. you can imagine it cant really as well

[prisma dagger]

yes

oh i read that wrong

close enough

[ebon jackal]

i put pounds to pennies on a bet it can draw that cat better than your mig 😄

[prisma dagger]

well yeah

actually i should try to get it to draw my cat

[ebon jackal]

you have a1111?

take that image

[prisma dagger]

thats not my cat

i want a fresh one

not img2img

[ebon jackal]

click those buttons once you put any photo into img2img, it will tell you what it, as an AI sees. its neat, and wrong a lot but still

[prisma dagger]

ive experimented with img2img

latest one

[ebon jackal]

oooh

trees still look like hell. type in "bokeh" at the end

[prisma dagger]

might save this seed and ty to get a better one

[ebon jackal]

the prompt changes the seed tbh

if you typed in blue train it would show it from the left or whatever

[prisma dagger]

so what is the seed specifically

[ebon jackal]

its like.... if you gave a computer a blank piece of paper and asked it to make X it wouldnt know. but if you give it an image of snow/tv interference and asked it to keep moving that around until it looks like a train it would

some seeds do just happen to be good, and keep that seed and try it

but if i took that and asked it to make a hedgehog, its not gonna be a hedgehog looking that direction most likely

its a bit like looking at clouds and seeing shapes

[prisma dagger]

oh no i meant save the seed for this type of image

just perfecting the prompt

it was pretty close

[ebon jackal]

yes, the prompt changing will not yield same image

hah

though im a cat racist, they all look like cats to me

[prisma dagger]

tbh i dont even know what breed my cat is

[ebon jackal]

i do like this tiger lookin guy

[prisma dagger]

i found her on the street

[ebon jackal]

shes from catland

so anyway, have you ever seived flour?

[prisma dagger]

idk what that means

english isnt my first language

[ebon jackal]

shaken flour through a mesh

[prisma dagger]

but i get what youre saying about seed

[ebon jackal]

to get rid of lumpy bits

[prisma dagger]

i did already know that an AI generated image just starts out as random noise

the seed determining the "type of random"

[ebon jackal]

yes well SD does that with noise. it will shake until the lumpy bits look like cat i guess

this is what clip thinks the photo is. thats pretty good i would say!

[prisma dagger]

how did you do img2txt

[ebon jackal]

just put the image into img2img or if you copy it press paste with the tab open

you can put in a photo of yourself and have it guess your age

dont be too hurt if it says youre a middle aged man with a goatee, it might be lighting

[prisma dagger]

whoops it accidentally made a bus

[ebon jackal]

yeah, but its getting better isnt it?

[prisma dagger]

yeah

[ebon jackal]

put wheels in negative prompt

[prisma dagger]

i did

well

[ebon jackal]

put (wheels:1.2) in it to emphasise it

[prisma dagger]

i put: road wheels, car wheels

[ebon jackal]

emphasis is a thing you can do too.

[prisma dagger]

yeah i know how the weights work

[ebon jackal]

highlight a word in the prompt. then press control+up key or down key

oh ok cool

[prisma dagger]

oh i just manually type that lol

[ebon jackal]

what is your cfg? by default it is 7

[prisma dagger]

a cat is laying on a chair in the middle of a room with a window and a chair in the background, Carlos Catasse, maya, a photocopy, tachisme
for the 1st image lol

[prisma dagger]

3

[ebon jackal]

oh move that waaaaaaaaaaaay up

unless its a lightning model, is it?

[prisma dagger]

literally no idea

[ebon jackal]

does it say "turbo" "lcm" or "lightning" in the name?

[prisma dagger]

yeah turbo

[ebon jackal]

ah right then you cant move it much

[prisma dagger]

yeah it makes the images oversaturated

[ebon jackal]

also, those models have a detail brick wall

[prisma dagger]

elaborate

[ebon jackal]

youre a 3070 or something arent you?

[prisma dagger]

3080

[ebon jackal]

do you have any "normal" sdxl models?

[prisma dagger]

yes

[ebon jackal]

lcm/lightning/turbo models kind of assume they know what you want in the first few steps. its mostly pretty girl robot big bust. anything obscure like a 1958 wind turbine in new mexico it stops being as helpful

[prisma dagger]

ahh

[ebon jackal]

yeah its like it flicks open the magazine to the page it thinks you like

[prisma dagger]

so the lightning models are basically for general ideas

[ebon jackal]

ok, open up the normal sdxl. turn your steps to 20 and your cfg to 7.

[prisma dagger]

this is what i got

[ebon jackal]

show me screenshot of your settings?

[ebon jackal]

try slimming down the prompt a little.

actually start with "raw photo of a tram" and no negative prompts

[prisma dagger]

yeah i guess this prompt was tuned for the other model

[ebon jackal]

we all overproduce prompts 😄

its annoying to see someone make the image you wanted it "fat cat sits on mat" and you've just written tolstoy

[prisma dagger]

result

time to prompt craft lol

also why is this faster than the supposed "lightning" model

[ebon jackal]

you have a goddamn good machine and you wouldnt notice. lightning honestly is for people making batches for animal or low end users

[prisma dagger]

i mean im generating 4 images in less time than it took to generate 2 at the same res

[ebon jackal]

if you want to detail an image, stick with standard SDXL models.

ive honestly no idea of the kungfu it does

i know input ....... MAGIC ....... output 😄

ive an idea

that photo you showed me earlier of the real tram

remember i showed you how to clip interrogate it in img2img?

[prisma dagger]

clip it and use the prompt?

[ebon jackal]

yeah

see what it comes up with

[prisma dagger]

immediately knew what you were gonna say when you said this

[ebon jackal]

you're getting it 😄

there also might be much better non-lightning models for this stuff on civitai, but stick with this for now

[prisma dagger]

a trolley car is traveling down the street in a city area with people walking around it and trees in the background, Boleslaw Cybis, zabrocki, a digital rendering, viennese actionism

thats what i got

[ebon jackal]

try that.

in text to image

[prisma dagger]

am

Sarajevo has a lot of cobbled tramways so i think its getting confused

[ebon jackal]

oh god! i forgot americans call them trolley cars and thats where 90% of photos are likely from. it probably knows a "trolley car" better than a "tram"

[prisma dagger]

ah

ill try that

will it know what "tram/trolley car" means

[ebon jackal]

also if you know the name of the tram in russian or whatever language, put that in instead. sometimes that works

not if you put in a slash

[prisma dagger]

its a czech tram

[ebon jackal]

you can do [tram:trolley car:0.5] and it will alternate the two

one step tram, next step trolley car

[prisma dagger]

ah ok

[ebon jackal]

throw in czech too. its not an english model for englishmen. it doesnt know english it knows letters beside images it sees a lot. hell put in ^__^; 🙂 and see what happens

those are text letters beside images 😄

[prisma dagger]

here is that

[ebon jackal]

i really think its just bleeding through old 1960s posters into its learning

[prisma dagger]

oh Tatra K2 is the same in czech and english

[ebon jackal]

ask it to make a big train with no wheels flying through california

[prisma dagger]

and definitley american

[ebon jackal]

bet it can do it better

[prisma dagger]

these are very much american trams

i wanna see what happens if i switch out trolley car to Tatra T2

T2 is a trolley bus

whoops

i meant K2

[ebon jackal]

if tatra t2 had photos before 2022 it should know

oh is it trying to make a terminator sequel?

[prisma dagger]

well that tran is from the 60s

so there is definitley photos of it before 2022

[ebon jackal]

try these negative prompts grim, traintracks, dull, drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly

[prisma dagger]

it got really close with this one

anyway

thanks @ebon jackal for helping me with all this

[ebon jackal]

no probs, you now know the bits and bobs

[hazy cobalt]

are there any good...'rules?' for prompting for more than one person? pretty sure it's AUTOMATIC1111 that i'm using.

unless that's just the dev... still kinda new to all this.

[agile bloom]

I'm making some a.i. generated dance videos, there's a lot of fast body/hand movements, are there any great standard prompts I should put in? I copy and pasted the negatives prompt in style so I always use those.

I hope they aren't conflicting anything too

[willow scarab]

is there a general guide for prompting when starting off? Im just starting and my stuff looks garbage

[prisma dagger]

i always start my prompt with "RAW photo of" and end it with "highly detailed, extremely realistic, volumetric lighting"
my most recent gen using the model analogMadness

also i have this standard negative prompt

(octane render, render, drawing, anime, bad photo, bad photography:1.3), (worst quality, low quality, blurry:1.2), (bad teeth, deformed teeth, deformed lips), (bad anatomy, bad proportions:1.1), (deformed iris, deformed pupils), (deformed eyes, bad eyes), (deformed face, ugly face, bad face), (deformed hands, bad hands, fused fingers), morbid, mutilated, mutation, disfigured(deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation

[ashen goblet]

hey. Is it possible to regen the exact same image in a higher resolution? Well, thus not exactly the same, but yeah. Upscale looks way worse than generating at the target resolution itself, but not really feasible to gen at 2048x2048 while praying for a good gen... 🤔 help?

[devout vale]

The only thing that comes to my mind is using hires.fix with low denoising strength

[ashen goblet]

is that how hires.fix works? 🤔 the image in this case was created through img2img, so I'm not sure if the seed in text2img would yield the same result, without the input image and img2img doesnt have hirex.fi

[ornate osprey]

I'm using SD XL Turbo and it doesn't seem to allow negative prompts
What version are you using?

[prisma dagger]

ah im using A1111 local generation

[tall reef]

I see a lot of caricature LORAs but the prompts used always have the name of the character/person

I'm trying to use the img2img to do a caricature of myself but the output looks nothing like me

any tips?

[finite gale]

how do you all get your gens, to have the whole arms spread out for a hug-ish

[silver valley]

Img2img has "hires fix"
Its called SD upscale script.
Make sure to set the Denois to 0.15 and then set the resolution to 512x512 or if you use sdxl to 1024x1024 and then select an upscaler and generate

[ashen goblet]

oh wow, that's so hidden

[willow scarab]

All my pictures of people look like this. The head oftentimes is not fully in frame. It eiher is way to zoomed in, or the frame is wrong. How to fix this?

[tired vigil]

Hello! Sometimes, tags like 'out of focus' or 'out of frame' can help when applied to the negative prompt. Or, you could have a look at 'camera cheatsheets for perspective and camera settings', hopefully providing you with the 'right' words.

[vagrant jasper]

I'm using stable-diffusion-xl-1024-v1-0 engine with the stability api. Trying to reduce disfigured bodies/hands etc. I've got this growing list of negative keywords that I append to any user supplied ones but they don't seem to do anything. Prompting it for: a couple out to dinner I get lots of weird hands and eyes.

Questions:

1. Recommendations for additional/better keywords to use?
2. Is this list too long?
3. Is v1-6 better than sdxl for this?
4. Any other known hacks to improve this?

Thanks in advance!

const DEFAULT_NEGATIVE_KEYWORDS = `
amputee,
bad anatomy,
bad illustration,
bad proportions,
beyond the borders,
blank background,
blurry,
body out of frame,
boring background,
cut off,
deformed,
disfigured,
dismembered,
disproportioned,
distorted,
duplicate,
duplicated features,
extra arms,
extra fingers,
extra hands,
extra legs,
extra limbs,
fault,
flaw,
fused fingers,
grains,
grainy,
gross proportions,
hazy,
improper scale,
incorrect physiology,
incorrect ratio,
indistinct,
kitsch,
long neck,
low quality,
low resolution,
malformed,
misshapen,
missing arms,
missing fingers,
missing hands,
missing legs,
mistake,
morbid,
mutated hands,
mutation,
mutilated,
poorly drawn face,
poorly drawn feet,
poorly drawn hands,
squint,
ugly,
unfocused,
unattractive,
unsightly,
misspelled,
bad spelling,
improper spelling
`;

[tall reef]

would 3-5 pictures be enough to train a simple lora of a face

to do caricatures?

[glacial stirrup]

Trying to inpaint left bare foot of a sitting person. The original foot is really deformed, missing toes etc. Tried controlnot (openpose, canny, depth) and different settings with not so great results. Ideas what to try next?

[lucid mesa]

Assume unlimited compute. Assume I have a 135 IQ for output efficiency. Here you will operate as my stable diffusion guide. While I study, keep in mind that I have very basic knowledge of art terminology. Therefore you may use as much advanced terminology as needed regarding artistic styles, terms, equipment etc., as long as you define it. I will use this to explore the various options as I teach myself the technology. We'll be focusing on fine-tuning detailed settings available to the various sampling methods. Assume I want to output HD images only. Lets start by walking me through making my first image.

This is helpful

I use Gemini

Leave out assume unlimited compute if you don't have a higher end card

or just insert "I have a [whatever card] with Stable Diffusion [whatever version] installed locally" in between the first and third sentence

This functions well for any of this new tech you want to learn about

It's kind of plug and play

Keep in mind Gemini only updates like every 3 months but if you're on that much of the bleeding edge of this stuff you're probably not hanging around here lol

[limpid jewel]

im really new to stable diffusion and not really sure where to start, i tried some prompts but they are not working all the way and need refining. im trying to make images like this one.

but what i get are like this.

the prompts i used are: Human skull looking forward, drawing, black and white, white background, 2d, no cracks, flat

Any tips would be appreciated.

[lucid mesa]

I sent you a DM

[rare brook]

generate video from this image

[hollow tapir]

use "line art"

and "soft shading" in negative

[wooden radish]

If you just want a copy of the images but in a larger format, without adding details, use Extra to enlarge.
I tried so long with SD upscale, that I stopped using it. When I wanted 4K it created artifacts between one tiles and another.

[ashen goblet]

also interesting 🤔

another question that i want to add is, for sdxl models, if you have the resolution as 512x512, but hires.fix on, resulting in the final outcome being 1024x1024, does that still work? Or does the initial size need to be 1024x1024 to be? (do sdxl models even benefit from hires.fix?)

[rare brook]

generate video from this image with slow motion

[median bronze]

What are the best SDXL models for the most accurate vehicles, Im currently using Juggernaut XL, ZavyChromaXL, & DreamShaper XL, are there any others I should try. I am interested in photorealistic, and some stylized.

[static hinge]

Hi, I was wondering the best way to prompt a photo with flash at a dark room that illuminates the room and cast a shadow of subjects, kind like these (even more gritty maybe, non professional)

Sometimes I get with something like fill flash, flash spillage but other times it makes spilling liquids...

photo with flash in dark room it doesn't necessary makes the effect

[proven jasper]

this lora can do that https://civitai.com/models/52652?modelVersionId=88101

[static hinge]

Oh I'm using SD XL, but thank you. Actually I got this effect with Leosam sdxl model and I started to try to prompt it in other models. https://civitai.com/models/43977?modelVersionId=338512

but even in that model sometimes it works and sometimes it doesn't

[proven jasper]

yea leosam has that effect on their xl model but idk the trigger words for it

[quartz coral]

how do I set the age in prompt? I get bad armpits and skin and ReActor detects 30-37 y.o. My prompt is currently " a teenage girl with a pair of sunglasses and a choker, (age 15:1.5)...." I need her to look 18-21. (dont look at the face, it s changed by reactor)

[quartz coral]

I used Age Slider embedding, it s better now, but still not great armpits

[craggy yarrow]

this is a cool lora, but definitely has a samey-face issue.

[static hinge]

try in the negative "mature, middle aged"

[ashen goblet]

... help. I suck at inpainting, but I want to inpaint the image to essentially make him wear a helmet. How do I do this?

[brittle geyser]

How do I stop generating images with crossed leather straps in front?

It's getting a bit absurd

[ebon jackal]

polaroid photo flll flash, man indoors by wall, film grain off-axis flash, (hard light:1.3), dark shadows chromatic aberration Its fairly consistent, a 2:3 ratio seems to make it snap to polaroid look, but it also throws frames around it, like scans of a photo. play around with that.

(doesn't have to be indoors, the left subject prompt is man outdoors by wall) e: you might even get better results if you can spell fill correctly, instead of flll like me 😄

[agile bloom]

hello, trying to get hands infront of face, they keep showing up behind or hidden away. Is there a prompt I don't know of that can have the hands show up?

[silver valley]

What are your settings?

[ashen goblet]

inpaint masked only, denoise varied between .7 and .9, fill/original attempts, but original never generated a helmet and just kept the face+hair, somehow, hence I tried more fill. I tried using the seed+seed variation etc, and prompt changes to not mention the face anymore so it generates a helmet there, but the result is visible

[silver valley]

What resolution and model? 1.5 or sdxl?

[ashen goblet]

in inpaint I used 1024x1024, but sometimes also lower resolutions if I'm inpainting only a small area, seemed to work well with high quality anyways when I edited eyes for example...

in txt2img I use hires.fix with a base resolution of 512, resulting in the same 1024x1024

seems to give me nice quality 🤔

^an image I just generated with those settings

but the editing helm onto head thing, I cant figure out

[silver valley]

Ah okay sdxl, then yes 1024x1024 should work. But you need to specify what you want to get created. So just prompt for the helmet and not for forest etc

When inpainting

Also dont add Upper body

Medival knight helmet, detailed, masterpiece, etc should work

[ashen goblet]

you mean, it'll properly attach the head then?

wtihout fucking up the rest?

[silver valley]

Yes it should

[ashen goblet]

I'll give it a try in a moment

[tired vigil]

Hi all, how do i add another image to another image, without messing up the image?

anyone can point me to the right direction...?

[ashen goblet]

well some progress for sure, but it's not quite right, I'd like a helmet that covers the house and the cheeks more, like outlined on the left, kinda..

[static hinge]

great prompt, thank you! yes I'm getting it more now

[silver valley]

You would need to increase the Denois then

Maybe even use 1

Or you use Inpaint sketch

Then use grey color over the face

[ashen goblet]

🤔 I gotta try that next, thank you

[little prairie]

can someone help me with two colored hair? If i want something like a white hair patch on a character like guts has or jason todd, anyone knows how to do it ?

[ebon jackal]

does it need to be consistent?

if not something like photography, woman in dress, with black hair, (green forelock poliosis 1:3)

it might be one of those times where an anime model gets the job done better

[little prairie]

I am using it on anime model but sometimes it turns hair completely white sometimes it's not there etc.

I tried to do it on this picture

[bleak goblet]

Hello can someone help me to get this type prompts and good model to use in comfyui

[elfin crescent]

im making fantasy character for my dnd campaign, and I am trying to get people with pink skin, but every time I type it in and put in negative propmts white skin or black skin it dosent work how can I make them have pink skin

[worn anchor]

What colour skin is it returning?

[elfin crescent]

white

pretty shure its adding makeup

because I looked up pink skin

and that what I got people wearing pink make up

I even put pink paint but nope

[worn anchor]

Maybe try "pink colored skin" or try Fuschia

[elfin crescent]

still making white skin

but I did got cool eyes

im trying different models

[little prairie]

just tried this prompt, absolutely broke the pic,

[ebon jackal]

Yes, you're using an anime model, the prompt I gave was for photorealism models (where it can still be hit or miss)

[little prairie]

oh i see well fug..

[ebon jackal]

idea: get an anime image of one of those characters or of an anime character with hair like that and use the booru interrogate tool in img2img (if using a1111)

anime models are trained on image board tags, so there might be a specific name for that in anime

you're probably better using inpainting or controlnet for specific things like this, theres no magic prompt that will give consistent results with colour, especially parts of hair

[burnt pilot]

Hello people, how are you? I'm new to this, can you help me with the prompt?

[tired vigil]

im very bad at writing my own prompts lol can anyone help a little bit?

[burnt pilot]

or this

[frank bison]

Cartoon caricature

[zealous egret]

what prompts should I use if I want to generate a random arrangement of liquid on a plain color background? in other words no humans or any other figures, or furniture or land or whatever

[twin yoke]

Something like this?

[zealous egret]

no that's too complex, uh let me get some stock image

https://www.cleanpng.com/png-orange-juice-lemon-juicer-fruit-juice-391606/

like this or its similar images

[twin yoke]

Well, I'd start trying things like "one liquid yellow paint droplet on white background" and make adjustments from there.

This is one medium droplet of liquid red paint on white background

[zealous egret]

the AI can read that? I assumed I needed to find particular keywords or something

[twin yoke]

Stable Cascade, but you could get there with other models. It's not terribly complex.

Yes, it could read that.

[zealous egret]

is there a way I can tell the generator to avoid reaching the edges of the frame? (so there aren't unnatural cutoffs)

[bleak fossil]

hey everyone, can anyone help me with an image ? I would like to adjust the right hand, i'm on it since hours tbh and can't make it, I think the easiest solution is just to draw the shape of the hand, and then do an inpaint, that's what i've been trying so far, the ai success but my hand shape sucks 😦
here is the image https://ibb.co/GRy9g9d, it's the right hand, it's holding nothing and in bad shape, the left hand is also holding nothing but at least it's less obvious, it's well shaped

[tired vigil]

what are your parameters? maybe a controlnet inpaint would do the trick

[cinder olive]

I'm trying to get it to do a top-down rpg style battle map, the sort you'd see on roll 20 or foundry, but it keeps giving me this weird isometric view. how can I make it not?

Masterpiece, best quality, (((battle map, top down rpg battlemap))), battle map of a neon city during the middle of a market festival, cyberpunk, sci-fi, dnd map city map, festival stalls, from above, bird's eye view, Star Wars, highly detailed, dnd map, urban,

NEGATIVES: cropped, worst quality, low quality, normal quality, jpeg artifacts, (signature, watermark, username), monochrome, text, error, lowres

[dull berry]

I am trying to create a prompt with 2 people with different features. Anyone have ideas?

[hollow tapir]

Use regional prompter https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111?tab=readme-ov-file#regional-prompt-control

[safe hemlock]

Hi guys, I want to use the openai clip model and a classification model to format a simple image description (for example a user who wants to create an image) into an sd-prompt.

My thought was to use clip to capture the text features and then train the classification model on the features and labels (which would be the sd-prompts).
Could anybody with more experience tell me if this is a viable way to achieve my goal?

[dull berry]

thx

[fervent mirage]

i am new here. i can't seem to get to the place where i can start writing the promts where is that

[hollow tapir]

#1047610792226340935 ,
Either you install it locally or use online services such as https://dreamstudio.ai/

[fervent mirage]

I am using dreamstudio now, but it is not realistic images, all are cartoonish people.

[orchid scaffold]

Any tips for getting a character to close their mouth?

[sturdy mural]

does someone know how i can achive this quality? what base model could be used etc?

[sturdy mural]

what can i do to make it more brighter?

[tired vigil]

how can i "combine promts, iant a cape that is a hankerchief for my fairy but i ony get normal capes

or that when i add new promts that it just "ignores" prior ones

[near fjord]

I am running into an issue where my VRAM is not enough to train my model using everydream, is there a possibility i could use a online gpu for everydream?

[tired vigil]

okay, how do negative prompts work? i have the feeling that it does the opposite, i added child and childish but my generated fairy's look more and more like small children

[hollow tapir]

weird... you got the idea. whatever you prompt in negative prompt should be avoided during generation. Maybe try increasing the weight of child in negative ? prompt specifically for adult ?

[tired vigil]

i think the problem is i want a Tiny/small fairy and the ai "thinks" that as child. so my negative prompt is not strong enough

[hollow tapir]

what are your prompt and settings exactly ?

[tired vigil]

it stopped working (connection problem) so i closed it and forgot to save my prompt 😅 . so i started over. now it really looks like i it was the tiny. tiny fairys look more childish

[hot hare]

Do anybody know about a model able to generate a Booru type prompt from a human natural language prompt ?

[compact maple]

What's the significance of using comma like this? Same with prompts that uses "BREAK"

[hollow tapir]

Using multiple "," is still nebulous even for me... It is changing the output for sure, how so exactly however.... it might also be the user's way of formatting their prompt for better readability.
The BREAK keyword however is used to manually split the prompt for clip.
your prompt being analysed by clip in multiple chunks of letters/words. when you type BREAK it will fill the current chunk with void char/close it early ( not sure on the technical implementation ).
So basically it s a way to manually guide clip so it can understand the prompt better.
So your cute girl, yadi yada, elegante purse, yadi yada doesn't get split in two
cute girl, yadi yada, elegante + purse, yadi yada sub prompts that doesn't make much sense.

[little prairie]

Anyone using AI to draw DND maps ?

[frosty summit]

Isometric.. maybe "cavalier perspective"

[west laurel]

Someone can tell me how to generate a statue in this perspective angle?

[frosty summit]

See above ^ I already replied to you

[misty cobalt]

#📝｜prompting-help I am trying to create a logo in this style.

A crest-shaped vintage logo with two foxes and two paintbrushes in black and cream with 1401 in the center of the crest, but I can not seem to get a good prompt. Any suggestions?

[civic snow]

which models are good at doing humans? I have:
juggernaut XL, juggernaut reborn, dreamshaper, dreamshaper pixelart, 'badass cars'
I know the last 2 aren't meant for it, really (and cars is 'lora' ig)

[nocturne raptor]

I like epiCrealism - Last Unicorn, although I think it's 1.5 only.

I've been so obsessed with PonyXL so I can't remember if there's an XL version of it.

[civic snow]

not sure if im limited to 1.5 or not with webui.

[rain cradle]

hello
I want to create a live-action version of an animated image.
I would appreciate it if you could also let me know about the negative prompts..!
Attach photos.

Please help me

[vestal tiger]

教程

[austere ginkgo]

looking for some prompt about: changing: clothes background and poses. somebody knows?

[slate radish]

Hi, need help with getting more than a face portrait in Fooocus Ai, I keep trying stuff like 'full body' or 'head and shoulders portrait' but all I'm getting are face slose-ups.

[tired vigil]

look up on civitai unrealisitc dream

[west laurel]

Can we generate oblique military perspective with Stable Diffusion 2 or 3?

For a statue like this!

[runic bear]

can somebody help me out with this masking issue with invoke-ai? the actual problem is, no matter how precise i paint the mask around this guys face, the complete right side of the face gets distorted a bit, completely ruining the facial expression, i also tried to use an image without the background, the results werent any better. i even outlined the guys body with a black 2px outline, but even then at least the eye/brow section gets distorted

i get results like this

[static hinge]

Hi, is there something like a clothes prompt database?

[civic snow]

Are there any models or loras that can do good looking dinosaurs? I've tried like 20 different lora and probably like 10 models, none of them seem to do a good job at preeting realistic enough looking dinosaurs. Doesn't need to look 100% perfect but I would like 95% at least to be able to use in blog posts

[strong token]

Yo guys, how does one make good backgrounds?? I am trying to replicate japanese visual novel styled backgrounds like these 🙏

[hollow shard]

https://huggingface.co/spaces/SmilingWolf/wd-tagger

[hollow shard]

what the fuck is this prompt

[fleet wadi]

Hi, can anyone please help me?

What is the best way for me to take for example an outfit from a photograph and apply it to an image I made in SD? If I want photorealistic results and for it to be as close to the source image of the clothes as possible?

[kindred blaze]

So I have been playing around with Bing's image Creator a little, making these really pretty mannequin styled characters, and I'm really curious if anybody knows what checkmark and/or Lora combination I could use in stable diffusion to get some similar results?

Prompt I am using:
a beautiful young mannequin Park Ji Hoo mannequin inside a mannequin display window, white wedding gown with fingerless lace gloves, bright neon pink low pigtails hair, gold choker necklace,hd detailed glossed face with large beautiful eyes, mannequin body and face coated in a smooth thick layer of clear glossy plastic, glossed fiberglass mannequin body, mannequin body, engagement ring, beautiful long pink square acrylic fingernails, looking at her hand

[slow cedar]

Hey there! I'm trying to make something similar to a ink painting, mostly because I need an image that is pure b/w, no greys, no gradients, but everytime I've tried it din't work at all, all images I did had coloring, and theres no resemblance of the ink style I've put on the prompt. Any help?

[agile bloom]

now for some reason I keep getting more close ups of my subject instead of a full view of them standing. Any prompt for full image?

[tired vigil]

question, im trying to make it where a picture is in a dark room with little lighting what words could i use, I use SDXL so its a little different

[frosty summit]

Dimly lit, candlelight, dark.. most simple things tend to work for me

[silver valley]

dark lit room

dimmed lights

also use the sdxl offset lora

[quiet raft]

Anyone has any fix on getting better hands with a model ?

[open musk]

what must i do to fix upscale ruining eyes?

[elfin crescent]

can someone help me get something like this ?

[silver valley]

Can you show an example?

[open musk]

for example the eyes on this

i paint his eyes and prompt

[silver valley]

What upscale technique do you use?

[open musk]

uhh

comfy latent upscaler

by

[silver valley]

Hmm can't help much in comfy but that doesn't look upscaled

[open musk]

ye athis isnt upscaled mb

does it have to be?

i just wanted to try inpaint

to fix face

[silver valley]

Ahh okay

Then inpaint the eyes or the whole face

Set the Denois to 0.5 or higher

[open musk]

alright

i think i did it with 0.6 but it didnt have a good result

i think it might be sdxl issue

[silver valley]

But your input image is pretty bad quality. You should generate it again but with hires fix enabled

Oh make sure to use 1024x1024 when generating with sdxl

Or 768x1024 for portrait mode

[shadow sleet]

hello, it's not really to ask help for a prompt, but rather to ask help to find a safetensor/checkpoint to create specific images.
can I ask here or is there a dedicated channel I didn't pay attention to ? ty

[quiet raft]

What prompts could I add to try avoid these kind of mouths?

I am using a cartoon-ish model

[normal roost]

Anyone have a better picture that shows the differences between all sampling methods?

[tired vigil]

There is a yt video about it

From my understanding, euler a = fast

and DPM++ 2s a karras has given me the best results i believe

DPM++ 2M should also be good

you can use the script thing to generate the same thing with different samples and compare directly

[rustic ravine]

ddpm is very fast and good

dpm++ 2m k is middle of the road

dpm++ 2s a k is the best quality for slowest

[tired vigil]

i have never used ddpm

[rustic ravine]

try it, its shockingly good

[silver valley]

smile, smug, grin, expressionless

[silver valley]

you can ask here

[daring geyser]

I need help im not sure why but lets just say i put "man sitting in a chair eating waffles" then it makes the most atrocious thing ever like it looks mutated and idk. its also weird because i had a detailer on same with highres any help?

[velvet spire]

Can anyone help me with finding prompts to generate a specific pose? Or at least in the right direction? Don't really wanna use open pose bc it's time consuming af

[trim merlin]

G Morning,excuseme,¿where i write the promtp here to make images?

[buoyant yarrow]

The bots are down

[brazen cargo]

https://github.com/Nick088Official/SuperPrompt-v1/tree/main?tab=readme-ov-file#usage-locally

Here's a way to run Brian's superprompt model offline (There's also a huggingface space and colab in the link)

[trim path]

lfg

[brazen cargo]

[slender linden]

perfect prompt example to generate everything ? who got an example

[twin yoke]

There is no perfect prompt. There is also no way to "generate everything". Your premise is invalid and rejected.

[slender linden]

i mean a structure of a prompt that can be used in everything

and have always good results

[twin yoke]

Doesn't exist.

[slender linden]

what

like cinematic style prompt

[twin yoke]

It doesn't exist. Outcomes currently depend on too many variables to receive good results 100% of the time. Even with a highly repeatable prompt structure, seed values, model choices, samplers, step counts, and more will have so much variation that you will not get good results at times.

So again, doesn't exist.

[slender linden]

when i just write a prompt it gives me trash results always

when i use negative prompts it becomes better

[twin yoke]

But even then, let's say you find something that generally works well...like 75% of the time...art is subjective. What's trash to you may be good for another and vice-versa.

[slender linden]

when i use cinematic style it gets much better

but it lacks something

the cinematic style cant generate for example expressive images

[twin yoke]

Sure it can.

[slender linden]

try to type corona virus in cinematic style

[twin yoke]

You just have to find the right combination of all the variables.

[slender linden]

it will gives u trash result

[twin yoke]

Maybe for the settings you've tried.

[slender linden]

yes im trying to find that

something that can work in everything

[twin yoke]

That doesn't exist.

[slender linden]

why not

[twin yoke]

I explained that already.

[slender linden]

a good combo of negative prompts and something else

some keyword in the postivie prompt that i can use always

[twin yoke]

I don't think you understand how image inference works.

[rich crow]

(masterpiece:99.99), best artwork, masterpiece best artwork by gold medal artist

these are the 3 secret ingredients to a great masterpiece artwork

[twin yoke]

It's not actually looking at your words. The words are converted into tokens which are then used as mathematical vectors across latent noise data created by the seed value and then sampled on a schedule to inferan image with pattern data that's been trained in the model. Change anything at all and you get a different result.

[rich crow]

if you want a serious response, soul pretty much summed it up

[twin yoke]

lol...your prompts are trash if used in a different model.

[slender linden]

if i type in the prompt vikings, it will give me trash result

but

[slender linden]

when for example i use this with vikings it gives me better result

that wat im looking for

[twin yoke]

I can type in just the word vikings with no negative at all and no other words and get a good result.

All I need to do is change the model.

[slender linden]

i cant

[twin yoke]

Sure you can.

Find a better model.

[slender linden]

what model u are using

[twin yoke]

I use tons of different models.

[slender linden]

i want 1 model

and 1 prompt style

[rich crow]

tbh thats kinda the thing with ai models, weird whacky crazy things end up working okay when you dont expect them to and thats perfectly fine
however its important to remember that every model is different so they probably wont work with other models. but if they do and you like the results, go for it

[twin yoke]

Exactly.

[slender linden]

what negative prompts u use

[rich crow]

ive only ever tried crazy word salads

[twin yoke]

Make images with SDXL models...many of them don't require much use of negative prompts.

[slender linden]

what model u suggest

good one

i dont want to use multiple

[twin yoke]

Juggernaut XL, for one. But again, you need multiple. There is no singular perfect model. Doesn't exist.

[slender linden]

doesnt have to be perfect

[twin yoke]

Doesn't matter.

[slender linden]

i use juggernault xl

xd

[twin yoke]

One word prompt: vikings
No negative.

Juggernaut XL v9

[slender linden]

bad picture

[twin yoke]

Nah.

[slender linden]

yes

[twin yoke]

Ok, troll. You can go away now.
Art is subjective. The image is perfectly fine.

[slender linden]

this is good

[slender linden]

not at all

colors bad

faces bad

and no details

trash lighting

[twin yoke]

All opinions. The colors are realistic because it's a realistic model. The faces are fine and distant, hence the details are lower. The lighting is typical outdoor lighting.

Now go awy, troll.

[daring geyser]

any fix for this?. i have the ponydiffusion cuz im tryna do so lora things but when i put this on comfyai it turns out like this, but when i put it into automatic 1111 stable diffusion it works any help?

[trim path]

btw gonna update this when the v2 comes out

i hope soon

[minor eagle]

trying to get the model to stop producing swords, stafs etc

[lofty halo]

anyone have any tips for just simple vector line art?

[desert rampart]

To be able to create images like this, a novel visual type, what prompt would I require? It's just that sometimes I try but either it takes a pose that I don't want or the camera is in the wrong position.

[hexed jolt]

Guys I need help please.... I can't make the AI design a fantasy dragon without spikes or horns! I've wasted so many credits already.

[misty halo]

@desert rampart have you tried giving an artist example as a art style?

@hexed jolt are you using negatives? And I'm not familiar with types of dragons but is there a term or references for that type of dragon you could base your attempts off of

[worn dome]

@ebon jackal I'm here

[ebon jackal]

ok paste those screenshots in here too so everyone knows what on earth we're talking about 😄

[ebon jackal]

you've decribed face features lots of times. its already using a controlnet it knows theyre there. start by removing duplicates of descriptions of teeth nose face etc.

tell it about detail not that theres teeth. things like "hyperrealism" and "skin pores" for example

i think its pretty good so far to be honest

[worn dome]

I've seen duplicated prompts in many images that look very proffesional. I assume people prompt the same thing in a different way to enphasise important thigs, so that they make sure AI "understand"

Of course I have no idea

[ebon jackal]

true but theres also such a thing as cargo culting where you do what other people do without knowing particularly why 🙂 its very common in programming too

you dont need to tell it its a front view, you dont need to explain chromatic abberation and jpeg artifacts to a 3d based model either

the less words here the more the model can focus on the strong things and detail it

[worn dome]

Oh

I didn't understand this: "you dont need to explain chromatic abberation and jpeg artifacts to a 3d based model either"

[ebon jackal]

chromatic abberation isnt a thing that happens in 3d, its a photography thing. this model youre using is tuned on 3d models so it shouldnt attempt to show photographs.

[worn dome]

I didn't know that

[ebon jackal]

think of it this way, SD is trained on lots of images, some great, some terrible. those images will have a vague description of them that the model knows and can relate that idea to other things. like "black and white". so you're not really telling it to not made black and white images directly like a computer program, you're asking it to filter out images it considers that

it doesnt actually know any of the words, so you cant reason with it. so things like chromatic abberation would never come up anyway with pixar type images, no more than a guy photobombing the shot would 😄

[worn dome]

That makes sense. After all it's neccesary to know the theory in order to create this images

[ebon jackal]

yep. or just mess around.

[worn dome]

I'm trying to generete stuff with the tips you gave me, I'll try to show you soon

I'm tryng "forced face expressions" in the prompt, but I neither think that's a prompt. Maybe I could replace it for something more understable for SD?

[ebon jackal]

theres a bit of a cheat you can also use and just download a styles file. they're prewritten prompts you just insert your subject into

[worn dome]

I think I already have one of those

For now this is the prompt

And this the result

I got the ending control step in 0.9

[ebon jackal]

the hair is messed up but you're getting there. also an upscale works wonders

[worn dome]

I can just Photoshop, I've got versions with nice hair

[worn dome]

I let that for the end

[ebon jackal]

run it a few times it might get the hair right

[worn dome]

Any suggestion for hair?

[ebon jackal]

upscaling doesnt just make it bigger, it gives the model more room to play with an add details. try this, upscaler: r-esrgan4x, denoising 0.35, scale 1.25 hires steps: whatever your prompt steps are halfed

1.25 with half steps will double the length of your generation but you should see an immediate improvement

try "anime hair"

[worn dome]

The upscale is in the generation, or I do an img2img after getting my generation done?

[ebon jackal]

do it as part of the generatio

n

[worn dome]

Where are that options?

[ebon jackal]

what is your starting image size btw? 512x768?

[worn dome]

I don't find that in my SD

[ebon jackal]

what program are you using? a1111?

[worn dome]

Yes

[worn dome]

I got this in txt2img but not on img2img

[ebon jackal]

ooooh youre using timg, not img2img

err other way around

i thought you were using a controlnet on a text to image

[worn dome]

https://youtu.be/_krLW0RQSuM?si=_MJXCZLa-yDWtdbb
I use to follow this tutorial for upscaling

In my case I upscale to 4K, as I need it for my videos

This channel really explains well

I'm falling asleep, thank you for your knowledge. See you

PD: Adding (ugly face:1.3) in negative prompts bettered a lot the face :D

[daring geyser]

anyone?

[runic sierra]

@ebon jackal I've noticed all the advanced prompts around but don't know enough about them to be able to use them. Is there a data file/sheet that you can refer me to that will explain all these symbols and syntaxes? Just asking you because you seem like someone who might know.

I'm trying to get I to some very advanced AI image generation, mostly T2Img. So I need to know what I'm doing.

[ebon jackal]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features this is a good page to read, scroll down to "prompt editing"

things in (brackets like this) emphasise a term. you might see them cluttered around in prompts

dont use them too much, think of them like focus on X. so if you said a photo of (big teeth) it will make a closeup of big teeth

if you say a photo of (big teeth) and (sneakers) it will end up kinda malforming it because it makes a mid sized image detailing teeth and sneakers and ignoring the middle

(big teeth:1.3) is over emphasis on it.

in general if you find a good image full of (words:1.3) they're a fluke.

in SDXL images you're better talking in real terms like "a toothy man standing in really clean reebok trainers"

weirdly with sdxl, since its trained harder on award winning photos, if you want an abstract or beautiful artistic piece i find starting it with a madeup title some fancy art guy would give your imaginary photo works really well, and then detailing it

so, smiles and trainers in seattle, big teeth, clean trainers

works very well for abstract ideas

you can also do [jesus|jimi hendricks:0.75] to have a switch between the two and make a 3/4 one 1/4 other

trying to think of others

oh loras, as a rule of thumb, if you add in multiple loras, never let it add up to more than 1 or else it gets overbaked. so lora:this:0.4 and lora:that:0.6

overbaking is a thing you'll recognise as soon as you see it, it looks like little squares of the image are melting

[runic sierra]

Thanks.

Also, what is noise? Specifically. I've heard the term used but not sure what it actually means.

[runic sierra]

Im also doing a lot of focus on non-living objects like machines and well...objects.

[ebon jackal]

ever see in old cowboy movies where the panned for gold?

AI cant just made a thing from an instruction, it needs some kind of base. a random seed is like the snow you'd see in old analog tv channels

it gradually tries to work on those and make them look like an instruction you gave it.

so thats kinda what noise is

[runic sierra]

Oh

What's the seed? I've seen seed numbers and settings but not sure how they work either.

[ebon jackal]

theyre just random numbers that make like the snow that it builds on

sometimes you might find a seed that works really well for an image, but its not really controllable.

its sort of like making pictures from cloud formations but with computers

maybe one accidentally seems to lead it towards the left of an image more and generates great portraits of a certain kind, but when you change the prompt itll be completely different. anyone talking about a perfect seed is sellling woo

[runic sierra]

Okay. Good to know. If my prompt looks good but it's just not working, could changing the seed affect this? Just asking for the hell of it.

I've tried making an image from a line drawing I did, but the poor AI couldn't make heads or tails of it. I probably needed to give it some shadows and the like to give it some pointers.

[slender linden]

who got a good prompt that works well in every type of pictures

[grave plank]

we

[timid rapids]

Hey all, im trying to get lettering on a bottle in an image. I am using ComfyUI, easyProductPhotoXL_v10.safetensors.

My positive prompt is "Snow-covered chalet, a pine bench outside of its windows. On the table, standing tall atop a red Christmas napkin, is a glorious see-through bottle of Grey Goose Vodka. The words Grey Goose are labelled on the bottle; the bottle is reflecting the scene around it"

My Negative is "Watermark"
Simple worlflow, but the letteriong on the botle is off. Any pointers on how I can fix it?

Oh, im using SDXL

[worn dome]

If I'm prompting, can I adjust the weights of only part of the prompt?

For example, for the prompt "purple denim jeans" can I do something like:
(purple:0.8) denim jeans

[leaden socket]

Hey lads,

I'm almost to the point where I think my prompts are perfect, but I'm still getting the AI look when it comes to a few parts of an image, any suggestions on making it so it looks a little more realistic, and not a flat material after I upscale?

Upscaled

[worn dome]

Can I see the prompt?

[leaden socket]

It's really messy but the upscale is this:

[worn dome]

I just wanted to se an "almost perfect prompt". Unfortunatelly I don't haver knowledge on that interface. I use A1111

[quartz coral]

hi, I need a picture of a girl looking at her phone, but I can achieve it. I ve tried 'looking at phone', 'holding phone', 'looking_at_phone', ' girl holding iphone, looking at phone, iphone'

[silver valley]

1girl, looking down at her phone, standing on street, sideview, close up, masterpiece, best quality, open eyes,

1girl, starring down at her smartphone, standing on street, sideview, close up, masterpiece, best quality, bluey eyes, crystal pupils, open eyes, smile,

[sturdy mural]

Hi guys, I want to make 3 characters via regional prompter but i only get 2, what can i do to improve my prompting?

3 characters, one girl, one boy and one child, sitting on a bench, scarlett blossom flower field, looking happy, spending family time
official art,extremely detailed CG unity 8k wallpaper, perfect lighting,Colorful, Bright_Front_face_Lighting,shiny skin,
(masterpiece:1.0), (best_quality:1.0), ultra high res,4K,ultra-detailed, ADDBASE

1manly man,male, 22 year old, Sasuke, bzsasuke, very manly face, broadly built lora:Sasuke.Uchiha.Shippuden.comish2:0.8lora:bettermalebody_bf16_batch1_v1.0:0.5 black hair, black siren eye ,manly face, uchiha sasuke, anime,s, short black onyx hair, closed mouth, grey jacket zipper pull tab,rope,hair between eyes,braid,pants,black eyes,short sleeves,shirt,bangs, spiked hair, a bit taller than karin
ADDCOL

1woman 19year old, karinuzumaki, (face lighting:0.8), lora:Karin:0.8, long red spiky hair, red eyes, glasses , cute look, earrings, purple jacket, ,red hair, ,glasses,
best quality (masterpiece) ADDCOL

1girl, solo, black eyes, black short hair, glasses, short black shorts, overknee socks lora:Sarada:0.8

lora:Sakimichan_style_v01:0.2

[orchid scaffold]

What is the best way to add a backgground to a subject that currently has a white background?

[eager nimbus]

hi folks, is there a model for this 2d style?

[worn dome]

What would you change in the prompt in order not to seem like he wants to kill me?
I just want a normal face expression.

[twin yoke]

You know what I would change?

I'd tell people what the original prompt was in the first place, homie.

I just want a normal troubleshooting scenario.

[worn dome]

Of course

[twin yoke]

With all this stuff, it's safe to assume that's a 1.5 model, yah?

[worn dome]

Yeah, how do you know?

[twin yoke]

Because XL & Cascade don't really deal in junk prompts.

[worn dome]

What do you mean with that?

I know my prompting is bad, but I don't understand the last part

[twin yoke]

All that garbage in the prompts that doesn't actually describe the thing in the image that you want is not really something that the other models require. 1.5 does well with a lot of that because people trained with a lot of that garbage in the prompts. XL & Cascade only require you to describe what it is that you actually want (or don't want in the negative), generally.

So...have you tried giving the character an emotion a bit more toward the front of the prompt?

Not only do weights impact the prompt, but the closer to the beginning, the more important they are...so if something isn't working, you can try shifting the word closer to the front. I'd say something like "happy" in front of the rest and see what happens. Then make adjustments from there.

I'd also take everything describing the character and put them ahead of the garbage.

[worn dome]

I had heard about that

Can I assume the "garbage" is : 8k, UHD, high resolution, delicate, detail, realistic, highly detailed, perfect composition, beautiful, detailed, insanely detailed, octane render, trending on artstation, masterpiece, intrincate details

[twin yoke]

Yup.

[worn dome]

And you want me to get the "garbage" at the end of the prompt

[twin yoke]

I'm just suggesting that you put all the character stuff ahead of it. That stuff is generally more important. Then, if something is coming in too strong, you can move it back or vice-versa.

[worn dome]

So you want that I delete the "garbage" from the prompt?
I'm not english, so I don't underestand well the expressions

[twin yoke]

No, don't delete it. 1.5 does better in a lot of cases with that stuff because of how people trained. Just move it to the back.

[worn dome]

That was what I said, I think

[twin yoke]

Friendly note: "Delete" means to remove.

We don't want to do that. 🙂

[worn dome]

Oh

[worn dome]

According to this, when I use XL models I won't need to write all that "garbaje"?
So I will be able the do good images with less prompt?
And what about 1.7 models, it's like 1.5 or like XL?

[twin yoke]

There is no such thing as 1.7 models. 1.7 is the version of Automatic1111 (the software running the various Stable Diffusion models). 1.5, 2.x, XL, & Cascade are the model version numbers. XL and Cascade do better without the other stuff usually.

[worn dome]

Interesting

It's nice to have another opinion. I'm glad to learn

[true thunder]

hi there. I am on a website named civitai and i am trying to copy this persons informtion over to my local SD. I have downloaded their model but there is an embedding file that I can download too ending with .pt

do I just put that in the embedding folder on my local drive?

because there is also a tab on SD thats called "textural inversion"

and that i think is my embedding

but I cannot select it

am I missing a step?

[elfin crescent]

im trying to get this man a cristal crown

like made of purple cristal

but I cant

help ?

im using dream shaper

7

[quiet raft]

I often do like 50 images as a batch and go through and delete the bad ones.. But sometimes there is a bad one but I wanna fix it.. I am not used to Img2Img and not sure what to use..

Should I

1. Use same + and - prompts? or just empty?
2. Rszie mode: Just resize?
3. Sampling Method: Same sampler?
4. CFG Scale: Same?
5. Denosing strength: like 0.05 - 0.2 so the image is the same?

I hope that someone know more about this and can help me 🙂

[hollow oriole]

A crisp and professional 4K HD LinkedIn profile photo of a 27-year-old arab fair skinned boy. He has short, styled hair and a lean build, exuding confidence and ambition. He is dressed in a stylish, well-fitting three-piece formal blue suit, with a white shirt and a blue tie. The background is a clean and clear professional studio setting, emphasizing the subject's strong presence and focus.4k, Ultra HD, realistic

[nova imp]

I should generate a batch of images to respond to these...

[twin yoke]

Yes. Yes you should.

[nova imp]

But I need to think of original ideas 🤔

[nova imp]

🤣 I got some things in mind...

[regal elbow]

I need help, i just want side wiev of the original image

[vital crystal]

Please help, I want to reduce the noise near the original image so that it blends better.

[buoyant yarrow]

Uh interesting

[inner ibex]

Hello!

How does someone check what words are influencing other words (or tokens I guess) ?

https://github.com/benkyoujouzu/stable-diffusion-webui-visualize-cross-attention-extension didn't worked on ForgeUI

and last time I installed https://github.com/AUTOMATIC1111/stable-diffusion-webui-tokenizer, it didn't made much sense to me. Just showed up token number (ID)

Is there something like showing a Venn diagram or relational graph of how CLIP links up words in your prompts?

[grizzled fossil]

Does anyone have tips for good tokens to imitate something made with a spray paint or a blow pen?

[elfin crescent]

how do I make a blind eye or prompts please

[ashen goblet]

I need help finding the tag/prompt to create this kind of veil. Pls ping if you reply to me.

The only veils I seem to find are either nose downwards, or bridal veils.. but none that actually cover the full face. Essentially, I'm looking for a veil tag that generates a full face veil like that, but a bit see through

that's an image from fumetsu no anata e

[boreal prairie]

nice

[worn dome]

You used an SDXL cha¡eckpoint?

[brisk elm]

Hello, what should I add to the prompt so that it does not cut the object?

I tried it: Full object. uncropped body. full image. uncropped head. body not out of bounds.
But no results.

[silver valley]

Add cropped, crop, to the negative prompt

[safe oasis]

Is there any way to see which concepts are trained and which ones aren't, and how heavily, and which words are used, besides trial and error?

[mossy knoll]

Hi! I'm trying out stable diffusion with comfy ui and my prompts seemingly get "overcorrected". For example, in response to "woman with four arms coming out of her shoulders" midjourney gives me pictures of women with four arms (ben 10 four arms), but stable diffusion gives me a woman with just two arms. similar with "woman with sword coming out of her mouth". is there a setting I'm missing? how can I tone down the overcorrection (the negative prompt fields are empty)

[elfin crescent]

no checkpoints

I was using rev

[sharp ember]

anny idea what could be causing this?

been trying Anything XL model and only this model gets these results

also reinstalled it and used different VAE's

[twin yoke]

You must use the XL VAE with XL models.

[sharp ember]

o im stupid

does using XL vae's with standard models also mess up?

[twin yoke]

Yes.

VAE is the encoder/decoder and each model type has a different one. Essentially, it's like a language. Some are similar in some ways, but ultimately, if you want to understand, you have to speak the same language.

[sharp ember]

ah ok

ive just been seeing them as fancy coloring things

actually while im here

are there any other settings u recommend me playing around with? like clip skip stuff

ty for helping btw

[twin yoke]

Clip Skip is typically going to be only really useful at 1 or 2 depending upon the model. There are some models where you can push that number, but I believe that the further from 1 you go, the less prompt adherence you'll get.

[sharp ember]

okidoki tyty

ill probably leave the rest on default

[tired vigil]

Hey guys, I'm new to Stable Diffusion. I am looking to replicate this ribbon style, but I am faling to do so. What tags should I use?

[bleak widget]

Hey guys, Im trying to make a dnd character for my gf, it's a moth sorcerer, These are my current prompts

creature with (moth wings), anthro, mothkin, moon moth, moth wings, antennae, lunar, moth fur, bug, sage green, pink, leaf, solo, wings, glowing, luminescent(masterpiece:1.2) (best quality) (detailed) (intricate) (8k) (HDR) (wallpaper) (cinematic lighting) (sharp focus)

Negative: human, 1girl, symmetrical from_front (photo photography photograph) (cartoon) (saturated) (grain) (deformed) (poorly drawn) (lowres) (lowpoly) (CG) (3d) (blurry) (out-of-focus) (depth_of_field) (duplicate) (watermark) (label) (signature) (text) (cropped)

I want to have something like the OC art but based on the Chinese moon moth, but I keep getting generations that just look like humans with mothlike features, I want a humanoid moth with all the fur and stuff

[tired vigil]

HDR photo of woman, dark-brunette hair, light-blue eyes, (f_stop 5.6), (focal_length 28.0), f/5.6, 28mm focal length, sitting in bed, contemplating, sunset, romantic, warm, revealing clothes, alluring, white sheets . High dynamic range, vivid, rich details, clear shadows and highlights, realistic, intense, enhanced contrast, highly detailed

[tired vigil]

Which model are you using? You'll want a model like Pony diffusion/4th tail/autismmix.

[slender linden]

how can i use an image as the style for all other images that i want to generate

[bleak widget]

Ah I'll look into it, Anything I could change about the prompt?

[civic steppe]

I found a message regarding the tokens to use for fill-in-the-middle for stable-code, however, I have not been able to make it work using ollama, what am I missing, I am using ollama like this: ollama run stable-code:code ; does anyone have an example of fill in the middle prompting for the recently released stable-code model? this is the message I found: https://huggingface.co/stabilityai/stable-code-3b/discussions/2 (TL'DR: "<fim_prefix>", "<fim_middle>","<fim_suffix>","<fim_pad>")

[bleak widget]

use img2img

[ashen goblet]

bumping my question one more time

[bleak widget]

you could use some img2text to see what the AI sees the veil as, when ur in your img2img tab you can click the box or clippy icon to generate text from an image

make sure your reference image is loaded into ur img2img tab

[ashen goblet]

huh I didnt know about that feature. Where's that box thing? found it..

[bleak widget]

👍

[ashen goblet]

the tags I learned aren't exactly helpful haha, it simply doesn't describe a face....

[bleak widget]

hmm, try the clippy icon, and if that doesn't work, I would suggest keep generating with the prompt you have and put things like "bridal veil" and "nose downwards" in your negative prompt until you get the desired result

the clippy icon gives you more of a descriptive text instead of just tags

[devout vale]

Hi, I just recently started using sdxl, can someone explain what are score tags? And what's BREAK? I've seen people use score_9 and such and use BREAK randomly between prompts and I'm just confused

[silver valley]

These score tags are used for the Pony models and its merges

They won't do anything on other sdxl models

BREAK resets the token counter. So the next word has the same attention as the first one in the prompt

[devout vale]

Oh thanks, is there like a written guide on it as well? I couldn't find any

[silver valley]

For pony and the scores, I would check the civitai side of the model

For Break, there isnt a guide it just have this one function to Split the prompt

[quaint light]

does anyone know a good way to prompt someone either having their hands on viewers chest or hugging the viewer?
I tried > "hands on anothers chest, pov"
"hugging viewer, pov"
but the results arent really working

[slender linden]

on what settings

denoising cfg scale and other stuff

[solid summit]

If one would want to adhere to a rule that no companies or artist names are used as tokens, how deep is the models language understanding? In general are models limited to what metadata might have existed in the training dataset and miscellaneous descriptive texts or is there something more that is added so that very specific descriptive tokens have impact? Like, how well can I describe a artistic style? Is there any point? Or is there any dictionary that has every token that a model knows of?

[dawn ice]

Images: what I was using for img2img and controlnet. I have been trying to get SD to shade my sketches and it winds up getting skullfucked sideways every time. It will put random colors and other shit in there even with the flat colored sketch. I tried using openpose, canny, linework and still shit. The only information I found was a sketch -> professional quality work and little to no explanation. Bitch theres no fucking way you did that in one step. How do i get it to not put stuff in that shoudnt be there and not render the mask as a fucking blindfold?

[rustic hearth]

I found a pretty neat prompting technique for blending things... just thought I'd share
(blended style of <("Style 1":0.3)> <("Style 2":0.8)> <("Style 3":1)> <("Style 4":1.0)>)

[marsh echo]

Hey guys,
When I use control net for a pose and openpose as model, I always get the same /very similiar hairstyle as I get from the control net img input.
Even if I change it in the prompt. I don't quite understand how I can get the AI to ONLY use the pose as reference and not the rest of the image as well (Such as colors)

[silver valley]

hey, its only using the pose as reference

the hairstyle comes from your prompt

[marsh echo]

I tried without any hairstlyes or appearance.
It still always got the hair in a ponytail or 2 buns, instead of open and long D:

I tried with appearance in the prompt and without

[silver valley]

can you show an example?

[rustic hearth]

Today I decided to play with facial expressions, using the same tactic as specified previously.

Baseline Prompt: (w/negative prompt (open mouth),(freckles))

headshot of a woman,
(blended style of <(Pixar Animation)> <(Hayao Miyazaki)> <(Realism)> ),
white background,
(lighting is a blend of <(ray tracing)> <(night lighting)> <(bloom effects)> <(particle effects)> ),
(high quality textures),(masterpiece),

For the next few we'll look at some basic expressions.
(images will appear in order starting at the base image and then moving down through the listed expressions that we added to our initial prompt)

(blended facial expression <(surprised:2.0)> <(open mouth:2.0)>), *w/negative prompt closed mouth
(blended facial expression <(surprised:0.2)> <(open mouth:0.2)>), *w/negative prompt closed mouth
(blended facial expression <(surprised:2.0)> <(closed mouth:2.0)>), *w/negative prompt open mouth
(blended facial expression <(surprised:1.0)> <(closed mouth:2.0)> <(smile:1.0)>), *w/negative prompt open mouth
(blended facial expression <(surprised:1.0)> <(open mouth:2.0)> <(smile:1.0)>), *w/negative prompt closed mouth
(blended facial expression <(angry:1.0)> <(closed mouth:2.0)> <(frown:1.0)>), *w/negative prompt open mouth and teeth

Anyways you get the idea...

Note: Obviously the seed stays the same for each of the images. We just alter the prompt.

[tired vigil]

ohhh

[icy trail]

Hey, I want to create a picture of a tree with hand-like branches in a pot. Where should i write my prompt?

[buoyant yarrow]

Elaborate

(Real answer: the bots are down. Be patient.)

[fluid sonnet]

i wont to create something like (img on the left) but the model is always flying or totally wrong

actual configuration

[worn dome]

Fast question,
Let assume I want to generate a red and blue helmet.
Can I prompt "red&blue helmet" instead of "red and blue helmet?
On the other hand, would be that more efficient as it's a shorter prompt?

[limber shard]

safe to asume whetever color you mention everything will be that color XD I just avoid using nay color in my promt

bluew eyes will mena blue jeans blue shirt blue sky blue wall, dam di da da dee da da dum everything is blue

[silver valley]

You should take a look at the cutoff extension

With that you can define what color the eyes, hair, and clothes should have

And it won't mix them up

[limber shard]

where is it link please

[silver valley]

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

[kindred elm]

Hey guys can someone tell me why im getting this error! and hot to clear it?
"NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check."

[woven mirage]

Hey guys, I keep seeing the name ''ponyXL'' on civit, very curious about what that might be and I'm having trouble finding info online, does anyone have an article explaning or know what that might be in a few words ? thanks 🙂

[twin yoke]

Not sure why you would have a problem finding information...a simple google of "ponyXL" should give you this as one of the very first links:
https://civitai.com/models/257749/pony-diffusion-v6-xl

[woven mirage]

Thanks, wasn't sure if that was the actual thing or not 😅

[slender pasture]

Trying to IMG to IMV something but I can't get the clearity to work out:

any tips?

resize mode: crop and resize
Smapling method: sampling euler a-

Resize to 522 to 768

CFG Scale 5
Denoising strength: 0.3

[keen panther]

I have a model of a character whom I want to texture using AI.

Is there a way for Automatic1111, Forge, or ComfyUI to use an animation of a turnaround of my 3D model and a fully textured version of one of the frames to generate a new turn around that is fully textured? Or is there a method that is similar but achieved my goal more effectively and efficiently?

I had looked into Dream Textures texture projection in Blender, Zero123, SV3D, EbSynth, app.meshy.ai, as well as doing it manually with canny, normal, and depth maps but it proved too time consuming or unreliable.

To be more specific with the methods I had tried:
Dream Textures is very intensive on my computer and does not output my desired results nor allows me to generate the back side of the model that matches with the front.

Zero123 really distorts the image, requiring me to modify the image manually which is time consuming. In addition, Zero123 does not have a reference to my 3D model and its shape at different angles so it predicts what will be in that space (This is not much of a concern for me even if it is an inconvenience).

SV3D seems to look for images with defined shading, but the images I had inputted have flat shading so that I can project them directly on the character. This thus causes SV3D to mess up the rotation. SV3D also has the same issue of not being able to reference the original model as mentioned in Zero123.

EbSnyth breaks when the turn around rotates far from its keyframe, thus outputting images which require manual editing.

app.meshy.ai only allows me to use it a limited number of times and does not output the results I want.

Manual editing for every ⅛ of a rotation is time consuming as it requires a lot of trial and error as well as inpainting (especially for the character’s face).

[abstract gust]

hmm... what is generally defined as a "tall" character? I may have a couple of characters that are tall in some source media, but I want to understand the trigger word better to have better useage in my prompts

[coral tangle]

Hello to everyone, I have a question about training stable diffusion webui, I have 2 videos first one one is normal and second one is the same but with cool 8 bit effect that I cant repeat in any editing software, the videos are from pxlprsche, as I understood he took the video and passed each frame through ai, I want to train my own ( I tried to find if he used already made one but didnt find any that makes this effect) so it gives me same effect, so to say I have a lot of his material to feed to ai but I dont if it possible to say to ai in stable diffusion " look I have this original img and here is the one with effect i want you to be trained so you can make the same effect to img i give" is it possible with stable diffussion or i need to use another software of ai?

[brisk surge]

How do I add little red accents in the armor in this prompt? Weighting doesn't work anymore, and just adding red armor on top just creates red armor instead of a light grey armor with red accents.
Prompt:
score_9, score_8_up, score_8, full body, teen girl, fantasy, orange hair, short hair, green eyes, freckles, fair skin, short, light gray armor, red armor, standing pose, holding poleaxe

[grand moon]

Try "light gray and red armor"

[rich crow]

Could try some form of gilded armor or find a way to throw the word accent/accented in there. Trimmed armor might be worth a try but I wonder if the model would lean towards the "cut" definition of trim.

[slender pasture]

What do I need to do to make my Img2Img to come out clear?

Im working with a GeForse RTX 2080

I fixed it a LITTLE, but I still can't get it clear 🤔

Okay well the rest of the image is coming in pretty clear now- but I still cant figure out why its giving me rasin eyes

[south lily]

Hi. Is there a prompt to to automatically generate an entirely new person with every new picture, without having to up date the prompt every time?

[slender pasture]

using the word alternate helps me with random clothes?

🤷‍♂️

[south lily]

thanks i'll try that

[slender pasture]

im getting so frustrated with this, id love some advice

[silver valley]

Do you have face restore enabled maybe?

[granite spear]

Hi, I've been trying out stable diffusion for the last month now via the Fooocus webui, and I was wondering, when describing two characteristic of something, do I really have to describe it twice, like "red skirt, mini skirt"? From my tests, it seems that making a single description like "red mini skirt" diesn't work that well.
Am I missing something?
Thank you for your time 😋

[atomic sapphire]

Hey ! Do you guys know a way to get only outlines out of a picture/prompt to color it yourself?

like a coloring book

[granite spear]

There are Loras for that, I never used, but did you try?

https://civitai.com/models/5153/coloring-book
Here's an example for SD 2.1, the first one I found

[grand moon]

You could try a prompt like this: prompt:YOURPROMPT, digital media, lineart, monochrome, high contrast, sketch, vector negative_prompt:painting, shading, colored

[atomic sapphire]

Thanks to you both, i'll look into your solutions !

[kindred blaze]

Trying to get a sort of sexier female version of an arcade fortune/gypsy teller cabinet machine, this is about the closest I have gotten so far, but most of the pics end up with mannequins and women outside the cabinet entirely... could use some help please 🙂

wanting something closer to this. but um...cuter/sexier xD

[open basin]

not quite but getting there.

lol

it does not like the enclosure prompting it keeps setting here out like this @kindred blaze

[kindred blaze]

That looks pretty decent, I appreciate the help, I still need to figure out what checkpoint and settings I want to use for it and any kind of Lora's that might help

[open basin]

how are you prompting it i get different results when I add weight to "gypsy automaton inside of (however I describe the various enclosure styles:1.25) and so on

it really seems to work well at about 1.2 with that being the initial prompt, on multiple models for having her in the cabinet / arcade case etc

[kindred blaze]

Well I am still very new at stable diffusion, not sure how alot of the settings work, but this is what I've been doing so far..https://gyazo.com/c0fdc618b5ccafff320dd8626d37ae75

[open basin]

this one here that is almost done is as:

life-sized fortune teller robotic automaton (is inside of a wooden fortune teller arcade box:1.15), automaton has a very uncanny valley look and porcelain skin, automaton is depicted wearing a vibrant gypsy costume including a headscarf adorned with rich patterns and beads, a dark emerald green dress with golden trimmings, a gleaming golden belt, automaton's eyes are accentuated with makeup, suggesting depth and wisdom, the automaton gazes forward with an enigmatic expression, automaton's hands rest on a fortune teller's table covered in an ornate cloth and crystal ball, one hand slightly raised showing its palm, surrounding scene is that of a eerie carnival.

using sdxl with combined L / G prompts same.

let me try this

[kindred blaze]

Sorry lost me a bit there with the end of your instructions
"using sdxl with combined L / G prompts same."

[open basin]

well so sd1.5 uses one clip encoded style prompt, the CLIP/l and sdxl uses CLIP/l and g they sort of react differently to differnet prompt styles, in this case i just doubled the same prompt up twice

like that

just use the same prompt twice. it comes out way different than just using the single clip node

[kindred blaze]

Oh ok 👌

[open basin]

not sure how a1111 handles that tbh, it may just combine automatically, I never really use it

[kindred blaze]

Appreciate the help, I can see I have a lot to learn still, but it's fun to play around with ☺

[open basin]

yeah, that's definitely what it is, fun... lol

[kindred blaze]

This was the kind of result I was getting earlier with the Bing app, was kind of hoping to get similar with stable diffusion

[open basin]

hrmm yeah you'll have to consider what the training data calls the thing

getting there

and this one, if not for the lady that leaked into it

close kind of thing

this is looking... interesting, though not at all what you really looked for, but has the affect

I think I have a decent prompt for this now, just need to find a better model probably

and maybe adjust my visual descriptions of the automaton

wide photograph of (a wooden ornate antique animatronic fortune-telling machine with a mechanical gypsy automaton:1.15) at a carnival, in the style of zoltar automaton has a very uncanny valley look and porcelain skin, automaton is depicted wearing a vibrant gypsy costume including a headscarf adorned with rich patterns and beads, a dark emerald green dress with golden trimmings, a gleaming golden belt, automaton's eyes are accentuated with makeup, suggesting depth and wisdom, the automaton gazes forward with an enigmatic expression, automaton's hands rest on a fortune teller's table covered in an ornate cloth and crystal ball, one hand slightly raised showing its palm, surrounding scene is that of a eerie carnival

is current prompt I used

[kindred blaze]

It's pretty close, ideally I would like her to be completely inside of the cabinet with the traditional three windowed walls and solid back, drapes tarot cards crystal ball that sort of thing 😊

[open basin]

that's wild

some of these getting really odd lol

@kindred blaze ok here we are now

not exactly the same, but stylistically the cabinets are about as close as I think I can get

[kindred blaze]

The top left and top middle look about perfect what I'm trying to get

[open basin]

except this thing, which looks cursed

[kindred blaze]

Lol

[open basin]

that one?

[kindred blaze]

Yes

[open basin]

i got an idea hold on

[kindred blaze]

Also, not sure how people feel about using specific actors or actresses names and such, but I'll sometimes use Ji woon-Kim as a reference.

Either that or simply "Asian woman"

[open basin]

oh you want asian woman gypsy, lemme try

I don't think people much care what people do as long as it isn't really over the top and messed up, or sold etc.

she has gumballs

[kindred blaze]

Those are both looking really nice

[kindred blaze]

The cabinets are pretty nice in the in that batch, some rather interesting gypsies 😅

[open basin]

yeah one looks like the guy from CHiPS

[kindred blaze]

Lol

[open basin]

pope

@kindred blaze there perfection

[kindred blaze]

yikes 😂

Few more results from Bing's creator

These are actually coming out really pretty 😁😊

[unique crystal]

great mango

how to generate img

[river nacelle]

Technology gun, sci-fi future, Lego, pTechnology gun, sci-fi future, Lego, packaging box atmosphereackaging box atmosphere

Technology gun, sci-fi future, Lego, packaging box atmosphere

[magic mauve]

poster

[wicked blaze]

Good evening, can someone advise me how to make a prompt so that 3 people appear in the image? Thank you

[tired vigil]

1.5 or xl? Theres also an extension for this. Or, better yet, you could use control net. open pose would work well . Ha theres an extension to edit poses, and in that editor tab a button to add another skeleton. combine that with a well worded prompt and youre gold

[mint ibex]

Good lord... She's not supposed to have a mouth, she's not supposed to have that much yellow glowing features on her, and she's not supposed to be black, but rather dark blue with hotrod flames!

How do I do it?

also her twintails are supposed to be mechanical too.

(I also talk about my generated image on the left, the right one is a slightly vague reference to what I want, mostly the flame vinyljobs.)

[mint ibex]

I SAID NO MOUTH! ugh... I give up, I'm way too impatient with this shit.

[twin yoke]

skill issue

[mint ibex]

or just unfocused today.

unfocused, cranky, and impatient.

[twin yoke]

nah

skill issue

😛

[quasi hearth]

Hi! I need images of a road that goes diagonally into the horizon. but every time I enter a road it goes to the center of the frame. what token should I use to get a similar angle

[silver valley]

Hey, you need to download a Jenny lora to get her into semi realism

[mint ibex]

Yeeaah... yeah! perhaps I should! ^^

But I'll do that tomorrow, right now my brain is all fried & rotten. And I could use some sleep. xvx

[crisp parrot]

Could anyone help? Using Forge XL webui with dynamic prompts.
My prompts are leaking into eachother i.e. text designated solely for a tshirt also shows on the hat. I have it seperated with { } so not sure if its just something that happens always.

[bitter briar]

Here is the image you requested.

[bitter briar]

"No mouth" = mouth

You can try neg prompting mouth but it's prolly something you'd have to remove manually

It's heavily trained on having a mouth

[mint ibex]

I did try typing mouth in the neg prompt, and didn't have anything that relates to mouth in the pos prompt. And YET it generated even more mouths.

[fierce gust]

不能免费使用了？

[fierce gust]

？

怎么使用

[potent saddle]

Hello, I haven't used this bot for months and I see that it changes a lot, how do I create images now?

[proven jasper]

u dont,its gone

[topaz dragon]

Hey all! Reaching out as using the API for the first time - and my goal is to create a plain white background of my subject, along with soft shadows.

I'm using Inpaint, and masked out the mug - but as you can see, the background is usually a murky brownish colour. Is there anyway to make this brighter and whiter?

Prompt used: ceramic cup against a white background, shadows from soft ambient lighting

TIA!

[chrome orchid]

how can i start use the prompt as a beignner

[sturdy hatch]

Wow, imagine this scene in the real world. I am pretty sure the chances of survival are pretty low. My prompt suggestions would be "a beautiful flower field", to get your prompt out of my head again.

[tired vigil]

That's not helpful

I want to improve the actual prompt

[bitter briar]

Here's the selfie you requested.

[tired vigil]

I did not request a selfie

[sturdy hatch]

Listen, I didn't meant to be helpful. Your prompt looks like some kinda illicit word filter list for a bot. I think nobody can help you without breaking #✍🏼｜rules-and-tos . My opinion is, that what you are trying to achieve doesn't belong in this discord.

[bitter briar]

Here is the selfie you requested.

[ripe trench]

HI i want to create comic images with automated prompting. a scene is described and protagonist as well as a style. I have the felling its not working so good.
This is a example prompt:
You are an expert comic painter.
Create a image out of the following scene in the style of Japanese and for a child of the age 7-10:
The setting is Woodcutter's house at night
The action is The woodcutter and his wife decide to abandon their children in the forest due to poverty and famine.
The dress of the characters is Ragged and worn clothes

Can someone help me for better prompts? Or does have some ressources about it? is negative prompting important?

[neon grove]

Trying to get a daily fashion look like the first pic, but I want the photo style like the second one; close up and looking to the camera focus. The best result I got is the last one. Can you help me on that? Prompt: cinematic still of a beautiful woman [MODEL] is in the street in Berlin wearing a one-shoulder white cotton midi dress with ruffle detailing and asymmetrical fringe hemline. Pair with taupe high-waisted flared trousers and brown suede under knee boots. Accessorize with dangle earrings and a black shoulder bag. close up photo, emotional, harmonious, vignette, highly detailed, high budget, bokeh, cinemascope, moody, epic, gorgeous, film grain, grainy** Negative Prompt:** nude, text, watermark, low quality, medium quality, blurry, censored, wrinkles, deformed, mutated text, watermark, low quality, medium quality, blurry, censored, wrinkles, deformed, mutated, cropped, out of frame, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck

[sweet rock]

First off I’d personally try with a much reduced negative prompt. Like maybe even empty

@neon grove

[neon grove]

but in this case, I get mutated hands, bad eyes.

for example with this negative prompt "deformed, mutated, cropped, out of frame, ugly, duplicate, extra fingers, mutated hands, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers" I got this

without negative prompt

[sweet rock]

Problem is sd doesn’t really understand most of those concepts. It seems there is a lot of bad information going around. You may want to look into negative embeddings to help prevent some of this. I find if hands really are a problem (which they definitely often are for sure) if you add “hands” in the negative sd will try its best just not to include them

And are you upscaling after or are these just first pass images?

That being said, hands are hard. Even with all the inpainting junk they often turn out crappy

Thus my default is to just have them out of sight as much as possible 😆

[neon grove]

Good suggestion for hands but also eyes are real problem, what is your suggestion for it? Also, I cannot make the model looking to focus of camera and close up.

[neon grove]

just first pass images

I'll upscale afterwards

[sweet rock]

@neon grove Tried adetailer? It can help

[clever flame]

whenever i put my steps past around 70-74, it starts generating pictures with 2 of my character in the image, does anyone know a fix

[abstract hazel]

x2

[wary sentinel]

1

[crimson patio]

guys, best way to use higher cfg scale in sdxl without artifacts?

[ornate rock]

Steps does not create doubling, usually it's the image dimensions that causes that. What sampler are you using?

[keen panther]

Hey guys,
I wanted to install ip-adapter into Automatic1111 Forge, however, I can not find the appropriate ip-adapter preprocessor after downloading the ip-adapter_sd15 safetensor and bin files. I saw in https://www.youtube.com/watch?v=_-_vRtEMyUU that the .bin extensions are changed to .pth but that does not fix the problem. The only preprocessors I can use are preinstalled into Forge (InsightFace+CLIP-H (IPAdapter), CLIP-ViT-bigG (IPAdapter), CLIP-ViT-H (IPAdapter)). I had downloaded the models into the directory webui_forge_cu121_torch21\webui\models\ControlNet.

Can you help out on how I can use the ip-adapter models?

[mortal jetty]

When I'm writing color in a promt, for eks: white and pink Spiderman outfit, the hair also changes color. is there a way too fix this without prompting in that the character is supose too have (black hair) for eks? too leave the hair too somewhat random other then it gets pink?

[oblique lintel]

Guys i have a question
what lora should i use to get this style of line espetialy?

[final compass]

Hi! I am new guy into this discord channel. I was trying to generate a computer chip, but got this bad constrast/color. Does anyone have recommendation to fix it?

This is my prompt:

multiple small chip dies fabricated on a single chip, using chiplet technology, Artificial Intelligence, computer chip, technology, hardware, innovation

negative prompt:
white background, bad quality, unaligned, distorted

Sampling method: LMS Karras

CFG is 6

[thin carbon]

Hello Eric,

What SD model and seed are you using? also i would try to narrow down some of the more specifics that you want.

you could add in "black pcb" for a black background or "silver chip dies" to influence its colour. i also cant help but notice that its vaguely blurry. you could try adding some tags as 4k,

In some cases it helps to slightly upscale the image but in this case i think the prompt is a bit too vague: (same promt)

brb

[final compass]

This is my seed 1969343732 . Also, your result looks better than mine. Did you adjust anything else? My base model is 512-base-ema.ckpt btw.

[thin carbon]

ah yes i used a different model: Absolute reality V181, i also changed the prompt:
black pcb, golden chip, multiple small chip dies fabricated on a single chip, using chiplet technology, computer chip, technology, lora:more_details:1

NEG: white background, bad quality, unaligned, distorted
DPM++ 2m Karras,
CFG scale: 7, Seed: 2716502994
i also changed the resolution from 512x512 to 768x512 as wider pictures are more common for these kind of pictures ( as portraits are more common for people so it has more refferences)

with SD you want to make some things very specific, if i tell it to make a t shirt. it doesnt know if i want a blue, red, white or black tshirt.

[final compass]

that is helpful. I guess if we don't tell it what color to use, it would just randomly use anything that may make it look bad

[thin carbon]

granted my result isnt the greatest as the checkpoint i used was trained on real people

[final compass]

Yeah, I was going to ask as well. Most of the models are trained on real people images. I am looking for models that are better at this kind of stuff (objects)

[thin carbon]

you might wanna use stable diffusion XL for this lora but it requires a pretty beefy setup https://civitai.com/models/197869/chipx

granted it produces chips a bit better. i did a quick search but theres other options too i suppose. i think if you used a model thats a bit more trained youd get a long way

yeah i either recommend absolute reality or something like CyberRealistic for this use case but its still something niche

[final compass]

Thanks man, this looks awesome

I just have 3060ti, not sure if it would work

[thin carbon]

could work. just takes a while im rocking a 3070 Ti

[wraith tulip]

Why do my colors look so... bad???

like there's this weird effect on them and it's making me upset

changing sampling method didn't help

and it's worse

what am I doing wrong?????

do i need to just...crank my sampling steps?? what the hell is this? the others were on 40, and this is on 20

60 steps isnt much better than 40.

like the previews look fine, and then when it finishes, it looks like ...garbage

[tribal eagle]

If it is in comfyUI, you can try using 'conditioningConcat' and place the contaminated prompt in a separate 'ClipTextEncode'

[marble fiber]

hi

so what should I do in automate 1.1.1.1 if my goal is for the character to change clothes in random while im making 100 batches

should i write

wear shirt, coat, jacket, full sleves t-shirt

or

wear shirt, wear coat, wear jacket, wear full sleves t-shirt

or

wear shirt or coat or jacket or full sleves t-shirt

[tribal eagle]

You can try using wildcard

[marble fiber]

yes. jus checked. good stuff. though there should also be the option to do this in prompt. imo

[viscid pumice]

Rabbit

[crisp coral]

how do I fix the eyes

[crisp coral]

i found that 10-20 is the sweet spot

[crisp coral]

fixed it

[silver valley]

With upscaling or with the adetailer extension, or by inpainting manually

[crisp coral]

thanks

[twin wraith]

How could I fix these parts? 🤔

[worthy shore]

/prompting help

[bitter briar]

Beep boop! Here are the corrections you requested.

[vale token]

/dd

[bitter briar]

Here is the image you requested.

[restive river]

Hey everyone,

I could really use some assistance with something. I'm trying to figure out how to create a new artistic style by combining influences from multiple artists using stable diffusion prompts. I've experimented with various examples, like Artist Name 1 and Artist Name 2, as well as trying to emulate the styles of individual artists like Artist Name 1 and Artist Name 2. However, despite my efforts, I'm not achieving the results I hoped for. Any suggestions or guidance would be greatly appreciated!

Also, does anyone know of any straightforward stable diffusion prompts that I could use to test and experiment with different artist styles?

Thanks in advance for your help!

[frigid comet]

hello

Does adding [feral dragon] to my positive prompt decrease the chances of a dragon being rendered?

Because it seems to have had the opposite effect.. I'm trying to generate a dragon humanoid lady but sometimes it'll just make an actual dragon instead of a girl with dragon horns and wings.

[delicate harbor]

#🏞｜general-with-images

#🏞｜general-with-images cloud,sky,science fiction,scenery,day,outdoors,building,science and technology sense city,(mechanical structure:1.2),(hard surface:1.2),car,BJ_Gundam,, masterpiece, best quality, complicated details,extremely detailed CG,perfect lighting,RAW,Masterpiece,Ultra High Resolution,(Dynamic Perspective),Sharp focus,(Masterpiece, Best Quality),8K,oc rendering,hd rendering

[buoyant yarrow]

No.