📝｜prompting-help

[hollow shale]

Yes

[slate prairie]

#📣｜announcements message

[hollow shale]

Thank you, I see

[slate prairie]

[royal yoke]

❓ ❓ ❓ Anyone have a good, well-rounded set of positive and negative prompts for 2.1? Something that's been honed to work with various genres and sizes of pictures? ❓ ❓ ❓

Please respond directly to my message here so I can see it later in case I don't get to it right away.

[smoky totem]

Thanks but these show what images were used to train SD on recognizing a type of jellyfish (crown jellyfish). I wanted cauliflower jellyfish, and tried many models, many combinations of positive and negative prompts, but never even got 1 image that remotely resemble a cauliflower jellyfish. Even though a crown jellyfish look somewhat similar, I want to learn how to make a very specific prompt of a subject, kind of like having an "adjective-subject" that SD1.5 or 2.1 will recognize and treat them as one subject, instead of giving me a jellyfish with some cauliflower stuck on it. Because next time, I may wish to draw another subject that is very specific that has no similar looking equivalent, that needs SD to really try to search for images of that specific subject.

[slate prairie]

Apologies, I thought they were the same thing. I was just going off wikipedia (https://en.wikipedia.org/wiki/Cephea_cephea#:~:text=crown jellyfish%2C or cauliflower jellyfish)

mmm

[smoky totem]

but is there a way to use quotation marks for prompts in SD that anyone here knows about?

[slate prairie]

If it doesn't pop up in that dataset, then it's likely that it just wasn't trained at all (or, if it was, that it wasn't labelled as such)

[smoky totem]

I tried looking on the wiki

because this time it was a jellyfish, I don't know what trouble I'd run into next time

[slate prairie]

the only way to get it to generate pictures of it would be to train your own dataset or find someone that trained a dataset on the topic

[smoky totem]

I see...so I have to train it if I really run into trouble next time.

Thanks for the tips.

[slate prairie]

yep np

[cursive fiber]

How do we generate images with no bots?

[keen quiver]

https://images-ext-1.discordapp.net/external/3M918jj1L9muJ09DhTZ_Oi9Qvj7XWq7z8KPb99iU9nI/https/s.mj.run/_zjKrEai2N0, retrofuturism style, A highly detailed and hyper realistic painting of a steampunk armour trending on artstation, sharp focus, studio photo, intricate details, highly detailed, by H.R.Giger, Tvera and wlop and artgerm and greg rutkowski, insane detail, oil on canvas, moody, eerie, mystery, artstation, sharp focus, cinematic lighting, studio photo, intricate details, highly detailed, concept art, trending on artstation, sharp focus, studio photo, intricate details, highly detailed

#https://images-ext-1.discordapp.net/external/3M918jj1L9muJ09DhTZ_Oi9Qvj7XWq7z8KPb99iU9nI/https/s.mj.run/_zjKrEai2N0, retrofuturism style, A highly detailed and hyper realistic painting of a steampunk armour trending on artstation, sharp focus, studio photo, intricate details, highly detailed, by H.R.Giger, Tvera and wlop and artgerm and greg rutkowski, insane detail, oil on canvas, moody, eerie, mystery, artstation, sharp focus, cinematic lighting, studio photo, intricate details, highly detailed, concept art, trending on artstation, sharp focus, studio photo, intricate details, highly detailed

[tired vigil]

So um, I am new with Stable Diffusion and currently running into trouble: how do you describe this hairstyle?

[silver valley]

Brown hair, long hair, curly sideburns,

For the eyes its heterochromia

[tired vigil]

oh damn, it's work ! thank you so much. The eyes color is wrong tho lol

[silver valley]

Np, looks good!

[fair girder]

sd doesn't do as i say. "no background" "Empty background" and on top "background" as a negative and i still get background :/
ideas?

[smoky totem]

all black background or all white background?

[fair girder]

will try

[smoky totem]

all white/back background as positive not negative prompt. empty/no as positive as well

[fair girder]

better. still shadows even though i have em as - promt but i will simply change cfg to try.

[burnt notch]

How these 3 different arts called and how you can trigger them in SD. What prompts or is there any model for art like these or a specific name how they different arts called?

How these 3 different arts called and how you can trigger them in SD. What prompts or is there any model for art like these or a specific name how they different arts called?

[lusty crow]

Not tested, but try "Oriental abstract landscape illustration", "Japan traditional sumi-e painting"

[hidden saddle]

Tomb raider,rain, train behind,forest,guns,

[lusty crow]

"circular canvas, round white frame, oriental abstract Japanese landscape, illustration, foreground sakura tree, background Mt Fuji"

Not even close 😄

[burnt notch]

Damn but nice try, maybe there is a model for this or a Name how These Art called?

Tried create Pictures in the Style of the second even Not Work

[lusty crow]

Try https://civitai.com/ and the Illustration area.

[burnt notch]

Thanks i will watch in this section when im at home

[lusty crow]

I can not get close to the style @burnt notch asked for, but Sumi-e style and Ink take me a bit on the way, but I will try this a bit later, the simple prompt was: "fantasy landscape river mountain temple pink sakura, ink (sumi-e style), sunset"

[burnt notch]

Alright! Thanks alot. I try out now some prompts and models 🙂

created already nice one 😄

damn

if you wanna try out use these model

https://civitai.com/models/4694/lessmagifactorygreater-t-shirt-diffusion

[burnt notch]

works perfectly now 🙂 but anyways thanks for help and if you have another solution with better results, tell me @lusty crow 🙂

[hidden saddle]

/unflip forest, train,sky

[lusty crow]

@burnt notch I just had to take a break, I thought I got closer but I ended up with this and understood that I had forgot the objective.

[burnt notch]

dont worry 🙂 nice which model do you use?

[lusty crow]

The last one... Delibrate v1.1 I think

Normaly I work in SD 1.5

[burnt notch]

okey, i try out different models at the moment with same prompt and different sampling methodes and let generate each 50 pictures and then ill do fine tuning when i found the best in my mind

at the moment is dreamlike-diffusion 1.0 checkpoint the best but i have some left

[lusty crow]

Nice.

[burnt notch]

works great with this model

[lusty crow]

Indeed.

[burnt notch]

but i test some more and some fine tuning 🙂

What was your last project, that you created?

[lusty crow]

I mostly do simple prompts to spit out ideas, last thing was something like "small studio apartment, scandi style, sleeping nook, natural warm colors, bed decoration pillows", then just let it make 50 images to see if I can find something to inspire for fixing at home.

And as always, backgrounds.

[burnt notch]

awesome! as printable for home?

Dreamshaper also good model for that creations with that style what i wanted

looks more detailed and more sharp with dreamshaper

and still wanna try 6 different models and each model 3 different sample methods and each 50 pictures lets wait whats the ultimate 😄

[lusty crow]

I so often have no plans, I do and test until I get an idea, I save images for later, rework them, add styles at random and just do prompt spamming, like "master religious art, in the Candy Land a bunny follow the (lollipop road) to the psychedelica rainbow forest where a funny bunny female does classical art pose for a evil plastic tyrant ", I am new to this, so no prints yet, but dound some persona texture work for a art Gallary in Sansar and VR-chat.

[burnt notch]

oh great!

Im not into VR at this time 😄

[lusty crow]

All those silly random things are good to keep creativity going, it most pointless BS 😄

[burnt notch]

yea i feel you 😄

imagine what comes with the next years

with all this technology

Can i ask where you from?

[lusty crow]

No. But I live in Sweden.

[burnt notch]

Okey 😉

[lusty crow]

Last for today, now I need to do important things at home... Netflixing.

[burnt notch]

which show you watch?

[lusty crow]

The Good Place, for the philosophical question it actually have baked in with the simple story, Alice in Borderland for I need something vanilla, Dead End, for some reason I cant explain, then it is mostly the Movies, picked at random.

[burnt notch]

alright 🙂

[lapis pebble]

panda realistic

[fiery estuary]

Which negative promp should i use to not get repetitive stuff? Like a set of mountains behind a mountain etc

[lapis pebble]

hola | panda

[burnt notch]

doubled mountain or cloned mountain something like this maybe?

Anyone know an generally prompt to get better quality or sharper paintings? Just use "sharp"?

actually i create my picture with 512x768, maybe higher resolution would give better output?

[undone rivet]

Hello everyone, I would like to make a question, maybe a little bit dumb

The seed

The purpose is to keep the same pose of a character, or to keep the same character....?!

[blazing quartz]

it's to generate the same image if everything else stays the same. Note that any change to other parameters can affect the output. For me, it's useful to keep the seed the same to check how other parameters would affect the image, like sampling steps.

[lusty crow]

The seed generate a "starting pattern", but change in prompt or program can alter the outcome of the image. That is how I understood it.

[undone rivet]

Understood, so in that case, how can I keep a same character in different poses y clothing???

[lusty crow]

That is a good question, I do not know that is possible unless you create a own model with that character.

[whole axle]

/imagen

[burnt notch]

what you guys use for upscale?

[sweet nimbus]

Hey so is it better to just type it dumb style like "Korean Woman walking towards an elevator in an indoor mall" or write it that looks like tags such as "beautiful korea, woman, motion, mall elevator, perfect quality, detailed, crowded"

[silent marsh]

run a comparison and see. the latent spaces are largely uncharted

[burnt notch]

Doese ever Batch have an different seed or every picture? So is it better when i wanna generate much different pictures to lower batch size and higher the batch count or set batch size to 8 and then betch count lower?

[burnt notch]

and can anyone tell me the difference between DPM++ SDE Karras and DPM++ 2M Karras?

[sharp storm]

I would also like to know too

DPM++ SDE and 2M Karras are the ones I have had the most success with.

[burnt notch]

I use the most time Euler_a, DPM++ SDE and 2M Karras and DDIM

[sharp storm]

ive had to many variables with Euler_a and DDIM

i have to hard of a time getting anatomy correct with those, that is why I use DPM++. but I am no expert

[silver valley]

EulerA, DPM++ 2M Karras and SDE are all good, they are just different sampling methods

[burnt notch]

aaa okey depends what you create yea 🙂

[silver valley]

So you have to compare them

[burnt notch]

Always when i start a project i do 50 pictures of Euler, 2M Karras and SDE with same prompt and settings and then i do fine tuning 🙂

Thats how i work

Or have another and better and faster idea?

[silver valley]

Yea, you can use X/Y/Z Script

[burnt notch]

never used

Script X/Y/Z Plot?

[silver valley]

There you can generate 1 image with the same settings with the three different samplers, it will then generate a Matrix style comparrison image

Choose X = Sampler

Then choose the samplers you want
Make sure to seperate them with , but not at the end

[burnt notch]

aaa okey interesting! thanks

you use then Y/Z too for anything?

[silver valley]

Sure you can insert cfg scale or steps for example

You also can compare models with this tool

[burnt notch]

almost everytime use 7 as CFG

[silver valley]

Yea same xD

7-12 is the goto

You can make Y = cfg scale and then try 7, 10, 12

[burnt notch]

and then Checkpoint name on z and do all models inside i want and then i need change batch count or batch size? because i always do there count to 50 pictures

[silver valley]

Dont use z for now

Batch count and size to 1 for the test

[burnt notch]

Alright! Does the order matter at x/y/z?

[silver valley]

Its like coordinates

Like a graph

[burnt notch]

Yea just wondering if its better to setup anything to X what is more important or like that

[silver valley]

No that dont matter

[burnt notch]

aa okey good

Damn, awesome

Works great, thank you so much!

[silver valley]

No problem! Its a powerful tool and perfect for comparing anything and all 😄

[burnt notch]

yea really!

U ever tried with sampler, CFG and Checkpoints?

[silver valley]

Yes i also compared vaes and even highres fix upscalers

Oh you mean these three in XYZ?

No

[burnt notch]

why not, are there issues?

[silver valley]

The overview and waiting time mostly 😄 it needs much more images to compare everything with everything then.
Mostly i use X/Y

[burnt notch]

Oh i understand

Not tried anything with resolutions. At the moment i do 512x768

and upscale after

[silver valley]

Thats good

[burnt notch]

or better with hires. fix?

[silver valley]

Did you upscaled with Extras tab?

Then yes Highres Fix is better

[burnt notch]

Im just upscale with gigapixel AI

[silver valley]

You can also upscale with img2img

Ah need to check that out

[burnt notch]

Topaz Gigalpixel AI but dont know if thats the best and upscale always to max resolution 6x

[silver valley]

Okay idk, heard good stuff about it

You should compare it with highres fix and img2img upscale then

[burnt notch]

Oh never tried

[silver valley]

Dont know how its in topaz but in SD you can finetune the Upscale realy good

Also you can download and add Additional Upscalers you want too

[burnt notch]

Oh i should give a try! What setting did you use?

and which upscaler?

[silver valley]

Depends on the images you make

[burnt notch]

oooh thats different, hard to say 😛

[silver valley]

For Anime for example is esrgan4,+ 6b Anime and for normal images i use 4x Ultra Sharp

But i have a lot more xD

[burnt notch]

Alright and then you set hire.fix to esrgan with an upscale to x4? Do you change anything on hires steps or denoising?

[silver valley]

Mostly i do 2x because GPU vram...
Yea 10 upscale steps and 0.5 denois is good

[plucky fjord]

anyone knows how to get images in centre and symmetrical?

[burnt notch]

i used on my last project just "centered" worked fine

[plucky fjord]

ok

when i do on faces it always gives me angled though

m generating some icons for my game

[river violet]

Hi. Any tutorial or tips to generate videogame assets for a point and click game?

[burnt notch]

Hmm dont know, sorry

Never created by myself or watched about

Maybe try to find something on Openart or a model on https://civitai.com/

[agile sapphire]

hey guys any ideas how to get front view of the character using img2img?

i got character which is viewed from the side and I dont like it

😔

[chilly kelp]

I'm trying to do a group shot. I've got some pretty good prompts to get two individual characters generated, but what I'd like to do is generate an image of them interacting. Is there a way to seperate the two descriptions in one prompt, like to specify that a particular piece of clothing or hair color goes on one of the characters and not the other?

[blazing quartz]

might be easier to generate the scene, then use inpainting to change things like hair color and clothing

[plucky fjord]

how can i create something close to my ref image but just different style to it??

example my ref image is this

but the ai is giving result like this

its not centered

i did type centered

these are my promts
masterpiece, armour, ((white background)),((medieval)), single, very detailed, icon,centered

[blazing quartz]

that seems like a job for img2img

[plucky fjord]

yeah doing img2img

its cutting my images in half

[blazing quartz]

denoising too high?

Using your prompt, denoising at 0.5

[plucky fjord]

7 it was

oh i see set it low like 5

[blazing quartz]

This is 0.75

sometimes you may have to generate a batch of images if SD wanders off

[plucky fjord]

hmm you getting better results than mine

[blazing quartz]

try different models, maybe

I have uberrealisticDreamy loaded right now

[plucky fjord]

ok

i was using anything v3

[tired vigil]

Hey, Question about prompting as it relates to image composition. Many of the images I'm producing are awesome, but parts of the image sometimes run out of the frame. What would I have to include in the prompt to ensure that everything in the composition is contained within frame?

[blazing quartz]

you can try negative prompts like "out of frame", "cropped", and so on

there's no guarantees though

sometimes it just does what it wants

[tired vigil]

Ok, thanks. I'll play around with those.

[blazing quartz]

alternatively, you can use outpainting to add in the cropped bits

[tired vigil]

I'm just a week into SD, lots to learn.

[plucky fjord]

is centered giving the character in symmetry in centre?

idk why it always gives darn angle look the one u got

sd is helping to fill in my weakness

i used to and copy for concepts

sometimes would get said that my stuff are copied so kind of annoying

now sd is good saver in that 😄

i am having fun and using for 2 weeks now

first week trying to install this week m getting little more used to it

[tired vigil]

The left side of the image. 😭

[blazing quartz]

what's wrong with it?

[tired vigil]

I'd like there to be more "breathing room" on the edges of the image. I haven't learned inpainting or outpaintiing yet, maybe that is the fix.

Also, can a prompt be crafted so that the images is created with a white border around it, or would it just be easier to do that in an image editor?

I'm using InvokeAI mainly, fyi.

[blazing quartz]

I'd just do the frame in an image editor

[plucky fjord]

image editor like krita wld be easier

or ps

krita free 😉

[tired vigil]

Never heard of Krita. I gave up my Adobe subscription a while ago, I just wasn't using it often enough.

[plucky fjord]

there are things called adobe alternatives

https://www.youtube.com/watch?v=nIZwPKB_KWE&t=9s&ab_channel=Gamefromscratch

[tired vigil]

Thanks!

[plucky fjord]

https://gamefromscratch.com/free-alternatives-to-commercial-graphics-applications/

doc version of it

btw thanks for the help its getting improved result after i changed to another model

[burnt notch]

Is anyone here and can check one prompt of me and can help me with one thing?

[cedar cedar]

fantasy world

[atomic flume]

what model did you use for this one ?

[dusk hill]

is there a way to get this type of image using stable diffusion?

[vocal arch]

Peter Paul Rubens #58
mythological painting
landscape art
religious art
animal art
portrait
genre art
diplomat
retrieved
religious painting
flemish baroque painting
baroque
portrait painting
history painting sacred sunflowers, mongolian goddess, emiting sacred sunflowers, mongolian goddess, emiting mystical light, intricate, elegant, highly detailed, my rendition, digital painting, artstation, concept art, smooth, sharp focus, radiant light, illustration, art by artgerm mucha ¯_(ツ)_/¯

[burnt notch]

Different models and prompts hard to say

output is so different sometimes 30 pictures not even look like this

[burnt notch]

Is there a "prompt" what makes flowers and clothes designs more detailed?

From this

[vast smelt]

Parents and children share fun with ChatGPT learning AI, Parents and children are smiling and staring at computer screens like photos taken with a full-body camera, Campus where fun is conveyed while chatting, Bright and fun atmosphere, realistic photos, shot with Canon EOS R5 camera with 50mm f/1.8 lens, --ar 16:9 --v 4

[burnt notch]

I generate different pictures but all time clothes looks like this

You cant create pictures here @vast smelt

[vast smelt]

where can i

create pictures

[burnt notch]

Run stable diffusion on your local computer

[vast smelt]

thank you

[burnt notch]

or in the internet

but there are enough tutorials or videos about

[vast smelt]

Please tell me the URL

[burnt notch]

Youtube and search for Stable Diffusion Local

[full dagger]

hey when a prompt has (word:number) what does it mean

[burnt notch]

more important

like 1.5 is more iportant than 0.8

you can also mark that word what u wanna set and use ctrl+ arrow up and down 🙂

[idle glen]

How can i start?

[burnt notch]

With that?

anyone have idea to fix this with a prompt or an negative prompt? i often get pictures like this with different seeds

[signal cradle]

Can you please explain to me why AI can't generate a cyclops? The simplest task,it's not hands where there are a lot of joints

[burnt notch]

What prompt, version of SD and model do you use? I can later try an cyclops

[trim vale]

It’s possible but in midjourney

[trim moth]

doctor preparing for the future

/imagina

#🏞｜general-with-images doctor preparing for the future

[signal cradle]

Since I do not have the opportunity to use the sd on the computer, I used the site to generate the image. For example, I wanted a "rock poster of a blue-skinned cyclops" instead of this gave me nonsense

[burnt notch]

not in SD? 😦

[burnt notch]

What page do you use?

[signal cradle]

I just wrote a promt

[burnt notch]

Yea but which site do you used to generate?

[signal cradle]

https://stablediffusionweb.com

[trim vale]

Nah man

[burnt notch]

Okey damn, so just try generate different seeds and try my luck to best result

[knotty lintel]

Is stable diffusion working with both [] and with ()? Or just ()? I heard before that it works only with ( ) to increase strength of the promt. Or maybe it actually uses [] to strengthen prompt? For Example if I do (prompt: 1.5) it's going to be increased? Or do I use [ ] in this case?

[hasty jasper]

Can anyone give me tips for negative prompts that would steer the ai to generate more graphic images?
ie I'm doing some Warhammer inspired landscapes and especially with the nurgle characters, they tend to have open wounds and what not.
I assume sfw and censored but I'd like to know if there are more.
Maybe PG-13 and such?

[plucky fjord]

how to get good looking eyes?

[crimson bay]

does anyone know how I get these eyes/expression?

[tired vigil]

hi im trying to get better back grounds behind a person ive messed around with a bunch of prompts but im still just getting like morphed background nothing special im trying to go for like super realistic looking but anything i put its just washed out color nothing amazing any prompts for nice backgrounds?

[blazing quartz]

I have better luck getting the foreground subject right, then use inpainting to mask my subject and ask the AI to inpaint the not masked regions

that way I can focus my prompts on the subject initially, then when impainting, focus the prompt on the background

for example, in this picture, I generated the woman first, and added everything else by repeated inpainting

[deft pilot]

Sexy Indian women

[narrow frost]

Hello

Mejdh#

[wheat swift]

delillah, what's it like in new york city?

[narrow frost]

How do I make image

[wheat swift]

I think most of us are using local installs, but there are some web services. Someone more experienced with the community than me should probably field this one.

[narrow frost]

Dang I thought that’s what this server was

[tired vigil]

How do I write a prompt for an animation where I’m zooming into a landscape, but want to avoid hitting into anythinf on the way. I want to have alot of things surrounding the “camera”, but always have a distance to the objects

[icy shoal]

Hi, trying to control/change the eye direction of my character via prompting but doesn't really work... I tried "look left", "look direction", "eye direction", "looking down" and some others.

Any ideas/experience what might work better?

[hazy whale]

some prompts for deforum? i have no idea how to make one

[ionic swift]

How come I dont see the prompt channels anymore? Is the AI gone?

[topaz summit]

is there a way to add weights to key words? like DD?

[tired vigil]

Tips for faces? I hate these deformed ... things ...
I want full body photographies, not close up portraits (these are mostly fine).

[topaz summit]

Ive noticed famous faces work. Anything that it can train on; more data is better. So famous people work great. I kind of look like Christian bale (according to some website) and I used that to help generate art on images of myself. lol

[tired vigil]

Ah damn, that sounds so easy, haha. Will try that.
Thanks.

[topaz summit]

I used this site (https://starbyface.com/), idk if there are better ones. but its an interesting tactic to try. you can also use (https://replicate.com/methexis-inc/img2prompt) which is great for deconstructing images you want to try and copy, and factor that into your prompt.

[prime cave]

New day

[tired vigil]

Thanks, sites are saved 🙂

[plucky fjord]

what are some prompts to get consistent game assets images?

[heavy holly]

can you say what was ur prompt to generate that woman?

[blazing quartz]

let me see if I have the original image still on my drive

[tired vigil]

anyone have any good prompts for super realistic photograph backgrounds behind person?

[blazing quartz]

hmm, seems I've purged it, it was from my earlier experiments. Probably something like "post apocalypse, beautiful redhead woman standing on road, from back, wearing tank top and pleated skirt, holding a large high tech rifle, wearing boots, pale skin", then all the usual tuning words, like "hyper realistic", "detailed", "photo realistic"

I remember generating a batch of 20 or so images and picking the one I like from it

[blazing quartz]

I would take a look at some of the online image shares that have photorealistic images and borrow from their prompts. Usually I'll use something like "35mm photography", "photo realistic", "high quality", "dslr", "photo", and so forth

negative prompts are important too

again, try to borrow from existing

[tired vigil]

ok thank you

[blazing quartz]

people usually have their favorite set of negative prompts

here's a negative prompt I got from somewhere that was useful for me to make photorealistic pictures:

(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck

for automatic1111

other UIs may have different syntax

[tired vigil]

yeah i probly gotta change up my negitive prompt

[old sorrel]

i took a survey the other daay, it seems like dreamstudio is better at giveing you what you prompt for, is that true ?

did SD just come oot with a new version

[tired vigil]

so i keep getting my pictures of people like washed out color any prompt to fix this?

[blazing quartz]

do you have an example?

[broken mortar]

Bro where i can generate images

And whats the command

[blazing quartz]

https://m.youtube.com/watch?v=VXEyhM3Djqg

[lusty crow]

Anybody here have a suggestion for prompt to add image noise? Like grain size in old 1600 ISO film or digital High ISO noise to generated images.

[fading current]

mmmh .. can you please help me with that? i said only one person and in negates more than one person ... but i have so many pictures with twins xD i really would like to get only one girl at picture

and what is this thing with the hands please? xD i have no picture where she is able to hold a magic wand in the hand .. its horrible

the most of the prompts are from a picture i saw at civitai, just copy pasted .. and i wrote some extra prompts to get, what i want

i get so many pictures like this:

[tired vigil]

cosmic night,moon outdoors, hyper-detailed, wide angle shots, beautifully color-coded, insane details, intricate details, beautifully color graded, wide angle, Unreal Engine, Cinematic, Color Grading, Editorial Photography, Photography, Photoshoot, Shot on 35mm lens, Depth of Field, Shutter Speed 1/1000, F/2, White Balance, 32k, Super-Resolution, Megapixel, Pro Photo GB, VR, Lonely, Good, Massive, Half rear Lighting, Backlight, Natural Lighting, Incandescent, Optical Fiber, Moody Lighting, Cinematic Lighting, Studio Lighting, Soft Lighting, Volumetric, Contre-Jour, Beautiful Lighting, Accent Lighting, Global Illumination, Screen Space Global Illumination, Ray Tracing Global Illumination, Optics, Scattering, Glowing, Shadows, Rough, Shimmering, Ray Tracing Reflections, Lumen Reflections, Screen Space Reflections, Diffraction Grading, Chromatic Aberration, GB Displacement, Scan Lines, Ray Tracing Ambient Occlusion, Anti-Aliasing, FKAA, TXAA, RTX, SSAO, Shaders, OpenGL-Shaders, GLSL-Shaders ,--ar 3:2

abstract planet

[drowsy oak]

a panda with kungfu outfit, enjoying a shiny jug of honey

[fading current]

mmmh next question .. why i dont get girls anymore? only hogwarts castles and dupe castles and mor dupes and dupes in one picture? 😮

every picture is like this example now

[silver valley]

to your first questions, models are trained on a resolution of 512x512 or 768x768, generating images that go over these resolution will cause many duplicates, so most people choose 512x768 (portrait) or 768x512 (landscape) and then upscale the image

To the second question, the AI reads the words from left to right and priors the first words it sees, so if you make 1gril the first word you will likely get a girl in the image, also you can use () to give a word more strenght

[fading current]

oh i see 😮 but its really pixeled in this small resolution i guess .. can i upscale it and make better resolutions? i would like to be able to print my pictures in din a1 size ... it is possible without losing quality?

[fading current]

ah thanks, nice to know

[silver valley]

yes highres fix and img2img sd upscale can use latent upscale which changes the image a bit but gives a high resolution output

[fading current]

can i also make this with photoshop? because i dont want that a great picture i find will change

[silver valley]

i dont know if there is a tool in PS

[fading current]

still no girl 😦

[silver valley]

here is a comparrison of 512x768 and upscaled to 1024x1536 with highres fix:

[silver valley]

yea your resolution is still to big

you can also add woman as tag

[static rapids]

i tend to get weird faces with restore faces on, using dreamshaper

[fading current]

mh .. i entered this in my prompts ... ray tracing reflections, anti-aliasing, FKAA, TXAA, RTX, SSAO, shaders, OpenGL-Shaders, GLSL-Shaders .. before i wrote that, i always had girls (twin girls ... but girls ... xD)

[silver valley]

you can also upscale an upscaled image again, i made a 8k image with that

[fading current]

face restore is enabled

[fading current]

how? 😮

when i use img2img the image changed so hard, that the girl looked in an other direction etc .. for example here:

this was my first img:

and this is the img2img picture:

[static rapids]

are you marking?

[fading current]

so its completely different

[silver valley]

yea thats img2img, look at the bottom ther is a dropdown with scripts, there you can choose SD Upscale

[fading current]

sry, i am completely new in stable diffusion and never made ai work before

[fading current]

ok, i found the sd upscale, and what to do next?

[silver valley]

choose an upscaler (for example real esrgan4+) and set the denois to 0,25 and the resolution to 768x768

scale factor on 2 or 4

[fading current]

i try. just a sec

where are the differents between the upscalers?

i have this now

[silver valley]

every upscaler is different, so you have to test them

[fading current]

ah you mean they could change the picture? like in my example?

[silver valley]

set width and height to 768x768

[silver valley]

yes but not like in your example more like in my upscale example above

minimal details will change

[fading current]

ah i see. that would be fine i guess

[fading current]

wouldnt this be a downscale? i am already at 1000x664

[silver valley]

no thats the tile size, the image will be seperatet in multiple squares and every square then gets upscaled

the square size is then 768x768

my upscale has finished

thats upscaled from my image above (1024x1536) and now its 2666x4000

[fading current]

so there, where i have 1000x664 now i have to set 768x768, right?

[silver valley]

yes

or 1024x1024 if your gpu can handle that

it just need to be a square

but if your image is under 1024 i would use 768

[fading current]

okay i see. and what i have to do next?

[silver valley]

if you adjusted all settings you can click generate

[fading current]

i have these settings now

[silver valley]

denois

on 0,25

[fading current]

what does this make, if i may ask?

[silver valley]

denois is how much the new image will look like the source image, a higher scale gets more abstract(more details more changes to the whole image)

so you want to keep it low for upscaling

[fading current]

oh i see

[silver valley]

thats also the cause why your img2img pictures where so different because denois was to high

[fading current]

and sampling steps and cfg scale doesnt matter?

[fading current]

oooouh xD

[silver valley]

yes leave them as they are

[plucky fjord]

how can i get something like this in SD?

[silver valley]

they can matter but shouldnt now

[plucky fjord]

what prompts will get me results like those

[fading current]

okay 🙂

i wait until my picture is rdy 🙂

[plucky fjord]

and what model to use for it

i have tried dreamlike dreamshaper open journey and few others

[fading current]

what will i have to do next to scale it up more, like your 8k for example?

[plucky fjord]

u can just use illustrator to trace it and scale up @fading current

scale to infiinity u like

no limit if u turn it to a vector art

[silver valley]

you can run it again as source in the upscaler or you can use the Extras Tab, there you can upsclae images too, but they just get a higher resolution and the image will not change at all

[silver valley]

try redshift diffusion maybe and merge it with openjourney or dreamshaper

[plucky fjord]

https://youtu.be/w-wMXn-EsGw?t=1208

easy as cake doing that

[fading current]

mh i worked very much with my illustrator cs5 but always, when i turned a picture into vector it was not really nice … no smooth shadings etc

[plucky fjord]

works nice on my AI

m using 2022

[fading current]

ok, that was too much information x) so its better to use the „extras“ tab then?

[silver valley]

depends on your liking i prefer the sd upscale for large images and then use extras after that, dm me for a explanation if you want 😄

[fading current]

oh yeah, that would be so awesome

thanks

[lucid tulip]

What software is that?

[coral socket]

That's AUTOMATIC1111's Stable Diffusion WebUI.

[lucid tulip]

What do you find better Automatic1111 or Invoke?

[coral socket]

I never used Invoke.

[restive perch]

hello guys, I logged in after many days, a lot have changed here, I can't see the channel where we used to prompt

[shadow spade]

does anyone know why some models turn out super saturated images? like this

in the smooth preview before it ends rendering it looks like it's going to be great

but then it does this super saturated image and i don't know why

[manic crow]

Can someone tell me why SD seems to mix up my color prompts, even with a high CFG? EG trying to do a 'Homer Simpson', with "white shirt, blue pants, black shoes".. and I get all white, blue shirts, white pants... the colors are generally all there, but never where I actually ask for them, even with negative prompts to remove the unwanted color/clothing combos. What can I do to get that right? I don't want to use img2img if I can avoid it.

[shadow spade]

i see, i'll try that

[lapis coral]

Is it possible to get a nice looking face out of a character that's far away from the camera? (without having to edit in post or use img2img) It seems like they're always at the very least ugly and lacking detail if not completely twisted

[shadow spade]

try adding perfect_face?

also maybe tick the "restore faces" option?

[plucky fjord]

nice outcome btw would be nice to get it to turn around and make a 3D model from

[plucky fjord]

can you share the prompts and what model used?

[shadow spade]

if you download the image and put it in the png inspector you'll get the full tags

[shadow spade]

Negative prompt: low resolution, sharp image, burn colors
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7.5, Seed: 340174865, Size: 400x656, Model hash: 79886b6484, Model: babes_11

[plucky fjord]

oh babes is a good model?

i am using dreamlike and open journey or stuff that makes me game asset stuff

m using sd or this ai tools to generate refernce for my 3D modelling or 2D vector game art 😄

[shadow spade]

neat

i'm just using it to generate fanart for webserial characters

[plucky fjord]

yeah easy way to get my colours lol

before people complained i used some one else art to make stuff

now they won't complain m using AI hah lol. yeah i suck at concept art lol

m just good with blender and vector art tools

m still looking for the best model to get game art concepts

[shadow spade]

guess it depends on your game

[upper thistle]

I want a little help for prompt to generate image with stable diffusion

a human shape suit and head is replaced by round shape tank where fish live and control whole suit and image is in portrait like this

[blazing relic]

a man or rather the spirit of a man sitting on couch confined in a torn down home, with blood on the floor, the scene captures the pain heartbreak and mood of the dark vibe, the pristine light outside the windows showed the outer beauty of moody but inside its white walls is actually suffering and distress, cgsociety, serial art, cinematic, aftereffects, cinematic view

#📝｜prompting-help a man or rather the spirit of a man sitting on couch confined in a torn down home, with blood on the floor, the scene captures the pain heartbreak and mood of the dark vibe, the pristine light outside the windows showed the outer beauty of moody but inside its white walls is actually suffering and distress, cgsociety, serial art, cinematic, aftereffects, cinematic view

a man or rather the spirit of a man sitting on couch confined in a torn down home, with blood on the floor, the scene captures the pain heartbreak and mood of the dark vibe, the pristine light outside the windows showed the outer beauty of moody but inside its white walls is actually suffering and distress, cgsociety, serial art, cinematic, aftereffects, cinematic view #📝｜prompting-help

[haughty cobalt]

image

[ancient flicker]

If someone is interested I created a chatbot on character.ai that can generate prompts. Just tell it some key words and it will try to generate one. If you want a negative prompt, it can also create one. I recommend this chatbot for experimenting. You can include some parts of your own prompt and it will try to add some more details to it (I will also try to update it in the future for more accurate results)

[frail hamlet]

Hi friends. Can anyone explain the correct settings and prompt style for inpainting new subjects into a scene? I am trying to make a second kitten in this photo with the prompt "close up of a kitten in a field" but it just changes the grass and the trees? I am using an inpainting model, and denoise strength is set to 1.

[robust musk]

Beautiful girl in cg render

[heavy holly]

hey please anyone can share prompt that looks like photos from Unsplash/envato elements etc

taken with professional camera

[eager pier]

I dunno what that looks like, if you'd provide examples instead of making us run errands to help you. More people would be inclined to chime in mate

[vagrant swift]

can someone help me render this photo:

i made it with dalle 2

and the prompt was something like : Minialist 3d render of a cartoon man wearing headphones in a orange room

any way of making it in stable diffusion

[heavy holly]

as I said below, look of the stock image, photo taken with professional camera.
Example,

as you can see there is some castle, winter lamdscape etc, they are different but both share that natural camera look

[heavy holly]

the problem i am getting is images often have weird palette of colors, not sharp edges, not realistic textures etc

[full sapphire]

I had the same problem, I found out that it’s the cfg scale that do that, less on the scale, less color and if you go to extension you can install, cfg scale fix. It generates images before it turns to shit 🙂

[proven cedar]

try "two little kittens" instead of "a kitten", sometimes it works like that

[frail hamlet]

It seems to only work when prompting for two of anything. For example: two laptops in a field

[simple palm]

I am just now trying my hand at landscapes. Any tips to get realistic HD images of natural landscapes?

[fallen owl]

/prompt

[glass furnace]

Check out #1072220168534642768 🙂

[silver tree]

It's odd that no matter what I try, even some prompts on the internet, all my images come out looking horrible. It really does feel like it's just taking what I ask for and going, "Yeah naaa... you don't get that."

When you finally do get a decent image, is there a prompt trick to not getting the head cut off? I mean.. to tell it how to keep the bust/head in frame about 2/3 of the way into the image?

[crimson ether]

I want a little help for prompt to generate image with stable diffusion

a human shape suit and head is replaced by round CEO shape tank where live and control whole suit and image is in portrait like this biomédical

[smoky totem]

Anyone has good experience and advise for creating a image of merging 2 subjects? I tried to create a castle shaped hot air balloon, but all SD does is drawing either a balloon holding a castle in the bottom, or a castle with a tiny balloon on top, or sometimes just straight up drawing just balloons and castles, ignoring the other subject:

Matte painting, photorealistic, ((side view of a very large castle hot air balloon)), centered, bohemian, desaturated colors, pastel colors, art deco, architectural model, insane details, intricate details, hyperdetailed, plain background, sharp details, sharp focus, soft lighting, soft shadows, 100 megapixels, fujifilm xt3, hyperdetailed photorealism by simon stalenhag

tried also: hot air balloon, shaped like a castle ; a castle shaped hot air balloon, balloon transformed/morphed into a castle etc

i even tried to put negative prompts castle and balloon so it wont straight up draw those

[silver tree]

Not ssure.. are you wanting a hotair balloon that looks like a castle?

[smoky totem]

yes

[silver tree]

I'm new.. like started today new, but I'll give it a shot.

[smoky totem]

CFG from a relaxed 7 to very strict 20

[silver tree]

Not sure what that means.

[smoky totem]

same prompt in midjourney gives this:

[silver tree]

hehe.. what is that. A magic the gathering card?

[smoky totem]

sorry the prompt cut off the image of the bottom of that balloon

[smoky totem]

but midjourney is just so much easier. very lazy prompting yield surprising results. I try to reproduce this on SD. No amount of prompt engineering, CFG scale, different models: dreamlike, protogen, realvision, and others produced anything even remotely correct, let alone getting towards midjourney's aesthetics.

[silver tree]

Closest I've managed to get from the SD illuminati_diffusion model

A hot air balloon in the shape of a European castle, [2] its grandiose design resembling a classic medieval fortress, complete with turrets, a drawbridge, and a crest emblazoned on the side

[smoky totem]

I cant even get to yours (see the long prompt I had up there). I have castle below a balloon, or a big castle with a tiny balloon up top, never a castle within a balloon, and a far cry from the midjourney marriage

[silver tree]

Added that little bit and got that.

I heard there's an openjourney model you can put into the SD system, but I haven't tried it.

[smoky totem]

I tried openjourney before, it tries to get the "aesthetics" of midjourney, but it is not as "smart" as midjourney if you know what I mean

[silver tree]

Haven't used midjourney so not really sure. People have gotten the illuminati model to do some amazing things. So far.. all mine come out looking like they belong in a Monty Python movie.

[smoky totem]

midjourney just understands language better. when I say hot air balloon castle or castle shaped hot air balloon, it understands that I want something in the shape of another

I can never get good "morphing" images in SD. I can do all those characters wearing another character's armor, etc... or jewels on some objects type of transformation on SD, but never a total transformation.

[silver tree]

Luck you. Mine always look like nightmare fuel. I can image a DALL-E ai in 10-20 years being sentient and in an interview someone asking, "So do you remember your early years? And what were they like?" THe answer, "Mostly nightmare"

[smoky totem]

you need lots of negative prompts to get rid of badly drawn things in SD

by the way i can't find illuminati diffusion on civitai

i only see illumination diffusion

[silver tree]

Yeah.. but when you put those in, your original image goes away when you generate it with the negative prompts.

[smoky totem]

try use the same seed and cfg etc. your original image may still change a bit, but better than a dog's breakfast image

[silver tree]

@smoky totem When you use openjourney, do you only need the safetensors file?

I mean.. what are the model.safetensors file for?

[smoky totem]

safetensors are safer than ckpt files

they do the same things

[silver tree]

Yeah but there is a mdjrny-4.safetensors and there's a model.safetensers

Do I need both or just one?

[smoky totem]

mdjrny-4.safetensors

i still cant find your illuminati_diffusion

[silver tree]

K.. I'm pulling that one now going to see if it does a better job.

[smoky totem]

only illustration_diffusion

[silver tree]

One sec... I'll send a link in DM.

Ha.. I used the very prompt they had on the page for illuminati and i get this:

Did 3-4 refresshes, looks nothing at all like what they posted. :/

Sorry.. my keyboard is throwing duplicate keystrokes in there, and I don't catch it all the time.

[smoky totem]

I cant even get to your image above where the balloon and castle merged...I dont know your CFG and steps so I cant replicate it

I still get lots of separate balloons and castles, never together. simple prompt, complex prompt, all kinds of cfg and steps, negative prompts, no negative prompts, nothing works. midjourney gives that image with no fuss

[silver tree]

@smoky totem My CFG is completely default from the installation.

I didn't know to capture the seed, or I'd have put it in there.

So I'm on prompthero trying to replicate some of the things there to try and learn how to craft a prompt and what to do after I get one I like.

[smoky totem]

i guess you run SD locally on your pc, what graphics card are you using? i have 1060 only so I use google colab instead, using their gpu

[silver tree]

Tryiing to figure out what this k_euler_ancestral is, as I don't see it in the sampling methods, and all the stuff I read doesn't say anything about it.

[silver tree]

1080 regular.

[smoky totem]

how long does it take to generate 1 512x512 image usually on a 1080?

[silver tree]

15 seconds for 30 sample steps.

[smoky totem]

oh thats not so bad...maybe 1060 can get 1 out in 30-45 seconds then

but google still faster lol 30 sec gets me 4

[topaz anvil]

Is there way to make prompts on this server?

[silver tree]

google? You get google to make images for you?

[smoky totem]

i use google colab notebooks as my webui and use their GPU

thats how i tolerate 100s of failed separate balloons and castle

still cant get what you got up there, let alone midjourney level

[smoky totem]

there is a youtube video talking about the "best" samplers... you will learn something there. I just use the faster ones. There is a summary near end of video showing the fast ones in group a and group b. Samplers within each group show similar results, whilst group a samplers and b samplers look different

If I keep doing this hunt for a castle balloon image, I cant get to sleep. I am going for a shotgun approach, using the whole CFG and steps range in a prompted sequence, X/Y plot, etc to do thousands of images until google colab kicks me after 2 hours or so.

I spent over an hour and not 1 god damn balloon looks like a castle.

[silver tree]

If you have any youtube links or tips, just DM me. I'll gladly take any input anyone is willing to give.

What is a PT?

[tired vigil]

#📝｜prompting-help love

How can I do prompts?

[smoky totem]

Thanks for introducing me to this model... none of the other models, even the popular ones like protogen, dreamlike etc gave me anything remotely close to what illuminati gave me. At least it gave me something I can work with thru image2image for further work. I got up for a bathroom break and found google finally kicked me out but I got images I can work with.

[limpid dawn]

chains coming off a pregnant woman

[modern bane]

trying to make a xilography style of two women both facing the same direction, not facing each other, can't seem to make the model work

Positive:
(black and white) pen xylography comic style of a portrait of (two woman faces) (cheek to cheek:1.1) (both facing left:1.3), with big lips and a ((hooked nose:1.4)), pronounced cheekbone with (eyes on camera:1.2) and light color pupil. Majestic artwork by Apollonia Saintclair, sharp face features, lot of pen strokes, woodcut, long vertical face, lot of details, intricate

Negatives:
naked, nude, nudity, nsfw, cropped head, 3d, render, (realistic), watermark, ((picture frame)), (smiling), (trunk), (half-trunk), (perfect face), simmetric, make-up, ((small nose)), ((facing each other))

still kinda facing each other, any suggestion?

[drowsy vector]

You guys had me curious about the the SD illuminati_diffusion model, tried it and was surprised not too bad at 30 steps, low CFG.

Even lower settings...

[glad coyote]

Noob here. Does anyone have a guide on how to make better prompts? Idk how to use operators and punctuations like "():{}[]" etc

Thankies~

[ancient flicker]

Check ur dms

[glad coyote]

Aye, tysm

[drifting rampart]

could i also get some kind of tutorial for this please? super lost !

[still jasper]

hey can someone help? i'm using DreamShaper 3.31 and it keeps making me some strange results

[fair girder]

any prompt ideas on how to get a full body instead of just parts?
trying to get an alligator but i keep getting only parts of it. "full body", "complete body" doesn't help

[still jasper]

cond:bust up oil painting of brown skin man with black pointy ears basking in the sun sitting in a chair in a porch, casual, thick brushstrokes, warm lightning, unreal engine 5, masterpiece, symetrical, 4 k quality, raytracing, anime, anime painting, trending on artstation
Uncond(negative):shirtless, sweaty, muscular, nsfw, hyperrealistic, photo, villain, ugly, bad anatomy, asymetrical, mutated, melted, bad anatomy, bad quality, bad art, medieval

it keepsgiving me this kind of pic

[ancient flicker]

Add some weights like that (shirtless:1.4) - example

Don't overdo it (like 1.9 or more) - it will totally ruin the image (my own experience)

[still jasper]

i'm using a notebook and currently prompt weightning is unavailable outside of positive and negative prompts

[tired vigil]

{light | dark} skin, should work as a prompt right?

or am i doing this wrong?

[waxen sable]

/SEXY WET ANIME GIRL

[wise meadow]

How would you prompt something like that: https://stock.adobe.com/stock-photo/id/567222022 ? Most of the images generated by Stack Diffusion with my prompts have some problem (3 years, only 1 eye, headless, rabbit/bunny not made of chocolate, not that fancy eggs, etc).

I've already tried to use negative prompts.

[stone igloo]

@waxen sable

FAQ: How do I upscale an image?

At this time, there’s no native upscaler for Stable Diffusion. Your best bet is to download your image and upload it to some other upscaling model, such as ESRGan, SuperRes Diffusion, ClipDrop, or SwinIR. Check the pinned messages in #1003034183716835418 for links.

[blazing quartz]

you can always create a img2img pipeline and/or inpainting pipeline to fix things you're not getting with the initial prompt

[wise meadow]

Thanks for your reply! Can you inform any videos and/or guides about img2img pipeline and inpaint? I don't even know what inpainting is. I've just using txt2img until now.

[blazing quartz]

img2img is to generate images similar to the input impage. Inpainting is to mask out regions of the image and only have SD fill in that region (or outside that region)

[stone igloo]

FAQ: What is Stability AI?

Our vibrant communities consist of experts, leaders and partners across the globe. They are developing cutting-edge open AI models for Image, Language, Audio, Video, 3D and Biology.. AI by the people, for the people. Learn more here

[blazing quartz]

from txt2img and img2img, you have the option to "send to img2img" or "send to inpaint"

basically, once you generate an image you like, you "send to", which then becomes the input for the next iteration

the "pipeline" would just be a continuation of "send to" to refine the picture by modifying specific parts or styles of the image

and each iteration you can change your prompt, so let's say your bunny didn't turn out very well. You can send to inpaint and mask out the rabbit, then just have a prompt that focuses on the rabbit

hopefully that makes sense

[wise meadow]

Interesting enough, thank you!

[wise meadow]

It does, thanks a lot!

[sage zodiac]

Make an young boy showing interest in science and technology

[smoky totem]

illuminati was the only model that i got any hybrid/morphing subjects to work. not even models like dreamlike diffusion, protogen, etc could do it with lots of positive, negative/no negative prompts.

I used weighting too... 0.3-5 balloon 0.9 castle (i am on a colab notebook too, so i dont know if it worked but the images did work on illuminati)

all the other models gave me separate balloons and castles, and all kinds of extra castles and balloons despite negative prompts 1st 2 phrases were extra balloons and extra castles... i dont get the logic on SD sometimes.

[smoky totem]

where did you learn about prompt weighting not working on colab notebooks? I type in weighting and it "seems" to work (look at my balloon castles), although I have not tested them really scientifically.

[still jasper]

i'm aware they do i'm sayin the one i'm using(which is deforum) doesn't have it right now

[vital heath]

#1072220168534642768

[quartz reef]

beautiful girl, straight hair, blue eyes, white skin color, realistic semi nudes

[smoky totem]

@silver tree you also need to use at least 768x768 for a decent image using illuminati as it was trained on SD 2.1 with 768x768 resolution, your image here looks like a 512x512 with coarse details

[late hearth]

Dashboard for financial statements analysis

[real swallow]

@late hearth Please read the announcement here: #📣｜announcements message

[silver tree]

Ahh.. that was probably the issue.

[smoky totem]

i finally got my bunch of balloon castles thanks to you.

you showed me a model that does this stuff well

[silver tree]

I'm glad I could be of help.

[smoky totem]

mushroom space station

finally effortless morphing model

much longer prompt but thats the simple version i typed...at least illuminati understood i want a space station that look like a bunch of mushroom without fuzz...most of the prompt involved style, background, lighting and details, instead of wrestling to get the model to understand deforming a space station into mushrooms

[silver tree]

I'm glad it worked out. I downloaded a ton of SD1.5 models, and am going to start playing with those to see what I can get it to do.

[smoky totem]

how did you find that model? the huggingface 1st page just showed a portrait...a plain portrait. I would not have thought it can do stuff like this

I looked at all the fantasy ones to try this stuff but they cant do this

[silver tree]

When I go into something new.. I dig deep and fast. Looked at things being posted, here in 2-point-1, and other discord channels. Tried to emulate what they did. Found some of the websites where people post pics and prompts for it.

[smoky totem]

oh i see, 2-point-1 channel up there.

this kind of lighting and vignette is very much like midjourney

i did not ask for vignette, illuminati gave it to me. I asked for a mix of sunset:0.4 and night sky with stars:0.6

even this interior, very much like the midjourney ones with lazy prompting

i dont know if they used some midjourney images to train their models but i get those vibes from it

[cosmic bone]

/prompts I'm glad it worked out. I downloaded a ton of SD1.5 models, and am going to start playing with those to see what I can get it to do.

[smoky totem]

can anyone get an image to show a pattern like this? I tried the literal honeycomb ball, "honeycomb-ball", ball with honeycomb, ball with diamond pattern, I just get honeycombs in hexagon, or a real-world honeycomb, some diamonds on balls etc, but nothing exactly like this organic looking, relaxed honeycomb shape. I want to apply this pattern on an organic looking surface:

[wispy portal]

/ faq

[real swallow]

#1072229020520947753

[royal yoke]

❓ ❓ ❓ How to change trigger-words for embeds only?

[smoky totem]

illuminati model, 20 stepsm cfg7, 768x768, dpm++ 2M Karras ; prompt: ((honeycomb-ball)), in a bookshop centered, desaturated colors, daytime sky, insane details, intricate details, hyperdetailed, plain background, sharp details, sharp focus, soft lighting, soft shadows, professional photograph, 100 megapixels, fujifilm xt3, photorealistic, by beeple, trending on artstation
neg: extra balls, honeycomb, blur, haze, (((close up))), (((out of frame))), (((b&w))), (((out of focus))), (((soft focus))), bad quality, worst quality, thick details, drawing, painting, oil painting, render, badly drawn, over saturated, text, logo, signature

not close to the pink paper ball that i want

((paper ball with diamond shaped holes)) replaced the literal honeycomb-ball description ends up looking like this:

more literal honeycombs on a ball instead of like the pink one

[smoky totem]

illuminati model 20steps 11cfg 768x768 dpm++ 2M Karras, prompt: a paper ball, made with soft tissue paper, ((with diamond-shape holes)), inside a bookshop centered, desaturated colors, daytime sky, insane details, intricate details, hyperdetailed, plain background, sharp details, sharp focus, soft lighting, soft shadows, professional photograph, 100 megapixels, fujifilm xt3, photorealistic, by beeple, trending on artstation
neg: extra balls, blur, haze, (((close up))), (((out of frame))), (((b&w))), (((out of focus))), (((soft focus))), bad quality, worst quality, thick details, drawing, painting, oil painting, render, badly drawn, over saturated, text, logo, signature

best i can get is a ball with triangle holes...paper still look stiff and not soft. if i prompt with the word toilet paper, it keeps giving me the toilet paper rolls within the ball LOL. i give up for the night. will have to tackle this later. That elusive pink ball....

I checked laion database, cant find images associated with the literal object called honeycomb ball. I dont have time to train a model just for that pattern on google colab. Maybe I have to wait and see if the database will get expanded and eventually we can just call this pattern out as it is called the honeycomb ball

[scenic pilot]

Hi

I am new to prompting where can i know more about prompts

[dusky forge]

Trying to do something of two elves "in love" or "kissing" but it keeps flagging

[modern hedge]

#1072220168534642768

[smoky totem]

I go to sites like prompthero and playground.ai to look/search for subjects I want to try

[smoky totem]

your service/website may have a not safe for work filter on. your prompt may have triggered some NSFW images so they get flagged and got cancelled during generation

I finally found a way to get close to the pink honeycomb ball I wanted...I stared at the ball, and figured that it looks something like a sine wave, remembering from the math books that tortured me in my youth. So here I changed my prompt:

A perfectly round soft paper ball, with sine wave pattern tissues, inside a bookshop centered, desaturated colors, daytime sky, insane details, intricate details, hyperdetailed, plain background, sharp details, sharp focus, soft lighting, soft shadows, professional photograph, 100 megapixels, fujifilm xt3, photorealistic, by beeple, trending on artstation

not perfect like that pink ball, but much better than the rigid honeycomb-hexagon balls I was getting.

these images represent my rage in my youth with math books, which i can now tell the AI to tear into pieces to make paper balls LOL

also, from Laion image database, I finally found 1 image of these paper balls, called "Wish Paper Decorative Ball Ornament - Off White, Gold, and Grey - Medium", calling it out on the prompt "paper decorative ball" with or without wish or ornament do not work however...nor any combination of those words.

[drowsy zinc]

How to use it?

[green wave]

#1072229057707659404

[agile sapphire]

hey guys, is there any other way to increase weight of the keywords than just adding "()"?

[smoky totem]

(subject:0.x to 1.x) below 2

over 2 i get weird results

[agile sapphire]

tkank you 🥳

[blazing quartz]

you can also brute force it with multiple (((subject)))

the subject:number just lets you have finer control

[smoky totem]

yes, each (is like a 10% or something like that, but I'd run out of prompt digit space with many (((((

so i use the colon and numbers

[heavy holly]

What is the 2.1 equivalent to Greg Rutkowski?

#1005627952261775390 message
This prompt doesn't work at all right now

[raw ridge]

Hi, I'm trying to generate something specific. that in theory sounds simple, but for some reason I keep failing. I would love some help, and I'm hoping this is the right place to ask for it.

I want to generate something that would look like the entrance to the attached tower, a doorway that would look slightly open, and would keep the ominous atmosphere of the original image.

[heavy holly]

k nvm just default sampler is really bad

[smoky totem]

inpainting, just paint the doorway area and do your prompting

[raw ridge]

You misunderstood, or I wasn't clear enough. I want an image of just the entrance

Like if someone would have made their way to the entrance and took an image of the door from up close

Trying to generate such image, I often get either: architectural-style image, where everything is very lit, or a boring building door, or something that is more like a video-gamish cartoonish dungeon-style gateway portal.

But nothing that captures the same atmosphere of the original tower image

[smoky totem]

ominous atmosphere from a horror movie, heavy fog day....put those way up front even with ((( ))) and colon with numbers if necessary, even before you prompt the door

normally, prompt the door first, then describe the scene. if that does not work, i move the atmosphere stuff forward

use gothic door, normally associated with horror, might help. if the atmosphere work by moving those prompts forward, with or without (( ))) and numbers, go back to tweak the door to the style and size and views you want

sorry when you said a slightly open doorway thing, my mind actually drifted off...you know...i was looking for new models to play with and stumbled onto places i should not be, i am off work now.

so maybe i first misunderstood your question.

[blazing quartz]

alternatively, focus on getting the door right first, then use img2img and inpainting to get the atmosphere, or maybe even try the new controlnet stuff

[crimson bay]

does anyone know how to get good tanlines? anything I try, something weird comes out. [anything diffusion]

[tired vigil]

((((more appendages than a typical human)))) how would you refactor this

[candid dome]

Any tips on using multiple Loras together? I haven't been able to get any decent results, usually very blurry and messy, and I've done a lot of tweaking

[safe marsh]

does anyone have prompt templates

[hardy trail]

Have u checked the pins

[safe marsh]

yes

[raw ridge]

It helps a little bit, in the sense that it doesn't look brightly lit anymore, but I now get mostly black and white images, and rarely actually ominous atmosphere
Also: I can't get the surroundings to look like an entrance to a tower/skyscaper, and I can't make the doors slightly open.

for reference this is the best version of my prompt:

(((ominous atmosphere from a horror movie, heavy fog day))) huge gothic doors, slightly open, entrance to a skyscraper, low angle shot | arch: -2.0

[fair girder]

🤔
whatever model i use, i always get breasts jumping in my face.
i added covered cleavage and highly locked noble dress but that doesn't do anything. interestingly enough an anime tenor is actually the most covered one

[raw ridge]

trying inpainting be like:

[fair girder]

yeah...
currently trying to get inpaint upload to work but only getting errors.

sadly people almost never write guides anymore and just make youtube videos :(

[fossil pivot]

hi, i'm playing with stabledifussion since a couple of months, last week i tried to design a character in the style of the movie "bug's life". I try to get a ladybug but... it was almost impossible. Is there any notice that can help with this challenge in the end i made a collage with some correct images (and gimp) and img2img but i want pure promt.. or almost pure

[smoky totem]

you need negative prompts: ((b&w)), positive prompt ((door slightly open))

[fair girder]

what am i doing wrong here? 🤔

[unkempt cloud]

Is there any repository for specific styles prompts? Maybe with some examples. Like anime of specific studio and bunch of prompts for sd to follow them? Basically I’m in search for library of some sorts for this kinda things.

[swift notch]

Hey guys, i'm trying to get oil painted brush stroke pattern. Anybody knows a good prompt?

[river warren]

(((ominous atmosphere from a horror movie, heavy fog day))) huge gothic doors, slightly open, entrance to a skyscraper, low angle shot | arch: -2.0

[river warren]

hi, how do i use prompt to generate an image?

[dense basin]

Does anyone know a prompt that can produce this cartoonish style?

[tired vigil]

That Taskbar is awful to look at, sorry. haha

• I would try less steps, maybe fill instead of latent noise, a bit less cfg
  Maybe also just on point of blood instead of multiple could help too.
  Im not the best at inpainting so far - just my stupid ideas.

[tired vigil]

Like different models, textural inversion etc?
You know https://civitai.com/ or https://huggingface.co/models ? Maybe there is something for you.

[tropic aspen]

Does anyone have experience getting prompts that are "vertical" aspect ratios (like 512x704) to look good? I get a lot of doubling effects like this face for example. (Ignore the fact that it's Epstein)

My theory is that this is caused by the fact that the network was trained on square images so it likes to try and do two squares in the output. I'd love to get better quality outputs of this shape, but so far it's pretty hit or miss.

Question two is one that I'm sure gets asked a lot. Anyone have success with getting stock watermarks out of their outputs? I'm doing this in large batches, so hoping not to have to do inpainting-- if anyone has a prompt solution that helps even a bit, that would be stellar!

[blazing quartz]

That's pretty much it. Try hirez fix or outpainting

[tropic aspen]

hirez fix? I haven't hear of that.

[tropic aspen]

Thanks

[atomic flume]

So I've noticed that the image I'm trying make seems out of focus any suggestions ?

modelshoot style, (extremely detailed CG unity 8k wallpaper), full shot body photo of the most beautiful artwork in the world, (female military guard captain:1.2) in (detailed:1.2) (sleek scifi cyberpunk armor:1.2) lecturing her soldiers, professional majestic oil painting by Ed Blinkey, Atey Ghailan, Studio Ghibli, by Jeremy Mann, Greg Manchess, Antonio Moro, trending on ArtStation, trending on CGSociety, Intricate, High Detail, Sharp focus, dramatic, photorealistic painting art by midjourney and greg Rutkowski, (military parade:1.2), (a lot of soldiers:1.2), (detailed faces:1.2), she has short military style cyberpunk hair, detailed sci fi background, white blanked, (focused:1.2)

Negative prompt: canvas frame, cartoon, 3d, ((disfigured)), ((bad art)), ((deformed)),((extra limbs)),((close up)),((b&w)), wierd colors, blurry, (((duplicate))), ((morbid)), ((mutilated)), [out of frame], extra fingers, mutated hands, ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed))), ((ugly)), blurry, ((bad anatomy)), (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))), out of frame, ugly, extra limbs, (bad anatomy), gross proportions, (malformed limbs), ((missing arms)), ((missing legs)), (((extra arms))), (((extra legs))), mutated hands, (fused fingers), (too many fingers), (((long neck))), Photoshop, video game, ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, mutation, mutated, extra limbs, extra legs, extra arms, disfigured, deformed, cross-eye, body out of frame, blurry, bad art, bad anatomy, 3d render, horses

Steps: 100, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1614103279, Size: 536x384, Model hash: 13dfc9921f, Model: dreamshaper_332BakedVaeClipFix, Denoising strength: 0.7, Hires upscale: 2, Hires upscaler: Latent

[smoky totem]

100 steps is alot, usually between 20-50 is plentiful. If you watch the rendering window you will see how the poses and scene changes as the steps are processed. I will then go back and select an earlier step to get the scene/pose i want

536x384 is low res too. 512x768 better for half body portraits

you can also use 512x768 and click on highres fix for a 2x resolution boost with low denoise like 0.3-4

[atomic flume]

I used hires fix

[smoky totem]

but a base 512x384 is really low

at least start from 512x512 and if you want portrait orientation, 512x768

[atomic flume]

Ah I can't go much higher I get the out of memeory Cuda

[smoky totem]

or 768x512 landscape

[atomic flume]

though if I set it to --medvram

I can get it to that

[smoky totem]

there is also lowvram (i run colab so I am not very sure here)

i have gtx1060 so I use colab, have not installed auto1111 locally yet

[atomic flume]

medvram works great for me but I noticed I got alot more bad images

I'm currently using a 2090 super

[smoky totem]

yeah i suspect your base images are too low res

2090 super should be good enough no? 8gb vram right? you only need 6 as bare minimum

[atomic flume]

when I don't do --medvram it say something is taking up 6gb of vram

[smoky totem]

I regret not buying a new PC during black friday...I was not into SD/ai art back then...there were those 1000 dollars lenovo desktop with 3060 12GB vram

[atomic flume]

I'm thinkg about buying a 4090

[smoky totem]

I'll buy a house instead at 4090 prices lol

i wonder how much the 4060 desktops are coming out at, I have not looked in a couple of months

[smoky totem]

you should be good with 8gb vram on a 2090 super...

[celest ice]

Finally got it to replicate a posemaniacs pose

[smoky totem]

I envy you guys with controlnet running locally

i want to draw some building outlines and details and let SD "render" the surfaces for me

using control net

[blazing quartz]

that's a nice one

[celest ice]

Yeah it seems to like some models better than others, earlier using dreamlike diffusion I got this

[smoky totem]

I've made comparisons using the same seed and other parameters for the same building scene with 512x512 vs 768x768, there is already a big difference in straighter lines and sharper details...which then i can highres fix. I can sometimes go 1024x1024 without the scene/building changing too much. Going beyond, 1280or1536 squared...then the building/scene look completely different, sometimes they still work well even if it is not the look i anticipated at first

[celest ice]

Which is still cool but nothing like the reference image

But with stably diffused wild mix its outputting much better

Actually the scribble model is probably the best results at the moment, like my input was basically a stick figure in a pose and it more or less perfectly replicated it (ignore ugly face and hand horrors its a low res low sample image)

[blazing quartz]

hmm, I couldn't get scribble to work

it placed my character in roughly the correct spot, but didn't follow any of the lines

I may have to try again

[fair girder]

I specifically modded my win 11 to have that taskbar again. Couldn't work with the normal 11 one, too much lost time.
And yeah, inpainting isn't always as obvious as it seems.

[tired vigil]

I understand the left side taskbar, i also changed that. But the whole name visible ... 🤣

[fair girder]

That's the point. You lose so much time when you have like several explorer windows open and can't pick the right one directly

[limpid schooner]

Does anyone know if there is a good way to generate proper looking guns with stable diffusion or if there are models trained for this?
Because most of the time i try to generate people with guns, the guns look like complete trash.

so i just omit guns completly most of the time

[limpid schooner]

rumors say only 8gb vram which is a shame considering 3060 has 12gb 😦

[agile sapphire]

hey guys, does the SF have parameter like "chaos" in the Midjourney?

[tired vigil]

WIN+TAB for that xD

[grand ocean]

Does anyone know how I can get the name of the model used as opposed to just the Model hash in the generation parameters txt file of an image? (using A1111)

[tired vigil]

You want to know the used Model or you want that the explicit clear name of the model is listed in the txt file?
If you just own the hash and want to know the model you can try to search with civitai, the searchbar can handle the hash.

[grand ocean]

Yes I'd like the clear name of the model. I do see people posting the settings and the name is there. I just wonder how they managed to get that in there

[blazing quartz]

maybe they pulled it out of PNG Info instead?

Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 68207042, Size: 512x512, Model hash: c35782bad8, Model: realisticVisionV13_v13VAEIncluded, Denoising strength: 0.9, Mask blur: 4, ControlNet Enabled: True, ControlNet Module: canny, ControlNet Model: control_sd15_canny(fef5e48e), ControlNet Weight: 0.7

that's from PNG Info for one of my images, it lists the model name as well as the hash

[tired vigil]

Nah, its a setting, i send him the answer in #🤝｜tech-support
"Add model name to generation information"

[blazing quartz]

ah, ok

[grand ocean]

Thanks for the suggestion, it doesn't list the model name for me when I load a png in the png info tab. I did ask in tech support but no one answered

[blazing quartz]

Tirak did. You need to turn on the setting, "Add model name to generation information"

[grand ocean]

Thanks I swear I scan the settings and couldn't find anything

[blazing quartz]

no worries, it happens to all of us

I literally debugged the same issue yesterday because I forgot that I solved it the day before, and what the solution was

[tired vigil]

First thing i always do - run through settings and change stuff 😄
Not always the best idea, but sometimes there are a few QoL changes.

[dire crypt]

Calling all AI artists, Prompt Engineers and Synthographers
We are field testing a prompt builder and image rendering tool for people living with ALS, who have lost the use of their muscles. The entire interface operates through eye tracking (and int he future, brain-computer interface). We are asking you to please donate your killer prompt to this important charitable effort.
As you can imagine, we need to minimize the steps to built a great prompt, as it can be exhausting for the eye muscles. So we are introducing prompt presets. These are prompt template that will deliver a sick image in an amazing style, all you have to do is insert the subject.
If your prompts are your secret sauce then maybe this isn’t for you. But if there’s one reason to share a prompt this is it.
These people are locked inside their body. They cannot move, and they cannot speak. This tool will give them a means to communicate their feelings with nuance, work through some pretty dark stuff, and just have some of that awesome dopamine that we take for granted.
Details: We are using Stable Diffusion 1.5, so prompts should use the appropriate syntax for the model.
I’m not asking anyone to share publicly. But if you’re open to it, then please share your prompt preset in the comments, with imagery! Otherwise, use Google forms link to donate your prompt privately.
https://forms.gle/t2rzXzmhrQ3TJg486

[frozen glen]

I'm not finding the dream bot can someone help me?

[smoky totem]

I think SD developers are gradually censoring sex and violence, at least for models 2.x, maybe in 3.x too. I made images of real life jet fighters, and none of them carried missiles and bombs despite I specifically asked for those. 1 of dozens of those jet images carried what look like a gunpod/rocketpod, i think the ai confused itself with those as fuel tanks. All my ai jet fighters love to carry half a dozen external fuel tanks.

[limpid schooner]

in 1.4 and 1.5 guns also looked weird

[smoky totem]

I have not seen the kind of gun-porn images in the SD 1.x community either. Browsing instagram etc, no guns.

[limpid schooner]

they look like distorted stick things

but maybe the training dataset just not ccontains enough weapon image data

[limpid schooner]

someone should finetune a model to better do weapons

[smoky totem]

bad images and cartoonish, but some of the closest i've had

[smoky totem]

I think those maybe censored/rejected by SD

[limpid schooner]

they cant censor or reject stuff, everyone can post models everywhere on the internet

[strong finch]

is there a way to to do a "or" prompt like red shirt or blue shirt?

[smoky totem]

I prompted real life gun models for these cartoon ones, but they dont look anything like the real life ones

[smoky totem]

yes...i forgot how. you can also use x/y plot with numbers so that after a certain number of images, your subject changes from red to blue

maybe search along the x/y plot on google to get some info, sorry i am not that good.

[limpid schooner]

this looks way more than a gun than the guns my S.T.A.L.K.E.R. characters have 😄

[smoky totem]

like i said, i called out real life gun models

use easy to identify names like kalashnikovs, colts, beretta, etc

i was making ww1 guns (even though the images above dont show anything like those), names like bertier, enfield

but ai rendered most of them brown/wood butt stock...so it is doing something right

[limpid schooner]

i did and they looked like mix of metal and wood scrap parts

if i get my hands on a 3090 i will finetune a model to procude better weapon images

[smoky totem]

fight it with controlnet....bring in images of real guns, force ai to conform to their shapes and porportions

[limpid schooner]

i think buying 3090 and finetuning requiers less effort probably 😄

also idk if controlnet works with 6gb vram, never used it

[smoky totem]

share with us on civitai when you are done training a rifle lora

i have not seen a gun lora yet

[strong finch]

i started checking out the x/y plot. well to much for me to handle

well i was gonna ask if there was a place to ask for prompt help but this is it

[smoky totem]

when i first read it, it does not seem too bad...just plugin some numbers. watch a youtube vid on it and follow along. i did it once, not hard. I just rarely has a use for "or" statements with what i make

[smoky totem]

this is the place to ask, just that you ran into me who is not good and noobish lol

check back in a few hours or tomorrow, someone may answer you in full

also, post your prompt and a couple of results so we see what you are trying to do

[limpid schooner]

probably x/y plot and put in x type = prompt s/r , x values = red shirt, blue shirt
y = whatever

https://github.com/AUTOMATIC1111/stable-diffusion-webui-feature-showcase#xy-plot

that would generate one image for red shirt and one for blue shirt, and whatever you put in y, like cfg or something

[strong kernel]

#👥｜roles

[low ivy]

hey, what´s the best way to prevent double faces, arms... ? i use negative words duplicate, ugly, additional arms, additional legs, additional head 🥹

[wheat swift]

here is my entire standard negative prompt

comic book panels, split image, border, seam, signature, ugly, duplicate, morbid, mutilated, extra fingers, mutilated hands, poorly drawn hands, poorly drawn face, mutation, bad anatomy, extra limbs, cloned face, gross proportions, malformed, missing arms, missing legs, mutilated hands, too many fingers, long neck, poorly drawn feet, poorly drawn face, out of frame, disfigured, deformed, body out of frame, blurry, blurred, watermark, grainy, signature, cut off, draft

[low ivy]

thx

[tame tapir]

Is there is a way that you can Prompt two character's explaining who they are individually?

[tired vigil]

anyone knows what prompt is used for nsfw?

[silver valley]

depends on the model, nude and nsfw working good xD

[tired vigil]

no matter what i do, my images are not detailed and out of place

characters arent rendered properly and they end up having deformed body parts

[silver valley]

You need to use negative tags

Like blurry, deformed, low resolution, multiple limbs,

[tired vigil]

oh i never used them

[silver valley]

That will make a big Difference

[tired vigil]

i dont even understand what that means

and how should i use it

[silver valley]

There are two fields

The top for Positive tags and the one below for negative

[tired vigil]

yea i know that but should i just copy the same things in the negative?

same as positive?

[silver valley]

No

Copy my example

You need words you dont want to see in your image

[tired vigil]

oooooooooh

i understaaand now

[silver valley]

Good ^^

[tired vigil]

thx thx!

[silver valley]

No problem 🙂

[tired vigil]

seems like the characters are clear and detailed but in 1 pic there are 2 same chars everytime

[gaunt jungle]

Maybe someone has an idea. All my pictures have the same style of colors like in this picture. What's weird is, some pictures in the midle of the processing are looking good and then so aftger 50% it gets always this colors. Does anyone know, what I'm doing wrong maybe?

[tired vigil]

no matter what i do, my pics are trash

no matter what model nor any settings

it will always generate bullshii

or it will crash and say error

arent there any tuts on prompts?

or anything??

im waiting hours just so it can render a 1080x1080 picture which always turns out into bullshit and nonsense

[lyric badge]

First mistake

You can always upscale later

I don't tend to go above 768

[lyric badge]

At the top left you are using a specific model

Change it

[tired vigil]

so should i go with 768x768?

[lyric badge]

I do 512x 768

Portrait or landscape

[tired vigil]

can i pleaase see your results?

im working with this all day

[lyric badge]

Search ai___portraits on insta

[tired vigil]

this one?

[gaunt jungle]

Hey thanks for your help. I had also the same problems with other models. But now I have seen in a video tutorial that I should not go below 768x786 iage size. This did the trick for me. With 512x512 the weird color distortion is showing. With bigger resolution it's okay.

[lyric badge]

No

[tired vigil]

damn bro ur renders are so good

[lyric badge]

Its all about how you prompt and negative prompts are your best friend

[tired vigil]

brooooo

[lyric badge]

They were done on 1.5 with out negative prompts

[tired vigil]

crazy

[lyric badge]

https://tenor.com/view/yes-wink-gotcha-you-tube-thanks-gif-3467082

[tired vigil]

look at my shit

i cant anymooooooore

[lyric badge]

What's your prompt and cfg scale?

[tired vigil]

cfg scale 4

prompt : beauty, face

[lyric badge]

Keep your cfg scale between 5 and 7

[lyric badge]

Too basic

[tired vigil]

hmmm

[lyric badge]

The more descriptive you are the better the outcome

[tired vigil]

really?

[lyric badge]

It's key word based

[tired vigil]

damn didnt knew that

[lyric badge]

So if you wanted a pretty face you would write something like:

A hyper realistic portrait photograph of a beautiful woman in her 20s with long brown hair and bright blue eyes. Kodak gold, cinematography, 8k, epic lighting.

Then in negative prompts you would write:

Deformed hands, warped features, ugly, nsfw, frizzy hair

Etc.

[tired vigil]

yoooooo

[tired vigil]

ill try to copy paste this one

but rn im rendering one

[lyric badge]

Set the cfg to 6.5, steps to 50

[tired vigil]

i did alien, space, dark, horror, spaceship

but seems like nonsense

[signal brook]

Hey guys I need help here... I have this prompt "Owl flying towards the horizon, over the shoulder shot, twilight, realistic, hyper detailed" but no matter what I do all shots are from the front

[tired vigil]

was at 7 cfg

[lyric badge]

Again not descriptive enough

[lyric badge]

Specify at the beginning, A photograph shot from over the shoulder of an owl flying....

Or A POV over the shoulder shot of an owl....

[signal brook]

Still nothing

[lyric badge]

Increase cfg?

[lyric badge]

Recent render on 2.1

[tired vigil]

craaaazy

[signal brook]

Cfg set at 18

[lyric badge]

Aha nah don't go higher than 9

[lyric badge]

I can only give it a go when I'm home on Monday

[tired vigil]

bro i copypasted ur prompt before and it turned out like this

wtfffff

[vapid moat]

I would say you can probably go up to 12 cfg

[tired vigil]

this is actually pretty goood

not perfect but much much much much better

thx g

[lyric badge]

Yeah just a rough example the more details (tokens) the better the outcome

There is a token limit however

[vapid moat]

try "from behind" as well

[tired vigil]

where

is token limit?

in the prompt?

[lyric badge]

Each key word = a token

[tired vigil]

what is the limit,

?

[vapid moat]

are you using a1111?

[tired vigil]

oh 75

ah i see

damn

hope they upgrade it sometime

[lyric badge]

Doesn't really need more

[vapid moat]

a1111 doesn't actually have a limit but the further past 75 you go the less coherent it stays to the prompt

[lyric badge]

^

[tired vigil]

maybe pros could think of something bigger,, or idk

[lyric badge]

It's "engineering" the right prompts

And trial and error

That's why it's better to render a batch than just one

[tired vigil]

yea true

[shadow spade]

how do i get img2img to change the background of images

I see people on youtube doing prompts and teh background changes with prompts and my doodles remain with a blank background everytime

[blazing quartz]

you can use inpainting, mask out your subject, then inpaint unmasked area

or, if you don't mind changing the subject but want to keep the general shape/pose/whatever, you can use controlnet

[shadow spade]

i'm doing that

i mean controlnet

using the scribble script

[blazing quartz]

ah, yeah, I've been having some issues with scribble, not sure what I'm doing wrong

[blazing quartz]

still need to do more testing

[shadow spade]

it's doing it, but i tell it to draw a kitchen in the background and it does nothing

[blazing quartz]

I would not try to do it in one step then. Feed it to regular img2img or inpainting and create a pipeline process

so take the output, use the "send to" options, and loop around using a different method to put in the background

if that makes any sense

[shadow spade]

yeah

i was thinking on that line too

manually adding details to the doodle also seem to help, but the kitchen still remains unpainted XD

[blazing quartz]

could also try doodle -> canny, or one of the other controlnet models

[vale jay]

Hi, on colab i just downloaded a LoRa, but it seems that it isn't activated, how do you activate it?

[quaint coral]

click that and a new section will appear, click the "Lora" tab and choose the lora you want to use

[hasty solstice]

I’m using dream booth to train a model on a face and when I prompt I notice that the backgrounds are always dull, even when I specify that I want a nice background. Any suggestions?

I’m using SD 1.5 btw

[wheat swift]

do you say you want a nice background or do you describe the background?

[hasty solstice]

I want a nice background with my character and I’m always getting a dull background. Even when I’m putting in my prompt something like “beautiful city background”

[wheat swift]

is there a background in your dataset images?

[hasty solstice]

No just images of a photoshoot in the training set

[wheat swift]

is the background the same in all of them?

[hasty solstice]

Essentially

Yes

[wheat swift]

If it is, that is your problem. it thinks the background is part of your character

[hasty solstice]

Oh ok

[wheat swift]

you need to edit your photos so the background is inconsistant throughout them

[hasty solstice]

Wow

Thank you

[wheat swift]

np

[hasty solstice]

Do you think captioning the images helps? For example “Charactername in front of blue wall”

I never used captions

I only trained like 3 times, do you think captions are helpful?

[wheat swift]

You can try that. I caption all my images

how many images do you have?

[hasty solstice]

Is it supposed to be like that? For example if all my images are Johnny, Johnny (1), Johnny (2).

Then my caption should be like “Johnny in front of a red car”

I only have 10 images right now

But they are all from the same photo shoot so.. same clothes and same background

[wheat swift]

that's going to mess you up every time

[hasty solstice]

Assuming I have different images from different settings. Am I describing the captioning process right?

[wheat swift]

change rooms, change outfitss

yes. captions have the same filename as the images with a txt extension, and be as specific as you can. describe what you don't want the model to train on

if your person is wearing glasses and you don't want the model to see them then the caption mentions them

[hasty solstice]

Johnny wearing glasses in bathroom

Is that right?

[wheat swift]

describe the clothes too

[hasty solstice]

Wow

[wheat swift]

if they are in the picture

[hasty solstice]

Ok cool

[wheat swift]

otherwise, they will always be wearing those clothes

if they are smiling in every picture, the ai will only know how to make them smile

[hasty solstice]

Img

Omg

That’s nuts

Ok

Makes sense to add captions now

[wheat swift]

it only knows what it's told

or what it sees. it you want variation, it has to see variation

if you want proof, ask it to change the mona lisa

let me know when you're done banging your head against a wall on that one 🙂

[hasty solstice]

Makes sense

Thank you this was very helpful

[wheat swift]

I can pop her off the poplar one in 16 times

you're welcome

[hollow lagoon]

anyone know how to stop this with any wheeled objects?

also had similar things happen with bicycles

[sick cipher]

@hollow lagoon try some cars textual inversions they might help

How do I make the AI zoom out ? Like i have a character but some of the hair above the head is out of frame or the legs aren't showing.

[wheat swift]

outpaint all 4 sides and regenerate a smaller size

[sick cipher]

Is there an easy way to install an outpainter ?

[wheat swift]

what frontend are you using?

[sick cipher]

automatic 1111 web ui

[wheat swift]

send to img2img, click on scripts, outpaint mk2

already have one 🙂

[sick cipher]

Oh sick

Damn bro you are a life saver @wheat swift

Is there any other cool things comes with it ?

[wheat swift]

have you looked at the features page? It's loaded

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features

[sick cipher]

Uhhhhh

[wheat swift]

You may have to do some inpainting

[sick cipher]

My dumb head installed Pix2Pix manually not knowing it's supported and the first feature in the page

[wheat swift]

It's a new feature that just came out a week ago LOL

The features page is moving at the speed of light

[sick cipher]

Oh good that makes me less dumb

Do you know why my textual inversion might not have thumbnails?

[wheat swift]

Because you have to generate and set them

if you hover over the ti, there will be a label (link) to set the thumbnail from the active image in the gallery