📝｜prompting-help

[south solar]

I am thinking that it might be because the edited elements do change the prompt length, and maybe it changes where other tokens end up in prompt, but I really don't know enough about how the program works to say with any confidence.

[pine spindle]

I dont think so.

I do it all the time. I understand that sometimes it feels that it does, but I'm pretty sure it's just in our heads.

[sour scarab]

is there something special with , commas or can I just separate everything with spaces with the same effect?

[pine spindle]

It won't be the same as everything you write will change the image somehow.

But no there's nothing special about commas.

A comma is still 1 token.

[sour scarab]

sure but it's parsed and the webui has different syntax from dreamstudio right?

[pine spindle]

I don't know, I never used dreamstudio.

[austere perch]

how do i remove the white bands with negative prompt ?

[feral fjord]

Go to paint and crop the picture

Or photoshop if you're feeling fancy

[feral fjord]

You can also try words like "wallpaper" "full resolution" "full image" "borderless fullscreen" or stuff like that

[indigo peak]

hey all, new stable user here. Would anyone know why my images are appearing as so....
I'm using Dark_Gemini V2 model and it seems that a few of the models i've downloaded including this one are giving the same results. Is it just a bad generation luck or something more crucial?

[frosty summit]

You probably need to use a VAE. Looks like that model has a required one.

[indigo peak]

A VAE?

[indigo peak]

is it a model?

[frosty summit]

No, it works alongside the model. It goes in Models/VAE folder. You can download it from here https://civitai.com/models/6209/dark-gemini-v21

then select it in your SD

[indigo peak]

ohhhhhh okay okay I see, i'll give it a try and hopefully all works well, yeah I noticed a few of the models had that so i'm assuming it's just VAE related? thank you for helping, a lot to swallow when doing all this the first few times

[frosty summit]

That's generally why images might turn out too psychedelic like that. I know it's a lot haha

This VAE file just fyi

[indigo peak]

ohh I see yeah I got that one, and its right that it's a ckpt. files, just throw it in the VAE folder. Yeah def will take time to get used to all of this

I've put it in the folder but can't seem to find it in Stable itself? sorry for asking so many questions

[frosty summit]

No problem, are you using WebUI?

[indigo peak]

hmmm yeah I believe so

[frosty summit]

I can't remember if it auto chooses the VAE by default but I like to go to Settings > User Interface and add these to the Quicksettings list

then you get dropdown boxes right at the top of the WebUI

to change models and vaes

[indigo peak]

sorry bout the late response, i think I have it sorted? I realised that I had the 1.5 pruned-emaonly.ckpt and not the full 7gb version, so I'm redownloading that

[frosty summit]

That shouldn't make any difference. You won't need that VAE with the 1.5 model anyway.

[lone badge]

by the way, any option is possible to add up top, as long as you know its name. The full list of options and their names is here https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/master/modules/shared.py#L311

[indigo peak]

ohhh okay okay, I see, I think I sorted it out? I'll keep playing around with things, how or what do i write to add vae to the quick settings list?

[frosty summit]

sd_vae 🙂

[indigo peak]

okay okay cool!, alright i'm so sorry about taking your time, it's very daunting doing this the first time, but thank you so much for helping, I'll keep going at it and hopefully i'll get a good result soon, thank you again for the help guys!

[frosty summit]

No problem at all - that's what this place is for! It's fun once you're familiar with a few things. I have lots of models and really none of them require any special VAEs so it's just some out there that make things tricky.

[indigo peak]

yeah thats great to hear that not everything uses VAE thank god, but again thank you guys heaps for the help, it's done a lot!

[silent copper]

is there any way to feed SD a color palette?

[tired vigil]

I've found that feeding a solid color image or color themed background into img2img and then prompting works

[silent copper]

oh interesting

say if I wanna make a character with a few colours, could I then just feed it a blob of colours

[proper geyser]

creative design logo vector a simple heart minimalist with zaha hadid inspiration

[autumn widget]

Hey, I have a problem.
In txt2img I set a Batch Size of 4, with Seed -1.
Then i choose a generation I like and grab its Seed to focus on it.
So I set a Batch Size of 1 and set the wanted seed
Result : the image is completely different. What I did miss? 🤔

Sometimes I get the image back when playing with the Batch Size. It’s like it has an influence. Any idea ?

[autumn widget]

After little more research it seems to affect SDEKarras only lol rip

[teal slate]

Heya, I've got a bit of a strange issue here. I'm trying to generate a sheet of different angles of characters, having tried both charturner and a controlnet turnaround image. The issue is though, that for some reason it just about refuses to give me the front angle. In fact, it's very adamant on showing me butts.

Even if I have (((front view))) as positive and (((back view, ass focus))) as negative, it still just gives me butts.

Sometimes I managed to get the body from the front, but then it had pasted a butt where the face should be.

It's almost like there's some invisible embedding which is 100% butt focused that has snuck in

[smoky totem]

America's Ass. You generating Captain America?

[smoky totem]

Serious answer: are you using a lora for that character? Some loras are trained to bias a certain angle/view (depending on the dataset images used)

[teal slate]

Nope, no LORA. Using RevAnimated_Reveal as checkpoint

[tacit shell]

I'm trying to generate a character that has a more sunken-in features and has lost a lot of muscle mass/body weight, but no matter what I try I can't get it to give me anything other than pretty healthy looking guys. Specifically I was using (gaunt:1.3), (frail:1.3) to try and accomplish it, but is there a better keyword I could use that would work better?

[smoky totem]

did you look at the seed number to make sure it is -1? (I've made that mistake before wondering why all images look kinda the same)

kinda easy to overlook sometimes

[teal slate]

Yep 🙂 I've gone through all the beginner mistakes twice over

[smoky totem]

try an img2img using some front view, refresh/relaunch auto1111 and try that, i find that sometimes clearup some stubborness

[teal slate]

The thing I'm not getting is, ControlNet limbs have different colors, so you'd imagine that it could differentiate a front view from a back view based on those colors, but it seems it's not doing that for me. If I have a T-pose it's always from behind

[smoky totem]

try referencing a known gaunt character, then scale back the :#.# on that character to get what you want without blatant deep faking.

also put gaunt way up front in prompt right after you call the subject

[dusky aspen]

i have a super fast question and then i'm off. i have a really simple prompt that just has a couple of mentions of her outfit. and in txt2img i am getting the same generations. for almost every image. Which is fine, but I am a bit confused. What do I need to do to get varied images?

[gusty veldt]

Yo ya'll, I've been trying to generate something like this with Stable for so long now but I can't seem to figure it out

I just want to generate a full body creature like this with a white background

[smoky totem]

try this lora:

https://civitai.com/models/13531/monograph

[gusty veldt]

Yoo, I'll try that, thank you!

[smoky totem]

will give you some nice pen strokes

[gusty veldt]

exactly what I'm looking for, thank you!

[smoky totem]

if you get it right, show us how.

[gusty veldt]

will do, of course

getting somewhere, ig

[smoky totem]

there is a v1 version (older?) that the author showed some images of white background with just the subject. Look at the user's images too. Pretty sure you can nail it.

[smoky totem]

yea, getting there. just do a bunch of styles and subjects and keep re-rolling

[gusty veldt]

yup, downloading it right now, tysm

hmmmmmm

I CAN use controlnet and just put the image I want to re-create but I want more variations lol

[muted steppe]

hello

can someone tell me what is the difference between tags and trigger words?

[tired vigil]

From a training staining standpoint or from someone who's just using them to generate stuff?

[grand juniper]

anyone have any tips on using the SD upscale in img2img?

I feel like it barely does any variation since if you go over like 0.25 you end up getting a trippy collage

[muted steppe]

just basic edits to generate stuff

on civit both are included which confused me

[tired vigil]

Tags are used when someone is training something, say you have an image with a cat and a dog in it. The person training it would have three tags in it "dog, cat, cat and dog"
trigger words are used when you want to enable an embedding or lora. Say you download an embedding that makes everything darker, to have that embedding enabled in your generation you would put whatever trigger word(s) the creator has created to your prompt

[muted steppe]

do i also need to add the lora

lora:fashionGirl_v47:1

[tired vigil]

yes

[muted steppe]

noted

can i use multiple loras at the same time or do they "conflict"

[tired vigil]

You can use multiple ones at the same time, they may conflict. however if you want you can lower the strength of any particular lora by doing this.
lora:fashionGirl_v47:0.5

Some lora's are more forgiving than others

[muted steppe]

noted

[lone charm]

ayo, does anyone have decent negative prompts for environments/objects? Most/all that ive found are made for characters

[lean island]

Honestly, I just asked chatGPT to “create a very long list of single, negative prompt words, separated by commas, I can use for Ai Art that describe poor image, bad, blurry, “add a few more”

/shrugs

In your case I’d add environment/obj to what you’re asking it

[undone rivet]

Hello everyone, I would like to ask you something, regarding Stable Diffusion ⚠️

Is there any guide, to learn about the right way to build a prompt...?

I´ve been learning a little bit and generating good images but, when I add or remove just "one word", I get misshapen limbs, misshapen faces, or bodies, or when I modify the order of the words

So I would like to learn what is the right way to make a prompt, to recycle it and reuse it, just making small adjustments

[tired vigil]

There is no "right way", but generally you want to keep things organized in regards to the prompt specifications.

[tacit shell]

I'm not fully sure myself, but the creation guide for this model has some nice tips in it. It definitely doesn't have everything since I am still very confused, but it was the first thing I found that actually tried explaining things without being a 30 minute youtube video
https://www.reddit.com/r/StableDiffusion/comments/10uwxni/rpg_v4_user_guide_now_live/

[tired vigil]

"character specification" , "pose" , "at a location ", " wearing a type of clothing, "type of haircolor " , " type of skin " , "photography details"

stuff like that

in regards to the negatives, you can try this
(worst quality, low quality:2), monochrome, ng_deepnegative_v1_75t, blurry, low resolution, low quality, out of focus, painting, cartoon, anime, 3d, cgi, digital art, blender, asian, piercing, old, , tall, deformed, disfigured, extra limbs, missing limbs, fat, curly hair, hands,

this is if your doing realism

[tacit shell]

does anyone know of a way I could get "black fire" in the generation image sort of like how it looks like in the fromsoft games
reference: https://preview.redd.it/1l794907gecz.jpg?auto=webp&s=62948e7b7a452daf5862ffe8934ec668cd6a8ae8
reference2: https://static.fandomspot.com/images/01/11692/black-flame-dark-souls-3-screenshot.jpg
I've been trying things like dark fire black fire black flames ect, but it's just giving me normal red and orange flames. It kind of seems like it just isn't something it can sample, but I was wondering if anyone had any ideas.

[tired vigil]

in the end you gotta experiment on what fits your taste

[tired vigil]

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

This part might help you achieve your goals

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing too

[undone rivet]

Thanks buddy...!!!

[undone rivet]

thanks a lot...!!!

[undone rivet]

and how to keep the face in frame....?

[sturdy scarab]

anyone know what kind of prompt I should use to get this asthetic?

[rapid slate]

Man how can I create beautiful images by using prompt and where can see the prompt section

[dawn lagoon]

Stable Diffusion is not run through a discord ui. you may install and run it on your own gpu using this guide "Install Stable Diffusion Locally (In 3 minutes!!)" by Royal Skies on youtube

[crystal dome]

Tell me how to start using Stable Diffusion and create an image? I have no rights to leave a message in Dreambot chat...

[lone charm]

o ok ty, in my experience chatgpt isnt the most helpful but maybe bc it hasnt learned all the new tricks, this is what bing ai gave:

blurry, pixelated, distorted, noisy, grainy, low-resolution, dark, dull, faded, washed-out, overexposed, boring, plain, unoriginal, uninspired ugly, unpleasant, unappealing, messy, chaotic, inaccurate, inappropriate, offensive, glitchy, broken, corrupted, jagged, stretched, crooked, cluttered

[lone charm]

lookup "install automatic1111 stable diffusion"

on yt

[glossy vessel]

anyone know how to remove text from an image?

[tired vigil]

Prompt: portrait photo of a asia old warrior chief, tribal panther make up, blue on red, side profile, looking away, serious eyes, 50mm portrait photography, hard rim lighting photography–beta –ar 2:3 –beta –upbeta –upbeta

[candid egret]

i want to make photorealistic images, what model should i use?

[lean island]

like a watermark?

[glossy vessel]

yeah

[lean island]

one sec

[glossy vessel]

fyi, simply putting watermark inside the negative prompt doesnt work for me

[lean island]

https://www.watermarkremover.io/?utm_source=upscalemedia&utm_campaign=cross-sell&ref=upscalemedia

[glossy vessel]

thank you, but is it possible to remove it using stable diffusion?

[lean island]

honestly, Ive never gotten an image with a watermark so im not sure

from stable diffusion

[lean island]

What stable are you using that is putting watermarks on it?

[glossy vessel]

oh

[glossy vessel]

im using webui, inpainting (?)

[lean island]

ah, I'm using web-ui txt2img

[glossy vessel]

oh, im trying to remove this text. but it just doesnt want to

so I tried inpainting

keeps on creating new alien texts

[lean island]

is the photo private? I could try on my end if its something you dont mind sharing

[candid egret]

what model should i use for photorealistic image

[lean island]

f99 SUPERB

not superb... it IS superb.. but its not in the name

and

one sec

[candid egret]

can u link it

[lean island]

yea

[tired vigil]

There are many models, I personally suggest realistic vision, deliberate or chillout

[candid egret]

also what should i use for anime?

[lean island]

also gotchu

[candid egret]

thx

[tired vigil]

Abyss orange mix is popular, anything is popular

[candid egret]

which one

[lean island]

sent you a dm

[tired vigil]

up to you. abyss has more nsfw mixed in it if you want

Pastel and counterfeit are notable mentions

[glossy vessel]

it's fine, thank you for the offer, I really appreciate it. I've already changed the base of it

[jolly night]

I can't create

[lean island]

Do you mind elaborating a little?

[vague stone]

is there a reliable way of using inpaint as a "content-aware eraser"? Like photoshop - i want to remove an object from an image. I know you can inpaint and then write a prompt on what should be behind it. But i don't find this to be very reliable, and hard to automate.

Ideally it could just detect the pixels around it and "fill it in" similar to how photoshop's feature works. Any ideaS?

[feral fjord]

Maybe use control net and provide a depth map or a scribble where it doesn't have the object you don't want but it does have the rest of the picture

[vague stone]

does controlnet support inpainting like this?

[feral fjord]

Of course

Though, the task will be easier or harder depending on what you want to erase

Is it a minor detail or a large part of the picture?

If it's something small and you want to have an eraser like the one in Photoshop, you can just try using Photoshop haha

[rustic birch]

Hey guys, i want to turn simple prompts into good (or better) prompts using AI. I already know roughly how I’m gonna do it but I‘m in desperate need of a top tier prompt collection. Like really the best of the best, elite stuff. Does anyone here happen to know where i could find some great prompts? I‘m thankful for any tip 🌚❤️

[hard wolf]

Hi! I am trying to add 2 LoRA in an image, but I have results where they cannot be clearly differentiated, does anyone know a trick to make them differentiate from each other? or some prompt to make the LORAS work better, it would be very helpful, thanks.

[feral fjord]

You can go download images you like that people are sharing here (in channels like #🏞｜general-with-images ) and then run them through PNG info, it'll show you the prompts, sampler, model, etc. However if the image has been upscaled in Extras the info will be absent. But like half the times it isn't so it's very useful.

[random mortar]

hi all. I was looking for a prompt guide for 2.1. I found some links to https://stability.ai/sdv2-prompt-book but that link is broken. Does anyone know where I can find that guide

[chrome cave]

I can't for the life of me get SD to make a Battleaxe with an axehead made of fire

[south solar]

SD doesn't like weapons of any type. Unless you can find a model to add them back in, there isn't much prompting can do.

[spring gate]

Tech-related question: If this is the wrong chatroom for this, please let me know which one I should use:

I'm very new to SD and am temporarily running tests on a 2015 Macbook Pro--left my PC in a different location while I moved :(. This is obviously not ideal for GPU, CPU...anything, for that matter. The consensus I've researched thus far tells me that running dreambooth for SD would probably be impossible on my current hardware, so I wanted to hear from y'all if you had any ideas as far as workarounds. Is Google Colab a service I could pay for that would make running training models possible on my hardware?

Apologies, I just saw the "tech support chatroom". Sorry, guys!

[feral fjord]

Try inpainting

[rustic birch]

Oh so the prompts are saved in the metadata? That makes it a whole lot easier, thanks for that tip!!

[feral fjord]

Yeah. Download the pic, and send it to the "PNG info" tab in Automatic1111

[rustic birch]

But it’s actually metadata and not reverse engineering like clip interrogator or something, right? Cause we tried that already and while those do give good results it’s nowhere close to what actual human made prompts do - at least for our specific workflow

[feral fjord]

No, it's literally metadata

It's 100% the specific data of that image

[rustic birch]

Awesome! Thanks man, I’ll try it later today

[feral fjord]

But as I mentioned some people remove it

You're welcome!

[crimson bay]

Where can I find a catalog of possible tags to be used in the prompt?

[finite prairie]

is anyone aware of successful generations of loose charcoal sketch style portraits? Eg:

[glossy vessel]

does putting "not nude" in the positive prompt will prevent nude, cus apparently putting 'nude' in the negative prompt doesnt work for me

[orchid moss]

Highlight nude in the negative prompt and press ctrl-shift-up-arrow to increase it's weight. The webui should add brackets around it. Increase the weight to 1.3 or 1.4
add more keywords in the bracket such as nsfw, nudty, naked, exposed breasts, etc. Maybe that will do the trick. Or you prompt positively in favor of a clothing style (wearing red dress:1.3). Should do the trick, even if the underlying model was trained on nude models. If you use a negative embedding that might also contribute to nudity, try removing it if not strictly necessary

[glossy vessel]

ok, thanks a lot

[glossy vessel]

but the ctrl + shift + up arrow doesnt work

[orchid moss]

Does CTRL-Up arrow work? without shift?

[glossy vessel]

nope

[orchid moss]

Damn, unfortunate. Then just type it manually I guess 😆 You can add extra emphasis to prompts with round brackets as in (((on the moon)))
You can read about the math and more info on here https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#attentionemphasis
But nowadays, instead of spamming brackets, you'd increase attention to certain prompts by colon + factor (on the moon:1.0) -- 1.0 is default strength, anything below it has reduced emphasis and something such as 1.5 is considered high already 😉 Hope it helps

Of course all of this implies that you are using that webui, application differs slightly depending on what you use exactly ⚠️

[glossy vessel]

yeah, used to spam brackets. just use one with different weights should be enough

thank you so much for the help. I really appreciate it.

[lucid knoll]

is it fine to combine multiple Lora/embeddings together?
ive been getting some strange results while combining them, im not sure why

[muted steppe]

small question, regargin prompt formats

does it matter if i separate them using "enter" to have them be on different lines or must they be one line

for example

beautiful face, pure face, picture-perfect face, absurdly long hair, (shiny hair, wet hair),

or

beautiful face, pure face, picture-perfect face,
absurdly long hair, (shiny hair, wet hair),

[tired vigil]

Anyone knows correct wording for a mask like this?
just "mask" is too wide of a word with lots of variations I feel like

[tired vigil]

Its fine, you just have to finetune the weights

[lucid knoll]

thank you!

its so weird, i see people getting good results from 20 steps, but i need at least like 90 to get something comprehensible

[tired vigil]

That's pretty high, what sampler are you using?

[lucid knoll]

DPM++ SDE Karras

i wonder if its my SD VAE?

what do you use for that btw? and thank you so much for the help 😄

[tired vigil]

Ya with that you shouldn't have to go over 30-40

I make anime stuff, and I use 25 steps for it on sde k

[lucid knoll]

what happens if i do go over that? 😮

[lucid knoll]

sde k?

[tired vigil]

diminishing returns

[lucid knoll]

i see

[tired vigil]

Its dpm++ I mjust too lazy to type it all out

What model do you use

[lucid knoll]

ah ok i got it

[lucid knoll]

im using my own mix rn. but usually i use a DosMix or similar models

[tired vigil]

Do you have restore faces enabled? I found that having that enabled on anime models washes out the color and makes the eyes cringe

[lucid knoll]

is taht in settings?

[tired vigil]

What UI are you using

[lucid knoll]

1111 pronanly 🤔

probably*

the normal UI? im not too sure

[tired vigil]

Are you able to create generations fine without the lora?

[lucid knoll]

theyre underwhelming usually

ive gotten some good ones

but they took 1.5 hours on my gaming pc

got 12gb of vram and a good cpu

[tired vigil]

That's probably why ,genning on CPU takes forever

[lucid knoll]

because of all the steps and rescaling steps

[tired vigil]

Do you have the karass settings enabled in settings?

[lucid knoll]

maybe not 😮

[tired vigil]

If not I suggest you do

[lucid knoll]

where is that setting?

thank you 🙏

[tired vigil]

I don't know off the top of my head but just control f karass or however you spell it

[lucid knoll]

this one?

[tired vigil]

The creator page for the model says to use https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/vae/kl-f8-anime2.ckpt or
https://huggingface.co/AIARTCHAN/aichan_blend/tree/main/vae
or
https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.ckpt

No it should be the one under it

Creator page says clip skip 1 works better for the model too

[lucid knoll]

yah, i use the first one, the anime one hmm

[tired vigil]

It's a semi realistic model so you really have to work your prompt-fu

Can you get similar generations to the sample page?
https://civitai.com/models/6250/dosmix

[lucid knoll]

i cant at all actually D:

thats what i struggle with

i see people getting crazy results

and i always doubt them

because i dont get stuff like that ever

unless it takes 1.5 hours

i couldnt really find that karras setting

🤔 *

[tired vigil]

Enable quantization in K samplers for sharper and cleaner results. This may change existing seeds. Requires restart to apply.

[lucid knoll]

ohhh ok

thats great! thanks a lot!

[tired vigil]

Depends on model and what you're trying to make.
As guy below mentioned - anime models usually gets good results at 25 steps and even below that.
There's not much space for improvement later on, even DPM++ samplers don't do much there, that's why most prompts you see usually using Euler a - it's faster , while not much different from DPM ++.

Edit: everything above applies ONLY TO ANIME MODELS

DMP++ is good for realistic models and also changes with higher steps amount, I'd say 25- is decent starting point for it, but it goes higher.

[smoky totem]

it really depends on the model too. I get good results starting at 15 using dpm ++ 2m/karras. beyond 20 steps, you will start to see the pose/situation change every 5 steps. I usually just look at the result, and if the quality isnt good, i try to step up/back 1 step at a time or so, using the same seed and other parameters until I find the image with the best possible quality for upscaling. If I just zoom by to super high steps like I see sometimes, such as 50 steps or above, I see lots of good poses skipped, and the quality is not greater than if I step back to 21, 22, 23. Maybe quality is not the best at step 20, but it is the pose that I want, not the ones at 25, 30 and above. I don't believe in higher step = higher quality in photoreal situations nor anime.

[winter harbor]

how do you describe depth? or achieve depth.

I'm trying to place a character in a dirt grave but it tends to put it only 2-3 ft into a grave

I've tried describing deep dirt grave

[tired vigil]

It's not just about quality, it adds and changes details with more steps.
If you want specific pose - it's easier to use control net, rerolling steps trying to get pose randomally is kinda weird

[winter harbor]

Sonar blows controlnet out of the water

[tired vigil]

No idea what's that

[winter harbor]

I'm sure controlnet has its uses

but for quicky copy cat of posing

or scenes

sonar quickly reads depth and replicates your reference image

almost 1:1

in a few seconds

[tired vigil]

okay

[smoky totem]

I've experienced in the past where going super high on steps do not help. I lot of things start to change, getting less close to the prompt. Like wavy hair getting too frizzy despite frizzy hair on negative prompt. So not just poses, and the supposed more detail on clothing etc not present with high steps compared to just 20 or so steps. It is just different at high steps. I see no benefits.

[winter harbor]

Heres a reference image

Here it is copying it

[smoky totem]

I only use high steps to see what the transformation looks like, and then I go back to a lower step when I see the scene transform to a state that I like

[winter harbor]

https://github.com/Kahsolt/stable-diffusion-webui-sonar

@tired vigil

[tired vigil]

I'm doing most gens at 25 steps too , but mostly cause I don't want to wait for too long 😄
Especially since most of gens are just seed hunts 😦

[smoky totem]

Me too, I only go high step when I see a good seed, and wanna bet on it to see what higher step bring me...usually go very high, and watch the transformation, and then go back down to narrow in on 1 that is good.

[tired vigil]

Alternatives are always good, how is is better tho?

I mean , why do you think it is?

[smoky totem]

Thanks for the intro, gonna try it sometime. I don't do enough characters so I don't pay too close attention to these tools, but I use control net for objects, mostly the sketch model. Maybe I can find a way to use this for objects and scenery.

[lean island]

Sorry if its been asked before.. I saw a setting in SD-webui that is for "old emphasis implementation"

what is the old method and what is the new? I prefer using the ((*)) method and I guess I just need to know if thats the old or new method

[silver valley]

You dont neet to enable it.
It was basicly both. Its the old and new one but it got a bugfix pretty early that it could change old outputs if you try them now again.
Here is the info:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Seed-breaking-changes#2022-09-29---new-emphasis-implementation

[lean island]

you're a god. thank you

[winter harbor]

You can run it straight out of the box (from the script area) and it will work 90% of the time as long as you properly control the denoise tab.

In comparison to controlnet, I would have to play with the amount and mess with all kinds of different depth and a lot of the time some of them would not load and would give me an error when I wanted it to process at higher amounts.

Sonar I've noticed is super light weight. I prefer it because you save time fiddling around with controlnet settings or the type (like openpose and etc).

I played with controlnet for a while to try and replicate poses and it was a real struggle

that was driving me crazy

sonar has been a huge time saver for me

[tired vigil]

This thing has plenty of sliders to play with too ¯_(ツ)_/¯
But aight, I'll stick with controlnet for now, it does what needs to be done, maybe when I'll need to do something more complex I'll try it out

[winter harbor]

As I mentioned

it works out of the box without any tweaking

I actually never touch any of the settings

but hey man, you do you

and it has literally 2 sliders. 3 for upscaling (which you don't need to have turned on)

In the IMG2IMG tab you can drop a photo and then using your prompt you can generate the same type of photo, by controlling the Denoising strength to match the prompt more or the image more

not sure I follow with how its more complex lol

In comparison, controlnet has you messing with all kinds of settings

meanwhile Sonar, just pings your reference image and applies that to your prompt

composition/posing within seconds

[willow nest]

is there any way to get characters to pose like this? i've included this in control net and added a bunch of words like standing straight etc. but the characters always have kinda tiled head/body

[sacred canopy]

That's called the A pose, like arms straight out is the T pose.

https://www.linkedin.com/learning/3d-animation-and-rigging-weekly/a-pose-or-t-pose#:~:text=Now A pose basically has,a number of different factors.

[willow nest]

tried A-pose but still getting stuff like tiled hips and face :( (same with T-pose)

[rare rapids]

use canny

[frosty summit]

Install OpenPose Editor (Extension) for ControlNet
https://github.com/fkunn1326/openpose-editor

[willow nest]

that worked!

[nimble grove]

has anyone tried training a lora model with faces/heads cropped or hidden? what prompts should i use?

[tired vigil]

I have.

[lucid knoll]

thanks!!

[pulsar valve]

Where can i find prompts that works and and make really visible changes in Render? Im mean prompts what Stable difussion understand.

[unique lodge]

Hi, at https://replicate.com/cjwbw/hasdx why are some of the keywords written in parenthesis?

[wheat swift]

it makes them more important in the prompt

[unique lodge]

Cool, thanks!

[wheat swift]

np you can also add weight by adding a colon and a number inside the parentheses

(word:1.4)

[unique lodge]

Nice

[pulsar valve]

U guys write always in promts box anything like 4k, art, masterpiece etc.? It is rly works or dont make any changes?

[unique lodge]

What prompts would you guys use to recreate:

• https://media.gq.com/photos/5cdb0d39bacbf2c41121be01/1:1/w_1296,h_1296,c_limit/Flying-Lotus-collaborator-breakdown-Flamagra-gq-1.jpg
• https://i.scdn.co/image/ab6761610000e5eb8d1330444720d6edffa07068

[pulsar valve]

I think best option is to make own Lora and then start to recreating it

[wheat swift]

I do that all the time

makes a huge difference

[pulsar valve]

Ok ty then i need to make my own set with these keywords

[wheat swift]

Learning the keywords will take you a long way towards customizing and developing your own style

[pulsar valve]

Yeah ik. Its time to start haha

[unique lodge]

Those photos are from Warp.net the world's largest record label. I'm trying to create my own record label and would like a similar way of presenting my artists.

[nimble grove]

what did u use for prompt/lora

[wheat swift]

as an artist who is also a musician, I think you may be going in the wrong direction. consider the mode you want to portray and put that into your prompt. You definately want to include the words "Head and shoulders portrait" but be descriptive about the picture beyond that. match what the album in question makes you feel to the prompt

one thing I have learned since I have started doing this is that sometimes you have to reverse your way of thinking

[unique lodge]

That's a great idea, but for now we haven't really released anything so just looking for something general but awesome.

I guess I should just brush up on my adjectives then 🙂

[wheat swift]

If you are just doing demos, then grab an album from an artist you like and try to craft a prompt from what that makes you feel

and don't forget your subject 🙂

but you don't want a collection of canned prompts for album covers. each one needs to be as unique as the album

[unique lodge]

But I want to craft a prompt like those images above just to get the hang of it.

True indeed!

Yeah luckily our music is quite unique.

You got any music I can check out by the way?

With accompanying images? Just curious, I know it's none of my business.

[wheat swift]

I have one off my album that didn't get released with it's album cover

stupid loss of hearing

[unique lodge]

You lost your hearing?!

So sorry man

[wheat swift]

Just above 12Khz. lost my ability to master so I couldn't finish the album

I'm not certain, but I think soundcloud is down

[unique lodge]

SoundCloud has been real spammy lately

[wheat swift]

https://soundcloud.com/djdragonforged/the-beast-within/s-10b1UJqWg4o?si=07d6f0b2421e4fb4897ea4f76f4f8964&utm_source=clipboard&utm_medium=text&utm_campaign=social_sharing

[tired vigil]

hello world, hopefully this is the right place to ask questions....um I'm trying to have two anime characters that are from lora into one picture and i followed this yt guide https://www.youtube.com/watch?v=uR89wZMXiJ8 but doesn't work. I have no problem with one character in one image. any experts can help me? Thanks in advance! 🙂

[wheat swift]

Do you have any other image editing software? When I have trouble getting 2 characters together, I end up using the brute force method and compositing the pictures

[tired vigil]

I have affinity photo but I never used it and don't know how to use it (and too lazy to learn it right now LOL)

[wheat swift]

Ok sorry then I'm not going to be much help to you.

I can tell you how to bind after compositing, but that's a completly different ball of wax. I'm just not experiencecd much with controlnet

[tired vigil]

i could get a "free photoshop" from somewhere if I can be bothered downloading

[unique lodge]

GIMP is awesome

[wheat swift]

I use gimp personally

[unique lodge]

Especially the Elsamuko plugins

One of a kind

[wheat swift]

but I will take the composited photo into img2img with low noise and the same prompt. it will do a slight regen gluing the composite together

[tired vigil]

i could probably ask my best friend to help me with it lol but i want to know how to make it work on stable diffusion w/o using photoshop or sth else (IT IS SO FRUSTRATING!!)

[unique lodge]

DragonForged: Nice music

[wheat swift]

Thank you!

[unique lodge]

If you help me generate prompts like the two pics I showed, I might be able to help make your music even better!

(and then maybe I'll sign ya)

[wheat swift]

I have given up on the music

My hearing can't handle it

and bethovan I am not

[unique lodge]

lol

[wheat swift]

but I can give Da Vinci a run for his money

[unique lodge]

You probably had the sound on full?

[wheat swift]

with headphones

I know it's my own damn fault

[unique lodge]

I can show you how to be Beethoven

With a few simple tricks 🙂

[wheat swift]

you can't fight with the AI. you have to cooporate with it

[wheat swift]

I'll think about it

right now. just trying to get my art out in front of people

[unique lodge]

Cool

[unique lodge]

Idk

Doesn't look realistic, nor anything like those photos above

[wheat swift]

But it is start. You can tweak the promts from there

[unique lodge]

And why is it that SD doesn't look 100% realistic?

Any tricks you know of?

Maybe import into GIMP and run one of Elsamuko's plugins

[wheat swift]

I didn't add the hyper realistic keywords

[unique lodge]

aaah

[wheat swift]

That's not the style I usually go for so I didn't think about it

But I embeded my prompts so you should be able to grab them, drop them into png info and tweak them from there

[unique lodge]

The guy himself isn't that important, more the vibe and ambiance of the photoshoot

I'll figure something out though, cheers!

[wheat swift]

I just borrowed your subject 🙂 you need one for that vibe

[placid mica]

Any ideas how I could get a pants to be unbuttoned as a prompt??? Might be I just don't have any models that have that trained, but it doesn't seem like such a weird thing does it? 😄

[feral fjord]

I'd use a model focused on realistic images instead of using base SD

You can try to figure it out with Stable Diffusion but it's quite the struggle

With models made for specific types of images it'll be way easier

[unique lodge]

I'm using Dreambooth (https://replicate.com/replicate/dreambooth) of my friend's photos. Trying to create some nice artist shots for him before his album release 👍

[tired vigil]

Yea, most models not trained on it, tried few and not a single of my realistic models can do it

But few anime models can, it seems

[tired vigil]

Try nsfw (xxxx , yea) related models

[placid mica]

Yeah, that's what I thought... Guess I'll either try to train a lora for it, or just wait until someone does that and uploads it. Seems things are blasting forward with some insane speed these days!

of course, then I need to figure out how to make a lora, but I should learn that soon anyway.... Been playing with SD now for over a week already, time goes fast when we're having fun 😄

[tired vigil]

I think there should be models already trained on images like that , just need to find one...

I'll download one more thingy to test in few mins

nope, not even nsfw realistic ones I tried can do it , weird

Have you tried unzipped pants

same deal

hm...actually...

[tired vigil]

(unbuttonned:1.2) (open) jeans pants, (unzipped pants:1.2)
This (kinda) works on arteros model

[placid mica]

Oooh, nice, guess I'll find that model and test it

[tired vigil]

Probably will take some time to get decent results tho...
Try changing prompt around a bit , see what works better, maybe it'll make it more consistent

[placid mica]

well, the fact that you got something shows that it's possible, so that's a good start 🙂

[tired vigil]

alternatively it should be possible to make somewhat realistic looking thing out of anime models , but won't be as good as actual photo-realistic models

[placid mica]

yeah I got it already

thx 🙂

[lavish surge]

why there are sometimes some images that are divided into two?

[placid mica]

in my experience that happens mainly when the ai has "too much space to work with" so to speak... if the canvas is very wide I sometimes get two somewhat similar images on the same picture

[lavish surge]

oh true

[silver valley]

Do you mean 2 images as output?

[placid mica]

I'm suddenly struggling to get SD to focus where I want it... I want a closeup on a hand, that shouldn't be so hard right?!?!?

[weak barn]

any good expert online? xD

2 girls hug, 1st one, that back to us have long black dress, standing on one leg, one leg up... no panties... 2nd girl standing face to us, grab 1st girl dress and lifted to the waist, showing nude ass, grab ass second hand....

my stable diffusion go crazy xD can't combine prompt xD any help?

[willow nest]

is there a prompt for pencil sketches or line art?

[weak barn]

can make one

1 sec

more clean

[willow nest]

oh sorry that was a separate question XD

[weak barn]

T__T

[willow nest]

have you tried control net tho?

[weak barn]

OpenPose Editor, ControlNet, Latent Couple and etc xD

[gaunt sphinx]

has anyone seen any fix for the same face being applied to everyone in the picture on the inpaint? would an anti-tween textual inversion be possible perhaps?

[strong quarry]

Hello, I'm not sure if I'm in the correct place to ask for help with that.

I'm trying put a fur robe in that sweet lady
But I don't know what I'm forgetting or what I should do in the prompt or in the settings bellow.

• I'm using:
  Yesmix model
  Prompt: masterpiece, best quality, granbluefantasy inspired woman, Silver long hair, with thin deer horns woman, covered by white (((fur robe))) and hood over shoulders, portrait, half body, mystic forest background, fireflys

Negative Prompt: (painting by bad-artist-anime:0.9), (painting by bad-artist:0.9), watermark, text, error, blurry, jpeg artifacts, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, artist name, (worst quality, low quality:1.4), bad anatomy, watermark, signature, text, logo,worst quality, deformed, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, disgusting, poorly drawn hands, missing limb, floating limbs, disconnected limbs, malformed hands, blurry, ((((mutated hands and fingers)))), watermark, watermarked, oversaturated, censored, distorted hands, amputation, missing hands, obese, doubled face, double hands, b&w, black and white, sepia, human ears
Sample: DPM++ 2S Karras
CFG Scale:10
Denoising: 0,6

[silver valley]

try an other model maybe

getting this first try with my merge model

[strong quarry]

OMG so good

[gusty marlin]

anyone know if there's a dataset of prompts (including negative prompt and image rating would be the best)

[sudden surge]

Hi guys, do you know how to get something like this? I can't find the right prompt, I spent almost a full day with only a few good outputs (I need to create different objects and animals)

[olive wedge]

Has anyone else experienced CUDA errors due to removing parts of a prompt? I can consistently generate with some settings without any issues, but if I remove "a beaming smile" from the prompt, it seems to randomly try to assign huge amounts of VRAM (21GB on a 12GB card).

[vapid shore]

yall I been lurking and searching -- is there a handy cheat sheet for some prompt guideance? there's a great sticky with negatives

[unique lodge]

How would you improve these to be more cinematographic like http://www.dannyhiele.com/?

  "save_sample_prompt": "male, brown hair, frills, detailed skin, perfect face, fashion photography, ultra realistic, photo realistic, 20 megapixel, canon eos r3",
  "save_sample_negative_prompt": "cartoon, 3d, disfigured, deformed, bad art, poorly drawn, extra limbs, close up, strange colors, blurry, boring, sketch, lackluster, big breast, large breast, huge breasts, face portrait, self-portrait, signature, letters, watermark",

[charred crow]

Hello everyone, was just wondering... is SD able to extract prompts from image? Like input an art and then it output the closest prompts that would generate similar art style to that image

[tired vigil]

It can only extract prompts from an image if the metadata is in the image itself. The person posting the image may or may not have deleted it,

On automatic1111, there is something called img2img, where you upload an image and with a guided prompt it will produce a similar image depending on the noise level you want.

[charred crow]

hmm, this seems more complicated than I expected, I will need to look into it more to understand what you said, either way thanks for help ;)

[rustic vapor]

Can anyone tell me or point me to a guide on how to train a lora in colab? I've done it before with a model and it worked, but I want to do with with a lora and all the stuff I'm finding are guides to locally run training and I don't have the hardware

[frosty summit]

You could use something like CLIP Interrogator where you input an image and it will tell you what tags it thinks would recreate that image

[floral leaf]

hey anyone here i have a question regarding getting a room to be dirty like its for a image but the tag dirty room isnt working, and ive tried adding ((dirty room)), is there any other way or tags i can use to describe the room to get it to look dirty

[frosty summit]

How do you envision this dirty room? Is it trash? laundry? oil stains?

get detailed

[floral leaf]

just stuff everywhere whether it be clothes, anything

[lone badge]

Describe the dirtiness more in your prompt. The things on the walls or in the corners,... It should help SD populate the room with more dirt stuff

[floral leaf]

alright thanks.

[lone badge]

Even just add at the end "dirty clothes, piles of dirt, old food, dirty dishes, garbage, ..."

[floral leaf]

that actually sounds like it would work, thanks. i never looked at it that way

[lone badge]

Your approach can also work sometimes, but the "dirty" token doesn't seem strong enough on its own, so yep, we need to go the descriptive way 🙂 hope you get some good dirt!

[frosty summit]

I like to go to relatedwords.org and see what other words are related to "dirty" for example

or a thesaurus

[floral leaf]

ah so relatedwords is a good source that would work for synonyms for any words

[lone badge]

Great tip

[floral leaf]

thank you so much i didnt even ask for something like this and it would help me even more, you guys are amazing wordsmiths

[frosty summit]

Good luck!

[silver valley]

messy room

[floral leaf]

i ended up getting a room i wanted

thank you

[frosty summit]

That's great! now.. clean your room lol

[floral leaf]

the room was basically for a small story i was thinking of, the protagonist found someone who was extremely technologically literate, but the issue was he needed his help and the guy wouldnt help and the protag decided to pay him a visit, the rumor was he was a neet that hadnt left his room in forever

[lavish surge]

2 divided images in one as a single output

[lone badge]

That does tend to happen. Especially when you have some concepts that SD has a hard time mixing together, it can separate those like that. It can also come from some tokens sometimes, that inspire split output like that. It should happen rarely, but it does happened still regularly to me

[crimson bay]

what are these prompts (bad-hands-5) exactly? Is it some pack of negative prompts or something?

[silver valley]

Its am embedding

[crimson bay]

how do I use this?

[silver valley]

You download them and put them into the embeddings folder, than you mention the name in the prompt to activate it

[olive wedge]

I think I've seen this created using "messy room" (implying it is untidy) rather than "dirty room" (which implies it is unclean).

[pastel ether]

So i was building a prompt, and from some point onward, all generations started having weird colors and deformed faces and for the life of me i cannot figure out why that is. Reverting settings to a while back and trying to remove tags i think it has issues with doesn't help. Why does this happen and how do you debug it?

[olive wedge]

Do you have any previous generations saved or Image Browser? If you load up a last "known good" generation, does it generate correctly?

[pastel ether]

It does seem to do that, so maybe i am missing a tag that upset it or something? I tend to not even really fiddle with setting much, mostly prompts

I just don't understand why a specific prompt it doesn't know will suddenly cause all sorts of mangled faced and limbs

[olive wedge]

Can you use the "Test My Prompt" extension? That can be useful to narrow down what is affecting your generation. It runs your prompts but removes each section of it in turn.

Perhaps it is related to the behaviour I'm getting where removing parts of the prompt almost doubles the VRAM the generations take!

[pastel ether]

Isn't that due to token length of the prompt?

Anyway, is test my prompt an extension?

[olive wedge]

It is: https://github.com/Extraltodeus/test_my_prompt

And removing parts of the prompt shouldn't make it take more VRAM. It should make it take less if anything (though really, I'd expect it to have no effect at all).

Certainly not drop my batch size from 12 to 2..

[pastel ether]

Huh, that is indeed a very strange behaviour, sorry i misunderstood 😛

Some sort of extension meddling maybe then? I have never seen that in base SD generation

[olive wedge]

No worries and certainly a possibility! It's just odd that I've only just started experiencing it which makes me suspect it has something to do with subprompts, which I've only just started using. Still, going from being able to generate 12 images with my settings on my 12GB card to being able to generate <6 because it randomly tried to allocate 21GB because I removed "a beaming smile" from the prompt is bizarre. Especially because if I add it back in, it works fine at 12 batch size again.

[pastel ether]

It almost feels like it is not prompt related and something causes it to leak memory/not cleaning up something during the generation of the last image

That is quite a stretched guess though

[olive wedge]

I'd expect I'd need to restart to fix it but that isn't required. Just undoing the prompt change is enough. So something is in a strange state, that's for sure. I should probably test it out by restarting and trying with the removed prompt right away.

[pastel ether]

Yeah that sounds like a sane step, then you know it is not the prompt that is to blame

[olive wedge]

#blamethepromptnottheprompter

I really should look into getting this GPU into my second computer so I can have 2 SD instances going at the same time. 🙂

[pastel ether]

You have 2 GPU's good enough to do SD? what a luxury problem to have 😛

i'm sitting on 8Gb here like a casual

Also, isn't there something out there, so you can link both together? like 1 master multiple slaves to subdivide the prompt requests to?

[olive wedge]

I have a 12GB 2060 in my primary computer which I am currently using. I purchased another 12GB 2060 to fit into my secondary computer but I haven't done so yet as that machine is still busy and I need to do some considerable setup on it (e.g. making it headless, setting up the VM and VFIO).

I'd imagine Stable Horde could do something like that, which I do hope to have both machines connected to once I can get it running.

[pastel ether]

i guess so

[olive wedge]

My plan is to have 1 machine doing generations while the other is set up for training. I want to get some TI training working well. My current attempt seems to be training too fast..

[pastel ether]

training is a time consuming game, it's neat, i got it to work decently, as a Lora anyway, hypernetwork failed miserably on my end. But in the end i get frustrated in what it does or does not do well enough.

Then again that complaint goes for full models as well

It's so strange how it can do so many things, but you ask for like a "sun dress" and it just completely gives up on you.

[olive wedge]

You've have decent results on Lora training?

My current idea for TI training is to set a "time limit to likeness". So I will get a learning rate where the likeness is just starting to come through e.g. after 5 minutes of training. Then another 25 minutes for it to train on the details.

In my testing, I start getting a likeness after an incredibly short time (10 or 15 steps!) which I think is just priming the entire thing to overcook.

[pastel ether]

decent, yes, but i don't fully understand all the learning rate setting stuff, i know getting it wrong ruins everything, and i happened to hit the right example online. But experimenting with "slightly more, slightly less" takes forever

because whatever training i used, with about 100 images, took me a good.. 3-4 hours

depending on how many steps per image i used

[olive wedge]

That's a lot of images! I've only been trying TI training with 16.. 🙂

[pastel ether]

Well, i figured since i did not get specific images for the training, i just yeeted in 100 photos of my wife, auto tagged by booru and manually cropped to fit her as the subject. See where it goes 😅 and it went well enough, images generated look like her. But it also tends to veer randomly, and if the prompt gets slightly confusing stuff goes off the rails

so i am now trying improved tagging, which has seen improvements, but that takes a lot of time

[olive wedge]

It certainly does.

[pastel ether]

so yeah, just like other prompts, it's fun when it works, but annoying when it doesn't understand a word for the thing you want, or it has a low sample size on it :/

[olive wedge]

Very true. I found it surprisingly difficult to get a "hands behind head" (as in, relaxing) pose which I'd expected to be well represented.

[timid kindle]

How do I specify colors I want in the prompt? I wrote "light-blue" and it turns out a little too dark. How would I be able to describe more in detail what color I want?

[placid mica]

lora training is dreambooth, right?

[olive wedge]

They are separate things as far as I am aware. I think Dreambooth can do Lora training, but it isn't the only way?

[pastel ether]

funnily enough, it's been giving me a lot of images with hands just above head for the lora i trained on my wife 😛

[pastel ether]

i used a separate tool and in that tool it is called dreambooth lora, so i guess yes? 😛

[tired vigil]

How do I inpaint a teeshirt on a topless man?

Seems to just tint the skin colour when I use something like yellow teeshirt

[olive wedge]

Denoise too low perhaps?

[south solar]

Try higher denoising strength.

[pastel ether]

@olive wedge If i have that tool you suggested, where should it show up?

[olive wedge]

test-my-prompt shows up in the 'scripts' dropdown.

[tired vigil]

Thanks you 2 i will try that

[olive wedge]

Personally, I would set it straight to 1.0, see if that creates a yellow t-shirt, then go halfway between that and the denoise that only tinted the skin. Binary search to a minimum denoise that reliably does it.

And don't forget you can always send a result back for another round! That can work a lot better than pinpointing a denoise (depending on your goal).

[pastel ether]

guess installing it didn't work properly then

[olive wedge]

Oh! Yes, the author does not not have the directory setup correct for being an extension. One moment.

[tired vigil]

Ok thanks.

[pastel ether]

oooh, kek

so i need to move the script file

gotcha

[olive wedge]

diff --git a/test_my_prompt/scripts/test_my_prompt_custom_script.py b/scripts/test_my_prompt_custom_script.py similarity index 100% rename from test_my_prompt/scripts/test_my_prompt_custom_script.py rename to scripts/test_my_prompt_custom_script.py

They have it in a subdirectory when it shouldn't be. Thank you for reminding me to send that PR in to them.

[pastel ether]

yeah, found it, it works now 👌

thanks

[olive wedge]

Excellent! 🙂

Very interested to hear your results. If I was going to throw a guess out there, I'd say you are pulling in a broken (perhaps negative?) embedding. But perhaps that is too obvious an issue..

[pastel ether]

i don't even know what an embedding does yet, so i assume i do not 😛

[olive wedge]

Ah, if you have no embeddings downloaded, then it is indeed an unlikely cause! 🙂

It would say so in the parameter info under the generations. Something like: Used embeddings: easynegative [119b]

Just for reference, this is why you will often see embeddings with strange names, such as replacing letters with numbers. This is so they are not used accidentally!

[pastel ether]

ah yeah i ahve seen those around, did not use them yet

looks like the cause in my case is my Lora, something ticks it off something fierce 😥

[tired vigil]

is there something I can put in a prompt that will be completely ignored? I'm trying to do an XYZ Plot with an empty value on the first column

using Prompt S/R

[olive wedge]

There is not. I did look into adding a "first value is a placeholder" option to Prompt S/R but have not finished that work yet.

[tired vigil]

guess I'll have to put it on the last column then 😒 thanks

[olive wedge]

I suspect I will have another look soon as I want it for testing my TI training results. My first attempt got into the weeds a bit due to how the grid is generated.

[timid kindle]

Still relevant. If there are any relevant resources they would be really appreciated

[olive wedge]

Have you tried finding a "common name" that is close to the shade you want? e.g. "azure", "sky blue", "pastel blue".

"aquamarine" might also work.

[pastel ether]

When you sue a Lora, what kind of weight do you give it? What is normal for those?

[olive wedge]

Really depends on the Lora in my experience. I see a lot of Lora authors suggest between 0.6 and 0.8.

[pastel ether]

Yeah, it's strange, some seeds my results are fine, and i can put my lora on 1 or 0.9 with no issues. Other images i have to put it down to 0.6 to not mangle the images, and then it doesn't perform the task as well as i would like

[timid kindle]

I thought about that, is the model capable of recognizing the different names of colors rather than generic terms? My intuition tells me that the more general I am with the prompt the more likely the model will know what the hell I'm talking about. Maybe I'm wrong.

[pastel ether]

I think this is the same issue i have with sundress, if the model was not trained on a word that matches your color of blue, you are just out of luck

[olive wedge]

I usually use Prompt S/R for exploring around concepts if it doesn't show up in my prompts right away. Sometimes SD just knows it under another term. Like having to use American spelling 🙂

[timid kindle]

Interesting. So it's a matter of luck

I wondered if it's possible to use hexcode in some way SD recognizes. Perhaps I can use the paint or controlnet features to do so?

[olive wedge]

If you are trying to generate just a single picture, you could paint over it and then img2img it.

[timid kindle]

Looking at some other creations, they might have not given much attention to the coloring and they simply got lucky, but it seems like the coloring is precise

It's not like painting yourself where you can actually just change the brush color

[olive wedge]

It doesn't help that SD struggles with colour isolation. Though perhaps Latent Couple could help with that, if that is part of the issue?

[pastel ether]

you named red somewhere? Let me just randomly color any particular item red

[olive wedge]

Haha, pretty much!

"You said pink hair and red clothes? You said pink first so.. pink everything!"

Though that reminds me that I did see someone suggest about putting all the colour prompts together so they got similar weight helped with that..

[pastel ether]

really? i tend to go "red_hat, blue_shirt" and that sort of thing tends to go... well enough, in my experience it is more often correct than wrong

but boy howdy will you have like 20% red shirts in there

[timid kindle]

Yeah it's kind of interesting actually. Both SD and chatgpt have incredible capabilities but sometimes they overlook very obvious details and it's the main cause they don't actually feel like humans. It sometimes takes me 3 times to correct the exact same error chatgpt makes over and over or a negative prompt I specifically said not to generate by SD doesn't really care

[olive wedge]

I've had ChatGPT tell me something that made no sense whatsoever (initialising a CPU register to 0 by ORing it with 0 or itself). On the flipside, I have told it a bunch of information about a character from a story I've written, omitting a detail and then asking it what view my character would have on that missing bit of information and it got it 100% correct.

[pastel ether]

It's because in the end, you have to realize, chatGPT is just a very advanced auto complete text generator. And stable diffusion is just a machine that makes fuzzy image less fuzzy, taking suggestions and hints from input words.

If the output you are looking for, was never part of the input, or not a big enough part to know how to deal with it. It is bound to just return complete bogus. And the only way around, is to train a bigger network, add capabilities to improve said network in some way, teach it new concepts without fucking up the original.

ChatGPT is at a point where it can GUESS at the right answer based on all the context it has on a topic. Which is what generates those "seems plausible, but totally false" responses. But a few years down the line, that gap in knowledge could be filled, especially if they allow it to search online for some hints on what does or does not exist, basically fact checking what it knows already.

[olive wedge]

I do realise that. That's actually what makes it so strange that it would suggest OR to set a value. There's no reason it should have ever seen that in its corpus.

[pastel ether]

well, that is where we go with the bogus, it just formed some random sentence, really off the rails i guess? 😛

[olive wedge]

Well this is ridiculous. I reduced the image dimensions on HighRes Fix and.. it's running out of VRAM. The more I try to reduce VRAM usage, the more VRAM it wants!

[orchid moss]

Any idea how to prompt for crayon style drawings? If I do something like (using crayons, color pencils:1.3) I often get crayons depicted within the drawn image itself... I can't think of anything good

[orchid moss]

Sometimes it hogs your VRAM and won't let go. Try loading another checkpoint, then back. If all else fails you have to restart the webui quickly

[olive wedge]

I have the inverse problem. Doing things that would require less VRAM cause it to use more and run out of memory. I increase the target resolution, it generates without issue..

[pastel ether]

and just for shits and giggles, you are not running a game on the side or anything? 😛

[olive wedge]

No. I run SD in a VM with GPU passthrough. There is, quite literally, nothing else using the GPU.

[pastel ether]

that is nuts, so strange

again it must be something leaking VRAM, meaning it tries to claim new vram due to lack of released data or something

i can't think of anything else

I ran into an interesting issue. I did some testing with that plugin about what is wrong with my Prompt

and every individual prompt excluding stuff (besides my LORA) faces are scuffed, however if i remove 2 or 3 things it seems to have slight issues with from the prompt (none of these have any to do with faces or facial features) it suddenly starts generating the lora implied face again without disfiguring it

and i don't understand why

[olive wedge]

In that case, I would expect it to get worse and have to use lower and lower batch sizes over time but that isn't what is happening. If I revert the changes, I can put the batch size back up again as well.

It isn't a memory leak.

[pastel ether]

there must be a pattern to it, surely

[olive wedge]

So far, it is "reduce parameters that should require less VRAM -> more VRAM is allocated" 🙂

[pastel ether]

I wish i knew more about lora parameters, i don't think there are any real good guides out there, only guides by people who "got it to work once" very ad-hoc

[olive wedge]

Batch size 9, 320x512 -> 600x960 fails torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 21.73 GiB (GPU 0; 11.75 GiB total capacity; 3.03 GiB already allocated; 7.44 GiB free; 3.17 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Add "a beaming smile" to the prompt.. succeeds: Time taken: 3m 11.86sTorch active/reserved: 9960/10840 MiB, Sys VRAM: 12007/12028 MiB (99.83%)

First 2 generations of a started instance.

Did another run after: Time taken: 2m 35.38sTorch active/reserved: 6250/7524 MiB, Sys VRAM: 8691/12028 MiB (72.26%)

Yeah.. for some reason, removing "a beaming smile" from the prompt just seems to make it require a lot more VRAM, which isn't released until the next generation succeeds.

Like I am so close to some sort of limit that the weights of the rest of the prompt push it into another envelope..

I can see it in nvidia-smi too. The first stage of generation take around 8GB of VRAM when "a beaming smile" is in the prompt, but just over 10GB when it is missing.

How.. utterly odd. If I do a generation with "a beaming smile" at batch size 12, it works. If I change it to "a furious scowl" it works.. but only the first time. Running it again it tries to allocate 30GB of VRAM!

[pastel ether]

that is so random :/

and i'd understand it, if 2 runs in a row with "beaming smile" would run into the same issue ,but it doesn't 😛

[olive wedge]

Okay, I think I've figured out what it is.

[tacit shell]

Can anyone help me get this prompt running smoother? I found a prompt online and was trying to get it to generate an Almiraj (rabbit with a unicorn horn) with a judicious amount of blood. I was able to brute force just the horn with enough repetition, but trying to add the blood makes it get really inconsistent. I had to add blood multiple times for it to actually take. I can't attach the images because the bot keeps deleting them, so they are at the google drive link

https://drive.google.com/drive/folders/1PSaWLe1edBtoaY52hFfOxJQj6Cn8m7fD?usp=sharing

Negative prompt: (woman:1.2), (female:1.2), clothes,jacket, coat, car, human, person, canvas frame, one ear, ~~~```

[olive wedge]

It doesn't happen if I remove the Lora in the prompt.

[pastel ether]

so the lora getting triggered causes a spike?

[olive wedge]

No, I've had that Lora in there. There is something about "", "a beaming smile" and "a furious scowl" that triggers the Lora in different ways, making it require differing amounts of VRAM.

The VRAM usage is consistent when the Lora is missing, but varies wildly depending on which of those prompt fragments is used. Though it doesn't explain why sometimes they work once..

[pastel ether]

that is nuts though, how big is that lora even?

[olive wedge]

And now I can run it as many times as I want with "a furious scowl". 😖

And indeed, as soon as I remove that part of the prompt.. Tried to allocate 28.97 GiB

[olive wedge]

On disk? ~75M

[pastel ether]

how does that even work then

[olive wedge]

I changed that part of the prompt to "a nothing nothing" and it generated. It really must be something to do with the prompt weights triggering some strange Lora VRAM behaviour.

[olive wedge]

Okay, new theory which I am hoping someone would help me test. I think this has to do with prompt splitting. If anyone has a prompt that is 72/75, could you add a Lora to it as well as the phrase "a beaming smile" to the prompt somewhere (right at the end before the Lora is fine) and generate as close to your VRAM limit as possible and then remove the "a beaming smile" and try to generate with the same settings and see if you OOM on VRAM by a large amount (like I see 20-30GB on my 12GB VRAM GPU).

Okay, yes, I think the prompt split is involved. I can do "a photo of a person <lora>" and it runs of out VRAM, but "a photo of a person,,,,,,,[snip],,,,,, <lora>" works fine. As long as it is > 75 tokens, it works.

As I thought, doing a photo of a person BREAK BREAK <lora> also works. Thanks for being a sounding board for my thoughts! 🙂

[pastel ether]

that is still hella odd though

I sadly do not have such a prompt right now, and i have to get going, but i'd help out tomorrow 👌

[olive wedge]

In theory, if you make your prompt a photo of a person BREAK BREAK <lora> and then set it up to use as much VRAM as possible, removing either or both of the BREAK's should cause it to run out of VRAM.

Maybe pick a ~75M Lora or use the asoulEileen one I was testing with.

[fast spindle]

Hi newbie here, may i ask if there are any threads or websites that probably have a list of prompts that gives you a head start of what you may want to create like " so here's the prompt, you can play around these and more."? Thanks!

[olive wedge]

It might be worth browsing CivitAI. A lot of the images uploaded there contain the generation parameters, so you might be able to find some things you like there. Though to get the exact same output, you'd need to use the specific things they use (e.g. checkpoint, lora). Still worthwhile for the general prompt though.

[fast spindle]

Sounds cool thanks! 👍🏼

[olive wedge]

Happy to help. In my experience, if you see an image you like and you can speak with the person who generated it, they are usually quite fine with sharing at least the prompt, if not the full parameters.

[fast spindle]

Wow.. that's good to know. Okay, that's a good tip. Thanks again.

[craggy steppe]

Hello, I'm working on a video and trying to get the style of right image for the entire video, the right side style and character was trained in lora and I used inpainting to get better result for this one frame. How can I get similar result for whole video? Tried ebsynth but it's not looking good as the video has lot of movements

I can see a batch inpainting option but not sure how to use it

[tall wasp]

What's the best way to get photorealistic goosebumps on skin?

[steep jungle]

As non native english speaker, can someone throw me a couple prompts that lets me generate "full picture" of human?

[crimson bay]

what negative prompt do you use to remove this shirt/jacket from her arms? i have tried everything, but something appears, sometimes it is not even clothing, sometimes it is just a fabric or whatever

[olive wedge]

You can use "full body", or "wide shot", "long shot". Describing both hair and footwear can help too, as well as generating in portrait aspect ratios (e.g. 512x768).

[steep jungle]

Thanks

[olive wedge]

Quite welcome. Another way, if you want really fine control of the pose is to use ControlNet and its OpenPose feature.

[steep jungle]

I am aware of this, but I am trying to generate training set, so its not applicable at this stage.

[polar adder]

Hi all! I'm attempting to make some game backgrounds of a gas station using stable diffusion and control net and I'm hitting a wall...

I'm having a really tough time getting the architectural elements to not be all wobbly. I'm basing this off a photo and using canny edge detection, but I can't seem to find settings that don't output wobbly lines. Any thoughts?

[placid mica]

feel like I keep coming back to this problem.... Any good tips on having a person drawn further away? I'm trying to make a black background with a person pretty small in the center, but I can't for the life of me get SD to understand that 😄

[silver valley]

fully body, wide view,

[silver valley]

you could try controlnet with mlsd for architecture

[placid mica]

aaaah, good point... I've really only used controlnet for posing, I totally forgot it had other modes as well. Guess I'll read up on that again 😄

thx

[silver valley]

ill sometimes get persons very small when describing the background to much so it got more focus

[placid mica]

yeah, problem is that I need a black background and it's limited how detailed I can describe "black" 😄

[silver valley]

getting then stuff like that:

[silver valley]

does it need to be a plane black background?

[placid mica]

hm, I could always photoshop things I guess... I also keep forgetting that photoshop often is a quicker way of solving SD issues 😄

[silver valley]

yea that too or try controlnet or img2img

but yea depending on what you want ps can be faster xD

[placid mica]

sd is generally not listening to me today, I feel... I think she has that time of the month or something... :/

[polar adder]

You absolute hero! That worked wonders

[silver valley]

Np, looks very good! 🙂

[placid mica]

Hello.... I'm back.... grumble Now... Why won't my subjects EVER listen to me and look away from the camera? Whatever I write, positive or negative, their eyes seems glued to the camera lens 😦

[remote bone]

Luz roja y azul

[south solar]

It's funny. I was actually having the opposite problem earlier.

[fickle remnant]

How do I make my art more lifelike

[left olive]

photorealism:1.3 or photoreal:1.5

[polar adder]

Stable diffusion seems to ignore "text" as a negative prompt. Any ideas?

[vital jolt]

automatic1111, latest, with multiple models, still getting a lot of istock, getty and other proprietary watermarks when i try to create landscape scenes involving gardens and temples, etc. Ugh. Does anyone have a negative prompt string that will disallow watermarks?

[vital jolt]

It depends on the prompt. For instance, if you want to do a tarot card, or a gas station, invariably, they'll have text in the photo shot. You're trying to erase something that's already there. It's a bit like what's happening to me with landscape scenes. I see the Getty and iStock watermark or something evocative of them among many other proprietary watermarks. It's a bit frustrating.

[vital jolt]

I've had some limited success by adding these terms to the negative prompt in Automatic1111:
kanji, text, logograms, writing, hanzi

[polar adder]

thanks I'll give that a shot!

[vital jolt]

You can also try editing one of your images and doing img2img with a version where you've manually edited out the text.

[wind rapids]

Help me guys that how can I make my text to a particular picture

Do we also have to write the "/imagine" before any post?

[high shadow]

does anyone have any luck making characters lay down etc? (not with NSFW models/for NSFW purposes)

Have wanted D&D based characters sleeping on their bed, or laying on the ground near a camp fire. It seems like a mega fight to not have the characters coming out of the ground like a skeleton/zombie or being contorted in an awful way.

[placid mica]

Haha yeah, I hear many people say that. I think it's probably because I often train TI's with my subjects so that I can make different pictures with the same person and my training pictures probably all look at the camera.... So my own fault, in other words.

[placid mica]

No expert, but make sure you have a wider than higher resolution first of all.

[violet olive]

I am looking to generate very different looking images from the same prompt, i.e. if I type woman right now typically all the women look the same.

Would anyone who a way where I prompt "woman" and get outputs with blonde women, asian women, african women, all from the same prompt?

[placid mica]

that would be all dependant on the model you use I'd think. If you use a "small" model with few input images you would get a much less varied response than if you used one of the big ones

[violet olive]

I see interesting, I'll try swapping back to sd 2.1

Thanks

[placid mica]

go to civitai or something and look for the most downloaded checkpoints. That's usually a good pointer

[violet olive]

I've been using deliberate, which is supposedly the highest rated but I'll try some others

[placid mica]

yeah I would also think deliberate had a good base... Hm, do you have much stuff in the negative prompt?

[violet olive]

No it's completely empty

I'm trying to do like iterative generations based on preference, so a high level of variation is necessary

[placid mica]

Weird... gimme a sec, I'll try to load in deliberate and have a quick look

I see what you mean.... 😮

[violet olive]

I've been playing around with seed variation but it doesn't seem to make much of a difference

[placid mica]

yeah tried that too, even on max variation they look all the same... guess it was a very limited photo pool then

[violet olive]

I guess because the model thinks of a "women" as exactly one thing, it's creating like it's platonic ideal woman

but that's an issue for my app idea haha

[placid mica]

just tried with a realistic model (hrl32_hrl32) and those look the same as deliberates women 😄

[violet olive]

There does seem to be significantly more variation inbetween models but that's quite impractical to run as it'll be really slow loading up the different models over and over

[placid mica]

yeah, it is interesting to see how similar everything is though....

tried 3 different models now and they all have alot of the same features. Guess they all stem from the original model

[violet olive]

Yeah :/

[placid mica]

tried to run with 1girl in positive and child, childish in negative, and got a different woman, asian this time, but all looks the same still lol

I give up

[violet olive]

okay thanks anyway brother

I'll try the devs

[high shadow]

yeah i usually try that and it comes out really janky. thanks though!

[tired vigil]

Hey, I've been trying to make full body character designs using blue willow and they always are out of frame or are distorted, either in the face or the limbs are very off. Is there any way I can fix this? Tried deleting the negative prompts and it eased the issue a bit but the faces are still distorted.

[vestal rose]

Hey guys, im trying to run some img2img using 2.1 - 768 model but all results kind of look blurry and a bit ugly, any tips to improve ?

[placid mica]

anyone know why I suddenly get a slighly different image when I turn on Hires.fix now? I used to just get a clearer image, but now there are some small but noticable changes (and sometimes big ones)

[sturdy moss]

Hey! Can anyone share a nice all around negative prompt under 75 tokens?

Or at least some good tokens to use in a negative prompt?

[olive wedge]

Dynamic Prompts and wildcards are very helpful for this. Your prompt would then look like photo of a __type__ woman and each image will have a different type. Even in the same batch.

[violet olive]

Okay cool, I'll look into that thanks

[placid mica]

ooh this new for me! interesting!!

so i could make a huge batch with various types of clothing for instance?

[olive wedge]

Yes!

If you are looking for pretty good faces too, it can really help to have a wildcard file of celebrities and mix it in near the end of the prompt: (__celeb__:0.4)

Or, even better, mixing them together to make the likeness not be so strong: ([__celeb__:__celeb__:0.6]:0.8) which will move towards one celebrity for 60% of the steps, then switch to another for the last 40%, at 0.8 weight.

[olive wedge]

There are some negative embeddings which "combine" negative prompts into fewer tokens. For example, EasyNegative: https://huggingface.co/datasets/gsdf/EasyNegative/blob/main/EasyNegative.safetensors

You just put EasyNegative in the negative prompt and then you can add anything additional that comes up in your generations that you don't want.

[sturdy moss]

Thanks! I've already tried some negative embeddings with various results. I'll give this one a try.

[olive wedge]

There are a few others, like bad_prompt_version2 that I've used and you basically have to add (asian:1.2) to the negative prompt as well, if you want anything other than that ethnicity.. it weights really heavily towards it.

[olive wedge]

I see. Apologies if I had come across as patronising in that case. Never really know what other people are already aware of.

[sturdy moss]

Oh, not at all, thank you, I actually didn't know about this one.

Although, the results are interesting.

[olive wedge]

In a good way or a "hmm" way?

[sturdy moss]

The first picture:
masterful picture of a woman in fantasy magic armor, bokeh, by Greg Rutkowski, intricate detail, glowing runes, dramatic light
Negative prompt: zombie, deformed, alien, monster, kid
Steps: 60, Sampler: DPM++ 2M Karras v2, CFG scale: 7, Seed: 3821791220, Size: 512x768, Model hash: e28b2e61fb, Model: RPGArtistsTools15_sd15V1
Second: - Same, the neg prompt is replaced with SimpleNegative
Third is ng_deepnegative_v1_75t
And fourth is bad_prompt_version2

[olive wedge]

Are you trying to get a portrait or just a torso shot?

[sturdy moss]

With a sample size of 1, it looks like just some basic negative prompt works better than negative embeddings, maybe this depends on the model, of course. I wasn't really going for anything with this picture, just something to test the prompts and embeddings on.

[sturdy moss]

Oh, and the thumbnails are cropped by Discord, if that's what you mean.

[olive wedge]

Oh yes, so it is. I thought you'd posted part of a grid!

[sturdy moss]

I'm just a bit sceptical about the ultra-long detailed prompts with ((((three legs)))) (((((multiple breasts)))))) and all that stuff you can often see in Civitai example pictures. I thought someone came up with something better. 🙂

[olive wedge]

Generated so many grids it's what I see by default now.. 🙂

[sturdy moss]

You mean X/Y grids?

[olive wedge]

No, just the normal grids. I always generate in batches of 16/12/9/6/4 (depending on what will fit in VRAM)

[sturdy moss]

Oh, I see, so you use different batch sizes, not just batch count?

[olive wedge]

Why those numbers? No empty spaces!

I don't use batch count at all.

[sturdy moss]

Have you tested at what batch size you no longer get a benefit in terms of speed?

[olive wedge]

Yes. I actually have a TODO to make a proper benchmark script that figures it out so I can run it while I sleep.

But it can depend. Sometimes I have a lot of tabs open so I will time it so that by the time I have cycled through the tabs (making edits, saving really nice ones), it will have just finished generating.

I have a bunch of these. This one was from when I added medvram which, while it allowed for larger batches, makes the generation time very inconsistent (probably because it causes swap).

[sturdy moss]

I've made a post about this on Reddit some time ago, if you're interested, or maybe you've seen it. So, tldr is that with 4070ti (12g) I got up to +40% speed at batch size 3, and then it plateaus, but doesn't get worse either. Hmm, looks like you get more interesting results with medvram.

[olive wedge]

As you can see in batch 1, the maximum took twice as long as the minimum.

This is on a 2060 12GB.

40% speed in terms of seconds per image?

I'm not sure if s/image is actually a good measure or not, but it is what I use! 😄

[sturdy moss]

Yeah, as in 22 seconds per pic at batch size 1, and 13 at batch sizes 3-8.

[olive wedge]

This is one from quite a while ago (as you can tell, I wasn't using HighRes Fix as the time/image is way lower!)

[sturdy moss]

And also the speedup is smaller with larger pics (768). What's interesting, a 4090 girl with tons of vram reported basicaly same results. No speed increase after batch size 3.

[olive wedge]

Both of these are without xformers too, which might make a significant difference. It's fairly even though.

I mean, if I were using xformers, perhaps the difference on batch would be more pronounced.

[sturdy moss]

Yeah, xformers must play a big role here, as they prevent you from running out of vram and slow things down.

I mean, it's actually less pronounced with xformers.

as it seems

[olive wedge]

I can do batch 12, 512x512 -> 1024x1024 without xformers, so I don't really have a need for it and I've heard it interferes with reproducing images.

Turns out, so does batch size, which is extremely annoying.

[sturdy moss]

Thanks, I never thought about disabling xformers. I'm definitely going to experiment with it now.

[olive wedge]

I just reformatted my data to the same as yours and I see pretty similar numbers. 33% going from 1-2 and then gets less and less improvement over batch 1 (again, likely because of swapping) until it batch 12 where I had to actually restart the UI because it OOM'd, where it is then at 40% savings.

I saw some very, very odd VRAM usage in this benchmark run.

Yes, that's right.. the VRAM usage did not change until I had to restart, where it promptly started varying wildly between runs.

I guess I changed something in between "by accident" (ready for the next start but forgot about it when it OOM'd.. might be medvram causing it).

[placid mica]

Wait... 16?12??? My max per batch is 8???

[sturdy moss]

I've just tried turning off the xformers, but it's way slower for me. With XFormers its 10 secs at batch 1, 6.6 secs at batch 8, and W/O XFormers its 14 at batch 1, and 10 at batch 8. So not really worth it for me. I also overlayed the pics with the same seed in photoshop, and they are indeed not the same, but the difference is just a few pixels, and there's no way to see it with a naked eye or tell which one is better.

[placid mica]

so xformers make the rendering go quicker?

without a significant loss of quality?

[sturdy moss]

Pretty much

[olive wedge]

Useful to see. Is that with overlay layer mode?

[sturdy moss]

Difference. Oh, and that picture is wrong, it's the difference between batch size 1 and 8. Here's the difference between Xformers on and off. Definitely more pronounced, but still too little to matter imo.

[olive wedge]

Tempted to throw that into ControlNet and make a landscape scene with it.

[sturdy moss]

Hah, sounds like a nice idea

[olive wedge]

Visible difference when flicking between the two but indeed, probably not something you'd notice without the comparison.

[sturdy moss]

Yeah, my main point is that I can't tell which one is better.

[olive wedge]

Okay, I need to figure out which extension is causing this issue. My UI is freezing something fierce since I added a bunch of them today.

Thankfully, I have my SD installed in a way that makes that super easy!

Submodules FTW.

[sturdy moss]

I heard the one that lets you draw prompt masks caused problems in chrome.

[vivid saffron]

Does anyone know of any tips to get a fully focused image? every model I use always has such a shallow depth of field. No matter what I've tried I can't seem to get a large depth of field.

[sturdy moss]

Try putting bokeh in the negative prompt

[vivid saffron]

I've tried that, no luck

[placid mica]

I've seen that before... what the... is a "bokeh" ???

[vivid saffron]

I tried deep depth of field, all types of camera lens settings.

[vivid saffron]

look at the image above with the girl, do you see the blurred out background, see how the lights are circular, that is bokeh

[fickle vector]

So i am trying to generate a woman looking at the sunset and it keeps generating(albeit very pretty) houses and landscapes. How do I make it like, not do that?

Prompt:
Picture of a young redhead woman watching the sunset, girl in focus, petite, wearing bathing suit, figure, beautiful face, beautiful body, wet, covered in water, soaked, (artstation:1.2), (epic realistic:1.2), (hdr:1.4), intricate, detailed

Using Deliberate model(although I had similar issues with other models)

Even with negatives I am still getting houses and such

[placid mica]

Never heard that word until I started with this ai-lunacy... 😄

[vivid saffron]

Its a photography term. You should check out some photography basics since it could really help your artwork and understand a bit more about lighting, focus,etc.

[fickle vector]

nvm i fixed it lmao, easy diffusion doesn't accept (hdr:1.4) or similar prompts hence the issue

[sturdy moss]

That's really weird, I've just tried that prompt with deliberate, and it makes women. Are you sure you don't have some lora on or a control network with a picture of a house?

[fickle vector]

it's because i use easy diffusion

easy diffusion doesn't accept (hdr:1.4) or similar prompts, which is what confused it

works now!

[sturdy moss]

Can you give an example of the prompt? My girlfriend had a problem once, SD generated all kinds of weird shit for her. Turns out she had Gorillaz in her prompt, as in the Gorillaz music videos art style. And SD decided that it needed to try and sneak in parts of actual gorillas into the pictures.

[olive wedge]

Interesting bug. If you put in no prompt, the batch size gets ignored and it only generates 1.

[silver valley]

depends mostly on the denoise strenght and the upscaler used

[cinder olive]

Masterpiece, best quality, a slender young male warrior wearing heavy furs and wearing a helmet, fantasy, dungeons and dragons, detailed tundra in the background, arctic background, detailed fur cloak, detailed heavy furs, (wearing a detailed steel helmet:1.41), digital painting, character painting,

Negs: shield, Big Breasts, NSFW, lowres, bad anatomy, bad hands, bad faces, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, monochrome, (face, showing face:1.41), gold

I keep trying to get it to give me an enclosed helmet, but...

What could I be doing better?

[silver valley]

try iron visor as tag

[olive wedge]

Perhaps try a specific helmet type e.g. kettle helm or bascinet helmet

[fickle vector]

I am struggling to get the ai to create what i want when it comes to multiple people. For example like someone fighting someone else, or someone being eaten by zombies. For the latter example, it just makes all people in the picture zombies, regardless of what I put

this is a problem with all models i have tried so far

[kindred garnet]

any advice how i can remove the 3rd ear?

from the small capybara

[fickle vector]

photoshop and/or inpainting

[kindred garnet]

mhh i never used inpainting, probably have to test it 😄

thought i could do it with a neg prompt maybe

[fickle vector]

I have hardly used it, photo shop is likely easier

[fickle vector]

that would also work!

[kindred garnet]

yeah thats why im asking here 😄

if u have any idea for a neg prompt let me know

[fickle vector]

three ears, extra limbs, wrong anatomy, etc.

[kindred garnet]

tried those allready, mhh... :/

[fickle vector]

oof, your best bet is to fiddle with the prompts and negatives, try to learn inpainting, or just use photoshop/other photo editing tool

[kindred garnet]

these are my prompts atm

ultra-realistic, highly detailed,Fantasy image of glorious capybara warrior holding a giant diamond sword, metal amor, critter, animal, horror, full body view, fantasy style, illustration, uhd, 8k, raytracing, highly detailed, perfect artwork

extra ears, bad face, Mutilated hands, Poorly drawn face, Deformed, Bad anatomy, Cloned face, Malformed limbs, mutated, ((mutation)), Missing legs, Too many fingers, low details, low quality, text, watermark, logo, signature, author name, cropped, not centered, out of frame, humanoid, happy, 3d, render, human hands

[fickle vector]

maybe trying "critter with two ears" or something along those lines

putting two ears in the prompt

[kindred garnet]

k

[fickle vector]

I'm still relatively new to this, but fiddling around and throwing stuff around until stuff works is how I've generated my favorite things

[kindred garnet]

now idea why the warrior get long hair now, but looks kinda cool xD

[olive wedge]

You should look into ControlNet and perhaps Latent Couple if you need that level of fine control.

And perhaps join some poses together from this pose pack: https://civitai.com/models/11573/zpose-fighting-styles-1-pose-helper

[fickle vector]

I am not familar with contol net or latent couple, do you have any links to videos or articles explaining them, or are they self explanatory

[olive wedge]

ControlNet is a set of preprocessors and model that use the processed images to direct the generated content (OpenPose is the specific model you'd be looking for this): https://github.com/Mikubill/sd-webui-controlnet

Latent Couple allows you to target subprompts (via the AND keyword) towards specific regions of the image. Useful for generating different people in the same prompt e.g. an oil painting, 2 women AND 2 women, blonde hair AND 2 women, brown hair would (generally) generate images with a blonde woman on the left and a brunette women on the right: https://github.com/opparco/stable-diffusion-webui-two-shot

They are very powerful together and this tool for creating the regions is useful too: http://badnoise.net/latentcoupleregionmapper/

Fair warning, Latent Couple is a bit odd because you do have to specify that there are 2 women in each part of the prompt or it will merge them together and you will get 1 person who is blonde on the left and brunette on the right.

[fickle vector]

funky, but I will try that!

Speaking of me and struggling with getting the ai to do what I want, I am struggling to get the AI to do gore of any kind. this is an issue with multiple models and settings, and prompts, any advice?

Like fighting zombies, and horror and anything like that

[olive wedge]

I'd imagine you'd need a specific model (or embedding/hypernetwork/lora) for that kind of content.

Are you aware of http://civitai.com ?

[fickle vector]

yup!

I am using it to find models

The problem is there isn't really gore/horror models i like. Maybe I learn to train my own model or something

[tranquil shard]

Hello. WebUI now have this cool loras catalogue where you can make preview images for them. And I wounder - is there a plugin or something to create similar catalogue for prompt words and preview of what they do? Making track of all style/subject/object prompts become kinda hard, as there a lot to experiment with and I'm currently just tracking them in txt files.

[rose crown]

anybody know the difference between the beginning and end guidance strengths in depth controlnet?

[lusty crow]

I have a long prompt I try to shorten down and optimize, I saw "blue hair" and then remember that I saw "blue_hair" mentioned as tag for Anime, but when I added the underscore it jumped from 80 to 81 tokens, and to remove a comma is reducing the prompt with one toke each, So is it better to use (blue hair) (green eyes) rather than blue_hair, green_eyes as I seen in some example prompts?

[lucid knoll]

super quick question, is there a way to make SD make multiple pieces? I want to leave my house and let it make like 10 pieces while im gone
or do i have to keep pressing generate after each piece is done?

[tranquil shard]

set the batch size to whatever you need

batch count 10 will generate 10 pieces

[lucid knoll]

ty 🙏

[tranquil shard]

if you are experimenting with prompts or params, you can also try matrix x/y/z script to create multiple images with different params.

[lucid knoll]

thats pretty advanced stuff eh, thanks !

[zenith anchor]

Hello

i am having some trouble generating images of a "chibi" artstylr

can i use a template or smth similar?

[tranquil shard]

I doubt there are models that are good at chibbis. You probably will have to use some lora for that.
look here for example
https://civitai.com/?query=chibi

[zenith anchor]

ye i already did

[tranquil shard]

also, look at this lora if you want it it to be really simple art

https://civitai.com/models/7267/childrens-drawing-lora

[zenith anchor]

i cant figure out how to import the model

[zenith anchor]

i dont want it to be simple but cute

is there a way for me to turn a specific image into that type?

[tranquil shard]

you put models to the stable-diffusion-webui\models\Stable-diffusion folder. and select it from the models drop down list
and you put loras to the stable-diffusion-webui\models\Lora folder. and select it in loras catalogue (given you have latest version of webui)

if you did used correct lora and it still dont generate intended style - try to lower or rise it's weight and weight of prompts that triggers them.

if you use lora that was trained on realistic model on anime-style model or vice versa you may have trouble making lora to change anything without playing with the weights first.

[zenith anchor]

will try this!

gimeme a few minutes

[zenith anchor]

https://civitai.com/models/6889/chibi-style

i am supposed to put this into the lora folder right?

[tranquil shard]

yeah, than click that icon and select lora tag and click on your lora - it will add lora word to the prompts

refresh if you don't see your lora. if you don't have that icon and tab - you need to update webui client.

and add trigger word to the promts. in that case - 'chibi'

[zenith anchor]

its generating

[zenith anchor]

it worked!

Thanks for this!

now is there a way for me to convert a already existing image into this style?

[tranquil shard]

no problem. you may try img2img to change existing image. but it may be hard to find right params for specific styles. you'l need to experiment with that.

[zenith anchor]

ohk

thanks for the help!

[sturdy moss]

Using Control network alongside img2img might help as well.

[smoky totem]

This is prompting-help, mental-health-help is in another channel.

[craggy steppe]

Hi, can someone help me with this. The style is applying properly but there is some blue shade around the waist. How do I remove it? Tried negative prompts like blue, blue shade but they are not working

[wheat swift]

inpaint it out

[craggy steppe]

I need to do batch img2img

[west blade]

I would like to ask why my RENDER will be blue?

[wheat swift]

#🏞｜general-with-images message

[tranquil shard]

I guess it's VAE issue. may be wrong. Try selecting different VAE.

[craggy steppe]

what's vae?

[tranquil shard]

I don't know all the rocket science behind them, but they corrects image color. Some models comes with .vae file - in that case you should use that vae with model. You can change it in Settings - Stable Diffusion - SD VAE.
You may need to adjust precision when using VAE. I have to use --no-half-vae starting argument or gen throws errors time from time.

Issues with VAE are usually dithered image or with colored spots on them.

[lucid knoll]

hey, so im up scalling my pieces from 544x960 to 1088x1920p, with like 30 hires steps
that way i can get 1080p results
each pieces takes around 45 mins, but i end up getting 1080p results which is great and sharp
BUT
i noticed that the regular steps are great and have no issues, but as soon as it switches to hires steps and it starts upscalling, it starts adding belly buttons or adding fat folds or extra legs
anyone else having this issue? anybody got a fix for that?

[past dawn]

Hi, does anyone know why SD would generate the exact same person repeatedly? I'm trying to generate classification images for DreamBooth, so I'm playing with txt2img to get it right and it's like it's not even randomising the seed or something. Images looking similar is fine, images featuring the same person though, why!?

Actually, the seed is almost the same, just counts up 2-3 digit increments... Why does random seed not mean random seed!?

[silver valley]

thats not the same person for me

[past dawn]

You sure it's not prosopagnosia? 😆 Just kidding. They look very similar to me. Also, the seeds are almost identical and sometimes actually identical... Very frustrating.

[silver valley]

you need to set the seed to -1 for randomising

[past dawn]

The seed is set to -1. As stated above, the seed changes as it iterates, but only the last 2-3 digits change and sometimes the seed is entirely unchanged, despite the -1 seed.

[silver valley]

do you generate in batches ?

[past dawn]

Yes

[silver valley]

yea then it uses the seed for the whole batch i think

[past dawn]

So if I want to generate thousands of classification images, how am I supposed to randomise the seed for each image. I surely can't generate one at at time to randomise the seed? lol

[silver valley]

okay not matter what i try, the seed just get +1 higher then the one before if i use batch count

[past dawn]

Yes, same for me

In fact, even between gens it seems to create the same person again

[silver valley]

did you used names in the prompt ?

it may recognize someone

[past dawn]

No, I don't use specific names

Just 'photo of beautiful woman' and such

[silver valley]

hmm

[past dawn]

Here are two separate gens, batches of ten.

So apparently, generating a completely new set doesn't make it more random...

[silver valley]

Maybe try a different model

[past dawn]

Yeah, the model is a merged from two checkpoints, I have different ratios of the two and seems to produce a similar phenomenon... Maybe there's something wrong with the merge settings, or perhaps the checkpoints used. Not sure honestly, I'm going to have to just play around with it I guess!

[olive wedge]

You're not pulling in an embedding or hypernetwork unknowingly?

[past dawn]

I'm not sure I could do that unintentionally, could I?

I haven't made any embeddings and my hypernetworks have nothing to do with human faces/photographs...

[olive wedge]

Embeddings are pulled in just from their name being in the prompt, so if you have an embedding called "beautiful woman", it'll be used.

[past dawn]

Oh, I had no idea... well I don't think I ever created an embedding anyway, so I don't think so.

[olive wedge]

It will say in the param info box under the generation if it is using an embedding:

Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 3604834669, Size: 512x512, Model hash: abbb28cb5e, Model: elysium_Elysium_V1, Denoising strength: 0.8, ENSD: 72, Hires upscale: 1.5, Hires upscaler: Latent



Used embeddings: easynegative [119b]

Time taken: 4m 49.70sTorch active/reserved: 5447/6074 MiB, Sys VRAM: 7295/12028 MiB (60.65%)

[past dawn]

Oh, nope no embeddings apparently...

There's no conceivable way checkpoints can contaminate each other is there?

Kind of a stretch, I know...

(besides actually intentionally merging them, obviously!)

Or training DreamBooth for example, nothing actually gets written to the base model?

[placid mica]

Ive honewstly wondered myself a few times 😄

but it shouldnt be possible

[past dawn]

Yeah we humans have a tendency to connect unconnected dots.

[olive wedge]

It shouldn't and I've never seen it (I have integrity checks on my files which would have picked it up).

And if they somehow merged/half-loaded in VRAM, I'd presume any generations would be garbage!

[obtuse quiver]

One message removed from a suspended account.

[olive wedge]

, has no specific meaning in a prompt. | outside of brackets creates a prompt matrix: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-matrix

[minor field]

Anyone know a way to prompt specifically what colour each clothing item should be? Would clip skip help me here maybe?

something like white shirt, black shoes, white socks tends to just result in random other items changing. like in the image above, the jacket is white

will be training a model to try create a consistent character, it's just interesting how it seems like having white elsewhere in the prompt makes other things white.

[orchid scaffold]

hey guys, can someone help me with this duplicate issue?

a cat wearing a top hat and monocle playing a grand piano

[sweet jacinth]

a photograph of circus people living in the 70's, during their show, detailed costumes, flame flower, Kodak portra pro, detailed pictures, realistic skins, reflection, realism, professional color grading, clean sharp focus

[atomic flume]

I'm a little confused my SD keeps creating multiple characters in my image. Any suggestions

(one character:1.4), (modelshoot style), (from_bottom:1.3), punk rock clothing, 1980s, punk coat with glowing collar, (ripped jeans:1.2), (combat boots:1.2), 25 year old, female, (a rpg character portrait:1.2), art by artgerm and greg rutkowski and magali villeneuve, analog style, (mdjrny-v4 style), highly detailed face and eyes, thick thighs, fully body, interesting character poses, thick lips,

Negative prompt: lowres, (bad anatomy), (error body), error hair, ((error arm)), ((error hands)), ((bad hands)), error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, ((missing legs)), error lighting, error shadow, error reflection, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, (watermark:1.3), username, blurry, ((error eyes)), ((bug eyes)), ((bad eyes)), bad mouth, error mouth, (error face), (((ugly))), text, (midriff:1.5), (cleavage:1.5), (multiple characters:1.2),

Steps: 50, Sampler: DPM++ SDE Karras, CFG scale: 12, Seed: 1427471437, Size: 800x1080, Model: neverendingDream_bakedVae

[olive wedge]

Having a quick look at the model you are using, I think it would respond to booru tags e.g. 1girl

[atomic flume]

Ah ok thank you

[olive wedge]

Also, I think something in your prompt is hinting at multiple people: interesting character poses

[atomic flume]

Ah

[olive wedge]

Should probably be interesting pose

[atomic flume]

it does one person now but the poses are just crazy lol

https://i.imgur.com/56a6aky.png

lol

[olive wedge]

Heh, "be careful what you wish for" comes to mind.

[kindred garnet]

why its so hard for sd to hug animals

i tried several prompts but always get double animals and not a human

made this one with midjourney but had no chance to replicate something similar with SD

[proud anchor]

Using a1111 webui if matters. Is there a way to write the seed into the prompt? as well as other settings?

[meager girder]

is there a model for firearms?

[idle tinsel]

Very simple question. How to get stable diffusion to create only a fully white, completely plain, background behind the subject? Basically want a png

[tired vigil]

For the location, put photoshoot studio, solid color background

[sharp aspen]

anyone has any idea how to make this effect with SD? watercolor? white background? but how about the fading, blurry layer?

[cunning arrow]

what is the best negative prompt for nature

[kindred garnet]

@sharp aspen u probably need a fitting model

[past dawn]

Anyone know how to prevent dreambooth generating new classification images and just use the ones already present in the directory?
There doesn't appear to be a setting anywhere...

[placid mica]

any good ideas on how to get a woman to swim instead of just standing in water?

[past dawn]

My assumption was that one just put classification images in a folder, enough to cover the 'Classification Images Per Instance Image' number chosen in the settings and it would just load it up, but even if I vastly exceed the required number, it loads them all up and then proceeds to generate 30-40ish of its own classification images.

[placid mica]

pretty sure this isn't the right channel, unless there's something you can do in the prompt to fix that. Since I have no clue what you're talking about I wouldn't know, just trying to be helpful here 😄

[past dawn]

Sorry TinyTim, we're both talking to an empty room it seems.

[placid mica]

lol it takes time. It's sunday, people are buzy with various things or in different timezones. I pretty much always get a good answer eventually, but it can take hours sometimes

[past dawn]

Yeah, these Discord channels can be pretty good, but sometimes you just want answers now, lol

I'm ready to click 'train' and just want it to not fck around, but need to wait on an answer before I can begin!

[placid mica]

ah yeah I can see the frustration... I've only tried dreambooth training one time, and the result was... not great... Havn't gathered enough motivation to try again yet 😄

pretty sure my main wrongdoing was to not make any prompts at all for the dataset

[past dawn]

You training a hypernetwork?

[placid mica]

no, a checkpoint probably, was planning to learn to make lora's but didn't really get there 😄

then I got sidetracked by women in cutoffs.... I have a weak mind...

[past dawn]

When you say you didn't make prompts for the dataset, you mean the txt file with tags in them?

[placid mica]

yeah. I just had face profile pictures, but all the tags was just the filename, no caption of any kind

I assume now that the ai probably can benefit from knowing what i'm serving it

[past dawn]

Ermm.... yeah

lol

[placid mica]

😄

it worked great that way when I did TI so I just carried on in the same manner. Not great 😄

[past dawn]

You can autotag, but it's awful and doing it manually is so labourious!

[placid mica]

Yeah, figured I'd wait until later next week when I have most of the week free so I can make a proper dataset and try again

and probably watch a youtube guide or something first! 😄

[past dawn]

Can't know too much!

[placid mica]

yeah, learning by doing works great for simple things. SD is not a simple thing....

[past dawn]

No, it certainly isn't.

What was the issue you were having with the Jesus women who walk on water?

[placid mica]

no, I'm trying to make a woman actually SWIM in the ocean, but they all seem afraid of getting wet and just stay with water up to their thighs or so

[past dawn]

I see, so you want the water neck height?

I suppose you're just going to have to be creative with your prompts, or play with different models.

[placid mica]

tried simple things like "swimming in the ocean" or "water up to her chest" but every time the result is the same

yeah, I have a feeling it could be the model or TI I'm using that's at fault... I should learn to try different models immediatly when trying a new concept

[past dawn]

and also combining models that both have aspects you like

It's all about the training data at the end of the day, it won't show you what it hasn't been trained on

[placid mica]

ah yeah, model merging is another thing I havn't really looked into yet. Just started with SD a week or so ago

[past dawn]

It's useful, I'm not exactly sure what witchcraft goes on under the hood, but trial and error has produced interesting results.

[placid mica]

yup, as soon as I removed my TI the woman was swimming. Guess I trained it wrong

[olive wedge]

Have you given SD an few images of women swimming in water to interrogate? Might give you some hints what it wants to be prompted with.

[placid mica]

Didn't even know that was an option! So is THAT what the interrogation-button does??

Man the learning curve.....

Makes sense when I see now that there's "interrogate clip" and "interrogate DeepBooru" in the img2img tab...

[placid mica]

Man this is a gamechanger... Such a small thing, and yet such a huge impact on how I can get the images I want... Thanks!!!!

[olive wedge]

You're very welcome! 😄

[kindred garnet]

Anyone has an idea why hugging animals is nearly Impossible or any tag i could use? 😅

I always get 2 animals or 2 humans but never the scenario i want

[kindred garnet]

tried several times to replicate something like this but had no chance

[past dawn]

Hey guys, anyone know how to start training in dreambooth without it generating class images? I'd like it to use the images already present in the folder, which it does, but then it still it generates another 30 or so extra images! How do I skip that step, there doesn't appear to be a setting anywhere...

[lone badge]

It generates pictures when there is less class image in your directory than in the settings usually. For me, it was usually that i didn't set right the path to the class pictures, and it would not find any so it would generate them. Outside of that though, if i had enough class pictures i never had that problem

Maybe more appropriate for #🔧｜finetune though

[past dawn]

Always generates images, no matter how many images are in the folder. I will jump over and ask there too. Thanks!

[spice delta]

Hey , anyone know where to find the best quality prompt examples