📝｜prompting-help

[hybrid thunder]

models -> Lora

then reload the SD webpage

or update

[tired vigil]

yes

most changes update after you do the little reload on the page

[hybrid thunder]

idk doesnt work

should have like diff eyes and hair

Did the lora activate?

I put in green eyes, glowing eyes

I think it would be similar without the lora anyway, no ?

[tired vigil]

@hybrid thunder did you go in the lora tab and click on the blood elves lora

it will type some stuff in your positive lora:something:1

[hybrid thunder]

oh

thx

Just click on it once enough ?

Holy shit

result is a lot different

wtf

where does that hand come from if both hands are at the table? xD

[tired vigil]

blood elf power

[hybrid thunder]

But for real how to fix things like that

[dusky stream]

Is there a way to save specific configurations for x/y/z plot scripts?
If possible, it'd be convenient to load up a specific list of my checkpoints without having to load all of them and then trim out the ones I don't want.

[tired vigil]

use inpaint if you really want

watch some video on it, not that complex to use

otherwise just generate a new image

[hybrid thunder]

What are x y z plot scripts ?

[red hemlock]

if i want to use comfyui with a v1.5 model, what node do i need to take ? "CLIP Text Encode (Prompt)" ?

[lyric goblet]

Any advice on how to generate lever-action rifles?

[lyric goblet]

Why are plague doctor masks so hard to generate?

[rancid willow]

Any idea how to hide one eye behind the hair? I keep getting "slit" of the eye showing through the hair no matter what I try (I'm completely new to SD :D)

Img2img
Prompt:

((Eye glowing blue color)), ((silver hair)), ((best quality)), ((masterpiece)), (detailed), (lighting:1.2)```

Negative Prompt

3d, cartoon, anime, sketches, (worst quality, bad quality, child, cropped:1.4) ((monochrome)), ((grayscale)), (bad-hands-5:1.0), (badhandv4:1.0), (easynegative:0.8), (bad-artist-anime:0.8), (bad-artist:0.8), (bad_prompt:0.8), (bad-picture-chill-75v:0.8), (bad_prompt_version2:0.8), (bad_quality:0.8)

[gaunt sail]

Hi everyone, how to promt art in which there will be 2 characters, which are dofferent Loras, and they should be on the same art?

[silver valley]

Hey therefore you should try the Tiled Diffusion extension. There you can activate the regional prompt control.
In it you can make squares and choose each lora for each square so they don't mix up

[tired vigil]

hair over one eye is the prompt for it

bunch of anime models are trained with danbooru tags

also just encapsule your prompts inside one thing lmao looks weird like that :p

(Eye glowing blue color, silver hair, best quality, masterpiece, lighting:1.2), (detailed)

And I'd do ''Blue eyes, glowing eyes'' instead of one prompt for both

and you might have too many negatives, 3 or 4 textual inversions is usually what you need for the best quality, more than that it can cause your prompts to not be as responsive or lower the quality

but can also depend on the model, thats why x/y/z plots are nice, you can test a combination of prompts for the same image

[gaunt sail]

interesting, thanks)

[frosty aurora]

Does anyone know how I can take this boolean and bind a true statement so that Ksampler stops if there are multiple persons in the image?

[gaunt sail]

My pc freze from using regional control ((

[silver valley]

Its to late for me now sry. But you can post your issue with the info what's your GPU and how much RAM you have in the #🤝｜tech-support channel.
I can look at it tomorrow if nobody answers before

[gaunt sail]

I don't think I can use it at all

I use this in order for stable diffusion to work, I can upscale only to 768 , and well, seems like regional controls are too much
set COMMANDLINE_ARGS= --lowvram --always-batch-cond-uncond --opt-split-attention

[prisma wren]

how do i make a negative prompt on here? i understand how to on other ai engines but idk the way to on this one

[gaunt sail]

Well, I figured my problem out

[hollow igloo]

What model is this guys

[tired vigil]

Could be Dalle or SDXL

On 1.5 dreamshaper might be able to get that look, hard to tell

[hybrid thunder]

What does easynegative, bad artist and bad prompt prompts do ?

[rancid willow]

Not a clue, I've been copy pasting stuff i've read online

Trying to grasp how things work 😄

[tired vigil]

those are textual inversions that you have to downoad and put them in emebedding folder for them to take effect in your prompt, merely typing those words wont do anything, you can download those textual inversions from https://huggingface.co/nolanaatama/embeddings/tree/main

[rancid willow]

Ooh okay thank you for the explanation! I'll look into that 😄 Thanks

[tired vigil]

here is an example of how to use them in your prompt box... i used the easynegative and bad hands-5 in this example, they go with other regular prompts seperated by comma, just make sure textual inversions names match the actual file name that you have downloaded and put them in embeddings folder

textual inversions are basically trained on image data sets not really text words, easynegative basically covers most of the basic fixes and bad hand is for fixing weird anatomical features

also something to consider, when you use too much textual inversions that can affect your actual model from how its intended to work

[rancid willow]

Ooh cool, thats very good information. Thank you!

[tired vigil]

welcome

[rancid willow]

I'll give that a shot right away

[keen sky]

hi everyone, nice to be here. Is there anyone who could help me at one step , i am integrating the API of dreamstudio with my telegram bot , but the issue i have is that i dont really know how and what to set into the section "text prompts" ? i have this as sample : {"text": "text_prompts", "weight": 0.5} -but like i said i dont know it so everytime i use in telegram prompts to create a img file , it happens the same img file comes as an art of text documents , what means the prompts that are typed in telegram group it doesnt takes , it uses only : {"text": "text_prompts", "weight": 0.5} - using it for workflow on pipedream

[keen sky]

SOLVED , THX

[finite pendant]

My generations are always kinda blurry like this, what do I put into he prompt to fix that? I already have "blur, blurry, out of focus" in the negative prompt.

[blissful sapphire]

when I am upscaling how should the prompt than look, the general direction. Is it best just to mention the details?

[austere rapids]

how can i make results look like those?

It's a glue gun

I made it in Bing AI Image Creator

I need these types of creative results

no local model I can find has it

best ai generated apple I've ever seen

[tired vigil]

What do you mean by upscaling

Where you doing that

[tired vigil]

Send a picture of your settings

The ones under the negative prompt

[finite pendant]

ControlNet is disabled

[tired vigil]

Did you crop this image

It's not 1:1 like 512x512

And can you not hires?

512x512 is pretty small, things might not have enough resolution to look sharp

[finite pendant]

I did crop it

[tired vigil]

If you can hires try latent, 2x, 10 steps

[finite pendant]

Rest at default?

[tired vigil]

Yeah

I mean those for the hires settings

You're settings are looking normal

What positive and negatives do you have

[finite pendant]

Yea. I don't do 1024x1024 because A. I was told SD 1.5 was trained on 512x512 and B. my GPU isn't strong enough to output it consistently

[finite pendant]

Pos: Three graves, three gravestones, desert, afghanistan,
Neg: blurry, blur, out of focus,
Extra prompt:

• • Pos:
• • Neg: ||disfigured, kitsch, ugly, oversaturated, grain, low-res, Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, ugly, disgusting, poorly drawn, childish, mutilated, mangled, old, surreal, text, blurry, b&w, monochrome, conjoined twins, multiple heads, extra legs, extra arms, fashion photos (collage:1.25), meme, deformed, elongated, twisted, fingers, strabismus, heterochromia, closed eyes, blurred, watermark||

I'm currently trying to make icons so maybe I also have to use a different moel? Using dreamshaper_8 currenly

[tired vigil]

Yeah stick to 512 or 768, and let hires upscale it

[tired vigil]

I'm confused, where is this extra prompt going

[finite pendant]

Here in "Negative Prompt"

I saw a YouTube tutorial use that with those tags so I made it too

[tired vigil]

Try adding high quality, photograph, 4k in positives

[tired vigil]

Oh okay those are styles :p

[finite pendant]

Apologies if my messages rn are full of errors, I type kinda quickly ad my GPU is struggling it rn

[solemn needle]

how i merge the characters?

[tired vigil]

Also @finite pendant I probably asked before but what is your gpu

[tired vigil]

Oh I thought you wanted a real look

Didn't see that message

Then put what your want in the positive

Unless you have a very specific model for it, you need to specify

Dreamshaper should be able to

[prisma sparrow]

Is there a clear best way on changing art styles for the same picture? Should I switch models, loras, or something else entirely to get the same image in different art styles

[finite pendant]

Radeon RX 570 (8GB VRam)

Also,

RuntimeError: Could not allocate tensor with 1073741824 bytes. There is not enough GPU video memory available!

[tired vigil]

https://tenor.com/view/harold-pain-pain-kiszkiloszki-harold-meme-gif-27704917

Friend we are on the same boat

[finite pendant]

Ik ik

Still surprised I managed to run SDXL, even if the image was nonsensical

[tired vigil]

@finite pendant what are your webui arguments

[finite pendant]

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=--medvram --no-half-vae --opt-sub-quad-attention --opt-split-attention-v1

call webui.bat

[tired vigil]

Do you need no half Vae?

[finite pendant]

Idk what that is

[tired vigil]

Try without it
And --lowvram instead of med should save more vram but it's slower

[finite pendant]

I was just told to put it in

[stray aurora]

what reason is there to use this instead of dalle3?

whats the difference

[finite pendant]

One is free and uncensored, the other is paid and censored

[tired vigil]

You're limited with dalle

Basically that

[finite pendant]

Ok so I ran it once more and got this with these settings

[stray aurora]

this one is uncensored?

no limits at all?

also dalle3 is free on bing though

[tired vigil]

Stable diffusion is local

You install and it doesn't use the internet

[stray aurora]

oh lol
can i do it on phone or pc only

[finite pendant]

Depends on the model, (you can get them from, for example civitai), but generally uncensored.
Dalle3 on Bing isn't free

[stray aurora]

i mean u get 25 free uses

[tired vigil]

PC only

[finite pendant]

And ST is unlimited

[stray aurora]

aight ty

[finite pendant]

PC host only. You can still access it through phone.

[stray aurora]

how?

[finite pendant]

There should be a guide of it on the site

[tired vigil]

But you'd be hosting to yourself no?

[finite pendant]

Well... you could rent a Server

[stray aurora]

how accurate is stable diffusion compared to dalle3 though? quality / prompt accuracy etc

[finite pendant]

Heavily depends on the model

[tired vigil]

If you want icons then remove the photograph in the prompt

And put what your want

[finite pendant]

SDXL appears to be very close to DallE3, while 1.5 is worse, yet has a ton of customization through community made models

[stray aurora]

wdym? there are a few models?
stable diffusion xl is the latest one isnt it

ah ok

[finite pendant]

SD 1.5 is the base model. There are a lot of models that are trained on SD 1.5 + extra

For example divineAnimeMix is trained on anime, while dreamshaper is more general purpose

[stray aurora]

which one do u recommend? im a little confused
thought theres just one version thats the latest and best
but apprently not

[tired vigil]

@finite pendant if you want to upscale images you can do tiles with controlnet
You should be able to find a tutorial on YouTube explaining

[finite pendant]

I have Gigapixel AI, but I'll look into it!

[finite pendant]

If you want the newest and have an NVidia card, use SDXL.
If you want flexibility with community made models, use SD 1.5.

Changing modules is easy, as it's just a drop down menu in the webui

[stray aurora]

dont think i have an ncidia card, i might just use firebase to host for me lol

[stray aurora]

ah i see

[stray aurora]

are all of those unrestricted?

do u have like a picture of how a local stable diffusion model looks like? im confused
is it like a web application?,

[finite pendant]

SDXL I think yes.
SD 1.5 base model yes, community model depends on which model you use.

[tired vigil]

Also do at most 30 steps with that samplers, it's not necessary

[stray aurora]

like when running it does it run in the browser or what

[finite pendant]

Here is how it looks on my end right now. It runs on 127.0.0.1, or in other words it runs on your PC

[cursive delta]

anyone know if it's possible to do qrcodemonster/illusion diffusion style gens with 2 existing images? maybe in img2img? it doesnt seem to work for me

[stray aurora]

ah ok
i think ill be able to run it on firebase then i think? idk

what should i do if i dont have an nvidia card

[finite pendant]

On the Github, there's a guide for non-NVidia cards. You basically have to edit a file once you pulled it.

I haven't heard of firebase.

[stray aurora]

https://www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=https://github.com/CompVis/stable-diffusion&ved=2ahUKEwjUxc7Og_6BAxWugf0HHfNxAVoQFnoECBUQAQ&usg=AOvVaw1W5Rmuvz2Cs3Hj7upWBUMT
this one? there are a couple

of compvis

[finite pendant]

This
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs

That one is for AMD

[stray aurora]

amd?

whats that

[finite pendant]

Another huge GPU card creator

[stray aurora]

ah ok ty

is there any online stable diffusion by any chance

[finite pendant]

I don't know, you're better off asking in #1072238304042438758, and when you want to install it, you can ask #🤝｜tech-support for further instructions ^^

[hybrid thunder]

Stable Diffusion lets you modify almost everything, and install a lot of different configurations, you have much more freedom basically

[hybrid thunder]

I like them better for CPU tbh

I have combination of AMD Ryzen 9 7900 cpu and RTX 4070 graphic card

[worn copper]

Hmm... any suggestion how to force main character not to look into camera? I was trying with negative prompts like "looking at camera","... viewer" etc but in no avail. I guess it might be model related but maybe there are some photographic-related terms or tricks I am not aware of? 😄

[orchid ore]

probably just put more attention on it. (looking at camera:1.5) or more.
What model are you using btw? If not secret?

[worn copper]

Ghostmix atm. 🙂

[orchid ore]

i have different model, but "man is observing horizon" works

[tired vigil]

I have gone through that question before :p

"Shy, looking away, thinking, posing" are the answers me and some other people decided on

"Commercial pose, athletic pose" too

[worn copper]

Haha, actually observing horizon works too! 😄 Thanks you both guys! Will be experimenting.

[tired vigil]

That is specially good if you want the character completely looking away from the camera

like a back view

[wide copper]

https://tenor.com/bI9Sy.gif

[finite pendant]

I got a Ryzen 5 2600 and a RX 570 (8GB VRam). I do plan on getting something better, though I don't have much money

[celest musk]

any tips for getting a specific item being held in a subjects hand? i'm trying to generate a fantasy character holding a dagger, but it always ends up embedded in clothing, hair, or just floating somewhere random. i've put just dagger, in the prompt, and i've also tried holding a dagger, and wielding a dagger, which all seem to produce the same results and never actually in the character's hand.

[finite pendant]

Perhaps holding, holding weapon, holding dagger, dagger might work?

Anyone know how I can recreate HoI4-style elements with a prompt? They look like this:

[viscid vale]

Hey guys- anyone can help me I have a product for example shoes - I want to do inpainting like a girl is holding that shoes - how can I do it in auto1111?

[tired vigil]

Try holding a dagger by the hilt

It's the only prompt that minimized those artifacts for me

Still not perfect

SD is very bad with weapons in general

[celest musk]

Yeahh any time I manage to get a generation with it actually in the subjects hand it's always being held by the blade lol. Thank you sm for the suggestion though I'll give that a try!

[dusky lotus]

Hey everyone! Been a while since I've been here but I could use some inputs! Any ideas on how I could prompt hairs like that ? I can easily have the colors but it's only streaked, can't have them on the bottom part of the hair. (blurred the face to avoid any flags ^^)

[tired vigil]

Are you using an anime model or realistic ?

Also its pretty hard in general to specify colors and specially if you want them a certain way

[dusky lotus]

tried with both really, natural prompts and keywords based

[tired vigil]

@dusky lotus if you don't really mind the effort, you can use regional prompting for that, just specify the area you want the blonde hair and below the other colors

Otherwise its pretty much luck to get it right through prompt

[dusky lotus]

i really don't mind at all, and i've installed Regional Prompter 2 hours ago, but it's kinda hard to figure out how to split the image

didn't think I could use it for this though ^^

can I use the regional prompt of TiledDiffusion without using TiledDiffusion ?

[tired vigil]

yeah there is a checkbox for the region part

I like regional prompter more

It doesn't make the image weird like tiled diffusion

like they are generated separately

[dusky lotus]

let's try with that then ! just need to figure out the split

[tired vigil]

(gradient hair:1.1), blonde hair, blue hair end

haha

Pretty all over the place

[dusky lotus]

gradient and hair end... tried with tips, tones, multicolored, and pretty much anything but that !

[stable tiger]

hi guyz, i am trying to make a meeting photo shot around a table with some people but mid-close shots seems faces are bad how can i fix this
close-up shot photo of midage man sitting around table with 6 other people inside a futuristic retrotech concrete rusty old loking shelter one big white square ligth on ceiling , instruments and a tv screen shoving desolete ground on the wall ,cinematic, photo realism, DSLR, 8k hdr, cinematic light, film still <lora:JuggerCineXL2:1.0>
negatives peopleneg poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, body out of frame, blurry,, blurred, watermark, grainy, signature, cut off, draft, text, logo, armor, plastic, , twin, double, copy , deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime, text, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck
sdxl 1.0 base10vaefix checkpoint and refiner i used and juggernaut lora face restore codeformer:0 (max) Dpm 2s a Karras method seed:3195027069 i also tried juggernaut5 checkpoint for same prompt but results were similar how can i fix that faces ?

[tired vigil]

No gradient really

[dusky lotus]

yeah but the principle is here ^^

[tired vigil]

Gradient is the anime tag :p

[tired vigil]

Is adetailer a thing in SDXL ?

It basically inpaints each face after the generation automatically, those far away faces get more detailed

[stable tiger]

@tired vigil an extension ? ok let me try

[dusky lotus]

@tired vigil thanks for the regional prompter tip, that's awesome x)

[tired vigil]

hey looks like its working
I have no clue how to use that extension also my pc is really bad things don't just work

[dusky lotus]

most of the options i don't get, just used the tutorial on the git page, and tried to split the image by typing random values x)

[blissful lake]

I'm trying to make some images of Park Jihyo from kpop girl group Twice. I downloaded the Koreandoll lora and started generating. I can clearly see SD is adding Jihyo's facial features, but the output just looks too artificial and not enough like her.

Would it be better to not use a lora at all?

[twin yoke]

The LoRA is probably going to undo a lot of the prompt work you might be putting in by coming over the top too strong. You could attempt to lower that strength, but you might just be trying to have it render someone that it doesn't know well enough to get right. A lot of times, if you want a specific person that the model doesn't really know, you would want to train up a LoRA with images of that person so that you could apply that to the model you're working with and get much better outcomes.

[blissful lake]

SD must know who she is, because I can clearly see her facial features starting to come through.

A lora of Park Jihyo would be great, but how do I go about building one?

Could I perhaps type "park jihyo" into google images and download the first 10,000 results, would that work?

[twin yoke]

It probably does know at least a little, but not likely enough to infer a perfect match. (Depends on the model that you're using, too.) There is a ton of information on the net about how to train a LoRA up, but it does require some juice. I don't know what you're running or what model you're working with, but that may be difficult on your own hardware; just kinda depends.

You should also do the reading/viewing on what images to use. The wrong set of images will give you poor results.

[blissful lake]

Alright, sounds like I have a big project on my hands. I just installed SD on my 1080Ti last night.

I had some good results from the Koreandoll lora to produce images of beautiful korean idols, but I wasn't targetting anyone specific.

[twin yoke]

1080 Ti was a nice card back in the day, but you're going to have a lot of time invested to train a LoRA on that. You could look into using a farm and renting GPU for pretty cheap and much faster. Depends on how much it's worth to you. You might get away with a LoRA for SD 1.5 models, but you'll hate life trying to train one for SDXL on that card.

[blissful lake]

I have a 1080Ti and a cold flat. I could either turn on my space heater and waste the electricity to make heat, or I can start doing this Stable Diffusion thing and maybe get some nice pictures.

I can't make my own loras on a 1080Ti, that sucks. 😦

I used to have lots of GPUs when I was mining, I shouldn't have sold them all

[twin yoke]

You might be able to with SD 1.5 models like I said...just will be a bear with SDXL.

[blissful lake]

I guess I have all winter to do it. Might as well do something productive with my GPU instead of using the same electricity on a space heater which produces nothing but heat. As long as I can pause and resume it so I can game on my GPU or pause during hot days, it should be fine.

I wonder what my wife will think of of the pics.

Last winter I was mining ethereum, but can't do that any more. 😦

[hot wolf]

I'm using that on my second computer and I had absolutely zero issues with speed or quality whatsoever. It's a little slower than my 4090 but it's not a huge time difference any way. I think the most important things are sys ram, ssd and cpu.

[blissful lake]

Does this look "photorealistic" to you, SD?!?!?

[finite pendant]

How do I stop the AI from putting b4re br34sts? I already put "cl34vage" into Neg, but it still puts a massive cl34vage on the pic.

[finite pendant]

Found out why it happened - The "gur0" tag. However, not it refuses to add bl00d or w0unds on the generated image, so Idk what to do.

[tired vigil]

What model are you using

[tired vigil]

Did you see people making realistic stuff on the model page with it

it has a game name :p

also photorealistic isn't actually realistic, try just "photo, RAW photo, photograph" one of those

realistic models will have realistic looks with little prompting for it

also those asian girls kinda look like dolls, its trying I guess

[blissful lake]

ffs

[tired vigil]

we need the flamethrower

[blissful lake]

It's so uncanny

[tired vigil]

lmao

Try a normal image

like just a random girl

none of that lora/idol girl

see if the model can do it well

[blissful lake]

I didn't have a lora selected

It's still weird

[tired vigil]

try another model

like realistic vision or epicrealism

(its the only ones I have for realism, they do well)

[blissful lake]

Can you give me a torrent file for it?

[tired vigil]

t-torrent ? huh

Don't you get them from civitai

[blissful lake]

My internet connection is getting hammered, and torrents are pretty good at long duration downloads

[blissful lake]

Which of these two do you recommend?

I only want asian girls to be generated, no other races.

Korean, specifically actually.

[tired vigil]

hmmm

Then try chilloutmix ?

search that on civitai, it's very nsfw but its the asian model from what I know

[blissful lake]

This is the one? https://civitai.com/models/6424/chilloutmix

[tired vigil]

yes

[blissful lake]

But the images don't have to be lewd, right?

I just need the face to stop being so artificial

Also, I found with some of these NSFW models, the body proportions are unrealistic

[tired vigil]

it does anything yeah, just prompt for what you want

[silver valley]

Check the model named Beautiful realistic Asians or Bra

[blissful lake]

That looks really good, thanks.

I will download chilloutmix and bra

Can I switch between models easily?

[tired vigil]

well, it just takes 5 seconds for me

I don't know about how your UI works

[blissful lake]

I'm starting to think that I should put a GPU inside my NAS so I don't have to stop gaming to mess around with this

Would an RTX 4060 be okay?

Or is it underpowered? I'm using a 1080Ti in my gaming PC ATM

Does VRAM matter?

This one has 16GB https://www.scan.co.uk/products/msi-nvidia-geforce-rtx-4060-ti-ventus-2x-black-oc-16gb-gddr6-ray-tracing-graphics-card-4352-core-262

[tired vigil]

VRAM is basically the thing that runs stable diffusion

[silver valley]

Vram is the important part here

[tired vigil]

and rtx are way better than 1080

[blissful lake]

So 16GB 4060Ti will be way better than 4060?

Is there a Stable Diffusion GPU comparison chart somewhere?

I'm looking at Tesla P4 too

It's so cheap, but is it powerful enough?

[tired vigil]

I dunno about those cards

[silver valley]

There is this old chart from last year

[blissful lake]

I wish the 1080Ti and 4060 was on there

[silver valley]

My 1080 is 1.9

[tired vigil]

someone sent this days ago

[blissful lake]

How does Tesla P4 compare?

Found this! https://www.reddit.com/r/StableDiffusion/comments/13hyn0c/selfreported_gpus_and_iterationssecond_based_on/

So Tesla P4 is poopoo IDK why someone in the Unraid Discord kept shilling it

[proven jasper]

because its cheap

[blissful lake]

It is yeah...

RTX 3050 looks like a good option

Faster than my 1080Ti

[proven jasper]

whats your budget

[blissful lake]

I can probably spend up to $400, but the bigger concerns are idle power consumption and space. I'm planning to run this in a Node 304

[tired vigil]

@blissful lake the 4060 16gb looks really good, is it cheap for you

with a lota of vram you can do multiple images at once

[blissful lake]

Twice the price of 3050, and might not fit in my case....

[tired vigil]

[blissful lake]

I'm not sure if I need that

I want to run LLM too

Do we know if these GPUs are good for LLM as well?

[proven jasper]

yes they are u could probably buy an used 3080 12gb but idk if u want an used card

[blissful lake]

I think I could claim this as a business expense on my IRS tax report

[silver valley]

Most LLMs need more then 8gb vram

12-16 would be good

But they work on CPU too

[blissful lake]

I tried, it's painfully slow

[proven jasper]

u could also buy an used quadro rtx 5000 16gb for like 400$ idk if u want used tho

[blissful lake]

Oh wow LLM hardware requirements are insane

People are saying I need a rack full of RTX 4090s wtf

[tired vigil]

https://tenor.com/view/boburnham-inside-netflix-internet-computer-gif-21887509

people and exaggerating, fits the internet

[vast bobcat]

depends which model

I did need 64 gb of normal ram to convert one LLM

Well 48gb I guess. Just went all the way.

went to terminal only and then did the conversion

but actually once it was in the GPU it was very small due to optimizations

[blissful lake]

Not terrible

Looks more like Momo but okay

[blissful lake]

Alright, I'm making some pretty girls now

I keep getting lewd pics though, how do I stop that?

Is there a prompt I can use to remove the nudity?

[proven jasper]

(naked,nsfw:1.5)

[blissful lake]

As a negative, or a positive?

Why does she look nothing like Jihyo

[proven jasper]

negative

[blissful lake]

didn't work...

I see a full rack of boobies in front of me right now

[proven jasper]

depends on the model

[blissful lake]

Is this not right?

(crowd)+, (worst quality, low quality)2, monochrome, zombie,overexposure, watermark,text,bad anatomy,bad hand,extra hands,extra fingers,too many fingers,fused fingers,bad arm,distorted arm,extra arms,fused arms,extra legs,missing leg,disembodied leg,extra nipples, detached arm, liquid hand,inverted hand,disembodied limb, oversized head,extra body,extra navel,easynegative,(hair between eyes),sketch, duplicate, ugly, huge eyes, text, logo, worst face, (bad and mutated hands)1.3, (blurry)2.0, horror, geometry, bad_prompt, (bad hands)+, (missing fingers)+, multiple limbs, bad anatomy, (interlocked fingers)1.2, Ugly Fingers, (extra digit and hands and fingers and legs and arms)1.4, (deformed fingers)1.2, (long fingers)1.2,(bad-artist-anime), bad-artist, bad hand, extra legs ,(ng_deepnegative_v1_75t),(naked,nsfw:1.5)

[proven jasper]

your negative is way too saturated

[blissful lake]

Yeah but it's producing good results

Haven't seen a single bad image

Too much nudity though

[proven jasper]

no the more saturated a negative is the less effective it is

try with this one (worst quality, low quality), monochrome, bad-artist, bad hand, extra legs ,(ng_deepnegative_v1_75t),(naked,nsfw:1.5), if it doesnt work try with another model,some models can only produce nsfw images

[blissful lake]

I've had many SFW images with this one

[blissful lake]

ffs no not working

Give me one just to ban the boobies

[proven jasper]

change model

[blissful lake]

Fucking chilloutmix is obsessed with boobies

[dusky lotus]

you should look into your embeddings to see what keywords are in them, and i'm not familiar with your syntax, but if you have to put the weight AFTER the parenthesis, do the same thing for (naked, nsfw)

[blissful lake]

I'm going back to bra model

[proven jasper]

try with this one https://civitai.com/models/81458/absolutereality

[dusky lotus]

this one is great too: https://civitai.com/models/25694/epicrealism

[blissful lake]

I'm using this one now https://civitai.com/models/25494/beautiful-realistic-asians

Hmmmmm now every image is a close-up and no body

I want body, but still want the ultra detailed face.

[tired vigil]

haha it is a very model

also can you install extensions in that ui ?

If you want more detailed faces even far away ones, try the adetailer extension

[steep canopy]

Hi, have anyone experienced getting malformed eyes when using face restorations?
This usually happens when I used restore faces while upscaling. Both GFPGAN and CodeFormer gave same results.
Hope someone knows what causes this.

[tired vigil]

the solution: don't use it

Really

close up gens like that you shouldn't need to restore faces or anything like it

If you're doing some with far away ones, try the adetailer extension, it automatically inpaints faces

[silver valley]

Hey, thats normal as you shouldn't use them together. Either upscale the image first and then fix the face in extras tab or better by inpainting.
Or generate without hires fix with active face restore. And then upscale

But like Cat 4gb said don't use face restore because highres fix is enough or inpainting or Adetailer extension

[steep canopy]

Alright got it. Will try Adetailer instead of face restore. Thanks guys!

[finite pendant]

Is it possible to make two people and describe each separately on SD 1.5? Like 1girl: blonde hair, blue eyes | 1girl: black hair, green eyes?

[tired vigil]

Regional prompter extension can separate prompts for each region you tell it

[finite pendant]

:O Gonna look that up!

[tired vigil]

Normally, through text to image only, it's pretty hard to do

[finite pendant]

I assume I have to use a color-coded image?

Ok so how do I now put in two prompts?

[tired vigil]

check the guide

There should be some examples to follow there in the github

Makes it simpler to understand

[finite pendant]

Ah okay thaks! Found it!

[eager widget]

hi can one tell me best prompt for making fake nude images

[blissful lake]

I think there's an 'Unstable Diffusion' server dedicated to this

You'd get the best support there

[eager widget]

can u tell /

[jovial phoenix]

does SD support negative prompts? my tank girls are all M1 Abrams even when i specify.

trying to make anime waifus in a Girls Frontline style but as tanks instead of guns

also any advice for faces? my faces are always smudged and the eyes are dead af

[blissful lake]

Yeah you can use negative prompts.

[blissful lake]

I've found good luck using koreandoll lora and chilloutmix and increasing the number of samples or whatever its called. I went from smudged faces to beautiful photorealistic girls

Also prompts like "ultra detailed face" and stuff like that

[vagrant sky]

Hi, is it possible to generate specific art style like madhubani and murals

[tired vigil]

https://civitai.com/models/53130/madhubani-style-lora-ancient-indian-art-style

maybe this helps

[frozen meteor]

I wish to make a picture of someone lying on their back, completely submerged in a shallow water
all the images i'm making have at least their head out of the water though. I don't want the subject to have anything outside the water. any thoughts?

using SDXL
a naked pale-blue-skinned woman with white hair, lying on her back fully submerged underwater in a ritual font of ice water, inside a wooden viking temple,
realistic detailed soft digital render fantasy

as you can see she's not quite lying down, most of the results are like this, i want her face completely underwater, like she's drowning

[pulsar oyster]

If I'm trying to use the controlnet:scribble to create a man holding a fist. Is there a prompt I can use to help generate a fist with inpainting?

[opaque musk]

https://civitai.com/models/67174/900-hands-library-for-depth-library-or-controlnet

[pulsar oyster]

I'm a little confused. Am I suppose to use this with depthmap in controlnet? Can I use this model for inpainting the correct hand position?

[tired vigil]

Anyone know what model this might be?

[tired vigil]

greek art style, drawn, oil paint, of a broken statue, cracks in statue, fragments, right hand in the air, lower body inside lake, upper body over water, underwater visible, black hands pulling lower body of statue, lots of black hands, detailed underwater, sea life, fish underwater, darkness underwater, detailed sky, cloudy grey sky, realistic,

could someone help me wit prompts im new

what im imaging is basically a statue that is in water. the upper body is above the water but the lower body is underwater and inside the waters its being pulled by hands

and the statue wants to be saved so it has its right hand in the air

this is what i generated

the first one is close but i would want the statue to be further away and smaller

and ofc the hands that are pulling it

help

(sdxl)

[burnt wedge]

anyone have any good resources to learn what this does and how I can use it to improve the faces I generate? Very lost rn and many of the videos or links I find are using some different version of SD. Any bit of help will do (If this is the incorrect channel please correct me since this is my first message)

[tired vigil]

cat friend

Try using the adetailer extension

it automatically inpaints faces that it detects in your generations

[burnt wedge]

I don't know what that is nor how to add it unfortunately LMAO

It's my second day messing with SD so far

[tired vigil]

Hmmmmmmmmm let me load up SD for some quick screenshots

[tired vigil]

try leaving hires steps at 10

if you leave at 0 it will copy your txt2img settings, and it is probably 20, thats too much for general use on hires

also 0.5 denoise should be good

[burnt wedge]

my sampling steps were at around 27 and once at 40, still trying to see how that messes with generation quality

i also never had hires steps more than 0 so that might've been causing some problems ??

let me load it up rq and show you my usual settings

[tired vigil]

At 0 it will copy your other steps settings

its most likely too much for hires

Extensions > Install from URL > and put that link there and install

https://github.com/Bing-su/adetailer

[burnt wedge]

this is what I usually use/do

[tired vigil]

might need to go in installed and "apply and restart"

[burnt wedge]

I encountered a problem not long ago where using hires gave me some sort of like red and blue 3d glasses type of effect and didnt really change face quality or anything by much or anything at all

would that have been caused by hires steps being at 0?

[tired vigil]

hires is basically just using your current image and upscaling it

too little denoise and it might still look blurry from the lower resolution original image

and too much it can cause weird stuff (not those type of 3d glasses) :p

idk what could be causing it, too many things c_c

[tired vigil]

and at 0 it will copy the other steps settings, its not actually at 0

@burnt wedge what is your GPU

[burnt wedge]

6800XT

[burnt wedge]

it happened even at relatively low denoise which is weird

like 0.2

will retest it with hires steps at like 10 and maybe adetailer after if there's no difference

[tired vigil]

Did you set commandline arguments in the webui batch file ?

[burnt wedge]

yeah but I don't really know if they're optimal or any good yet

[tired vigil]

What do you have in there ?

[burnt wedge]

--upcast-sampling --medvram --opt-sub-quad-attention --opt-split-attention-v1 i think

maybe all that w/o the last opt

[tired vigil]

Those look fine. I saw cs1o (one of the experts around here :p) say that he uses those but instead of upcast sampling he uses --no-half-vae on a 6700xt
You could try later to see if one of those makes gens faster but shouldn't be a big worry

[tired vigil]

try to stick to 0.45 - 0.6 maybe even up to 0.7 but depends on what you're doing

and what model are you using ? I've seen those artifacts on faces before being caused by models

[burnt wedge]

For the one that got the 3d glasses type artifact was the fanitex model

though that was after hires

[tired vigil]

I can't find it on google, got a link ?

[burnt wedge]

no link but ik i got it from civitai site

fantexi I think it's called

when I use hires should I change anything in the prompt or does that change the whole image generation??

[tired vigil]

hires only takes your generated image and upscales it

its like a second generated image but using the old one as base

thats why too much denoise can lose the original look (which should be coherent) and cause weird stuff to appear

and too little denoise it can still look blurry like low resolution

0.5 denoise should work for all images

[tired vigil]

found it, many booba
Anyway, I don't know about that model but if it keeps causing issues, look for another one like dreamshaper, its a very good model and it does many art styles

[burnt wedge]

I havent tried dreamshaper yet

i think only ghostmix, one kinda realistic one but I forgot it's name and fantexi

will give it ago later, maybe tomorrow actually

just need to see how I can fix the 3d glasses artifact if possible

gonna try generating with this and trying hires with it after, hopefully no problems

[dusky lotus]

why 27 steps?

[burnt wedge]

ngl I'm not sure either

im messing with anything between 20-30

[dusky lotus]

have you tried running an xyz plot with different steps to see how much it'll change your generation?

[burnt wedge]

not sure what that is yet so I unfortunately have not done that

[tired vigil]

Good afternoon, has anyone taught SD how to make Sprites of bodies or objects? How to cut an image into layers on the front layer, back layer, body parts, clothes?

[inner acorn]

I'm attempting to make rocket boots right now - There's a surprising lack of rocket boot art online - But the AI can't seem to get flames jetting out from beneath the boots. Here are my current settings and prompts, as well as the lora and model I'm using:
Positive Prompt:

dnd magic item, pair of boots with flames coming out from beneath, leather boots, burnt leather, raggedy, magic boots, thrusters attached to the boots:1.3, flames shooting out from beneath the boots, flames shooting out from the thrusters, flames propelling the boots, blast off
Negative Prompt:
legs, human
Model - https://civitai.com/models/4713/fantassified-icons

Outputs have been more along the line of cowboy boots with flame decals, which isn't really what I'm going for. I'm more leaning towards magical jet-boots of sorts

[polar glade]

Hey, I'm having trouble with what is essentially a very basic prompt

full body portrait
absolutely 0 images I generate show the full body

these are at best half-body

[tired vigil]

Try describing the feet

Or boots whatever

Generally full body works for me

Exclude the portrait since that is half body

"a girl wearing black boots and blonde hair, full body photo"

Or if you're using an anime model "1girl, full body, blonde hair"

[burnt wedge]

y'all know how I can improve face quality? these are what I use and the face quality is not very good even after img2img/inpaint. Anything I'm doing wrong?

[silver valley]

Hey, do you have an example of a bad face?

Upscaling should fix it

And inpainting too

[burnt wedge]

will try scaling this with img2img maybe, not sure. these were the settings. will send how it looks after img2img

unsure if it's due to parameters or maybe certain models aren't fit for certain prompts

ghostmix model was used for this image ^

one of the images from img2img @silver valley not sure if I can make it any better in inpaint or anything else

[silver valley]

That looks okay. But for the first picture you could have used highres fix.

Be aware that the smaller the face the less pixel it gets to be good quality

For example try 512x768 and then use a tag like
Photo, portrait
You will get better faces

For full body stuff you always need highres fix or inpaint

[tired vigil]

@burnt wedge did you try the adetailer extension ? it should fix those faces automatically, its really just a case of them being super low resolution for small details
Doing 768x768 images will help too if you want to keep the square aspect ratio
Did a 2024x2024 img2img :p

[burnt wedge]

tried it but there's a lot of settings in it so idk how to really optimize it for what I want yet. Need to look up couple videos

[tired vigil]

The default settings should cover what you need, so you just enable it

if it isn't detecting faces as much (like you have multiple people in the generated image) you can up the detection value (don't really remember what is called)

[burnt wedge]

even with hires it's unfortunately not very good. maybe 512x512 might be the problem but it's usually fine on some other prompts

after hires i still need to go through the img2img process + inpaint and after I'm done it doesn't really change much for some reason

inpaint seems to do literally nothing for some reason which is a shame

[silver valley]

What are your inpaint settings?

[burnt wedge]

scale being 0.5 to upscale it later by 4 in extras. Not sure why that's done but that's what my friend recommended.

only masked instead of whole picture tho

[silver valley]

Yea only masked and don't use resize by

[burnt wedge]

no resizing at all?

[silver valley]

Yes

You only want to inpaint

Inpaint is like generating an image into an image

[burnt wedge]

inpainting sometimes changes my image and sometimes it doesn't do anything at all

it's weird

[stiff bison]

Need help with a promt for a person making a selfie. I want it to look like the person is holding the camera 😄
When I promt "selfie of a girl", I get to see the girl holding a phone 😄
Any suggestions?

[oak flax]

like this?

[stiff bison]

exactly

[oak flax]

first person perspective woman taking a selfie with her phone

doesnt work 100% of the time

[stiff bison]

thank you! will give it a try 🙂

sadly not successful.. at least with cyberrealistic 😄

[oak flax]

sdxl is way better its like 1 in 5~ chance of getting what you want, sd 1.5 is like 1 in 20

[stiff bison]

any recommendation for a realistic checkpoint? 😄

[oak flax]

i like this one https://civitai.com/models/139562?modelVersionId=169921

yeah sdxl is way more consistent

[smoky gull]

Is this a prompting issue?

wth!!

[quick thicket]

When using lora's in webgui, where should the activation words associated with a lora go?

[silver valley]

Does it work without controlnet ?

[silver valley]

In the prompt field

[smoky gull]

Yeah it does

[silver valley]

Okay then try with 512x768 as resolution

[smoky gull]

I'll try that tomorrow, thanks. Is it a bug?

[silver valley]

I can't say for sure

But also test an other model

[orchid ore]

@quick thicket i think in lora tab when you click on lora you want it makes <lora...>

[maiden folio]

Is this considered a good prompt: Bustling Spanish city during sunset, cobblestone streets lined with colorful buildings, adorned with flower-filled balconies, lively plaza with a fountain, surrounded by outdoor cafes, people enjoying traditional Spanish food and drinks, vivid blend of orange and pink hues in the sky, impressionist style painting, reminiscent of Joaquín Sorolla

[golden marten]

Struggling to output sometimes with ComfyUI, just getting a "reconnecting" error as if it has failed

[terse ether]

Hey! I am searching for an expert who can create a 10 seconds long video for me. Of course I will pay for the service

[tired vigil]

Did you close the cmd window

[tired vigil]

looks like a normal prompt for realistic models, don't see anything wrong

try "in the style of Joaquín Sorolla"

instead of reminiscent

but both should work I guess

[orchid ore]

Also would put in it some BREAK keyword as words at the end can probably has very low attention? Dont know.

[golden marten]

No I never closed it, was I supposed to?

[tired vigil]

no, if you close it you'll get connection error... if you didn't, I dunno, not familiar with comfy #🤝｜tech-support is better for errors and stuff like that

[golden marten]

No worries thank you for helping!

[brave fulcrum]

I am not sure this fits here but question, is composable lora broken nowadays? it crashes when i try to use it with loras and latent couple

[compact path]

Hey all! Are there diffences that I should keep in mind when prompting for Midjourney, SD, and DD?

If so, what are they?

[granite holly]

anyone has a workflow to create an AI influencer?

[maiden folio]

a what?

[granite holly]

a female model

[novel bolt]

Hi gang! I'm painfully new to SD and wanted to ask if there's a way to regenerate an existing image with new variations. Thanks

[muted reef]

you can use img2img for that yes

generate with the original seed,prompt and settings and then send to img2img

[brave fulcrum]

Hello there I am trying tiled diffusion, and while its great for having 2 characters from 2 diferent loras, is there any way to make them """interact""""? (between many quotes as I know it doest work like that)

[next hound]

can someone send me his config file of a deforum + controlnet settings please?
i have problem with my generation it doesn't work

[muted reef]

think using controlnet_openpose works well for it

[violet nimbus]

Hello, wanna ask if we can actually generate the same image with same prompts from civitai. I tried it along with the same seed and same cfg etc but it doesn't come close enough, is it the difference in resolution or graphic cards? Or civitai just doesn't list everything used? Or it's just impossible to generate same image or close enough image?

[sterile wind]

What prompt makes my image horizontal

[tired vigil]

Its pretty hard to generate someone's exact image, there are a lot of settings that can be different, including ui, resolution, sd version...

and if they do one thing like controlnet, it will not appear in the png info

[narrow needle]

yea

fake advertising sometimes at civit.ai 😆

[tired vigil]

Hi guys I need help please. I am training my own model and I am not getting the results i expect this is how i do it:
I am using this script:
https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_lora.py

like this:
accelerate launch --mixed_precision="fp16" train_text_to_image_lora.py --pretrained_model_name_or_path=stablediffusionapi/cyberrealistic-v32 --dataset_name=anjakuzev/michael_scott --dataloader_num_workers=8 --center_crop --random_flip --train_batch_size=1 --gradient_accumulation_steps=4 --max_train_steps=800 --learning_rate=1e-04 --max_grad_norm=1 --lr_scheduler="cosine" --lr_warmup_steps=0 --output_dir=diffusers/output --checkpointing_steps=3000 --seed=1337 --validation_prompt="Generate an image of Michael Scott with short dark hair and deep-set eyes. He should be wearing professional attire, including a dark suit jacket and tie. His expression should range from jovial and smiling to deep in thought or surprised. The setting should hint at a workplace environment, preferably an office." --image_column=image --caption_column=additional_feature --push_to_hub --hub_model_id=anjakuzev/name

what am I doing wrong?

my dataset:
https://huggingface.co/datasets/anjakuzev/michael_scott

thank you in advance!

[viral wolf]

What does putting prompts in (parenthesis) do? I see that a lot in example prompts

[orchid ore]

((guitar)) put more weight on word guitar. Also can be (guitar:1.4) for example. Better to use second form imo.

[snow musk]

how can i prompt better eyes?

[tired vigil]

Have you tried adetailer

Higher Resolution helps too

[snow musk]

i havent im quite new to the whole thing, where and how can i add it? and with higher resolution, do you mean just increase the height/length scale?

[orchid ore]

depends on model. if 1,5 768x512 and if sdxl 1024x1024
for 1,5 lower than 512x512 got not sense. If sdxl try keep resolution on about 1MPx

[tired vigil]

I think a quick Google will get you to the adetailer github

There they have how to install it

[snow musk]

ah gotcha

[tired vigil]

You just simply click on enable in it and it works with default values

It will help with faces specially far away ones

Like bernix said up the resolution to 768 on 1.5 and use hires if possible
That should get you good eyes

[snow musk]

i'll give it a shot, thanks both of you!

[brave fulcrum]

Hello there I am trying tiled diffusion, and while its great for having 2 characters from 2 diferent loras, is there any way to make them """interact""""? (between many quotes as I know it doest work like that) Someone told me to use open pose but that hasnt worked

[sharp olive]

A friend of mine asked to be zombiefied.
(I started with img2img, but in my opinion, the results are fairly low quality)
I'm trying to load his image using ControlNet as a reference only image, but I don't quite understand how to control the strength.
I've seen it essentially copy an image with slightly changed angles and poses.... I can't seem to do that here.

For now I'm using xxmix9realistic with the settings from the first screenshot
The prompt is "male zombie"
and control net is set up for the second screenshot.
The images look nothing like the input image (which is a stand-in for my friend, while I try to get this to work and share public screenshot).

[dapper tree]

You looking for something like this? https://youtu.be/9UOdXLYwhYY?si=aNS-NN6C7w5zC8fW

[sharp olive]

I don't want to inpaint. I'm looking for something like this
https://github.com/Mikubill/sd-webui-controlnet/discussions/1236

@dapper tree

[dapper tree]

Never used that before.

[terse ether]

I am searching for a long term partner who can make a video like this for me. Of course I will pay for your service. If you are interested just text me!

[tired vigil]

i wish i had a lava lamp gf

[orchid ore]

@sharp olive try put his image in img2img and prompt Turn him into zombie.
Play with models and denoise strength. Lower value, smaller change, higher higher change.

[sharp olive]

That I was able to do, but at the moment where I get variation in the poses, he becomes unrecognizable. It seems that with controlnet I might have better control over this... except I don't, because at best I can do open pose

[orchid ore]

just lower denoise strength somewhere about 0,2-0,4. Default 0,75 is too heavy for this sort of things.
It is feature which used to be as well extra, but probably so good, so it is now implemented in A1111.

wait moment

[dense stream]

hello guys sorry for the silly question but how I can create image based on another image and prompt?

[sharp olive]

for img2img, yes. But again, either it looks like a cheap zombification filter or it changes the image too much, there is no right balance for denoise strength here, as the face information is lost before the pose information.

I'm hoping ControlNet can do a bit more, as shown in what I linked.

[sharp olive]

can you go a bit more into details on what you want to do?

I think you want img2img or controlnet as well Aisern

[dense stream]

I have a photo of mine and I just want some magic to create a beatiful picture with my face

but how can I use img2img in the Discord? I see only dream fuction here

[sharp olive]

I could probably try to help if you were running it locally. I don't know how or if it works through discord

[dense stream]

yeah but I don't have enough GPU to run it 🙂 but I can use midjourney then? It seems it's avaiable only, no? I'm looking for an advice 🙂

[orchid ore]

@sharp olive if you need to use controlnet, try instructp2p
here is some person turned in zombie by it. I will censor it
Strength only 0,35 to more it making more changes. But is it still easy to know who it is.

[sharp olive]

I don't "need" to use it, I just think that it's most suited to what I need - but my knowledge is limited.
What you posted looks promising. I'll try it later today.

Thanks

[orchid ore]

no problem hope it will helps you!

[dense stream]

Hello! Can someone kindly ask with creating image2image by running locally Stable Diffusion? I don't have enough GPU to do it and Discord bots don't have this capability 😦

[bright pagoda]

hello, how can I make someone point with his finger like he is showing something? I couldn't create such usable image yet

[dapper tree]

What GPU do you have?

[sinful yew]

I've looked at several pre written prompts and ive even got a model that was popular but no matter what i do they all come out looking terrible is there anything i can do?

[tired vigil]

show your settings

[tired vigil]

yeah

you're on sdxl

You need at least 1mp resolution

like 1024x1024

[sinful yew]

ok i will try that, is there anything else i need to change?

[tired vigil]

can you use hires on sdxl ?

50 steps seems way too much, specially for hires, you can do hires 10-20 steps

[sinful yew]

idk what that is sry, im new to all of this

[tired vigil]

What gpu do you have @sinful yew

do you know that you're using a SDXL model even

[sinful yew]

yes i know that one

[tired vigil]

okay 2060 doesn't seem like it will be good enough for hires on sdxl, close that tab :p

hires upscales images further

bump up the resolution to 1024x1024 and try to gen something

I don't think you need negatives on sdxl, unless theres some good embeddings for it now, I heard about them but not sure if they are effective

[sinful yew]

ok thanks i will try to gen something, it takes like 5 min

[tired vigil]

are you sure you want to use sdxl lmao

its much slower than 1.5

[sinful yew]

well tbh i though sdxl was the newest thing, what should i use instead

[tired vigil]

You can just use 1.5 models

sdxl is really cool and can understand some prompts much better but its super demanding on gpu

sd1.5 can still look just as good but not as smart I'd say

@sinful yew do you want to generate realistic images or anime, just get a 1.5 model from civitai and slap in your folder, should be good to go

[sinful yew]

ok so look for a 1.5 and put it in the same place as the xl one?

[tired vigil]

yeah... maybe.. I never used sdxl on my pc, if its the same as this webui\stable-diffusion-webui-directml\models\Stable-diffusion

then good

[sinful yew]

yeah its there but with out the "-directml" part

[tired vigil]

I think thats because I installed the AMD version

ignore it :p

[sinful yew]

i have amd not intel, did i install the wrong thing lol

[tired vigil]

@sinful yew no, my GPU is AMD

not CPU

[sinful yew]

ohhhhhhhhhh

sry lol

[tired vigil]

for 1.5 I'd recommend you start with something like this

don't forget a vae, it makes colors better

[sinful yew]

Vae?

[tired vigil]

euler a, restart and dpm ++ 2M Karras are good samplers you play around with them see what you like the most

[tired vigil]

some models don't have vae baked in, you'll need to select one

[sinful yew]

what is a vae tho and how do i know if its my model has it or not, sry for all the questions you've been a massive help thank you

[tired vigil]

you put them in your models\VAE
Download this one and leave it there if you need, its pretty much the global anime vae xd https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.ckpt

[tired vigil]

usually the author tells you if the model you downloaded has or not

without a vae your images will come out kinda of grey

and desaturated

[sinful yew]

ahhhh ok thanks

[abstract gust]

hmm... I keep seeing pictures of Link and Zelda in hylian-themed space suits on DeviantArt, made with the Bing Image Generator... Would something like that be possible with SD as well? If yes, what Lora's do I need and what could an example prompt look like?

[crimson patio]

Best lora/embe for better hands for now? (sd 1.5)

[gritty pollen]

any idea how this was made? curious about the prompt and model used here

[polar glade]

Hey I'm noticing that SDXL doesn't work as well when specifying ages.
For example images for "25,35,45,55, or 65 years old" all produce nearly an identical image. Upon hitting "75 years old" then I notice suddenly the character has gray hair and is old.

I'm trying different ways to blend age words instead. For example
(toddler:0.0)(child:1.0)
(toddler: child: 0.5)
(toddler | child)

I'm running experiments... but I'm wondering if anyone else has guidance on how to blend two age words in some kind of smooth progression

failed experiment showing different weights for the words "Toddler" and "child" which seem to demonstrate (toddler:0.0)(child:1.0) this kind of blending doesn't work

I thought maybe I'd see a transition along the diagonal, but (toddler:0.0)(child:1.0) looks almost identical to (toddler:1.0)(child:0.0)

[tired vigil]

new to stable diffusion. not sure whats causing the faces to blur?

[teal nacelle]

anyone know how to promp image to image to get rid of an extra leg

[tired vigil]

go into negative prompts

and try "extra limbs, extra leg, third leg"

[hazy trout]

cause: it sucks (idk)
solution: use a vae. idk what it does to stable diffusion i just know it helps with faces, hands, quality, and sharpness

[tired vigil]

gotcha. thanks for the feedback. ill google vae

ah looks like it only enhances colors but still good

[hazy trout]

https://stable-diffusion-art.com/how-to-use-vae/ heres a good article

[teal nacelle]

slowllly getting it out and getting other stuff, trying to find a good way to match the background it had before

[tired vigil]

what's your positive and negative prompts?

i can try running it in mine and mess around with it

you can post it here or dm me

your choice

[half nymph]

Hello all. I want prompting guides on fooocus and which model to use? I am using realistic.bat which uses realisticStockPhoto. I also have RealVisXLV2.0 on Fooocus. Also do I need to use refiner and LoRA? I want to generate images which produce highly detailed facial features and body features. Thanks in advance

RealVisXLV2.0 is the SDXL version of Realistic Vision

[tired vigil]

Downloaded Ghibli Diffusion but all my pictures are coming out like this:

Any thoughts?

[tired vigil]

better, but still kinda messed up looking

[tired vigil]

can you use hires fix ?

[neon solar]

Hello ! I'm pretty new to SD and i have trouble getting this kind of view ( picture 1 ). I tried : isometric view / from above / etc but i get this kind of results ( picture 2 ). Is there a prompt that can help me ?

I don't care about the style etc. All i want for now is the top view

[polar glade]

make bigger images. It needs to be around 1024x1024

[polar glade]

Try interrogating the CLIP

to get what words it thinks match the first image, then try those words

[verbal olive]

anyone know what the name of this type of clothing is to prompt for?

i've tried reverse image searching it already

i want to create something with that sort of dress, the shoulderless lacy thing

[polar glade]

Why are these two images different
Crystal containing elemental (fire:1.0|ice:1.0)
Crystal containing elemental [fire:1.0|ice:1.0]

I thought () and [] get overridden if you put a specific weight number

[verbal olive]

different seed?

[polar glade]

Same seed

[burnt wedge]

how is this done? that seems really helpful

[polar glade]

There are guides online, try googling and come back

It's in the Img2img tab

[burnt wedge]

alr

[half nymph]

Hello all, anyone work in realism stuff? I want to ask about few things. Thanks

Anyone?

[polar glade]

https://dontasktoask.com/

[half nymph]

Okay. Are there any great models and LoRAS for both SDXL and 1.5? (LoRAS which details facial features and body features if available)

[kindred garnet]

i really like the result, but is there any way to make the eyes dont pop out soo much

i think they are always extremly sharp and kinda too much in the focus

[viral wolf]

Is there a way to do two characters from different Loras? Or rather two characters in one image while maintaining the authenticity of their respective source material? (Example: Spongebob in the same image at Batman on a rooftop)

[tired vigil]

Regional prompter extension

[minor crypt]

A new Chinese tea drink store called "Chajue". The store decoration style is simple and elegant, with a Japanese style.

[old tiger]

whats up G's, does someone know how to generate good tattoo designs in comfy? it does not really work out for me aswell as I imagine. my promt is this rn: Tattoo for right arm, all around the arm at a height of 5cms, a graveyard valley with one side with a RIP Stone, another side a man walking, fossils (skeleton) under the graveyard valley and also have vulture on top of the man

I do get alot of good stuff but not what I was searchinf for, here is an example. Not exactly what I wanted.

[snow musk]

what prompt can i use to have the entire upper body and not using " full body"?, my character has hair vents but it usually never shows them unless i use "full body"

[short laurel]

Does anyone else prompt like this or would anyone know if there are disadvantages to prompting in this way versus using one text box? Just making sure I'm not making a crucial mistake. Creations usually come out as prompted. Secondary prompt is for SDXL.

[tired vigil]

try cowboy shot

[old tiger]

try to use lenses, like 40mm lens, 10mm lens and so on

[snow musk]

testing this one out right now

[snow musk]

lenses? where/how do i use that?

[old tiger]

just write for example: tall man, blonde hair, jojo pose, 40mm lens, strawhat

but you can take whaterver lens you want, smapper lenses mean more zoom in

maller*

Smaller*

[snow musk]

oohh i'll definitely play around with that

[sullen rover]

How can I achieve sharp edges on buildings?

they're always a bit wonky and not straight, sometimes it just goes from sharp and visible and gradually blends in with the background

[teal nacelle]

reccomendation for a prompt to get a sideways picture? Like facing the character sideways

[silver valley]

Profile view or profile of ...

Worked for me

[buoyant agate]

hey guys, so with ipadapter do i not need to write a prompt? or does it still require you to come up with a good prompt

[molten ether]

oes anyone know which models are better at vertical (9x16) versus horizontal (16x9). Some models, when I make the image wide, end up making two heads and has other similar artifiacts.

[orchid ore]

try with sdxl 768x1344 it is close to 9/16 and sdxl model was trained on it

[neon solar]

Thanks for your answer but i already tried it ( even copy past the whole thing but get something different haha )

[silver heath]

I use 3 photos in my IPA. If I set Strength to 35 or below, the more the text input expresses itself - the higher the Strength - the more influence from the images. Text input on my ComfyUI setup seems to be almost negligible after Strength 45+

[polar glade]

You need to use your brain. Those words are a place to start

Try multiple images

Not just one

[buoyant agate]

Yea this is what I was finding, I pretty much just get the same result regardless of prompt w strength over .4-.5 . Thanks for your input! I appreciate it

[neon solar]

Wow really ? not only one image ? no shit sherlock. Just say it if don't know instead of embarassing yourself xD

[inland tusk]

Hello - I've been doing some inpainting to change outfits on some people, but where it has a new exposed area of skin, the tone rarely matches. For example someone in a dark room has long sleeves replaced by short sleeves, but their generated arms are lit like they are in the sun and stand out. Does anyone have any suggestions for prompts, scripts or tools to mitigate this?

[lunar yoke]

What are some good pos/neg prompts for hands?

[short laurel]

In ComfyUi you can use Ultralytics Detect Provider to mask and improve in a second pass, not sure if it is in Auto1111. I also use negative prompts like (Disfigured Arms:1), (Disfigured Hands:1), (Disfigured Fingers:1) etc.. can lower or raise the weight on the negative prompts.

[lunar yoke]

Sorry, I'm using Automatic1111. Thanks for the reply.

[polar glade]

I consider my job to be telling other people how to fix their own problems, not doing the work for them

If it didn't work, show what you tried

Maybe someone else will have a better idea

[tired vigil]

can't really do much, its RNG to get them perfectly xD
try some textual inversion like this one https://civitai.com/models/56519/negativehand-negative-embedding

[compact path]

Are there any modifier, artist, and style studies for SDXL and v2.0 that have images for visual comparison?

[teal nacelle]

would anyone be able to help me upscale my project, my laptop is hitting memory failure when I try to upscale

[kind elbow]

Hello, I am not sure it is the right channel to ask this question.
Basically I want to generate some art to create short stories, but none of the models/styles I have found suit the artstyle I am looking for.
So I would like to know if it is possible to choose an image with a specific drawing style and to have this style recognized by the AI to generate images with a similar drawing style.
I know you can generate similar images / alternate versions based on one image, but what I want is really just generating a style/model from an existing image, is it possible ?

[hard elm]

If you want to do something in a specific style, you might want to look at img2img or training. You can find information about this in #1080946152318443610 as well as searching on YT for training on LoRAs with Stable Diffusion.

You can also put in specific keywords for a given style, like vector art, cartoon, things like "inspired by Lisa Frank," etc.

[kind elbow]

I will do that, thanks for the advices !

[hard elm]

np!

[inland tusk]

If I have an older picture that is maybe a little grainy, is there a way to get SD to make any inpainting match that grain?

I have some pictures in low light that have a little grain to them and want to inpaint an area, but that bit comes back very clean

[rigid wind]

any ideas why this is happening when I try to inpaint my image? Like it's not even close to blending and I've tried about 30 images with varying settings but I keep getting these spots that just don't blend at all, I've tried with high mask blur, low mask blur, high and low padding, whole picture and mask only but for some reason it's just not wanting to match.

[inland tusk]

That's my most common issue with finishing off images

I've taken to recoluring in photoshop or GIMP and then running it back through the model with low deionising

Sometimes I just send it back over to img2img with the original prompt and set a really low denoising strength and run it over a few times

[compact stone]

I could use some help with a really weird problem. I have this one particular face that is persisting across different models and seeds

First one is EpicPhotoGasm, second is AmIReal v4.1

different seeds

photo of adult brunette girl,tight spaceship quarters,pragmatic space uniform,thinking,RAW photo,high quality,35mm photograph,film grain,bokeh,professional,4k,highly detailed,lora:epiCRealismHelper:0.7,lora:more_details:0.5,
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 5, Seed: 2059615960, Size: 512x768, Model hash: 06c19426e3, Model: amIReal_V41, Lora hashes: "epiCRealismHelper: c5e24d9bc4b5, more_details: 3b8aa1d351ef", TI hashes: "girl: 6d9690084e3f", Version: v1.6.0

These are just two examples, but it happens across... half? Most? A third? Of the images I have been generating.

In the preview, I will see a normal image generating, and then this face will come in at the end. I think I have face detailer and ADetailers off

I have restarted Automagic

Sometimes even if I try to make an image of a man, or e.g. a blonde woman, I will still get this face and hair

[brave lion]

How can I get such a color transition in the hair always

[tired vigil]

Lora for that character/hair or regional prompter extension

Maybe look up controlnet reference model

It's very hard through text2image only, you can specify gradient and stuff but hard to work

[short laurel]

I've had decent results just saying two toned hair in the prompt.

Made a few custom CSV list nodes adapting a custom node I found on GITHUB which I believe was repurposed from the WAS Suite Prompt Style Selector.

[short laurel]

https://github.com/theUpsider/ComfyUI-Styles_CSV_Loader files can be found here, I posted my changes in the Issues section as a proposed enhancment. Up to the creator if he wants to include it but you can find the zip there and my changes are listed.

[short laurel]

Ended up uploading the ComfyUI CSV Loader to CIVITAI if anyone is interested in downloading it. https://civitai.com/models/186048?modelVersionId=208913

[peak stone]

hey guys can someone explain in simple terms how i use a lora to alter an image? what directory do I put the lora in?how do I use it in img2img? What are the steps? I want to apply a lora to an existing image via img2img... is that possible?

[inland tusk]

A question. When inpainting, should I include the prompt for the whole image?

For example, if my prompt is "a kitchen table in a country cottage, sunlight" and I want flowers on the table, if I inpaint where I want the flowers, should the prompt be "flowers on a table" or would I get better results with "a kitchen table in a country cottage, sunlight, flowers on a table"?

[muted spruce]

hey trying to do img2img iterations of a contruction site to generate high quality looking final renders but when you use img2img instead of txt2img the architactual lines are all wobbly and unstable, is there a better way to do this?

[inland tusk]

What model are you using?

[wooden ore]

Hi, I am new in the world of stable diffusion and I am looking for advice. I would like from a video recreate the face of the person but on fire. And I can’t do it on deforum. Have you some advice:)

[queen urchin]

hi good day to all, can someone help, about the latent coupl

when i try to make a normal without a lora, everything is fine

but when im adding lora models, this is the result, the image is not clear, in colors ar not good, and more thisng, why there adding offset? lora:XXX_offset:0.9

[boreal pecan]

hey ppl, which is the best negatice prompt to

avoid double images

or double scenes

[hollow tapir]

use a resolution closer to the resolution from the dataset used for training (512x512 for sd1, 768x768 for sd2 and 1024x1024 for sdxl) and upscale/use hires.fix from there

[boreal pecan]

OK OK

Also any advise with

amniotic tanks, xd

im having problem with the liquid inside

for now result is perfect

but I want liquid inside the tank an im having PROBLEMS

I tried to look for loras haha but no luck

[upbeat grove]

Does changing lora in the prompt to a higher value, while keeping the same seed, prompt and other settings do anything? Can seed override it?

[silver valley]

here is one i used
https://civitai.com/models/82528/cultivation-tanktransformation-platform

[wraith tulip]

Does anyone know why this is happening? The face keeps...messing up. I'm not using anything special besides glasses and ears in negative.

[tired vigil]

are you using adetailer* and what sampler are you using

also tried another model ?

I've seen models do that

Meant adetailer not inpaint

[wraith tulip]

I'm not using adetailer and I'm using the... dpmpp-2m-sde sampler with the karras scheduler.

Removing glasses from the negative seems to have fixed it though which is really weird. I'll run another batch and see what happens

[tired vigil]

okay I have no idea what sampler magic is that from comfyui, but I think thats a popular one shouldn't be the issue

[wraith tulip]

Yeah it's comfy ui lol

I'm trying another batch without glasses and seeing if that is the issue.

I mean it kind of is but i'm still getting issues... I'll try a different model

Was using JernauMix, now trying little monsters.

Yeah, it was JernauMix..damn.

[wraith tulip]

...Okay new problem, the images are coming out way oversaturated.

Gonna try increasing the steps, maybe that's it

adding vae as well

that did nothing, shit.

[tired vigil]

are you using loras or something

[wraith tulip]

For once, no

Just negative embeddings.

[tired vigil]

It doesn't look bad...

it is kinda of low resolution looking

can you not do 2x hires ?

[wraith tulip]

I'm using comfyui, but i'll try upscaling it by 2. I normally don't do it by 2.

I do it by like 1.2-1.5 since I'm used to my laptop not being the best for upscaling

This rig I'm using though is good for it.

I'm also using UltimateSD upscale with the 4x-UltraSharp upscaler I found on youtube.

[tired vigil]

Those shouldn't affect the quality of the image, just how things look I guess

Pretty minimal difference

[tired vigil]

Yeah it's probably just too low res

Hires won't help that much at 1.2

Try 512 width x 768 height with 1.5 maybe

Or 2

[wraith tulip]

That's the exact width and height I have for the base image.

[tired vigil]

Kinda of standard size for models :p

If it's still looking weird try 2x and without ultimate SD upscaler

[muted spruce]

is there some way to do gender detection in controlnet and pass that mask off to Adetailer or something like that? ReActor has gender detection built-in

[velvet compass]

how can i get the keyword/vectors from a model? Like how do i know which word i should use for model X.

Is it "Kittie" or "Kitty" or "Cat" or "CAT" or "Feline" or "meow". You get the idea. how i know which keyword is matching the best for the model? is there a way to find out? also does case sensitive matter at all?

[tired vigil]

If you're using an anime model its most likely using danbooru tags primarily

otherwise the AI is smart, it will understand things :p

[velvet compass]

and specialy for "scaling keywords". like whats the ranking. Is "Big" above "Huge" or below? what about "Tall"

or is Small below Tiny, or above

like its all a bit confusing xD

[tired vigil]

thats a thing I'm not sure of, with danbooru tags it goes like small, (nothing/medium for normal sized), medium, large, huge, gigantic

SD is not suuuper good with perspective of size on things sometimes so these keywords can vary a lottt

[gaunt hemlock]

Hello...I'm seeking help around generating and using updated loras in the prompt. Subsequent generations seem to use loras from previous prompts instead of current loras.

[thorny slate]

Hello everyone, I would be looking for someone who could help me in person with the use of stable diffusion. Is there a kind soul who would assist me, perhaps in a voice channel?

THANK YOU VERY MUCH!

[shut falcon]

Hi all! I'm trying to figure out how to generate some RPG art for some fantasy races. Full details in the link below, but the short version is I'm trying to make beast people and dragon people and SD is not understanding what I'm asking for.

Would love some advice on the matter.

https://www.reddit.com/r/StableDiffusion/s/MBwxa49BEp

[muted bone]

hey guys I am new to stable diff. How can I create a single environment from different perspectives and keep the same style on different generations as well?

[viral wolf]

Does anyone else get purple-ish shadows on places like skin? What is the best way to remove them?

^Example on the eye

[vivid sinew]

If it doesnt know what something is, you need to teach it. This is called creating a model.

Like if you wanted it to generate an image of a world of warcraft kobold, it would not know what this is or how to generate an image of it. I had to teach it in order to make images like this

[cloud glade]

Prompt wizards I need ur help! I'm looking for someone that can help me find the best high quality cyborg prompt for an image to image project im making.. Also what model do u think works best??? The right side of the face has to be like a high detailed cyborg model that blends nicely! Let me know I really appreciate the help!!!!! 🙂

[shut falcon]

Awesome. How do?

[glass solstice]

prompt fat spiderman

[delicate tusk]

Trying to figure out prompts for general fantasy character portraits to use in TTRPG's. Any suggestions?

[toxic dragon]

Good negative prompt to get rid of the "back of person staring into landscape"? It's ruining my landscape pictures and it's triggering me

Current Negative prompt: Humans, people, person, face

[tired vigil]

What are your positives?

[toxic dragon]

I'm using dynamics, let me find those three specific images

Alright nvm, found the culprit, need to separate out a wildcard file with "girl" and "people" terms in it

[earnest pawn]

when I'm inpainting am I supposed to write the prompt as the thing I'm painting over (e.g. "sword", "chair", "mug") or the image in general (e.g. "knight in a tavern")?

[tall prism]

Any idea on how to improve the Colours in this? Right now, it feels a bit too dull. This is my Prompt:
(masterpiece:1.2) (illustration:1.2) (best quality:1.2) (cinematic lighting) (sharp focus), (Sandhiya) is a 5-year-old girl with Fair skin, black eyes, Black Pixie cut hair. They prefer blue colored costumes, <lora:style05:1>, multiple views, comic, line art, comic network, fighting against kraken, sea monster, sea, ocean, pirate ship, canons, storm

[dusky lotus]

Maybe try adding vivid color or colorful? Great picture by the way !

[tall prism]

Okie will try and thank you!

[tall prism]

This is not working 😔 Same quality

[tired vigil]

Come on cat friend

You sure you aren't missing a Vae?

[tall prism]

What is a Vae

[tired vigil]

the thing that makes pretty colors

[tired vigil]

https://civitai.com/models/76118/vae-ft-mse-840000-ema-pruned

[tall prism]

Oh oh

Which folder do I add it to?

[tired vigil]

Check the description, if the file is missing from civitai download directly

[tired vigil]

stable-diffusion-webui/models/VAE

Something like that

[tall prism]

Ah thank you will check

[tired vigil]

If the option to select a Vae isn't on the top left near models

You can try looking up on settings/Google how to enable that to show

Makes things easier

[tall prism]

Wait wait

I have hosted my SD WEBUI as an Image on Dockerr

So, I cannot directly change how it works if ykwim. So how do I integrate the VAE with my model?

[silver valley]

In the webui go in Settings>User Interface>Quick Settings
There add sd_vae

Then hit apply and reload ui

Then get a VAE like clearvae and kl-f8-anime2 and put them into models/VAE folder

[tall prism]

Is it possible to do it through code? Like when building Docker Image?

[silver valley]

The VAE download?

[tall prism]

Since I will be accessing the Docker Image through Runpod by sending it a JSON Input

[tall prism]

No no, enabling the Vae. I got the VAE to work when I use the WEBUI but in my actual case, I will be using it as an API through Runpod.

[silver valley]

Oh okay

I don't know how to give it the VAE in json

[next hound]

when you do negative prompt is it better to put "positive sentence" or "negative sentence"?
for exemple i don't want a fringe, so i have to write "fringe" or "no fringe" for the best result?

also, is there a way to do something for this wacky face?

[snow musk]

what are some good prompts for poses? ive been using "dynamic pose" all this time but wonder if there are more

[vivid sinew]

aDetailer

[vivid sinew]

Do you know about ControlNet?

[next hound]

thank you i will try

[snow musk]

A little bit, I got stuck on trying to do openpose stuff

[vivid sinew]

Its really easy lemme find tutotiral

[vivid sinew]

https://stable-diffusion-art.com/controlnet/

[dry sphinx]

Hey Guys I had to ask can i style my own images using stable diffusion here on discord

If not can anyone suggest me the best tutorial to download Stable Diffusion Locally

[silver valley]

Hey what's your GPU ?

You can use the bots to create your own images

[dry sphinx]

NVIDIA GTX 1660 Ti

[dry sphinx]

I want to style an image assigned to me using SD

[silver valley]

Ah okay then you would need a local install. I can link you my guide in #🤝｜tech-support

[dry sphinx]

Sure Please Link

[tall prism]

What is your preferred Sampler. I've heard some of my friends say Euler a but they shifted to DPM++ SDE Karras now. I am not really sure now 😭

[gritty harness]

Hello everyone. I'm cant seem to persuade SDXL to create top down sprites of people and other beings. It just keeps showing me side or front perspective. BTW. The planets are coming out okayish. Any help appreciated.

[vivid sinew]

idk how anyone can decide

[storm pecan]

PLEASE , I NEED a HELP . Does Anyone knows some PROMPTS that I can to use correctly for to create some images of the female characters " WITH STRECHED ARMS TO THE FORWARD " more or less " as if wish to be embraced " on Stable Diffusion ?

[still pine]

What happened to the eyes at the end? I no longer have the strength to figure out what the problem is, everything was fine until the last step

[spring burrow]

#📝｜prompting-help me ajuda

[tall prism]

Understandable

[tall prism]

This model https://civitai.com/models/64544?modelVersionId=69192" shows outputs in that style. But I am unable to get my output in that artstyle. Why is that? Mine look like this

[warm valley]

Anyone ever gotten SD (either 1.5 or XL) to have a human vomit?

dall-e does a nifty job and I cant seem to get SD to replicate anything near this

[slow mist]

open examples, find the one you like and click copy generation data

check if the selected model, image size, sampler, cfg scale, number of steps and seed match

slightly different settings might result in a totally different image

[static hinge]

for some reason I don't dig Euler, it is steady but dull (?!) and needs more steps

usually I try stuff with DPM ++ 2M SDE Karras

then I use any of the DPM++ 3M (SDE, SDE Karras, SDE Exponential)

[static hinge]

is there some video of someone actually inpainting not simple stuff? I mean struggling and resolving stuff, not just magically doing it in one go.

[obtuse pilot]

Anyone ever gotten SD (either 1.5 or XL) to have a human vomit?
dall-e does a nifty job and I cant seem to get SD to replicate anything near this

[tired vigil]

any idea what is causing this?

[warm valley]

@tired vigil - mainly the aspect ratio causes that

[tall prism]

I've tried that, the art-style completely changes.

[tall prism]

I see. Lemme try

[strange jasper]

main coz is aspect ration, i think --ar 4:3 or 9:16 will be good
try this.

[strange jasper]

another thing is that ai generation is not for very big sizes. there is a limit for generations now dive in to that first please.
thanks

[tall prism]

DPM++ 3M is changing the art style completely. DPM ++ 2M SDE Karras looks feasible! Thank you so much!

For some reason, DPM++ 3M makes good outputs but when ADetailer starts working on it, it breaks

[river apex]

I'm creating a comic and I've run into a couple of issues that are baffling to me that you might be able to help me with. First off, for some reason the AI will simply turn off ALL the inside lights in a scene changing it from a pleasant ambient look to harsh lights and shadows. I see that there's an 'ambient' prompt but I've found no instructions on how to use it. I hope bikinis don't violate any rules:

http://www.oliviacrossbooks.com/templar/tba/AI/Nakitomi/Agents/Gunne/gunne_lights-out_1024.jpg

I should preface by saying that this has happened in two of my scenes so far, I have not changed any lighting settings and yes, I've tried ALL the lighting prompts that I could find - they have the effect of bringing a 40 watt bulb into a pitch black room

My other issue, while I'm here is how the heck do I prompt out other people in the scene? It's a strange phenomenon that the more I run the same prompts, the more people start to filter into the scene until it's practically a mob.

[cunning skiff]

prompt for generating images

/promt foir generating imaging

[river apex]

Here are the positive prompts for the girl on the right. The negative take up all 1024 chars but I can post them if you like.

revAnimated_v122(full body shot) RAW photo, detailed textures, sharp focus, ultra-high pixel detail, intricate, realistic, movie scene, cinematic, high-quality, full colors, incredibly detailed, 4k, 8k, 16k, hyper-realistic, RAW photo, masterpiece, ultra-detailed, professionally color graded, professional photography, Diffuse light and create a soft, even glow., soft, natural light, Bounce light and soften shadows, an extremely beautiful 18 year old woman, (tall slender body, tall narrow waist, , long narrow waist, high navel, small round hips:0.3) medium breasts, big blue eyes, long curly blonde hair, ((extreme narrow diminutive small size green thong bikini)), green high heels, green hand gloves, green thigh quiver, (((firing a gun))), ((standing and aiming a gun)) in a (((very wide well-lit office hallway))),

You can see where I was getting pretty desperate on lighting.

Hete's the positive prompt for the first girl: revAnimated_v122((full body shot)), ((diffuse lighting)), ((masterpiece)), ((best quality)), ((extremely detailed)), an extremely beautiful 18 year old woman, (tall slender body, tall narrow waist, , long narrow waist, small round hips:0.3) medium breasts, big blue eyes, green flower in hair, long curly blonde hair, ((extreme narrow diminutive small size green thong bikini)), green high heels, green hand gloves, green thigh quiver, (((firing a gun))), ((standing and aiming a gun)) in a well-lit office hallway (curious, careful, moaning) , (((diffuse lighting))), girl, kristinapimenova, realistic, GdClth, taaarannn, lora:GoodHands-beta2:0.800000 (BTW: 'moaning' is just meant to open her mouth so she can be talking.)\

Obviously I should just copy the working prompt, right? Unfortunately that didn't work.

http://www.oliviacrossbooks.com/templar/tba/AI/Nakitomi/Agents/Gunne/gunne_lights-out02_512.jpg

[soft wadi]

Hi, I'm having a problem with a prompt. I've tried it on several notebooks, different seeds, different models, different sampling methods, different everything... I always get two women instead of one.
This is the first time I've had a problem like this... I've changed the prompt to something like "one woman alone in a field of flowers" or something like that, but it doesn't work.
It's terrible... I don't understand. Can you please help me?

"A beautiful dark hair woman, lies down in a red flowers field, half closed eyes, soft red bud lips, hyper realistic, hyper detail, cinematic light, trending on artstation, sharp focus, studio photo, intricate details, highly detailed, by greg rutkowski and James Jean" and the negativ prompt is : "nude, two women, ugly hands, ugly arms, two faces".

Here I tried : dreamlike-diffusion-1.0 / DPM2 / 25 steps / 1280x720 / Seed 1234561388

I always have two girls stucked together. I loose all my ressources trying to get one woman

I had this one but with the same prompt, if I generate again, I got two ladys ...

"A beautiful lady lying in a red flowers field, intricate details, cinematic lights, high details, by greg rutkowski and James Jean"

[river apex]

I tried your original prompt and did get two women. I changed the descriptive part of the prompt and placed all the description of the woman first and the action second.

A beautiful woman with dark hair is lying down, eyes half closed, soft red bud lips, in a field of red flowers,

It worked in DreamShaper just fine.

[soft wadi]

Thanks for your help. Can I ask you what stable diffusion you are using?

[river apex]

1.5

[soft wadi]

It doesn't work for me with Dreamlike Diffusion

[river apex]

That prompt? That's strange.

[soft wadi]

Local stable diffusion? I'm on a notebook google colab

[river apex]

Worked fine in ArtUniverse

[soft wadi]

Can you see? It does not work... so strange

[river apex]

6 renders in RealCartoon 3D worked

[soft wadi]

Is there so many differences from a platform to another one?

[river apex]

Must be. I just did 6 renders in GhostMix and 6 in RevAnimated. Just fine.

[tired vigil]

Are you copying the prompt from someone on civtai ?

[soft wadi]

No, I'm not. I wrote it myself

[tired vigil]

It could have been done using things like controlnet, getting aiming poses are pretty hard on SD

well

You should try some controlnet open pose stuff then, those are hard

wait

Not you chloe x3

[soft wadi]

So I don't understand.. I'm not lucky...i loose all my ressources... I really need the local stable diffusion to learn and test and do it as much as I want

[tired vigil]

Try lowering your resolution to 768x512 and do hires 2x

above 768 things can go... that way

[soft wadi]

Ok, I try

[river apex]

Cat: I use a lot of OpenPose but I'm not sure how that will help my lighting issues.

Chloe: I use Tensor.Art and Sexy.AI for all my renders. They're mostly free and have a LOT of options also you don't have to worry about having a 12 GIG GPU!