📝｜prompting-help

[lone badge]

yes

[fathom lance]

is it any good?

thx!

[lone badge]

but you need to install it at home for it

#1080946152318443610 has some nice tools for that

frankly, you're asking a mod that has been a fan of this specific tool for more than 6 months now
so I'm a little biased but yeah, it's good

[cobalt coyote]

bros its only putting out legs

and nothin else

i've had enough panty shots im trynna make sakura here

[stable tiger]

i need to convert a photo to hyper realistic pencil drawing style, img2img and prompt "hyper realistic (black and white:1.8) pencil drawing" style:"dpm ++ sdekarras"what else i should use , any artist ?

[versed radish]

5 kids smoking weed

[sullen quartz]

hello, does someone know how to fix eye separation on realistic faces?

[south solar]

I am not familiar with that model, but I have had luck with "robot" giving me ball joints in some other models.
That said, from my experience you are going to have trouble getting both clothes and robotic parts. I have tried a few different models, and generally the only consistent way to get both was inpainting, then upscaling to get rid of the rough edges.

The fact you have a hero forge model might make it easier. I think you were on the right track with img2img. Try a lower denoising strength. Also look at other models. I think Dungeons and Diffusion knows what warforged are by default, and there may also be LoRAs or textual inversions.

I find with clothes on robots, txt2img really doesn't work, but if you use inpainting to put both in the image, img2img can figure out which parts are which.

[vagrant igloo]

I appreciate you getting back to me. I was able to create something I wasn't satisfied with but my friend fell in love with.

[south solar]

Sometimes that is all you need. I have been experimenting with robots a lot for similar reasons, and I have gone through a lot of failures. Hope the campaign goes well.

[fluid arch]

need a good prompt search website

[silver valley]

Try maybe Dreamshaper

[cobalt coyote]

hello im trying to make dua lipa somewhat in this style

shot from the side, looking down at phone, not at the viewer

instead i get results like this

i have stressed in the prompt how desperately i want her to not look at me but she won't do it

anyone got an idea?

[forest folio]

https://youtube.com/shorts/aFaSKkSfemI

[cobalt coyote]

guess the channel is deadge

[teal yoke]

whats the difference of commas and periods in prompts?

i could not find any info about this

[silver valley]

This is done with the samdoesarts model or the lora

[cobalt coyote]

it's an original artwork

by samdoesarts

[silver valley]

Oh xD

[silver valley]

You have looking at car, looking at phone and looking away from viewer

Try only side View or from side

[next flint]

Does anyone know of a very general and basic prompting guide with some examples?

[silver valley]

Here:
https://openart.ai/promptbook

But its very basic

[cobalt coyote]

@silver valley since ur here

i've moved on to a different "project"

do u know what i can tell it to avoid making these weird limbs

[next flint]

thanks

[silver valley]

Add "multiple limbs" as negative

And deformed

Then you can try change the resolution to a 4:3 one for example 512x768

[cobalt coyote]

that helps with limbs?

aight im kinda cooking now

thank you

[silver valley]

Hey looks good

[cobalt coyote]

still trynna figure out a way to make her stop wearing this leather armor hahaha

but gettin there

[ebon axle]

{$sepehr_key_70} Flower

#1072220168534642768

[next flint]

anyone have a prompt generator for gpt? im trying to make one but im not getting amazing results

[next flint]

Its not great but ive gotten this to sometimes work

Is there a site that converts standard sentences into prompts?

[fickle linden]

Who has soccer pics?

[timber pebble]

If you just need to barely tweak a subject's pose, sometimes it works for me to import the image into 3D open pose editor, then adjust the post by hand, then try the original prompt with the adjusted pose using controlnet.

[cobalt coyote]

i just checked out that editor and holy

that seems far too advanced

[outer path]

whenever i ask for a coffee cup on a desk it gives me a top down view as if I were on the ceiling. I tried "low camera angle" but can't get the POV of someone sitting at a desk. Any ideas?

[obtuse torrent]

,,.. from_above, close-up shot of a coffee cup ... ?

[outer path]

hm, I tried from_side, from_front, closeup, but it still only gives me views like this

I'm also brand new at this so maybe there's some basic syntax I need to learn. Thanks @obtuse torrent

[obtuse torrent]

... try the rest.... full_body view of coffee cup, from_above ..

@outer path

sorry which view you need?

[outer path]

From the side so you can see what’s on the mug.

Not from the top

[obtuse torrent]

...yep..some photographic term is needed but i not have clue.. lents x,x etc.

[finite prairie]

hey I'm still trying to figure out img2img, and I'm not sure how to prompt it. Suppose I have picture of a blonde woman and I want it to generate that same picture but she's brunette. Does the prompt need to be basically a prompt that might generate the whole picture, then with the blonde as a base it'll turn out like I want, or is it more like I give it the blonde picture and prompt "make her brunette" or something similar?

[obtuse torrent]

@outer path it work?

[outer path]

yes! how'd you do that?

or are you asking me to rate your setup? 😉

[obtuse torrent]

ok can i pm you?

@outer path

to explain the base idea ..

[outer path]

totally thx

[next flint]

i think there is a guide for this in the auto1111 github wiki. its like (direction)(degrees)(direction)(degrees)

[outer path]

rock and roll. thanks @next flint

[next flint]

lmk if it works

[next flint]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#instructpix2pix im guessing this is what you're looking for

[next flint]

img 1. coffee cup on a desk,(wide angle shot), up 40, left (40 mm:1.97), unreal engine, trending on artstation
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 1583257515, Size: 512x512, Model hash: 7f16bbcd80, Model: dreamshaper_4BakedVae
img 2.coffee cup on a desk,low angle shot (no helmet:1.14), highly detailed, hyper realistic, trending on artstation,4k, chromatic aberration
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 4078738935, Size: 512x512, Model hash: 7f16bbcd80, Model: dreamshaper_4BakedVae

[outer path]

heck yea

[next flint]

https://unimatrixz.com/blog/latent-space-camera-positions/ this will probably help

[outer path]

i like the camera coffee in number 1

must really help you focus

[next flint]

this is probably my favorite so far
coffee cup on a desk,Long Shot, up50, left 50, back 50,cosy,theme,surreal,in a coffee shop, background is bustling city on the right, smooth, Sharp focus, high detail, (Soft light:1.5)
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 279237305, Size: 512x512, Model hash: 7f16bbcd80, Model: dreamshaper_4BakedVae

[outer path]

almost too perfect

thx to @next flint

[frigid kindle]

Help!

I tried to generate an old photo portrait of military soldier who doesn't wear any head wear, but so far SD keep generate full body photo, and some image has headwear, some has tiling, it seems doesn't stick with my prompt.

Positive:
*A black and white photo, colourized face portrait of male or female military soldier in modern era, they are not wearing any headwear, their age between 30 to 35 years old, good looking face, realistic, asian race, random expression, the timeline is in 2000s. *

Negative:
ugly, old, weird, painting, drawing, photo frame, wearing headwear, wearing helmet, wearing hat

Sampling method:
DDIM

Sampling steps:
128

CFG Scale:
7

What should I do?

[silver valley]

You can use pix2pix model to make her brunette, with that model you can say "make her brunette" and it will change only the hair

[silver valley]

Hey you dont have to say "who doesnt wear any headwar" if you say it like that ai will give him headwear. Because its in the positive prompt.
You just have to put
Headwear, helmet, Barrett, into the negatives.
Then a Trick is to describe his hair like short hair

[steady fern]

I'm trying to modify a family picture to change it to a ghibli style. I am using the spirited away general model. In img2img I set CFG to 25, denoising to 0.5, steps to 20 and sampling to Euler. This works not bad for a portrait, but it is terrible for a bigger picture. So I guess, has anyone been able to do something like that ?

[frigid kindle]

thank you very much!

[jovial iron]

Any tips for creating two characters in one frame?

[next flint]

ive seen someone do something similar with 0.55 denoise and 16cfg. I think downsizing the image and then using hirez fix might help

[next flint]

Inpainting and switching models. You can render a scene with something like a robot courtroom and then inpaint on the person on trial with a custom model to change it to them . Or https://github.com/opparco/stable-diffusion-webui-two-shot

[tired vigil]

#imagine

[next flint]

is there an extension that converts midjourney language into sd language for prompts?

[fiery pollen]

Please can someone help me understand the prompts that were used to create this? I've tried CLIP interrogator & ControlNet but I'm unable to recreate the lighting & shadows.

[warm bolt]

Do you need help with guessing the prompts to get this image? If yes, with or without the img2img feature?

[fiery pollen]

Yes please, maybe both?

[tired vigil]

#prompt

[stone igloo]

@tired vigil

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[jovial iron]

How do I describe this hairstyle?

[cobalt coyote]

how do i stop it from adding all these limbs

a girl, beautiful blue hair,
Negative prompt: (low quality, worst quality:1.4), multiple limbs, deformed, cat ears, EasyNegative, badhandv4, (bad_prompt_version2:0.8)

[silver valley]

Maybe Grey hair, long hair with side bangs

[jovial iron]

Yeah that partially did the trick

[willow quarry]

I was curious if you had any knowledge on how Pebblely.com is addressing this issue with their AI-generated product images? I was wondering if they are using any specific techniques or methods to prevent objects from sticking to the product images...any clue?

[cobalt coyote]

bros i need an inpainting expert

there is a part here which i want gone, to blend in with the ground there instead

it's the part on the left knee

that additional line for skin it added

sadly i cant zoom in and post the close-up cuz discord blocks it but yea xd

which settings would i need to tweak ideally for such a small change?

[lone badge]

sorry not the prompt expert but the TOS guy for a second...
The green hair here is really nice, but the previous one in blue, I got to delete for the NSFW rule there. The pose and intent of the character coupled with the partial nudity makes it too sexual for here, sorry.

[cobalt coyote]

oh sorry about that

[lone badge]

no problem, I clearly see it wasn't intentional, just wanted to give you the warning so you'd know

[lone badge]

You can try to use the "Inpainting area : Only masked" option, and change your prompt to something more fitting to just the knee. This will make it faster (smaller sizes) and give out more details

[cobalt coyote]

i've managed to do it like this since the only masked didn't deliver many results

now my knee is fatter but the issue is fixed

[lone badge]

that does work, yeah. It's really hard to force those kind of details sometimes

[jaunty echo]

Anyone know how i can make it look a little better?

Prompt: A visually striking image of a lone astronaut standing on a desolate planet, with a distant galaxy or nebula in the background, using a cool color palette to evoke a sense of isolation and adventure, 16k ultra HD,

Negative prompt: no watermark

[next flint]

can you be a bit more specific about what you want to be better

[jaunty echo]

The backpack seems to have replaced the head, apart from that it's the image I'm going for.

[next flint]

something like this?

[jaunty echo]

yeah

[next flint]

A visually striking image of a lone astronaut standing on a desolate planet, with a distant galaxy or nebula in the background, using a cool color palette to evoke a sense of isolation and adventure, 16k ultra HD,
Negative prompt: ng_deepnegative_v1_75t, verybadimagenegative_v1.3
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 4168358972, Size: 512x512, Model hash: 44f90a0972, Model: protogenX34Photorealism_1, Denoising strength: 0.5, Hires upscale: 2, Hires upscaler: R-ESRGAN 4x+ Anime6B

i think the deepnegative and verybadimagenegative help keep it from messing up the head and backpack

[jaunty echo]

I need to learn more about negative prompts

is there like a guide or article about negative prompts?

[next flint]

im sure there are lots of good ones but it relies on the same principals as normal prompts. you can research it but id recommend just getting some basic textual inversion ones

https://civitai.com/models/11772/verybadimagenegative, https://civitai.com/models/7808/easynegative, and https://civitai.com/models/4629/deep-negative-v1x are some good ones

[jaunty echo]

Appreciate it

[next flint]

if those dont work there is a pinned message that has a lot of all around good negative prompts. Doesnt work as good as TI but still works and can be fine tuned a bit

If you’re not sure about what would be good you can try dynamic prompts set to effect the negative prompt for low step counts to see what gets you closest

[tranquil shard]

I see sometimes grouped prompts like (tag1, tag2, tag3), tag4, (tag5, (tag6, tag7)) - does it really have any effect?

[floral leaf]

hey there if youre not too busy, can someone help me get a dark room, ive tried, dark,dark room, and i cant think of anymore ways to describe a dark room

(dark as in no lights)

[silver valley]

do you use offsetnoise or a lora for the dark room?
what prompt did you tried?

[floral leaf]

No I haven't used either. I just tried dark room and dark, alongside the room I was trying to get.

I know how to train Lora's however. I have done it for a few characters

[silver valley]

For dark rooms you can use https://civitai.com/models/13941/epinoiseoffset

[floral leaf]

Ok thanks

[trim ether]

anyone can help to generate a better face? always deformed like this

[sonic dagger]

Is there a way to generate multiple images from a single prompt, separated by a special symbol or something?
thinking of generating a storybook from gpt

[empty sluice]

Increase your batch count

[sonic dagger]

But I'll just end up with something similar

I want to gen 100 images, each unique, overnight

[summer light]

Golfer

[quasi timber]

Need help with the faces. I'm not looking for perfect handsome faces. Just some that don't look like Frankenstein's monster meets Rocky Dennis from Mask.

[sullen venture]

How would you define the art style of league of legends or bastion? There's a color richness, chunkiness, anime influence, and 3d watercolor/oil texture...but is there anything that defines all these together?

[umbral shell]

i'd define it as... lora:leagueOfLegends:0.6

[sullen venture]

is that a model extension?

i see the ahri page, but it's very anime

the trouble is when I prompt league of legends direclty i get a ton of artifacts and weird anatomy, as their splash art is really wild. lots of magic and mechanic bits flying around, wacky poses, etc, that confuse the model

[next flint]

It’s a Lora file. I think that one is on civit. They are called similarly to TI but need the <> signs around them

[next flint]

Try a different sampler

[south solar]

No guarantee it will work, but sometimes with full body shots upscaling is needed to fix the face. It seems like the smaller face means SD doesn't pay enough attention to it. Might not work, but worth a try.

[tough marlin]

Message # a boy playing video games

Prompt help

[sonic dagger]

Is there a way to generate multiple images from a single prompt, separated by a special symbol or something?
thinking of generating a storybook from gpt

I want to gen 500 images, each unique, overnight

I could use wildcard but I would repeat the images

[hushed spire]

hello everyone, i need your help...i'm losing sleep 😦 How can i get a "non lucid" skin in SD ? Could i post an example or the prompt i use ? thanks a lot guys, have a great day 😉

[next flint]

Im not sure what you mean by non lucid. could you send one of the problematic pictures?

[next flint]

you would either use a script or dynamic prompts. Dynamic prompts would add special modifiers. the scripts would be prompts from file or textbox or prompt matrix. the dynamic. the file or textbox option basically allows you to queue prompts while matrix allows you to experiment with different sets of modifiers.

[foggy python]

Sorry for the weird collaged picture full of arrows, it's just to point at the areas that are bothering me.

I've been stuck for a few days now trying to perfect a style. Which I can best describe as being "game concept art" or something. I have found the tokens and weights that get me good results but more than half of the time, the lighting is really extreme.
The problem is, all obvious tokens to help remedy this feel like they're really seed dependent and they might make it worse on other seeds.
For example using "diffused lighting" seems to make it so in some seeds it just adds lights because I used the word "lighting" or something, making the effect even worse etc.

I'm looking for a way to make it either consistently like my good results, or as frequently as possible at least.
It currently feels to me that in my bad results, the subject is "standing in a studio" with very heavy light in their face making their own facial features cast a very hard shadow, while also highlighting the skin and hair too much.

Any tips and tricks are very much appreciated, I've been stuck on this and it's really demotivating.

my current "if all else fails" route would be to generate a fuckton of them and then train an embedding on the ones with good lighting... that would work right(?)

[trim ether]

ah i understand, my gpu so weak can't generate hi res/upscale image well, thanks for the point out, i might try later with online generate 👍

[granite thistle]

Anyone know how to create a queue of prompts? I saw there was a script section but unsure of how to use it. What I want to do is run through a list of positive prompts each using the same negative prompt

[warped drift]

anybody getting good results for painted miniatures?

[lone badge]

"plastic doll, plastic texture, wax doll" as negative will help

but

sorry

this goes over what the NSFW rule (rule 4 #✍🏼｜rules-and-tos ) let us let through, I need to delete those

[old breach]

I know there's a PVC model floating around. Have you tried training your own model on a bunch of different painted miniatures from diff tabletop games? To get it to learn that "style" ?

[warped drift]

ive had a really hard time w training

i dunno if theres an up to date ideal explanation vid

[old breach]

I used dreambooth and started w the base 2.1 model then trained it on stuff i wanted for a style. There are a few good articles out there on it I like the one from ByteXD

[copper slate]

This was made in SD with A1111. Would anyone know the what extensions I could use to achieve this level and style of art?

I didn't make this btw someone I follow did

[silver valley]

You mean model ? Cause extensions dont give you a style

You can try the Dreamshaper or Dreamlike models

[copper slate]

Oh yes sorry I’m still super new to SD and wasn’t sure how to ask

What are dreamshaper or dreamlike models?

Where do I find the models? @silver valley

[silver valley]

You can find them here and a lot of other models, loras, embeddings and stuff
https://civitai.com/

[copper slate]

Are they safe to download usually?

Also thank you for the help

[silver valley]

Yes they get Scanned on upload there. You can also stick to .Safetensor files these are safer than .ckpt

[copper slate]

Ok cool. Is there any guide I can follow on how to install and set them up?

[silver valley]

Its pretty easy:
Models (all over 2gb) going into models/Stable-diffusion folder.
Everything below 2gb isnt a model.

Loras go into models/lora
Embeddings(Textual Inversion) go into Stable-diffusion-webui/embeddings
Hypernetworks go into models/Hypernetwork

[copper slate]

Thanks for the help I appreciate it

[silver valley]

No problem have fun trying out some models

[next flint]

Im not sure what model you're using but have you tried adding the epi noise offset lora?

it's supposed to be used for darker images but it seems to be able to get an even lighting for a lot of different images

[blazing belfry]

Trying to generate variations of my drawing to get some ideas but it's generating some weird stuff.

[next flint]

try control net.

try other preprocessors and models but canny might be the best for this case since there doesnt seem to be too much intricate detail

[next flint]

this might cause the head to not change a lot. You could try putting the original photo into paint and making the head area white so it doesnt get accommodated for in control net

[next flint]

For concept art you could try the chartuner TI https://civitai.com/models/3036/charturner-character-turnaround-helper-for-15-and-21 . Might not be exactly what your looking for but its good for concept art and whatnot

[next flint]

My prompt that i used for this wasnt the best but is this the lighting you're looking for?

[haughty nexus]

I want fluffy hair with a black or brown undercut? its unreasonably hard to get it and when i try to achieve it with mohawks its way to hard as they make it to much of a mohawk

for male

[next flint]

could you send a real image of what you want and then your output with the workflow?

@haughty nexus Is this more what you're looking for

Here's the workflow
masterpeice, best quality, 1boy, solo, looking at viewer, realistic, 8k, sharp focus, photorealistic, brown hair, full body, highres, highly detailed, ultra-detailed, intricate, illustration, standing, cinematic lighting, brown hair with undercut, undercut hair

Negative prompt: low quality, blurry, bad anatomy, worst quality, text, watermark, normal quality, ugly, signature, lowres, deformed, extra limbs, disfigured, cropped, jpeg artifacts, bad hands, error, mutation, missing fingers, username
Steps: 30, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 1574018582, Size: 512x512, Model hash: e04b020012, Model: rpg_V4, Denoising strength: 0.5, Hires upscale: 2, Hires upscaler: R-ESRGAN 4x+ Anime6B, Score: 7.36

idk if that's exactly what you're looking for but you could probably adapt that towards it

[tepid oar]

This was made in SD with A1111. Would anyone know the what model I could create this style of art?（Japanese style, Hayao Miyazaki style），I tried to use these models （SD1.5、anim-full、anyting v3/4、Rev animated）and the prompt control, but none of them achieved this effect，help me！

[hushed spire]

for this i asked ^^", thanks for the hint... i'll try that

[silver valley]

hey, im getting this with expmixline_v20 model.
Prompt:
1girl, in intricate kitchen, sitting, old studio ghibli style:1.2, Hayao Miyazaki art:1.2, pale, grain,

[foggy python]

thx for your effort, yes I have tried the offset noise lora. I have also found that the background is influencing the lighting a lot. Because I have solid 1 color backgrounds, it seems that stable diffusion thinks my subject is inside a studio for a fotoshoot or something, and then there are some results that will have that big studio light effect (with the hard shadows)... so I'm going to have to experiment with some "outside photography" style stuff or something 😄

[tepid oar]

@silver valleyThank you for your answer, I will download this model and try it out

[carmine nova]

what words can you use in the prompt to make the ai do full body shot lol

its really really focusing on the face in mine

[silver valley]

Try full body, cowboyshot, or wide view,
you can also describe the shoes

[next flint]

I think there is a studio ghibli Lora

[silver valley]

Ah right!

[next flint]

Stupid question but have you tried putting lighting things into the negative

[foggy python]

Yes and so far, I've not found anything that works consistently. 🤷‍♂️
So I'm now just creating a bunch of gens, only keeping the ones with the lighting that I want, and I'm going to train an embedding on that

[warm frost]

Full body, fisheye

[carmine nova]

AI art fighting the intrustive thoughts to put massive glowing circles on the titties

[next flint]

Do you need help so it doesnt do that or are you just sharing goofy ai stuff

[tepid oar]

I didn't find this lora

[sharp mauve]

is dynamic prompt exte good?

[silver valley]

Seems like a missing vae for that model

[next flint]

https://civitai.com/models/6526/studio-ghibli-style-lora

[tepid oar]

Thank you ！

[tepid oar]

I tried it, but the style of the picture is still not the same, let me try to train a lora

[feral ore]

Hello everyone. I saw many prompting suggestions mentioned that adding Resolution such as "8k" or "4k" as a part of prompts would lead to better quality. But I am confused does resolution really can affect the results?
I mean the resolution of an output image is obviously defined by controlled parameters outside the prompting words, so what is the point of adding that thing to prompts?

[next flint]

A lot of higher resolution images are more highly detailed. Saying a resolution in the prompt is kinda telling it to be accurate or have the same level of detail. having resolution in the prompt has nothing to do with the actual resolution

[glossy scroll]

1

[next flint]

How similar do you want it

[next flint]

Because you could always do canny control net

[tepid oar]

I don't want to directly copy the composition of the picture. ControlNet may not be a good solution. I prefer to obtain a picture style, color and character characteristics similar to this one (Hayao Miyazaki's childhood style)

I have seen this effect on many open platforms (China), but I haven't found his training model yet, maybe the developer didn't make it public

[tepid oar]

I found the source of the pictureBut he is packaged in the app, the model is not open source

Mantra (Prompt): A happy walking girl, long black hair, braids, happy smile, big eyes, sunflowers, flower wall behind, spring, yellow shirt and short skirt are very beautiful,
Model: Meiman, CFG scale: 7, Hires upscale: 1.5, LoRA: Ghibli (0.55), Ink-2 (0.5)

https://m.wujieai.com/s/IBY5L089
After one-click copying, open the [Unbounded AI] APP or click on the link to automatically fill in all parameters to create the same style

[feral ore]

Now I understand that. Thank you for explanation.

[timid acorn]

Hi is there a way to tell sd on the prompt to use either one of multiple options?

By default, without any scripts or fancy stuff. Something like:
A (green or red) apple.

I have tried some things, and i come up with mixes instead of random options chosen.

[lone badge]

yes there are, multiples options in fact :

• a|b|cwill try all combinations of those 3 https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-matrix
• X/Y/Z Plot script, using "prompt S/R" feature, lets you make a grid to compare results by swapping part of the prompt https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#xyz-plot
• using the built in extension "dynamic prompts" https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Extensions#dynamic-prompts
  example for that last one : A {house|apartment|lodge|cottage} will run each time using another random value in those 4

[frosty thorn]

hey guys can you tell me what a good prompt looks like , am new

[lone badge]

Andree, a guide here, made this little paragraph explaining how he makes good prompt
#📝｜prompting-help message

Usually, a "good prompt" will depend on the model you are using, but should cover :

• the type of media (a photo, a sculpture, ...)
• the subject (a cat, a person) and its description
• the background and its description
• the lighting, the textures of the surfaces
• "made by" some artists inspirations you want to find. "RAW photo, taken on iphone 6" or other things like that for more realistic results for example too
  You can also add moods, feelings, ... that can push the image in a given direction
  Finally, the negative prompt is somewhere you may want to add things that you don't want to see pop up, depending on your first results

[frosty thorn]

I appreciate it

[lone badge]

no problem. Sometimes, finding the right keyword/token to use will require to think outside of the box.
Using the name of an actual real life event for example can help a lot. you need to ask yourself, if I was to find my photo on Internet, how could it be titled ?
For example "wildlife photography" is good, but will work even better with "taken for National Geographic"

[cunning edge]

hey, I cant figure out how to create pictures that show a person in a specific position, for example "a woman sleeping on a sofa". tried many many times, also with negative prompts, but nothing seems to be working 😦
anyone has any advice?
or maybe SD isn't there yet?

[lofty forge]

Hello. Can somebody give advice how to write prompt in SD? At this moment I write prompt like this > "the woman, brown bob hair, blue eyes, smile, wearing pink bikini, hold watermelon piece, 8k, extremly detailed, hyper realistic, global illumination" with negative prompt > "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry,bad eyes,bad face" and "anything v3" model, but most of the time I receive broken face and artefacts

[obtuse torrent]

Hi, I need advice on getting a short pointed ear (Tolkien, elves) and not an extra long one (earth elf or gnome, like the ones in the H. Potter films), I haven't made much of a difference using words like elven, elf, pointed. .. it would help to mention some style or author to achieve them? .. oh is it a matter of the model used?, thanks ..

[shrewd owl]

What causes things to break down like this^

[stuck osprey]

too many or too little steps

depends on sampler

[shrewd owl]

LMS

[stuck osprey]

how many steps you have?

[shrewd owl]

I think that was 30

Bumping to 80 still gave weird artifacts

But less

[stuck osprey]

What about guidance scale?

[shrewd owl]

CFG?

[stuck osprey]

yup

[shrewd owl]

I think it was 8 or so

[stuck osprey]

hmm

[shrewd owl]

I'm running a script to parameterize through 20-120 steps and 2-14 CFG

[stuck osprey]

on LMS I got this weird result when I had too many steps

I think it works well between 10-20

[next flint]

Either add weights to the position or just use control net

[median fern]

I have a RTX 3070 and 've been waiting for a few minutes now, seems like nothing is happening still with these settings. How long should it take for an image to generate?

Or am I doing something wrong?

oh

💀

[next flint]

Keep the main idea at the front. For example, “A woman wearing a pink bikini holding a piece of watermelon” followed by the descriptors. Also try using textual inversion for negatives. The prompting guide on the auto1111 wiki is a good start

[median fern]

wtf

[next flint]

Maybe it’s your other programs on your computer. Takes me 1 min per 1000x1000 image at 30 steps each and i have a 2060S

[median fern]

I am running a game while I'm generating that's prob the reason lol.

[next flint]

Try fantasy like models. You can always ask gpt on how to describe them if you are confused or just in paint until you find something you like

[next flint]

Yeah. You probably are using just enough vram on the game to not stop the generation but aren’t giving it enough to go fast. Check the command prompt to see the it/s or s/it

[next flint]

At that resolution on a 3070 I’m guessing it might take 30s at most if you are just running SD

[next flint]

It might work to specify the type of style of ear you are going for. Try “Tolkien style ears”

[median fern]

I can see some goofy stuff, like uhh maybe an extra arm.

[next flint]

Have you tried textual inversion

[obtuse torrent]

@next flint thanks, so is more about author style and not model construction? --- same for eyes? you know some sheet style - eye type?

[next flint]

My last response was a shot in the dark so I’m unsure. I just know from previous experience specifying things to be in specific styles sometimes helps

[obtuse torrent]

Sure thing .. the main problem is SSD space, i can't switch or download many models :S . but i thank you again (and i fear of gpt and picking my number in some t-800 list) @next flint

[next flint]

Try Dreamshaper. Sounds like it might work for this and is all around a great model

[lofty forge]

Thanks

[knotty sigil]

anyone know a good model that can design a logo with letter?
or prompt maybe
(fantastic logo lol, for minecraft server..)

[carmine wind]

Anyone knows how to make photo realistic picture? I tried all the ultra realistic words but it’s not working that much

[feral ore]

I saw some prompts which include multiple words in one bracket such as (low quality, worst quality:1.4).
I guess it stand for (low quality), (worst quality:1.4) or (low quality:1.4), (worst quality:1.4) but I'm not sure which exactly is it. Does anyone know?

[obtuse torrent]

@carmine wind you are using photo realistic models?

[vocal portal]

guys how should i generate game ui like buttons. Can you suggest me a prompt

[silver valley]

Hey, here you found all information about this feature:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#attentionemphasis

[sacred wolf]

Hello guys! Can I get some help with prompts 😄

[silver valley]

Sure what do you want to know ?

[sacred wolf]

I want to make a Balenciaga meme video with my friends pictures. What do I need to do for the AI to make realistic photos and keep the faces recognizable?

Is it even possible?

What settings, do I need super clear normal pictures etc.

[indigo sedge]

So I may just be dumb, but how do disable hypernetworks/Lora in stable diffusion? there seems to be no way to just select none in extra networks and it always adds it to my prompt. I'm using automatic1111's gui

[silver valley]

If you want your own images as generation Material you need to train a lora.
But you can try first to put the image into img2img tab and play around with that

For realistic images you need a good model like realistic vision and a prompt like
Photo of cat, photorealistic, high resolution etc

[sacred wolf]

So I need to download a specific model ?

[sacred wolf]

How hard is it to "train a lora" idk what this is but yeah 😄

I want to create a similar Balenciaga meme video with my friends pictures, I can obviously do the animating/voice over part, that's easy but I'm wondering if its possible to generate these kind of high quality realistic images that keep the faces recognizable

[silver valley]

a lora is something that goes on top of a model and can be trained to do specific characters or styles.
how to install and use: (you need at least 6-8gb vram)
https://www.youtube.com/watch?v=70H03cv57-o

[sacred wolf]

it seems like i only got 4k

🥹

[silver valley]

oh then you would need to train in a google collab workspace

[silver valley]

you need to remove them from the quicksettings menu in the settings

then use the third button udner generate button to get to the additional networks menu

[inner ferry]

ok, but if I do this, the hyper stay applyed to the settings. Indeed don´t has a way to remove without a "none" option. This appear when install A111, but vanish when select any hyper. Diferent with Vae selector, it stay with none as a option.

[silver valley]

i dont know if i know what you mean sry

[frosty thorn]

pls what are the most known prompts to get better looking gen

masterpiece etc..

sometimes masterpiece ruins my gen completely

[next flint]

best quality,8k,highres,highly detailed, ultra-detailed,hdr,etc

[tough portal]

i finally donwloaded midjourney and am new does anyone know what the sliders do?

[haughty shore]

help newbie, why sometimes color t shirt not color i want

[tall carbon]

how can you know picture prompt?

[silver valley]

Try pink shirt:1.2

[silver valley]

You installed Stable-Diffusion Webui not Midjourney, you can hover over the names of the sliders to get more information

[empty tundra]

I have a drawing of a character that I want to generate more of but in different angles etc. Does anyone have any tips on how to generate the same character? Is there a good control net model for this?

[next flint]

You cant. there are some features bult in and some gradio interfaces that can get somewhere close

[next flint]

this sometimes helps https://huggingface.co/spaces/hysts/DeepDanbooru

[next flint]

Go to civit and look up either Charturn or 2D Sprite style

[empty tundra]

I understand I can use Charturn to generate a prompted-character from different angles. But if I have an image of a character which I want to rotate, how would I do that?

[next flint]

im not sure but i would think that if you did img2img and used an open pose thing with the camera angle from behind you could with enough denoising strength

[vale marsh]

Are there a list of command prompts for in depth commands for like "sharp focus", "70mm", "deep focus" and such on any site?
cause I can't think of any of them from the top of my head 😅

[next flint]

idk if anyone wants it but here are the top 1000 of both positive and negative tokens with comma separation https://pastebin.com/5nQ0MwzW

[next flint]

Most popular settings
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Size: 512x768, Model: chilloutmix

[obtuse torrent]

@next flint it is in revision only owner have access ... but thx

[next flint]

of damn i guess i messed it up. first time using paste bin and they flagged it because of the nsfw terms

Alright i switched to rentry
https://rentry.org/SD_Civit_Top_Data

[subtle osprey]

hey quick question about promts, i have seen a few styles of setting emphasis now and would like to know whats the difference between them: (girl with blue eyes:1.1) | girl with (blue eyes:1.1) | girl with (blue:1.1) eyes | (girl, blue eyes) | (girl:1.1), (blue eyes:1.1) | (girl), (blue eyes)

[next flint]

Holy shit it actually makes a decent result

[next flint]

ill make some examples

there should be slight differences

it will be a little bit but ill upload the grid. should show the differences. granted i dont have the base prompt/main variable so i went with nothing for it

[subtle osprey]

i looked at some workflows and was just wondering why, no specific goal

[next flint]

its a lot of experimentation. i think its like youtube's algorithm where you dont exactly know why and how it works but through trial and error you get the result you want

i bet there is someone who understands the difference but i feel the best way would be to just give a visual example

[subtle osprey]

alright, im gonna try it out - still copying alot because just started 3 days ago

[next flint]

ill send the matrix in a few

[obtuse torrent]

@subtle osprey you need search for emphasis sintaxis in SD ..

[next flint]

for auto it just explains basics of using brackets

[subtle osprey]

on the wiki there isnt an example with multiple words

[next flint]

yeah

[obtuse torrent]

I m not expert too,

:1.1 <--- how you call it?

[subtle osprey]

emphasis

it the same as ()

[next flint]

I’m uploading it rn but it’s 30mb and I have bad internet

[obtuse torrent]

well emphasis in a bracket?... apply to all token in the bracket ..

[next flint]

They seem to all be rly similar

You probably will notice slight changes tho

[obtuse torrent]

()= 1.1 (())= 1.1x1.1 (not 1.2 i think not sure)

[subtle osprey]

alr sec

[next flint]

bad computer moment

[obtuse torrent]

:S

[subtle osprey]

hmmmm

with the comma its more zoomed in but other than that now much difference

[next flint]

wdym

[subtle osprey]

i cant post the picture what i mean because bot deleted it, but only the gens with a single promt of girl blue eyes are relevant

[obtuse torrent]

@subtle osprey adding the word in the prompt is 100% of weight (in the specific model) a emphasis is a 10% more (if the weight in the model is very low ..10 or 100% more no is nothing) ..and you need add the effect the all tokens in the prompt (for a relative weight, i am speculating here), the position in the prompt (speculating again to see when the steps pick the token), etc, etc ..

[subtle osprey]

but thanks, that was what im looking for

[next flint]

So to answer the original question, there isnt rly a noticeable variation depending on where you put the brackets at least in the example i made

[subtle osprey]

yes thanks for the help 👍

[next flint]

it seems like with the absurdity of 1000 negative tokens every image is surprisingly good

i hope this counts as sfw

[subtle osprey]

i was generating batches of 4 of a simple girl but with a big negative promt to test out the models, they all got the same face because of that

[next flint]

i just added lora:easynegative:1, easynegative, ng_deepnegative_v1_75t, verybadimagenegative_v1.3,(nsfw:2) before the top 1000 and it produces suprisingly sfw things

wdym?

like take the image?

sure

you do you idrc

ill send workflow if you want

https://rentry.org/h5aqt heres the workflow for that one

[obtuse torrent]

@next flint sorry i not understand English and i am very slow using translators.. but yes thx ..

[next flint]

all good

[upbeat minnow]

Im trying to produce stuff on certain models and I keep getting this wierd output, sorry new to this type of AI so sorry if its an easy solution

[next flint]

can you give a little more detail and send workflow?

[upbeat minnow]

Using Stable Diffusion UI, loading in models, when they generate I can see they're working untill right at the end when they get wierd and red

On a laptop with 3070 RTX

Unsure what that means

Ah ok, im at this atm

[next flint]

Can you send a pic with the prompt and negative prompt

[upbeat minnow]

Ah I kinda gone from that sorry, might not be much help there

Don't worry aboutit, seems to just be a one model problem so idm and seems to be working now

Thanks for the help though

[next flint]

Ok

[next flint]

Can somone try the workflow and prompts in the link and give opinons. its just a collection of the top 1000 and its been working surpisingly well for me
https://rentry.org/SD_Civit_Top_Data

[rigid sorrel]

Hey there, I'm also pretty new to this and would need some help to get into the right direction. I'm trying to generate some wallpapers for my living room and want them to be in the style of Masashi Wakuis art.
I've tried to use some image2text-sites and the Interrogate CLIP function under the img2img-tab in A1111 WebUI and came along with the following prompt: street at night with neon signs on the buildings, cyberpunk art by Liam Wong, ominous vibe. According to https://stable-diffusion-art.com I've used this well-known negative prompt ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, bad anatomy, watermark, signature, cut off, low contrast, underexposed, overexposed, bad art, beginner, amateur, distorted face on v2.1 768 model. I'm using the DPM++ 2M Karras sampler on 20 steps, with a CFG Scale of 7. The images are generated at 1024x682. I heard that using other resolutions can lead to unexpected generations, but so far it looks fine.
I'm already getting some generations that lead into the direction I want them, but I'm not able to hit the style I want.
I want to have a mix out of the four ones I have already generated, the vibe and style Shappie generated and the smooth cozyness from Masashi Wakui.
Thanks in advance!

[next flint]

would you say this is close to his style

Probably some need for finetuning but for the prompt you can just put "Masashi Wakui". They are prominent enough that their style and photography is already known by the v1.5 SD model

[silver valley]

something like this?

[next flint]

probably varies from model to model of what youll get but just use his name

[rigid sorrel]

thanks guys!
yeah the second one of @next flint is pretty close to the art style I'm searching for, just a bit more of this bright neon flashing like in your first one. At least in my tries I wasn't able to produce the results you already have. But using SD1.5 I was able to get a bit closer than in 2.1. Unfourtently in 1.5 the structure looks more like copy pasted to fill the area.
the one from @silver valley has the simple calm structure but is missing the I would call washed out art style of Masashi Wakui
I also was trying to use Masashi Wakui, but in my tries it hasn't performed that much like cyberpunk art by Liam Wong suggested by these image2text-sites

[torpid surge]

hey guys how do i make my image good quality?

so its more sharp and detailed

[orchid saffron]

I think just type that sharp, high quality or 8k as a prompt not sure though

[torpid surge]

hmm okay

what the fuck

[orchid saffron]

how are u using stable ? on a website or what ?

[torpid surge]

on my ipod rn

[empty sluice]

https://prompthero.com/prompt/38a73eb8142

[true brook]

does anyone know if you can retrain on top an existing embedding?

[silver valley]

I heard you can train further if you load the trained embedding as source for the new one

But idk how or if it works that good.

[blissful sapphire]

Does someone how what numbers in prompt mean? For example (legs:1.2, long legs:1.2, slim legs, high heels:1.3, white socks:1.4,) Is that some prompt magic one should know?

[obtuse torrent]

@blissful sapphire think of the prompt as a request in which all your words have a weight, then SD fulfills your request but seems to ignore some of your words, that's where those numbers come in that basically increase the weight of your words

So if when you ask for blue hair he gives you hair of another color you increase the weight of your request to blue hair: 1.2, this is asking for blue hair with 20% more force

[blissful sapphire]

ah ok thanks very much!

[obtuse torrent]

@vital fern in local you can..

[vital fern]

local?

[obtuse torrent]

in your desktop or laptop no exist censorship @vital fern

[shell cosmos]

how to generate an image here

[next flint]

What are you using. Is it a locally run ui and is it running something other than the standard model?

Use auto1111 and use a model from civitai

[gleaming harbor]

Please read the #✍🏼｜rules-and-tos and rule #4 specifically. Your question is inappropriate for this server.

But to answer it, you need the local install. You can't generate nsfw stuff online.

[frigid kindle]

what does (()) use for in Negative section? Like if I put ((Close up)) in negative, what happen then?

[next flint]

Weighting is the same concept

[icy wedge]

no matter what I do with inpaint, I just cannot remove this one specific article of clothing from an image, why? am I supposed to be editing the prompt?

[rigid sorrel]

try to use fill under Masked content.
You can also try to put Mask blur or Denoising strength higher

[rigid sorrel]

Do you mind telling me which settings you have used? 🙂

[blissful sapphire]

How do I can get the prompt and settings from a pictures I created back into stable diffusion? If that is possible

[rigid sorrel]

you can use the PNG Info tab in A1111 WebUI

[blissful sapphire]

Thanks!

[next flint]

I did 700x400, dmp++ 2s a Karras, and 30 steps

[rigid sorrel]

Thanks, which prompts do you use?

[next flint]

All I put was “Masashi wakui”. The model knows the style and can replicate it

[next flint]

There is a also a great extension called image browser (https://github.com/AlUlkesh/stable-diffusion-webui-images-browser.git). it allows you to view all the different types of images made (with workflow included) and allows you to search using parts of your workflow

[next flint]

These are what you would want to have as a starting point. There are probably some better settings but ive found these work well

[sacred orchid]

Weird question but has anyone an idea how to describe the feet of a duck/sea gull to stable diffusion?

[rigid sorrel]

@next flint oh wow unexpected simply. will try it thank you!

[rigid sorrel]

seems like the 2.1 model doesn't know her style, only the 1.5 models
is there some known best case szenarios when to use which model, or is it just trial and error?

[acoustic chasm]

Hello everyone. I'm new at generating images, and now I ran into a problem that I can't solve by myself. I'm using DivineEleganceMix V4, which is made by my friend and looks awesome to me, but there's one thing that really pisses me off: this model designed to generate anime characters with realistic and detailed backgrounds of all sorts, but for me it generates awesome detailed background only if there's no anime girl in prompt. If there is, then background tends to simplify drastically, sometimes to the point of forest trees become plain broccoli, the grass is just flat green color, etc. I hate that. I want my anime girl to live in a detailed and colorful world, not low budget anime title. For sampling method I'm using Euler A, 30 steps, upscaling is Fatal_anime_50000, denoise 0.21, Clip Skip 2. Here's my prompts that I'm using for basic image improvement, exactly as recommended by the model's author:

Positive:
(Masterpiece, Best Quality, High Quality, Highres:1.4), Detailed, (Extremely Detailed:1.2), Ambient Soft Lighting, 4K, (Extremely Detailed Eyes:1.2),

Negative:
EasyNegative, Bad-Hands-3, (Low Quality, Worst Quality, Lowres:1.4), (Blurry, Blurry Background, Depth of Field, Bokeh, DOF, Fog, Bloom:1.4),

I'm also including one image without anime character, and one with. Thank you in advance.

My only guess is the more prompts there is, the more simplified image becomes. But I doubt it, because I'm getting simplified backgrounds even if I'm not going past 75 tokens. I'm usually sitting at around 100-120 though.

[acoustic chasm]

I'm at 512x768 and x2 Upscale now. When I tried to increase numbers, I got Out Of Memory error (RTX 3070).

[young bear]

if you dont mind changing models, that will help too

[acoustic chasm]

The thing is that my friend (model author) almost never gets the same problems, for him it's 90% of the time both the anime character, and extremely detailed background.

[young bear]

maybe try recreating some of his

prompts: 8k wallpaper, beautiful scenery, perfect lighting, nature?

[acoustic chasm]

I'll try those, thanks. Friend tried to use my prompts and figured out that the problem is mostly in Ganyu prompt, because it's only recognized as danbooru tag "ganyu /(genshin impact)/. As far as I know, the SD reads such prompts twice, as a complete prompt, and as a separate ones, so it makes goofy ahh background because of "genshin impact" in it. It even tried to replicate official game font in the corner of a picture one time, probably because official artwork is also included in data set, but I'm not sure.

[young bear]

i understand, its all genshin's fault

[craggy dove]

does anyone have a good negative prompt to keep pictures of a characters face from appearing as a graphic on their tshirt?

[obtuse torrent]

@craggy dove add tag, solo_focus, put a clip "about T-shirt whit plain color", negative: duplicate, cloned face

[fallen pulsar]

Hello

[obtuse torrent]

hi

[acoustic chasm]

I figured it out I think. I put "genshin impact" in negatives, I've changed the sampler to DPM Adaptive (very slow, but detail retrieval is unmatched), upscaler to 4x Ultra Sharp, and here's how it works now.

[craggy dove]

What do you mean by "put a clip" ? sounds useful

[obtuse torrent]

Clip is a normal descriptive phrase, (in opposition to DeepBooru tagging) you can see it drooping a image here (in Img2Img tab)

[next flint]

Most people like the 1.5 model because it has a lot better generations and freedom of generations

[dry void]

what prompts do yall use to avoid cropped faces like these?

[obtuse torrent]

PP= wide shot, full body ..... NP= cropped, out of frame @dry void

or add a description about the face or head of the creatures..

[dry void]

Thanks will try. Already have "cropped" in NP

[tired vigil]

Hello, can anyone help me fix this photo and make it more realistic and also fix the faces, hands and body of the people in the image. Any suggested prompts and stable diffusion settings or tools? Thank you.

[smoky gust]

#1072220168534642768 猫

[crimson patio]

Gentlemen and ladies, do you have a problem with putting a firearm in a character's hand? I try to put a "gun with sound suppressor" and there is never any weapon, just some thing I don't even know what to call.

[obtuse torrent]

@crimson patio control net

[next flint]

try and upscale it by putting it into img2img with only the negative prompt and a denoising strength less than 0.3 with the upscaling script enabled. after that its just inpainting

[next flint]

there is an h&k lora

[crimson patio]

I think ControlNet is only for poses or am I wrong? Is it also suitable for items?

[crimson patio]

Thanks, will check.

[obtuse torrent]

@crimson patio exist a especific for hands .. but IDK more .. but exist

[next flint]

https://civitai.com/models/12519/handk-hk416-lora

[crimson patio]

Yeah, I see but thats only this gun. I need small pistol with silencer. :c

[next flint]

maybe find an image of one you like and then put that into control net. is it for someone holding it or just the gun

[obtuse torrent]

@crimson patio Hand Pose Estimation .. https://github.com/Hzzone/pytorch-openpose#hand-pose-estimation

[spare whale]

I'm trying to get into Stable Difussion image generation, but I'm really struggling, any help? Both for handling models and prompting, I'm also finding it very difficult to do faces and hands well... :S

[next flint]

Could you provide one of your outputs for feedback or be a bit more specific?

i struggled at first in a lot of ways too but its hard to give general advice when i dont know where you are at

[spare whale]

Can I talk you privately?

[next flint]

sure

[next flint]

That was just one of the first images i found on google and threw into control net at about 1 guidance

[next flint]

Is that an extension cuz that concept seems fun to experiment with

[obtuse torrent]

and if you have a pic to guide Img2Img some times do the work too..@crimson patio

[crimson patio]

I think my skills in SD are not so great to deal with it so easily. I've only used ControlNet for poses before, I need to see what and how. In img2img I have the impression that the results are hardly ever satisfactory when I add something completely new to the photo.

But thanks for help guys.

[next flint]

is the pytorch openpose the same one built into the control net settings on the ui for auto1111?

[obtuse torrent]

@next flint i only search for it, sure are more models in control net to hand...

[next flint]

Yeah it takes a while to get used to working with SD

[next flint]

i need to finetune the controlnet settings and prompt but i was able to get this with a basic google image

[obtuse torrent]

I m no a expert ..https://www.reddit.com/r/StableDiffusion/comments/11gnkyt/hands_library_for_controlnet/ @next flint check..

[next flint]

Damn that is a gold mine

but alas i am too lazy to make my own skeletal structures for controlnet

i just downloaded most of the pose libraries from civit

[obtuse torrent]

Running it in a old 2060 .. a gen make me happy 😄 ..@next flint only reading about other tools...

[next flint]

OMG SOMEONE LIKE ME

im running on a 2060s

Idk if this should go in here but I feel like I ought to share it

Rev is on some different shit

[torpid harbor]

Anyone know how to keep freckles from being applied to clothes? if I include the word freckles, even (face freckles), it makes the clothes have a repeating dot pattern

[next flint]

Try adding freckles on clothes to negative prompt

[sick ruin]

@next flint how fast is the image generation now with the 2060S? i'm curious because i'm running a GTX 1660Ti and have to wait around 1 to 1,5 minute for a 512x512 image with 20 - 30 steps (you can send me an PM)

[next flint]

yeah. im in the middle of some other generations for someone but once im done ill send some screenshots of the console

[sick ruin]

👍 you're the best ^^

[sullen orbit]

hey guys, does anyone have any camera prompts like ''200mm 1.4f macro shot'' ?

[silver valley]

do you use --xformers --medvram ?

[next flint]

what does --medvram do

[next flint]

Ask gpt

[hoary quest]

makes it use less vram but runs slower, basically.

[next flint]

damn maybe i should be using that

[hoary quest]

are you using xformers?

im on 3070ti and after enabling xformers my gen speed and vram efficiency almost doubled

[next flint]

yeah

i only have 8gb vram,

[hoary quest]

same

it's enough for me though

[next flint]

I was able to have this gif have some continuity but im wondering how to keep it a bit more of the same without decreasing denoising strength

[merry wasp]

Anyone know of a good prompt to get a 3d model sheet as in front back and side?

[next flint]

charturner is good

there is a textual inversion and lora

[merry wasp]

Sorry i'm a bit out of the loop, what dose that second part mean?

[next flint]

textual inversion is a thing that bascially adds words into the prompt from a file. Its a set collection you can add. Loras are almost like mini models that allow you to mix styles

[merry wasp]

So I can find a textual inversion and lora for a modelsheet online then or will I need to train one?

Sorry my shift starts soon, I'll have to put the phone down for 2 hours

[next flint]

https://civitai.com/models/3036/charturner-character-turnaround-helper-for-15-and-21

[tired vigil]

Great thanks. I will try this.

[atomic flume]

How can I get this image to be in sharper focus ?

medium format portrait, beautiful face, a 20 year old black woman sitting on a chair, red dress, large red red hat, warm natural light coming in through a near by window, epic realistic, faded, (((hdr))), hyperdetailed, cinematic, warm lights, intricate details, muscle diffused lighting, centered composition 85 mm f 1. 8 lens with hasselblad photo by Annie Leibovitz and Steve McCurry in the style film Nizouveau watercolor inktobernetic illustration sharp focus face symmetry volumetric lights highly detailed Cinema, photo realistic, ultra details, natural light, light background, photo, Studio lighting, sharp focus on eyes, highly detailed green eyes, small smile,

Negative prompt: surreal, nrealfixer, nfixer, cgi, fake, render, painting, illustration, out of focus, blurry, bad eyes, freckles,

Steps: 70, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 3675278697, Size: 816x544, Model hash: c10a124704, Model: aZovyaRPGArtistTools_sd21768V1, Seed resize from: 712x712, Denoising strength: 0.55, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires steps: 40, Hires upscaler: 4x-UltraSharp

[fluid mason]

You used a model that is sort of designed more for illustration, but if you like this exact image, you can load a photorealistic model like Realism Engine or PRMJ and then do image2image with the same prompt and play with the image strength, further, you could use the "Sharpen Image" option in Upscayl and then downscale it back to your desired resolution

[atomic flume]

I got the same results from

medium format portrait, a 20 year old woman sitting on a chair, red dress, large red hat, 50mm, strobe light, natural light coming in through a near by window, epic realistic, faded, (((hdr))), hyperdetailed, cinematic, warm lights, intricate details, muscle diffused lighting, centered composition 85 mm f 1. 8 lens with hasselblad photo by Annie Leibovitz and Steve McCurry in the style film Nizouveau watercolor inktobernetic illustration sharp focus face symmetry volumetric lights highly detailed Cinema, photo realistic, ultra details, natural light, light background, photo, Studio lighting,
Negative prompt
Negative prompt: surreal, nrealfixer, nfixer, cgi, fake, render, painting, illustration
Steps: 70, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1628774906, Size: 800x800, Model hash: 7da43996bb, Model: rmadaMergeSD21768_v60, Seed resize from: 712x712, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires upscaler: Latent

[fluid mason]

some of those prompt terrms might also hit low contrast photography data...like medium format and annie leibovitz....all of her work is like muted low contrast

and medium format might be hitting "film grain" style source

[atomic flume]

ah ok

what prompt would suggest ?

[fluid mason]

you could keep trying prompting, but to me, its hardest to get the composition you want, once you have that, use image2image with other models to perfect the contrast and sharpness

for example i just ran that p rompt and similar settings image2image with realismengine

got this

then i used Upscayl and chose the Sharpen Image type of upscale

personally im liking the contrast from prmj

is this what you are looking for:

@atomic flume ^^^

[atomic flume]

yes

what is Upscayl >?

[fluid mason]

Model: realisticVisionV20_v20

Prompt:
photorealistic, photograph, stunning portrait of a 30yo african woman sitting next to window wearing a red dress and a large red hat, warm natural light, backlighting, slender, petite, skinny, seductive, masterpiece, 8k, hdr, cinematic

Negative Prompt:
picture frame, ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, bad anatomy, watermark, signature, cut off, low contrast, bad art, beginner, amateur, distorted face, blurry, draft

Seed: 1891299913

Noise Threshold: 0
Perlin Noise: 0
Sampler: k_euler_a
Steps: 30
CFG scale: 7.5
Width: 768
Height: 960

Upscayl is an app you can download locally and choose from different x4 and x 8 upscaling algorithms

its AI based upscale, some of the algorithms are similar to the ones in invoke AI

[fluid mason]

that the prompt ^^^ above and here is the model https://civitai.com/models/4201/realistic-vision-v20

[south solar]

Anyone know how to get an active camouflage effect? Like a partially transparent person showing the background behind. I would look at models as well as just prompting, but I don't really know where to start because it isn't something I have experimented with before.

[lone badge]

https://civitai.com/models/2386/double-exposure-for-sd-2x maybe ?

complicated concept but I think double exposure is close to what you describe here

[south solar]

Thanks. I will give it a try. Got a rad idea for the PoW that I hope I can make work.

[hearty plank]

With SD 2.1 base, I want to generate images without any eyes nor face. Negative prompting of "faces, eyes" didn't work for me, when the prompt suggests an animal. Help please

[vague flint]

it keeps giving me nude images even when i put nsfw on negative prompt. any ideas?

[lone badge]

you may have too many NSFW tokens inducing in the positive prompt. Add more into negative, like "cleavage, boobs, naked, nude", ....

[lone badge]

some kinds of monster ? without eyes and face animals are monstrous, so I would use this kind of tokens, "disformed, monstrous, blank face" and add "eyes, face" in negative, but this is a hard concept to get I think

[vague flint]

Assassin creed, illuminati and evil input

[silver valley]

Maybe try:
woman sitting at a fountain, wearing a blue evening dress, black hair, big earrings, silky dress, black eyeliner, masterpiece, 4k, high quality,

[silver valley]

A realistic one, like realistic vision or deliberate

Maybe Dreamshaper too

[wicked kettle]

Ok ,thanks
it's actually a game character , maybe it's done img to img

[silver valley]

Could be a lora then

[void patio]

how to prompt for a camera in the grave kind of photo?
pov in the casket kind of thing...

gpt called it worm's view but i could get it

[fluid mason]

you may need to think of how a real world photograph/art would be tagged with that...maybe "looking out from the casket"

[vernal raft]

My suggestion would be find a similar photo online and feed it into the stabled diffusion and it can return a group of keywords

Kinda reverse engineering

[void patio]

into interogator? that's a good idea

[vernal raft]

It’s amazing how these AIs communicate with each other

I feed some of my generated images back to get the keywords and some of the key words phrases are exactly the same as what I wrote

[void patio]

that's nice

what clip model you use?

[vernal raft]

What’s a clip model? 😂

[short bay]

hello ! I've been testing Stable Diffusion for a few days and it works quite well. However, when I try a scene with 2 characters, I end up with more characters than the number indicated and despite a negative prompt, I can't correct. Does anyone have any advice on this sort of thing ? (soryy for my very bad english, i am french)

[vernal raft]

Now I know what you mean. I’m using stable diffusion

[fluid mason]

sometimes this happens if you try higher resolutions than the model. IF you reduce the resolution to get the composition you want, then you can feed the image back to img2img with a higher res and refine it

[short bay]

ok, I will try to reduce the resolution

[short bay]

After some tests, I have much less errors, thank you

Are there more suitable models if the resolution is increased (above 1024)?

[fluid mason]

as far as i am aware most models top out at 768x768 and this means you can generate up to 1024x768 (or 768x1024) without too much trouble. If you want higher res, first generate at a lower res with text2image, then take the output of that and use image2image selecting a much higher resolution. As long as it has the composition already defined usually you can push it higher

another option is to just take a 1024x768 image and use an upscaling tool like Upscayl which is AI based and produces pretty great results

[short bay]

tank's for tips 🙂

[marsh vessel]

Hello! How can I make something like this? https://media.discordapp.net/attachments/1084938009843093674/1095763066437775461/image.png

I've been advised to use abyssorangemix3 to do it but I don't know the prompts that can get me something close to this

[daring scaffold]

As I looked out from the balcony of my cozy little cottage, I couldn't help but marvel at the picturesque view in front of me. The nocturnal beauty of the garden was nothing short of enchanting, with its lush greenery illuminated by soft flickers of light emanating from the delicate fairy lights hung around it.

But what truly made the scene come alive were the two figures sitting on the garden bench - a young man and woman, lost in their own world. They were holding hands, their fingers interlaced, and their faces were mere inches apart, lost in a deep conversation, their eyes twinkling with delight in the soft, warm glow of the night.

As I stood there, soaking in the romantic ambience, I couldn't help but feel a wave of envy wash over me. How lucky were they to have found each other amidst the chaos of life, to have found solace and comfort in each other's arms. I wished I had someone next to me too, to share this magical moment with.

But for now, I was content simply savoring the beauty before me, letting the scene etch itself into my memory, so I could relive it whenever I wanted. The gentle rustle of leaves, the sweet fragrance of flowers, the faint sound of their whispers - it was a moment that seemed to transcend time, a moment that overflowed with love, contentment, and an untainted sense of joy.

As I finally tore my gaze away, I knew I would never quite forget this magical night, and the beautiful, romantic view that had captured my heart forever.

[obtuse torrent]

@marsh vessel loras ..

[marsh vessel]

what?

[silver valley]

Hey jodio, these images are done mostly with loras.
Loras are little files you can load on top of the model that are trained to get an specific character or style

So someone trained the ai on images of them and made a lora

If the model knows the characters you can try it without using loras too

Like:
1girl, hatsune miku, guitar, highres
Would be enough to get hatsune miku images

But thats because she is a well known character

[marsh vessel]

Oh ok

And the negative prompts are pretty much the same everytime?

[silver valley]

Yea pretty much, in this channel there is an negative prompt pinned you can try

[tawny wraith]

I'm having trouble prompting mass army in wide shot scene; is there any good prompts or more accurate prompts for me to make it happen?

my current prompts are:
(masterpiece, best quality:1.2), scenery, (wide shot:1.2), lord of the ring, epic, (extremely populated:1.1), dawn light, mountain, light shinning through clouds

model: dreamshaperv4

I've tried army, mass army, epic war, but to no avail

what my aim essentially is to generate a mass war between 2 races; but so far the best i can achieve is just one race amassing a large group of armies in a scene, not a war

[slender pasture]

Its to my understanding that you can train stable diffusion to remember faces n stuff?

Is this true?

can train it to remember themes and faces and other things

[silver valley]

Yes for that you need a tool to train.

[silver valley]

Exactly, the best way are loras for that

[slender pasture]

huh ok thats a model?

i literally just started this yesterday, tho ive been using holara and novel, so i understand (promopts) I leave the technical computer stuff to my husband.

[silver valley]

If you have 6-8gb vram you can train a lora.
If you have 12 or more gb vram you can train a model with Dreambooth

A lora is a little file that was trained on an specific style or character and goes on top of a model

[tired vigil]

Is this the appropriate place to ask questions about how to generate better output images?

excellent -- I am trying to generate a seamless image such as this to use as texture on a fireplace in a 3d model I am building -- I cannot seem to get it to generate the mortar lines between the blocks -- mortar, grout, concrete dont seem to work -- any suggestions?

[slender pasture]

can you use images made from other places like noval or holara? for theme, then train it to remember a face from a stable gen?

and how do you make a lora or whatever owo

if theres a page you can point me to with all the info thats fine ❤️

[silver valley]

You can use every Image. You even can train a model or lora of your own face to generate images from you in like a Knight armor etc

@slender pasture this is a guide to get the Programm running where you can train stuff: its called kohya
https://m.youtube.com/watch?v=70H03cv57-o

[slender pasture]

THANK YOU!

youre very kind and helpful n_n

[tired vigil]

THis is my current prompt -- https://gademasonrylandscaping.com/wp-content/uploads/2021/02/stones-770264_1920-1024x683.jpg Stone wall 4k Texture seamless

[silver valley]

No problem, if your very new to sd i wont recommend starting woth training instead get familiar with the sd webui, get some loras or models and try them out 🙂

[slender pasture]

i have dream uh

dream S lol

[tired vigil]

thats my output

[silver valley]

Like Mortar of a brick wall?

[tired vigil]

yes

[silver valley]

Hmm maybe try cobblestone brick wall

Im not at pc rn but tomorrow ill try

[tired vigil]

cool -- i appreciate it

[silver valley]

? What do you mean sry

[slender pasture]

dream shaper

sorry 🙂 I couldnt remember the name

[silver valley]

Ahhh xD

Very good model

[slender pasture]

so i can download other's loras from civitai.com?

agreed- im liking it

[silver valley]

Yes they go into the models/Lora folder.
Then you can select them in the webui by clicking the third button under the Generate button and select lora

[slender pasture]

ok so what is a textural inversion then?

[hard elm]

If you go to #1072220168534642768 and #1080946152318443610 there's actually a lot of information on SD, how things work, etc.

There's also several community guides as well

[slender pasture]

coooli ty

[upbeat ledge]

is it possible to get the character to point at "you"

holding out hands towards you, etc

stuff like that

like directly in the middle of the screen i guess, idk how to word it

ia?

be careful?

wut

oh, well generally using that prompt doesnt work

anyways

cant seem to do it, i will keep trying tho

thanks for help

[tawny sierra]

Hey all, I love how the community works together to make great things now and in the future. I have a question on a project I'm working on. I have people in front of a greenscreen, where they take a picture with a webcam. I want to have a few options in a selectlist on a page where they can choose a background. Kind of like the virtual backgrounds in Zoom and Teams. But what should I use as a prompt or as a setting to generate backgrounds like "hills", "city" or "seascape"? Whatever I try with img2img and Controlnet Canny/Depth/Segmentation just keeps giving me back green backgrounds...

[honest raptor]

is there any prompt what i could make little more colorful instead bit dull ?

[lone badge]

"rainbowshift" has always worked great

[cunning stag]

hi, how to wrote good prompt if i want a photorealistic photo? i put the negative prompt. Model is a Delibrate V2, sampling method DPM++ SDE Karras, cfg scale 1, sampling steps 100.
My generated images looks like shit from dalle mini

realistic santa claus riding on his chopper, snowy background, realistic, 4k, ultrahd, photorealistic
my prompts looks like this one

and looks like this

[wide phoenix]

can someone please help me with stable diffusion? for some reason it wont let me launch it

[cunning stag]

what is wrote in the cmd?

[honest raptor]

is there any guide how i can force prompt like white_dress ? i cant produce that

[tired vigil]

hi folks - -this is the prompt, and init image that I used to generate the image on the right -- im trying to get something more like the init image (Showing the mortar lines) can someone suggest a better prompt? (VERY new at this)

[next flint]

Ig you could take the image, outline/mask subjects, invert mask, and then do a background

[next flint]

there are a few ways to do that. either increase weight of white_dress, inpaint the dress to be a different color, sketch the dress to be a different color, use instruct img2img to change what they are wearing

[honest raptor]

ohh i found issues, now its working

[next flint]

what was the problem

[next flint]

Positive prompt: Santa riding a motorcycle, snowy forest background, photorealistic, highly detailed
Negative prompt: easynegative, ng_deepnegative_v1_75t, verybadimagenegative_v1.3, badhandv4, badv5, By bad artist -neg, verybadimagenegative_v1.3
Try using DPM++ 2S a karras or Euler a

[grave chasm]

Santa vs dragon

[muted crow]

Elephant walking on a Red ball

[next flint]

que

[cunning stag]

wow nice

[next flint]

If you want a better image just use loras and textual inversions in the negative space

[ancient cove]

Temple #1029055412764422214

[true agate]

Hello

[undone rivet]

Hello everyone

Do you know if there is a guide, reference or tutorial, about the right way to make the "structure" of a prompt, for the image generation, without being complex or excessive, and besides that having the right weight for each part....?!

For example to describe the background, clothes, expression, poses, lighting etc.

[next flint]

https://stable-diffusion-art.com/prompt-guide/

[undone rivet]

Thanks a lot 🙌🏻

[tawny sierra]

Thanks. I played around with this and found some proper results with the Inpaint Upload possibilities.

[atomic flume]

I've seem people doing this on a few prompts

(James C. Christensen:1.2|Jeremy Lipking:1.1)

I'm a little confused by what that is doing

[obtuse torrent]

@atomic flume names of artist, illustrators, photographs, for a Art style orientation ..

[smoky ore]

Guys, is this using a Lora to put the this kinda specific dress?
I dunno how people dress up their character using a very specific yet detailed dress

[left flicker]

Does anyone know any good prompt tips to improve eyes and facial features? I keep getting fairly "indistinguishable" features with this prompt:

Negative prompt: [(ugly painting):easynegative:,0.9], (bad_image:0.7), (low quality, worst quality:1.4), (bad anatomy), bad proportions, bad face, extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000, short hair, military honors, epaulettes, billowing clothing, billowing coat, umbrella, exposed chest, dramatic pose, elf ears, modern clothing, business attire, wings, beard, pale hair, facial hair, vest, multiple people, white hair, old men, colored hair, pointed ears, knife ears, large ears, black and white, monochrome, magic, wizard staff, glasses, halo, horse, horses, asymmetrical face

Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 354110623, Size: 512x512, Model hash: 4d651c7638, Model: expmixLine_v20```

[crisp oasis]

microscope sitting on a desk in a lab, close up, watercolor

[stone igloo]

@crisp oasis

FAQ: How do I generate images? Is there a bot on the server?

Currently, there is no bot on the server that generates images. However, there are plenty of other ways such as the official https://beta.dreamstudio.ai/ website or by running Stable Diffusion locally using your own hardware! Check out #1080946152318443610 for more details! You can also stop by #1025467151206854736 for any issues you experience while using DreamStudio or #🤝｜tech-support for any problems you encounter while installing it locally!

[silver valley]

hey your missing the vae for that model, anime models need that for color correction

you can also try to set the resolution to 512x768 for a portrait few to get better faces

[left flicker]

How do I get/apply the vae?

[empty tundra]

I want to generate fish/game icons but I can't for the life of me figure out a good prompt which does. Here are two example images of how I wish they'd look. Does anyone have any tips? What type of keywords should I put in my prompt? Would I need to make a custom model(I do know how dreambooth works) for them to actually look good? Please help!

[left flicker]

I admit I'm struggling to really understand what a vae is, or rather, how to put it in

[empty tundra]

Download the vae, then put it in .\stable-diffusion-webui\models\VAE

[night heart]

I am trying to create Melina from Elden Ring. She has a black scar on her left eye and I don't know what prompt to use for that. I am using img2img and ControlNet with depth and canny models (Guess mode on both).
I drew this scar on the eye in Photoshop, but the AI still ignore it
There's 5 photos

[left flicker]

so, my model is expmix_line, but the huggingface doesn't appear to have a vae?

[silver valley]

a vae is needed for color and detail correction,
most models have it already included, but most anime models not

expmix is a mix, so it has a model in it that needs a vae

so the vae of the model will work with exp mix too

[left flicker]

Looks like I may be in trouble, then. It doesn't say what the model is, and the civitai link is dead/404'd

[silver valley]

ah it shows exactly which vae you need:
https://civitai.com/models/35207?modelVersionId=41560

select this then press download

[left flicker]

oh! I think that's a different one than what I have

Thank you for this.

[silver valley]

put the kl8-anime2 vae inside the models/vae folder

[left flicker]

to be clear, the vae is a .cpkt file?

[silver valley]

then you can select it in the settings of the webui,
to add an ez dropdown you have to go into settings -> User Interface -> Quicksettings and add , sd_vae, to it, then hit apply and restart

[silver valley]

mostly not, vae are .vae.pt files but this one is a special .ckpt one trained on the official sd vae

[next flint]

what image are you using for canny

[next flint]

also https://civitai.com/models/6863/melina-elden-ring-lora

[night heart]

from img2img

[night heart]

this is a .. model?

like this?

[silver valley]

nope thats a lora

an addition that goes on top of a model to give a specific char or style

[next flint]

its a lora

can you send the image you put into canny

[sullen orbit]

Guys, what prompts do I use for the eye to look like this, in a happy expression?

[night heart]

Tell me a site where you can find a lot of these lora?

[silver valley]

civitai.com

[left flicker]

That definitely made a difference, thank you! My faces are still a tad weird, though

[silver valley]

the nearer the face to the viewer the better the faces

[silver valley]

you would need upscaling for that because the face is far away

upscaling fix faces

[left flicker]

Ah, that makes sense. Are there any good prompts for getting just a chest-up portrait of someone?

It seems to be resisting anything that would zoom in on the face

[silver valley]

portrait, close up, close up portrait,

you can give it more weight with (portrait) or portrait:1.2 or if you put it at start

[honest raptor]

should i use prompt pink_hair or pink hair ? or it dont matter ? booruprompt does with space only

[silver valley]

i would use pink hair,

[honest raptor]

so it dont matter much ?

[silver valley]

booru tag autocomplete extension uses also only spaces

[honest raptor]

ohhh alright

good extension

[left flicker]

Brill! Though I've now encountered a peculiar problem in img2img, where no matter what I set the batch size to, it only spits out one image.

[silver valley]

Hmm sounds like a bug, maybe make a Screenshot of it for #🤝｜tech-support

[left flicker]

Yeah, it does seem like a bug, since it shows five images generating, but only shows the result of one. Thanks again for the help!

[next flint]

it normally only outputs 1 image when there is nothing in the prompt box. just put something like a space in there. if that still doesnt work there is a script called run n times

[left flicker]

there is stuff in the prompt box, though!

I was upscaling it with this

[next flint]

take out the positive prompt

if ur upscaling just add a lot of negative prompts

[left flicker]

...I instantly regret doing that. Oh dear

it turned him into a girl

[next flint]

Yeah that sometimes happens

Marshall they arent great but i was able to get these

[next flint]

i mean it does look better

@left flicker when upscaling id recommend troubleshooter history making it like this

[left flicker]

Troubleshooter history?

[next flint]

huh

wdym

[silver valley]

@next flint Your Keyboard had a problem and typed that it seems

Troubleshooter history

[next flint]

im so confused

[next flint]

What do you mean by this

[next flint]

oh

[left flicker]

it was your comment 😅

[next flint]

i have no idea how that got there lol

[cobalt sluice]

Is this the right place to ask about how to create prompts with python?
Trying to figure out how to add negative prompts.
This site shows two textarea inputs.
https://stable-diffusion-art.com/how-to-use-negative-prompts/
GPT-4 wrote.

negative_prompt = "trees, green"

# Combine main prompt and negative prompt
combined_prompt = f"{prompt}; negative prompt: {negative_prompt}"

# Pass the combined_prompt to the function generating the image with Stable Diffusion 2.0
generated_image = stable_diffusion_generate_image_v2(combined_prompt)

However, I get a full image of green trees.

[lone badge]

ok so there are multiple things here. I have not seen this bot yet, is it stable diffusion ?
You seem to not be passing the negative prompt for it correctly, it's using green as positive prompt, thus the color
You should ask the admins on your server, this isn't a bot coming from here, i'm not sure how to use it

[next flint]

what he said

[cobalt sluice]

This one.
https://huggingface.co/eimiss/EimisAnimeDiffusion_1.0v
It gives this as an example.
Positive:a girl, Phoenix girl, fluffy hair, war, a hell on earth, Beautiful and detailed explosion, Cold machine, Fire in eyes, burning, Metal texture, Exquisite cloth, Metal carving, volume, best quality, normal hands, Metal details, Metal scratch, Metal defects, masterpiece, best quality, best quality, illustration, highres, masterpiece, contour deepening, illustration,(beautiful detailed girl),beautiful detailed glow Negative:lowres, bad anatomy, ((bad hands)), text, error, ((missing fingers)), cropped, jpeg artifacts, worst quality, low quality, signature, watermark, blurry, deformed, extra ears, deformed, disfigured, mutation, censored, ((multiple_girls)) Steps: 20, Sampler: DPM++ 2S a, CFG scale: 8, Seed: 4186044705/4186044707, Size: 704x896
I'm confused where to put CFG, Seed, and Steps. The example looks like it's in the prompt.

These essentially my config.
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)

pipe(prompt).images[0]

[lone badge]

this is a model, not a bot you are linking me.
Your params seem ok to me, looking at those, but like I said, I don't know discord bots for this, so asking the one who implemented it on your server seems the best. If you implemented it, checking the doc of the bot itself, or the API you are calling if you are writting the bot yourself
If you are calling a local base install of diffusers, I'm looking at the pipeline documentation right now, and I think you are just passing everything as prompt right now, not as params :
https://github.com/huggingface/diffusers/blob/main/tests/pipelines/pipeline_params.py

[cobalt sluice]

I'm programming it.

Do I add the different prompts to the pipe function? I saw example someone putting in guidance_scale.
pipe(prompt, guidance_scale=7.5).images[0]

I'll research this pipeline function more. It seems to be the key. GPT-4 has been useless helping me with this.
pipe = StableDiffusionPipeline.from_pretrained(model_id, scheduler=scheduler, use_auth_token=True)

[lone badge]

This is how I call my pipe in my personal diffusers tool images = model.pipe( prompts_to_run, height=H, width=W, negative_prompt=negative_prompt, num_images_per_prompt=1, num_inference_steps=config["steps"], guidance_scale=config["cfg"], generator=model.cudaGenerator ).images
https://github.com/Guizmus/DreamboothSimpleUI/blob/main/code/scripts/lib/txt2imgBenchmark.py#L135

this was some months ago, the pipeline interface may have changed a little

I set the pipe on line 26

[cobalt sluice]

Thank you for the context. I just found this.
https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/text2img Is has all the parameters listed. 🙌

[lone badge]

it's what I was still looking out for you

I had it when I made that script I linked

well done finding it :p

[cobalt sluice]

GPT-4 is much more helpful with the documentation passed to it.

``negative_prompt = "lowres, bad anatomy, bad hands, text, error, missing fingers, cropped, jpeg artifacts, worst quality, low quality, signature, watermark, blurry, deformed, extra ears, deformed, disfigured, mutation, censored, multiple girls"
steps = 30
guidance_scale = 20
seed = 4186044705

generator = torch.Generator(device="cuda")
generator.manual_seed(seed)

image = bot.pipe(
prompt=prompt,
num_inference_steps=steps,
guidance_scale=guidance_scale,
negative_prompt=negative_prompt,
generator=generator
).images[0]
``

generator = torch.Generator(device="cuda"): This line creates a random number generator on the GPU. The generator is used for creating deterministic random numbers when generating images. By specifying the device as "cuda", the random number generation will also be performed on the GPU, which can be faster than generating random numbers on the CPU.

[lone badge]

damn it's good

it understood the whole thing and piped it for you

I think it called bot.pipe instead of generator.pipe though

yeah some errors in there

but much closer

[honest raptor]

its possible in prompt specify to generate one thing not multiple ?

[lone badge]

your best bet is to describe more of what is not that 1 thing.
What I mean is, SD will try to fill your picture based on the prompt, so having more things describing the background instead of the character can help on this side, making more background
If you generate big pictures though, the model can tend to do duplication, and there aren't a lot of response to this, mostly generating in lower resolution and upscaling

[honest raptor]

ohh..

[shadow crypt]

Hello, I am new to Stable Diffusion prompts,

Is there a way to have universe/galaxies in a fog in a humanoid shape? (in a vector style or close to it as possible?)

[shadow crypt]

Is this normal? I didn't even ask for text? I just thought it was weird.

[next flint]

yeah sometimes that happens

just add text in the negative prompt

[next flint]

there are a few ways. you can always specify them forming in that shape ( like say "galaxies and nebulae coliding to form a human figure") or you can put something into controlnet (like a stock imae of someone or a standard depth map of a person)

[next flint]

there is always controlnet. You can also specify it to be 1 thing. for example, one of the most common prompts on civit is "1girl". Ive tried probably thousands of images using that and i always get 1 person

[shadow crypt]

Thank You Neo! (also sorry for not responding immediately I was looking at something else)

Also I only have a 3? gb graphics card otherwise I would try to train my own art style. I just don't have a good computer for that.

[next flint]

there is no need to train anything for thaht

ill send you a picture in a second

[honest raptor]

oh damn

[shadow crypt]

I wanted it in a art style but as is looks really cool.

[next flint]

even without controlnet i was able to get a great picture of a cat made out of cosmic materials

[shadow crypt]

Thank you for the prompt idea Neo! It does have a different feel to it.

[next flint]

Ill send an example with and without controlnet in a sec

just taking forever for my computer to calculate sha256 for all my models

First one is without controlnet. They arent good but it gives and idea

i used depth for the controlnet model but you might have better luck with others ones

honestly for this kind of thing anything other than openpose might just make things worse

[shadow crypt]

that looks really cool Neo! Thank you! I have to go but I will look into this more.

[dusky pagoda]

Mistress Mommy sexy in a black latex suite whit a whip above a slave petguy , 8k resolution , ultratexturized realistic scene

[upbeat ledge]

wut? lol

[night heart]

so you use universal lora model or something else?

[next flint]

It was something else. Ill send it in a bit

I think it was Darksushimixmix_brighter

[next flint]

https://civitai.com/models/24779/dark-sushi-mix-mix

[night heart]

wow this looks incredible, thank you

[next flint]

Try both the darker and brighter versions

Good for the Elden ring anime aesthetic

[dark matrix]

Does anyone know the sizing to use for making iPhone wallpapers when generating images?

[next flint]

Depends on the iPhone

An 11 is roughly 828 by 1792

They should all be similar ratios

[dark matrix]

Would you put that in the image size scale?

[next flint]

19.5:9

[dark matrix]

So you would put this size in the prompts?

[next flint]

No that would be in the size part

So instead of 512x512 it would be roughly 270x585 and then upscale that

[dark matrix]

Okay thank you 🙏

[night heart]

I wanna change eyes on character by inpaint. Why it have an another skin color

[next flint]

since your image is is only 736x680 there is probably around ~9000 pixels for their whole face. first you are going to want to upscale it

keep it in img2img. Make everything the same as when you generated it (seed, sampler, sizes, model, cfg, steps). Set denoise strength to 0.1-0.25. Scroll down to the scripts area and select SD upscale. Have tile overlap set to 64 and set the scale factor to how much you want to upscale your image by. Scroll back up and only leave the negative prompt in the textbox and then generate the image.

This will not only upscale how many pixels there are so you can inpaint better but will also improve the quality of the image

In img2img it tries to apply both the positive and negative prompt to every tiling of the image. By only putting the negative prompt you can have it fix all the stuff you dont want without any chance of it adding new things

The first image was the original image
(Negative prompt: easynegative, ng_deepnegative_v1_75t, verybadimagenegative_v1.3, badhandv4, badv5, By bad artist -neg, verybadimagenegative_v1.3
Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 3560508312, Size: 768x768, Model hash: 099e07547a, Model: darkSushiMixMix_brighterPruned)
The second image was just and upscale of the first
(Negative prompt: easynegative, ng_deepnegative_v1_75t, verybadimagenegative_v1.3, badhandv4, badv5, By bad artist -neg, verybadimagenegative_v1.3
Steps: 30, Sampler: Euler a, CFG scale: 7, Seed: 3560508312, Size: 768x768, Model hash: 9aba26abdf, Model: deliberate_v2, Denoising strength: 0.25, SD upscale overlap: 64, SD upscale upscaler: R-ESRGAN 4x+ Anime6B)

[next flint]

Granted i could keep doing this until i get an image around 10000x10000 but this is a good illustration of my point. If you open the image in browser and try to zoom in you will notice it a lot more

[next flint]

For the denoising strength the more you set it to, the more it changes. If its set to 0.25 it will be fixed more than 0.1 but will also change the image a bit. if you want only upscaling keep it closer to 0.1

[magic wharf]

hi guys

how do i emphesize a combined prompt such as [x|y] ?

can i put the entire thing in parantheses?

does (x|y) count as an emphesized combined prompt?

[next flint]

My bad. The second image was darksushiMixMix

[next flint]

probably

[magic wharf]

won't using something like ((([x|y]))) read as ephesize x3 then deEmphesize x1 the combined prompt of x and y ?

is (((x|y))) not better?

[next flint]

idk i havent tried that concept before

[magic wharf]

ya wish there was a wiki for this shit xD

thanks though

anyone with experiance with this concept? help :D?

[next flint]

best thing to do is to just experiment

[magic wharf]

i tried that but unfortunately its unclear, the difference is too minute in a short experimental prompt to actually know how the code reacts

what about using , , ,prompt, , , in a long prompt?

how does it affect the generation?

[honest raptor]

(masterpiece:1) does this counts as 100% ?

[magic wharf]

i think it counts the parantheses as emphesis

so it would be more than 100%

not sure xD

[hasty pumice]

Not sure where is the Positive and Negative Prompt is

[honest raptor]

bc (masterpiece:0.9) aint 90% ?

[hasty pumice]

What happened?

[magic wharf]

i don't think you can give negative prompts in whatever that is

[hasty pumice]

I got negative prompt and positive prompt seriously

Kind of weird

It was more realistic when i used the Negative Prompt though lol

[hard elm]

https://huggingface.co/spaces/stabilityai/stable-diffusion

The demo, right?

[hard elm]

Also, I have guides: https://atypicalconsortium.carrd.co/

Although the bot is no longer running, you can still view my Friendly SD Guide: https://docs.google.com/document/d/1aHJ9RBt_vlCwJQBVUUsb7VghKB-wynv7WGTxm9ozL1k/edit#heading=h.uh133k3c2aq

[hasty pumice]

I realized i was using Stable Diffusion 2.1💀💀

[hard elm]

2.1 is great 🙂

[magic wharf]

doesn't include an answer to any of my questions

[hasty pumice]

2.1 Is Kinda Better

[next flint]

That is a weight of 1

[night heart]

What is the name of this style? Any ideas

https://www.youtube.com/watch?v=6riDJMI-Y8U

[obtuse torrent]

@night heart ?? use all you can see there..... toho animation, jujutsu kaisen, search for illustrator of manga, TV show, search for Lora or TI all ready published in Civitai, etc.

[next flint]

not great but this is what i was able to get just specifying "jujutsu kaisen style"

[obtuse torrent]

@next flint is already no bad ... people need almost make a try .. 😄

[next flint]

wdym no bad

[obtuse torrent]

(I don't think it's bad, it's just that we don't know what he's looking for...@next flint ... manga style .. anime pic, etc ..

[next flint]

ok

i was just giving an example of what happens when just saying the style of something. People often forget or dont know you can do that

[obtuse torrent]

@next flint if you can think of 1 or 2 words ...try it first with that !! 😄

[copper shell]

What is the best way to achieve "similar image"? Like I upload an image of a woman in a modern living room using laptop on a couch with a small yellow dog, and I want the output to be a brand new image but with same elements and sam description. Something like in Midjourney, using "describe" to describe a picture, then using the description as a prompt for a new image. Thanks!

[lone badge]

here, it will be more discussed in #1011634831467221033 , and the current best method is called "controlnet"
example #1011634831467221033 message

[next flint]

Similar in what way

if you have the workflow that would be optimal but otherwise controlnet or just a high denoising strength on img2img

[tropic nexus]

Hey, I am looking to create character sheets that are used for computer games with dreamstudio, but all I can achieve is very far off. Can anyone give me recommendations for a) what model to use and b) what style and c) hints for a good prompt?

I am not too picky, but something like this

[next flint]

a: There aren’t any specialized models for this (that I know of) so go with what you like
b: whatever style your game is in
As for c, there aren’t rly any all around solutions for prompting per say. Your best bet is using the charturn Lora or embedding on civit and/or open pose controlNet

[tropic nexus]

Can I do that on dreamstudio?

[next flint]

I’m not sure. I’ve never used it

Does it let you download you own models?

[tropic nexus]

I’d need to upload it, right? I mean, the reason I want to use dreamstudio is that I do not want to install it all locally.

[next flint]

I’ll take a look and see

Sadly I don’t think so

[tropic nexus]

Thx

[next flint]

I bet there are some sites that would let you though

Keep the image aspect ratio wide (idk 4:1) and add things to you prompt like “1 person”,”character sheet”,”character turn around”, etc

If you are looking for options try happyaccidents.ai. It seems to be integrated with civitai which means there will be tons of great models

[tropic nexus]

Thanks, I will give it a try. I will continue playing around with dreamstudio also, and will give an update here if I achieve what I want.

[hard elm]

I have made quite a few character sheets

[hard elm]

You can use a prompt, or you can start with img2img.

[tropic nexus]

What prompt did you use to get it to work?

[hard elm]

I have used several. It depends on what you want to achieve, and the style you are going for. You can use an image with front/back/side, etc.

Just a minute, and I will give you some of the keywords I have used in the past

beautiful fantasy character sheet, in rogues armor, front view, side view, back view, turn sheet, intricate detail, realistic color palette, soft transitions, all figures are facing different directions, both figures are facing directions, orthographic, game concept art, indian ink, fantasy mythology, senior concept artist, award winning

[tropic nexus]

Which model did you use?

I get this with sdxl, which is not really what I prompted

[hard elm]

3 d fantasy anime girl, soft features, long pink hair, mage outfit, character turn sheet, facing the viewer, two views, facing left, facing right, long view, mid shot, octane render, view from the side and front, bust, game art concept, with wireframe

I have been using these kinds of keywords, relatively, the whole time. You can use these for 2.1, SDXL, too--1.5 as well

[hard elm]

female with long hair and detailed armor, game concept art, front side and back view, arms outstretched, reference sheet, lineart, orthographic view, grayscale, lineart with varying thickness, manga pen, 3d modeling sheet, grayscale illustration, traditional medium, inspired by Final Fantasy IX, art by Final Fantasy etc works, too

[tropic nexus]

Doesn’t really work for me :/

[next flint]

Try changing it to a wide image ratio

[tropic nexus]

Hmm, 1/4 was a bit better (4:3 ratio)

3:2 works better. Thanks guys, I will try more the coming days and give an update.

[hard elm]

Yeah, use wide, and also, when you find an image you like, that starts in the right direction, you can use it as a base.

I think these were...cinematic and fantasy?

You can try the different styles to see what works best for you.

Etc

And go from there

I used analog film for the last two

[thin plume]

Any suggestions on getting image clarity, or more details? What I mean is, the eyes particularly, look kinda.. Muddy? Is the best word I can think of. Fingernails are also a bit weird. Ignore the fancy pinky finger poking out too.

I used Lanczos and R-ESRGAN 4x+ upscalers for the image. The later being used at 0.069 visibility (too high and it sharpens the image too much. So I keep it low.)

Any other tips would be appreciated!

Prompts that probably affect the eyes that are used are: best quality, masterpiece, unreal engine, highly detailed eyes, realistic, ((art by greg rutkowski))
Negative promts: (worst quality, low quality:1.4), poorly drawn, bad anatomy, deformed, disfigured, warped, misshapen, mutated, poorly drawn eyes, warped eyes, distorted eyes,

Oh, this happens regardless of the model I use. I was using Sunshine mix for that, but I often get the 'muddy' eyes in Dreamshaper as well.

[tired vigil]

Whats the initial resolution of the generation

[thin plume]

512x712.

[hard elm]

Change the eye color to something like green

[thin plume]

I think it's because I use an AMD graphics card, despite having 12GB of VRAM. But I can't really go too far above that resolution.

[hard elm]

Or brown

[thin plume]

I need to learn how to install and use that "color cutoff" extension
otherwise, half the time I specify eye color, the clothes change to the same as well! I'll give it a try though!

[hard elm]

Use blue eyes as negatives

[thin plume]

so like, add "brown eyes" as a positive, then "blue eyes" as a negative as well?

In the same prompt.

[hard elm]

Yeah. Something like that. Just try it. Or use blue eyes specifically as a positive

[thin plume]

Ok! I'll give it a try in a bit. Thank you!

[hard elm]

Np

[obtuse torrent]

@thin plume https://www.artstation.com/rutkowski ... check here style (for me) not match whit "Any suggestions on getting image clarity, or more details? What I mean is, the eyes particularly, look kinda.. Muddy?" ... PP: "((art by greg rutkowski))"

[thin plume]

Ah! Good to know! I see the tag on so many images, I recently started applying it myself. Still learning through all of this!

[obtuse torrent]

@thin plume I think that his painting style is relatively classic, I don't think he produces works in the realistic and clear cut that you are looking for..

[thin plume]

I have nooo idea what I'm looking for! I'm a computer nerd more than art nerd. Two of my sisters can draw really well! But I've never practiced enough at it.

I'm just having fun playing with AI things, and learning as much as I can about it.

[obtuse torrent]

me too.. you can try the basic .... Anime or Cartoon or Photographic or 3D or Paint , etc and later pick some artists or concepts in the style..@thin plume ... 😄

[thin plume]

Thank you for the tips! All are appreciated!

[dusky lotus]

what are your other settings ?

I'm gonna need more like how many steps, which sampler, model, any restore faces or hires fix ?

[heavy pendant]

I keep getting this error and I'm not sure what to do. I've tried various fixes but nothing works. I have a 3080 but the most I can do at a time is 1 image and sometimes it will crash at 1024

OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 GiB (GPU 0; 10.00 GiB total capacity; 6.09 GiB already allocated; 0 bytes free; 8.07 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation

[real swallow]

@severe chasm please be mindful of our nsfw rule in #✍🏼｜rules-and-tos! thank you

[light fractal]

a boy

Prompts that probably affect the eyes that are used are: best quality, masterpiece, unreal engine, highly detailed eyes, realistic, ((art by greg rutkowski))
Negative promts: (worst quality, low quality:1.4), poorly drawn, bad anatomy, deformed, disfigured, warped, misshapen, mutated, poorly drawn eyes, warped eyes, distorted eyes,

[timid totem]

portrait of a beautiful woman, cyberpunk, 8k

/midjourney portrait of beautiful woman, cyberpunk 9k

[tall carbon]

i try to implement a body into a picture but not work i do fullbodyshown

how i can do it ?

i get only closeups

nvm it was the checkpoint

[unique lodge]

https://github.com/pollinations/stable-diffusion-audio-reactive only has a prompt option. Should I put my negative prompt in there then like "NO deformed iris, NO deformed pupils, NO semi-realistic"? Like is it a universal ting that prompts can be assembled that way?

[jolly bloom]

I'm using 2.0 and a midjourney embedding. I often get double faces when I do 512x768, what negatives would you suggest to not get double faces? I tried "double faces" but that was not very effective 😄

[silver valley]

You should use the 2.1 model and then you would need a higher resolution like 768x1024 because 2.0 and 2.1 are trained on 768x768 resolution

[tender basin]

How to I get rid of an extra hand?

I told it too many hands

and it just made more people for the hands, but now there are more people

[next flint]

Send the image to img2img. Use only negative prompts ( preferable textual inversion but anything works) at about 0.25 denoise.

[tender basin]

so (too many hands:0.25)?

[next flint]

change the seed and sampler to the one used when creating the image

[tender basin]

or is it with these []

[next flint]

No it would be (extra hands)

In the negative weighting works the same

[tender basin]

ah because its a negative

if you put "not enough hands"

would it make more hands because of the negative

or does it not work that way

[next flint]

negative tells it what not to do or what to stay away from

putting not enough hands might make one an amputee

just go with
"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, (extra hands),( extra fingers), (too many hands)"

[tender basin]

I guess I need to learn inpainting then

[next flint]

Or actually it might give one another hand

[tender basin]

I tried it but it made everything blury

smudged

but I did try it with water so 🤷‍♂️

[next flint]

Good idea. keep in mind that inpainting applies your entire prompt to the covered area. just put only what you want replaced

[next flint]

Can you send me a screenshot

[tender basin]

ah its ok im basically just messing around learning the stuff for now

dw

thanks for the advice

[next flint]

ofc

if ur still confused lmk in dms

[magic wharf]

can you use underscore inside emphesized prompt like:

(((this that this_for_that that_not_this)))

??

[jolly bloom]

Thanks. It's the 512 model tho!