📝｜prompting-help

[next flint]

i mean try flowgpt (has some sd prompt makers) or poe.com ( rate limited gpt 4 )

if u want i can send you mine

i just mixed other templates. doesnt work well on gpt3.5 but it kinda does

[solar dust]

Thought about getting GPT4.0 but it doesn't look worth 20 dollars a month to me.

[next flint]

agreed

i mean it works well with the correct instructions for prompt gen but its just so expensive

id rather pay for api tbh

[azure plinth]

i use claude

[next flint]

oh cool

then i dont have to lie

yeah i use it too

[azure plinth]

lol

[next flint]

i just dont want more people

cuz im getting rate limited SO HARD

[azure plinth]

I open source all my prompts anyway

[next flint]

and now i cant even use temp emails cuz phone verification

[next flint]

wdym

[azure plinth]

it's not!

[azure plinth]

especially with llama2 and codellama out now!

[next flint]

now claude is worth at least 10$/month

[azure plinth]

agreed

[solar dust]

How does one access llama2?

[next flint]

but they fixed my jailbreak for special img prompts

now it only works for like 3 messages

[azure plinth]

easiest way is use lmstudio.ai

otherwise llama.cpp, koboldcpp or oobabooga

[next flint]

i think the people at anthropic actually hate me at this point

[azure plinth]

no they are just catching on

[next flint]

yeah but 4 days after i started making tons of accounts with temp mails they added phone verification

and now i have to figure out a new way to get my jailbreak to last longer

[azure plinth]

oh no

[next flint]

legit there are 4 messages and then it starts getting mad

[azure plinth]

yeah i noticed that

are you using a prompt template

[next flint]

whats that

[azure plinth]

just a long prompt that helps make random prompts

[next flint]

kinda. like i made a long prompt (10kb) but it doesnt just make random stuff

[azure plinth]

oh ok

[azure plinth]

😦

[next flint]

something wrong w it?

[azure plinth]

looks good

[next flint]

why sad face?

[azure plinth]

wrong one

lol

😄

[next flint]

lol

mostly i use it for just improving basic prompts from civit

[azure plinth]

yes

[cloud pine]

Theres a lora for that it has crayon in its name but i forgot the name

[granite crescent]

are these last 2 negative pompts add ons or is there something i have to add to my stable diffusion folder?

[next flint]

they are embeddings. to use them you would just have to search those on civit, download them, and put them in the embeddings folder in stable-diffusion-webui (or any other name)

[granite crescent]

thanks!

[next flint]

https://civitai.com/models/72437?modelVersionId=77169
https://civitai.com/models/72437?modelVersionId=77173
these are the ones used

[half current]

How to create images similar to this style usind SDXL

[twin wave]

I have 2 people in a pic, I only want to swap out just one of the faces. How do I do that?

[next flint]

using roop right?

[twin wave]

That changes both faces to teh same one 😦

[next flint]

oh

[twin wave]

I just want to swap one face

[next flint]

so i just looked at this guide since i dont use roop that much and it says comma sepparated faces are for the output image ( i think )

https://youtu.be/QGrCkjfWpfU?t=436

[next flint]

okay i did some gens with roop and it only swaps one face. do you mean you have multiple faces in the input and output and you want to swap corresponding faces?

[crystal fjord]

so now i am in the SD camp

at least I know what goals I am shooting for

[twin wave]

I mean, I have one picture with 2 faces. I want to swap 1 of those faces.

[quiet zodiac]

@half current I can get close using a prompt like this with varying strengths of supplying your reference image as an image guide.
an astronaut wearing a modern orange space suit walks toward the camera on an asteroid with the sun spiraling behind him, small parts glow, walking pose, thousands of tumbling background debris, highly detailed, intricate detailed, volumetric lighting

[lean depot]

/subscribe

[floral yoke]

when writing a prompt and applying weights to different concepts, how does ading a negative weight to a concept differ from including that concept in the negative prompt? they are handled at different stages of the process right?

[next flint]

is this for a normal keyword or for a lora because most of the time its used for loras

[floral yoke]

just normal words and phrases

[next flint]

gimme like 5 minutes and ill get a grid to show the difference

[floral yoke]

from my own testing i can see that there is a difference, i was wondering at the why of the difference and if some concepts may be useful used as a negative weight

[next flint]

so im not sure exactly why there is a difference but my understanding is that negative prompts tell the ai specific to avoid something while negative weights still tell it not to do something but in a different way that leads to a worse result. things like loras are useful at a negative weight. for example, if you were to use the add details lora at a negative weight you could get something like a vector art drawing or other forms with less detail

[floral yoke]

add details lora as a negative weight is pretty cool, that was the one that got me started on testing other loras as a negative weight

[next flint]

here are two different grids using the gen parameters. the one on the left is having it in the positive and the one on the right is in the negative. ranges -2 to 2 with 0.5 intervals

when you have a weight of a thing at -2 in the positive it completely ruins the image but a weight of 2 in the negative doesnt ruin it

[floral yoke]

yeah i noticed the more weight and concepts that are negatively weighted the quicker it just makes trash

but lets say you add just one of the concepts, like if it is producing one bad quality, on those it would be blood or face paint on her face. just put that in the neagtive prompt and weigh it negative in the main prompt

[next flint]

i dont think i get what you mean

[floral yoke]

anime, anime art, drawing, demonslayer art style,(facepaint:-0.5)

anyways it was jut something i was messing with and wondered if anyone had an idea as to why it is so different

[next flint]

do you want to try one with facepaint and one without?

[floral yoke]

i was just using that as an example, of lets say you get a generation you like but it could be a little better without one detail, if you try it again but weigh that detail negatively vs adding that detail in the negative prompt, whch is often just not strong enough to deter that detail from showing up anyways

[next flint]

i mean im guessing there is a difference but also there is a reason why people havent been talking about using negative weights in the positive prompt

i think just to slightly guide it away from something, putting something in the negative prompt at a low weight would be best

let me try something

so imma specify demon (since it adds horns) and then test those two options for removing them

yeah big difference

took a sc of the top half of the pics cuz i dont want to face the wrath of mods

but the left is negative weight in the prompt and the right is just in the negative. both are with the same weight

[floral yoke]

yeah very different results

[next flint]

much better keeping it in the negative which makes sense

[half current]

@quiet zodiac I really appreciate for your help.
If I am correct, volumetric lighting is key

[jagged kindle]

How can I make it look more like something is under the surface here? I've been trying photobashing with basic image to image so far.

[next flint]

i mean you could alway specify something rising from the acid in the prompt

[carmine wasp]

Help, i want to generate a person from bottom view like this but it didnt works

[frozen raptor]

hello

im trying to get the ai to generate a logo for my clan but its not actually doing it the way i would like it to

prompt: Professional Logo design gaming clan called 99 Gods, cyberpunk, dark background with neon orange futuristic lines, highly detailed, sharp focus, SLR, holographic, 4D, hyper realistic, photorealistic, octane rendering, delicate details and textures, intricate design and details, 8k, 8D, cyan, blood red, orange aspect:1:1 style:Digital Art

this is the prompt

and this is an example

[delicate sand]

what negative prompts would i use to have the desired amount of limbs and making hands not look like spaghetti also if i want multiple subjects probably looking the same or literally the same how would i prompt that? i have tried 3 subjects but it fails im using triplets now but sometimes i get 4 subjects also theyre all different in some way and im not talking about faces and poses in this case i mean the heights and whatnot

also this is what sometimes happens

[quiet zodiac]

@frozen raptor No one can know what way you want it, but here is my revision to your prompt. SD can't do text, so it's better to ask for a boiler plate or flat area where you can drop in the copy using photoshop or something. Requesting multiple color spread throughout the prompt is hit or miss. Try grouping all color requests together.
Professional Logo design gaming clan with boiler plate for copy, cyberpunk, dark background with futuristic lines, highly detailed, sharp focus, hyper realistic, photorealistic, octane rendering, delicate details and textures, intricate design and details, dominate colors of cyan with blood red and neon orange, in the style of vantablack

[candid grotto]

Does anyone know why it seems that whenever I generate something, the preview image is exactly how I want it to end up, but the output is always same pose, different art style and just worse quality in terms of form?

[quiet zodiac]

Probably because of the size change, right? Changing the size is like changing the seed. I don't use WebUI, so I always generate at my final resolution. No surprises.

[glad elm]

Is it possible to make art of the same character I created on AI but now with different poses? I know img2img is a thing but it doesn't give what I wanted.

[burnt crown]

When I generate images using sdxl with controlnet&canny edge, the background is always blurry, how can I fix this? / besides increasing sampling steps

[jagged kindle]

I've noticed this too. I've tried putting blurred background, depth of field, out of focus in the negative yet it still does it anyways sometimes

[quiet zodiac]

@glad elm Consider using a control net with OpenPose. You supply a picture with a person in the pose you want your final character in.

[flat pewter]

Can I use a sentence like i was talking to chatgpt when i am making prompts or does it needs to be words like this " male soldier, tall, green hat " ?

[torn cliff]

Depends on the model somewhat. Some can use sentences, some work better with certain words.

[flat pewter]

@torn cliff When u are talking about models it's this ?

[torn cliff]

Yes.

[flat pewter]

@torn cliff Where can I find this one ?

[torn cliff]

Is that in your settings or the main page where you generate the images?

[flat pewter]

main page

im watching a video for training, but i dont have this model

I want to do like him

[torn cliff]

I don't think there is any way to know what model that is. I'm guessing they changed the name of it to Model

[flat pewter]

ok

it might be the defaut one maybe

[torn cliff]

I think the default one is named v1.5 but you can change the name of the model file to anything apparently. It's like if I saved that image and named it "Discord Sample Image" while you named it "Screenshot"

[flat pewter]

i have these two

[torn cliff]

You can still mimic the video, but the prompts you use might need to be different since you'd use a different model. And yeah thats the default one

[flat pewter]

ok ok thanks

the video is from 2022, even his settings are not the same as mine

[torn cliff]

Yeah things have changed since then. (Funny I am saying that since I only started getting into it in June 😛 )

[flat pewter]

well I m having to much errors with it, i think im gonna stop trying to make this software works lol

since this morning i am having so many issues with

from make it launch to make it works x)

[torn cliff]

It's a bit tricky because there are so many variations of it.

[candid grotto]

size change? what do ya mean?

ive been changing the width, height, and size the whole time, and reset it to what it was originally, but its not doing anything to fix it - preview image almost ALWAYS looks better

[orchid ore]

@candid grotto if preview looks better probably you have problems with VAE

[candid grotto]

interesting, alright then whats that? and how do i fix that in the online client?

[orchid ore]

hm would ask probably some support for it.
If not online, models have their vaes. In A1111 or comfy it is simple, how in online i realy dont know

[candid grotto]

hmmm, alright. yea i defo need support for this i just dont know the proper place to ask. all i know is that im using both a ckpt for the stabl;e diffusion checkpoint and a lora thing

[solar dust]

I've never understood that as well, sometimes when looking at a preview the image looks pretty much perfect, but then when SD finishes rendering it, something will ALWAYS get fucked up

be it the hands, legs, etc.

[candid grotto]

ive tried different sampling methods, sizes, batches, hi res fix, restore faces, upscalers, strengths, all that

[candid grotto]

yea exactly!

plus it changes its style and coloring too

like

[orchid ore]

make screen of preview and final image, to have something

[candid grotto]

itll go from the artist style to something completely different

[orchid ore]

it looks like high cfg isnt it?

[candid grotto]

its at 9 currently, and ive tried 12 as well

[solar dust]

If you have a lot of different prompts, is it best to put more steps in the gen or keep it at 20?

[orchid ore]

higher cfg will not help you, 9 should be o.k. as well as it doesnt change in preview and final

[solar dust]

What about steps?

I see people usually go for cfg 7 steps 20.

[orchid ore]

less steps can hurt images, but will not making image look worse than in preview

@candid grotto you are using bots here?

also you can try without lora, if you havent tried it

[candid grotto]

no, im using the online client and ckpt/loras i found on civitai. the lora is of the specific character im wanting (which ive tried without), and then the ckpt is basically of the body type and artist reference

[orchid ore]

not sure about loras personaly, never used them. I know but can be mistaken simply not sure, how is it with using more loras.

[candid grotto]

hm, ive not tried that, guess i can lets see

[candid grotto]

it made it WAYYYYYYYYYYYYYYYYYYYYY worse

[orchid ore]

o.k. what model are they using, do you know?

[candid grotto]

i do know the model yes. only reason i havent mentioned it specifically is because it violates the server rules (im generating nsfw art references so i can become a better artist myself in that nature)

[orchid ore]

o.k.

[candid grotto]

its nothing super crazy nor anything actually bad but i dont wanna break rules of course

[orchid ore]

check if models and lora are same version

[candid grotto]

both are based on SD 1.5

[orchid ore]

and model as well.

[candid grotto]

yes

both the model and the lora

[orchid ore]

i dont know. really. I think the reason is vae, but they should have this solved. Sorry

[candid grotto]

You're fine fam, thanks for atleast trying to help lol. You've been the only one in the past 24 hours basically xD

[prisma grail]

I'm having trouble doing headless people, or just anything replacing the head with something else, is there a way to manipulate it?

[quiet zodiac]

Have you tried decapitated?

[prisma grail]

I haven't I suppose but I had head in negatives and headless in positves with a high weight on both of them

if that wouldn't work then I doubt anything really will

I'm trying a lora

[hexed atlas]

#🤝｜tech-support message

@marsh wraith It's simple, you have your current image on the right side after generating it, right?

Under the image you have the button "Send to inpaint", click that and it will send you to the inpaint tab

In the inpaint tab change the following:

• Positive prompt: add your lora (for example lora:biting_lips:0.7), add the trigger word/words for the lora to work (for example biting lips)
• Inpaint area: Only masked (Only what you mark with the pencil will be inpainted)
• Denoising strength: Change it to something between 15 and 25 (This means how much the original image will be taken into account when inpainting)

Here is a small short tutorial, which helped me before:
https://www.youtube.com/watch?v=No1_sq-i_5U

You can first start with generating one result to see if it's actually doing anything coherent, if you like what it's doing, do a batch of 5 until you get what you need.

[marsh wraith]

Thank you so much bro

[kindred blaze]

For the life of me I cannot get a doll with a windup key in her back, tried a few variations, but the key ends up floating around or in her hand or anywhere beside her back xD

[tired vigil]

It doesn't sound too hard to inpaint one, with a rough drawing of one in the image

Maybe there is a lora for a character like that

That can help generate the key

[kindred blaze]

Okay thanks, I'll try both 🙂

[tired vigil]

The only character I can think of that has one is Oriana from lol xd

[hexed atlas]

Yes, getting a windup key on the back is hard, so I guess its up to the inpainting like Cat Bot stated.
Enjoy this cursed one:

[kindred blaze]

lol, not bad! I have a few LORAs for ball jointed doll characters and that sort of thing, at this point it's the wind up key that's giving me the most issues and trouble 😭

[hexed atlas]

the one I found is cute_joints_dolls_set999_v1.0

set999 is the user name

I name them like this so I know what the hell they are 💀 (backtrace them to civitai)

[left yoke]

Hi, I need a concept art for a video game, and I am finding it extremely hard to generate this image.
All I need is a garage with few snowmobiles, but we also need a gas station alongside the garage where a person is lying on the ground(dead), it should be a night sky, and the environment should be snowy...

I have tried to generate this image on SDXL, Leonardo AI, blue willow, everywhere, I cant generate where all these conditions are met
Can someone help me with the prompt?

[proven jasper]

try with something like this scenery,outdoors,gas station,snowmobile,(winter,night sky:1.3),snowy,horror theme, 1 person lying on floor,dead, blood,

[solar dust]

Is there a checkpoint that can generate characters in play dough style

similar to Pingu or one of those stop-motion animations?

[gentle lantern]

Oh man, I've got a great pingu horror animation

I think there's an SDXL prompt template for claymation if that helps

claymation style {p}, sculpture, clay art, centered composition, play-doh###{np}, sloppy, messy, grainy, highly detailed, ultra textured, photo, mutated

...it might even be the same prompt used for craft clay on the bots. 🤦

[eternal forge]

fix anyone?

[broken crystal]

Can anyone reproduce this type of image? It was also generated by an AI

[rich crow]

Id start out with a canny controlnet of an airplane and prompt for an outdoor or exterior restaurant dining patio or rooftop scene or something along those lines

Could even get a 2nd canny controlnet of a rooftop outdoor restaurant and combine the two

[final briar]

Hi, I would like to recreate an illustration. how could you describe this position in prompt? Because I tested many prompts (lying, hands on legs, looking up etc...) and it gives bad results. The Lora is good though.

[broken crystal]

So it's complicated in Text2Image, so it's better to do it in Image2Image ?

[rich crow]

No not image to image, you can still use text2image with controlnet

But you could use img2img yeah

[tired vigil]

Use controlnet poses extension

Open pose

Through prompt only I did manage that with "leg up, hands on leg" but it's very not good, rare to make it look like what you want

[hexed atlas]

what @tired vigil said
https://www.youtube.com/watch?v=ALnmwCpM2xE

[final briar]

ok thanks, I will check open pose then

[left yoke]

Hello! I am not able to generate this image, I have been trying for two days now. Will be trying on comfy UI tomorrow.
Can someone help me with this scene?

"view from outside of a big garage with snowmobiles, gas station in front of the garage, a man sleeping on the ground near the gas station, outdoors, nightsky, north pole"

It always misses out on the dead person on the ground, I have tried on all the ai image generators

View from inside a larger hangar with some snowmobiles. The hangar contains cabinets and tools for mechanics and a small pump station for refueling the snowmobiles. A lifeless person is lying near the pump station. The gate of the hangar is visible, outside there is twilight. The hangar is part of a research facility at the North Pole and is set in a barren snowy landscape that is in a perpetual twilight.

i m looking to generate this image ^

[next flint]

Interior view of large aircraft hangar in arctic research facility, snowmobiles parked inside near refueling station, lifeless body lying on floor, hangar contains cabinets of mechanic's tools and pump for refueling, large metal sliding hangar door open revealing twilight arctic landscape outside of barren snowy tundra and minimal sunlight, highly detailed and photorealistic, cinematic lighting, by James Gurney ?

[proven jasper]

something like this?

[left yoke]

Yup.. But we also need a garage alongside the gas station, where some snowmobiles are parked.. This looks amazing by the way

[left yoke]

I will try to use this image, I will add a garage with snow mobiles myself with photoshop, on thr right side of this gas station, or just try to adjust somehow.
But, I was trying to put the dead body near the gas station, and the garage with snow mobiles beside gas station

[wraith echo]

How to get stable diffusion to stop generating dresses that reveal the fcking legs? Ive entered legs, short dresses, short skirts in the negative prompt and they still show legs like they have a leg knk

[wraith echo]

This stops me from generating specific fashion styles

[tired vigil]

When I want that, I put revealing clothes, revealing outfit or skimpy outfit in the positive, so try that in the negative ?
And full dress in the positive helps with less revealing dresses when I want them

Which model are you using ?

[torn cliff]

Could also try a lora for long dresses

[wraith echo]

I’ve tried typing revealing clothes in the negatives, haven’t tried “skimpy outfit “ tho, thanks

[broken crystal]

I want to generate a house but with the dimensions 1080 x 1350 the ia generates me an image cut in several times whereas I wish only one image.

I tried to put "split" in negative prompt but nothing changes.

Does anyone have a solution?

[hard elm]

What's your full prompt?

[broken crystal]

modern style, design futurism, villa built into desert with pyramids, Mediterranean, Cirrus clouds, night, golden hour, lights, (curved:0.5), (brutalism:0.5), raw concrete, metal, highly detailed, hyper-detailed, extreme detail, award winning, (Archviz), visualization, ambient occlusion, Octane render, 4k, 8k, HD, photorealism
Negative prompt: 3D render, soft, burry, divide,no split

[hard elm]

You could try "a series of images" as a negative

[broken crystal]

unfortunately it doesn't work

[hard elm]

I would re-write that as...."A modern style villa in the desert at night, Mediterranean, cirrus clouds, golden hour, golden ratio, brutalism, raw concrete, metal, professional photography, award winning, cinematic, photorealistic, ambient occlusion, unreal engine 5 quality render, ambient occlusion" if you wanted to throw everything in

And you are using sdxl

[broken crystal]

It works perfectly, thank you very much. Nop i use stable amtrix UI

[hard elm]

Np!

[broken crystal]

As soon as I add "with swimming pool" it starts all over again.

[flat flare]

Hi, im getting error about ur still dreaming. How to i stop dreaming?

[quiet zodiac]

@broken crystal It might be your size resolution. Try 768x1024 as a test.

Using a modified version of Sunny's prompt and your reference, I got close to your airplane house.

[tired vigil]

You should try lower resolution, when I go wide/tall it just duplicates things everywhere :p

like cut both in half and upscale 2x later if you want exactly that resolution

[weak seal]

how do you i prompt to change color hair for having random between brune and blond for instance ?

[quiet zodiac]

I use python to generate my prompts, to randomly choose from a list of hair colors, body types, emotions, poses, etc...

[marsh wraith]

How do you make an image similar to this?

[quiet zodiac]

I'd say that's a tough one. You might try generating it as a reproduction and reference the original artist's name. SD has a hard time making believable characters at a small scale. Higher output resolution does help, but expect vague faces.
a reproduction of the school of athens painting by Raffaello Sanzio da Urbino, detailed architecture, detailed inhabitants

[marsh wraith]

That's gorgeous. Thank you so much

[marsh wraith]

I'll attempt to do impainting

[torn cliff]

There is an extension called Wildcards, you can put various things, such as colors, into a text file and when you call the wildcard it will randomly select a line from the file. You use it by putting a file like "colors.txt" and then the wildcard would be "__colors__ hair" in the prompt to generate a random hair color.

It's two underscores on either side of the name of the text file

[manic juniper]

Can someone help me in prompting 2 loras for one pic, never did it before and somehow it doesn't seem to work for me when I try z_z

For example, I'm trying to do this with 2 loras, but no idea if I should use LatentCouple, Controlnet or Regional Prompter

This pose is also a lora of it's own: https://civitai.com/models/131969/shared-umbrella-or-lora

[candid dome]

I seem to remember a syntax where you could have a keyword active for X steps then stop, does anyone know what this was?
And is there a place more advanced syntax stuff can be found? Couldn't find a breakdown on the github

[solid bramble]

i need help with a prompt, can't get the girl to look out over the balcony at the lake. seems like no matter what i do she always has to look back at the viewer.

[twin yoke]

describe in the prompt the back of her head

[lean pecan]

hello, sorry if this isn't the right channel, i'm new to ai image generating and was wondering how to do good prompts? is there a specific syntax or something? i have seen people putting words in (), and others even doing something like <something=0.7> in their prompts but i really don't understand very well, so i'm wondering if you have any tips/tutorials

[quiet zodiac]

Just google. There are all types of markups all over the web. Decide what platform you're going to use first. This helps narrow down what special characters you can place inside your prompt. The UI systems support different features. Overall, try well formed sentences first.

[solid bramble]

what on earth are you talking about... "describe the back of her head"? it's the back of a head what is there to describe? if you have specific phrases that remedy this then suggest those.

[twin yoke]

You have to understand how prompting can work. Telling SD that the subject of the image is looking away is only one way to try to accomplish what you're looking for. However, if you describe the back of her head such as literally saying "back of head" or something similar, that forces the sampling to render that thing you're describing and thus it would infer that she's looking away from the viewer.

The other thing to consider is that the model you're prompting against may be poorly trained for looking away from the viewer.

[solid bramble]

i understand that, but describing something that is already in frame won't have any effect. not to mention this doesn't always work. i have had plenty of instances where i have asked for a specific item or plant ect to be in the image and it doesn't show.

[twin yoke]

describing something that is already in frame won't have any effect
That's simply not true. Every single adjustment to a prompt that has any impact on a token will have some effect. Depending upon what changes you make such as where in the prompt you put the words, how strong you make them, anything you put into negative and how strong you make those, etc., could potentially land you what you're looking for. Also, simply attempting this against a bunch of different seeds could also give you the outcome you're looking for.

[quiet zodiac]

What if you tried something like...
looking away, across the waves
Sometimes you can't fight the seed. Try another one.

[solid bramble]

this isn't my first day prompting i understand the concept plenty well. your not telling me anything i don't already know. in the mean time i have tried your "describe the back of head" suggestion to no avail on my checkpoint of choice and base sdxl.

[solid bramble]

sadly i've tried looking away, looking at lake, looking at forest, negatives, looking over shoulder, looking a viewer, facing viewer. it's rediculous.

[quiet zodiac]

New seed or image guide, maybe openpose. Perhaps tweak a negative. Sometimes just switching things up can cause a change.

[twin yoke]

Literally added "(back of head:1.4), " to the prompt and got this:

[solid bramble]

i'll try openpose or controlnet on comfy as a last resort. imo if i can't do this without 3rd party help then there is something fundamentally wrong with SDXL.

[twin yoke]

Now what?

[solid bramble]

idk what part of what i'm trying to do your hung up on. i want her to look stright forward at the lake. no facial skin visible period. no ears, neck or face.

[twin yoke]

So describe the hair, put things like ears, face, eyes, etc. in the negative.

[quiet zodiac]

What about..."posterior or distal view"

[twin yoke]

Good suggestions as well.

[solid bramble]

lol what?

[quiet zodiac]

I googled "what is facing forward called"

[solid bramble]

i'll try it just for the laughs at this point. better than any suggestion "soul" has had so far.

[twin yoke]

When you come asking in an open forum for help and someone's trying to give you ideas, maybe don't be an asshole about it just because you don't understand or like the advice.

You originally said:

can't get the girl to look out over the balcony at the lake
And I'm pretty sure the results I posted above give you exactly that. You didn't say "no ears, neck or face" until after I solved your original problem by following the advice I gave you.

[solid bramble]

sadly no dice, thanks for trying.

[rigid wind]

So my friend and I are talking about negative prompts and we ran into a question that we're not sure the answer to. We both use words like ||underage, child, children, kids, kid, shota, loli|| in the negative prompt whenever we're dealing with something that may turn out NSFW content but we were wondering how models know what these inappropriate subjects are? Like surely they aren't trained on these subjects so how does it know to avoid them?

[quaint cairn]

https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2Fknn.laion.ai&index=laion5B-H-14&useMclip=false&query=loli

You can just check for yourself

[quaint cairn]

I've ran into that issue too, most models always try to put a face in there. Haven't had much luck with prompting but a possible workaround is to roughly edit the face out and run it through img2img

[stoic fjord]

idk if this is the right place to ask but is there a way to make the bitcoin logos come out better? they're a bit janky. or is it better to just clean those up on photoshop?

[rigid wind]

couldn't you use OpenPose to try to force the face forward?

[quiet zodiac]

@solid bramble I tried a patched image guide as well. It kind of works. Even with raw prompting, the pose systems tries to not give you that 90 degree looking away.
facing away, posterior view,a red haired woman in a green dress standing at a railing overlooks a fjord looking at the castle in the distance, realistic skin, volumetric lighting

[solid bramble]

i'm guessing this is using a1111 inpaint rite?

[solid bramble]

thanks i will try out this prompt myself and see what i can tweak without breaking.

[quaint cairn]

Yeah, that's also a good idea

@solid bramble yes, img2img in a1111

[frozen thorn]

Hi all.. brand new to stable diffusion..
Installed Automatic1111 last night. Played with a few different models and prompts.

My goal is to get a crest/logo of a grizzly bear done in a very minimal/flat style. Similar to this (which was done by bing image create).

But the images I end up with are still just far too detailed like:

Here is the prompt info I used (logged out by auto1111):

"(stylized mascot logo, sports emblem, vector art),  (angry grizzly bear mascot:1.2),  (3 solid colors, thick-linework), white outline, black backdrop, black background, (common sports logo emblem and heraldic elements:1.4), HDR, UHD, vibrant, dynamic, (detailed:1.36)",32016053,1024,1024,Euler a,4,16,00008-32016053.png,"3D, realistic, modern, photorealistic, turn-based strategy, text, title, worst quality, low quality, ugly, blurry, signature, watermark, monochrome, (grayscale:1.4), shadows, 3D, render, photo"

and I've used crystalClearXL model and sd_xl_base_1.0 as well.. both end up kind of the same results.

How do I get it to be 'simpler'.. more flat vector style?

[unique bane]

can we add the reference image?

[tired vigil]

Try a detail tweaker lora

It can add details but also simplify it with negative/low values

Also HDR and stuff like that prompts could be hurting your desired image making it more "realistic"

[orchid ore]

also higher cfg, cant read what cfg are you using. Also i believe there are for sure loras for 1.5 maybe for sdxl as well
Also you can finalize it in inkscape where you have more control and can vectorize it properly.

[frozen thorn]

ok.. I'll look into those.

I had removed the HDR and UHD tags last night.. didn't seem to make much different.

[frozen thorn]

That's kind of my plan.. want to get it generating something close and then figured I would pull it into inkscape.

[queen ember]

Hello, someone have an idea how i can have this kind of result please ? prompt &/or checkpoint ?

[quiet zodiac]

Kind of getting there. Your best bet might be to supply that reference image to extract the style.
(lit in a sepia tone:1.5), a high contrast pencil drawing of a romance book cover male model wearing armor by Jesper Ejsing, extra detailed,2D, sketch style, concept art, inking, orange and white,vignette fade around edge
Model: galaxytimemachines/GTM_semirealV11

[queen ember]

Thx a lot for your help, i gonna try

[unique lodge]

you ever realize we kinda look the same?

😁

[queen ember]

kind of 🤣

[charred root]

hey guys, I'm using Photon v1 checkpoint, and for some reason I'm getting the same or a very similar person each time.
Here are 2 batches of 2 images
with prompt a photo of a 26 y.o. woman, Sarah, white skin, bangs, bare shoulders, brown hair, jacket, long hair, off shoulder. sitting, sleeveless, solo, twintails, very long hair (close up face portrait)
How can I make them all different?

[wise vigil]

What kinds of ckpts/loras are good at making the kind of painterly/surreal images that Midjourney has characteristically been known for

Stuff like this for instance (real Midjourney gens I've done)

[tired vigil]

Avoid names in your prompt if you need different faces every time because prompting names will give you similar faces.

In your case "Sarah" is the name of a person and that's the reason you are getting the same face in each generation.

Also make sure your seed is set to -1.

[static hinge]

How can I prompt so I get an image that looks like taken from a TV program?

[charred root]

I've added that name as a last resort just to see if it's gonna be different from others.
One dude recommended using some decloner or something like that, and it kinda works

@tired vigil

[vital charm]

Has anyone using hypernetworks getting worse results after updating A1111 to 1.6 ? i am now getting blurry images like using the same hypernetwork i used before. Cant seem to understand what is the issue

[hard elm]

What exactly are you looking for? You mean like a movie, a tv show, a news broadcast?

I just cobbled these together with

a singular crystal, with gold on one side, surrealistic brush stokes, complimentary color palette, digital watercolor, surrealistic art movement, abstract art background, salt technique for stars, a detailed oil painting

You could get more finite details, probably, by using the refiner, maybe adding more details about brush strokes, textures, canvas, etc

Using texture overlay, canvas texture overlay, etc, might help

a singular crystal, with gold on one side, surrealistic brush stokes, complimentary color palette, digital watercolor, surrealistic art movement, abstract art background, salt technique for stars, a detailed oil painting, on canvas, knife palette, textured canvas overlay

[quiet zodiac]

@static hinge Try using this model:
https://civitai.com/models/9114?modelVersionId=142173

[quiet zodiac]

@hard elm Your prompt seems to work well in several different models.

[hard elm]

Yup! It should 🙂

[sweet lily]

what is the prompt for the anime long spiky hair?

[atomic wigeon]

Try "animepunk spiked hair" maybe.

[static hinge]

like a tv show (news, talk show), infomertial, history channel crap. Both like real escenes in the style and composition and feeling of being film for it but also kind of a screen shot of how it looks in the TV, and also a photography of a TV. Any of it but usually I would I get none.

[frank imp]

I'm trying to create a fantasy map for a game. SD is giving me really nice images, but none of them are actually a "map". How can I get the camera angle from directly above? The 3d looking icons and colorful graphics are nice, but it needs to be an actual map that can be used with coordinates. Also I'm trying to get several key features on it, 4 different forests, swamp, snowy mountains with a volcano. Feature rich eh? Any prompting help would be huge! Thanks

[quiet zodiac]

Something like this might get you going. You may need to push the CFG scale a little more than usual.
an overhead view of a paper world map for a dungeons and dragons campaign, intricate, highly detailed, epic, infographic, marginalia, cartography

Ooh, here we go. Switching to model Realistic Vision v51.

You may want to generate a few sparse map candidates, manually add in your key features using photoshop, then submit it back as an image2image render to remove any clumsy editing.

[granite crescent]

how do i remove this style? it always starts when i load up A1111 as default and i forget that it's there.

i mean obviously i can click the X and remove it temporarily but when i start up A1111 next time it will be there again

[torn cliff]

In the main folder where the bat file is, there is a file called styles.csv, edit that and you can delete that style from it.

[granite crescent]

thank you!

[torn cliff]

Welcome. ^.^

[tired vigil]

You want those but more realistic ?

also I'm trying to gen some and I'm getting eyes, just the eye
Have some big eye

Lmao

SD doesn't want to cooperate with something I'm doing

[fluid solstice]

Idk if this is the appropriate channel to ask; but when I increase batch count, all the different images in a batch look very similar. All this time, I thought they got the same seed but it seems it is not the same but the previous one +1.
I am using wildcards, so I would like that each image with a certain iteration of wildcard prompts gets a random seed.
Can I do this with the extra menus in the seed field somehow?

[kind kayak]

Nothing can make model generate two colored eyes. It's impossible

Please help me

[twin yoke]

Try adding the word "heterochromia" into the positive and adjust the strength accordingly

[kind kayak]

I have to put parenthesis like ((((((((((heterochromia)))))))))?

Cuz my checkpoint is really stupid.

[twin yoke]

Just start without that, then if needed, use (heterochromia:1.2). If that doesn't work, bump it to 1.3, etc.

[kind kayak]

What is 1.2 1.3???

[twin yoke]

That's the strength.

Same idea as the (())

[kind kayak]

Oh ty

even worse.

It's imposible

Oh it worked!

[twin yoke]

Is it?

[kind kayak]

Daaamn

btw

[twin yoke]

Keep in mind that a lot of this relies on random stuff like the seed value. You have to tweak values sometimes and just try, try, try.

[kind kayak]

how to make it generate a different color?

[twin yoke]

Just play with the prompts until you get what you want.

[kind kayak]

but I am bad at English

[twin yoke]

Use translation services.

[kind kayak]

Ah cmon, there should a better advice than this

I expect so many things from Stable Diffusion becuz I cannot draw

[twin yoke]

better advice for help with English?

Translation services are free online and do a great job

[kind kayak]

facepalm

prompting, duh

[twin yoke]

You said "but I am bad at English". My response was to that.

[kind kayak]

fine.

[kind kayak]

You can also even generate a blue brim of a red cap

what are you?

A god?

a god of stable diffusion?

I'm trying to generate this son of a bitch

[twin yoke]

It's just random. Keep generating and changing the prompt around and you will get similar output.

[kind kayak]

I guess I am a man who have no patient then when it come to waiting for A.I. to generate everything right for me

[twin yoke]

It is not perfect. It is not capable of what you are asking it to do without a lot of patience. It is still very early technology, but it is getting and will continue to get better rapidly.

[quiet zodiac]

Can you recommend a model that supports heterochromia? The ones I'm trying don't seem to recognize the term.

[still hawk]

I mean I want more like those. Not anime or cartoons for example.

[proven jasper]

skill issue

[kind kayak]

stop

[proven jasper]

😔

[kind kayak]

Anyway

basically

the more RAM you have

the faster you can generate?

along with graphic driver?

[proven jasper]

its all about vram and the gpu, like a on quadro has a lot of vram but cant generate fast because its an old gpu

3xxx and 4xxx series are the fastest ones

[hoary galleon]

hey, i'm very bad at prompting, anyone could tell me how to get an image similiar to this?

[lean pecan]

hello, idk if this right channel but can anyone help me with openpose? i don't understand how to pose the head, what are all these points supposed to mean? i understand the blue would be the neck obviously, but what are the other four? eyes/ears?

[static hinge]

I was trying the Dynamic Prompt in order to use the Magic Prompt, which I used in the past but now it doesn't work https://huggingface.co/spaces/Gustavosta/MagicPrompt-Stable-Diffusion

The thing is, I don't like that it generate the prompt and then you have to put the image in PNG info to learn what it added, so it is very annoying to test, also it doesn't give different options.

[static hinge]

yes the biggest circles on top are the eyes and the other two at the sides are the ears. I don't know about the blue circle, it may be the nose?

[lean pecan]

oh ok, thought it was something more complicated, thanks, i assume the blue one to be the chin? maybe

[spark rune]

godamn man, thats exactly what i need

any good tutorials on openpose to link me?

[static hinge]

yes could be the chin also, but it is the nose, mouth or chin I guess

[spark rune]

I wasnt aware about it till now

[lean pecan]

if you are using comfyui you can use this https://github.com/space-nuko/ComfyUI-OpenPose-Editor

[spark rune]

thanks

I was going to literally train a entire new custom model based on a dataset just to get the position i need lol

nice to see there is an easier way

[static hinge]

also "don't generated images" generate images (I thought it may just generate the prompt)

[lean pecan]

speaking of wich, you have good links on how to train models yourself? am a beginner in all this so no idea how to do it but would like to try

[static hinge]

oh no, you can also use two or three controlnet models on top of each other if you want more control, OpenPose with Depth/Canny

If you are using a reference img

[spark rune]

Two different aproaches for using dreambooth: https://www.youtube.com/watch?v=eh1LOS7TFZ8 https://www.youtube.com/watch?v=usgqmQ0Mq7g

[spark rune]

what do you mean?

[lean pecan]

thanks

[lean pecan]

i think he's talking about this https://www.andreszsogon.com/comfyui-workflow-sample-with-multiareaconditioning-loras-openpose-and-controlnet/ ?

[spark rune]

I see

not sure if it applies well to my use case

my goal is to generate different characters always on a fixed position, to enable it to be animated later

from a text prompt

so such fine tunning details are not that important

[static hinge]

if you are using a photo for reference, you can (in Automatic1111) enable several ControlNet modules, put the reference photo in each, use OpenPose, Depth, Canny in each. Play with the Control Weight, so you will get the pose and also details of depth and others

[spark rune]

I will try openpose out. if I can always generate different characters always at the same position, it will work great in my project

so lets see

[static hinge]

Also it seems this extension is better https://youtu.be/30b2k1p2CiE?si=yl748HBsbaEccYqi

It has to be installed from URL, havent try it yet

[spark rune]

thank you, guys

Im so luck I randomly stumbled in this chat today lol. straight up saved 1 week of unecessary work in my research project

[hoary galleon]

thanks for helping me everyone ig

[tired vigil]

so im trying to change the hairstyle of an image to another from a lora but it just doesnt work what am i doing wrong?

i get the hairstyle but it doesnt look like i want it

this is the lora itself

this is what im getting

[royal fractal]

Heya all, does there is a database of most trained words / character in SDV / SDXL ?

I’m using vanilla SDXL / SDV 2.X

[crystal echo]

Does anyone know how to disable highres fix running by default on the new A1111

[tired vigil]

If the box is expanded then it is on, if it looks like this then it is off.

[crystal echo]

it runs both ways on me. i have been watching it in the cmd window and it stays on

[granite pulsar]

I'd like help to recreate this image. Which model, prompt, settings, etc.

[tired vigil]

The answers you will recieve for this will be very vague, as this piece of work is very abstract. Your best bet would be to find very similar works like this, make it a LORA and apply it to something like dreamscape/RPG/unstableinkdream

Another option is to see other peoples creations of works of art, and follow their prompt. There is no wrong or right answer

[granite pulsar]

I'm taking a look to this: https://github.com/comfyanonymous/ComfyUI_examples/tree/master/unclip

[manic juniper]

Hello, can someone please help me generate 2 characters from a lora in 1 pic? I want to understand what I'm doing wrongly z_z

I have a ControlNet pose for that

And if I have a ControlNet, do I still need to use things like LatentCouple or RegionalPrompter ?

If so then which one, because there's so many options I don't know which to check @_@

This is the pose I want to have 2 loras on left and right, what else do I use besides ControlNet?

[quiet zodiac]

@granite pulsar I do recognize the effect. Do a google search on Bokeh and Plexus. They are visual effects plugins. Try adding those words to your prompt.

[warped drift]

im trying to make concept art for ork guns, any tips, ive been struggling w this for a few weeks

[lapis warren]

I saw someones prompt have a weight column in the middle of a grouping, (beautiful, high quality, highres:1.1, aesthetic)
Is that equivalent to (beautiful, high quality, aesthetic), (highres:1.1)?
I thought if I did (a, b, c:1.5) it would apply a weight of 1.5 to a and b and c, but does it only do c?

[quiet zodiac]

I believe it will only apply to c because of your use of commas. (beautiful high quality aesthetic:1.5) would apply to the entire phrase. But (beautiful, high quality, aesthetic:1.5) would only apply the push to aesthetic on top of the other outer parenthesis push applied to the entire group.

[lapis warren]

ooh well ive been doing it 100% wrong all this time

misused (a, b, c:1.5), should have been (a b c:1.5) ?

but then are these 2 equivalent? (sky:1.5, moon:1.5, clouds:1.5) vs (sky moon clouds:1.5)

what if I use (large moon, blue sky, shining clouds), if I do (large moon blue sky shining clouds), wouldnt that result blue moon, large clouds, shining moon, etc?

[quiet zodiac]

Your first example, I'd say yes, but with a comma being a delimiter, that changes the meaning of the prompt. While "weight wise" these two are the same (sky:1.5, moon:1.5, clouds:1.5) vs (sky moon clouds:1.5), they may produce different images.

[obtuse torrent]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#attentionemphasis ...
I'm not an English speaker, maybe that's why I understand this in the opposite way... (in the sense that the absence of separators like the "," was what caused the new notation to only affect one tag, obviously I can be mistaken)
I suggest you do a little testing with (x,y,z) ..to see the effects as they really are
... regarding crossed attention, this is natural to the model, since the tags at the end are translated into a function that tries to fill as much of the latent space as possible... things like the order in which the tags appear can be translate into variations...

[lapis warren]

does the model treat synonyms (i.e home vs dwelling) the same, or will home show modern home and dwelling show rustic home?

[lapis warren]

this is interesting, does crossed attention mean multiple tokens in 1 set, ie 2 words no delimiter (i.e large moon)?

[obtuse torrent]

crossed attention is typical of the model in the sense that internally it does not work with words or letters but with vectors... that is, a data matrix, it can be thought that when executing an SD prompt it decomposes it into vectors, then looks for those vectors in other images and recreates something similar to those images it finds... that's why in many models you can put a "cute cat" prompt and get a cat girl... because in the initial data set there is often a catgirl, was tagged as "cute"... that's crossed attention... mathematically they are functions fighting to fill a whole space that for them has no limits... hence a red cap "wants to dye the pants red also" ...

but the examples above will not eliminate possible crossed attentions or "bleeds"... what you ask and discuss is how the model interprets the signs that reinforce attention...
In short, you increase the "weight" of a tag... when you verify that in the initial prompt it is not shown as you want.... that is, you increase the "weight" for the tags that by themselves lose the fight in the latent space and do not show up in your image...

[quiet zodiac]

Does anyone have any tips on how to make a font/word appear like what it says?

Professional font made of gears spelling 'MachinE'```

[hollow dagger]

i am using control net and my prompts arent really taking into shape around the image im using

[high leaf]

hi! Comfy user: how do you apply a process like ImageFilterSharpen to an image with a mask? so you don;t sharp the background for instance? is it even possible?

[tired vigil]

I tried to generate Goku and Naruto clashing. But the AI is generating two narutos, with one wearing goku's cloethes. Why is this? Is Goku forbidden or something?

[wise belfry]

I noticed a lot of times SD1.5 models at least will tend to confuse where to put specific colors, for example mentioning blue eyes and green skin would make the skin blue-ish, mentionign red eyes will make the arround part of the eyes except the iris red. How to prompt around that and make the desired results respectively?

[silver valley]

Check the cutoff extension. That can help with colors

[torn cliff]

Just note though that apparently the Cutoff extension doesn't work with AMD cards for whatever reason.

[half bone]

My prompt only draws a lion, instead of what I asked for. Help?

faces of an eagle, lion, ox, and man on points of a compass. heraldry, symbolic mural, (foreground amber illumination:2.0), a detailed painting, by William Morris, arts and crafts movement, 1 9 2 2, black velvet painting, nsfw:2.0, renaissance style, mysterious, haunting, ornate, detailed

[sacred wolf]

Hi guys, I have a specific question. The first image is the original art and the second image is what I want to achieve - I'm looking to keep the same art while basically just "leveling it up" like it was on the 2nd image. How do I get such high quality image with vibrant colors that simultaneously changed the art a bit (head is different, belt is different, fur on the legs etc.)

It doesn't seem like the person worked on different body parts but it seems like they had the entire image changed

[lavish trench]

I'm trying to get a photorealistic ... photo starting from a paint, for example Monnalisa.
I'm trying with img2img but as far as I get sill looking like a paint, even if I use (paint, draw, sketch, illustration:1.5), in Neg

😕 any Advice?

That's the best that I get 🤔
(close-up:1.3) photo of an orc and a human knight facing and looking into each other's eyes, in front of a crowd of people in a battle scene. ((masterpiece:1.3, best quality, absurdres)), epic realistic, award winning

(paint, draw, sketch, illustration:1.5), ((mutilated)), ((poorly drawn face)), (((mutation))), (((deformed))), ((ugly)), blurry, ((bad anatomy)), ((extra limbs)), cloned face, (((disfigured))), child, childish, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), jpeg artifacts, signature, watermark

[pine lintel]

Does Lora positioning matters within prompts?

[half bone]

I suspect "no", though I havn't proper evidence to back that.

[pine lintel]

it seemed to me like lora positioning changes the output.. i have done a test on that and found out

[spark rune]

hello, is there any good tutorials on the best way to modify dw openpose's parameters?
I need to make the dw openpose skeleton more impactful during generation

I mean, really...

how the hell did this skeleton

generated a image like this???

I doesnt make any sense. There must be some small detail that I forgot

[spark rune]

like look at this image. all limbs are in the exact desired position. Not sure what I'm doing wrong, but I cant get results like this

[obtuse torrent]

Conntrolnet and the model should be active...

[static hinge]

Does anyone use Magic prompt? I still figuring out which model made what, some just add "by beksinski" and yes some cool (maybe generic as I think I saw that name a lot in prompts). Other just add random inconsistent stuff unrelated to the original prompt (which could be good if you just want random stuff but usually is nonsense and unrelated)

I liked the original magic prompt from the web browser that gave you 3 options and you could take from that what you want, this system where you generate and then has to figure out what it add by inspecting the image I don't like so much.

[tranquil shard]

A question about emphasis syntax - (a, b: 1.4), c
what will be a? 1.1 or 1.4?

[quiet zodiac]

I would think a would be 1.1 and b would be 1.5 because it's inside the parenthesis and will inherit the 1.1 boost as well. If you remove the comma, token a b would be 1.5.

[inland tide]

ok I'm stuck I want to know how did it generate to the complete opposite of what I wanted

[torn cliff]

Because SD has color bleed issues where if you put in a color it will tend to affect anything else that normally has a color. Like you might have gotten a black hoodie, green hoodie, black eyes, green eyes, green hair, etc.

The Cutoff extension supposedly helps with that, but it apparently doesn't work on AMD cards

[inland tide]

your what your saying is SD doesn't want to work what I got

[torn cliff]

What you can try, is to try and generate an image that is close in looks to what you want, then use inpaint to mask sections of it individually and put in specific colors. Like mask over the hoodie and remove most of the prompts except for "black hoodie" then when you get a result from that which is decent, send that result to inpainting and mask over the hair or the eyes and put in the color for them instead.

[inland tide]

can I have some examples

[torn cliff]

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

This is how you inpaint.

[inland tide]

another thing why is it being realist when that was the oppsite of what I wanted

[torn cliff]

That's possibly due to the model you're using

[inland tide]

is there a good anime model

[inland tide]

like have any examples

I want to a have a alittle bit of a dark senery but with a light anime character

[torn cliff]

https://civitai.com/models/64548/fcanimemix-fc-anime

This might be a good one for anime style

You can also use Loras to get it closer. If you download from Civitai, check to the right side of the images, it will say "Type" and beside that it has a letter "i" inside a circle. Click that and it tells you what to do with that file.

[inland tide]

what should I search then

[torn cliff]

Anime or other specific things relative to what you want to get.

Anything marked as "Checkpoint" is the base model. Anything marked as "Lora" or "Lycoris" is basically like a theme.

[inland tide]

is check point the base so will it always be what it will be

[torn cliff]

You can change the checkpoint in the box at the top of the image you posted where it says "model.ckpt".

[inland tide]

ok

[torn cliff]

Loras and Lycoris can change the results you get somewhat, sort of like the checkpoint is a model doing a photoshoot while the loras are the clothes the model wears.

[inland tide]

what folders do I put the checkpoints in

@torn cliff

[torn cliff]

If you are using Automatic1111 then it's: stable-diffusion-webui-directml\models\Lora

Oh wait nvm

You said checkpoints.

Should be the same place the bat file is located

In a folder called Models

Cant say for sure though cause I am not sure what version of UI you are using

[inland tide]

https://civitai.com/models/6891/dark-style how does this look

[torn cliff]

Looks fine if you're going for a dark theme

[inland tide]

that's what I'm doing but when ever I do it includes a bug

[torn cliff]

What bug?

[inland tide]

a inscet

[proven jasper]

sounds like a skill issue

[inland tide]

well I'm trying my best ok

[digital dirge]

hi friends [img2img] I want this image to be in steampunk comic art style however I can't seem to figure out prompt or checkpoint model for it 😄 can anyone help out?

I do get it to be in steampunk style however the art style doesnt change

something like this however the art style should be different

[torn cliff]

You will need to find a model/checkpoint that has the style you want. For example you can prompt "Realistic" all you want but if the model you are using is an Anime style model then it's not going to help much.
Maybe this:
https://civitai.com/models/54009/pirsus-comics-style
Might help

[ornate turret]

anybody knows how can I change tiling to false?..

[gaunt shuttle]

It's either here

Or remove 'Tiling: true'

[ornate turret]

Thank you ❤️

seems like the new interface is different and I cant find it

[gaunt shuttle]

Np

[ornate turret]

the UI lacks the tiling option since I've updated. where would I set 'Tiling: true' to false? do you have any idea?

removed it from settings under "stable diffusion", thank you!!!!!

[gaunt shuttle]

Hi, I have a question:
How do I generate black hair? I never seemed to have a problem with that but now most of the images sd generates have this weird effect where hair is colored differently on one side. It's either green because there is green eyes in prompt or blue without any blue color specified. 90% of images have something like this across 5 different anime models
Do I have to inpaint it every time or is there any way to prevent this?

[silver valley]

Your missing a vae file. Its for color correction

[gaunt shuttle]

Thanks, I will get one then

[gaunt shuttle]

I got a vae, it works but and the problem still occurs
Should I get any specific one?
Is there something else I can do?

[silver valley]

That could be a specific prompt tag that triggers that or its model depended

What are your hair prompts ?
Also a trick. Describe the hair color before describing the eyes

[rigid wind]

I'm a bit confused if this is the right channel for it or if #🔧｜finetune is the right place for it but I am having a hell of a time with inpainting. Everytime I try to add something to my initial generations it never quite fits in like I want it too. Like this image I started maybe like 3-4 days ago. The initial generation came out fantastically, I really liked it (just in case you can't tell what was in the initial generation I added the frogs, reeds, birds on the head of that cat and the "caterpillar" in the top right). I've watched a couple of videos on inpainting and found some parts where I was making mistakes but even after all that I can't quite get things to feel like they fit in. The way I've gone about adding things after watching videos is by inpainting the section I want to add something, then use a high denoising strength (0.9 to 0.95) with a lower resolution on masked only (in this case from 1920x1080 down to 1080x720) with latent noise on which does a pretty good job of getting the actual object I want into the image, then I will find a version that has sort of has what I want I will go over the same area with a lower level of denoising strength (0.4-0.7, depending on how close it is to being what I want) with original masked content, still at 1080x720, then if I'm happy with it overall I just want it higher quality I'll turn the resolution up to 1920x1080 and a really low denoise strength (0.2-0.3) but as you can see it never really ends up fitting in

Basically I feel like I've done nothing but make this image worse since trying to make it better xD

This was the initial generation

I'm also quite new to stable diffusion (I think this is like 10 days or so since I've started) so it might be that I'm doing something really dumb

[tawny beacon]

Hey everyone, I used dreambooth to train on photos of my deceased mother in law. I want to use SD to help recreate the memories my children have of their grandmother. I want to print these photos along with a short story in a small book for each of my children. Stable diffusion does a good job drawing my MIL’s face but I find it difficult to get the poses I need. One is based on a funny story where she fell asleep sitting as our cat sat on her head. Any ideas on prompts to have it look like a cat is sitting on someone’s head? I get giant cat ears at best. This is using in painting btw

[gaunt shuttle]

(masterpiece, best quality:1.2), beautiful woman, (black hair), straight long hair and bangs, green eyes, pale skin, black dress with sleeves, frills, black pantyhose, earrings, forest
Negative prompt: (worst quality, low quality:1.6), belly button, badhands4, watermark, signature, <some adult content prevention here>
Steps: 35, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 3818822201, Size: 512x768, Model hash: 428a35ae0d, Model: rmhf_animev4, Denoising strength: 0.5, Clip skip: 2, Hires upscale: 2, Hires steps: 20, Hires upscaler: ESRGAN_4x, Version: v1.5.1

[amber fable]

Hey peepz,
Is there a website where i can find different workflows? or maybe some educational videos on youtube u peepz can suggest?
thanks in advance

[rigid wind]

https://www.youtube.com/watch?v=4u-Ytioi3DM

I watched this yesterday or the day before and it had a lot of points in it that seemed to make sense, but I'm still very new so idk if this is like the ideal way of working but it's relatively short and gets you into making some decent stuff.

[silver valley]

Ah okay can you try with clip skip 1 ?

And move green eyes to the end of the prompts. Not near your hair tags

[rigid wind]

anyone have any idea for where I'm going wrong?

[agile haven]

Im trying to get more realistic proportions on a female character. right now the hourglass figure is very exaggerated. I want it to be a bit more realistic, closer to the red line. anyone have some good keywords for this? nothing i tried seemed to work

[static hinge]

Really don't know if it will work, but maybe to try with thick, bulky, robust (or inpaint that part with that), also try some negs like (hourglass figure:1.4) or Curvaceous figure, Shapely figure, Feminine figure

[upper inlet]

is there a prompt better than large breast?

I want even bigger

[sick anchor]

So I spent quite a bit of time trying to get wet hair but not a wet body. The single positive wet hair works great for the hair but also wets the body. No amount of negative or positive or combination will get rid of the wet body. When it does it also removes the wet hair. Any suggestions?

[silver valley]

You can try Soaked hair

Or try an other model maybe
Also mention the hair after the clothes can help

[sick anchor]

I have tried that also. No luck. It is also already mentioned before and after cloths.

It seems I will have to find another seed. Although the only situations it does not seem to affect the body is in a closer up photo but that is just because the face generally does not get those water lines.

You can still tell it wants to put water everywhere. Edit: adding waterdrops to negative with a 1.3-1.5 weight helped in some situations but not most. It did reduce the visible drops but left many trails which is still much better but not what I wanted to see was possible. It seems using a lora will be the only way to achieve this properly as any amount of water removed from the body and skin results in dryer to completely dry hair like all of it is tied together. It happens in every sd model I have tried. Even taking the prompt back to the basics does the same for me. Trying to use a prompt to override the wet skin with another texture like sand also did not help as that only resulted in sandy wet skin occasionally. Thinking the embeds or lora's I am currently using are the problem I removed those and it just results in a less detailed but still very wet body in the photo except this time a far more opaque viscous substance that looks just like what you are thinking.

[wise belfry]

when we get those artefacts what does it mean and how can we avoid it?

[agile haven]

Any help on getting rid of the cloth on the arms and legs? i tried (sleeves:1.5), (leg warmers:1.5) in the negative but it didnt improve it that much

[tired vigil]

Sleeveless

You can also try "showing part of the body"

Or just plain prompt for the body part, also works for the models I use at least

[agile haven]

ohh thats a good suggestion thanks

[quiet zodiac]

@wise belfry Try another seed.

[torn cliff]

If you send it to img2img and mask off the areas you want to change, like just the legs to start with, then only put in prompts relating to that, like "bare legs", then it might help

[wise belfry]

shouldnt it however be okay with all seeds? also i thappened with many seeds.

[orchid scaffold]

Hey guys, is there a good prompt for getting "full color" when doing img2img with a black and white image

adding something like colorful seems to make it colorful but not the right colors

making human skin blue for example

[orchid ore]

@wise belfry what is your setting? Try lower cfg for example. More steps. But one dont know without knowing what settings were used.

also probably reason people are using best quality tag and in negative poor or low quality. Is settings in image?

[wise belfry]

hmm? yeah used it in the images prompts

[orchid ore]

o.k. and those things still remain, what cfg are you using?

[wise belfry]

the usual is 6-7.5
steps 30 usually for the default ones in comfui and others

[orchid ore]

should be o.k.

[fathom spruce]

What are some good settings to change the color of clothing items on inpaint? I keep trying to up the intensity to 1.5, and denoising strength to varying strengths but the color refuses to change

[neon briar]

Could you share the prompt and the model used to generate this image?

[quiet zodiac]

@wise belfry No, some seeds are just bad, and no amount of prompting can fix them. That's why batching a bunch of images is the way to go. Pick out out favorites from the generated candidates.

[rigid geode]

I need some help trying to fix this image.

Here are my prompts.

Regular Prompts:

[Candid:1.4] ultra-realistic, 4k , photograph of a cute goth Asian girl, with long black hair that reaches her shoulders, black lipstick, black mascara, wearing black leggings that hug her curves, wearing loose white t-shirt, gazing into the distance [full body framing:1.5], with brick wall in the background, natural lighting, shot from the front, eye level, picture taken with Canon EOS 5D digital camera

Negative Prompts:

<bad_hands_5><bad_prompt>, ((((ugly)))), (((duplicate))), ((morbid)), ((mutilated)), out of frame, extra fingers, mutated hands, ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed))), ((ugly)), blurry, ((((((low resolution)))))), ((jpeg artifacts))), ((bad anatomy)), (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))), out of frame, ugly, extra limbs, (bad anatomy), gross proportions, (malformed limbs), ((missing arms)), ((missing legs)), (((extra arms))), (((extra legs))), mutated hands, (fused fingers), (too many fingers), (((long neck)))

348x768 (width and height)

35 steps

Seed Mode: Scale alike

Sampler DPM++ 2M Karras

Clip Skip - 2

And I am using the Epic Realism model.

[tired vigil]

maybe 348 is too little ?

I always use that model with at least 512 resolution which I think is the minimum that SD likes

[rigid geode]

I could try 512x512

[wise belfry]

or you can use that one as the img2img for the next one with the same seed and increase the resolution while keeping the aspect ratio the same

with .30 denoise

[celest quartz]

anyone know the right prompt for getting that attractive V lines on someones waist, just under the navel? like is it hip bones? doesnt really seem to work. cant even find anyone that actually knows the proper name for it

[quiet zodiac]

@rigid geode I often add "natural pose". Perhaps that will drop her hands and allow your negative fixes to take place..?

[rigid wind]

try v-cut abs.

[noble charm]

Anyone know some good general purpose models for making memes?

something kind of like the old dall-e mini shitposting days, but more coherent and modernized?

[dapper harbor]

i've tried so many prompts and i can't get a single seed to show a sailboat crashing into a rock in the ocean

[tired vigil]

"sail boat crashing into a rock in the ocean, (accident, crash:1.4)"

Is working for me, not sure what you want specifically but some of the images are a sail boat crashing into a rock in the ocean :P

[dapper harbor]

thanks I'm not very good at this 🙃

[twin yoke]

P:

highly detailed, 70mm, IMAX, HDR, 4k```

**N:**
```(worst quality, low quality, normal quality, lowres, bad, oversaturated, undersaturated, overexposed, underexposed, grayscale, bw:1.4), (watermark, autograph, trademark:1.2) (blur, blurry, grainy), morbid, collapsed, cropped, missing, deformed, imperfect, conjoined, deformed, ugly, imperfect, skewed, unnatural, error, mutated, out of frame, cut off, censored, jpeg artifacts, glitch, deformities```

[past sluice]

Are there any tricks to roughly preserving a similar color grading through a series of images?

My understanding is there is some similarity between adjacent keys, but wanted to check if there is anything more specific that might be worth trying.

[tired vigil]

Styles 2 image?

Look it up, see if it's what you want

[whole olive]

Anyone know what model can best create something like this?

[rigid wind]

it's probably going to be more about your prompting/controlnet than being a specific model

[quiet zodiac]

Yeah, most models can handle that. Try something like stroke only, no fill. Also coloring book style might help.
clean line drawing of an elephant crossed with a zebra, coloring-in sheet style children's coloring book, no fill, just line

[past sluice]

There's a civitai lora called "anime pose sketch" that pushes it toward that style hard

[orchid ore]

I made this sort of storyboard.

[torpid bramble]

do y'all have some quick tips to make an overall better prompt? something like words that generally improves quality?

[quiet zodiac]

If you're talking about characters, getting them closer to the camera helps improve quality. I often use "detailed faces, realistic skin, natural pose" in every character prompt. There's a certain distance where SD 1.5 just fails and makes blurry faces. SDXL is better because of the higher resolution dataset, but SDXL images are all non-commercial.
If you add tokens (i.e. words) to request detail or realism, that often triggers a closeup, just keep that in mind.

[torn field]

Add --raw to your prompt to disable our backend prompt assistance

[orchid scaffold]

what are some good words for facial expressions?

[wintry oriole]

literally the "mood" you want followed by "faceexpression". like for neutral stuff I use "neutral faceexpression". of couse you can use "happy" or "sad", "smiling" and stuff, but with those you have to be a bit carefull if you don't want you character to look like a crack-addict. 🙂 negative promts like "open mouth" or other positive prompts like "perfect teeth" followed by a slight step-increase can do wonders here. good luck. 🙂

[ocean timber]

So, how do you tell SDXL to not put an object in the front and center of your photo, (in my case it's a spaceship) I want it a bit back in the distance and not taking up the full screen real estate on the photo

[pastel quest]

Hey dudes, how to draw group scenario decently like staffs in the large office?

[orchid ore]

@ocean timber if you are using A1111 there is extension regional prompt which can help you very. I believe it is as well for comfui, but not sure if easy to create workflow.

[ocean timber]

I am using A1111 yes

how is that extendion regional prompt used?

[orchid ore]

you make grids and tell what should be in what part of image. But some tutorial will be needed i think, it is not 1 click thing

[ocean timber]

Noted. I'll look that up later

[orchid ore]

https://github.com/hako-mikan/sd-webui-regional-prompter
this should help

[spark rune]

how multiple controlnet units works? does the order matters? if I have many of the same units, do they stack?

[spark rune]

this documentation is very good https://stable-diffusion-art.com/controlnet/
but still dont answer my last two questions

[spark rune]

yes. my bad 🤦 thank you

[pulsar summit]

question how do i make sure the negative prompt is working? im using txt2img and putting anime girl sitting down and i add skirt, socks, shoes in the negative but she still has them?

[violet ruin]

Can anyone recommend a model or lora for such detailed armor? (sorry if i ask at wrong place)

[silver valley]

Use barefoot in the positive prompt

[silver valley]

A good lora would be the more detail lora. Then prompt for: intricate knight armor, detailed, engraving etc

As model you could try Dreamshaper v8 for example

[chilly karma]

How do I make something like this

will well defined edges

and such high details

like the sweater, every fabric is visible

[chilly karma]

Best I can do is this, the details on the dress barely visible, and the whole image looks like a painting none of the borders are well defined

[crimson saddle]

Hello. I'm trying to create 1 pumpkin, but it always creates more than 1 image. Can you help me?

[thin creek]

yes

batch count 1

batch size 1

there u go : )

[tired vigil]

Have you tried those detailer loras?

[rigid wind]

So I'm trying to make a capybara driving a NASCAR. So I did what I always try first which is just typing in what I want to see and hoping it works. But that resulted in the first image. So I thought okay well I'll start with generating a NASCAR image and inpaint a capybara into the driver seat so I generated a NASCAR image (the 2nd image) and then when I tried to inpaint it it became real obvious that there was a problem, whenever I try to inpaint something into the car it doesn't seem to understand that I want something to look like it's inside of the car and instead will just inpaint it as if it's on the windshield. I don't know if it's a matter of me just finding the right denoising strength and mask blur or if there are some prompt tricks to help with it but some advice woul be great.

[solid bramble]

i am struggling to take a older prompt and put a new spin on it. the idea is to use this glass rose concept and turn it into a wine glass with the same style/looks. so essentially take the center pedals out, smooth out the sides and make the bottom stem flat ended. however this sounds simple but haven't got anywhere close to what i want.

[rigid wind]

try blending the 2 prompts together ie [glass rose: wine glass: 50], that might help

[solid bramble]

it's not that simple, to get the glass rose to work is a semi complex prompt.

[rigid wind]

You can do the same with a complex prompt. I just simplified it. You could also try using Dynamic prompts

Could also use controlnet to use that image as a reference.

[torn cliff]

Maybe try inpainting with "capybara behind glass"?

[kind kayak]

Are there any prompt that make model not generate character that her head is at the edge of image?

make she stay in the middle

or on the middle idk

[proven jasper]

you are probably using a height thats not compatible with the model

[kind kayak]

768

I used 512x768

and it's SD 1.5 model too

[proven jasper]

try to reduce the size of that giant negative prompt salad

[kind kayak]

huuuuuh?

[proven jasper]

too many words on negative

[kind kayak]

I'll try only text and watermark

[proven jasper]

try only with low quality,worse quality then whatever negative embeddings you use

[kind kayak]

FastNegativeV2

and badhandv4

[proven jasper]

yea the shorter both prompts are the better

[kind kayak]

btw

what can embedding do?

[proven jasper]

negative embeddings like badhand4 are for improving quality of hands and fastnegative are for removing artifacts or some deformities,they are not perfect tho

[kind kayak]

Horrible

[proven jasper]

your positive prompt sucks too thats why

[kind kayak]

How?

[proven jasper]

anime artwork is not necesary because its already an anime model that can only produce anime, remove these too: female focus,female,

[kind kayak]

A little bit better

[primal yacht]

image//Background: African village with huts, greenery, and blue sky.
Foreground: Joyful children playing with kites and traditional games.
Community Engagement: People attending malaria prevention awareness, distributing repellents, and healthcare support.
Theme: Hope, unity, and transformation.
Banner: "Join the Fight Against Malaria - Your Support Can Save Lives!"

[weary veldt]

Can anyone explain how Clip G and Clip L of SDXL works?

[chilly karma]

yes I'm using more_details to add details but with too high of a weight it messes up the whole imges

[obtuse torrent]

@chilly karma In your example image... I don't see any mention of quality or realism in the clothes,...you use a Lora, but it's one that adds details for the eyes.. It's basically a prompt that leaves a lot in the hands of the AI , it's not strange that she doesn't treat him the way you want...
If the prompt at least outlines what you're looking for, things like increasing the steps to the 50-100 range can consolidate and refine the details and style and get rid of some of the generation errors.

[compact maple]

what's a prompt to define low waist, low-cut pants/skirts? All generations so far is of pants and the like going above covering the navel. As i've not found a prompt yet that i can adjust that height.

[twin yoke]

lowrise

But your model would need to account for that.

[quiet zodiac]

@compact maple Try crop top or bare midriff

[lucid willow]

Hey, i need Help. I created this avatr using a low quality pic and would like to know how can i make him in various other poses and at a higher quality.
i also would like to know if its possible to make him with a chroma key background, so i can superimpose him over videos.

[tall prism]

To remove the BG, you can use websites like remove.bg

I guess

[tall prism]

Can anyone recommend some keywords for creating a comic book front cover

[orchid ore]

comic book front cover? Or rather comic book frontpage?

[chilly karma]

ok thanks I'll try that 😄

[dull beacon]

Hello !
I was trying to do an image generation about an anime characters from black clovers (Yuno Grinberryall)
And no matter what I do, I can't get someone who looks like him
The best I could get was that

I used a lot of prompts and I only got this with that prompt :

yuno from black clover, yuno grinberryall, a young man of average height with a lean build, gold or amber eyes, and messy black hair. His main outfit consists of a pitch-black shirt with a high collar and black pants

When I try to use an image as example I got a girl

[rigid wind]

I did end up trying this, as well as "capybara behind windshield", "capybara behind window", "capybara behind windscreen" and none of them worked. I ended up giving up on it and maybe when I get better I'll be able to do it. Turns out I misunderstood what my friend wanted anyways and they wanted capybara running around a nascar track. I am thinking that probably my big issue is my stubbornesss of moving off of Dreamshaper_XL10 model. While it creates amazing images I do find it's much harder to work with to get good results of weird things and I find it's particularly bad at inpainting. I was just trying to add a detective in a trench coat in the background of an image I generated last night and no matter what settings I used the best I was ever able to get was a guy that was missing feet and half of his body was missing. even though I put in full body and close up into negative.

[torn cliff]

Yeah DreamShaper is a bit better with realistic things.

[rigid wind]

I also probably need to get an actual inpainting model. Any recommendations?

[tired vigil]

https://civitai.com/models/103729?modelVersionId=111118

This guy?

Models most of the time don't exactly know the character you want, using a lora helps it

[dull beacon]

I used this model

But I don't get Yuno

Literally an other character

[dull beacon]

So I tried but I got nothing like him

[tired vigil]

@dull beacon did you add the Lora correctly ?

To the prompt

And the folder

[orchid ore]

it is not model it is lora. If you use it as model on other hand you wouldn't end with sensible image.

[dull beacon]

Oh I don't know about that

Can you explain what is a lora pls

[orchid ore]

ask @tired vigil

i used lora probably once.

you put that what you dl into proper directory, and in prompt you trigger it by word you should find in description

[tired vigil]

@dull beacon you need to put the Lora you downloaded in the right folder

webui\stable-diffusion-webui-directml\models\Lora

If you're using 1.6 I don't know where you go to activate the lora,
on 1.5 I click on the little picture icon under the generate button and select the lora, so it appears in your prompt like this <lora yuno something:1>

[dull beacon]

I have an old version

I downloaded it a long time ago

is that the version ?

I think I will need to update my stable diffusion

[tired vigil]

Did you try the lora

It should appear in the Lora section when you click on the little picture icon under generate

[dull beacon]

I don't know which one

There is no "lora section"

[tired vigil]

click on this

[dull beacon]

oooh

Ok

And then

that ?

But what do I need to put in models ?

[tired vigil]

use a normal model

not the lora

[dull beacon]

I only have these

[dull beacon]

Should I need to download one ?

[tired vigil]

Hmm I don't know what model to recommend for that black cover anime look

Try with rev animated

on the lora page you'll see these they help on the prompt

Trigger words

do a test with a prompt like yuno, black hair, yellow eyes, hair between eyes, fur trim, anime, lora:yuno:1

[dull beacon]

Ok thanks !

I will try

[sullen orbit]

Can anyone help me, is there any LoRa or prompt that creates this style?

[tired vigil]

Should be doable with prompt only and a general model

[sullen orbit]

that's the problem, knowing the prompt and the model 😭

[tired vigil]

@sullen orbit found a similar image on civitai
The prompt was
shadow puppet templates, kirigami, papercutting, shadow puppets, flat print, line art, monochrome, silhouette, batman, high quality, 8K

Maybe something like dreamshaper or realistic vision can do it

And anime ones but I don't know many :p

[sullen orbit]

thank you very much

[dull beacon]

@tired vigil it did work !!

Thank you

I just need to add some prompts to get a better image but it's going in the right direction

[cinder olive]

((exposed skin, Big Breasts, Big Breasts, NSFW, lewd, risque, nude, sexual,))

What other negatives can I use to tell the bot I want my character wearing god damn clothes.

it keeps giving me old MMO girl armour smh

[tired vigil]

try skimpy clothes/outfit
maybe full armor in positive ? Full something makes them wear more clothes for me like full dress

also thats funny the literal korean mmo armor

maybe that model really likes showing skin ? :p

[kindred minnow]

you could also try using the kind of armor you want in your prompt (and if you want good unique armor, AI probably isn't the best for that anyways)

[cinder olive]

I asked for scale mail

[cinder olive]

it got worse X.X

[tired vigil]

you could do like, chest plate, armor skirt to focus on things the model should show

[tired vigil]

which model are you using

[kindred minnow]

well you made sure to put what you don't want into the negative promt?

[cinder olive]

clearly the wrong one. It's a shame because it gives some amazing dynamic poses.

[kindred minnow]

also nice, that your character has 4 ears

[cinder olive]

never mind the ears, what's going on with her hand?

[tired vigil]

Try soft lighting

[cinder olive]

Switched to Mangled Merge, much better - on the clothing front at least.

[tired vigil]

@cinder olive try using dynamic pose on your prompt if the model does the boring poses, on my model its doing some crazy stuff like yours was haha

[cinder olive]

wut model 👀

[tired vigil]

From KizukiAI

I dunno if I can post the link here

just search that on civitai :p

it works better on 512x768 from what I know

[cinder olive]

ooooh the henty one?

[tired vigil]

This image yes

both are good

and similar

like this is the same model but anime on prompt, it does a lot of things :D

And I don't know how to use these "EasyNegative" textual inversion things

@cinder olive not gonna lie I'm impressed with how well "dynamic pose" is working lmao, im testing some settings while doing these

[cinder olive]

[potent coral]

Hi. Is there a way which words was trained a negative embedding.pt ot .safetensor?

[lucid willow]

thaxs, it works.

is there a guide so i can generate image of my avatar in different poses?

[vital jolt]

looking for a prompt method to batch produce images with flowers so that each iteration has a different flower. How do I do this?

Nevery mind. Figured it out.
{ cherry blossoms | carnations | gardenias | sunflowers | poppies }

[tacit kelp]

AI绘画 生成

[tired vigil]

Yes

It's not perfect but

Search for stable diffusion reference only on YouTube

[buoyant axle]

Quick question -- when using a lora, do I need to specify both the lora's triggerwords, and the <lora:file_name:weight> syntax when using automatic?

[tired vigil]

the triggerwords help triggering the lora

if needed, most of them need

and the <> stuff is what activates the lora to be used

[manic hill]

Where are the Pad prompt/negative prompt to be same length and Batch cond/uncond settings? CTRL + F in the Settings tab didn't show them.

[lucid yew]

Hello, I just made this from a picture using controlnet, anyone have any idea how can prompt it such that the grid goes yellow and electrified like hot golden electrified and moon becomes yellow electried too

I tried using the recolor control type but that dosen't change

[static hinge]

what are some good prompts for making non-regular humans, I mean just freaks or weirdos. Like high fashion zombies or undead

[lucid yew]

just use the regular human prompts in the negative prompts lol

[static hinge]

I will try, but it is that easy?

[lucid yew]

probably..make sure to put the non regular characterstics exactly in the positive prompt

[static hinge]

you mean in the negative?

or how exactly positive and negative prompt?

[orchid ore]

what you want in picture positive what not negative. You want irregularities so in positive those

[sinful oar]

It helps to put the key tokens in the prompt usually, but sometimes you can get different results with and without them so try it both ways

Same for the location, at the beginning or end

[spark rune]

about the pose part, you can use dw open pose to generate the image exactly in the position you want

https://www.youtube.com/watch?v=30b2k1p2CiE&list=PL-4VJa3jNVK3J5QA42Pc0lUU3tibV4rGK&index=5

if you cant find the right image to use as a source, make your own at https://webapp.magicposer.com

[spark rune]

but yeah, to bring your character into the custom pose this might be the way to go

the trick might be to use 2 control units, one for dw openpose and other for reference only

if Im not right, someone correct me pls

[lucid willow]

this is super helpful, thanx

[clever shadow]

How possible would it be to take an existing Forward facing character model, then generate a left, right, and rear facing version of the same or very close to the same model? If that can be done would I use a combination of controlnet/openpose with img2img or is there some other way to go about is? I basically need a character turn of a character already generated. Ive been at it for days, starting to tear my hair out..

[weary sonnet]

The middle image is te original and the first is the crappy output I get from sdxl on replicate.com. The last one is a great response from the model using Bot 1 here on this server. My question is though, both prompts are exactly the same. Does anyone know configuration of other variables to get realistic outputs from this. sdxl model as compared to whatever I'm getting out from replicate

[clever shadow]

Check out civit.ai and filter models by photorealistic. Pick one then run the prompt. Careful when going to civitai as they also allow nsfw models

[tired vigil]

Have you tried Reference only ?

controlnet stuff

[clever shadow]

havent heard of reference only, i was using open pose but the whole image dimensions thing was just a nightmare wiith distortion, if you have any advice towards the reference only concept, please im all ears

[tired vigil]

Its an extension like open pose, it basically uses the input image as a reference, in your case it would be the same person
look up some youtube guides on how to download it

[clever shadow]

Yeah thank you. Im getting much closer! Step 2 is 3d mesh. Still going uphill it seems lol

[uneven lantern]

Someone can explain me why I got this with img2img prompt with 0.95 denoising strength (so really close to original picture on the left) ?

[twin yoke]

Because 0.95 is almost re-doing the entire picture. You'd want something like 0.2 or less to stay close to the picture on th eleft.

[crisp lily]

While I've got no proof other than getting more equivalent results, I suspect 'No Style' in the bot adds the 'Enhance' style from the list of style prompts that went out ages ago which had the bot / clipdrop styles and 'Enhance'.
which would add ... "breathtaking {prompt} . award-winning, professional, highly detailed"
and a Negative prompt off "ugly, deformed, noisy, blurry, distorted, grainy"

[uneven lantern]

Any way to simulate this easily ?

[tired vigil]

can you explain it in words ?

[stable sigil]

not sure where to ask this but how could i go about converting a large batch of image files into openpose files? i can manually "detect from image" in the OpenPose editor but I wanna do this automatically to each frame of a video (200+ pics) so it's not viable.

[upper inlet]

I need some help dunno if anyone is around

[tired vigil]

Just say it

[plain mulch]

Hi. I'm currently using an image2image process and struggling with some prompts. I have various close up clips of people holding hands, 5-10secs and I want to change some bits into wedding cake models of the hands, some as knitted toys of the hands and some made of pebbles. How do I use the prompt to define the person or object that needs to be made "new Ai real" and not stylise the whole image with the prompt? As in, making a crochet blanket portrait of the hands, instead of making a 3d knitted toy of the hands. I am an animator using Diffusea in After Effects, some Stable Diffusion webui and Discord server apps but I am no expert at Ai use, coding or prompts. Any help will be much appreciated.

[twin yoke]

After that, explain it in interpretive dance and post the video.

[tired vigil]

[topaz terrace]

What prompt could I use to get a very close result to that? Im not able to get the hand like that even though I'm using pose to image

[static hinge]

closest I got in the first tries, then I wasn't so lucky

portrait bust close-up, a young light black woman with a fist up in the air with a yellow background behind her, white t-shirt, a stock photo, short hair, subtle smile, professional photo

[topaz terrace]

That's pretty good! I got results pretty close to that one. But I'm trying to get the hand as in the same position as in the picture, do you think something can be done about that? I'm using pose to image.

[static hinge]

I would try "knocking door fist gesture", or just use some hand of the depth library to inpainted it

[topaz terrace]

how is that about the library? I have 0 clue

[static hinge]

is this extension https://github.com/jexom/sd-webui-depth-lib but last times I tried it was kind of buggy and messed up Automatic1111. You can just download the hands depth maps, go to an image editor, open the base img, paste the corresponding depth maph, pose it (flip it, rotated it, scale it), then use it in ControlNet Depth

[topaz terrace]

thats awesome, probably will do that, there is any video about the addon or something that I can look for myself?

[static hinge]

you can just download this depth maps and use it externally https://civitai.com/models/67174/900-hands-library-for-depth-library-or-controlnet

[static hinge]

https://youtu.be/Oqf-93TpykI?si=b1EgkNoWV0CiD7Wg

there area some more videos

[topaz terrace]

Thanks for the answers!

[static hinge]

only thing it takes some time to understand which depth map is the one you need, as sometimes they are flipped or in some other angle, but in this case it seems pretty straight forward

Really only non frustrating way of fixing hands imo, it takes some time to set up

in this case this may be it, not exactly but for starters

[topaz terrace]

yep thats basically it

[static hinge]

well after some general inpainting tries and the some corrections it looks decent. I hate inpainting, practice is good I think

I'm a masoquist so I inpaint with SD models, should try with SD 1.5 as you have ControlNet inpaint that has even more power

the arm is kind of messy, hand is in a strong fist. Another way would be to look for a similar picture, cropped in the same way as with the depth maps, then use pre-proccesing. I realize I f%($( up as I use the fist depth map but didn't disable the ControlNet preposesor (you don't need as it is already a depth map). But it kinda worked, maybe it would have worked in the first tries like that.

[static hinge]

well I actually got a better faster result inpainting with SD XL and ControlNet depth. With 1.5 it delivered the pose but with poor quality, and I had to make some final passes with SD XL. So I don't know, that's why I try to inpaint with SD XL... the hands are always a no no (I use various neg embeddings and a "good hand" lora for 1.5, cyberealistic inpainting)

I almost got the finished hand with SD XL

But maybe I prompted wrong, I used "girl's han" and I should have use "woman's hands/fist"

But well need to practice so I can get it done faster, there is so many steps when inpainting, the most annoying, changing the img2img resolution when you send img to inpaint (I forge that may be avoided if you just drag the img to inpaint)

LOL I though another way, so you take the base img with good seed (the original with the fist up but not in the right position), send it to txt2img to re-generated, but this time using ControlNet depth, so you force it to draw the fist there and as you use the same seed, sometimes it delivers something similar to the original image, sometimes it changes the pose enterirely, it always respect the fist position, in some way.

In a way better than inpainting afterwards as it merges the pose better, but the fist also needs inpainting soo

In fact maybe in some cases better with other seeds and just pose the hand where you wanted, play with ControlNet Depth weigth and ending step control. Also maybe add another node with OpenPose

[night sphinx]

guys how do i generate anime school uniform without the bowtie or necktie, i even strengthen in neg prompt but it still trying to make the bowtie

[torn cliff]

Are you using a Lora? If so the tie might be part of that

Anyone know good prompts to use to get rid of an extra finger in inpainting? To just make that space part of the background

[night sphinx]

no, it just consistently makes shirt and then includes a tie

[torn cliff]

Maybe try finding a lora like that. If you know of any animes that have the kind of uniform you want you could look for a character from that.

You could also try inpainting, mask over the tie and try prompts for collar and/or buttons

[ivory dawn]

Hi. any idea how these were created? There are thousands of these. My guess is stable diffusion? Trained model? I'm relatively new to SD and try to figure out how to create something similar. Reminds me of someting cartoonish, popart style-like but it can't just be a simple prompt since they are all relatively similar and there are so many.

[torn cliff]

I'd guess its done with Prompt S/R since the backgrounds in each row are similar colors

[ivory dawn]

hmm yes could be. Any ideas on the prompt? And what model was used?

[torn cliff]

No clue on the model, but I kinda think it may have something to do with Sonic the Hedgehog

[ivory dawn]

oh yes absolutely. and it reminds me of 1940's/50's cartoon or something like that

[torn cliff]

Yeah, sort of looks a bit like the early Simpsons style

[orchid ore]

cant it be just lora? I think it will be very difficult to find out.

[tired vigil]

it looks like a mix of the frog meme and sanic

try putting those 2 loras together and get those colors if you want them through prompt D:

Pretty cursed image

[topaz terrace]

Damn, thanks alot for all your tries and help!

[ivory dawn]

Yes agree. How do I put 2 loras together? Also aren't loras something i download? So which specific two loras do you mean?

[tired vigil]

Well just search them on civitai
frog lora https://civitai.com/models/19006?modelVersionId=45722
Theres this sonic model https://civitai.com/models/1493?modelVersionId=96182
And this sonic lora https://civitai.com/models/82047?modelVersionId=87105

to use 2 loras together you simply put both on the prompt and try it out :p most of the time I try multiple loras they work, fiddle with the strength on them and they should adjust

you might want to take one of the images in that image you sent and use it for img2img so its easier to reach the desired outcome if you want to copy/make more of those

[wintry oriole]

Heyho - Hopefully someone can help me with my hairstyle-problem... 😅 As having english as my second language I struggle with defining different hairstyles in my prompts. Does anyone has a good resource for some basic styles? When I google that stuff I end up in various fashion-sites which wanna tell me the newest trends, but that is not what I'm asking for tho. Thanks. 🙂

[ivory dawn]

Thanks for your help. Will try that. And if it's not working i will try to train my own model

[orchid ore]

@wintry oriole
https://en.wikipedia.org/wiki/List_of_hairstyles
how abou this, or more in depth?

[wintry oriole]

Nope - That helps a lot! Thanks. Is it best to combine it with "hairstyle", like "Mullet hairstyle" for example or are there better ways?

[orchid ore]

oh, i dont know, you have to test it 🙂

would start to use it with hairstyle

[wintry oriole]

will do ^^
I need it to build my charactersheet workflow ^^ need to create several consistant images and I'm struggleing with noses and hairstyles... 😄

[winter cloud]

How prompt for audio

?

[wintry oriole]

try out this one: https://github.com/riffusion/riffusion

[modest bronze]

How can I position objects in the locations specified by the prompt, I mean absolute location or relative location such as left side, right side, upper left corner, and bottom right corner?

[orchid ore]

there is extension regional prompt, that can place object in premade rectangles. In A1111 way easier, in Comfui, not sure how well nodes works with it.
Just some time is needed to learn it

[modest bronze]

thanks, I'll try it later😆

[spark rune]

hey yall. Im looking for some new good negative promps to decrease poor generations. My current negatives are the ones bellow. Any good additions?

(bad-artist), (bad-artost-anime),(bad-hands-5),(bad-prompt),(bad_prompt_version2),(easynegative),(ng_deepnegative_v1_75t),(worst quality:1.2),(low quality:1.2),(lowres:1.1),(monochrome:1.1),(greyscale),multiple views, comic, sketch (((bad anatomy))), (((deformed))), (((disfigured))), watermark, (blurry), (((strabismus))),(wrong finger), multiple_views, mutation hands, mutation fingers, extra fingers, missing fingers, watermark

also, Im using them on ReV Animated checkpoint. Would they work in other checkpoints like DreamShaper or even v1.5 pruned emaonly?

[quiet zodiac]

@spark rune What model are you using? A lot of the new ones are much better at generating quality images. The fact you have multiple views in your negatives a couple of times means you might be over driving your resolution compared to the model resolution you're generating from. That can happen if you request too high-res of an image.

[spark rune]

ReV animated. But I use Dreamshaper in some generations too

But yeah, these negatives were made with ReV in mind (found in a reddit post)

[quiet zodiac]

If you check the model card on CivitAI...

512x512
512x768
768x512```
Try sticking to those resolutions, or choose another model.
It looks like the author has quit maintaining it.

[spark rune]

ok, thanks. I will look into it

but anyhow

I want to learn how to make better negative prompts with any model. Is there any good documentation, reference or tutorial on how to get started?

Yes, I know prompt engineering is much trial and error

But there must be some kind of guide to ease the learning curve

[tired vigil]

Hello! im trying to make clothing designs with ai. But i had some problems prompting. I want my designs to be like stickers, 1 whole image that i can put on center but whenever i try to prompt it as i shown in image i only get pattern like images

Model im using is sdxl 1.0

[tired vigil]

I'm liking this
(KHFB, AuroraNegative:1.2), negative_hand-neg, (Worst Quality, Low Quality:1.2), border, grayscale, (watermark, multiple views:1.2)

You might have too much going on in the negatives

They do work in other models, as far as I have tested

I use dreamshaper sometimes too

[tired vigil]

Multiple views might not be necessary for you
I have it on 1.2 cause the model I'm mostly using is pretty bad not doing multiple views xd

You can find khfb and Aurora negative in the aurora checkpoint page in civitai

Try it out :d

[static hinge]

Here's some and I think it is a good, liberating practice (for SD XL) https://civitai.com/articles/1826/sdxl-prompts

Sometimes I check the recommended negative prompt from the model creator, but most ot the time I leave it blanck and use it for what is it meant for, I mean, when it is easier to prompt a negative than a positive. Like, I want a circus, don't want elephants, possitive: circus, Negative: elephants

Then recommended negative prompts make sense, if you want a realistic photo: painting, drawing, sketch, cartoon, anime, manga, render, CG, 3d, if you don't want garbage text watermark, signature, label

For SD 1.5 I use some negative embeddings you can find some on civitai

[finite ginkgo]

hey can i get some help trying to write ridley from smash as acruratley propmted as possible? like Emanciated? inhuman? humanoid kinda difficult

[cunning compass]

Guys, help me create similar models, as in the links below, I really ask
https://ru.pinterest.com/pin/24206916743368705/
https://ru.pinterest.com/pin/69876231720390056/

[static hinge]

anthropomorphic (maybe lowering the weight if it is too human), Skeletal (may be bad)

[spark rune]

I will try it. thanks

[spark rune]

ohhhhh that makes sense. Im definitely overdoing on negatives then

[static hinge]

yeah everything you add kind of hinders the generation directly or indirectly

[chilly karma]

https://www.instagram.com/kapibagi/

Anyone know what Checkpoint this guy might be using?

and how do I create this type of art?
Which is anime style but more kinda realistic or high budget anime

This is what I got when I tried but the lighting's off and kinda cartoonish

[opaque notch]

Hi, I am trying to turn this pixel art spaceship from starsector into a higher res version but all I am getting is blurry images or unrelated things, tried img2img, extra upscale and controlnet. Also tried finding a model for spaceships but couldnt find one. Kinda lost with this one. Anyone got tips?

[tired vigil]

The face looks like it's unpainted to look more anime

[proven jasper]

looks like breakdomain

[cinder olive]

I'm trying to get a porcelain mask but it just keeps giving me covid masks smh

[still hawk]

Have you tried using something like (porcelain:1.3) or something like that?

Maybe if you say face mask made out of porcelain

[tired vigil]

how can i generate picures ?

[still hawk]

Well you can download stable Difussion or find a website or app that allows you to make ai images

[tired vigil]

i cant generate it here?

[still hawk]

Go to the bot channels

[tired vigil]

That is probably way too small
And use controlnet tiles for it, I have upscaled 78x78 images into 1024x1024 before and it keeps the same look

Your best bet I think

Also use ultimate sd upscale

I can do that for you when I'm not in bed being lazy

[tired vigil]

@opaque notch

The original is just so tiny and looks blurry ;-;

best I could do

[chilly karma]

any idea what model this guy might be using?

[tired vigil]

It does looks like breakdomain

[chilly karma]

must be a merge with chillout mix?

and what about the lighting?

it looks pretty well lit and not too over exposed

[tired vigil]

could be a vae or some lora, hard to tell, try the model for a bit and see if you like

not sure where to find it

[chilly karma]

you mean breakdomain?

I already got it 😄

lemme try it out

[opaque notch]

WTF how? best I managed to do yesterday was this

thats really great thanks

[tired vigil]

I just said how you can do it :<

[blissful sapphire]

So I know brackets [ends at this::10] and [starts at this:5]. but can I also do some think from step 5-10 for example? [ [from to ]:5::10] seems not to work which is not surprising

[tired vigil]

Glad it helped

[opaque notch]

ik i am just surprised 😄

what model did you use? I tried the one spaceship lora I found which did make it less blurry but it also really changed the image

and was it in txt2img or img2img? (I am guessing the latter)

[opaque notch]

Welp, it do be higher res, lmao why is it so hard to find a model without women in it 😂

[opaque notch]

Welp It's better then it was but I am still far

[stable valley]

looks like classic midjourney to me, not SD

[plush sphinx]

How should I structure a prompt to get a view of a character from above and behind? I'm trying to make a character image for a mobile game, a fairy with a wand, but most of my results only show the front.

Cartoon, cute, colorful, mobile game sprite, whimsical, illustration. Magical fairy with wand. Top down game character. Top down view, birds-eye-view, high-angle view, from behind

Negative:

face, background

I'm getting really really good pictures, but only from the front.

Using SDXL through Dreamstudio right now because I'm not at my computer

Alternatively, should I use image to image?

[tired vigil]

Dreamshaper

[small pulsar]

I'm new to Stable Diffusion and from my understanding: interrogate Deepbooru work best for anime art, and CLIP for realistic art?

[tired vigil]

Excuse my poor pictures

my prompt was: looking at the horizon, view from above, view from behind, girl, fairy, wings, wand in hand, forest
and negative looking at viewer
The rest are TIs and quality stuff that I always use

Specially the looking at the horizon geerally works for me

try that

[tired vigil]

I read about that and you're right

I never used them before

[small pulsar]

interrogate to extract prompt from imagine, helpful to build a correct prompt.

[minor harness]

Is there any way to upscale images in SD while using Discord?

[rigid wind]

Is there anyway to get SD to create images that are only black and white? I don't mean greyscale

[static hinge]

most easy straithforward is to look for an artist that produce images like that, in the most similar style, and prompt it

also try to reinforce some positive prompt "black and white", "color" in the negative, something like that

[tired vigil]

no idea what I'm doing but this is pretty cool
using this lora https://civitai.com/models/551/black-and-white-design
prompt with: (black and white image, monochrome:1.2)

I'm not sure if it can do normal looking images but black and white, like these don't have background and all that, I did try putting river in the prompt

[potent coral]

please interrogate this clip. Is there a way to ENTIRELY BLUR THE IMAGE?

Stable Diffusion REFUSES to blur the entire image, it keeps focusing the face. Even in negative promppts "focus face" or prompts like "blurry face", it will not obey.