📝｜prompting-help

[sour ether]

you can try inpainting or outpainting, cause i think the dog is somehow getting confused with the blanket

cause in the youtube tutorial it had a portrait, and changed settings

[delicate arch]

Try a more clear picture too

[sour ether]

so maybe have to play around

i see the background in the youtube video barely changes also

of the potrait

[delicate arch]

@sour ether You heard of ranbooru extension?

[sour ether]

nope

[delicate arch]

Ah, nvm then

[silver valley]

In the settings

[delicate arch]

It's Restore Faces right?

[open cipher]

I will look into that one, thanks 😄

[sour ether]

oh it gives tags for the model

[sour ether]

thanks

[delicate arch]

Still comes out blurry

[sour ether]

yo anime waifu

[delicate arch]

Oh

My resolution got reset

[sour ether]

your anime waifu pretty good resolution

[delicate arch]

How many sampling steps is too many?

From far away, but when I open it, the eyes are not sharp detail enough

[sour ether]

yo it got an extra finger

[silver valley]

Don't use 1.5 EMA pruned. Try Dreamshaper v8 for example.
Then use a lower denois

[sour ether]

6 finger waifu

[delicate arch]

Lmao, I just noticed that

[silver valley]

Over 50

[delicate arch]

Fingers and hands always come out wrong

[silver valley]

30 is mostly enough

[delicate arch]

I even got a safetenor called Perfect Hands to use in inpaint, but the results are very mixed

[sour ether]

i mostly do simple prompts anyway to make it somewhat realistic or yolo with one word

[delicate arch]

I load up more negative prompts than positive

[sour ether]

ah makes sense

[delicate arch]

Yet the hands still come out horrid

I literally have ((poorly drawn hands:1.8)) typed in negative prompt

[sour ether]

lol

[delicate arch]

Or I get images like this

[sour ether]

lol

[delicate arch]

It's like rolling a dice, I will get amazing stuff and then 10 more generations they're bad

[sour ether]

did you check the lora works for the model

sometimes the lora is only for a certain model

[delicate arch]

Uhh, no clue tbh

I'm using these two together:

https://civitai.com/models/8685/eyesgen-lora-wip

https://civitai.com/models/6755/cetus-mix

[sour ether]

it says 1.5 on the lora and cetus mix is 1.5 so yeah idk

[delicate arch]

I'm gonna try lowering the weight to 0.5, recommended by the author

[sour ether]

i'ma try using it too

[delicate arch]

The quality has gotten better

But the hands are still bad

@sour ether Anything good for you?

[open cipher]

OMG changing the resolution make a whole world of diff

[remote badger]

I'm getting much better results with a different picture and the Dreamshaper model. Thanks for the tip!

[hexed shoal]

hello any idea why copying the exact settings and seed are giving me different results?

I'm not new to SD but I haven't been active in the past year finally have the time to get back to it

I didn't include it in the screenshot but I am using the correct model, clip skip and have installed the Textual inversion easynegative

so really I have no idea why my image is so much worse

is it possible there are other settings not included in the "Copy Generation Data" on CivitAI

[woven willow]

Any suggestions as to how I can inpaint this better so that the woman is holding liquor bottles in each hand? Here's the best I've gotten so far out of a few hundred generations, and the base image, with the Comfy workflows. Never been able to get the inpainting down.

About 50% of the time it makes her hold boxes

[silver valley]

Yes, all the settings are stores in the "Copy generation data"
The image for example is upscaled and yours not

[honest flare]

guys

how do i use BREAK and what do it do

[woven willow]

or even just regular hands at this point

[vital jolt]

SDXL Lineart using Redmond Coloring Book. I like what I'm getting, but I want to get rid of the blocks of black area that are generally heavy shadow and high contrast. Any ideas?

[dull portal]

Prompting help on getting feather wings instead of what looks like from a statue, the arm blending into the wing lol

[fresh pollen]

yo, dont really know where to ask but im trying to inpaint a insect on the wall of a picture someone took, however it looks very out of place not too sure what settings to use etc

[vital charm]

I am trying to find out how to add a texture to the simple input image with the letters ABC.

I pass the input, i tried, control net, clipvision, depth , img2img to try to copy a certain style but it looks really bad.

Has anyone tried to generate decent text on SD ?

[thin owl]

Anyone know a good prompt for Dreamshaper_8 to make it compose hands correctly?

[remote badger]

You might find adetailer helpful. https://www.youtube.com/watch?v=Fef5ia8ly0o

[worn dome]

I wanna create this character from lego Fortnite

And I've got this prompt. Now I need to add something to describe SDXL the curious details of his head. Any idea?

[remote badger]

Have you tried lego hands or c-shaped lego hands?

[worn dome]

That's actually not a problem now. But his head is

[quiet raft]

If I have auto1111 installed, what is the extenstion called to be able to generate "pose" for generation ?

[silver valley]

Controlnet

Or controlnet editor for creating a pose

[quiet raft]

I do this right right?

The correct one is? 😄

[silver valley]

Oh you already have controlnet installed.
The other Extension is called
Openpose editor

[quiet raft]

Ohh, I tried that one just now. Would not work and did not send to the txt2img :S

[silver valley]

Then click save and put it into controlnet manually

There are multiple openpose editors

The simple one is the best imo

[quiet raft]

Okay nice 😛 Well, you got any recommend settings? 😄

There are quite a lot 😄

[worldly acorn]

ask for outline?

[silver valley]

Depends on what you want. To use controlnet you also need to download the controlnet models first.

[quiet raft]

I got that 🙂

[silver valley]

Then select preprocessor to None.
And model to Openpose

Then enable controlnet

[quiet raft]

I can't select a model? 😮

[silver valley]

Like I said you need to first download the controlnet models + config files from here:
https://civitai.com/models/38784/controlnet-11-models

Choose the openpose file, if you dont want the others.
Each file is 700mb

And the config file

Put them onto models/controlnet

Then restart the webui

[worldly acorn]

what does ((( ))) do
and is there like a guide to how to write current prompts
I might've tried to start too ambitious with 3 subjects in a picture ( a boy and a lynx on a fishing trip witha big catch)

[silver valley]

() gives the word more attention in the image

[worldly acorn]

ok that explains things

ooh yeah now it's actually composing stuff I want lol

I had the latin name (lynx lynx) and it kept wanting to give me only 2 lynxes in the pixture

xD

[silver valley]

You can also use (word:1.3) to give a word more strength

[worldly acorn]

ahhhh
nice nice

[silver valley]

Its like (((word)))

But don't go over :1.5
That will force it

Can lead to bad images

[quiet raft]

Quick check, looks okay+ 😄

wait, you mean control type: openpose and Prepocessor: none right? 😄

[silver valley]

Exactly

If the input image isn't already a pose. Then you need to select openpose as preprocessor

[quiet raft]

Why does the pose follow to the img? 😄

[thin owl]

Thank you, I'll check into it

[remote badger]

It's not helping me on my current image. 😦

I'm struggling with hands too

It's so slow just inpainting...I think I could use gimp to paste a hand in faster and it would look better...

[maiden geyser]

Does anyone know, could admins invite the Metadata bot to the server? It would allow people's renders to display their prompt with the mag glass reaction.

[grand tide]

I'm going to get flamed for this but the best way to learn is by making mistakes - Year: 2305. Cyberpunk anime drawing of five friends in their late 20s watching a plasma TV on New Year’s Eve. Two couples: an Asian girl with long purple hair and a data analyst with a cybernetic eye, a petite dentist with a pink dress and a Singh businessman with a turban and a beard, a dark-skinned Indian cyber security analyst with a caboodle puppy on his lap. They are sitting on a couch in a cozy apartment with neon lights and holograms outside the window. Medium detail, 2k resolution, full color

This just doesn't work the way I intend it

How do I go about refining the prompt so I can create a poster for me and friends

Any criticism is appreciated

[thin owl]

I can't get inpainting to work wtih dreamshaper_8, it gives me weird amalgamations of hands when I inpaint, its kind of funny. also, the color never matches.

[worldly acorn]

how do you deal with a picture that has multiple subjects? should I just be doing it part by part?

If I try to get a picture with a boy, a lynx and a fish they caught it keeps giving me fish-lynxes and boy-lynxes

[remote badger]

Let me know if you figure it out, I'm watching a video on fixing hands now. I might just copy and paste some hands in with gimp if I can't figure it out soon.

[quiet raft]

How can I add points to the OpePose Editors?

ALso, how do I know what point is what?

[thin owl]

you can turn them on and off on the menu to the left of the image, there is drop downs for "person 1" or "person 2", which contain all points, visible and not. Click the eyeballs next to them to turn them on and off.

[quiet raft]

If you are talking about the OpenPose Editor tab .. Then I can't do that

[quiet raft]

What model should I use to get prompt like these?

a lady model dressed up as an elf wearing a green dress on some wood, 1girl, christmas tree, solo, blonde hair, dress, breasts, blue eyes

I only get
a woman in a green dress and a gold fairy costume, pinup body, sitting on santa, blonde short hair, realistic soft lighting, pixie cut with shaved side hair, j scott campbell, young sexy elegant woman, photorealistic disney, pale complexion, wood ornaments

[stoic wave]

i need help with my prompt i am trying to get stable defusion to produce a copy of an image i found online and i cant get it to work my image is always different

[spark wyvern]

Sorry for not being a ping about help but what is that tool you are using?

[elder pebble]

you need to use imgtoimg with denoise under 0.45, or txt to img using control net reference or cany. You will never get an exact copy, that takes a lot of work around.

[sweet basin]

hello. what model is running on the current boots on this server ? how to find it ?
I am looking for a way how to generate exact same pictures (with exact same seed and prompts) on my local environment instead of discord server. (sorry for repeating same question in general chat 😇 )

[silver valley]

hey, its the sdxl 1.0 model

[sweet basin]

I thought it some modified model.

[wraith tulip]

is there a model like Bing AI's model? Trying to recreate some images, but I don't really feel like using bing ai.

A model that looks like this, for reference

[hollow mesa]

Hi guys. Hope you all have a good day. I'm super new to SD. i want to add a completely new object to an image and it should fit well with the lighting and surrounding. I try couples of time with inpainting but didn't success. Can someone help please?

[jagged fog]

need help generating a specific image of a man in a grey suit, and a full-sized building instead of the man's head.
pls send me if u succed.

[jagged fog]

looking good but i need the building to be full sized, its either you can see a huge building for the head and small man for the body or regular man for the body and the head is the building but there is no room for all of the building

[marble aspen]

I thought that is a full building. Care to show me an example?

[jagged fog]

i wish i had example, but i'll try somehow to show you what i mean

this

or

this

[feral radish]

Are you high bro

[jagged fog]

lmao

[marble aspen]

is it going to be a piece of art or just a humorous photo?

[jagged fog]

i hope humorous photo

i succed doing a pretty good photo of what i wanted i just need to make few things disappear

photoshop have somthing with ai that can do that?

if someone can make the things i marked disappear that would be nice

[glad mirage]

Can someone help me get something like this but not buried in the ground? The blade length should correspond to the handle length, and I've been trying to add a dragon eye between the hilt and blade (image 2).
Also, please tell me, does the ([text here]:[number]) work here? And does the order of words matter? For example, if I put long blade at the beginning of the prompt will it be longer than if I put it in the end?

[glad mirage]

I can't get it either, but I did find this.

[marble aspen]

Photoshop has a remove tool, just paint any unwanted area.

[stoic wave]

i am using comfyui and i have seen people input an image and get the same out i can link to the image i am trying to get

[twin quartz]

Does anyone have ideas on what prompts can i use to fix ears? they always come out a bit deformed for some reason lol

[glad mirage]

Anyone?

[slow mist]

how to properly add weight to multiple keyword, for example Watermark, Text, censored
how to write this shorter (Watermark:2.0), (Text:2.0), (censored:2.0)?
(Watermark, Text, censored:2.0) or something else?
i need to add more weight when i am using lcm lora that requires low cfg

[slow mist]

and much more stupid question, should weight for negative prompt in a1111 be + or -?

[glad mirage]

I have the same questions

[pseudo elm]

Unable to generate a robot with missing limb, any solutions?

[slow mist]

be sure that you do not have something like bad anatomy, missing limb, disconnected limbs in negative prompt

[dull portal]

What happened here?

I think the little guy is trying to hand me a purple cabbage, but anyways, why is he there? This is ComfyUI, I prompted for Joker in a purple suit, I see one, but then Joker face paint (cause I wanted civilians and for them not to have face paint, that cause the second Joker? This was 1 hour and 15 minutes on this mini PC in CPU mode.

[remote badger]

I found a really good way to fix hands. https://www.youtube.com/watch?v=sCtRgc21CjA

I'm running an AMD card on linux so it was a pretty big hassle to setup, but using the Dreamshaper model with a low strength, I'm getting a very nice result.

[thin owl]

Thank ya, i'll check it out

[remote badger]

It still struggles with small hands

[slow mist]

krita should be installed by default by any kde based linux (at least it was like that in the past, but i haven't use KDE in last 10 years)

btw, there is a photoshop plugin that use a1111 https://github.com/AbdullahAlfaraj/Auto-Photoshop-StableDiffusion-Plugin

i stoped using text inversions related to hands, since some of them totaly change the picture

just find a decent model that do not mess things too much and use something like this in negative prompt

deformed, bad anatomy, disfigured, poorly drawn face, mutated, extra limb, ugly, poorly drawn hands, missing limb, floating limbs, disconnected limbs, disconnected head, malformed hands, long neck, mutated hands and fingers, bad hands, missing fingers, cropped, mutation, huge calf, bad hands, fused hand, missing hand, disappearing arms, disappearing thigh, disappearing calf, disappearing legs, missing fingers, fused fingers, abnormal eye proportion, abnormal hands, abnormal legs, abnormal feet, abnormal fingers

and add a few keywords related to quality, for example

worst quality, low quality, poorly drawn, by bad artist

that fix most of the issues with human anatomy

[slow mist]

can someone please answer this? i am sure someone knows

[silver valley]

Yes that will work

I think aryetis tested it sometime ago

[slow mist]

so (watermark, text, censored:2.0) should behave the same as (watermark:2.0), (text:2.0), (censored:2.0)?

and even negative prompt should have positive weight? right?

[silver valley]

Yes

But be carefully with anything above 1.5

Thats like forcing it, and can cause random behaviour

[slow mist]

make sense

[slow mist]

how to use "layered" prompt? like subjet + subject style, backround + backround style
somthing like 1girl + anime style, background + realistic cyberpunk style

[forest orbit]

What am I doing wrong and this object gets replaced with wrong background?

[silver valley]

multidiffusion with prompt control or regional prompter

[slow mist]

hmm... i understand how regional prompter devide image to sections, but how to separate complete background

[silver valley]

Tiled Diffusion works perfectly for this:

[slow mist]

nice, and how would you add additional character on the left with this?

[silver valley]

making a new region, you can add up to 8 regions

[slow mist]

ok, thanks

i am stil strugling with regional prompter

i am getting merged visual instead of a two separate objects

[silver valley]

xD i tried it a bit, but tiled worked better for me as you can visualize it better, but its for different purposes

[slow mist]

red cube BREAK blue sphere
gives me a mix of those two
even i used 1,1

i am confused about this part

in matrix tab

[slow mist]

i do not have feather in mine

oh, i didnt switch to foreground, thats why it was not there, i see it now

[orchid scaffold]

Do you guys think this prompt is well written or no?
an ultra-detailed hyperrealistic caricature of (Vladimir Putin wearing a suit and tie:1.3), Film grain, masterpiece, skeptical, doubting, cynical, with Pixar-like rendering. The skin texture should be rich and tangible, with fine details like pores and subtle blemishes, and the lighting should be complex, with soft shadows and delicate reflections to add depth

is it too wordy?

Using SDXL

[twin yoke]

Personally, I wouldn't write it that way, but it depends on what you're going for. The problem with words like "hyperrealistic" is that it implies that it is art, but with an attempt for the art to look realistic. If you're actually trying to generate a realistic looking person in the image, that word can actually be a problem for you.

Also, because the subject matter is the most important thing, I would suggest moving words that are in front of it to a spot behind it.

I'd also suggest toying around with the order of some of the descriptors depending upon the outcome of your image and what you're looking to generate.

[slow mist]

as expected hell opened when i tried to use LCM with this 🙈

generation is too slow for a proper experimentation

[orchid scaffold]

thanks!

I removed hyperreal

Any tips on making my character not cross eyed?

I am curretly using "cross eyed" in the negatives but it doesn't help

[dull portal]

'detailed eyes' maybe, gotta tell it to pay attention sometimes

[slow mist]

try this #📝｜prompting-help message
it mostly works as intended

[remote badger]

What model would you consider to be a decent model for making hands? I've tried a bunch. The picture I'm working on right now has five people in it, so their hands are pretty small. I found dream shaper to be the most reliable so far but it has a hard time with smaller hands which is problematic for my picture.
I'm using a ubuntu studio which came with Krita pre-installed. The difficulty was getting ComfyUI set up, getting the AI diffusion plugin setup, and putting all the proper models into their proper folders. I also had to make sure all the permissions for the folders allowed me to put files into them and allowed ComfyUI to do it's output without being permission restricted. I literally spent all day making that happen.

[slow mist]

the one that can produce something like this without using all text inversions that exist on civitai 🙂

Steps: 6, Sampler: Euler a, CFG scale: 2, Seed: 65316718, Size: 768x1024, Model hash: ec41bd2a82, Model: photon_v1, VAE hash: 735e4c3a44, VAE: vae-ft-mse-840000-ema-pruned.safetensors, Lora hashes: "LCM_LoRA_Weights_SD15: aaebf6360f7d, SDXLrender_v2.0: 3925cf4759af"

with a dash of black magic

i like this model very much https://civitai.com/models/84728/photon

[half nymph]

hey guys
which model is the best one to generate greetings cards on social media

[polar lava]

hi all, can we do image to image generation with the bot ?

[slow mist]

i have a strange question (as usual) 🙈
what to add to negative prompt to fix fingernails?

hand are good, but fingernails are strange

[quartz zephyr]

need help to make it more clear prompts are "a close up of a man with horns on his head, Handsome, rugged, golden eyes, beard, maroon hair, white horns, looking at viewer, (4k, masterpiece best quality:1) lora:add_detail:0.5"

Negatives are "(bad anatomy, bad generation, bad quality:1) 1girl, blue eyes, red eyes"

img 2 img

original photo is this

[valid wasp]

sliding over here @slow mist

[slow mist]

you sent this prompt
pink eyes,beautiful detailed eyes,1 girl,smile,cute,hair bangs,long hair,wearing robot armor,red armor with silver accents,lora:gundam:1,

you have only one lora there, but you have Lora hashes: "Japanese-doll-likeness: 0450e1183d1a, gundam: ef694f6150c8" in your generation info

Japanese-doll-likeness is triggered by something, or that is not the whole promt

[valid wasp]

i removed the second one by the time i copied the output

[slow mist]

ok, https://civitai.com/models/22470?modelVersionId=255349 states use the weight from 0.5-0.8

try with lora:gundam:0.6

[valid wasp]

wrong lora

https://civitai.com/api/download/models/31131

[slow mist]

that is download

i need page with description

[valid wasp]

https://civitai.com/models/26002/gundam-chillout-mix

[slow mist]

as i said 0.6

[valid wasp]

ok, changed

generating

[slow mist]

1 is too strong

[slow mist]

let me try

[valid wasp]

pink eyes,beautiful detailed eyes,1 girl,smile,cute,hair bangs,long hair,wearing robot armor,red armor with silver accents,lora:gundam:0.6,

[muted reef]

are you using the same checkpoint as the lora is trained on? : chilloutmix_NiPrunedFp32Fix

[valid wasp]

yes

[muted reef]

lol .. always fun when the example images with the Lora don't actually use said Lora

[slow mist]

cfg is 11.5

use 4-8

[valid wasp]

copied from teh examples

[slow mist]

40 steps on mac on batery takes a eternity

[valid wasp]

lucky

lol

[slow mist]

let me try another example

and than i will try your prompt

my images will look different, since i am using another model and i do not have nvidia

[valid wasp]

i'm on amd

[slow mist]

the only differenc is that the same seed will not produce the same image, but quality should not be affected in any case

[valid wasp]

i get taht, jsut wondering why with teh same settings, i icant get non-onnx to look like with onnx

[slow mist]

as expected with cfg 11.5 i got a total mess

[valid wasp]

looks like mine

lol

[slow mist]

lower cfg to 5 or 6 and try again (leave lora at 0.6)

[valid wasp]

i'm going back to the onnx install, the speed is well worth having to rebuild when we get rocm support

i'll just need to fix the png info not being included

[slow mist]

my model does not like pink eyes

Steps: 8, Sampler: LCM, CFG scale: 2, Seed: 3760521634, Size: 512x768, Model hash: 0a902e8ee6, Model: detailAsianRealistic_v45LCM, VAE hash: 735e4c3a44, VAE: vae-ft-mse-840000-ema-pruned.safetensors, Lora hashes: "gundam: ef694f6150c8, SDXLrender_v2.0: 3925cf4759af", Version: v1.7.0

but i am using black magic 🤣

[valid wasp]

how do you specify a vae?

[slow mist]

go to setting, than to user interface and add this

[quartz zephyr]

why does the woman look better but the guy more realistic?

[slow mist]

any change in the prompt will affect the picture, so it is posible that you get a slightly different image

you can try to sligtly change cfg and see will it gets better or worse

[glad mirage]

Can someone help me get this but better? I need a dark background and to make it more epic somehow (glowing, on fire)

[valid wasp]

i'm going to bed, I'll try to figure out the prompt i used tomorrow, really sucks that onnx isnt saving teh png info

[half citrus]

Hello! I've seen a lot of tutorials about getting consistent characters in AI, but does anyone know of resources for getting consistent environments?

[valid wasp]

i miss linux level logging

[dull portal]

How do you prompt for a pose, using ComfyUI, I try 'crouching with hands on knees' proceeds to grow hands on knees, trying to use OpenPose models looking ok at first, but then decides that foot needs extra toes or squished to fit the length of the supposed skeleton it was trying to get the pose off of (shortened perspective angle)...oh and that's besides skin starting to look like it was smeared on for some reason by trying to apply this OpenPose thing.

[lime estuary]

im using img2img with ReActor. i'm trying to use the inpaint to faceswap only the face but the facial features keep coming in oversized and slightly large, what am I doing wrong?

[small pulsar]

how to write prompt for SD to display a box of text?

[glad mirage]

What's wrong here?

i had put a space before the command thats why

[slow mist]

can someone give me a link to a good photo to anime tutorial, that use controlnet? img2img + promet gives realatively good results for simple images like person face, but for a complicated images like landscape i am getting something that is totally different, or it looks as a photo with some overlayed anime details if denoise is too low

[flat gorge]

Is there a syntax to include refiners in a prompt file - i.e. --refiner "blahblahblah"

[tropic dome]

(masterpiece, best quality, 1boy, male focus:1.2)
is this the correct way to enhance prompt strength?

[slow mist]

yes, but in most cases masterpiece, best quality (and similar) will actually do nothing

[tropic dome]

oh

[slow mist]

my advice is to try create an image with them and then another one with exacly the sama params and seed and to compare the images

[tropic dome]

do you have any tips for ai not listening to negative prompt, like extra limbs

[slow mist]

this is something i am testing

[deformed | disfigured], poorly drawn, [bad : wrong] anatomy, (([extra | missing | floating | disconnected | fused | malformed | mutated | abnormal | huge | disappearing] [limb | arm | leg | head | hands | fingers | feet | calf | thigh])), (([short | fat | ugly | crooked] fingers and hands)), ((two thumbs on one hand)), ((two nails on one finger)), four fingers, six fingers, long neck, blurry, ugly, deformed, noisy, low poly

in some cases i had much better results than with badhandv4 and ng_deepnegative_v1_75t

i am sure it can be better, but this is the version i made yesterday

with badhandv4

with ng_deepnegative_v1_75t

with my negative prompt 🙈

but i changed some things since then

[tropic dome]

there are loras to improve bad prompts?

[slow mist]

did you mean to improve hands for example, or something else?

[tropic dome]

overall, toes, clothes merging with everything and so on

[slow mist]

you have textual inversions for some fixes
https://civitai.com/models/16993/badhandv4-animeillustdiffusion - fix hands nicely in anime, but not so good for realistic images
https://civitai.com/models/4629/deep-negative-v1x - fix many things, not just hands, for "normal" not-disfigured people

i usually use my negative prompt first, and then try to combine with deepnegative if needed

[tropic dome]

where would i put these? and how would i use these?

[slow mist]

in embedings folder

[tropic dome]

and they apply automatically?

[slow mist]

you need to add trigger word to negative prompt

[tropic dome]

okay thanks so much!

[slow mist]

like this

but as i said in many cases my prompt is enough, and in some cases it actually worked better (as you can see in example)

[tropic dome]

last question, what are hypernetworks

[slow mist]

basicly most of those (loras, hypernetworks...) do the same thing, but in diffrent way

i use this hypernetwork for example https://civitai.com/models/140117?modelVersionId=155236
it can modify the shot without too much play with prompt

[slow mist]

can someone give me any advice?

[valid wasp]

if i have an image of a person sitting, and i want to have teh exact same person standing, all I should need to do is a prompt of standing with the sitting seed, right?

[muted reef]

what is the best method to not have a person in the image trying to do prompt blending (separating the background and subject with two prompts and blending them afterwards with ClipVision and IpAdapter) .. but i get a person in some of the background generations).

[slow mist]

i was adviced to try tiled diffusion for that https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111

you will have better control what is what than with some other solutions

like this

background and two subjects

[muted reef]

am using comfyUI so A1111 extensions won't work but thanks

[slow mist]

https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/issues/177

Unfortunately, i do not think it will be made anytime soon

[kindred blaze]

Trying to make some nifty android lady pics, the bodies are turning out really nice and techy, but I am wanting their head/faces to also match the metal/plastic look and not just a human head plopped ontop of a robotic body...any prompt advice please? ^^

[strange fox]

If you are still having this issue, it could be a bunch of factors. You might have your codeformer weight too high or low. The resolution of your starting image and face image might be out of wack. You might need to crop the face image you are using more.

What UI are you using? ComfyUI, Automatic1111, InvokeAi, etc.?

[frigid tree]

Well i need some help, i want too make a image of a bald guy standing in a desert with back to viewer, standing infront of a car

but, do u guys make it in parts?

like first the desert, then the person, then the car or all in 1 prompt?

something like this but i cant get it too work

best way with inpaint or something?

or some tutorials

[compact maple]

Hoi, what's the proper way of separating colors and themes per item/thing/subject?

For example:
red and black dress, dark red drapes, galactic purple hair with cyan hair ends, black and gold vampiric furniture, rustic brick walls.

As whenever I try different colored stuff, it makes 2 separate clothing's black if I specify hair another color, then hair will turn into black instead.

And what's the purpose of "BREAK" I see used in images?

[quiet raft]

How do I get blue / green hair? I set it to Blue|Green but nothing happening?

[wispy jetty]

how to use prompt in bot section

what do I write there€??

[quiet raft]

What do you mean?

[silver valley]

You Start with / dream then you get little blocks like prompt: or format: to select

[wispy jetty]

ok

[pliant solstice]

Hi.
I have no way to get normal hands.
Please help me with fixing this point.

[raw sparrow]

are you new to anime 😄 ? Typically men are drawn with more details and facial features ( therefore more realistic ) and woman are drawn more like young cute models ( therefore look "better" ). You need are more detailed prompt to counter act those stereotypes to get the same level of details for men and woman

[dull portal]

Is there some embedding or something I should use in 'negative' prompt to not get squished limbs as if it's trying to force to get it into the frame? This is what I'm using right now:
(ng_deepnegative_v1_75t:1.6),(worst quality, low quality:1.6), (BadNegAnatomyV1-neg:1.2), (EasyNegative:1.2),(hair between eyes),sketch,duplicate,ugly,huge eyes,worst face, (blurry:2), horror, geometry, bad_prompt, (bad-artist-anime), bad-artist,squint eyes, logo, text, (watermark:1.2),username, lowres, normal quality, less people

[valid wasp]

If I have a full body image, but I want to generate just headshot photos, would it be easier to crop and upscale, or is there a way to txt2img or img2img to do this?

[dull portal]

I'm not much help on using it but I just saw this mentioned here the other day: https://github.com/Acly/krita-ai-diffusion/ was kind of a pain for me to setup, and haven't actually used it yet, but it's not showing any errors like it was missing anything. Had to install, refresh, run again, find missing, install refresh, run again, till it stopped giving me messages. LOL This is the supposed list of what it's going to be searching for. https://github.com/Acly/krita-ai-diffusion/blob/main/doc/comfy-requirements.md

[dull portal]

try 'portrait, headshot focus' maybe?

[valid wasp]

reusing my existing promts on teh image i like?

[dull portal]

Yeah, depending on the prompt it might still try and get more...I've also been playing around with this LoRA https://civitai.com/models/114460/zoom-slider-lora

[valid wasp]

ok, i'll try taht first, then maybe the extension

thanks

[spark wyvern]

Can ayone help me out to set up my webuii again :c

The first image is a result from webuii using the same model
advolomptuos smth like that and orangemix vae

I have no idea why it looks so bleak, and that happens with every vae and the model here, for some reason it produces bleak images, even at 40 steps it produces low quality images

The second image is a result from comfyUI, same model, same prompt.

I don't get if its an issue due to comfy's optimizations, or maybe the fact that I am using latent couple in webuii idk, please help :)

[silver valley]

Can you show the settings you sued for the bleak image? Also can you try kl-f8-anime2 and clear vae and compare them

[spark wyvern]

uh I already closed sd webui, Im not entirely sure how to get them back :c, if you look at the metadata im sure the latent couple should show up or at least the prompt, vae and model I was using :)

Or does webui have a log like comfy does?

AND (science fiction, neon, bright lights, city) a neon futuristic night time city filled with skyscrapers, (skyscraper:1.2), cyberpunk, future

AND (1girl:1.3), adult_female, female_cyborg, mechanical arms, mechanical legs, green mechanical body ,translucent_clothing, latex bodysuit, robotic_headwear, headwear, yuri, long hair, (blonde hair:1.2), long_blonde_hair, (green ey:1.3)es, standing_straight, upright, looking at another, looking down to the left corner of the image

Negative prompt: (mecha:1.5) {{{young_female}}}(bright_colored_room:1.3), daylight, day, (sunlight:1.3), high contrast, high saturation, high color, cropped, mutant, scary, conjoined, lowres, blurry, worst quality, medium resolution, low res, low resolution, (deformed:1.4), mutated, blurry, bad, incomplete, signature, text, jpeg artifacts, compression artifacts, getty, getty images, text, bodyless faces, extra_faces\n\nzombie, (incorrect_body_torsion:1.1), (incorrect_body_orientation:1.1), (twisted_torso:1.1), (twisted_neck:1.2) , (extra_legs:1.4), (extra_arms:1.4), (extra_fingers:1.3), too many legs in image, character must not have more than two legs, character's body must not be deformed, (futanari:1.3) EasyNegativeV2 BadHandsV5NegativeSamael1976 FemaleHandDetailerPlusPlusKaleidia
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1446358145, Size: 700x500, Model hash: 104e7066cc, Model: AdvoluptuousV1AnimeAndHyperrealisticImagesAdvokat, VAE hash: 0b204ad0ca, VAE: orangemix.vae.pt, Clip skip: 2,

Here is the information about the generation, all that I could give you from the metadata, for the latent couple

I used a) the literal predefined settings, like the background division left, division right, same weights same divisions

And before you jump at the negative prompt being longer, I am using the same info from comfy, and in comfy I set the nodes up such that the tokenization uses **length + mean ** and is a style similar to a1111 :)

[dull portal]

Uh, you have 'high contrast, high saturation, high color' right in your negative prompt, isn't that what you don't want?

[spark wyvern]

wait really? how strange, I figured copying the same negative would do the same as in comfy.

But if i change that the image itself is looking bleak, the colors are one thing, but most of the composition is weird.

but I will try removing that and see how it goes, maybe webui works too differently idk

[silver valley]

Good catch there, yes try without them

But the first image looks like no VAE got applied

[spark wyvern]

also , removing --no-half and --no-half-vae damage image quality?

[silver valley]

Nope

[spark wyvern]

I haven't touched the VAE naming or anything in them, I changed model names for organization and that hasn';t killed them yet

[silver valley]

And you select the VAE over the dropdown?

[spark wyvern]

give me a sec the ui is loading i can show you

im generatin once it finishes ill send the result

[silver valley]

you should try the kl-f8animev2 vae for more vibrant colors

[spark wyvern]

Okay here is base, no changes in prompt and with orangemix vae

Imma try changing the negative as noted above and then kf

But kf makes the image look less sharp all around, I have tried in comfy with DPM++ 2M and all anime vae I have and the worst results come out of kf

Changed pompt, (i have no idea where that man came from but yeah he appeared there)

kf-8 Anime VAE

[silver valley]

yea but colors are much brighter with kl-f8

you can try Clearvae too

[spark wyvern]

I would say yes but the clothing lost detail for instance. And where can I get that vae?

[silver valley]

https://civitai.com/models/22354/clearvae

[spark wyvern]

Let me get it and try it

[dull portal]

ey, you got '(green ey:1.3)es' in your prompt

[spark wyvern]

dang let me change that :c, idk how that happened but oh well

Im going to try rn with same kf vae and upscale see if the issue is that it desperately requires a hires.fix

[spark wyvern]

well either webui is not my friend or I am doing something wrong, this is with clearvae. :c

[silver valley]

hm not as vibrant as kl-f8

[spark wyvern]

Im rn learning a method to do latent couple in comfy, I will try the same thing there and see if it works better

[silver valley]

the image needs to be upscaled too

but colorwise it looks okay

[spark wyvern]

yeah but upscale in webui alwways fails

[silver valley]

lol why?

[spark wyvern]

even 2x models fail

out of memory

[silver valley]

what was your gpu again?

[spark wyvern]

RX 5500 XT

8gb :c

[silver valley]

ahh right :/

even with tiled vae ?

512x768 upscaled by 1.5 ?

with tiled vae enabled

[spark wyvern]

most upscales work in comfy but I have no idea how to do tiled vae in webui

tiled works well in comfy too with ultimate sd upscale

[silver valley]

you need to install the tiled diffusion extension

[spark wyvern]

oooofff alright let me boot up the server again

[stuck dirge]

I just want a men with a sweat with flowers on it@muted reef

[muted reef]

man wearing a sweater with flower patterns, standing, boots

[spark wyvern]

boots :0

[muted reef]

or footwear etc

anything that tell the feet should also be in the image

[spark wyvern]

can you guyide me through how to use this tiled thing please? I got it but im not sure how to use it , should I turn hires.fix too or what?

[stuck dirge]

Now i have these kind of images...

[silver valley]

rn you cant use hires fix right?

[spark wyvern]

check #🤝｜tech-support I sent the erorr there ^^

[valid wasp]

so the prompting changed the character, but the LORA was a great idea. I'm using SDXL so i had to use a different one, but its doing exactly what i needed. thank you for the advice!

[zealous field]

Is there a tool that would help me improve my prompts? maybe a guide book, or even better a prompt generator ai

[quiet raft]

What does the Refiner do really?

Never really used it for anything

[subtle linden]

it adds detail with the base model... and a ton of problems with the others

[quiet raft]

Should I use it all or not?

[silver valley]

not needed imo

[subtle linden]

if you´re using base SDXL, it´s handy but not required, just don´t yank the weight up much

[zealous field]

I think it's better to use "More Detail" LorA's

less problems

[subtle linden]

that miiiight be problematic

on nightcafe i can´t use loras (no pro)

[zealous field]

I just started using invoke ai, and everything is so comfortable to use

[subtle linden]

and tensorart has a lot of models/loras focused on either random girls/anime/spicier anime

i saw a few "realism" models

but not a single high detail lora

sadly for me, not a single glow/magic/rainbow lora either ;-;

[quiet raft]

What are the best control models?

That you use?

I have only used openpose, canny and depth so far

[silver valley]

ip-adapter!

its amazing

[quiet raft]

Alright, I will try it 🙂

[quiet raft]

What is the best Img prompt generator?

[muted reef]

depending on the models .. you can use tags from booru sites for instance or just natural language

wouldn't use generators and just modify prompts as you go

[subtle linden]

booru sites...?

[muted reef]

https://safebooru.org/ for instance ..

[subtle linden]

Oh ok anime

[quiet raft]

Where is it that I turn off the preview window?

[muted reef]

in A1111 or comfyUI?

[quiet raft]

A1111

But I think I found it, I turned off 1 settings and nothing happen

But I found another settings and now it's not previewing.. Like you know as it working the images 😄

[unique solstice]

Hey people, I haven't found a way to replicate the faces of my OC to make multiple images. If I recall correctly, LoRAs should be used trained with more than one image, I tried ControlNet and ReActor but the faces look really bad. Any help or guidance would be appreciated.

[silver valley]

You can also use IP-Adapter or Reactor to generate with a specific face

Oh you tried reactor

[frigid tree]

can someone explain why my faces keep getting weird and blurry

negative prompt: EasyNegative , (worst quality, greyscale), ac_neg2, zip2d_neg, ziprealism_neg, watermark, username, signature, hills, text, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, bad feet, extra fingers, mutated hands, poorly drawn hands, bad proportions, extra limbs, disfigured, bad anatomy, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, mutated hands, fused fingers, too many fingers, long neck, deviantart,ac_neg1,text,anime, cartoon, line art ,monochrome, stickers, text, rocket, bombs, epiCPhotoGasm-colorfulPhoto-neg

[muted reef]

Limit the amount of Negative embeddings to 3 at most

get Adetailer if using A1111 to have it fix up faces and hands etc

@frigid tree

[orchid scaffold]

tips on how to get an angry character but keep their mouths closed?

[outer plaza]

What is the term to make a bird have spikes on its wings?

[full stream]

hey there! I've tried to prompt images a few times, but sometimes it will prompt out videos instead and I'm not sure why it will suddenly do that.

[silver valley]

hey, use the format:image tag

[full stream]

ohh okay, thanks!

[formal cypress]

Hi, I am trying to create sdxl lora for product images
I have 6 product images with white background
But it struggles in color
After I finish training it sometimes generate different color products
does anyone can help me?

[subtle linden]

6 images is a pretty small dataset, and they probably aren´t varied

aside from the white background ones you should try taking more pictures with the product in different situations

[thick pond]

Anyone knows how to avoid exaggeration on upscale? For example adding "freckles" to promp looks great at first but after latent upscale they become larger more frequent blotches?

[subtle linden]

maybe [freckles: -void token here-:0.9] could work? turns off the token when upscaling is incoming

might be just a property of the upscaler tho, idk

never tried making people so-

[molten ether]

Can anyone suggest a model that was trained on horizontal images?

[quiet raft]

If I would try to upscale IRL images, what upscaler is best for that?

[sour ether]

do yall know how to prompt better in majicmix to have a standing person but with a good face?

[buoyant isle]

hello
in stable diffusion version 2.1 there are unusual drawings in the images. It draws the images worse than normal. i have provided examples below. is it possible for you to check?

Stable diffusion v2.1

[slow mist]

this usually happens when you are using dimensions different than those that model was trained on

2.1 was trained on 768x768

can you please try everything same, just with 768x768?

[buoyant isle]

sure

640x640

[muted reef]

SD2.1 in general is a pretty bad model ... use SD 1.5 or SDXL

[buoyant isle]

2.1 so realistic

[slow mist]

100% agree with @muted reef

[buoyant isle]

but bad

1.5 or sdxl not realistic

look sdxl

[slow mist]

sdxl

[buoyant isle]

howw

can you give me prompt

I will try

[slow mist]

Steps: 5, RNG: NV, Size: 576x1024, Seed: 1555025749, Model: hephaistosNextgenxlLCM_v20, Version: v1.6.0-2-g4afaaf8a, Sampler: DPM++ SDE Karras, CFG scale: 3, Model hash: aeaaff8e1f, Hires steps: 3, Hires upscale: 2, Hires upscaler: 8x_NMKD-Superscale_150000_G

and see the number of steps

https://civitai.com/models/200061/hephaistosnextgenxl-lcmbf16hybrid

i love this model

[muted reef]

SDXL 🙂

#1103708504142925824 message

[buoyant isle]

We use it via API, why does it give bad results?

[muted reef]

make use of negative prompts

[buoyant isle]

wt hell 😄

[muted reef]

negative prompt: anime, ugly, contortionist, amputee, polydactyly, deformed, distorted, misshapen, malformed, abnormal, mutant, defaced, shapeless

style: photorealistic

[slow mist]

try mine, give me a second

[buoyant isle]

cinematic film still, strongman, circus outfit, man, mesmerizing beard, smiling, Top hat, victorian freak show, amazing details, dark atmosphere, shallow depth of field, vignette, highly detailed, high budget, cinemascope, moody, epic, gorgeous, film, Fujichrome Provia 100F, F/8, RTX, photolab, high quality photography, 3 point lighting, flash with softbox, 4k, Canon EOS R3, hdr, smooth, sharp focus, high resolution, award winning photo, 80mm, f2.8, bokeh

my prompt

[slow mist]

[deformed | disfigured], poorly drawn, [bad : wrong] anatomy, (([extra | missing | floating | disconnected | fused | malformed | mutated | abnormal | huge | disappearing] [limb | arm | leg | head | hands | fingers | feet | calf | thigh])), (([short | fat | ugly | crooked] fingers and hands)), ((two thumbs on one hand)), ((two nails on one finger)), four fingers, six fingers, long neck, blurry, ugly, deformed, noisy, low poly

put all this in negative

leave everything else the same, just use this as negative

what you get?

[slow mist]

lets see what i can steal from here 🙈

[buoyant isle]

not realistic

but nice

[muted reef]

yea, not using a realistic local model 😄

[buoyant isle]

2.1 realistic

But the images are inaccurate.

[slow mist]

have you tried your prompt with the negative i sent?

[buoyant isle]

trying

😄

[slow mist]

768x768 or 1024x1024?

[buoyant isle]

1024

[slow mist]

please try 768x768

btw, can you show all settings? sampler, cfg, steps?

[slow mist]

wait

that is not good

you should have 2 boxes

one for positive and one for negative prompt

it will not work like that

which software that is?

[buoyant isle]

neyapayi.com

our project

[slow mist]

2.1 must have negative prompt

in positive prompt you are entering what you want to get, and in negative what you does not want
negative prompts and text inversions are there to suppress bad and deformed images

if you do not want to have negative prompt as part of ui, you should add it trhough code

but without good negative, you will just get a mess

[quiet raft]

Yo 🙂 Is it possible to avoid these prompt somehow
a woman in jeans and a sweater posing, a hologram by Alexander Kucharsky, featured on tumblr, superflat, perfect shape, blonde cream, butterfly

With another Clip Model.

I keep getting, featured on tumblr, hologram by alexander kucharsky etc..

[buoyant isle]

okey guys thank you so much 🙂

[slow mist]

my advice for you is to install automatic1111 and to test many different prompts with some different models, and to see what works the best for you, and then to incorporate all that into your project

[near tinsel]

I'm trying to get a male hairstyle that looks like this

https://i.imgur.com/hlJ44qz.png

[radiant geyser]

how would I go about taking an outline of a character and having stablediffusion edit it to add clothing, faces, etc, to the outline while maintaining the exact pose?

when i try img2img it always changes the pose, positioning, etc

[cinder lark]

hi guys i have a question so so I'm trying to make a wallpaper by going through outpainting the problem is that between the images which are superimposed here is what I have, could someone help me to resolve this problem which is more than annoying ?

[agile ferry]

Ideas on how to get Sisters of Battle and other Warhammer 40k images? Checkpoints and prompts to try?

[surreal aspen]

Looking for help on how to get fortnite style icons/images like these, anyone know what checkpoints and prompts to try?

[woven willow]

Does anyone have any tips for prompting a limited palette or (effective) resolution? Like if I want to generate pixel art or an image for a sticker that should have only a handful of color values and be effectively 100-150px across

[vagrant sky]

How to do image to image prompt with discord bot

[vagrant sky]

Did you get any answer?

Did you get an answer?

[bronze tiger]

I prefer using Stable Diffusion but I'm asked by my company to use Dall-E 3 and Midjourney due to better text coherence and understanding of a longer sentences, are there any tips and techniques to make Stable Diffusion models to understand a longer Prompts better, are there extensions in Automatic 1111 or nodes in ComfyUI that help with this?

[muted reef]

if a sentence doesn't work .. chop the sentence up to keywords works 99% of the times

[rancid swallow]

which model is most close to midjourney?

[muted reef]

SDXL and all merged/trained with it as a base

[rancid swallow]

which one is sdxl? i downloaded juggernaut but cant get images that i want.need to be expert on prompting i guess

[muted reef]

For realistic people .. close to MJ i would say use Dreamshaper XL

Juggernaut seems to be very specific in prompt style, 10% positive, 90% negatives O_O

[rancid swallow]

tried that too but not worked well. I think i need to get better with prompts xd

[muted reef]

give an example of your prompts .. might help

[dull portal]

Why do backgrounds tend to look like sets, like I want the scene to be a room but it looks more like the idea of a room...but it's just a closet setup kind of like the 'room' I want. Alley, same thing, just looks like it's a box that kinda resembles an alley.

[muted reef]

what is the first image you get when you think of the word .. room or alley?

that is exactly what the AI will output

room with a kingsized bed and red curtains will give a different result

suburban Alley with graffity on the walls will give a different result than Alley

[dull portal]

Ok, so I'm not expanding enough on the idea of the background huh, need to give it some distance in the prompt

[muted reef]

the more details you can give the ai the better it will be able to generate your vision

just remember that the ai can only work with like 275 words before it loops around

[dull portal]

Did not know that

[rancid swallow]

what is lora ?

[muted reef]

Lora are mini models trained on a specific style or character

[rancid swallow]

look. I wanted a sinematic , realistic image but here what it gives me xd

https://www.tiktok.com/@thestellarsagas/video/7278022559755439392 i want photos like this ones

[muted reef]

Positive Prompt: 16:9, Roman emperor, Roman Army in background, Cinematic Shot, Dynamic Pose
Negative Prompt: contortionist, amputee, polydactyly, deformed, distorted, misshapen, malformed, abnormal, mutant, defaced, shapeless, buckteeth, toothless

[slow mist]

i have a strange problem with controlent and openpose, with 1.5 it works as expected, but with sdxl openpose_full does not detect the hand position, only face and sholders (while with 1.5 it detects the hand too). dw_openpose_full in sdxl detects the hand too. what is dw_openpose_full?

[muted reef]

different dataset used to train

[compact maple]

Hoi, what's the word formatting needed to properly tell SD to distinguish colors? Like when i say black shirt, purple pants, red hair for instance, it will make other stuff any of the mentioned colors.

[muted reef]

using underscores or plus maybe or ""

"red+pants" etc

[muted reef]

cherrypicked out of hundreds or thousands of failed generations

12 minutes for one image but it came out good 😄

[rancid swallow]

or just using midjourney ahaha

[muted reef]

still gotta cherrypick 😄

[spark wyvern]

Does anyone know if this is how long Miaoshouai assitant is supposed to take for installation?

[silver valley]

long xD, check the cmd too, i would let it load

[nova seal]

How should i config this?

[spark wyvern]

Ir went up to 2600s and nothing appeared down in the cmd.

:(

[nova seal]

xd

sometimes it bugs

Just restart and it should be fine

And if you dont have it just try reinstall again

[dull portal]

20 steps, 8 CFG when I usually do 7, low detail looking like a magazine scan

30 steps, 7 CFG, more detail, still looking a bit like a magazine scan

As far as I remember these were at 20 steps and 7 CFG, what makes that big of a difference?

[slow mist]

show me the info that you got just bellow the picture (with steps, model, cfg...)

[dull portal]

Below the picture?

[slow mist]

oh, you are using comfy, sorry

[dull portal]

Ah, yes. Should have mentioned that.

[slow mist]

what model this is? is this XL or 1.5 model?

[dull portal]

1.5

[slow mist]

in that case do not generate 768x1024

use 512x768

use upsacaler later

[dull portal]

But...my joker is at 768x1024 and it turned out a lot better than my 'Aztec goddess' at the same resolution.

[slow mist]

you have images with 2 heads beacuse of dimensions

use upsacaler to get better resolution later

for 1.5 use something like 512x512, 512x768, 768x512

and upsate with 2x

next, which vae are you using?

[slow mist]

things like this happes beacuse of vae or to big CFG

try to use something like 5 if that happens with 7

do not increase CFG to get mroe details, use something like https://civitai.com/models/153562/detail-slider-lora for more details

[dull portal]

Uh...didn't know there was a choice for that. LOL

[slow mist]

for 1.5 i usually use 840000 or flF8Anime

for xl i use only two that exists 🙂

can you give me link to that model?

[dull portal]

Been playing with ComfyUI for a little over a week, only thing VAE was this 'VAE Decode' with no selection box, didn't look further into it. Checking what else has VAE on it I see VAE Loader shows me having 3 options, TRCVAE, TAESD and it's XL variant.

[slow mist]

i want to check few things

[dull portal]

The model...epiCPhotoGasm? https://civitai.com/models/132632/epicphotogasm

[slow mist]

always check those two first

do not use clip skip 2 if 1 is recommended or vice versa

if model is 1.5 do not mix with SDXL loras and other way around

[dull portal]

Actually have another gen going, bypassed the clip skip to see if that was part of my issue. 😛

[slow mist]

read How to use

use simple prompts without "fake" enhancers like "masterpiece, photorealistic, 4k, 8k, super realistic, realism" etc.

don't use a ton of negative embeddings, focus on few tokens or single embeddings

so do not use those with that model

[dull portal]

And I'ma have to look into upscaling cause I set it up and my last try was amplifying blemishes and stuff and looking like crap, besides taking forever which I didn't know it was gonna do, 3x3 😄

[slow mist]

can someone recommend something like this for sdxl? https://civitai.com/models/153562

https://civitai.com/models/122359 maybe?

is there any with -5 to 5?

[nimble panther]

Backstory: I am trying to tranfer IMG1 to a different style, and I got IMG2, since the left hand didn't look how I wanted I used gimp to cut the hand from the original and put that in the image to get IMG3, then ran it back through SD with an inpaint to try and mesh the images together to get IMG4.

Problem: For IMG4, if I turn up the denoising strenght I get a grabled hand. Though when I get a correct hand it doesn't mesh well with the rest of the image. How can I make the colors and shading mesh better without garbling the hand?

[muted reef]

use controlnet and a mask of the hand with some feathering so the seam will get blended to the rest of the image, with a relative low denoise

[nimble panther]

I haven't used controlnet before, could you link me to some resources on how to use controlnet?

[muted reef]

https://youtu.be/WZg3e6B2yPQ?si=SFtmGM5_m_NMjeNJ

[nimble panther]

TYSM, sorry for not being very well informed!

[dull portal]

Hmmm, went with a 512x768 instead of the 768x1024 I was using, Skip Clip bypassed, but still my output seems a little worse than the preview? If I run this through an upscaler I have to keep the seed or is the upscaler itself going to perform some magic on it?

[dull portal]

What the hell happened at step 3!? It was looking great at step 2.

[quiet raft]

Hello, is it possible to use like re-szie while in the img2img tab?

In the Txt2Img you have the option to do Hires. fix

Is it possible to do that in the Img2Img tab?

[silver valley]

In the img2img tab you can use the SD upscale script to upscale

[slow mist]

how to stop "color leakage" caused by a long (and most probably not good enough) prompt. what i mean is that instead to use a color for one specific detail i am getting some other details in that color

example:

woman with long hair wearing a (black dress:1.2), a (black choker around her neck) and (glasses with red frame)

produces some red details on her dress, instead to produce compleatley black dress and glasses with red frame

[subtle linden]

¨how do you bypass clipskip? i can only set it to 1, not 0, idk

[dull portal]

Guessing you're talking about SD WebUI? Cause in ComfyUI you just right click it and hit the 'Bypass'

[subtle linden]

sorry i forgot what server i´m in

[subtle linden]

i´m talking about the tensorart UI

[subtle linden]

i checked its workflow mode (which apparently works like comfy) but its way too complicated for me to add/remove anything

i´ll just ask in tensorart server ig

[dull portal]

Man, got the perfect angle from a sky scraper earlier, but no joker in sight I was trying to get, tried it again, now it's a girl with some joker paint and it's a headshot to top it off, try again joker in a spiderman pose on a roof top, and finally a joker jumping backwards off some sorta building, not over the edge like I was trying to do, sigh, gonna let this one upscale anyways I guess

[near tinsel]

Any websites that help with prompting?

I installed this "Gender Slider" embed but I'm not entirely sure about how embeds work even after reading about them. For example, the embed page says "GS-Masculine: Place in your positive prompt at strength of 0.2 to 1.3. Conceptually pushes towards rough and strong seasoned masculinity. May vary by model, lora, or prompts."

So would it be something like (GS-Masculine:0.5), joker, falling from a building?

[slow mist]

check the image infor on civitai

[near tinsel]

I mean... it doesn't tell you if you're supposed to use it that way. It just assumes that you know how to use an embed.

[slow mist]

i just checked, you have the whole How to use: section

check those

and see what they did

[near tinsel]

Yeah that's what I read

[slow mist]

so for the right image

in positive they have (GS-Womanly:1.1)

and in negative they have GS-DeMasculate

[near tinsel]

So that is a keyword to put in the prompt?

With the paranthesis?

[slow mist]

so that will make a women (even if it is a man originaly)

positive

negative

[near tinsel]

So there is no comma after that the parenthesis? So it would be like (GS-Masculine:0.5) joker, falling from a building?

[slow mist]

you do not need it, since joke will be a man

[near tinsel]

Ok, thank you

[slow mist]

you would need (GS-Womanly:1.1) if you want do to gender swap

and GS-DeMasculate in negative

[near tinsel]

Yeah I understand the concept of how to use it, it was just the prompt formatting I was curious about

[slow mist]

trust me, inspection of other people promts is the best way to figure how those thigs work

[near tinsel]

Oh, I didn't know you could do that

[slow mist]

btw

when you click on that i

you will se the prompt

but you can copy all the settings

[near tinsel]

Ok thanks

[slow mist]

paste ALL that in positive prompt

and clikc this arrow near generate button

it will change everything for you (negative prompt, steps, sampler...)

but check everything before hit generate

[near tinsel]

Really? That's amazing

I'm not seeing that though

[slow mist]

https://civitai.com/articles/904/settings-recommendations-for-novices-a-guide-for-understanding-the-settings-of-txt2img

this part
How the Copy Generation Data function works

[near tinsel]

So many guides out there. A lot of them I've been reading (or watching) are dated. I didn't realize Civi had an extensive one. Thanks.

[runic jackal]

does anyone know why i get black images even if its sfw?

[slow mist]

what software do you use? automatic1111?

[runic jackal]

yep

ive read that i shouldnt use --disable-nan-check

so ig im trying that now

well still

[slow mist]

is that radeon maybe?

[runic jackal]

ive had some issues with it before but it didnt give me black images yesterday

[runic jackal]

i dont think so

[slow mist]

do you have Radeon or Nvidia

[runic jackal]

its nvidia

[slow mist]

when you get black square, what you get below the image (generation info with model, steps...)

[runic jackal]

ive tried a different sampling method and model but it stil gave me black image

[slow mist]

do you have any external vae enabled? i am asking since that model has baked vae

[runic jackal]

i actually dont know, how do i find out?

ive used stable diffusion in the past but only with voices

[slow mist]

go to settings, then find user interface in user interface section

and add those

[silver valley]

whats your GPU?

[slow mist]

then you will be able to change vae and clip skip anytime you want

[runic jackal]

1650 super

[silver valley]

okay, then you need to edit your webui-user.bat
at the line Commandline_ARGS=
you need to add: --xformers --medvram --no-half
Then save and relaunch

then you wont get any black images and it should be a bit faster

[slow mist]

so it is same s___ as Radeon and black square and needs --no-half

[silver valley]

only for RX 5xxx radeon

[runic jackal]

i'll try that now

[slow mist]

thats why i asked if he had radeon

but he caught me off guard when he said nvidia 🙂

[runic jackal]

whoops

[silver valley]

can you show your cmd when launching the webui-user.bat?

[runic jackal]

i just saw this

[silver valley]

yea that means you didnt editet the webui-user.bat correctly

[runic jackal]

a guy in a video said you should do it like this

[slow mist]

that is not good

[silver valley]

nope, it shouldnt look like this ^^

[slow mist]

that guy has no idea how to set those is he said that

[silver valley]

it should all be in one row:
set COMMANDLINE_ARGS= --xformers --medvram --no-half

[slow mist]

second one owerrides the first, and third override the second, so you just getting the 3rd one

do it like CS1o said

[runic jackal]

oh i see

makes sense now

[sweet loom]

help! why are my images being generated so garbage?

[runic jackal]

thanks guys it worked

but does anybody know why its so "difficult" to install reactor? ive tried it also multiple times and it didnt work out in the end

[silver valley]

you need to exactly follow the steps on reactors github page

then it will work

[runic jackal]

alr thanks

[restive socket]

hey wassup guys, which prompt should I use to remove the blur on the background?

[slow mist]

try to add
blur, bokeh, depth of field
to negative prompt

[restive socket]

thank you so much man 🤝

[runic jackal]

is it just me or does roop dosent work anymore

if i try to pip install insightface thats what i get

[slow mist]

did it work?

[restive socket]

Yup

[slow mist]

i have a stupid question (as usual 🙈) - what is the maximal weight of a keyword for a1111 that i can use? for example (white dress:1.5)

[muted reef]

don't go over 3 i would say

[slow mist]

i am using model with embedded LCM, so my CFG is low (or image will be overcooked), and i am using img2img with controlnet, so i cant find a proper balance between those

[silver valley]

Depends on the word. Someone fried his whole images with sharp:1.6

I usually stay under 1.5

[slow mist]

LCM is a great thing, but it can easily become a nightmare when you need to use with other resources like loras and controlnet ☠️

maybe i am trying to do things in a wrong way

this is an original photo, i am trying to make anime style image using img2img + controlnet, and that works great

but i want to change dress color to white, and that just not working, i have CFG 2 (becuse of LCM)

i have this

and this

but there is no middle ground 😦

[muted reef]

neckstraps maybe?

don't know what those dresses are called

[opal ridge]

Quero ajuda pra criar a prompt dessa imagem

[subtle linden]

what´s LCM btw?

[slow mist]

LCM is black magic that helps to get images like those #📝｜prompting-help message
in just a few steps

for example those are made in just 8 steps
Steps: 8, Sampler: Euler a

[subtle linden]

so its kinda like SDXL turbo from walmart?

[slow mist]

but the probelem is that CFG needs to be VERY LOW (like 1.5 - 2.5)

if you justa make an image, that is not a problem

but when you mix it with some LORAs or controlnet it can be tricky

[slow mist]

actually turbo is more like LCM from wish .com 🙂

turbo is faster, but has lower quality
lcm is slower than turbo, but has the same quality as "normal"

you can get almost identical image with lcm with 5-8 steps instead of 20+

https://civitai.com/models/202108?modelVersionId=247022

check ti gallery and see the number of steps

[subtle linden]

cool, ty

[slow mist]

you can use https://civitai.com/models/195519/lcm-lora-weights-stable-diffusion-acceleration-module with "normal" models, or you can download models with lcm embeded

if you use lora, you need to download a proper one for your model (1.5 or sdxl version) and you need to guess the right combination of cfg and lora weight

models with embeded lcm are easier to use

[restive socket]

hey guys, I found a picture of a woman posing on instagram. I wanna generate a full body woman, I want her pose to be the same as in the picture I found. how can I do it? (sorry if its a bad question cuz im a newbie)

[subtle linden]

controlnet pose-to-image

[restive socket]

thank you sm

[subtle linden]

np

[wispy obsidian]

hey guys using a model, I used some prompts that gave a certain style to the images. restarted SD, used the same model, same settings, same sampling method yatti yatta, but the images now look like they're in a different style. I also did some gens with other models prior to that, anyone have an idea what's going on

[runic jackal]

how long does it take for you guys to generate an image?

[subtle linden]

that just depends on your vram+some args

[runic jackal]

is there like some way to boost it?

[subtle linden]

ig you could lower the sampling steps-

[runic jackal]

ive noticed every time I use reactor with the same resolution as the image before, it loads much faster

[restive socket]

Guys i am using juggernautXL as checkpoint, im generating a realistic pic of a woman in 1024x1024, yes its good but not realistic, hair is not detailed, low quality, when I zoom its getting worse, its obvious that its AI. How do I make it photorealistic? What am I doing wrong

[twin yoke]

Without knowing your settings, nobody can give you any specific recommendations. From a general standpoint, follow what the model author says for the recommended settings:
Sampler: DPM++ 2M Karras
Steps: 30-40
CFG: 3-7 (less is a bit more realistic)
VAE is already Baked In
HiRes: 4xNMKD-Siax_200k with 15 Steps and 0.3 Denoise

Also, if you're doing a portrait, you'll probably get better results with a rectangular resolution, rather than square, like 768x1344, 832x1216, or 896x1152.

But you might have all those settings above and still just have bad luck with seed values that just happen to give you bad results.

[restive socket]

I'm a newbie, i should've shared my settings, i'll try those settings now, thank you🫡

[slow mist]

go to PNG info, open the picture you generated before (that looks like you want) and click send to txt2img button, that will copy all the settings used (prompt, steps, sampler, clip skip...)

[slow mist]

i usually use 768x1024 with SDXL models, but i agree with absolutely everything else soul wrote. i would just add that since some seeds might produce garbage (as soul also wrote) it is not recommended to use fixed seed during the initial testing of a model. the only exception is if you are trying to recreated example image, but be aware that you might get completely different image, since seeds on nvidia and other GPUs are not the same. i leaned both the hard way 🙈

[tired vigil]

Is there any way to specifically train a LoRA or sth to get the "style of selfie" right? Like I have some selfies of a person and I'd like not to just have the person herself trained but also the quality of the photos

The generated images always look way too good, the quality is too perfect and it's clearly obvious that it's generated

[restive socket]

I tried to generate a pic with these settings using juggernautXL twice, but my pc froze at %49 everytime so I had to restart my pc, how can I fix this?

I have a gtx1060 3gb and i5 6402p

16gb ram

I know my pc is not very good but how can I prevent crashing my pc everytime? Doesnt matter if its gonna take 30min to generare a pic

[silver valley]

you cant use sdxl models with 3gb vram

there is no way around other than using cpu mode, that will take really long

better use 1.5 models and upscale them in img2img

[restive socket]

Sorry im a beginner i dont know some terms, wdym by 1.5 models?

[silver valley]

there are 3 types of models you can use. 1.5, 2.1 and SDXL
SDXL models are the large models that are over 6gb in file size. These wont work with 3gb vram.
1.5 models can be as small as 2gb so they will work without crashing or freezing your whole PC 🙂
and you mostly can ignore 2.1 models as they are not as good supported like 1.5 ones from the community side

[restive socket]

You are much appreciated thank you 🤝

[silver valley]

Juggernaut Reborn is a good 1.5 based model, from the same guy as Juggernaut XL

[restive socket]

Okay i'll try it now

[hollow lily]

Can someone assist me with prompting using this lora?
https://civitai.com/models/120096/pixel-art-xl
If I use default 1024x1024 it gives me artifacts like warped fake pixels. However if I upres 2048x208 using hires fix, the pixels look fine, but the proportions of the body are way off.

4x_nmkd-superscale-sp_178000_G hires steps 25 denoise .7 upscale by 2

[silver valley]

Try using Denois 0.45

Or even 0.35

[restive socket]

which checkpoint model do you guys think is used here? (slightly nsfw)

[hollow lily]

Lower denoise does photos like this, odd.

[covert shore]

hi guys who can tell my Why does SD always turn my prompt into videos?

[silver valley]

Hey use format:image

[silver valley]

Can you show your txt2img settings?

[quiet raft]

I can only make 2048 as max?

Or can I do custom size somehow?

Like 1260x2160 ?

[silver valley]

you can set an other max resolution in the ui-config.json
Or you can use a lower resolution and upscale with hires fix

[quiet raft]

Hires is also just 2048 as max

I can like 5000x5000 images etc. But I wanna try do 768x768 or like 512x768 and then go to 1260x2160

But can only select 2048 as max

[silver valley]

it wouldnt make sense to gen any higher than that, because sdxl is trained on 1024x1024 and 1.5 models on 512x512

with 2048x2048 you likely get more deformation, thats why everyone uses upscaling

@quiet raft with these settings your output image will be 4K (16:9)

that wont work on AMD, and for Nvidia only when Tiled Vae is enabled

[quiet raft]

Well, I usally do 512x512 and hires to 1024x1024 when happy. But then I send them to Extra ..

But on extra it's still only 2048

But I have to change that in the json file?

[silver valley]

yea you would need to change that in the json file, but for upscaling from 512x512 i wouldnt recommend Extras. Better use the sd upscale script in img2img

[quiet raft]

Wait what?

[quiet raft]

I usally do like 512x512 to test then .. I do batch with Hi.res to 1024x1024 .. Then I put them in a folder and use the batch thing in Extra to get them all done

to higher res

Like I can't do batch photos upscale with Img2Img Sd upscaler right?

[silver valley]

you can batch upscale them in img2img

[quiet raft]

But I need the prompts and all? Or how, if you could show me how you would do?

I just did 1 image.. so my setup is very wrong

[silver valley]

You dont need the prompts but you can use them too. make sure to use ngeative prompts in any case.
You have to change the resolution to 512x512. (thats not the outpur resolution), its the tile size.
Then set the Denois Slider below 0.3 (0.1-0.3)
The lower the number the nearer it is to the original, the less will be fixed while upscaling.
Then you have to select the sd upscale script.
Last step is to set an upscaler and generate.
if you get artifacts in the image, remove every person specific word from the prompt like, hair, eyes, smile, etc,

[quiet raft]

I have changed nothing expect put SD scaler and

I have uploaded a img2img

But I get fuckt up picture..

Okay nvm, I set the Denosing str to 0.15 and resize mode: just resize.

But why is this SD upscale better than the extra tab?

[silver valley]

because it can do the same as hires fix, Extras tab cant add details to an image

[quiet raft]

Well. I only use extra tab to make them go from 1024x1024 -> 4000x4000 sizes

[silver valley]

would use sd upscale script after hires fix and then at the final step load it in extras

[quiet raft]

Do you use the scale factor in sd scaler then or just set it at 1 and use extra to upscale it?

[silver valley]

i use scale factor 2

[quiet raft]

Okay.. I might try the sd upscaler a bit then and see

Not sure if it will be any higher quality .. beside be able to add items or such

[silver valley]

i have an example i can share

[agile ferry]

Is there a suggestion for an AI prompt generator?

[quiet raft]

You mean, you upload a picture and get a text prompt?

[quiet raft]

Oki 👍

[agile ferry]

Actually I do need to figure out how to get "Interrogate" back, but I meant putting in basic prompts and having an AI expand out the prompt idea.

[quiet raft]

You need to install it with the extensions

[silver valley]

what do you mean with getting it back ?

[agile ferry]

Awesome, I'll do that now. I had to restart my SD and some of the things I did got lost on the way.

[silver valley]

its always there in img2img

[quiet raft]

https://github.com/pharmapsychotic/clip-interrogator-ext

[agile ferry]

It's not showing up.

[silver valley]

the little clip icon

[quiet raft]

Double slam 😄

[agile ferry]

Ooh, okay, I didn't see that. Previously it had a big grey bar saying "Interrogate" but yep, there it is! Thank you

[agile ferry]

Hey, I appreciate it! lol

[quiet raft]

That is the extenstion

I have it downloaded

[agile ferry]

Yeah, I had that previously

[silver valley]

maybe that thing here idk:
https://github.com/imrayya/stable-diffusion-webui-Prompt_Generator

[agile ferry]

Not sure that I used it, I probably just need basic interrogate is all.

[quiet raft]

Damn, this one looks better 🥫

[agile ferry]

Interesting!

[silver valley]

yea, but havent tested it

@quiet raft here is my base imgage: 512x768 with hiresfix generated to 1024x1536

[agile ferry]

I'm always leery about trying new stuff, but that looks fun.

[silver valley]

@quiet raft and here is it upscaled by 2, in img2img and in Extras tab.

i set the denois to 0.2

i didnt removed the person specific words from the positive prompt so there are some artifacts in the tree

[quiet raft]

The left one has better quality at eyes and so

But best way is to leave out the prompt and keep the negative and so sd upscaler then?

[silver valley]

that or a better way is to enable Just Resize (Latent upscale) and then lower the denois

nor artefacts and best quality compared to the others

[quiet raft]

So when you do your stuff

[silver valley]

no

[quiet raft]

wait

Lemme finish haha 😄

[silver valley]

thats in img2img with sd upscale script

[quiet raft]

I can't use img2img > to generate stuff

I need to do txt2img > img2img

Right?

[silver valley]

if you have a lot of vram you can generate in img2img

[agile ferry]

This thing is awesome. I really like it!

[silver valley]

but using txt2img with hires fix is always better

[quiet raft]

I only got AttributeError: 'NoneType' object has no attribute 'mode' Time taken: 0.0 sec. when I did

[silver valley]

cool, i need to try it too then ^^

[quiet raft]

How did you install it?

[silver valley]

show me the settings when you get that

[quiet raft]

?

[agile ferry]

https://github.com/imrayya/stable-diffusion-webui-Prompt_Generator.git in install from URL, give the model time to dl and that's it

[silver valley]

thats txt2img

[quiet raft]

Ahh okaty

[agile ferry]

I got some really good results first time.

[quiet raft]

Its img2img

[silver valley]

ah okay but i cant see all the settings there

[quiet raft]

That's the rest

[silver valley]

there is no input image

you can only generate when an image is loaded

[quiet raft]

Ahh okay :3

[silver valley]

tried the Gustavoosta model, yea thats exactly the quality i want!

[agile ferry]

I don't know what that means, sorry. 🙂

Oh! You can pick generation models, I forgot already

Isn't it great?? I'm gonna use the hell out of this

[silver valley]

its nice 😄 yes

but i dont like that it tags a lot of artists

i dont use any artist name in my prompts

[agile ferry]

I figured I'd let it do it's thing, I haven't been disappointed yet

I'm trying Gustavoosta next

[silver valley]

haha let me know what you get 😄

screenshot from a 2011 laptop camera

[agile ferry]

2tb drive.... man, I thought that was a lot of space way back when I was young and naive. 😅

[quiet raft]

I get these prompts ... -.-'

trending on patreon, big puffy lips, long flat hair, deepfake, very realistic effect, hungarian flag in the background, persian queen,

[silver valley]

same xD

[quiet raft]

I mean.. Why do I want trending on patreon?!

[silver valley]

who even can see that 🤣

[agile ferry]

Depends on what you get! The AI seems to think it's important so who knows, might be your favorite prompt ever. Lol

[quiet raft]

I use the Interragator clip model ViT-L-14-336/openai

I always remove those from the prompt but annoying 😄

[silver valley]

that one was nice, used the FredZhang one.
1girl, wearing cool outfit, sitting in front of the sydney opera house at night, soft lighting, realistic wide angle, wonderful lips and eyes, white skin, sharp focus, 8 k high, 64 megapixel, insanely detailed, intricate, elegant,

[agile ferry]

That's really great

I'm liking Fred more, the other model left some blanks for some reason

[silver valley]

exactly, will stay with Freds xD

[agile ferry]

Okay, Photon is giving me too many extra limbs, going over to some SDXL

What's your favorite checkpoints?

I'm going to JuggernautXL

[silver valley]

if your resolution is to high, that can happen

[agile ferry]

Ooh, okay, that's the issue

[silver valley]

1.5 modesl are trained on 512x512 so stay near that. for a portrait use 512x768

then upscale

sdxl is trained on 1024x1024

[agile ferry]

I got this on JuggernautXL using just 10 steps. It recommends 30 but we'll see what changes.

Eyes are a bit off

[silver valley]

yea Juggernaut is really good

[agile ferry]

And I use ReActor almost every time as it always fixes poor quality faces without the extra ADetailer time.

That was an AI generated prompt, btw. Which is why it's relevant in this channel. 😄

[agile ferry]

I can't get Gustavoosta model to give prompt outputs, I get repetitions, repeats of the starting prompt and like 1 real result.

[quiet raft]

Just got home, lets test 😄

Wish I could add a image to it 😦

[agile ferry]

If only

[quiet raft]

If I got 2 images .. like 1 male and 1 female.. How can I add them 1 image.. Can I do it with img2img or should I use photoshop?

[static hinge]

oh great I wanted that, I had Dynamic Prompts extension that had Magic prompt, but it just generated and I had to manually check what it has created

[agile ferry]

I'm not 100% sure about it yet, but I like it as an extra tool

[restive socket]

guys I have a beginner question, lets say I have an AI girl with a specific look, and she's 5'3" tall and fit, and she has a specific breast size etc. and I generate a full body picture of her, how can I preserve that model? like how should I generate the exact same body in a different image?

[terse jackal]

Save the seed #

[restive socket]

thats enough?

[terse jackal]

Should be. Same settings + same seed = same image. Try it out! If you're using ComfyUI you can click-drag a previously generated image/video into the UI and it will load the workflow that generated that result

[restive socket]

alright

[novel marsh]

I'm not sure if this is the right place, but how do I make images with multiple woman from different races? the ai keeps trying to make asian redheads instead of making one girl of each kind

[restive socket]

I installed ReActor, I've been trying to play around with it for a while but it's not swapping the face. not a single difference, I tried to switch to epicrealism from juggernaut resolve but it didnt change

in the tutorial videos I see people turning on a face swapping setting in ReActor panel but I do not have that setting here

[analog sphinx]

https://i.imgur.com/49qrM7N.png

how does this make sense?

i can't for the life of me get a night picture

what am i doing wrong?

[rich crow]

what base model? if its 1.5, i read that it's essentially trained to create bright images from here https://old.reddit.com/r/StableDiffusion/comments/13joe98/sds_noise_schedule_is_flawed_this_new_paper/
the workaround is to use a noise offset lora or something similar to create darker images
don't remember if sdxl has the same issue or not

[near tinsel]

This isn't technically a prompting issue. I'm trying to use ControlNet. I've watched a detailed video by Incite AI but I can't seem to get the model reference pose to do the pose correctly in the final generation. I've attached the image I want the pose for. Should I bring it into Photoshop and try to make it more detailed so the lines are more prevalent?

https://i.imgur.com/3fwLga0.png this is what comes out in the preview https://i.imgur.com/3fwLga0.png

[slow mist]

you get a face an sholders