#💬｜general-chat

1 messages · Page 187 of 1

fervent thunder
#

was wondering what diffusers does that is less memory efficient?
I've started using it but cos I use big servers I don't notice memory efficiency
is it like an issue of casting or an issue of loading?

serene mountain
#

Im not sure, Id have to reload it, try to duplicate the problem and come back.

Closing out invoke and restarting would help but not solve it. Matter of fact it got worse, Id get 1-2 generations and then black squares.

#

I have comfyUI installed and it runs, Ive just put off really digging into it. Felt like if I wasnt fine tuning the “easier” programs then Id be more lost in Comfy.

#

I have forge installed as well, had the most “luck” there. The images tend to look soft but it stays consistent at least

fervent thunder
#

there's different ways to mess up memory management, is why I was asking them

median jewel
#

i have been trying to download comfyscript for 2 days buy cant get it to work at all, i have followed the instructions on the github page and tried to google, if anyone knows how to download it, please tell me :D

oblique agate
verbal notch
#

Hey guys,

is kohya_ss still state of the art when I want to train my own LoRa or are there some other usable alternatives out there?

The reason I am asking is because kohya_ss is gving me a ton of dependencies error messages after clicking the Button "Start Training" which I cant seem to fix

RuntimeError: operator torchvision::nms does not exist
20:22:47-499620 INFO Training has ended.

vagrant crater
left terrace
#

is api pricing flat fee for stable video?

atomic mortar
desert dagger
verbal notch
#

Thanks for the suggestion, I actually got kohya working yesterday and trained my own LoRa successfully, yay thomas

Do you guys have suggestions for good tutorials on how to make the bust use out off stable diffusion?

desert dagger
hardy hawk
#

I have issue with stable diffusion running on AUTOMATIC1111

#

Everyone facing same issue or its only me?

candid light
#

hi

near seal
#

I do not currently have a computer, only an ipad. I am assuming I can use artisan? And only artisan? Is that correct? It looks like much of the site is devoted to people running stuff on their own computer systems, and that artisan is the only part of the site that is for creating on discord? Please ping me.

atomic mortar
#

However these are paid services

#

But using sdxl and 5 dollars should get you about 900-1500 images depending on you use loras or not

storm shard
tribal glade
chilly heron
#

hello i was wondering how good is stable diffusion at converting sketch to a nice looking image

#

in an accurate way chat seem to change it how it want it to look like not what i skatched

storm shard
tribal glade
sinful lantern
#

Okay I have some doubt so see in my training my sdxl model the images which i have used to train the model aree all of same size 1216x832 and when I used trained model to generate images previously I was using output image size of 1024 X 1024 but images were not that much good but when I used exact output size of 1216X 832 then the generated images were amazing why is it like this if anyone can guide me in this why is it like this

abstract quarry
#

resolution is part of the input the model might overfit on

#

if all your training images have same resolution, then the model associates the resolution with your training data (same way as you would associate a trigger word in the prompt)

#

just crop some of your training images to 1024x1024

jovial shoal
#

Any recomendations on where to start getting into generative AI art? Video by Shadiversity got me interested in more complex use of ai to get better results than with using just one prompt at some site.

oblique elk
sinful lantern
# abstract quarry just crop some of your training images to 1024x1024

Another thing what I observed is fir example if I try to generate my image of Suze 2048x 2048 than what it does like it makes bad image like for example if i want to generate image of pizza than in the generated image I will have multiple pizza just thrown away but if i generate exact size image than only one pizza comes up

#

Why is it like this

abstract quarry
#

that's an artifact of the convolutions/unet. But in general you have to train on similar settings as you want to do inference

sharp cairn
#

HYEEEAAAAAAAAAAAAAAAAH

sinful lantern
# abstract quarry that's an artifact of the convolutions/unet. But in general you have to train on...

So let me tell u what I m actually doing so my input image is actually of size 6969 X 4640 okay then when previously i was doing some experiments and making buckets with max resolution to 1024 x 1024 then the bucket size which is made is of size 1217X 832 which is 6 times smallee than my original image so what I did was i increased my max resolution to 4096X4096 so now the bucket which I got was 4096X 3792 which is better than previous ones and not in the training config file I m using exactly the same resolution of which is in my bucket because all of my images are of same size so what do u think is this correct way of getting high-resolution images at inference and not getting multiple pizzas

#

So i did wrong or right because for now my training is running

#

Wdyt

tribal glade
#

👀 gotta let all the channels know bout that fiddy dollaz huh

nimble light
#

What are the best AI Image to Video generators that accept watching videos to earn tokens

tribal glade
main snow
#

Is the stable diffusion or is the diffusion stable? Hence the question.

-Shakespear probably

main snow
nimble light
main snow
nimble light
#

What is that

main snow
nimble light
#

So a Local AI Generator? Like Comfgy UI or Stable Diffusion? I have an RX 570, not the best GPU

main snow
#

No clue if that one's good or not

nimble light
#

What is the name of the Loca AI you said?

main snow
#

Ya mean the app itself? It's really just called a1111

#

Serves me well for pictures, haven't tried videos just yet

#

Still, if a1111 can't do it I'm sure confy or swarm can

nimble light
#

How mean like how do I run it? COmfy UI or what?

main snow
nimble light
#

Comfy UI doesn't work for me? Any good alternative?

main snow
nimble light
#

So swarm uses my GPU, but is web based?

main snow
#

idk man, never used swarm, i use a1111, ya gonna have to ask that in tech-support

#

i simply heard it was good

nimble light
#

Gimme link to a1111

main snow
#

as i said i'm not sure, never tried

nimble light
#

Just gimme

main snow
#

make sure to donwload the version

#

for amd gpus

nimble light
main snow
#

happy to help

nimble light
#

Wait

#

What do I do now? Idownloaded it,

#

Do I run Windows batch file

main snow
#

now ya install it, i don't remember the exact steps and it's probably a bad idea to tell ya from memory lol

#

read it's page, i'm sure the instructions

#

are there

sinful lantern
pseudo forge
#

i have downloaded stable diffusion for the first time and have downloaded a model off civitai but i am getting an error when generating, it says the model is for Pony not sd1.5 like in the tutorials i'm watching, could someone give me some insight/tips? thank you

faint timber
#

Revisit CivitAI and check the upper right-hand corner on the models page. There is a filters icon. From there, you can filter which model type to use. Skip Pony for now and try out the SD 1.5 first.

ancient mauve
#

Guys is there any way of increasing training speed without affecting the model's quality?

#

I have enough vram

#

Wtf is that link

#

I just want to increase the training speed it's just that I don't know what values do I have to change to use more of the vram

lean timber
#

Does anyone have recommendation for a SD model that is good at creating things other than characters? There's virtually an infinite amount on Civit.ai that are trained to be good at making people, but I want one that can create interesting spaceships, buildings and the like.

atomic mortar
ancient mauve
fervent thunder
#

hey new here

#

does anyone here run stable diffusion locally?

ancient mauve
fervent thunder
# ancient mauve I do

Im in the market of getting a custom built PC and thinking of getting a GTA 3000 or 4000 graphics card series with a good number of VRAM so I can generate something like 90s cel anime like sailor moon or berserk or evangelion, or even something like aqua teen hunger force style

#

what are your recomendations?

ancient mauve
#

Idk what is your budget

fervent thunder
ancient mauve
#

But aqua teen it's extremely simple artatyle so you shouldn't need that much

#

There must be someone here better suited for answering this

fervent thunder
ancient mauve
#

It should work

#

If I'm not mistaken having realistic generations is harder than cartoons in general

fervent thunder
#

my current computer is using a gtx 1060 6GB graphics card from 2016

ancient mauve
#

I see people within the 3000 series making good generations

#

But I'm not sure

#

You should aim for the Max cost-perfomance gpu

#

But dunno where does it fit that limit today

fervent thunder
ancient mauve
#

It's an user interface that helps you to make generqtions

fervent thunder
ancient mauve
#

I like forgeUI for generations

ancient mauve
#

Start the program and works at least for me

#

There are some YouTube videos with simple guides

fervent thunder
ancient mauve
fervent thunder
ancient mauve
#

I just take one model I like from civitai, put it inside the stable diffusion folder and it's usually good to go

fervent thunder
ancient mauve
ancient mauve
#

But if you have the money try to get something good

chilly heron
#

@oblique elk i sent it in the other general chat can you take a look

quiet mason
#

quick question
i want to use the hunyuan v2v model where i upload a video of my own and give a prompt which it then changes accordignly and gives back to me

but to install hunyuan do i just install it normally as a model and then get a workflow for v2v or is there a specific v2v model i need to install?

atomic mortar
#

Higher vram however allows you to run it faster

#

My 8gb vram back then was about 20-30s per image with 1-2, loras

small citrus
#

(I am unable to send any image in this chat) Hello everyone I am technical officer at genotek, a product based company that manufactures expansion joint covers. Recently I have tried to make images for our product website using control net ipadapters chatgpt and various image to image techniques. I am giving a photo of our product. This is a single shot render of the product without any background that i did using 3ds max and arnold render.
I would like to create a image with this product as the cross section with a beautiful background. ChatGPT came close to what i want but the product details were wrong (I assume not a lot of these models are trained on what expansion joint cover are). So is there any way i could generate environment almost as beautiful as (2nd pic) with the product in the 1st pic. Willing to pay whoever is able to do this and share the workflow.

fervent thunder
#

int4 flux is 6.64 GB so 8GB is ok

atomic mortar
vale furnace
#

Does anyone have the detail++ Overall Detail SD1.5 embedding? It was removed from civitai and was curious where I can find it

oblique elk
exotic sphinx
#

It's pretty nuts the optimizations some webuis do

#

I myself used to use a laptop GPU but that was with SD1.5

#

2060 was with Illustrious

#

Tho admittedly no clue about Forge, I use ComfyUI

fervent thunder
#

I use CPU a lot even its fine

#

its doesn't have to be fast

shell peak
#

Hi guys, not sure if I’m allowed to post this but I am looking to hire a LoRA trainer

tardy swan
#

Yo

#

How can I get in contact with Ronaldo

opaque horizon
#

Hi

#

how can I make spiderverse style images with dreamstudio ai

#

??

#

and animate them

ashen sleet
#

yo

ancient mauve
#

do you guys have a good to go formula depending on the number of images?

sinful lantern
#

does anyone of you know how to generate image of width = 4096 and height = 2752 from sdxl model, and if it is then how i can do it and how to multi gpu infernce in sdxl

still glacier
sinful lantern
#

seee i generated the image of 2048 x 2048 with sdxl

#

liike i was using kohya inference sd-scripts tpo generate imag of this size

#

actuakky i generated image 2752 x 2752 too

#

so its not like its just limited to 1024 x 1024

#

have anyone of u saw inferernce scipt given in kohya repo

oblique elk
fervent thunder
#

these days the models are getting so big that the larger models like cosmos or stepfun come with their own multi-gpu scripts as well

sinful lantern
#

can u give me documenation or provide some link to it

fervent thunder
#

yeah its this

placid hatch
#

What is the current stable diffusion experience for supported amd cards? Trying to get an idea of whether or not the 9070 is worth it for me when it gets full support

fervent thunder
#

if you are willing to write custom kernel, driver and compiler code then amd can be ok

#

otherwise rly not

atomic mortar
#

iirc amd is useable upto sdx

#

sdxl but you gotta pull some tricks for upscaling etc

placid hatch
atomic mortar
#

tiled upscaling, zluda etc

weary socket
#

Umm

atomic mortar
#

scam dont click

weary socket
#

Yes

vagrant meadow
#

Hello, I'm new to SD and want to buy a GPU. Would it be better to buy a 12GB RTX 3060 or would it be worth spending $250 more for a 16GB RTX 5060 TI?

atomic mortar
#

More vram the better generally

#

And considering the 5000 series is newer you'd get warranty too

vagrant meadow
#

I don't live in USA and won't be able to claim warranty if anything goes wrong.

Would 4GB more be worth the $200-$250 more?

#

@atomic mortar

atomic mortar
#

Depends on how much 200 dollars is worth in your country

vagrant meadow
placid hatch
#

16GB can make a big difference, but may still not be enough depending on what you intend to do

vagrant meadow
slim jacinth
#

hello 😄

tribal crown
#

what is the fastest workflow for wan? i am running it through comfyui and the least i can get is 5 minutes, i have a 4090. any faster workflow please? and maybe same quality. thanks 🙂

merry ginkgo
#

Is there a quality program for downloading models off CIVITAI in Bulk? Looking to fill an 8TB hard drive with at risk models

ancient mauve
#

anyone up?

atomic mortar
atomic mortar
#

swarmUI has a downloader integrated with civitAI apikey support

merry ginkgo
atomic mortar
#

Well if you want to download all models you'd have a gigantic task ahead

merry ginkgo
heavy yacht
sinful lantern
#

do u people really trust comfyui?

#

i really feel that the comfyui generated results are not goood because when i used to generated images with the help of comfyui its results are not good at alll rather than if u use raw code to generate images they are very amazing thats what i have observed

#

we cant blindly trust comfyui at all

atomic mortar
#

Forge however uses obfuscated code to hide they used code from comfy

sinful lantern
#

yes ofcourse i m just saying like we cant blindly trust comfyui for genrating images, for example if someone is finetuuning a sdxl or sd3 model investing so much and they generate results from comfyui and if the results not come good they will think that training didnt went well

#

because thay happened with me

#

*that

#

but when i used raw code to generate images results changed so much

atomic mortar
#

Just a case of samplers and schedulers

sinful lantern
#

dis agreee

#

it cant change that much

atomic mortar
#

You know comfy is just a bunch of nodes (code blocks) transmitting data to one node to another

#

You just made your own workflow

#

Tedious but workable for you

sinful lantern
#

its not tedious i guesss, i just mean that we cant trust it blindly

valid aurora
#

hi guys does anyone know any checkpoint that goes well with Vroid models? im still noob and my loras suck, and when i use different check points for Vroid the generated image style changes so much. also how imporant are regularization images for loras? ive been at it for a week but all my loras look like deformed monsters especially the eyes xd

#

sd 1.5 btw

ancient mauve
#

can anyone please help me with some settings, Im getting crazy

#

and peopel instead of trying to help they just start yapping at you

#

I just need to understand somet things Im having trouble with

#

no one?

oblique elk
# ancient mauve no one?

If it is a technical problem feel free to ask your questions in the support channel if you have problems with prompting or prompt settings ask in the prompt channel. …. If someone can answer your question they will.

oblique elk
ancient mauve
#

I just want to compare some training settings

oblique elk
# ancient mauve may I assk you what training program do you use?

Great question and would fit perfectly into #🔧｜finetune
And pretty sure others already recommended you Kohya, onetrainer and fluxgym.
The parameters are not comparable as it depends on so many different factors. Amount source images, concept / style or object. Close to the source model or far away. Overfitting useful for later generation etc.

ancient mauve
ancient mauve
#

I find settings that dont work for another trainer

#

and its a bit confusing

oblique elk
# ancient mauve I find settings that dont work for another trainer

Yes, copy and paste does not really work for training. At least if you do not want the exact same result as the tutorial, blog,… and you use the same input files and descriptions.
So no other choice by reading and understanding the effects of the different parameters. And then start with trial and error with a small and good labeled dataset.

ancient mauve
placid hatch
#

Does anyone know why refine/upscaling uses up so much more vram on rocm compared to normal image generation or the equivalent process on nvidia cuda?

faint timber
#

I would guess because internally it is upscaling the latents, which uses more ram.

tight yacht
#

Anyone proficient with Stability Matrix and in general running models locally? I need some guidance on a couple of issues I'm encountering with certain models...

main snow
main snow
fervent thunder
#

especially the typical consumer rocm setup

placid hatch
#

it seems to fill up my vram even if i am only upscaling by a really small amount. It makes me wonder if its some kind of bug

fervent thunder
#

possibly yeah

#

only a tiny % of AI code is memory safe

somber gale
#

Anyone good with hunyuan?

#

I wanna make a 3d model but ive had no luck

plush gale
#

My name is Ziliah and I had a few breakdowns and pretty much decided to rebuild my life by being fit and healthy, yk

I like listening to music and reading novels
I currently own an online store that brings in some amount per week because I am passionate about helping my parents retire and achieving financial freedom.
I'm open to friendly conversations

sudden karma
#

does anybody know a easy way to ai generate images for a book? it doesn't even need to bee the whole szene, one object theoretically would be enough, although the whole scene would be more impressive.

#

it would need to be some kind of text2text generation first (book2prompt), afterwards prompt2image

#

i would need one picture per page

atomic mortar
#

hmm that would just be text to image though?

#

like you describe the object or scene

oblique elk
# sudden karma does anybody know a easy way to ai generate images for a book? it doesn't even n...

You might need to use a llm (ChatGPT, Gemini,..) to summarize the page and to focus on main subjects / main objects. Then let the llm create a prompt and your favorite image gen ai creates the image.
BUT for a bit of consistency I would add some style tags to avoid each image looking different (comic, real, anime, sketch,…).
Additionally you would need to remember the last object to avoid repeating images on each page if the text main object overlaps over many pages.
Another BUT if your book has a main object/ topic/ character that appears on multiple pages you will not get similar images. So a reader might be confused why the brunette woman from the start is now blonde. …

slim cove
sudden karma
#

flux.1-dev onnx fp4 (for scenes) and ssd-1b (for single objects) are the models i use

#

Would it be easier—or even possible—for an AI to pair a scene from a book page with the corresponding sharp frame from the movie adaptation? If so, how could that even be achieved?

sudden karma
atomic mortar
sudden karma
#

i don't want to manually prompt 300+ images

atomic mortar
#

Ahh thats the gist, yeah you need a llm for that to somewhat get a prompt out of it

sudden karma
#

that's what i'm asking for. an ai generating the prompt from a given book scene

atomic mortar
#

Chatgpt or a local one could do the trick yeah

sudden karma
atomic mortar
#

Even copilot could somewhat manage if its below 3k letters

sudden karma
atomic mortar
#

Hmm youd need a solid system prompt

#

What kind of style of image would you use?

sudden karma
#

not even when you include a manual description of every person/setting

atomic mortar
#

Realistic? Anime?

sudden karma
sudden karma
atomic mortar
#

Hmm flux and sd could work

#

But hmm extremely long prompts arent image gens strong suit

sudden karma
#

that's what i'm struggling with

atomic mortar
#

As it will just ignore certain elements after a certain point

sudden karma
#

book page 2 short image comma seperated gen prompt

atomic mortar
#

Afaik theres no solution for that yet as far as im aware

#

Ahhh thats an xl prompt

sudden karma
sudden karma
#

thanks for your help anyway : )

atomic mortar
#

Well the only way would be manual work

#

Hmmmm if its single item/people prompts however

#

Then you could maybe, ask the llm to highlight the most important person or object in the page

sudden karma
#

or just a simple szene, just needs to be a background image for every single book page

atomic mortar
#

Yeah you could do that

sudden karma
atomic mortar
safe cliff
#

Ironically for complex things I found it was easier to 3d model what I wanted manually then img2img to change the style

atomic mortar
#

Your prompt should start with the first sentence setting overall context and containing the model name mentioned by the user. The overall prompt should account for location, model overview, expression and pose, angle of the shot, placement when the model is an object/product, lighting, colour palette, and styling.

#

Could be a decent start as a base prompt for gpt

sudden karma
#

the scene thing... someway to get every different place/setting shown in the movie, image2image generation to remove everything in the centre of the image, and a simple python script to use those images as pdf background

sudden karma
#

i'll try later

atomic mortar
#

Youd have to edit it though a little but i hope that gives you a way further

oblique elk
# sudden karma tysm!

Still thinking you could look tools like dfiy. They allow you to chain up different ai tasks (llm, scrape, image gen ). So you would create on ai agent to determine the key element of a given text (system prompt etc. ) then pass the result to another llm agent creating outstanding sketch image prompts with the input of the first agent. (Could even be 2 different llm ). Finally send the prompt to the third agent which uses for example a predefined comfyUI workflow to create the image.

iron flax
#

Hello

alpine hazel
#

Hello

#

If you have the option to work at home, please contact me. I understand how things can be challenging now and I want to help you find an appropriate opportunity. There are many vacancies to apply 💌

pulsar crane
#

Hello

charred wadi
#

Hello

ancient mauve
#

is there any channel here to discuss trainings?

#

or another server for it

#

I dont want to spam this channel

atomic mortar
#

Fine tune channel got removed a while ago

ancient mauve
atomic mortar
#

And tbh even if you spam here nobody would mind

#

Its not like theres a big discussion going on

ancient mauve
atomic mortar
#

The only big active ones are anime and techsupport lmao

main snow
atomic mortar
#

I know, its one of the few channels im active in lol

main snow
atomic mortar
#

Once? All the timedeadbread

main snow
#

Well I only saw it live once

atomic mortar
#

I only wish that people like donny just use off topic for their vague resume job search

#

Instead of posting it twice a week everywhere including techsupport

main snow
#

Yes, not sure the chat is a good place for Job begging

atomic mortar
#

Just use fiverr or community's made for it ngl

main snow
#

Or apply to McDonalds

#

Put the fries in the bag as they say

atomic mortar
#

Oh cool literally what we just talked about

ancient mauve
atomic mortar
#

no?

atomic mortar
#

'>??

#

not the server

#

the sever is a scam

ancient mauve
#

Ahh ok XD

atomic mortar
#

the stable diffsion channel is not lmfao

ancient mauve
#

I just need some place where I can discuss training settings with people

#

I got good generations but they could be better I think

atomic mortar
#

iirc werent you banned in the one trainer server?

#

ah i see we dont share this one server

#

imma link it to youi

ancient mauve
#

but I just need to ask some questions

#

something with some back and forth

oblique elk
ancient mauve
#

this is the kind of thing that pisses me off

main snow
ancient mauve
#

Im gonna ask here too if its ok for you now that Im here

# 
do you know that some games put the same exact image but with different face expressions for dialogue? like a visual novel
I was wondering if feeding all of those images helps the training
because they are the same images but not really; if I tag the images all the same except for those face expresions, the model will learn to differentiate?```
main snow
ancient mauve
#

more than others

main snow
ancient mauve
main snow
#

Not what heard, but hey the best way to really find out is try it

ancient mauve
#

because otherwise, if I can use "variation" images as completely new ones, my dataset increases from couple of dozens of images to hundreds

main snow
#

I remember hearing like 20 or so

#

Should be enough

ancient mauve
#

if you want to do a full finetune you need more

main snow
#

Not sure(?) my stuff comes out perfectly well

ancient mauve
main snow
#

Tagged you on the chat with images for example

ancient mauve
#

so I suppose I have to change some tags and images

main snow
#

Far as I remember reading, that char's Lora had about 20 or so pics

main snow
ancient mauve
#

like 10-20 the bare minimum

main snow
ancient mauve
main snow
#

Ofc the more variety the better in Theory

main snow
#

Works fairly well

#

Not flawless I'm sure

ancient mauve
main snow
ancient mauve
main snow
#

Not optimal, but it's a way to do it if you got low sample numbers

#

What about those?

ancient mauve
#

like in these cases its clear that the artist amde 1 image and edited a bunch of face expresions with layers

main snow
#

Yes, that should work perfectly fine

#

In fact a lot of facial expression variation is good

ancient mauve
#

and its like 90% of the image the same, except for the faces

main snow
ancient mauve
#

dunno let me search for some quick examples

main snow
#

If it's a known char I suggest getting fan art, screenshots and much more

ancient mauve
main snow
#

Preferably in different styles

ancient mauve
#

not only teh face but a whole body

main snow
#

So it doesn't super glue to 1 style

ancient mauve
#

is it better to crop and get the face only or just feed the whole image

main snow
#

I'd feed it as is

#

I think it should be fine

ancient mauve
main snow
#

Ofc, assuming each of those is 1 pic

#

The same char multiple times like that in the same pic... Not sure that's a good idea

ancient mauve
#

then the thing is tagging so they have the exact same tags except maybe 1 or 2

ancient mauve
main snow
ancient mauve
#

just imagine that in the example I gave you they are separated in 4 diff images

main snow
#

Yes, that should be fine

ancient mauve
main snow
#

Indeed it do

ancient mauve
#

do you have any tips for tags

#

I just use WD14 then filter some things that could be wrong

main snow
#

Well, best tip I can give you is making the trigger a unique word

ancient mauve
main snow
#

Also the most words possible.

Some peeps only use 1 word for the whole thing and then it's a mess if you wanna change literally anything

ancient mauve
#

for a character

main snow
#

Like removing a hat

main snow
ancient mauve
main snow
#

Or if you're brave lol short hair

ancient mauve
#

do I remove black hair and spiky hair then?

main snow
#

Do whatever you find better

ancient mauve
#

do you have some examples you trained on?

main snow
#

Trial and error

main snow
ancient mauve
#

as a general rule,you have to remove the tags you want the model to learn right?

#

if you are training dunno, a goku character, you remove all the tags that define "Goku"

#

am I correct?

main snow
#

Remove? I don't remember doing so no

#

Honestly, ya should just try the same tutorial I use lol

#

Used*

#

No expert by any means, simply had a guide to make mine

#

Or use Civitai to train it

#

There it's not free costs buzz

#

But it has an amazing in-depth guide

ancient mauve
#

btw is there any reason characetrs eyes look weird?

#

like the bidy usually looks ok but the AI seems to have problems with the eyes, I see this with many models

main snow
#

it's amazing to fix eyes, give a lil more detail and fix artifacting

atomic mortar
#

Yeah lack of hires or adetail

ancient mauve
#

whats exactly highres fix

atomic mortar
#

And if it's isn't a close-up it's not gonna look great without it

main snow
#

assuming you're using a111 like me there's a lil box you can tick to use it

#

there's many upscalers tho

#

i suggest remacri

#

never disapointed me

ancient mauve
main snow
ancient mauve
main snow
ancient mauve
#

no I mean for a chgaracter lora

main snow
#

for example if a char has a hat and there was no word for it... that shit ain't comming off lol

ancient mauve
main snow
#

but ya can try both and see

#

which works better

#

sounds like a case of trial and error to me

fallow sky
#

Hi

main snow
#

Hello

abstract quarry
ancient mauve
#

you cant train on a single image only

abstract quarry
#

you can

#

you can also train iteratively: build a initial model by training on a single image. use this model to generate a larger set of training images (using e.g. controlnet to increase variety) and then train again

ancient mauve
#

I have tons of those and I f it helps if I tagg them correctly then Ill do it

abstract quarry
#

it might help to teach the model which prompt stands for which expression

#

but I don't think it helps much with training a single face

ancient mauve
abstract quarry
#

looks like rpgmaker face sets 😅

ancient mauve
#

many Vns and games do that

#

if I feed it woth more different images it should help the training right as long as tags are good

#

soI increase my dataset like 10 times

abstract quarry
#

what do you want to achieve?

#

I don't think that these face pics are complicated to generate with flux. I don't think you need a lot of training data

ancient mauve
#

I already have good results but I want to improve it

#

its just that I dont have that many images but I do have that many images if I include teh face expressions avriations

viscid hollow
#

hey

ancient mauve
ancient mauve
#

Btw does anyone know the purpose of image repeats in a dataset?

#

I know this option makes an image be processes that X amount of times during an epoch

#

But I don't know what does that really mean for the training

#

Why is it a thing?

oblique elk
# ancient mauve Why is it a thing?

You can use it for weighting. Let’s say 10 of your images are great you might repeat them 4 times. 10 are good you repeat them 3 times. And 20 image are for reference but not your training goal you might repeat them only one time. So you make sure the training is more biased towards your images.

ancient mauve
#

I've heard this is also good to increase if you have a small dataset vs a big one, to compensate

#

What I don't really get is that this logic points to a standard N° of steps threshold everyone is trying to imitate

#

But I thought there was no formula for this stuff

abstract quarry
#

let's say you train a model on a specific style and your training images are men and women. You have 30 images of men but 90 images of women. So you set repeat on men to 3 such that in average the model is seeing the same amount of men as women

#

if your model would see much more women than men then it would get biased towards women. You see this very often in anime models where the model generates a women even if your prompt asked for a man

#

however, anime models are extremely biased. Having a 1:3 ratio would not induce such a strong bias

livid gale
#

guys that the hell is DPM++ SDE CFG++

#

i cant fint it anywhere

#

krita diffusion seem to not have this sampler

#

there is dpm++ 2m cfg++ and dpm++ 2s a cfg++ but not sde

#

ohhhh its only comfy and reforge thing

ancient mauve
ancient mauve
#

I think in some GUIs you can only set repeats to ALL the dataset, not just some images

#

Is that the same time basically or the logic behind it doesn't exactly work like that?

abstract quarry
ancient mauve
#

So I should have that in consideration while prompting and while tagging because it will have a bias with those repeated elements, like character poses

#

I'm only assuming, but I had some generations copy the poses of my database if not specified

#

For example, if I say "looking at front" sometimes looks from a diagonal, because my dataset has those, it is *technically looking at front (dunno if I'm explaining it well)

buoyant oracle
#

is 4070 fine for creating lora?

floral umbra
main snow
#

If you're willing to spend a lil you can also train on Civitai, that way you're not dependant on hardware.

Think it's about 5 euros to train a Lora

#

And far as I'm aware they do come out well (?) dunno, never done myself

atomic mortar
mortal bison
#

Morning

#

What kind of systems are you guys running? Because mine crashes on wan video generating (3080TI)

atomic mortar
#

5080, takes bout 10-12min

fickle flicker
#

Hi, when I try to use any name for example "GLOBE" the ai generates an image with incorrect letters. eg. it would make it "GLLOBE". any idea how to correct this?

atomic mortar
vestal jetty
#

Sharing this in case anyone else is deep into messing with GPT model and figuring out jailbreaks prompts. Its a list of prompts and models that the prompts can be used on. uncensored/jailbroken.

Not mine. Just figured someone else might want to check it out:https://shrinke.me/Z6zVWp
If the link dies, I’ll try to reupload it later. No idea how long it’ll stay up.

fickle flicker
atomic mortar
#

Flux dev /sd3.5 large is better at it but still not perfect

fickle flicker
atomic mortar
#

I recommend a textless logo or photoshopping after ngl

fickle flicker
atomic mortar
#

Ngl = " not gonna lie"

#

It's internet language

fickle flicker
atomic mortar
#

Do you know photoshop?

fickle flicker
#

yes

atomic mortar
#

Well youd need to photoshop said result to try and fix the text probably

fickle flicker
fickle flicker
atomic mortar
#

What gpu do you have?

fickle flicker
atomic mortar
#

Hmm im not sure if you could use flux dev, Maybe someone else could chime in for better text results

oblique elk
# fickle flicker 3060 12GB

Well depending on how you want to integrate your text within the logo. Flux and SD3.5 are ok for generating images with short text even short sentences.
Another way would be a simple depth controlnet with the text as source and for example sdxl as output.

fickle flicker
#

@oblique elk thanks for the alternatives

ancient mauve
#

do you use basic and specific tags or only specific tags for prompts?
like eyes, blue eyes

#

with automatic taggers I get many clothing tags

#

and I dont know if I have to leave coat, white coat, white long coat or only white long coat

ancient mauve
#

@oblique elk you know when you use wd14 you get tags

#

Or while prompting, you can get tags that include other tags

#

Maybe eyes is a bad example

#

For example if a character has a white coat, do you put coat, white coat

#

Or do you only put white coat

#

What is better?

atomic mortar
#

Ahh bottom one

#

Personally

#

Both work

#

Its trail and error ngl

#

No one size fits all solution

ancient mauve
#

Maybe if it's a very complicated images, too many tags is not good

#

Or maybe more detail is better, but if a model is already trained it should at least have some kind of knowledge right?

atomic mortar
#

Or coat

ancient mauve
#

XD

atomic mortar
#

But honestly tag based prompting is nice for smaller prompts to keep it accurate as possible

#

But once you get complex prompts, cant beat natural language

#

Illustriousv3 will support it

#

Illustriousv2 works sometimes with it but also does tags

#

Illustrious v0.1 does only tags

nimble slate
#

Hey

#

i need a bit of help

#

ded chat awww

#

😦

atomic mortar
nimble slate
#

i wanna stream

#

i just wanna know how workflow works

#

🤷

atomic mortar
#

you wanna stream AI?

nimble slate
#

no i need help with setting up workflow

atomic mortar
#

well if you wanna stream comfyUI workflows im afraid your gonna need to dive into it a little

nimble slate
#

stable diffusion

atomic mortar
#

though if you use forge its pretty much working straight out of the box

#

yes thats a model

nimble slate
#

i just wanna make some spaceships 😭

atomic mortar
#

comfy/forge/swam/foooocus are the US's using the model

#

what GPU do you have?

nimble slate
#

uhh

atomic mortar
#

its important to know before i recommend anything really

nimble slate
#

NVIDIA GeForce RTX 2060 SUPER

#

and i have 16 gigs

#

i can run stable diffusion

#

but idk how to make images 😭

atomic mortar
#

a 8gb vram is working for xl

nimble slate
#

i just need help with one thing 😭

#

its not that serious

atomic mortar
#

its not just one thing lol

nimble slate
#

oh

atomic mortar
#

you need help installing the UI it seems

nimble slate
#

i wish i could post pictures

#

can u vc?

atomic mortar
nimble slate
#

k

atomic mortar
#

yeah i see the image

#

the video will help in this case 👍

nimble slate
#

posted

chrome jacinth
#

why is stable diffusion only capable of generating one image per prompt?

atomic mortar
#

I can make 3 images concurrently but i save about 0.7s or cost 1.5s on my 5080

#

So its not really worth it to do it like that on consumer hardware

thin helm
#

what settings are great for flexible character illustrious lora?

flexible as can handle more loras, styles, concepts, etc? and not getting baked..

#

im using civitai's trainer btw

noble tiger
#

Hello everyone!! So i have some questions about stable diffusion models. What model do you think is the best and can you recommend me one?

atomic mortar
#

You want anime? Illustrious type models, you want realism? 3.5 large/flux

#

Dont got the vram? Sdxl

#

If you got less then 6? Well i think sd 1.5 is pretty cool still

olive lintel
#

hey guys!

I'm akash and i head bd at Hive intelligence. We're building infrastructure for AI agents

have a few synergies in mind and would love to connect w the bd team at Stability AI to discuss this in detail

can anyone from the team help me connect to the right POC?

ancient mauve
#

I think someone here said it before, but tags are "aware" right? if I make a custom tag in my training or prompting likw potatocamp the AI knows about potatos and camps and it will have somme kind of influence over that tag

#

talking about wd14 tagging

#

so some tags dont have to resemble what is only shown on screen, but also concepts

floral umbra
ancient mauve
floral umbra
#

Yep. In this case, the trick is to ¨tell¨ the training that X thing is actually Y, so instead of saying itś actually porcupine needles, i tell it that the handfull of needles are ¨hairstraws¨, and the training learns that.

ancient mauve
#

so maybe randomly in one of my generations I get a random potato generated?

floral umbra
ancient mauve
#

with characters I suppose I could use a triggerword that already has the name of the character on it

#

like duno CustomGoku

#

it osnt really Goku but it will get characteristics from it right?

desert dagger
#

so you wouldn't use potatocamp - you would use p0t@toc@mp

ancient mauve
ancient mauve
#

ty

desert dagger
#

even a . in your prompt has an effect on what the AI draws. or a , or a ; <--- those are noise, but it's aware that: this is an apple. and this... is an apple aren't really the same phrase and don't really mean the same thing

ancient mauve
#

I have like a white greyis background for some images

#

do I use white background, gtey background, both?

desert dagger
ancient mauve
desert dagger
#

remember that the AI wasn't trained on color charts, or pantone colors. so try to stick with the common terms you'd find out there on the net

ancient mauve
#

using white background grey, background make sthe model mix concepts or does it confuse it

desert dagger
ancient mauve
#

it also generates more stuff in a simple model like sdxl

#

I dont know what to do, iare models able to mix colours

#

or is it better to stick to the color that resembles it more

desert dagger
#

stick with common color names that you'd find thousands of times in the google webscrape database

ancient mauve
#

but tags have information on their own as yous aid right?

desert dagger
ancient mauve
#

if the model you train on already knows what a white background or a grey background is

desert dagger
ancient mauve
#

like these are wd14 tags which mean there are already images with those tags trained on

desert dagger
#

and so that's likely a term in the database

ancient mauve
#

so if I have a greyiss white background, do I use white, grey or both

desert dagger
ancient mauve
#

if I want to recreate the background in my images as close as possible?

desert dagger
ancient mauve
#

concepts that are mixes of wd14 tags

desert dagger
#

think of the AI as a fancy sort of camera - you get raw footage out of it. you don't have perfect control and you'll need to do post production work

#

color matching like that is post production work

ancient mauve
#

you dont really have a tag to define them or the tags are too abstcat, in this case, I cant really knwo what grade of white or grey the tags reffer to

ancient mauve
desert dagger
#

give me one of those tags that you're using please.

ancient mauve
#

so while training, the model "changes" the meaning of tags a bit then

desert dagger
ancient mauve
#

so if for example you use white background in the base sdxl model, you get white backgrounds, but if in a training all images have a more greyiss background and then you train on those and you generate images with white background prompt, you will get more greyiss backgrounds

desert dagger
ancient mauve
#

and white hair

#

do I use white hair, silver hair

desert dagger
#

what do you consider a "wd14 tags"

ancient mauve
#

there are 2 ways of doing prompts

#

doing descriptions

#

or, doing, tags, like, I, am, doing

desert dagger
#

what do you consider a "wd14 tags"

desert dagger
#

that's not a phrase, with proper punctuation, so the ai won't look at is as a concept. it'll use the commas as noise

ancient mauve
desert dagger
ancient mauve
#

like, its something very common with AI image generations

#

for a while now

desert dagger
atomic mortar
#

pony and illustriousmodels are trained on danbooru etc no

desert dagger
#

forgot about illustrious - yeah, those are the models you want to use with those tags

ancient mauve
#

Was asking because that tag system isn't something fluid like a long descriptions

#

So you can't say greyish white if there is no such tag in the first place

#

So maybe mixing was the correct way to do it, if anyone knows please let me know

desert dagger
hybrid linden
#

Hello! In AI image generation field, what can you do with 24GB of VRAM you can't with 20GB? I have a choice between RTX 3080 Ti 20GB for 45.000 rubles (550$~) and RTX 3090 for 65.000 rubles (800$~). They are almost the same in performance, so the difference is in VRAM capacity and price.

desert dagger
fervent thunder
#

hey everyone how are you

neat solstice
keen parcel
#

UAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

#

UAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

#

UAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

#

UAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

#

UAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

#

UAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

#

UAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

#

UAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

#

UAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

#

UAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

humble dagger
#

Hey uu

main snow
#

waz good

analog ingot
#

I'm building a visually-driven lifestyle brand and I’m looking for an AI image expert to help me generate highly realistic, on-brand product photos using tools like Midjourney v6, DALL·E 3 (ChatGPT), or Flux AI.

Our products are luxury-inspired fashion items (bags), and we need a consistent, editorial-style visual direction that feels clean, elevated, and believable.

This job is part hands-on creation and part training

desert dagger
fervent thunder
#

Hi everyone

#

Hey! I’m building Claity ai, an AI tool that will integrates ChatGPT, Claude, Gemini & more in one place with smart prompt routing.

Just launched the landing page and would love opinion or support!

Join the waitlist here: https://claity.netlify.app/

golden drift
atomic mortar
#

You know i was about to make a snarky comment about all these advertising posts but it's the most we get here lol

velvet fox
#

Helloo

atomic mortar
#

Yo

main snow
#

so, a1111 can't do text to vid right? preferably being able to use ILLU stuff

#

if not, can forge do it?

#

i'm trying everything to not install comfy bro lmfao

oblique elk
#

Sooner or later we will get you. Come to the dark side of comfy users and node torturers

oblique elk
main snow
#

so forge can't either? fuuuuuuuuck

oblique elk
# main snow so forge can't either? fuuuuuuuuck

No quite sure. To be honest I am not very patient, so I prefer a “UI” which is designed to be able to add fast new model support. Long time comfy user. Especially T2V models seems to be released nearly every two weeks.

main snow
#

i shall google it

#

now that my masterpiece is finished

#

doesn't seem like it. videoless i stay

#

ain't worth the pain lol

still glacier
#

deforum, animatediff, etc etc there are ways to do video on a1111/forge. I don t think there is any support for wan and hunyuan stuff however.

main snow
#

shame, hunyuan seems to be able to use illu loras, stuff so i was looking forward to that one

still glacier
#

so yeah do it the "old way" or switch to comfyUI

main snow
#

neither, i shall remain videoless

still glacier
#

valid choice

atomic mortar
main snow
atomic mortar
#

Valid but your loss for video ai

main snow
#

a sacrifice i'm willing to make

kind yoke
#

Hey guys anyone got edu maıl? currently studyıng ın unıversıty can you reach me pls

arctic ridge
#

i'm saying hello because discord made me

cobalt hill
#

Sorry for mmy basic question but how do i generate a image? Which channel?

#

It isnt free?

main snow
main snow
tidal canopy
atomic mortar
#

While correct it's dependent on your hardware

floral umbra
still glacier
#

oh yeah that had me not touching comfyUI for a while too. I ended up making my own half baked node for previews. Never released tho, code was wack

floral umbra
chrome jacinth
#

why is SD 3.5 Large not working? It's only generating empty images (none at all)

main snow
#

been using hem for a while... rarely tho, since illu there's barely a need for it

#

pony is miserable without hem tho lol

main snow
floral umbra
chrome jacinth
main snow
chrome jacinth
peak musk
#

@karmic brook is there ever gonna be a v4 v 3.5 is nowhere close to the top anymore neither closed or open

spiral tartan
#

how does this exactly work

bleak matrix
#

Good afternoon everyone! 🌞 How are you all today?

honest prawn
#

i'm ok, how are you?

stiff night
#

Hello, has anyone using Flux NF4 All in One models in Comfyui been getting errors on their workflows since last updates? Error: mat1 and mat2 shapes cannot be multiplied (1x1 and 768x3072)?

vestal sand
night sonnet
#

Fooocus, it generated me an image based on certain prompt, how do I get more variations of this single model from the photo. This faceswap shi aint working or i just dont know how to use it. Anyone willing to help

livid gale
#

i literally do not believe that images in chat gpt image generation gallery on civitai are ai

#

i refuse to believe it is ai

#

it is literally not fair at this point

#

Results are absolutely mindblowing even if you write something like "make colorful manga with coherent story" in prompt

night sonnet
#

yeah crazy, maybe u know answer to my question

stiff night
# vestal sand your doing it wrong

They where working fine, before the update 5 days ago, do you have a simple workflow that I could test out? Have used some and all give the same error.

carmine heath
#

Hi

radiant escarp
#

I'm new to Stable Diffusion (SDXL A1111) and am thorougly enjoying figuring out all the ins and outs of the tech, installation, etc. After a couple of weeks of learning, I'm finally creating images that I'm please with. Here's my question: with the proliferation of online AI text-to-image sites, what are the advantages of SD beyond the offline/local genration and no censorship?

viral zodiac
#

Hi there y'all! Can I delete the .cache folder or move it to another drive? Right now it's taking almost 20GB of space from my C drive and I'm just wondering if I can move it to the same drive where I instaled SD WebUI Forge if it's not advisable to delete it.

main snow
#

it's always slower for us consumers and all that, and we are restricted to whatever our hardware can handle

#

but ya know

#

NO FILTERShabby

warm junco
hearty violet
#

howdy all 🙂 i have a few servers here at home i run proxmox on, and am thinking about making a vm to run stable diffusion with. im just wondering what GPU i should pass through to the vm. currently gpu's i have are as followed:

Tesla K80 12gb x 2 (Dual GPU card)
Quadro M4000 8gb
1060 6gb

id like to use the the tesla gpu but have seen talk of the Kepler architecture maybe not being supported for alot of things. ive debated buying the M40 instead because of this.

idk what do yall think? or should i be trying for something Pascal or newer?

paper nymph
#

hi hi !

viral zodiac
abstract quarry
echo inlet
#

I've been out of the loop for a while, have there been any large/notable improvements in the world of open source image generation in the last couple of months?

floral jay
#

should i put vertical borders on both sides of a vertical image to make them fit a square for lora training

#

and tag it as "pillarboxed"

#

or will that screw the training over and put out generations with borders

steady mountain
#

hi what is this servser about

atomic mortar
floral umbra
#

Just tested LXTV, and after a few gens, it feels like it's basically "animatediff done properly, while wanvideo being a full on video-only gen Thonk

trim garnet
untold moth
#

How to get model from seaart?

#

The non downloadable one

dawn mulch
#

What's the difference with all the recent ai video generators compared to creating videos using SD. I think SD usues animediff?

#

I am assuming with SD its only good using img to video, or video to video, basically it can't really generate animations on its own?

valid aurora
#

hi just wondering how important is resolution for training images for lora (im not talking bout the parameter setting) ? im using 768 x 1465 for some of my training images, is this enough ?

woeful pasture
#

Hhgg

wet kelp
#

Hello !

rocky dock
#

Hi! I’m new! Is it possible to create images from text directly here on discord?

abstract quarry
#

with animatediff you see the image (in particular background) changing all the time

#

also, video models can do stronger changes and animations without getting incoherent

atomic mortar
#

we mostly run it locally

abstract quarry
#

yeah, you don't want to use Artisan... ^^°

atomic mortar
#

imo the civitai site one is a billion times better

abstract quarry
#

the whole idea behind Stable Diffusion and Flux is that you can run it on your own gpu.
If you want to do it online, there are so many websites offering strong models like Sora, Midjourney, Flux Pro

#

Stable Diffusion 3 is not competitive to those in my opinion

#

if you want to start generating images on your own graphics card, look for tools like

  • SwarmUI
  • InvokeAI
  • Reforge
exotic sphinx
humble iris
#

where can I find 'Insert Anything'? - couldnt find in ComfyUI manager

floral umbra
feral root
#

Hello

night gladeBOT
#

how to generate a picture
No data source is currently selected. Please choose a data source from the dashboard and try again.

wind saddle
#

Hi there

round siren
#

Does anyone know how to take an image, preserve a central part of it, and then use the prompt to build things around it or something?

valid aurora
#

hi guys im just wondering im using sd 1.5 and my stupid 8gb vram pc keeps crashing when upscaling and i dont have the money to pay for cloud, do you think its better if i save money and buy a 16gb vram or a 24gb vram ? i know 16gb vram is enough, but if 24gb vram is much faster and can do some other things that 16gb vram can't i will consider buying it. any feedback appreciated 😄

warm junco
valid aurora
#

rtx3090 and automatic1111

warm junco
#

A 3090 24gb

valid aurora
#

ah wait might be wrong i will check again

#

but im sure its 8gb vram

#

rtx 3060 - its weird bc it says its 12 gb in google but thats what pops up when i task manager

warm junco
valid aurora
#

no, computer

warm junco
#

OK strange a 3060 should have 12gb and a 3060ti has 8gb

valid aurora
#

do you have any recommendation of what i should get ? i guess logically since 3090 is 24gb and its older than 4000 and 5000 models its the best cause only vram matters ?

warm junco
#

But we can fix your Upscaling problem too

#

You need to have --xformers --medvram-sdxl
In your webui-user.bat for the best performance

#

Then you shouldn't have problems with upscaling or using sdxl models

valid aurora
#

@warm junco omg ty so much i just checked the price its so cheap. you da real mvp bro. i thought i had to sell my kidneys for a new gpu xD

warm junco
#

Np ^^

warm junco
valid aurora
#

nice thanks for telling me this. i want to pursue it professionally so i have to consider this well. hayzzz why does everything need money haha

true canopy
#

can community guides remove ppl? because theres a scammer/phishing attempt in #🍥｜anime

main snow
#

true and real

#

tho if ya fall for that.. at that point ya deserve it lmao

compact thicket
#

Any one help me in stable diffusion fordge ?

#

Anyone know where to find the pre-github-removed reactor?

#

Where to find releases?

main snow
#

yall ever heard of "invoke ai" ?

#

apparently it's crazy good

#

been meaning to move on from a111 so might give it a shot

oblique elk
humble iris
# valid aurora <@161910921269084160> omg ty so much i just checked the price its so cheap. you ...

2080ti 22gb is almost twice cheaper than 3090 and its OC`ed so close to 3090 speed

https://item.taobao.com/item.htm?id=748089024629&price=2399-5300&sourceType=item&sourceType=item&suid=77cbc4c8-f9fa-4b5f-9f70-5101549950cf&shareUniqueId=30274277199&ut_sk=1.ZbpctLHuksMDAHBYyz1%2Fa4WB_21646297_1737560540959.Copy.1&un=1ee4f95d3b7867999eedd86203cf18d8&share_crt_v=1&un_site=0&spm=a2159r.13376460.0.0&wxsign=tbwOF5fKCvXW6J3Vh4X8IVinZRLhK8ed8D8BaYIOq7-pa8MQaCsidkWvIEYhtM_Xm0v277ohqUXPrF_3-FXgyFDDiUZtNviwoPkIr0jkfAeW6ew-qLt9PgolKZhev8_Kpoj&tbSocialPopKey=shareItem&sp_tk=cVZCYWVleVpNcU4%3D&cpp=1&shareurl=true&short_name=h.TONFhvv5woqGfdu&bxsign=scdpg4r_FHvue9stwmYeQNszoxADO7tGsp8BHmui8sljwxlnQm_8yDpPVx5pskjETPdXE3wCAc6btpyGaSAxuGZkXXWX1TBCPsyRYSTB9HotIh4qqvVtabRgW31Rp9tPdbE&tk=qVBaeeyZMqN&app=chrome

#

drivers work well too

worn musk
#

hello

livid gale
#

guys

#

my noobai just broke

#

i rechecked couple times and my settings are perfectly fine

#

theres a shit ton of artifacts on the image

#

my wai also broke......

#

what the fuck

exotic sphinx
livid gale
#

i never touched this setting

exotic sphinx
# livid gale checkpoint default

Yeah uh try grabbing the VAE from a checkpoint you know has a VAE embedded or just download the VAE as FP16 separately and load it (it's like 200MB it's not that big)

#

If you are using ComfyUI, also try re-making the nodes to see if there was maybe an update that broke the old ones

#

Especially the Sampler

livid gale
#

im using krita diffusion

main snow
#

Another scam

#

We better add some kinda anti -bot measures lol

atomic mortar
#

at this point id just block any link other then huggingface, civitai, reddit and github

stoic onyx
#

if any one can tell what settings should i use to improvise my time taken in generating images without costing quality in stable diffusion
currently using(automatic1111(ui),
models( juggernaut XL, realistic vision),
sampling method(DPM ++2m)
REFINER(checkpoint(juggerenaut XL) switch at 0.9)
sampling:- steps 80)

on rtx 4060 without using --medvram

atomic mortar
#

A1111 is hilariously outdated and forge is optimized for XL

#

80 steps is exessive imo

stoic onyx
#

what is best sampling method

atomic mortar
#

its different per model

stoic onyx
#

for juggernaut xl

atomic mortar
#

what version are you running?

stoic onyx
#

as it is producing better images from other models

atomic mortar
#

anyways for the jugg XL by rundiffusion recommends ```'
Recommended Settings(VAE is baked in):

Res: 832*1216 (For Portrait, but any SDXL Res will work fine)

Sampler: DPM++ 2M SDE

Steps: 30-40

CFG: 3-6 (less is a bit more realistic)

Negative: Start with no negative, and add afterwards the Stuff you don´t wanna see in that image.

VAE is already Baked In

HiRes: 4xNMKD-Siax_200k with 15 Steps and 0.3 Denoise + 1.5 Upscale```

stoic onyx
#

(juggernautXL_juggXIByRundiffusion.safetensors [33e58e8668]) where i can check version ??

atomic mortar
#

the name

#

use the settings i posted

stoic onyx
#

and should i run on --xformers --medvram

atomic mortar
#

no idea about the xformers but medvram could help

#

no idea though since your using a hella outdated UI

stoic onyx
atomic mortar
warm junco
#

And dont use any refiner

#

If you have 8gb vram then also add --medvram-sdxl

#

Then you have the best performance

barren ice
#

¯_(ツ)_/¯

late pawn
#

This looks fun

oblique agate
heady bane
#

Hello, I would like to use stable diffusion locally, does exist a package ready to install with all plugins and setup ready to create images ?

atomic mortar
#

For sdxl

round dawn
#

Hello

real path
#

I’m still pretty new to LoRA training, but I’ve had a clear plan in my head for months – now I just want to finally put it into action.
I really want to get things moving, but setting everything up (Kohya, WebUI etc.) is dragging on for way too long.

I’ve got 30–60 images, generated with Astria AI.
The face is nearly identical across most images, although there’s still a slight AI look to it. The poses are pretty repetitive – mostly frontal or simple angles, very little variation.

What I’m looking for:
Someone who can build a strong starter LoRA model out of this. Ideally, someone who can also give me a bit of feedback or guidance (I’d even be happy to chat on WhatsApp if that’s easier).

I’ve spent over 50 hours working with Kohya, Stable Diffusion, WebUI, etc. – but I’m honestly just not that fast with tech stuff. I need help speeding up the process and doing it right.

I’m definitely willing to pay!
Depending on the level of support and effort, I’ll compensate fairly and immediately.

I’m based in Germany, super motivated, and really ready to get things rolling.
Would appreciate anyone who’s open to help – thanks so much!

desert dagger
real path
#

Thank you for response can you explain that a Little Bit ?

valid aurora
#

question: does anyone know how to transfer drawing style into an image using controlnet ? for example image A(reference image): any ghibli style drawing, e.g., ghibli house, image B: keanu reeves = image C(generated image): keanu reeves in ghibli style

random reef
#

hello

floral umbra
#

Man, i love this SHA256 hash script that i had GPT make kek Scans all my models from checkpoints, embeddings etc to find duplicates as i'm in the 100's and 1000's of each lol.

#

Sadly it's quite slow because i'm scanning my server's model repo of the models at 100MB's, so it's gonna take a hot hour lol

oblique elk
valid aurora
#

@oblique elk thanks !

valid aurora
#

@oblique elk hi, sorry, could you go into more detail, like is what are the exact settings, like preprocessor and model. sorry im really bad and all the youtube videos ive watched seem to be about image reference + text prompt rather than my desired image reference + image

oblique elk
hollow pagoda
#

Whacky

modern badge
#

Morning yal

#

Just recently got into stable diffusion, learning all the tools on how to make it effective

#

Is it possible to have SD use an image onto another image creating a new one utilizing both?

#

I was able to create an image using chat gpt, and then merged the image that was created with a google image to create a new one

#

A detail that is important, is that the google image was a greatsword. I had replaced the ai generated greatsword with one I chose from a screenshot ( from an mmo )

#

Substituting aspects of an image with another is basically what im wondering.

main snow
#

through photoshop

#

than doing inpain

#

in my HUMNLEEE opinion

desert dagger
thorny bloom
#

is installing comfyui from pinokio the same as from the github?

real path
#

Thank you man can you recommend a Model for my project: I have about 50-70 Pics of a Woman generated as a lora from Astra.ai but I have Only the pics no safetensor data. And than About 20 pics further developed by me. Tried all sorts of programs, facefusion, kohya gui, comfy ui, but nothing really sensible came out of it 🤣

real path
desert dagger
rich ridge
#

hey!, I just joined and I am new here but I was wondering if anybody can help me out with stable diffusion, my pc is very capable for generation but things like hands and feet are still a bit wonky, just send me a dm if you are able to help me out please!

floral umbra
#

Oh, deleted as i was about to report lol

sour quarry
#

when making a concept lora how many images are you guys generally using for the training dataset?

uncut venture
#

Im trying to load flux 1-dev in cmfyui but whys it using my ram instead of vram?

atomic mortar
uncut venture
#

24gb

atomic mortar
#

Hmmm shouldn't be doing that

#

Did you happen to launch comfy in cpu mode lol

pearl sand
#

Is there a hub or place like a website or discord server apart from CivitAI where I can download Lora character models?

cloud notch
#

hello

abstract quarry
cerulean loom
#

Hi, I have a character that I generated and I want to use him for more pictures. How do I do this?

fluid jungle
#

Hello guys, i need someone help

#

who can teach him SD

#

Draw a raccoons with Hong-Kong egg waffles in they hands

#

SD don't understand what is "Hong-kong egg waffle"

atomic mortar
#

I'm gonna be honest and i don't know what a egg waffle is either but you could try image to image

#

Seeing its a cone shaped you could try generating a icecream cone at first, photoshop over a the egg waffle and do a img2img /inpaint to

#

Match the style

quick thorn
#

Hi everyone. is there someone who has experience in replicate API integration?

gaunt sparrow
#

are there any open source IMAGE GEN (not video) to SORA? that level of quality?

atomic mortar
#

Or you mean images at the quality of sora?

gaunt sparrow
#

open source models that are comparable to SORAs image gen capabilities

#

sorry i didnt phrase it properly

cerulean loom
#

I have a character that I generated and I want to use him for more pictures. How do I do this?

abstract quarry
#

but Sora is quite new. It's normal for a new model to be better than older models. Just wait a few months until new open source models come out

gaunt sparrow
#

can you recommend me a model to use? i am new to this

#

i set up an environment, but not sure what to download

abstract quarry
#

Flux is usually the best I would say. Except for anime stuff, where Illustrious and NoobAI are the most popular ones

#

But Flux + custom model/loras on civitai is usually giving you great results

gaunt sparrow
#

thanks 🙏

hearty wedge
#

Hi everyone, I'm new to generating images with AI, nice to meet you, I was wondering if anyone could help me get this ( animatediff_motion_module_v1.safetensors ) file so I can animate my drawings with AnimateDiff

floral umbra
real whale
#

My goal is to make a ai music video rhythmically ai arranged to my music. I am using a amd 6700xt but perhaps consider buying a used 3090gtx, when nessesary... you have suggestions what I can try to use for this? Thank you in advance

floral umbra
floral umbra
summer sedge
#

I came to ask about the best locally run video generation models of today. FramePack and LTX-Video, is it?... Everything standalone these days? No more SD-webui integrations like with animatediff?

odd matrix
#

hello all, new to this Stable diffusion. Still getting used to prompting. Been trying Layer diffusion with no luck yet. Seems it is broken. Maybe somebody knows a way to get it working better? You get the checker board backround and the image that is supposed to be transparent has the colors of the image smeared across the backround, but you do not get a transparency.

atomic mortar
#

But yes, stand alone nowadays

summer sedge
atomic mortar
summer sedge
#

the 1.3B model? (ComfyUI integration)

atomic mortar
summer sedge
# atomic mortar The 13b

I can only see 1.3b and 14b, and the latter already exceed 16 GB. Anyway, I'm gonna give the small one a try. Thanks for the tip.

atomic mortar
#

1.3b is comparable to the old ltx ngl

summer sedge
#

On RAM? Must be pretty slow alright. Maybe I'm gonna try that, too. Need to set everything up first. Switching from A1111 to Comfy just now.

atomic mortar
valid aurora
#

does tiled upscaling for stable diffusion 1.5 cause more deformation when upscaling ? im forced to use tiled upscaling bc of low vram, so ive never tried norml upscaling and want to know the experience of other people

ancient mauve
#

does anyone here use forgeUI?

faint timber
#

@floral umbra I took a look at the github for FramePack. I'm not sure what I need to download. Is installing from Comfy Manager enough to get up and running using the ComfyUI_RH_FramePack?

tight badger
#

Hiii guys, girls and queers !

flat spruce
#

Can anyone here help me set-up SD/ comfyUI on a Linux VW (win11) with AMD 6700xt GPU? 😅

#

Having issues!!! 😂

floral umbra
floral umbra
floral umbra
atomic mortar
worldly herald
#

Hi Everyone, 👋 new to this channel

atomic mortar
# atomic mortar i really should link this to people more ngl https://dontasktoask.com/ 

    Foobar123:

    Any Java experts around? 

This is bad form, for several reasons. What the person is actually asking here is,

    Foobar123:

    Any Java experts around who are willing to commit into looking into my problem, whatever that may turn out to be, even if it's not actually related to Java or if someone who doesn't know anything about Java could actually answer my question? ```
atomic mortar
worldly herald
#

Quick question - is there a recommended channel/thread on this server (or elsewhere) to look for freelance SDL devs for an upcoming project?

atomic mortar
worldly herald
atomic mortar
#

granted you might have more luck on fiverr if you want to hire somebody

worldly herald
#

Got it - thank you

atomic mortar
#

hmm the more specialized servers like L3 or Bandaco have more power users

#

here its mostly people asking basic questions or how to get started running it them selves nowadays

worldly herald
#

ahh ok

#

Really appreciate the advice

floral umbra
atomic mortar
#

But ngl i hate discord as the normal forums have died and you can't tell people to just Google it anymore

floral umbra
#

Nope. Literally 20% of the time when i google up an issue, i find the same problem asked from half a decade ago, and first comment says "just google it".. Like dude.. i did, it lead me to your useless comment lol

atomic mortar
#

I should make a basic flowchart / faq since its mostly the same question being asked over and over

#

Maybe a PowerPoint so they cant accidentally skip a step

#

Bros posting this 3-4x a daycryingbread

floral umbra
atomic mortar
#

I got a project that pays in experience and advertisment

#

So basically i want this app...

#

Woah the first message that isnt the advertisement

#

I thought he was a bot