#šŸ’¬ļ½œgeneral-chat

1 messages Ā· Page 174 of 1

copper crystal
#

You can do that with photoshop or any other traditional editor. Those kind of edits are what graphic design has been all about for decades

#

Stable diffusion isn't geared for such things. It's a denoising algorithm at it's core. Any changes have to add noise to the image and then "denoise it towards a prompt" if that makes sense. So it's fundamentally a destructive process

hidden sparrow
copper crystal
#

You'd have to learn all that to get good results with stable diffusion anywyas. it'll only do what you guide it to

hidden sparrow
#

Well thatā€™s what Iā€™m asking. If anyone done this

#

I could train my own lora model or similar and create an entirely fictional image, but thatā€™s the thing. I just want to make the quality better. Not change everything

copper crystal
#

yeah. that's actual skilled work and is not an automatable thing really.

#

not yet anyways

#

you could look into sd1.5 and sdxl tile controlnet models. They keep the original image VERY consistent, but affecting the color grading or fine tunes like you want.. that's not easy because it's trying to maintain the original image

#

you could also look at IC-Light, whhich is a whole different model that can relight an image. it's impressive but it's just one type of control over the lighting and you'll stiill need to know color theory and stuff

fervent thunder
#

Hey so when doing latent noise mask to change stuff with inpainting any tips? Can I use it to add stuff thatā€™s not where rather than just tweaking already existing content?

copper crystal
#

yeah with higher denoise values. You can set it to either start with the same image you're inpainting, or start with pure noise and fill it in from there

fast sage
#

Lad still doesnā€™t understand how all SD projects que up prompts. šŸ˜‚šŸ˜‚šŸ˜‚

Iā€™ve written many nodes for many applications, but please keep telling me I donā€™t know.

Maybe start by understanding what vram and ram is before you start giving advice?

copper crystal
fast sage
#

Yes you can upscale. Nudist doesnā€™t know what heā€™s talking about.

fast sage
copper crystal
#

you're not explaining anything. oh well

#

weird flex is weird flex

fast sage
#

Itā€™s not my job to explain to you basic understanding of programming.

copper crystal
#

til an array of strings will over fill your vram

fast sage
#

Vram != ram

copper crystal
#

20 prompts. that'll fill your ram so hard.

fast sage
#

And technically itā€™s json not strings.

copper crystal
#

(tfw json is just a string)

#

you don't need to explain basics of datatypes or programming to me. I clearly have a better grasp on it than you realise, since.. well.. this whole conversation

fast sage
copper crystal
#

he's not using nodes. he's using webui

#

you've lost the plot son

fast sage
#

Are you the kid that I argued with a week ago who didnā€™t know what an array was?

fast sage
copper crystal
#

here's a fun fact. a string is just an array of chars

copper crystal
fast sage
#

Lmao, you sound like that kid, who didn't know what an array was. at the very most it's an array of characters

fast sage
fast sage
copper crystal
#

this is a really weird argument. No i'm not @dasilva3334181

fast sage
copper crystal
#

i dont think i'm needed here. you got yourself your own little strawman best friend to argue with. if he only had a brain you might have a useful conversation though

#

the guy isn't using comfy. he's using automatic1111 webui.

All i said was forge doesn't use comfy backend. i don't use comfyui since it has a large attack surface.

#

i dont' argue about it since i'm not versed in it

fast sage
#

You should prob understand this before you start claiming it has anything to do with the vram, as it doesn't

copper crystal
#

caching 20 prompts (in json format no less) will fill a megabyte , maybe a few megabytes, worth of memory

fast sage
copper crystal
#

youve lost the plot and are spiraling. take a breather. this isn't pretty

copper crystal
fast sage
fast sage
copper crystal
fast sage
#

Do you even know what gradio is?

#

Anyways it's been fun.

copper crystal
#

we're well past a socratic teaching method. what are you trying to say?

#

i've yet to find a point

fast sage
copper crystal
#

dont let the door hit you on the way out big gunner

fast sage
#

Hey how can I use SD to improve an image?
You can do that with photoshop or any other traditional editor. Those kind of edits are what graphic design has been all about for decades

šŸ˜„

copper crystal
#

That wasn't his entire question. nor was that my entire answer

#

šŸ˜†

fast sage
#

Hey how can I use SD to make a image better.
"use photoshop"

that's how usesless you are šŸ˜„

copper crystal
#

QQ

fast sage
#

Now we're in the then claims some other garbage. going to do another full circle?

copper crystal
fast sage
#

Please pray for my friend here, he's deficent.

copper crystal
#

are we to be discouraged from offering advice to people here because of this one guy's ego being bruised? I mean, i will if thats desired

fast sage
#

goes to SD form, gets told to use photoshop

#

It's sad that you don't even know what those algorithms in PS are doing...

shell tendon
#

it should just be up to the model i think?

#

the current version of clownsharksampler doesn't have the shift stuff in it anymore

copper crystal
#

heyoh clownshark. you're a dev that i respect. offer illumination? how much memory would json with 20 prompts in it require ?

tawny bronze
copper crystal
#

if you want to stay completely outta this topic, i fully understand

tawny bronze
#

thanks for the reply

cedar salmon
#

secret twist the prompts contain all known literary works

copper crystal
cedar salmon
#

this fits

copper crystal
#

yeh i concede. it would definately fill all system memory in that case

#

i mean seriously though. If i was wrong about making any photo looking professional without changing the core aspects of the photo, i'd love to be wrong about that. Lay it on me

fast sage
#

Again, we don't know his workflow, he could be using base64-encoded images and many of them, just because you think a prompt is just text doesn't mean it is.

#

Either way it's not vram

copper crystal
#

wtf is a base 64 encoded image? you mean like how you encode it into an html string and use javascript to display it?

webui isn't compatible with that. it just takes pixel data in it's img2img tab.

did you notice when he said he's using automatic's webui? it's a big clue.

green sand
#

Ok, are digital tools to get an image's tags/prompt off-topic?

copper crystal
#

images only have tags if the person published them in the meta data. any metadata reader will see that but there are lots of specific ai generation solutions out there.

another way is to use a vision model to describe the image, if there are no tags saved in the metadata.

i'm reluctant to give more specific advice. it might offend people

green sand
# copper crystal images only have tags if the person published them in the meta data. any metada...

Of course there are always limits to what can be described, for instance, an image with a man walking with his dog in a park will most likely consider the foreground with more details and the background as just a place the foreground is. If the dog is a chihuahua or a dachshund will end up having more of a difference than if one of the barely visible people in the background is wearing a backpack, and adding that description could even interfere with the original focus.
It may not exist but a way of satisfying every parties would be to add a minimum to maximum amount of tags for the image, because I believe an trained ai would naturally strafe to descript the most important parts of the images instead of the harder to view ones. With tags I mean in the sense of online image boards, they usually describe their content with tags and so do some AIs. Natural language, to me, is only a last resort when I cannot find tags to describe what with clearance.
LLMs like ChatGpt already can analyse images, but as every non generalised ai it simply doesn't do that specific task as well.
The utility of this would be for when you've got a reference image but the ways you can describe it are way too broad, and the correct terms would have never went over your head. Ex: for poses, bird's eye view, actions.

#

Sorry if I wrote too much, I think slowmode would have gotten me there

fast sage
fast sage
copper crystal
fast sage
#

literally starts an argument with someone, then starts to cry when they are clowned on.

almost everything you've said in the last hour has been either a bad take, or wrong except for this first time where you seem to understand that meta is embedded in the image

copper crystal
#

^ relevant

green sand
fast sage
green sand
#

Still, I'm talking about getting a description to something you can't describe

fast sage
#

"digital tools to get an images tags/ prompts" I would consider classification, unless I'm missing something in your question.

fast sage
green sand
fast sage
#

If you ask GPT to explain what colors are in a image, it will tell you.

#

If you ask it how many men walking their dogs it will do that too.

#

llama3.2-vision is a good option if you want to run local imager

green sand
fast sage
urban scarab
#

i have some questions boyz

green sand
# fast sage It's local \

I know on-web is wack but it's kind of my only option, so if you have experienced any alternative that's server hosted it'd help me

copper crystal
#

llama 3 8B sized models can run on most local GPUs

#

koboldcpp is a good app to use for them. supports vision models too iirc

fervent thunder
#

open router or docker plus runpod

copper crystal
#

runpod if you want to run the 40B sized versions yeh

#

running local is ideal. using a service means they get complete distribution rights to your work, due to how service works. they have to store, copy, transmit your data back to you. And instead of taking specific rights, they just take a broad kitchen sink license

fast sage
#

I donā€™t really know what youā€™re asking I assume itā€™s how to classify things in an image.

copper crystal
#

i'm unsure if llm vision models fall into the category of classification models. but i'm not about to discuss that in depth because.. well, lets just not

fervent thunder
#

cloud is the cheaper option, contrary to very popular belief

copper crystal
#

depends on electrical prices in your area

fervent thunder
#

yeah cos Clown was saying the price for him and its like half

#

the US has lower prices in general for that which is partly why cloud is less popular there

proud dirge
#

Hello everyone! There's a LoRA trainer on Civitai, would the result work with Flux Dev GGUF Q8? Or is it only for the original Flux?

fervent thunder
#

yes it will work with GGUF Q8 if you have enough VRAM

#

with GGUF if you hit the VRAM limit the lora can fall off

#

you can fix this by merging the lora

#

you might not need to though

proud dirge
#

Thank you for the replies! Do you happen to know how much VRAM GGUF Q8 normally takes?

fervent thunder
#

12.7 GB

#

the VRAM usage is simply the file size https://huggingface.co/city96/FLUX.1-dev-gguf/tree/main

proud dirge
#

Ah, sure, makes sense. And LoRAs seem to be pretty small, so hopefully 16 GB would be enough for one or two of them. Thank you!

fervent thunder
#

you can always merge them and then they take zero

#

good luck

proud dirge
#

Thank you very much!

graceful dagger
#

Hello

agile schooner
#

Hi

green sand
fervent thunder
#

that works yeah

stoic vortex
#

hello

dusky moss
thorn hare
#

what program do ppl use to train AI on specific voices ?

copper crystal
fervent thunder
#

Is stable diffusion not freemium anymore?

upper plinth
#

What do you guys think will be the best uncensored LLM model to run on the new 5090 with its 32GB VRAM?

cedar salmon
#

check back when the card is released

upper plinth
#

lolol its dropping early Jan

cedar salmon
#

in AI time that like years

upper plinth
#

rofl

cedar salmon
#

ide like to see a Lama 3.3 uncensored

signal solstice
#

Do you guys have any model recommendations for spaceships? Iā€™ve been using dream shaper and juggernaut and SDXL.

mortal delta
upper plinth
#

do scalpers use a webscraper to buy GPUs on launch day or how tf is it done?

#

I ask not because I want to scalp (I wish I had the money), but because I want to make sure I get one before the scalpers do

#

I'm not in the US so sadly waiting in line outside a store is outside of my reach

fervent thunder
modern pagoda
#

Most sellers on amazon set 1 max limit now.

#

Price tags gonna be hefty

quasi shell
#

hello

modern pagoda
#

šŸ‘‹

copper crystal
opal hedge
#

I can't wait to pay 2500 dollars for a 5090, or whatever it costs

#

I'll feel dumb but the speed ups will be massive

upper plinth
#

The problem is that amazon itself is a scalper

#

Ur delusional if u think amazon will sell it at MSRP

#

MSRP is gonna be $2000 max which is not that bad

#

but amz will hella scalp it. U need to buy from newegg best buy etc

copper crystal
#

canadian stores mark up GPUs way past MSRP

#

all of the physical stores do it

brave cradle
#

halo

upper plinth
#

MSRP price is only available on the first 10 seconds of launch

#

after that, its sold out and everything else is a resale

#

only US will do MSRP tho bc of import tariffs

fervent thunder
#

for image gen you don't need RTX 5090 that badly
svdq-int4-flux.1-dev.safetensors is only 6.64 GB of VRAM usage

#

and that's one of the fastest flux methods out there
only a good tensorrt setup might be faster

slender vault
#

3060 is plenty for flux

#

maybe not dev versions actually

#

schnells and hybrids yea

lavish iron
#

3060 will kind of work with the dev version

#

i've generated images with flux-1-dev with 3060

fervent thunder
#

I've rented 3060 for flux before its ok

slender vault
#

oh? how i always get bsod when i try to use a dev model

lavish iron
#

i would guess that a 3060 might have problems training flux dev though

slender vault
#

FluxGym says they support 12gb

#

to train

lavish iron
#

i got comfy to work with it

slender vault
#

ya me too i use comfy

lavish iron
#

i put the --lowvram command line argument

slender vault
#

ooooo

#

i should

#

okay

#

ty

lavish iron
#

i tried fluxgym and that just gives error 0

fervent thunder
#

I actually do the opposite and put --highvram
but then spam "unload model" nodes

slender vault
#

how much of a hit does it take to gen? on this 8 step anime hybrid it takes me 30 sec for 1024 res pics and 1 min for abive that

lavish iron
#

most of the time is spent loading the ~20gb file

#

once you've spent the day waiting for that to load, it's a little bit longer than an sdx renderl I'd say

slender vault
#

niceeee. good info to know ty

lavish iron
#

I was also loading the model from a HD, not an ssd, so 10 time slower

slender vault
#

ah ya i use nvme ssd

#

shouldnt take me that long then lol

lavish iron
#

i finally bit the bullet and copied my models folder over to nvme the other day

#

it's much better

slender vault
#

how god how long did that take

lavish iron
#

180gb - not that long when you do something else

fervent thunder
#

on 3060 I tend to use distilled versions of SD 1.5 like SD 1.5 hyper rather than anything bigger

#

some Schnell stuff at 768x768 like Shuttle 3.1 for 2 steps is okay as well

lavish iron
#

if I'd spent the time watching it copy 180gb, it would have taken forever!!!

slender vault
#

why? 1.5 is ridiculously easy to run

#

for 3060

lavish iron
#

sd 1.5?

#

you mean flux 1.5, i'm still noob

fervent thunder
#

sd 1.5 is stable diffusion 1.5

slender vault
#

yeah no need to run that, just use xl models

#

3060 is good for it

#

illustrious models

#

for anime

lavish iron
#

i gave up running sd1.5, i had absolutely no luck with it at all, everything has either 3 legs, 4 arms or 12 fingers

fervent thunder
#

I personally like quite fast generations but people have different tastes on that

slender vault
#

thats cos its super early model, you needed alot of prompting know-how and loras

#

and negatives

#

can still make nice stuff with it but

#

impractical

#

if have the vram

prisma horizon
#

honestly, i prompt better on 1.5 than xl

slender vault
#

one of my merges at the time

lavish iron
#

sdxl seems far more reliable

#

i've not managed to train an sdxl lora yet though

prisma horizon
#

i really attempted to prompt on xl, but i simply can't feel it

fervent thunder
#

it depends if you use Ella or not
Ella SD 1.5 prompts well

slender vault
#

i did last night when sleeping, took like 3hrs with only 10images

#

so if you not planning on using pc lol

#

i can send you a ss of settings if needed exard

fervent thunder
#

I like IP adapter a lot

lavish iron
#

i had reasonable luck using florence 2 for sdxl prompting, I tried it for tagging train images, then take the prompt and put it into generate

fervent thunder
#

not sure if SD 1.5 or SDXL has better IP adapter

slender vault
#

also for flux

fervent thunder
#

florence 2 is awesome yeah I use florence 2 for every model

prisma horizon
#

what is Ella?

fervent thunder
#

Ella is this thing https://github.com/TencentQQGYLab/ComfyUI-ELLA?tab=readme-ov-file
it lets you use T5 with SD 1.5

#

it makes it prompt better

lavish iron
#

i had a weird experience with the resolutions, if i asked it to make an arbitrary picture of resolution not on that list, it makes a mutant

slender vault
#

out of all the uis i always tend to gravitate back to comfy

prisma horizon
#

oh, that's actually a good thing, i should try it

fervent thunder
#

maybe invoke as well as comfy

#

not sure about others

lavish iron
#

i had no luck with auto11, but tried it recently when i had more knowledge and got it to work, but now I have the issue, it randomly does not load lora directories

slender vault
#

invoke good? havent messed with it

fervent thunder
#

yeah invoke is under-rated probably

slender vault
#

i got comfy within krita running last night

prisma horizon
slender vault
#

super neat

lavish iron
#

i gravitate back to comfy

fervent thunder
#

invoke is like comfy but more stable and with canvas
downside is less nodes

prisma horizon
#

i actually gen stuff like that native on 1.5

#

that was a little messed up

lavish iron
#

Anyone use OneTrainer, I don't know where the trigger/activation tag is set for a lora

prisma horizon
#

i hit the step in the middle

slender vault
#

ye

#

uh sec i need to look myself

lavish iron
#

i just put my gpu on train for the last 8 hours overnight and got complete garbage

slender vault
#

usually to add a trigger you do it while tagging

#

your dataset

#

least thats the way i do it

lavish iron
#

i put the trigger as the first word in the tagging files

slender vault
#

how many imgs/repeats

#

and epochs

lavish iron
#

it was 18 epochs for 110 images

slender vault
#

repeats?

lavish iron
#

i don't know where that is set, so couldn't tell you

slender vault
#

so prolly 1 then

#

in concepts

lavish iron
#

i have 1.1 days of experience using OT

slender vault
#

you can open up ur concept

#

click it and youll see repeats on the rightish

#

(moving stuff from on drive to another so i cant open onetrainer to send ss rn)

#

usually i do like 2-4 repeats

#

and play with the learning rate

#

.003 is default i think? i move it to .001

lavish iron
#

I'm guessing this is it?

#

Hmmm, can't paste an image

slender vault
#

dm me

#

or paste in gen w image

#

right below

lavish iron
#

Balance

#

for regularization, i set the balance to (No. reg images) / (training images)

#

there doesn't seem to be an option to explicitly set them as reg images like in kohya

slender vault
#

oh shit dev isnt crashing and working now

#

and only 15sec longer

#

lma9

#

fantastic

#

literally 11999mb is being used lmao

narrow perch
#

Hi

lavish iron
#

I was dubious when upgrading my pc, should I go for 16 or 32gb or ram, thinking, I've almost never seen my pc use most of 16gb, but erred on the side of caution and went for 32gb. training SD has bought a new meaning to this when i see task manager performance and 15.99/16gb vram and 31/32gb ram being used constantly!

lavish iron
#

Why do they not just put the abstract at the top of the github page

slender vault
#

yeah 64 of ram isnt enough

#

for me

#

lolol

lavish iron
#

I'm guessing it uses the page file for the rest of the ram being used!

fervent thunder
#

you need 64gb

#

for some reason

#

can't get servers with 32gb to work

lavish iron
#

that should be what the page file is for, increasing your ram when it goes beyond the physical limit

fervent thunder
#

oh yeah that still works

#

but its too slow

lavish iron
#

If I use 0 reg images, the ram consumption and speed of processing increases massively

#

i'm not sure what effect having 0 reg images has

lavish iron
#

What might be a good item for gpu manufacturers would be a PCIE dedicated vram board, no processor, just VRAM

lavish iron
#

From my experience, changing the learning rate is useful for training convolution networks, can this be implemented with stable diffusion, ie.
If I train a model with a high learning rate for the first 10 epochs, then incrementally lower the learning rate every 10 epochs.
Is this a thing?

glass grotto
#

Should I consider using Comfy ui or forge?

#

I'm looking for speed. Also I have a 6gb ram gpu and I'm using flux v1

lavish iron
#

I remember someone saying that Forge automatically detects and switches to low vram, while comfy uses a command line argument in the batch file --lowvram

glass grotto
#

but i think that comfy is faster but the issue is that its hard to use control net with comfy

#

it's not user-friendly

lavish iron
#

at the end of the day, it's the amount of tensors/s that your gpu/cpu can process that determines the speed, not the application, as both applications use the same underlying code, just have a different ui

#

I've tried a number of different ui's and the speed it renders from a given model is very consistent across different ui's.

#

it's true that it's a bit more complex to make the control net work in comfy, but you do have a lot of flexibility with how you wire it all up

#

most ui's should offer a button or option to allow you to load the model into memory before you start the generation queue, this can help.

ashen sleet
#

wsh chat

serene lark
#

Hello guys, where can I download stable diffusion

snow perch
#

Hi guys

#

I want to use SD to draw color block textures like early cartoons. Is there any suitable model or lora? I only have 8GB Vram, is it not enough?

#

I can't paste pictures here, I mean cartoons like Tom and Jerry

slender vault
lavish iron
#

damn, wish I'd had that advice when I decided to learn SD

slender vault
#

I went from AMD linux to getting nvidia cos windows

lavish iron
#

that's quite a shift

copper crystal
#

i prefer windows but amd windows drivers are so bad that i took up proton gaming on my desktop for a couple years because of it

#

it worked out well though because i was genning on my amd card the year that SD came out. automatic1111 had it all set up to work with rocm on arch

#

also games were getting more frames

earnest stratus
#

Hi everyone!

lavish iron
#

Is it worth adding captions to my regularization images when training a lora, ie run them through WD14 etc, other than the class word?

slender vault
#

1dont use wd1.4 that is super outdated imo

lavish iron
#

I'm guessing that can't be added to OneTrain?

#

ignore that, i just looked, it's a complete ui

copper crystal
#

wd1.4 is just an sd2.1 model iirc. it was bad when it first came out too. The gold standard of anime when it released was novel ai's model and the refines based on that. shortly after wd came out, sdxl dropped and the few people using 2.1 moved on

#

you dont want to bother with refining 2.1. it's not very refineable. doable but something about it's architecture made it a lot more difficult to teach anything

lavish iron
#

you've convinced me completely to not use WD14 again!

#

back to the question though, is it worth properly captioning regularization images?

lavish iron
#

Just had a gander at my page file, it's only 30gb...

static cape
#

Hey guys,
What's a good Discord for chat about Local Video AI Models like Mochi, LTX & HunYuanVideo?

quartz siren
static cape
quartz siren
slender vault
ashen tundra
#

the pink plead is gone :C

steel fable
#

How to create photos for 16:9 full screen monitors without cut

slender vault
#

with stable diffusion

steel fable
#

Yeah but it is some extension?

tacit narwhal
lavish iron
#

what is a good model to use for tag generation in Taggui?

slender vault
#

florence

#

large

#

adjust the max tokens as you want

lavish iron
#

Thanks

#

tried to do 100 epochs overnight with 3 repeats. I managed to do 17 epochs! The computer decided it wanted to sleep and stopped processing.

#

i'm guessing that I shouldn't minimise the window

slender vault
#

power settings

#

turn off dim screen and sleep mode

humble iris
#

yo. Just rented a 4090 for the first time, I need to make some images illustrating songs for a small concert. Not very experienced with Comfy and never used Flux before. My Comfy is in fp16 currently. What exact Flux model do you recommend and maybe some of your favourite loras?

fervent thunder
#

if you can, try to get SVDQuant nodes working, and use their flux model https://github.com/mit-han-lab/nunchaku/blob/main/comfyui/README.md

#

its by far the best choice for 24GB VRAM and below

icy path
#

Hey how are you guys. I have a very tiny problem which is extremely big for me. I need Stable Diffusion XL Inpainting's finetuning script, but even after months of work, I can't find a working script.

I saw SD v1.5 and SD v2 Inpainting's script which makes random masks and I think that is awesome, but even extensive works on the script trying to convert it for the XL model came up dry. I need help, if someone could?

#

is this the right place to ask, or would the tech support channel be more suitable?

maiden orchid
#

Hell!! šŸ™‚

Iā€™m really excited to join this community! Iā€™m passionate about everything related to AI and always eager to learn more. I use AI to help with projects like book writing and content creation, and Iā€™m looking forward to connecting with others who share the same interests.

lavish iron
#

sounds like a scam to me

#

if it's not a scam, you shouldn't be asking to loan people's accounts, that's called phishing.

median jewel
#

How do I make an image prompt in this code? I want to be able to load a picture and create a picture that is inspired by the input picture

from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler
import torch
import os
import time

# Image storage folder
IMAGE_FOLDER = './images'

def create_pic(prompt):
    # Load the model
    pipe = AutoPipelineForText2Image.from_pretrained('lykon/dreamshaper-xl-lightning', torch_dtype=torch.float16, variant="fp16")
    pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

    # Optimizations for vram usage
    pipe.enable_model_cpu_offload()
    pipe.enable_vae_tiling()
    pipe.enable_vae_slicing()

    # Inference (should take a few seconds)
    generator = torch.manual_seed(0)
    image = pipe(prompt, height=768, width=768, num_inference_steps=8, guidance_scale=2).images[0]


    current_time = time.strftime('%H%M%S')
    filename = f"{current_time}.png"
    path_to_save = os.path.join(IMAGE_FOLDER, filename)
    os.makedirs(IMAGE_FOLDER, exist_ok=True)
    image.save(path_to_save)
    print("heer")

    # Return the relative URL path
    return f"D:/setup/Gamla bilder/website/{filename}"

print("saved to ", create_pic("man jumping from sky"))
modern pagoda
median jewel
#

yeah but what I am looking for is being able to load a picture so img2img. So when I write in the textprompt it will use the loaded picture as guidance @modern pagoda

modern pagoda
#

This looks like text2image? Afai can see?

#

You want image2image

median jewel
#

yes so lets say i have a mug i want to use that in the loaded picture, and then in text prompt i say dog holding mug, and it will hold the mug i loaded, ykwim?

modern pagoda
#

It should in theory depending on model

median jewel
#

yes but i am pretty sure SDXL supports that

boreal cosmos
#

lmk if us till looking for someone hosting open source 3d gens

copper crystal
#

latent diffusion coming to LLMs. so it'll generate entire passages of text at once instead of only the next token at a time. META published a paper on it and the next generation of models will likely be developed this way

jaunty osprey
fervent thunder
copper crystal
#

Yeah it's a generational leap (heh)

fervent thunder
#

there were some funny diffusion language models before they were not too strong but used for some stuff like artificial data creation

copper crystal
#

yeah it's not a new direction. it's been tried before. but Meta's new model seems to get it done

fervent thunder
#

Meta get the benefit of the doubt because scale yeahg

copper crystal
fervent thunder
#

ah nice they gave code that's cool

copper crystal
#

as far as i can tell the license is permissive too

junior solar
#

very good

steel acorn
fervent thunder
#

šŸ‘€

tranquil sedge
#

hi

fresh ruin
#

is it possible to reverse search an image on civitai?

slender vault
#

why

#

can use googles

fresh ruin
#

true

#

just wated to find similar images so i can maybe find the prompts for it

slender vault
#

try png info first

#

load an ai generated image and if it has metadata the png info will pull the prompt

fresh ruin
#

ill try rn

#

i put the image but nothing shows

abstract quarry
fervent thunder
#

ah okay I haven't read the paper yet I assumed it was one of those diffusion language model things

fervent thunder
timid spade
#

anyone know of a good stablediffusion finetune for pixel art that i can run on 12gb of ram? I'm trying to train a controlnet on a specific pixel art task and i need to make a good choice of base model

fervent thunder
#

12GB VRAM and you can run flux

timid spade
#

can i train controlnets for flux? that's what I've been using for straight up image gen. was trying to use a lora for my task but it doesn't really work for what I'm trying to do

craggy heath
#

I'm using omnigen on playground (https://omnigenai.org/playground). I launched a generation but it indicates "queue: 36/36" + a timer which increases over and over, slash 6349475.8s. There is also an loading icon which plays over and over. 6349475.8s, it's more of 73 days. This is just crazy. What's wrong here !??

copper crystal
#

i welcome being wrong

abstract quarry
#

yeah, it's a very different technique

#

but a cool paper, although I have the feeling it's not a theoretical breakthrough but rather an engineering one. Their main recipe is the n-gram encoding, not their fancy entropy based patching

glass grotto
#

can flux understand natural language ?

#

like full sentences

copper crystal
#

it uses the t5 encoder so yes

wicked root
#

can automatic1111 run flux models or i can't yet?

warm junco
wicked root
#

thanks for the info

flint nest
#

hi

slender vault
warm junco
#

and gguf models work there too

slender vault
#

interesting

orchid terrace
#

Hi guys; where can I go to get a refund for my subscription? I can't get through to support

copper crystal
wicked root
#

i might try it out

copper crystal
#

old extensions won't work as well. there are some compromises with it

wicked root
#

i don't really use extensions, the only one i use is tiled vae

copper crystal
#

i have a thing called "Stability Matrix" that allows me to manage and update a few different UI's, all of them with access to the same models and lora folders

wicked root
copper crystal
#

The matrix app manages it all pretty well for you

#

you could otherwise do it yourself with symlinks, which isn't so hard

#

looks like forge ui has multidiffusion built in. Which is tiled vae afaik

#

not sure if that works on forge though

wicked root
#

forge ui seem better than automatic1111 so i think i'll switch soon

#

thanks for all of the informations

copper crystal
#

"better" in some ways.. but others . .. i often find it being a complete memory hog

wicked root
#

then i might not switch if it take that much memory

#

i don't know if 12gb of vram would be fine

#

though i could try comfyUi

#

i heard that it's pretty good

copper crystal
#

12gb is tight. comfyui might allow more flexibility there, but its more advanced noodling

copper crystal
#

https://arxiv.org/abs/2406.02507 what is this ? negative guidance? "throw some poop at it. that'll make it better" lol. more involved that that surely, but summing it up that way seems hilarious.

slender vault
#

<12g 3060 user

#

Generation is quick, training however takes the night

#

if you use comfy/swarm youll have better speeds too imo

cyan sinew
#

is there some damn tutorial to launch the software? i look at stability.ai public releases, it has zero search results for "download". i download some random model, it asks me to download user interface for stble diffusion, i do that, there is not a single .exe file to launch the software....

slender vault
#

check out stability matrix, i use it to be able to use multiple uis while sharing the models across uis

#

easy to use

cyan sinew
slender vault
#

go to tech support

#

cs1o is pretty knowledgable when it comes to errors

cyan sinew
#

it has no errors, it simply closes itself.

cyan sinew
slender vault
#

read what hes saying, you arent launching the webui via the .bat

cyan sinew
#

i did. it had no exe files, so i ran the bat file, it simply flashes the explorer as if cmd window was opened but closed too quick to show anything

wicked root
#

what bat file are you launching?

cyan sinew
#

webui.bat

wicked root
#

if i'm not wrong, the right .bat file to open should be : webui-user.bat

#

for me, it's the file that i need to open

#

with automatic1111

cyan sinew
#

it behaves the same

wicked root
#

i can't really help then

#

sorry

#

but i can still try to find why

cyan sinew
# 
@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS=

call webui.bat

``` thats the webui-user.bat
wicked root
#

you have a nvidia gpu or an amd gpu? (or intel, if it's intel, i can't hel)

cyan sinew
#

nvidia i think

wicked root
#

alright. it says for nvidia to

Edit the webui-user.bat (right click, edit), At the line COMMANDLINE_ARGS= You add: --xformers --no-half-vae
Add the following command there too depending on your GPU Vram. (Check Task Manager ->Performance ->GPU ->Dedicated GPU-Memory)

#

i use amd so my webui-user.bat is

#

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --use-zluda --update-check --skip-ort --medvram-sdxl
set HIP_VISIBLE_DEVICES=1
call webui.bat

cyan sinew
#

couldnt find python error

wicked root
#

so after the place where it's written : set COMMANDLINE_ARGS=
you should add --xformers --no-half-vae so it should be :
set COMMANDLINE_ARGS= --xformers --no-half-vae

for you

#

(wait this is the wrong chat)

cyan sinew
#

blender runs fine with python, so idk.

wicked root
#

so if you use Nvidia, do you have CUDA of downloaded on your pc?

#

it's for nvidia graphic card

#

we should switch to tech-support chat because we are talking in the wrong chat right now

#

(sorry for flooding general)

copper crystal
# wicked root so if you use Nvidia, do you have CUDA of downloaded on your pc?

not needed afaik. in the past it helped because you'd have the newest versions of the cuda, but drivers and pytorch have since come along since.

Consider that you woudlnt' need to install the entire directx sdk to play dx. you only need the dll that works with the binary compiled for it. The driver has cuda built in for that end user purpose

fresh ruin
#

does illustrious work with forge?

#

im just getting noise

slender vault
#

yes

edgy notch
#

hiiiiii

#

i'm new here

slender vault
#

welcome

raw dove
#

hi happy to be here

restive moon
#

Heyy

acoustic moth
#

hi, very happy to be here with you!

wet tide
#

Hello. I am new with SD. How artists get so same looking ig models for example in ig pics with ai. What settings and tools they use. How same ai model with 100% same look you can create to be as model etc in internet etc.

tranquil thistle
#

hello! I just start to use SD and happy to be here

abstract quarry
cyan sinew
#

i needed an image generator which generates map layouts. the one i found used stable diffusion, so i had to go with it

abstract quarry
#

I use ComfyUI myself, but I would never recommend that for beginners. The same with Forge or Auto111. Yes, they are more intuitive than Comfy, but not by much. InvokeAI might have less plugins and need always a few months more to come with newest features, but it installs automatically, has an intuitive user interface and lots of tutorials

abstract quarry
cyan sinew
#

well, it said to download stable diffusion so idk.

abstract quarry
#

its just a model. You could run it in python if you are a programmer, but I assume you are not

#

you need an application that runs stable diffusion for you

#

there are plenty to choose from

#
  • comfyui
  • swarmui
  • auto111
  • forge
  • invokeai
  • ....
#

invokeai is in my opinion the easiest to use

#

when you install it, it will ask you which model you want

#

there you can select stable diffusion

#

btw. if you want to make battle map layouts, I don't think stable diffusion or image generation in general is the right tool for it... I tried several stable diffusion models for battle maps. I also made a model myself. Yes, they work somewhat, but they are not perfect. I would rather use one of the many tools you find on e.g. steam for it (like dungeon alchemist)

slender vault
#

should probably do abit more research before jumping into something new

abstract quarry
#

asking in a discord channel is "research"

lucid bobcat
#

Am I supposed to have the new hook nodes in comfyui? They are beta, but also they have been released 2 weeks ago. I have a stable build, so are they not part of it?

fervent thunder
#

probably need update

lucid bobcat
fervent thunder
#

I have those nodes but I don't have a stable build I just download the newest of everything as soon as it releases

glass grotto
#

Is there a guide explaining the diffrent prompts parameters, cuz sometimes is see things like this <0.9> or this (Prompt)

lucid bobcat
glass grotto
#

k, I'm using forge

lucid bobcat
#

Forge should have the same syntax as WebUI

glass grotto
#

I'll try to look for it

lucid bobcat
# glass grotto I'll try to look for it

Short answer is <> is used to add loras and () to increase the weight of a part of the prompt. But there's more cool stuff you can do so definitely read the documentation or watch a tutorial.

glass grotto
#

Is the t5xxl_fp8 encoder good ?

raw oasis
#

Hey guys!
Beginner here.
I've managed to get the hang of image generation and modifying those images by inpainting by following some tutorials, but I'd like to alter an image of my pet. And for some reason I struggle to get that right. Either it barely changes anything, it screws up the shapes, or... it turns it into something entirely different...
I was wondering, do you guys know of any (video) guide to do this kind of stuff? For example, I'd like to create something like this:
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/eae3f1f8-c95e-49cb-b502-e5b88fd76082/original=true,quality=90/00102-3865871024.jpeg
or this:
https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/562f5b5e-e84e-4c2c-862a-67915014e736/original=true,quality=90/41254735.jpeg
Thanks!

#

(I'm using AUTOMATIC111 if that matters)

warm junco
#

That should help

#

And set the denois to 0.5 at start, then adjust

raw oasis
#

I'm already doing those things. Very little effect unfortunately šŸ˜¦

warm junco
raw oasis
#

Sure, give me a couple of minutes. Thanks already! šŸ™‚

warm junco
grim mulch
#

this is very fun

young hazel
#

whats the best model?

#

i finally got everything working lol Im trying a few different ones

quartz siren
# young hazel whats the best model?

There is no "best" model, completely depends on what you want.
flux dev has the best prompt following, humans, and text rendering, however sd3.5 large is a bit worse in the above capabilities but is more creative, knows a bit more knowledge, and knows more art styles.

Pixelwavev3(finetune of flux dev) is great too, has flux dev level capabilities and is more artistic and has more knowledge.

If you want fast models, flux schnell is the best at 1step generation and shuttle3(finetune of flux schnell) is the best at 2-8step generation. Sd3.5 large turbo is good quality at 4-8step but shuttle3 is similar quality at 2step and can do much higher res like 2k.

young hazel
#

Buut yeah more depneant on what ur looking for

quartz siren
fair forge
#

hey any good toturial for installing stable 3D?

glass grotto
#

Does anyone knows a good upscaler for flux generated images ?

cedar salmon
#

Comfyui_TTP_Toolset

glass grotto
#

i'm on forge ui

cedar salmon
#

(ā•ÆĀ°ā–”Ā°)ā•Æļøµ ā”»ā”ā”»

blissful aspen
#

Not sure if I am dumb, but I am using WebUI reForge and I can't see the ControlNet tabs. What am I doing wrong?

blissful aspen
#

"sd_forge_controlnet" built-in extension is checked under the extension tab

#

and no errors in console
nvm, I have an error in console:

Path E:\SD\extensions\sd-webui-controlnet\annotator\downloads does not exist. Skip setting --controlnet-preprocessor-models-dir

and then

*** Error loading script: controlnet.py
Traceback (most recent call last):
File "E:\Data\Packages\reforge\modules\scripts.py", line 533, in load_scripts
script_module = script_loading.load_module(scriptfile.path)
File "E:\Data\Packages\reforge\modules\script_loading.py", line 13, in load_module
module_spec.loader.exec_module(module)
File "<frozen importlib._bootstrap_external>", line 883, in exec_module
File "<frozen importlib._bootstrap>", line 241, in _call_with_frames_removed
File "E:\Data\Packages\reforge\extensions-builtin\sd_forge_controlnet\scripts\controlnet.py", line 24, in <module>
from lib_controlnet.controlnet_ui.controlnet_ui_group import ControlNetUiGroup
File "E:\Data\Packages\reforge\extensions-builtin\sd_forge_controlnet\lib_controlnet\controlnet_ui\controlnet_ui_group.py", line 16, in <module>
from lib_controlnet.controlnet_ui.openpose_editor import OpenposeEditor
File "E:\Data\Packages\reforge\extensions-builtin\sd_forge_controlnet\lib_controlnet\controlnet_ui\openpose_editor.py", line 6, in <module>
from annotator.openpose import decode_json_as_poses, draw_poses
ModuleNotFoundError: No module named 'annotator'

modest smelt
#

Hi. Iā€™m new and i use stable diffusion (text2image) but itā€™s not always really realistic, can i do something to upgrade ?

warm junco
warm junco
modest smelt
#

You talking about, [Sampling method] ?

warm junco
#

Which one do you use?

modest smelt
#

V1-5-pruned-emaonly.safetensors
I just have this one

warm junco
#

You get better checkpoints (models) at Civitai.com
Then you drop the model files into the models/stable-diffusion folder.
Make sure the model file is 2gb or larger

modest smelt
#

I will try that

modest smelt
#

Itā€™s not free ?

warm junco
modest smelt
#

"Get access to this Model Version!

The creator of this Checkpoint has set this version to early access, You can download with this Checkpoint by purchasing it during the early access period or just waiting until it becomes public. The remaining time for early access is 3 days, 11 hours, and 10 minutes"

#

i need to find another one

#

3 days

warm junco
modest smelt
#

oh you right excuse me, i am a little nooby

warm junco
#

No problem

#

It can be confusing at first for sure

slender vault
#

god i hate they added early access

#

i remember crying the first few days of learning sd

modest smelt
#

@warm junco that's work, thanks šŸ‘

modest smelt
#

Can I swap the face? Like, put a face on the image I want?

warm junco
modest smelt
#

Should I go through an extension and paste the URL, or is there something better and easy ?

glass grotto
#

Any Upscaler recommendations?

modest smelt
#

i have face manipulations it's okay ?

#

(available)

glass grotto
warm junco
warm junco
#

Then search for Reactor

modest smelt
#

Itā€™s work

#

And I check ReActor
I drop the face, and i generate ?

warm junco
modest smelt
#

thanks

#

This is so interesting, I think Iā€™m going to get a lot more questions šŸ˜‚

desert dagger
wraith vine
#

best course to learn stable diffusion w/ a1111?

#

I like youtube tutorials but want series all put in one place

#

so thinking like online courses

slender vault
#

@karmic brook can you handle this spammer pls

slender vault
soft breach
#

hello guys I think I can ask here but I need help making a decision for a new PC. I can either go with a 4090 laptop or a 4070 TI Super desktop. Between the two which GPU would you go with for AI? Both are 16GB Vram

desert dagger
#

not a cpu with an integrated gpu

craggy trout
slender vault
#

desktop always ideal for work takss

#

laptops are nice for gaming

elder yoke
slender vault
#

optionally subscribe to me, might get image gens, maybe

#

for legal purposes this is a joke

lucid bobcat
elder yoke
lucid bobcat
hollow marlin
#

Hello!
-# this user is under scrutiny by the FBI, do not contact this person as he/she is a direct suspect for a case. for more info, check discords trust and safety policies.

winter anchor
#

Can someone make a Lora for me

eager shard
#

b580 looks goated

modest smelt
#

animation is possible ?

glass grotto
#

Does anyone know a good upscaler ?

pastel lynx
#

Anyone did install successfully Trellis thing for image to 3D on Comfy ?

modest smelt
warm junco
fervent thunder
#

the guy mentioning stocks is one of the scam bots yeah
need to be careful cos what these discord bots have been doing is sending hyperlinks to malware over DM

#

they can be persuasive now because they can use Claude etc to do a bit of "conversation"

slender vault
#

wonder if can expose the ai with some logical bs lol

modest smelt
fervent thunder
#

need to be careful

modest smelt
#

What link are you talking about?

fervent thunder
#

oh they send the link in DM after a bit of conversation

slender vault
#

do the ol "say potato" trick over an over

fervent thunder
slender vault
#

you keep asking them to say the word potato, usually works to weed out the fake tinder bots cos tll be obvious its a bot, theyll either ignore it repeatadly or be strange

modest smelt
#

what did u think about only fan IA ?

fervent thunder
#

Hey so I'm using sdxl with comfyui and trying to get the impact pack, it says install failed and I think it's because I need the requirements seg anything etc but I'm using a laptop which doesn't have WiFi is there anyway to manually get the dependencies maybe please?

#

oh I see
this test can be passed by LLMs now though

fervent thunder
#

So annoying lol everything looks great apart from faces

slender vault
#

use ADetailer

blissful aspen
#

Does anyone uses X/Y/Z Plot Script with Reactor Face Models?
I have the issue, that when running with the script, the generated images with later face models are looking completely different, than when I use them seperately.
In ReActor I have the setting "Save Original" and "Swap in source image" ticked on, because in some images I got better results, when the swap is in the source image, instead of in the generated image. But I think what is happening is, it takes the previous generated image as a source, instead of using the very first image as a source. So the more swaps happening, the worse it gets.

lucid bobcat
fervent thunder
#

these days they use LLM

abstract quarry
#

I think he has a good point. llms might be vulnerable to this kind if test

lucid bobcat
# fervent thunder these days they use LLM

Yea but they are instructed to play a role, right? I don't think they can distinguish between reasonable questions and unreasonable questions. Also LLM can't do math, ChatGPT used to fail at basic additions. But now the prompt is preprocessed and math instructions are executed separately.

abstract quarry
#

but let's be honest: just thinking about "is he a bot?" is already the most important step. In the end it doesn't matter if somebody is real or a bot. A real person can send you a scam link, too

fervent thunder
#

doing a Turing test on the scam bot is actually not necessary yeah

#

cos you can just avoid the humans as well

lucid bobcat
abstract quarry
#

yeah, but even there real humans might be more dangerous than bots šŸ˜…

#

but yeah, it helps to not waste time

#

I imagine it's extremely frustrating chatting with somebody for long time just to find out he is fake

#

but again thant happens with real humans, too

lucid bobcat
#

Well if it takes you a long time the joke's on you

abstract quarry
#

fair enough

fervent thunder
#

there's that thing where if a test gets popular it ends up in the training data

abstract quarry
#

that's why you don't do a specific test but test general behavior

#

llms can be very naive and dumb

lucid bobcat
#

But for me being able to detect AI is important when researching on the internet. I don't even wanna waste 10 seconds on websites where the information is AI generated.

abstract quarry
#

it's easier to trick them than other way around

#

that's difficult, though, when you cannot interact with the ai

#

text generated by ai just looks similar to human generated text :/

lucid bobcat
#

Not really. I'm now fairly good at it. It's the way the information is structed and the talking points. I only have to scroll through and can tell immediately if it's AI or not.

#

AI often uses alot of words to not say anything of importance.

fervent thunder
#

I've read blogs sometimes where I only realised near the end it was written by LLM

lucid bobcat
#

I sometimes read blog posts with instructions to do something. If it's AI you always see the usual bullet points with super weird topics where you go "what does that have to do with anything?".

modest smelt
#

amd user, i want edit files "webui-user" but after he said python not find etc

#

can you help me ?

subtle tiger
#

I can't tell if it is intentional to cause more wordcount and ads, or if it's just llm runaway

fervent thunder
#

BBC articles did that before AI lol

oblique ledge
#

Hello, I was trying to learn how to make my own Lora, what would be the best way?

steel burrow
#

hello guys

desert dagger
steel burrow
#

It's been a long time since I've been here. I'd like to know what model people are using these days! Is it called Flux.1-Dev?! What are the requirements for using it or if there is another one, what would it be?

lucid bobcat
# subtle tiger I hate the recent news article trend where the article slowly loops itself

It's LLM. It's all fake news anyway. I'm not in that business but from what I understand many such sites use freelancers that would write articles and get a portion of the ad revenue. There is no real editorial process. So it's most likely people that would feed an LLM with info and tell it to write an article. Might even be fully automated at this point where a script scans the internet for hot topics, gathers the information, makes an LLM write an article and then publishes it on multiple news outlets. Pro tip: Do not consume news from main stream media. Find some independent commentators that you can trust and let them gather and digest all the information out there.

abstract quarry
#

which "mainstream media" is using llms for writing it's articles lol

#

mainstream media are the ones that actually still do something like journalism

lucid bobcat
# abstract quarry mainstream media are the ones that actually still do something like journalism

I don't know what they're all called, or if they even have names. But there's these news feeds on various websites like your e-mail provider for example. Or any odd website that has these newsreels. I call this mainstream because it's what most people are exposed to. I'm not saying it's all LLM because I don't read this stuff. I'm just saying if an article is weirdly structured, like repeating itself for exmpale, it's most likely LLM. There's probably tools out there that take text as input and tell you the likelyhood of it being AI generated.

lean terrace
#

any word on open-source/open-weights alts to suno and udio? all i could find was riffusion and they seem pretty dead or inactive. in my understanding, at least partly, audio generation is somewhat in a similar domain to image generation and there have been a multum of new models since like sd1.5 yet no word on anything in regards to audio, apart from stable audio, which can't really do music all that well.

desert dagger
slender vault
# steel burrow It's been a long time since I've been here. I'd like to know what model people a...

an nvidia gpu with 16gb vram or more is recommended, you can get away with lower up to 6 but at heavy performance cost, personall i use a 3060 12gb and its been keeping me fine, training loras takes anywhere between 1 to 3 hours depending on settings, with flux dev you could get away with genning with the 3060 12gb on --med-vram command line arg to help but i personally use distelled versions of flux, runs satisfactory.

I recommend Stability matrix for an easy way to install and manage UIs and some of the pins in the server from CS1o

#

personally i use illustrious XL models for day to day stuff, lora training and whatnot, then flux anime distilled/newreality flux models for memes,realistic,more detailed stuff

#

I have a tool on my civit aimed towards beginners in lora training to help them get the right balancing args, if you're interested (or anyone else for that matter) you can click my profile here on discord and follow the civit ai link to browse my tools/models

gritty dust
#

i'm also back after a long time. i'm reading that forge is the best webui to use and it looks like it has a one click installer as well, i'm trying that. my past experiences have always been riddled with technical difficulties to trying to get stable diffusion to work

slender vault
#

Could also try Fooocus, Swarmui or invokeai

#

all good webuis

gritty dust
#

thanks, i'll make a note.

modest smelt
#

hello, live portrait work good ?

i have this error code
RuntimeError: No ffmpeg exe could be found. Install ffmpeg on your system, or set the IMAGEIO_FFMPEG_EXE environment variable.

slender vault
#

install ffmpeg

#

like it says

#

lol

modest smelt
#

yes but how ? šŸ„² (i'm nooby)

slender vault
#

np, windows?

#

or linux

modest smelt
#

macos šŸ˜­

slender vault
#

gonna guess windows cos nooby

#

oh ok

#

sec

#

youll proll ywant the second one

modest smelt
#

i just need to lunch FFMPEG files ?

slender vault
#

looks like it just that one file yeah. extract it to a folder outside of the desktop first

#

then run it

modest smelt
#

finish

slender vault
#

retry doing whatever you were doing that caused the error see if still happen

#

may have to reboot the device idk i dont use mac

modest smelt
#

euh ...

#

RuntimeError: No ffmpeg exe could be found. Install ffmpeg on your system, or set the IMAGEIO_FFMPEG_EXE environment variable.

#

I have to put it in a special location ?

#

File "/Users/buell/stable-diffusion-webui/extensions/sd-webui-live-portrait/liveportrait/gradio_pipeline.py", line 246, in execute_video
raise gr.Error("Please upload the source portrait or source video, and driving video šŸ¤—šŸ¤—šŸ¤—")
gradio.exceptions.Error: 'Please upload the source portrait or source video, and driving video šŸ¤—šŸ¤—šŸ¤—'

slender vault
#

prolly need to set the imageio (which im guessing is the file?) in the enviroment variable and thats where i wont be much more of help and should poke around down in tech support two channels below

fervent thunder
#

don't rly need much VRAM these days, 8GB is fine
the fastest version of flux is svdq-int4-flux.1-dev.safetensors which just uses 6.64GB VRAM

#

if you are less fussed about speed then flux.1-lite-8B-alpha-Q3_K_S.gguf uses 3.74GB

fresh ruin
#

anyway to batch inpaint, but still be able to select the areas you want to inpaint on each picture? (useful for removing watermarks)

copper crystal
#

news articles repeating what they've said isn't LLM slop. that's adwords in action. and small screens that you can only scroll slowly being aprimary form of information consumption.

you have to scroll past ads to look at it all. and more keywords in proper contexts allow you to rank higher in search results.

LLMs aren't the reason it's going on

abstract quarry
abstract quarry
# gritty dust i'm also back after a long time. i'm reading that forge is the best webui to use...

I always advertise InvokeAI here xD
It's in my opinion the most intuitive webui by far and its also the easiest to install. Forge, in contrast, is still the ugly gradio app we know since auto111. I think the reason its called the "best webui" comes rather from the fact that swarmui and comfyui are soo bad in terms of usability. If you want to install hundreds of plugins and know what you do, Forge or ComfyUI are the way to go. If you are new and still learn and search for a ui which is intuitive and easy to use, I would recommend invokeAI

fresh ruin
#

is training a lora on Illustrious the same as training on pony? How many images should I have on the dataset?

fervent thunder
#

InvokeAI are great yeah their canvas is very good
they are more stable than Comfy also
I like that they give an official docker image too

oblique goblet
#

hi

final viper
#

Can someone explain to me how the Lora zoom slider work? Is it just the Weight that u use as slider?

#

nvm found it

gritty dust
#

i installed forge and it works outside the box so i will stay with it for now. the model it comes with however is barely usable, do we have an index of good models?

warm junco
gritty dust
steel burrow
steel burrow
#

At first I used A1111, but I liked the modularity of ComfyUI more, I'm learning how to create the node chain. I can probably train some models later, currently these models are loras instead of entire checkpoints from what I understand. I followed the hype of the SDXL launch, but my 1060-6gb was not able to support the process.

#

Previously, in sd 1.5, I put an image and the image name as the CLIP description, currently I verified that for LORA it is exactly the same thing, but divided between an image file and a json with the description, right?

slender vault
steel burrow
#

followed you

#

can you recommend me some channels/worksflows for flux, thanksss

quartz hare
#

Hello šŸ‘‹

finite echo
#

I saw image made with a prompt that contain this"score_9. score_8_up, score_7_up" and i don't quite understand the mean of it may someone explain

abstract quarry
#

you only use this kind of prompts when you are using the Pony model

finite echo
abstract quarry
#

they give all their training images a score and then add this score to the caption of the image

finite echo
#

K thx

steel burrow
#

because:

  1. they're lazy
  2. to hide the process and avoid copy
#

i know who are the creator youre looking, ive saw the same prompts XD

copper crystal
#

astralite completely destroyed the sdxl text encoder with his genius captioning. i doont know why he's so praised. He's not lazy, because he went out of his way to do this broken quality tag approach, which was never needed. Then he brags about how he disaligned the text encoder and acts like he created a whole new base model.

The worst part is his primary customer base on his generation service are under aged kids. and he's held up like a hero for it. The whole community around pony is just so savage. Not a good look for generative ai work. Don't publically admit to anyone that you like Pony because people who know will judge you.

compact swan
#

does anyone have any recommendations for running local SDXL on ubuntu?

#

comfyui has some annoying errors I don't know how to fix

#

and forge ui doesn't have a linux version at all

copper crystal
#

forge ui has sh launch files. it's just a python environment. theres no reason it wouldn't work on linux.

I think your problem has more to do with you not knowing your system very well.

#

linux isn't known for "one button installers" so you won't find them there

#

comfyui should work with linux environments easily too. The UI wouldn't be the problem

#

One thing you'll learn quick with linux world is that you can't pass the buck as easy. You've gotta do the leg work. If it doesn't work, that's probably going to be a layer 8 issue

slender vault
compact swan
#

i didn't find them when i looked

#

i'll try again later

copper crystal
lucid bobcat
#

Does anyone know why sd ultimate upscaler does 3 passes in comfyui? Is this normal? The documentation says nothing about multiple passes and ChatGPT knows jack shit as always.

#

I guess the second pass is for seam fix but the third one is an enigma.

low moon
#

So I've been out of the loop for a couple of months. Is AI images and video perfect by now? If not I'll check back in summer 2025 or something like that.

lucid bobcat
low moon
#

I don;t have a precise barometer but these things move faster.

low moon
#

Ai moves quick.

#

Images are kinda good enough already...

#

Video not so much yet

lucid bobcat
# low moon Awww, that;s too pessimistic.

It's not. It's realistic. It can't possibly be, there's too many difficulties that can't easily be overcome, like lack of good training data and lack of good image captioning as well as ambiguous language. And don't forget that we are way past Moore's law.

low moon
#

Quantum computers!

#

stuff like that

#

Computing power is like money, you throw enough of it at a problem and it gets solved.

lucid bobcat
# low moon Quantum computers!

Quantum computer research, as well as fusion reactor research has been going on for decades. And it will continue for decades before we get anything useful.

low moon
#

Nah. This is the last generation. We'll get the best outcome. This is it folks.

copper crystal
#

i saw a cool device invented. quantum positioning systems. it cools a bunch of atoms into a bose einstein condesate state of matter, and then takes a reading from that bunch of atoms, and cycles them again. it's kinda a bazooka sized apparatus, and it can do accurate global positioning with no external devices. so no GPS satellite signals.

These are huge breakthroughs. but also scary. because now warfare machines can't be disrupted with gps signal jamming.

#

there's hard math problems too that quantum computers might one day solve. like does P=NP or does P!=NP. i'm hoping for the latter because then encryption is still mostly safe for now

low moon
#

I cna tell you with 100% certainity that P=NP because the world will not stay the same forever.

copper crystal
#

prove it and you get a nobel prize that year

#

they'll likely hold a special ceremony

lucid bobcat
low moon
#

i dotn need any of that

#

its juts a secret between the two of us

#

shhhh

copper crystal
#

Neither is proven yet. But there are suspicions and ideas that both camps have.

proving eitehr will have huge implications either way. I'm hoping that its not equal

finite echo
#

Hey i'm having problem with different thing when generating image i'm a beginner in this domain and i can't find what resolution should i pick because if i set a one to high it takes to long and if i set a one to low it looks like garbage and i cannot find wich one should i pick

Also should I use the Hires.fix when generating image and if yes how

copper crystal
#

How much vram do you have? Higher resolution images will use more vram.

sd15 models have a native resolution of 512x512. but They can generate higher images with hires fix. Essentially doing a small one first, then sizing it up and denoising it some more. But that second step will fill more memory.

sdxl has a native resolution of 1024x1024, 4x as much. So hires fix isn't needed, but sometimes i use it still just to get refinements on the second pass.

You could try loading your model with fp8 memory precision, which is fine for most cases but could take a little longer to load your model on the first run. Since more calculations. That means the weights fill half the amount of memory, so more memory is left over for the image generation.

you could also try something like tiled diffusion. Forge UI has this built in as multidiffusion. It does small patches of your image at a time. There are lots of these kind of extra solutions too. SD ultimate upscaler is another one, that'll upscale images using tiles.

finite echo
copper crystal
#

That's plenty but it'll be tight. i'd run sdxl with fp8 mode if i were you, and keep gens around 1 mega pixel (1024x1024 or other aspect ratios that have a milli pixels)

finite echo
copper crystal
#

helps to keep your task manager open on your gpu's performance tab. you can see how much memory is being used. if it's maxing out your gpu vram, it'll slow it down to mud speed

#

Detweiler's comfyui series

#

i use forge ui these days, but comfyui is a strong tool that allows a lot of flexibility and control

finite echo
deft horizon
#

Just a question/Suggestion : Shouldn't there be a Flux Channel in the "Stable Models"?

quartz siren
deft horizon
#

Oh.I didn't think This server was that exclusive. But that makes sense.

abstract quarry
low moon
#

Qubits r quick!

copper crystal
#

don't dm this guy. it's scam bait.

#

i dont even have to to know it. i can smell it

#

exactly what i said

#

šŸ‘

lucid bobcat
copper crystal
#

Marvy the failed scammer everyone. Bad at real business so he had to try scamming, but is bad at that too. Give it up for Marvy everyone!

#

literally idled on the server since months ago, just to spam scamm invites today.

If you're actually DM'ing this guy, reevaluate your life choices

gritty dust
#

i was able to use civitai.com for like 10 minutes and now the website isn't loading on my browser, other websites work fine and other browser works fine, is this a known issue?

fervent thunder
gritty dust
#

it was back a half an hour later, maybe just server issues, yeah

desert dagger
slender vault
teal hull
#

i am new to coding, developing an star app looking to mingle and learn. hello everyone.

lucid bobcat
astral acorn
#

OlĆ”

copper crystal
#

$250 inference machine

lucid bobcat
copper crystal
#

anything. low power $250 inference machine

wide flame
#

hey can you guys help me out a little? im tryna use a flux model for image gen and im using ae.safetensors vae but im still getting burnt/blank images

copper crystal
#

don't use adaptive samplers. use euler.

wide flame
lucid bobcat
glass grotto
#

Does upgrading ram and going from an ssd to an nve make sd faster ?

warm junco
glass grotto
warm junco
#

But ram upgrade never hurts

glass grotto
#

Whatā€™s that ?

#

Windows pagefile

warm junco
#

Thats the file that gets used when your Ram is overfilled. Its like virtual Ram and uses disk space of an SSD preferable

glass grotto
#

Is it safe to increase it ?

#

If so, how ?

warm junco
#

Make sure its only enabled for the C drive and not for any other drive.
And then set it to 16000 Min and 24000 Max.
Then apply and restart the PC.
Also make sure to have at least 15gb free space on C.

fervent thunder
#

in a handheld

#

its amazing

#

there might be some fine print though, that the numbers they gave are for Int4/FP4

#

which is fine, but would be deceptive

#

Nvidia's favourite marketing trick is to do comparisons where one thing is in FP32/FP16 and the other thing is in FP8/FP4

abstract quarry
#

Flux is a cfg-distilled model, so cfg does not work as it does for non-distilled models

#

you can use cfg, but you should use low values for cfg (1-3) and maybe start cfg not from step 0 but at step 3 or so

copper crystal
fervent thunder
#

if I were them I'd say the FP4 number so I suspect that's what they did

wraith ivy
#

Hello

cold frost
#

hiļ¼Œnew here怂good 2 everyone

barren mulch
#

what's the best local AI image generator right now ? last time I was doing this I was using SD1.5
with the stable-diffusion-webui

lucid bobcat
barren mulch
#

thx very useful

analog yarrow
#

I got flux running on local (yay) what kind of guides should I look up on how to copy the style of an image I provide?

Right now iā€™m promping and getting good images for a scene in my game

But I want to prompt for different scenes with the same style

restive tusk
#

Hello!

restive tusk
lucid bobcat
#

You don't uninstall stable diffusion. If you don't use the model anymore just delete it.

hollow surge
#

you can Reinstall the systemšŸ˜†

#

discord åÆ仄ēŽ© stable diffusion 吗ļ¼Ÿåƒmidjoureyäø€ę ·ļ¼Ÿ

silver robin
#

greetings all

analog yarrow
#

I think my comfyui is stuckon a style lora. Everything is coming out the same style no matter what I do tried loading sdxl and schnell same exact style I have the lora nodes all deleted and restarted too

ancient jetty
#

because its a scam

#

lmfao they set it up so i could ping @ everyone

#

so i just pinged and said "this is a scam"

#

its community based so it might be slow

#

wdym wallet

#

like bitcoin?

#

oh so its like

#

not even an advanced scam

#

its just like

#

give credit card and we double ur money!

#

they didnt even respond to me

#

probably because i started to send actual things that i "needed help with"

#

no no

#

they use the nfc chip

#

in your credit card

#

for processing power!

#

nfc chip is the thingy mcjig that lets you tap on

#

yeah i was trying to say that its using that chip that has the processing power of my pet rock as a gpu

plain karma
#

hi!

half quarry
#

"Hi, I'm an ethical hacker. I specialize in identifying vulnerabilities in systems and helping organizations strengthen their security and I can hack and recover all social media account, recovery of lost funds,unban, game hack hmu for your service

crimson sorrel
#

hello

low moon
#

If you get a dog call it Barkolomeo.

hollow anvil
#

Hi y'all, quick one: is there an established set of stable diffusion tools for asset generation that game devs use seem to favor? For example, generating isometric characters, etc. ?

lucid bobcat
hollow anvil
#

I'm a coder, not a graphics modeller, so looking for a good solution, doesn't have to be perfect

modern pagoda
#

SD has 3D models pretty sure. Looked chunkier than N64 graphics tho

lucid bobcat
hollow anvil
#

I tried itch.io , what others do you recommend?

modern pagoda
lucid bobcat
#

I don't know, used to create my own place holders graphics when experimenting with game ideas.

lucid bobcat
bleak matrix
#

Good afternoon, everyone! How are you all doing?

haughty gulch
#

Hello

silver moon
#

Hello everyone,

I am an artist looking to create a model from my existing images so I can quickly generate unique art to use for other projects. I have tried using Upwork to find people who could potentially help me with this, but everyone seems to find it tricky. If anyone has any knowledge on this, please let me knowā€”I would be happy to pay.

silent turtle
#

Hello

lucid bobcat
quartz siren
inner star
#

hello

slender vault
inner star
#

how the fuck do i get the diffusion soundvoard

astral nymph
#

hello everyone !

tender vault
#

šŸ‘€

#

ji

#

hi

astral nymph
#

what prompt should i use to achieve that vibe ? i think that image was generated with midourney

#

i want to make an image with that kind of texture and vibe

astral hearth
#

hi

quick wolf
#

Hey someone do here faceswap with A1111?

gritty dust
#

can someone tell the differences between SDXL and SD 3.5? Civitai seems to have a lot more models for SDXL available, is tehre a reason to still consider SD 3.5?

slender vault
#

3.5 is the newest model

gritty dust
#

ah ok, thanks

fervent thunder
#

@stray zinc

severe axle
#

Guys I tried to convert a digital image I have into a realistic photo but it takes such a long time (6 minutes so far). Did you guys know how to convert a digital image into a realistic photo in Stable Diffusion ?

warm junco
#

Try lower the resolution

severe axle
#

How did you find it ? @warm junco

warm junco
#

I use the local version of Stable-diffusion
Automatic1111 or Forge Webui

severe axle
#

How much time does it take to convert a digital image into realistic photo without using img2img ? @warm junco

warm junco
severe axle
warm junco
vestal coral
#

I just want to say, SD3.5 large is extremely based.

#

It just does the style I ask it to do, and doesn't do stupid style locking. For that, it is extremely based.

lavish iron
#

ComfyUI - I have 32gb ram and a 4070 ti super with 16gb vram, should I be able to use flux1-dev with a lora? When I use the Load Flux Lora node in comfy, it gives an error. list index out of range

#

do i just use the regular load lora node?

warm junco
lavish iron
warm junco
#

A little output difference

lavish iron
#

does either one produce any noticeably better results?

warm junco
#

for comfyui you also need the gguf loader nodes

lavish iron
#

I'm able to generate images using flux1-dev using the regular lora loader - LoraLoaderModelOnly node, will this output be any different if the Load Flux Lora node was used?

warm junco
lavish iron
#

is the restriction to fp8/q8 because flux1-dev is using more ram/vram than I have and paging it to HD?

warm junco
#

the fp8 / q8 is half of it and should work much better

lavish iron
#

Thank you.

#

It's nicer to know that it's a limitation of my system than thinking, it's just broken for me šŸ™‚

warm junco
#

np, yea the 23gb version is for the 4090 users xD

#

and even on that its "slow"

lavish iron
#

I can generate images, but it is slow, several minutes

warm junco
#

yea its using your pagefile for that because the model doesnt fit into the vram

#

and that slows down

lavish iron
#

do you know if i can use the same clip models, ie clip_1 and t5xxl_fp16?

#

and do i need to use a lora explicitly trained for the fp8/q8 models?

#

i'm getting a large blob of colour, not a decent image

warm junco
lavish iron
#

Thank you šŸ™‚

#

i think i got it working, i was using the regular k-sampler. As soon as i switched to the X-Sampler, it worked

lavish iron
#

if i upgrade to 64gb of ram, will this likely be sufficient to run the full flux1-dev perhaps?

#

it's 7200mhz ddr5 ram, so is pretty fast

ionic wraith
#

Has anyone already created an example/full script to request/generate and receive the output from an generated image? (Python)

lavish iron
#

this is basically what all the UI's do already. if you look at the nodes in ComfyUI for example, they are just configuring the parameters to be passed when the generate button is pressed. what you are asking isn't really as simple as it sounds. if you are a coder, you can review the git repo's for the components that make up the generation configuration you are intending

#

if you are doing remote SD, i'm sure there are scripts for executing that, just look for the api, perhaps there would be examples

fervent thunder
#

yeah its a client-server architecture

#

someone made a nicer input method called ComfyScript https://github.com/Chaoses-Ib/ComfyScript

#

it just writes an API call but in a nicer syntax

#

and some cool alternative modes where you can call nodes one by one

lavish iron
#

it has a transpiler, that should do what you want

swift isle
#

hello

lavish iron
# fervent thunder yeah its a client-server architecture

the comfy ui is exposed on a localhost port, can this be hosted on your local network somehow perhaps, so it can then be accessed from the client machine. I haven't done network hosting for a few years, so am really quite rusty about this.

#

look into Secure tunnel services for that option.

#

or use ipaddress of server:port

warm junco
fervent thunder
lavish iron
fervent thunder
#

Comfy is not really in a state where it would be a good idea to deploy it
its fine for experimenting or playing with new tools

#

or using in the way people use photoshop

#

if you're gonna make a custom server and network setup I think its better to switch to pytorch at that point

lavish iron
#

is this just for yourself, so you can run a server and access it via another machine on your network?

fervent thunder
#

was !InstantNameOfficial who was asking

lucid bobcat
lavish iron
#

lol, sorry bud, lost track of who was asking

fervent thunder
#

yeah --listen does work
Comfy has the functionality to be used that way I just don't think its robust enough yet

lavish iron
#

I thought about running a 3060 on a spare machine I had myself, but ultimately, the spare machine is dell, so they skimped and added exactly zero upgradeability to the pc, so won't accept a powered graphics card.

#

and of course, the mobo won't accept anything but a dell power supply

fervent thunder
#

ah yeah I had a lot of trouble with trying to upgrade Dell prebuilts

#

can be very tricky

lavish iron
#

it was a free pc, so can't grumble.

#

it makes for a very large oversized paperweight! šŸ˜‚

fervent thunder
#

if its free that's fine

#

I know someone who paid multiple k for Alienware

#

was not a good idea

lavish iron
#

any prebuilt pc is going to be built as cheaply as possible imo, regardless if it cost Ā£3k+

#

particularly if it's sold from a large corporation with shiny green badges and glowey fans

fervent thunder
#

the power supply and motherboard were pretty rough yeah

lavish iron
#

a low quality power supply is the one thing a pc shouldn't have, that's so long as you don't want to be buying a new one in a few years.

fervent thunder
#

I'd be worried about it harming the GPU yeah

lavish iron
#

i spent about 2k on a pc back in 2012/13, all the asus, corsair parts lasted until i replaced them a couple of months ago. the Ā£600 780ti was msi, that broke numerous times and bust completely after 2 years. The asus graphics card lasted 10 years.

fervent thunder
#

oh this matches my experience very well
I always buy Asus stuff and I have a corsair case still working well

lavish iron
#

that msi was the first time i didn't go asus, i regret that, lol

fervent thunder
#

msi can be a bit squiffy yeah

lavish iron
#

it actually cost me about Ā£1100, as I wanted to go the SLI route, for 2x 760's, however the mobo, id' bought was a crossfire, so had to buy a new sli compatible mobo. This proved completely useless as sli 2x760's did nothing. so just bute forced performance buying the 780ti.

#

gigabyte are usually good also

fervent thunder
#

oh yeah I remember SLI

#

people used to get dual/quad gpu for gaming but then they stopped

#

and now they get high gpu count again for llm

lavish iron
#

it stopped, because it relied on nvidia setting up the game profile for it to work

fervent thunder
#

ah okay I was never sure why it went away