#πŸ’¬ο½œgeneral-chat

1 messages Β· Page 162 of 1

urban storm
#

im new to this kind of stuff man 😭

#

and i hope i didnt installed the wrong thing

#

rn im installing this

fervent thunder
#

lora is where you trained the model but only part of it

#

on your images

urban storm
#

like photos of me

#

?

fervent thunder
#

yeah that's right

urban storm
#

will the photos of me be publically for everyone ( stupid question )

#

also will i be able to do such high quality and nice edits like that with my laptop?

fervent thunder
#

there are secure ways to use stable diffusion

#

but lots of ways that are not secure or private

dawn jewel
#

Hi my name is Alvin and I also do stable diffusion and mid journey Ik I'm not supposed to be here for ai so yes hello.

wet grotto
#

refiner is only for sdxl?

abstract quarry
#

in sdxl you would usually run it directly on the latents. When using with another model you would first decode and reencode the image

#

but honestly... the refiner is not that good. It might improve sd 1.5 images and sdxl base images, but most custom sdxl models are already better than the refiner

pale latch
potent spire
#

am i the only one whose Google Colab refuses to run Flux?

#

it worked before

urban storm
#

hey guys im back

#

so

wet grotto
pale latch
#

i like this

proud arch
#

hi

honest stirrup
#

hi

harsh fable
#

How can I use stable diffusion in my python project?

quartz siren
lean hull
pale latch
#

why would AMD kill off ZLUDA. Are they dumb?

desert dagger
pale latch
#

sure but... all the people with amd card who have been enjoying zluda over direct ml are now up shit creek today

narrow kernel
#

well it was always at risk of being canned and it has been known for roughly a year now that it was probably inevitable

#

part of holding a patent is that you have to enforce patent laws. if you don't, you can actually end up losing your patent on something

desert dagger
fervent thunder
#

for the last decade its as if AMD secretly works for Nvidia

#

and are trying to sabotage themselves

pale latch
fervent thunder
#

imagine if AMD had kept parity with CUDA 10 years ago

pale latch
fervent thunder
#

the biggest mess up was Intel cancelling Knights Corner line

#

they essentially had the makings of a good GPU division

#

and then just canned it

desert dagger
#

they might be rethinking that now

fervent thunder
#

oh yeah apparently intel have GPUs now

#

I haven't been following it

desert dagger
pale latch
#

they're not bad. they're contenders. i woudlnt' be surprised if one year intel dropped new resarch into their dgpu line

#

keep an eye on it

fervent thunder
#

I do still like AMD GPUs for certain linux stuff

#

not everything plays well with linux all the time

#

but AMD can be better than Nvidia in that particular aspect

pale latch
#

amd works great on linux and gaming does well there with amd cards

fervent thunder
#

yeah I have found it okay

#

I had trouble with early ryzen

#

first gen ryzen took many months or maybe even a year to work

#

on linux

lone dust
fervent thunder
#

this is just doing control net you can re-create yourself

#

Steps:

  1. download comfy ui
  2. find a workflow online that has control net
  3. set canny edge control net to strength 0.7 and end at 0.3
lone dust
fervent thunder
#

those were my numbers they aren't right for everything

#

the idea behind strength 0.7 is just to turn it down a bit because there are costs to high strength guidance (whether its CFG or control net)
and the idea behind end at 0.3 is that around 0.3-0.4 sigma is where the structure of the image has finished being made
so you can turn it off after that

lone dust
#

I'm not sure if your solution takes this into account. this sketch2lineart thing does it though

fervent thunder
#

oh I see
there are a couple of ways to get that

#

the main one would be using a lora, if you can find one

lone dust
#

(γ… .γ… )

fervent thunder
#

but a good IP adapter workflow would be good

#

there is a node called IP adapter mad scientist (terrible name sadly) which works block-by-block and its very good

lone dust
#

*link

fervent thunder
#

going block-by-block is particularly good for IP adapter because structure and small details are different blocks

#

is the code for this huggingface space public?

#

not sure if you can get hold of it

lone dust
#

ahh no code...

#

then what if I make python code to put image in the app, run it, and download the output?

lone dust
#

I would if I have to tho

fervent thunder
#

I just tried it and to be honest you could do better

#

their canny edge settings are a bit lacking

#

you only really have to learn 3 things- how to install comfy UI, how to use control nets and how to use IP adapter
its not easy for beginner but you could learn this relatively fast

lone dust
fervent thunder
#

its not easy but its not hard
its sort of medium difficulty

lone dust
#

and even better!!! damn!

fervent thunder
#

yeah they set the detail level a bit low

lone dust
#

not hard at all, as long as idk what IP adapter is but aslong as I don't need to try 20 different param combinations

fervent thunder
#

maybe 5-6 paramaters at most

lone dust
fervent thunder
#

good luck

#

there is also softedge if canny doesn't end up working too well

#

and there is scribble control net

#

but most people use canny

#

if you turn off canny half way through then you get softer lines anyway

lone dust
#

but just using ControlNet canny with dreamshaper checkpoint was giving me shaded areas instead of binary line image

lone dust
fervent thunder
#

the setting that they have low is the canny thresholds

#

depth control nets also have thresholds

#

if you use it a bit you will see quite quickly what this does

lone dust
lone dust
# fervent thunder if you use it a bit you will see quite quickly what this does

You said it's a simple setup with just an IP adapter node right? You said I'll figure it quite quickly, but I don't think line art making ControlNet tutorials are very ubiquitous. I've been searching for line art making tutorials for some days and they're very few, and now searching with IPAdapter even fewer. People show line art to colored but not the other way around

#

Can you give me some idea about how a simple setup with IPAdapter can be used to turn regular image to line art?

fervent thunder
#

its just a case of understanding what IP adapter does, particularly the ones that work block by block

lone dust
fervent thunder
#

the mad scientist IP adapter can go block by block yes

bold star
#

Is there no model.ckpt anymore? Or am I looking at the wrong repository

#

Sd3_medium.safetensors or all of them needed?
Sorry been a long while

pale latch
#

i'm renaming al my files to .sft

#

the old gods have spoken

bold star
#

Ty

dawn vine
#

pig

lone dust
#

Hello everybody! I wonder how I could tell if the code for this hf space is hidden or not.
https://huggingface.co/spaces/tori29umai/sketch2lineart
It has app.py, so I thought I could git clone and run it, but there's the part that says
dl_cn_config(cn_dir)
and what it does is:
config_path = os.path.join(os.getcwd(), file_name)
shutil.copy(config_path, file_path)
Does that mean they're running this app.py in their local computer directory where they have their config.json
and they're copying it over to model_dir/filename at runtime,
which basically means this code can't be run in my local computer, right?

narrow kernel
#

And that is assuming you have access to the HF repo or repos that contain the models and files

#

(in terms of being able to download stuff like the safetensors and config.json files and so on)

#

looking at their code, they pull a bunch of various models and store them locally in some folders like cn_dir = f"{path}/controlnet" which basically just means within whatever folder you have the app.py in. if you look within the utils/dl_utils.py file you'll see the various models they use. (They should automatically download on the first run)

#

but other than that, looks pretty standard and you really shouldn't need to change much. Just make sure to git clone https://huggingface.co/spaces/tori29umai/sketch2lineart the whole thing to a folder and make sure you have all the proper dependencies. highly recommend making a virtual conda environment for it, shit gets messy if you use a global python install

#

with conda and from within the app.py folder, you can run conda create -p ./venv python=3.11 followed by conda activate ./venv and from there, run pip install -r requirements.txt

#

that will get all the requirements for your virtual environment setup. after that, you can run python app.py and if you are lazy and don't want to have to activate the conda environment every time, you can make a launch.bat file with conda activate ./venv && python app.py in it and it will do it for you and launch the app.py

lone dust
lone dust
narrow kernel
#

I just threw in 3.11 off the top of my head. Not all libraries are built for newer versions of Python. That's why virtual environments are a must. The whole AI world is a clusterF of various reqs, so you end up with a million torch installs lol

fervent thunder
#

oh someone has added the thing thats meant to be a better PAG
https://github.com/pamparamm/sd-perturbed-attention

#

they did it nicely as well, with block by block, scheduling and rescaling

lone dust
lone dust
narrow kernel
#

oh sweet, did it end up working? like you're able to use the gradio frontend now?

#

and don't feel bad, python is super f'in annoying sometimes

narrow kernel
lone dust
lone dust
narrow kernel
#

well hopefully it works out. i have to get the hell off here before my eyes start bleeding

lone dust
wet grotto
#

ADetailer is only to people, don't exist nothing to buildings?

hearty yarrow
#

what is the best web version for mac for nsfw generation to commercialise? its okay if its paid

low moon
sudden ruin
hearty yarrow
#

I’m not posting anything here like that, just a general question πŸ€§πŸ™πŸ»

fervent thunder
#

unsure if this is the best place to ask about this: i'm a pretty old guy using a pretty old pc (purchased 8 years ago). a bit ago i upgraded the gpu and storage in it to be able to do more AI stuff. i now have a pretty bottlenecked pc using a 3060 12GB but only 16GB of 1600 ram. this is fine for general AI stuff, but i've run into stuff my pc just doesn't handle well and i strongly suspect the ram (and maybe the cpu?) is the issue. looking into it, my motherboard and cpu don't support higher than 32GB 1600 and if i wanted to get higher/faster ram then i basically need to upgrade the entire pc (better ram requires changing the motherboard and cpu, which when added to the new higher ram requires a better power supply as i currently use a 400w, which at that point i may as well replace the case too since it's pretty beat up...).

i realized i have what might be a problem others have previously run into: i'd like to upgrade these components, but i already have decent storage (4tb ssd, seems to run fast) and a good-ish gpu (the aforementioned 3060 12gb). my age makes me reluctant to assemble a bunch of parts myself, but there don't seem to be any retailers that allow you to "custom build" a pc without the storage, operating system, or gpu (which makes sense). i've checked locally and of the limited options available to me, none offer this service.

TLDR: any suggestions on a place that offers a "custom build" without including storage, OS, or the gpu? alternatively, any recommendations on how someone with somewhat limited hand range of motion can easily put one together? atm my best option appears to be getting a prebuilt machine with integrated graphics(so it doesn't include an additional discrete gpu), but that seems like a bit of a waste as i'd be paying for the storage, OS, and integrated GPU. the non-integrated prebuilts don't typically offer the better ram speed/capacity

opal bramble
#

Sounds like u need to use a vps right?

#

Im trying to find out how that works

#

I dont know if vps are charging by the hour of use or by the hour of runtime

sudden ruin
#

Especially since you already got a few components

fervent thunder
#

my family isn't around any longer and I'm pretty much a hermit outside of the internet. no irl/physical friends lol. appreciate the suggestion though! might try a "fiver" type service or something πŸ‘

pale latch
#

building a pc is relatively easy and thanks to the internet there are tons of references now. no more being deathly unsure of how hard you need to press on taht dimm chip to get it to socket properly. just watch a video of someone putting all their weight on it until it goes SCHNINK! it's like lego pieces really. gpu goes into the gpu slot. ram goes in the ram slot. it's mostly ez pz. sometimes nimble fingers are needed is all. tehre's concerns about touching the chips and stuff, just handle components like hockey cards or limited edition pogs.

#

i've oftne bought barebones prebuilt pc's and then added components to them. black friday sales are good to get clearance models

amber yacht
#

gm

wet grotto
#

did you have a workflow with sam to recomend me?

fiery wasp
hearty yarrow
fiery wasp
#

local install will work, but it will be slow for SDXL
it will work with decent speed for SD1.5

#

but models that require less steps like lighting based ones, will reduce time, simply by reducint the needed number of steps

sage compass
#

If I am reaching lowvram mode is it important to switch to the lower model?

cedar salmon
#

if its completing successfully, no. Other versions could possibly save time

ornate pollen
#

Hi guys is 900 credits monthly enough to use ?

solid owl
#

Hello there, am a website designer offering the creation of a website
designing service for free. DM if interested

sonic gazelle
#

Please do anyone know a way I can do face swapping on a pc with intel cpu. I'm willing to pay for your service

sonic gazelle
#

??

fervent thunder
#

reactor or roop

full lark
timid lava
#

imagine being such a talentless hack that you use AI to create "art"

pale latch
full lark
empty stone
vestal zealot
#

yo can u make videos w this

fervent thunder
#

yeah

#

Open Sora 1.2

#

edit for correct name

nova lily
#

Ive been trying to use noise inversion/control net tiled for detail, but tutorials are outdated or their settings dont apply to me (I get botched results) not seeming worth the hassle.
reference this video https://www.youtube.com/watch?v=44waH3sDYOM

fervent thunder
#

noise injection gets massively over complicated
all it is, is that the model expects the latent to have a certain noise level for each step
and if you put more noise in then it thinks that the shapes made by the noise are real image features so it turns them into more details
noise can be added either to the latent or to an image before VAE encode, doesn't matter which

honest stirrup
#

Hi Guys, anyone experienced in s/it jump before?

I changed latent size from 640x360 > 1280x720
the s/it jumped from 3 to 35

is it normal?

Requested to load Flux
Loading 1 new model
loading in lowvram mode 21931.88662261963
100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 40/40 [02:12<00:00, 3.32s/it]
Requested to load AutoencodingEngine
Loading 1 new model
Prompt executed in 146.34 seconds
got prompt
[rgthree] Using rgthree's optimized recursive execution.
Requested to load Flux
Loading 1 new model
loading in lowvram mode 21931.88662261963
8%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ– | 3/40 [01:59<22:02, 35.73s/it]

honest stirrup
#

now it's 168 s/it πŸ€¦πŸ»β€β™‚οΈ

fervent thunder
#

its going exponential !

feral pike
#

anyone here using amd gpu for SD? how is the performance?

honest stirrup
woven kindle
#

Hi, anyone know if it's possible to video to video in local, I mean I have a real video and I convert it in anime style?

#

similar to what goehance do because it's limited to 20s per video and it's expensive

main junco
#

Is it possible to use fooocus with controlnet?

#

And does anyone happen to know - what model/options does fooocus use to achive face swap functionality?

low moon
#

Go paint on a cave and leave the real artists to embrace the best tools.

wispy fulcrum
#

hello everyone. I have a Flux.1 question: can someone explain to me exactly what the max and base shift parameters on the model sampling flux node actually do? How do they work?

low moon
#

Who knows man.

#

Test them. XD No one knows much with this. emerging technology!

#

I envy the fools born now ( today like at this minute as I type this) who will be using this tech in 10 years and will have perfect models that will literally generate a 120 minute movie personalized for your viewing experience with 0 problems and consistency issues.

#

F U future assholes. XD

#

Here I am dealing with ugly hands.

night cipher
#

Hey admins, I just posted a video for the Gemini api dev competition on YouTube , do you mind if I post it here and get some feedback?

wise ginkgo
#

what should i download to compare? SD3 and Flux.1?

feral pike
#

gonna try ROCm

#

see how it compares with directml

#

@warm junco do u know which one would be faster ,ROCm on linux or ZLUDA on windows?

fast sage
unique robin
#

How can i generate images

#

??

warm junco
fast sage
#

I'm trying to create a story bord, but generations don't have cohesion, would you guys suggest how to keep a similar cohesion along different prompts

feral pike
warm junco
feral pike
#

how do i use it?

pale latch
warm junco
pale latch
#

fully supports? sure. sure.

Easier effort ot go work a full time job, earn enough for an nvidia card, and use that

#

and then pray that amd driver updates don't break your system

warm junco
#

xD its a linux subsystem and the performance of SD is the same as on native linux. we already tested that

pale latch
#

amd killed off zluda. stop supporting them in this field.

warm junco
#

yea the zluda move was really bad

#

but it still works

feral pike
#

is there a guide on how to do it?

pale latch
#

yeah. it just wont be maintained anymore. it's a bad move. perhaps the dumbest i've seen them make in some time.

warm junco
warm junco
#

he will now try something else with zluda

pale latch
#

original plan for it before amd funded him was to be for cuda on intel cpus iirc

proven temple
#

is there way to tell whether your stable diffusion froze up? the progress bar on my cmd is at a flat 0% for a good while now

pale latch
fervent thunder
#

I might be a moron but can someone tell me how turn on ''v-prediction'' mode?

#

Everyone talks about it as if its the most obvious thing in the world, but never how to actually do it

keen kraken
#

hey guys, anyone know how when i download the images with the download button it makes the name different than just Image (1), Image (2) so on

sudden ruin
keen kraken
#

im using forge

sudden ruin
#

In a1111/forge you can set that in the settings

#

So it uses the prompt as image name

#

At least thats what I did I think

keen kraken
sudden ruin
#

I dont know what it's called from my head, id just to to settings and search for save

#

Should be somewhere in there

keen kraken
#

if you can provide certain proof you arent trying to scam someone im sure its fine

#

also on civitai some people do take commissions, but in all honesty it isnt that had to make your own loras

#

the problem there is you need alot of different sample images of the character to feed to the lora trainer

fervent thunder
#

this is nice website for ai but

#

u have to pay to download simple lora

keen kraken
#

let me have a look at the people on civitai i know a few i follow offer commissions

pale latch
#

civit has tons of baldurs loras

#

probably won't find that kind of specificity

#

civit is a good place to find commission too. "bounties" . i don't think you're a scammer but i think it's super shady to hire for that sort of stuff through discord servers. to many variables. to much opportuniyt for a scammer.

civit bounties is where you offer up their site's buzz as a reward. protects you. protects them. everyone's happy

keen kraken
#

you can look at creators Novowels, Nochekaiser881 and marlosart, they make alot or loras i like and do commissions

#

look at their pages and they give links to commission forums

proven temple
#

what's a xformers? my cmd said I don't have em

cedar salmon
#

a package that was widely used prior that is unneeded for most now, not an error message, just an info message

urban storm
#

hey guys

#

anyone

#

why dont i have the img2img tab

#

on invoke ai

oak latch
#

can i pay someone to help me gen realistic images

river sparrow
#

I can help you with some realistic sexy anime for free? πŸ˜„

#

check general-with-images

oak latch
scenic garnet
#

Why can't I find the stable foundation community in my search?

pale latch
#

That's what this server is. You're here

#

They changed the name of the server back to diffusion I guess

shadow oyster
#

Check

nova lily
#

A1111: What live preivew settings do you guys use, It seems if you disable everything you do get 1-2 faster gen perhaps.

I was wondering, What s the point of the image format anyways? and then I was thinking are these individual frames of the preivew actually written/created on your drive?

dusky vapor
#

hey hey

#

i am new in server

warm cedar
#

Heya I haven't checked the server in a while. I only see an anime channel in the Dreamer Communities section. Am I missing something? Are there more channels dedicated to certain topics?

fervent thunder
#

only anime

#

anime and non-anime are the only two topics rly LOL

fervent thunder
hardy nexus
#

did the sai devs came back to this discord or still missing ?

chrome lichen
#

Hello guys, how are you? I am new here, trying to use Ai create my avatar used in my business prfile

#

Any angles to show me how to do it? It's a new journey to me, excited and frightened a little bit

slender field
narrow kernel
lone dust
#

Hey guys! I'm using comfyui inside stability matrix. How do I run comfyUi command line arguments? It says go to your comfy directory and put arguments like "main.py --force-fp16" but I get errors. Has anyone tried something similar? (Ubuntu user btw)

echo kite
#

anyone know some good generalist models that can handle weird unorthodox shitpost prompts? a lot of the popular models seem to be mainly geared towards portraits

stray bane
#

Sorry, I'm using paperspace, but no matter how many times I try, /usr/local/lib/python3.11/dist-packages/IPython/core/magics/osm.py:417: UserWarning: using dhist requires you to install the pickleshare library.
self.shell.db['dhist'] = compress_dhist(dhist)[-100:] appears and it cannot start properly. How can I solve it?

ashen acorn
#

Hey guys i’m new to the ai space think the potential is endless!

pale latch
#

I dont think AMD legal department know what they're doing

fervent thunder
#

can 5 ppl join please

#

dm me for link

#

i need 5 ppl

fresh stone
pale latch
mystic siren
#

hello I was customizing my web UI automatic1111 and I would like to know how to change the icons of the buttons in the ui like the ones of send to image to image,inpaint,etc, anyone knows how or where could I find more info about this?

fervent thunder
#

if you want to customise it you are probably better off just switching to comfy or diffusers

ashen acorn
#

hey guys is anyone familiar with canva?

floral umbra
#

How do people usually run flux? Just generate with 24GB cards at 512x512? Or everyone renting high memory cards from cloud like runpod, or do tiled generations?

loud solar
fervent thunder
#

do you guys know any models better then sd1.5 and can rn on 6gb vram ?

#

the only model i found is kolors with 5 gb size

floral umbra
loud solar
#

And I am using the biggest things .... it works here ...

native heart
#

Anyone know if ReForge is better than Forge?

loud solar
native heart
loud solar
native heart
#

do you think its worth try ReForge?

loud solar
#

Sure!

native heart
#

alright

floral umbra
loud solar
native heart
loud solar
native heart
#

Stock Forge means the old forge that i using?

loud solar
#

I think so

native heart
#

idk why i asking it

#

can just download the zip right?

#

dont need do the Git Stuff

loud solar
#

Didn't try it that way ... but if they offer it ....

native heart
#

there any advantage using git?

#

πŸ™ƒ

loud solar
#

Ask god .... I have no idea πŸ™‚

native heart
#

forget it just will download the zip

#

πŸ‘ thx

loud solar
#

Have a creative time!

floral umbra
loud solar
loud solar
drowsy robin
#

me on arc a770 be like

drowsy robin
fervent thunder
#

also you can put t5 on cpu if you want

wanton lotus
#

do you guys think were gonna get a anime flux model better than pony

abstract quarry
#

comfyui only supports fp8 quantization, but that's already enough to bring the model down to 13gb or so

wanton lotus
floral umbra
wanton lotus
#

im really upset theres lots of stuff comparing comfyui with a1111 but not with comfyui vs forge

native heart
#

what vae should i use if not using SDXL model?

native heart
#

will try now but its gonna take some time to rlly tell its worth or not

#

in my case

river sparrow
white parrot
#

helllo guys, i need help

#

i tried nyan mix with same stuffs liek author to test but smh it became liek this
how to fixx that

desert dagger
white parrot
#

i didn't see any required vae in descripsion

#

that's may my critical mistake

#

so what vae need to clear the blue stuffs
everything seem smooth till it come out

desert dagger
white parrot
#

nyan mix
also what is pruned model, do i need to mix it together, i heard that pruned just liek original but smoler stuffs

fathom gate
#

what is the image size of SD? dalle at bing has 1024x1024

desert dagger
desert dagger
white parrot
#

civiai sir

fathom gate
#

hm.. the newest one?

desert dagger
white parrot
#

i may using sd 1.5, idk

desert dagger
white parrot
#

sr im too new for this

#

let me dm you for details sir

desert dagger
mystic siren
#

Guys I think I found a way to make the image generations on automatic1111 go faster and It's something very simple,i don't know If everyone know It but I changed the live preview from NN to taesd which is the same settings I use on comfy , and now my generations match the times of comfy ui

fervent thunder
#

taesd is awesome yeah

#

you can also just use it as your VAE whilst testing

#

its very fast

#

at high res like 4k or above, the VAE decode times of a batch can be pretty long, like multiple minutes

mystic siren
#

thanks, I apreciate the info

white parrot
#

what positive and negative prompt i should use on anime stuffs everyone?

fervent thunder
#

no negative

#

positive: "an anime picture of"

#

and then pick a good model

white parrot
#

shoul i prompt by keyword or?

fervent thunder
#

these days its better to just write a whole paragraph description in plain text

#

especially for the current new generation of models like SD3, Flux and Auraflow

#

these can just take in text

#

I still do keyword prompts for SD 1.5

#

SDXL can go both ways

white parrot
fervent thunder
#

yeah sure

barren sentinel
#

hoi folks

narrow marsh
#

hello teachers! I want to turn comfyUI into other languages ,how I can change that ?please

wind verge
#

what python version does comfyui use?

silent glade
#

How do I use img2img? I ask it to do something but it just seems to create a new thing according to the prompt...

#

o wait it does modify the image

#

hmmm

#

lol

fervent thunder
#

its a matter of when to put noise injection, and what the type of noise is, followed by the choice of sigmas

silent glade
#

πŸ’€

silent glade
fervent thunder
#

if you are using something like A1111 then
the default is likely something like adding white/grey/brown noise once at the start
and then a slider called something like "denoise %" will set the sigmas

#

so if you slide "denoise %" to like 0.2 it will change it less

#

its confusing because every app names this stuff different things

silent glade
#

"denoising strength"?

fervent thunder
#

yeah that will be it

silent glade
#

so if its high then what should happen

fervent thunder
#

so in general 0.2 will change only a little

#

0.8 will change a lot

#

what this is doing is changing the sigmas

#

a second way to alter how much the image changes is to alter the noise injection but A1111 doesn't let you do this

silent glade
#

hmm its just not good

#

it doesnt just change the img a bit it changes it all

river sparrow
#

do take into account that inpainting is not perfect, it needs a certain amount of trail and error

silent glade
#

mhmmm

river sparrow
#

this is true even with the most perfect settings and prompt

silent glade
#

what does inpaint do

#

πŸ€”

#

ugh i cant send pics lol

river sparrow
#

uh, i thought you would have read the basics ...

silent glade
#

xDD

fervent thunder
#

I mean I agree that A1111 img2img isn't good
I would recommend learning noise injection workflow in ComfyUI
but that might take like a month to learn from scratch

silent glade
#

damn

fervent thunder
#

a lot of people seem to like it

#

so A1111 does its job

river sparrow
#

I have some kind of masking workflow in comfy that uses the Mask By Text node, it can auto mask things like eyes and faces (which makes it act like Adetailer). Or it can replace the whole sky or background. It does need 3 x a sampler: first the basic image, then the inpainted rough image, last a refiner pass to smooth things out.

fervent thunder
#

that sounds great yeah

river sparrow
#

still its not fully stable, needs some trails just like "normal" inpainting

#

Personally I rather have a more simple workflow, just trying prompt-to-image

fervent thunder
#

I'm a big fan of using a depth map to add noise

#

so you can get it to either do subject or background

river sparrow
#

The Mask By Text node needs to recognize something. So it works nice for example on "red dress" if the rest of the image is green or white.

#

Does work on "sky", does NOT seem to work on things like "full background, beach sand and sky"

#

Ill post an image with the comfy workflow included in the general-with-images chat

fervent thunder
#

it mostly comes down to whether its worth fine tuning Yolo

#

its quite an easy fine tune to do but I am a bit lazy

#

Yolo knows some things but not others

civic cradle
#

can someone cook me an image of something real quick?

#

if u have control net

fervent thunder
#

ok

civic cradle
#

freddy fazbear doing the gungum style pose

#

blue simple background too

#

cuz

#

fortnite

fervent thunder
#

sounds okay

#

what sort of control net would you like

civic cradle
#

damn i cant send images

#

open pose

fervent thunder
#

in the channel general-with-images you can send image

civic cradle
#

idk if u can delete bones but if u can

#

remove the head bones

#

i think its gonna make the head too small for a fortnite skin

fervent thunder
#

not sure if the model will know fazbear but maybe I can get a lora

silent glade
#

i wonder whats a good mix of sampling methods and upscalers to reach good realism

#

if thats how it works

silent glade
fervent thunder
fervent thunder
#

if you can use tricks to generate at the higher res then that is best

civic cradle
civic cradle
fervent thunder
#

oh I can't do pony so could you link the 1.5 lora?
I will do a refiner with realvis or SD3 to make it looks nicer than 1.5

silent glade
#

what abt sampling method tho

#

currently going thru them

#

lol

fervent thunder
#

there's a normal set of samplers and then a fancy set
out of the normal samplers, DPM++ 2M, DPM++ 2M SDE, DPM++ 3M SDE are all good
out of the fancy samplers, 4th order explicit adams, 4th order implicit adams or 2nd order gauss–legendre are all good choices

civic cradle
fervent thunder
#

sd3 has issues yeah

civic cradle
#

i have been loyal to 1.5

silent glade
civic cradle
#

like how people still stick to iphone 8s till this day

fervent thunder
#

oh yeah the iphone 8 is still ok

silent glade
#

is there really a difference between dpm++ 2m karras and dpm++ 2m sde karras

#

πŸ€”

fervent thunder
#

yeah these are actually extremely different

#

because one is stochastic and one is not

#

many samplers are similar to each other but these two are not similar

silent glade
#

im wondering in terms of realism

#

hmm idk really

fervent thunder
#

depends on how many steps you are willing to use

silent glade
#

thats right

#

i just wanna choose sampling methods that're a bit more realistic than the default one

#

but maybe ill focus on other things

fervent thunder
#

no I mean

#

the choice of sampler depends

#

on the number of steps you are willing to use

#

just as a rough example:
12 steps - unipc
20 steps - dpm++ 2m
60 steps - dpm++ 3m sde
1000 steps - 2nd order gauss–legendre

silent glade
#

does it auto-switch?

#

wdym

#

or thats a list of what fits the best

fervent thunder
#

thats a possible list of what might be good yes
but you could easily set this up to auto switch in ComfyUI or Diffusers yeah

silent glade
#

damnnnnn 1k steps is crazyy

fervent thunder
#

1k steps is actually what this technology started as

silent glade
#

wow

fervent thunder
#

which is why a lot of papers have charts with specifically 1k steps

silent glade
#

i keep forgetting i cant send pics

# 
# Configuration scale mapping with defaults
    cfg_scale_map = {
        "creative": 7,
        "regular": 15,
        "specific": 30
    }
    cfg_scale = cfg_scale_map[cfg]

    # Detail level mapping with defaults
    details_map = {
        "low": 20,
        "medium": 69,
        "high": 135
    }
    steps = details_map[details]

    # Dimensions mapping with defaults
    dimensions_map = {
        "wide": (1000, 512),
        "regular": (512, 512),
        "tall": (512, 1000)
    }
    width, height = dimensions_map[dimensions]```
#

These are my settings

#

Maybe I'll do matching sampling methods

#

Which would you recommend for 20, 69 and 135?

fervent thunder
#

the range that I gave before is fine

#

dpm++ 2m for 20
dpm++ 3m sde for 69 and 135

silent glade
#

ok so i set that thru the code i assume

#

yea

#

where can i see docs for that tho

fervent thunder
#

I actually am not sure if you mentioned what tool you are using?

silent glade
#

by which tool u mean which webui

#

?

fervent thunder
#

yeah

silent glade
#

automatic1111

fervent thunder
#

ah okay
I don't know this one, but maybe someone else can help

silent glade
#

alright thanks

fervent thunder
#

as far as I know A1111 does support extensions so there will probably be a way of doing it

silent glade
#

doesnt support at all?

#

πŸ€”

fervent thunder
#

I used A1111 a little bit a year ago
but I didn't click into every single menu

#

but I don't think it can do this by default

silent glade
#

im using an extension on it but maybe u mean something different

fervent thunder
#

maybe they call it plugin

silent glade
#

i mean theres an extensions folder

fervent thunder
#

basically whatever they call adding additional functions via code

#

yeah anything like plugin or extensions

#

implies that they have set it up to be able to take modifications

silent glade
#

hmmmm

#

im not sure frfr

regal snow
#

I have a 4070ti (12gb vram). With Flux dev model, generating images in 35 steps in ~40 seconds

wet grotto
#

controlnet union is only for comfy or works on a11 too?

fervent thunder
#

in theory it could work fine with A1111

#

as it takes many different inputs

#

but A1111 codebase is a wild west so you never know

wet grotto
fervent thunder
#

i have been using sd1.5 with no upscalers XD

fervent thunder
winged wasp
#

hey

fallow pine
#

anybody knows if flux model is supported on forge ui yet or no?

pale latch
#

yup i'm using it right now. the newest forge release is a big one

#

has nf4 support

civic mirage
#

hey everyone, i just started using stable diffusion and i dont know how to add flux to it

#

is someone available for a quick google meet to explain a few things

warm junco
fervent thunder
#

its in Diffusers also if you are ok with that

fervent thunder
warm junco
fervent thunder
#

yeah , i think its not supported in 2060

#
  • i have 16gb of rram.
silent glade
#

Anyone familiar with AUTOMATIC1111 and can help me? πŸ™‚ (@ me)

fervent thunder
#

maybe i can go head

wraith lintel
#

hey guys

#

im in trouble with control net

#

any one help me

fervent thunder
#

ok

wraith lintel
#

where am i wrong

#

why never use pose

pale latch
#

you're using an adapter for sd14 on an sdxl model

#

@wraith lintel

fervent thunder
#

its crazy how sdxl dont have final animate diff motion module.

#

yeah

fervent thunder
wraith lintel
#

yes i hate w11

#

w11 not stabil as w10

fervent thunder
#

should i downgrade ? cuz i think windows 11 takes more ram and vram compared to 10

wraith lintel
#

i was upgrade win 11

#

and i hated

#

nothing work like i want

#

actualy im going to instal ubuntu

#

linΓΌx is the best

fervent thunder
#

linux best yeah

#

tbh , everything works out of the box in win11 i just think its overkill for my 16gb ram an 6gb vram when using sd ai

fallow veldt
#

I use ubuntu for AI stuff, you know all is fully optimized there

fervent thunder
fallow veldt
#

as generating with animated diff with 512x512 took a lot

wraith lintel
#

i try it with run diffision web host

fervent thunder
#

you need the right models

fervent thunder
fervent thunder
#

you can use SDXL at lower resolutions now with tools like Resadapter

#

will that lower vram and make it faster

#

faster yes

#

but I don't know about VRAM

#

I use 40-100GB servers so I just never think about VRAM

#

I need to learn more about VRAM usage really

fervent thunder
#

cloud yeah

#

I totally understand people who use it locally

#

but i dont rly understand why cloud users go for low vram

#

on Vast.ai low and high vram cost similarly often

pale latch
#

win 11 is fine. i've used all the windows and this one is no worse than any other

fervent thunder
#

for sd ai in gpu there is vram and FP16 (half)
12.90 TFLOPS (2:1)
FP32 (float)
6.451 TFLOPS thats my rtx 2060 spec for fp16.32

#

if u buy a faster fp 16.32 gpu but low vram u r wasting ur money.

#

more vram almost always wins yeah

#

anything to avoid neural network layers being offset to cpu

#

yeah cuz the entire model must be loaded into vram ... why is that , why not load it to ram

pale latch
fervent thunder
#

thanks I like videos like this

wanton lotus
#

is it possible to finetune flux like pony diffusion with sdxl

fervent thunder
#

possibly but not certainly

pale latch
#

potentially. just liek a lot of research had to be done for sdxl before pony was made, people have to play with flux more before we find out. flux was distilled from the pro version , as it was training, it uses a secret set of weights to inform the model we got. we don't have those weights and they may be needed for refinement. but there might also be other ways to do something too

fervent thunder
#

I recently went through the papers on distilled models
their trainability is not really certain yet

#

distilled models are VERY hyped in the papers at the moment so I expect more distilled models are coming

#

I kinda dislike them cos you lose some control but I can see the appeal

pale latch
#

thats why i've always avoided lightning or turbo models. though i'm begrudgingly loving it now since flux is just such an exploration

#

they seem to collapse a lot of classes imo, which is seen in a lot of ways with flux. default blur on photographs, default smoothiness to all paintings, thigns like that

fervent thunder
#

yeah I avoided turbo and lightning for the same reason

#

they can't handle much perturbation at all

#

not even really a single noise injection

#

and I like image variety

abstract quarry
#

I'm currently training Flux

#

from all I can say it responds better to training than SDXL, even though it is distilled

#

but I just trained on my face yet. Someone else trained on a specific anime. Both worked great, but of course we have to check more art styles and stuff

desert dagger
#

in case you don't have them, Kajai has released a number of loras for flux

last bay
#

can anyone simply explain the lora weightts

trail lion
#

I thought about trying to get flux working this weekend, but decided I'd rather wait for something beyond a base model

last bay
#

???

desert dagger
#

wrong discord.

modern oak
#

???

lone dust
#

I downloaded the Files and tried to click "Load" in comfy, then select the config.json

#

nothing opened

#

all the nodes and connections remained the same as my past setting. Didn't load

fervent thunder
#

this is what config.json contains:

  "_class_name": "ControlNetModel",
  "_diffusers_version": "0.27.2",
  "act_fn": "silu",
  "addition_embed_type": "text_time",
  "addition_embed_type_num_heads": 64,
  "addition_time_embed_dim": 256,
  "attention_head_dim": [
    5,
    10,
    20
  ],
  "block_out_channels": [
    320,
    640,
    1280
  ],
  "class_embed_type": null,
  "conditioning_channels": 3,
  "conditioning_embedding_out_channels": [
    16,
    32,
    96,
    256
  ],
  "controlnet_conditioning_channel_order": "rgb",
  "cross_attention_dim": 2048,
  "down_block_types": [
    "DownBlock2D",
    "CrossAttnDownBlock2D",
    "CrossAttnDownBlock2D"
  ],
  "downsample_padding": 1,
  "encoder_hid_dim": null,
  "encoder_hid_dim_type": null,
  "flip_sin_to_cos": true,
  "freq_shift": 0,
  "global_pool_conditions": false,
  "in_channels": 4,
  "layers_per_block": 2,
  "mid_block_scale_factor": 1,
  "mid_block_type": "UNetMidBlock2DCrossAttn",
  "norm_eps": 1e-05,
  "norm_num_groups": 32,
  "num_attention_heads": null,
  "num_class_embeds": null,
  "only_cross_attention": false,
  "projection_class_embeddings_input_dim": 2816,
  "resnet_time_scale_shift": "default",
  "transformer_layers_per_block": [
    1,
    2,
    10
  ],
  "upcast_attention": null,
  "use_linear_projection": true
}```
#

this isn't a setup for comfy with nodes and connections etc

lone dust
fervent thunder
#

oh yeah this is Gradio

#

you recognise the Gradio boxes and buttons after a while

brisk urchin
#

How i can train a lora? Anyone have a good tutorial?

honest stirrup
fervent thunder
#

LOL

#

you are spelling volcano wrong

honest stirrup
#

Yeah xD, vocalno and got a Japan's Prime minister.

fervent thunder
#

the reason is that with transformers, if a word is misspelled then sometimes the attention heads give it a really low attention score

#

so it kinda fades into the background

#

in the same way that words like "the" and "of" do

#

your prompt is basically gonna come out as something like "wide angle photograph"

#

after the self-attention calculation

remote cosmos
#

anyone here using flux on forge?

vestal dew
#

just read about it...trying to get live portrait to work

brisk urchin
#

existe una opcion para borrar el cache de koyah?

#

local

#

is that when I tried to create the description with the utilities section it no longer generates the description for the same images when I try to repeat the process.

oak latch
#

how do i fix very reflective skin?

fervent thunder
#

for the most part, choose a model that doesn't have it

drowsy mica
#

Yibin City, Sichuan Province; Crossing the Minjiang River Grand Bridge; Suspension bridge; Golden Gate Bridge; Bridge tourism integration

honest stirrup
honest stirrup
oak latch
oak latch
honest stirrup
hollow marlin
#

HI

lone dust
young needle
red skiff
#

whats the current best most precise state of the art style cloning in sd? without training

fervent thunder
#

IP adapter mad scientist node

#

in comfy ui

#

this video shows it https://www.youtube.com/watch?v=ewKM7uCRPUg

fervent thunder
#

requires some experiments cos its not well known what each block does

#

but this will outperform methods that don't target individual blocks

cold yoke
#

A1111 is not run on intel?

#

"Found no NVIDIA driver on your system." how choose gpu?

#

or is just not work on intel

warm junco
#

Iris XE and Intel Arc can work

#

Everything else runs on CPU mode only

cold yoke
#

why is it saying "found no nividia driver"??

warm junco
cold yoke
#

Wow holy mongus pickles it seems to be workinf after a couple of reinstalations

#

Tha k you omg

#

How to find more webuis that can run xe? Or how to know if webui can run xe

fervent thunder
#

looking for someone that can match my freaky😘

cold yoke
pseudo pumice
#

yo

reef dew
#

hi is there like some trick to using SD automatic111?

#

i'm been generating images for the past few days and only maybe afew of them are decent

#

tried different models and loras and values

#

prompts

#

but i just can't get anything constant

kind thorn
#

Use keywords specific to the Lora if applicable
Also make sure you're using good negative prompts and an appropriate sampling method for the model

reef dew
#

hmm how do i know what are good negative prompts?

#

i tried taking some from galleries aimed at anime style

#

but they don't always seem to work

#

WIth Midjourney i could just use prompts and it'd produce something really good

sterile raven
#

Does Flux have a Discord

#

?

reef dew
#

but SD seems like a lottery πŸ˜„

fervent thunder
#

whens stable diffusion gonna be able to generate 3d models? πŸ™

storm shard
river sparrow
placid citrus
#

question HI, is there a way to change the ui in stable diff , to view past generation and bring the prompts back ect ect kinda like civtai?

rain crest
#

Do periods do anything or have any use in A1111 prompts?

placid citrus
#

yeah

#

kinda like with civtai where just click hit remix and stuff

#

and plasters the prompt back in the prompt with the negatives and such

river sparrow
#

in comfyui you can use a file save node that included the workflow in the image, so you only have to drag-n-drop the image onto the comfy canvas to load it

#

it stored all settings, whole workflow with all nodes

river sparrow
#

then replace with . or other symbols

rain crest
#

ohh ok

river sparrow
#

yeah, sorry

#

same test method should aply

#

also nice to see the difference between , or .

rain crest
#

Can you still like put it into words what it does

river sparrow
#

same as in language, it groups and separates

#

everything between , is a subgroup, all between . is a bigger (stronger) group

rain crest
#

I see

river sparrow
#

but, there is always some prompt bleeding since models cannot pull everything correctly out of the text. though sd3 and flux are much better in that regards

narrow kernel
#

Report the loser spamming swastikas with the exclamation warning mark

brazen dome
#

Can I generate an image of a "color palette" with Stable Diffusion? but one that not only has color but also gradients. Basically 5 long rectangle shapes above each other that are either filled with a color, a gradient and could have a colored stroke around the shape. I want to use them as visual style reference for my motion design projects.

#

since most people I see on the internet use SD for generating paintings / illustrations / photos. I am curious if it can be properly used for "graphic design"

willow shore
#

hi everyone

#

how r u doin

real pewter
#

hi

#

How can I prevent my Google Collab runtime from not stopping?

#

I'm trying to use Stable diffusion in colab

simple heron
#

anyone use Reactor with Forge or A1111? the dropdown on my face models option has disappeared. anyone know why?

quartz siren
# real pewter How can I prevent my Google Collab runtime from not stopping?

not sure if its really possible, colab will automatically stop runtime after some time it calculates or if you are not executing new code for a long time.
I would recommend kaggle instead, you get 2 t4's and a better cpu and you get 12hrs runtime. You just need to be sure to be active for 40mins at least.

fervent thunder
#

when trying to install stable fast 3d, hugging face asks for a token, in my profile there is no tokens, however I have been granted access to the repo. Where do I find my token?

abstract quarry
#

you generate one in your profile settings menu

#

settings -> access tokens -> create new token

#

select READ permissions

fervent thunder
#

now after installing 2345908235097235 more things, I can't get python to find my CUDA installation

#

"OSError: CUDA_HOME environment variable is not set. Please set it to your CUDA install root."

#

i dont even know what to do here

abstract quarry
#

cuda should get installed when you install pytorch

fervent thunder
#

i got it working. I really appreciate you helping me

stark halo
#

hello

frail sonnet
#

How often does google collab actualy give A100s? I want to make a lora πŸ™‚ However, everything I've read says barely ever 😦

left hamlet
#

hello

left hamlet
#

im downloading an AI model

#

and

#

im getting issues. copilot is actually helping very well

proven temple
#

can anyone help pastebin or upload a caption .txt with Keep and Exclude tags?

#

I want to know how to format for it

low moon
#

So flux wont work on the latest forge

#

object type zero or some nonsense

#

--no-half-vae --no-half --disable-nan-check <--- tried addign these - pointless

#

any ideas?

low moon
#

total wipe new reinstall worked...

#

the new install is 7 gigs the old one was 40, i didnt have any models in there

#

what the hell happened

#

these things tend to get out of control...

sly sinew
#

/η”ŸζˆηΎŽε₯³

silver zealot
#

how can i train my face with flux ?

turbid shale
#

Hey everyone

#

Anyone familiar with StreamDiffusion?

#

been having an issue with installation... as my background in coding is 0

#

not 0, but im not big on certain concepts

#

anyways... i did install the whole thing from Github, but im lost on how to launch it on TouchDesigner

#

like how do i get it show up?

proven temple
#

is there a prompt I could use to stop SD from generating a person?

#

things like nobody, no one, etc wasn't working. person or people on negative prompt didnt stop it too

white parrot
#

hi everyone

#

how can i upscal then sprerate 2 image

#

do i need to do it in pts lol

sick flint
proven temple
worn rune
#

hello! every one , friends

spare root
#

hey guys sorry total noob here but could anyone point me in the right direction for this?

not sure if this is the right discord but want to learn how to put my face on stuff using flux realism.

apologies if dumb question but very new to all this.

tried fal.ai but didn't give the best results (looked a bit cartoonish / unrealistic)

bleak matrix
#

Good morning,everyone! How are we all today?

fervent thunder
#

good

worn rune
#

good evenings,every one,

#

Which section should I choose for making videos? Can you recommend it?

fervent thunder
#

are you willing to rent a cloud server or does it have to be local

tawny halo
#

hey guys whats up how can i run a flask website on colab and see the output

old blade
#

I have a question regarding training LORAs, I know there needs to be an associated text embedding file, but I can not find a quality example of what it should look like. Any recommendations? "filename; name of scene; weight; description" is what I have deduced as the format.

fervent thunder
unkempt hatch
#

Is there a list of online services using flux somewhere?

brittle oyster
#

curious if anyone has an tricks on how I could get SD/Flux to add my company logo into it's images? I Presume this would have to be done using ControlNet which is still being worked on.

low moon
#

Some hard noise injection? image to image thing

brittle oyster
#

I tried img2img but the results don't seem to include the reference image at all, basically just follows the prompt... I have denoising up to 0.95 and it doesn't seem to affect it.

hasty hornet
brittle oyster
#

ah ok, I'll try again with 50% to see if anything works better.

river sparrow
old blade
#

Yeah unless your logo is 50% of the image πŸ™‚

hasty hornet
#

even if it is, that's still very high

#

doesn't really matter how big area that is tbh

wintry kelp
abstract quarry
#

with sdxl it was impossible to train on logos

#

flux can do this however

#

so just prepare a few files, can be 512x512 and train a lora on them

#

make sure to train all layers including the feedforward ones

#

flux not only learns the logo but it can also place the logo on shirts or buildings accurately

boreal notch
#

Hey guys, does anybody knows where I can get good information on ComfyUI's fundamentals?

teal shard
#

Is it better to have more CUDA Cores or RAM to generate images with SD? Is it better to use the GTX 40 or the A4000?

little lark
#

Hello, I'm looking to create a logo for a motorhome sales company, and I'm a bit lost. I think I've found the right Lora, but not the right stable diffusion model. And maybe not the right prompts either, as the results I'm getting are all over the place, far from what is shown on the Lora pages. Could you help me?

pseudo pumice
#

gn

wispy osprey
#

Hey guys, what is up?

novel raft
#

so hoe exacly would i run it on AMD?

warm junco
real pewter
quartz siren
novel raft
pale latch
# novel raft so hoe exacly would i run it on AMD?

easier to buy a new nvidia card. there was a man who made a library called ZLUDA which allowed AMD users to use CUDA. Earlier this week, AMD got ahold of him and told him to kaibosh that library and to kill all support for it.

He's still making it but it no longer works on AMD hardware. I think he's making it into a way to run CUDA code on the cpu now

#

situations like taht demonstrate AMD isn't keen to expand the ML capabilities of their consumer cards

#

when the manufacturer themselves are making it hard, find a new brand.

novel raft
#

I will have to work with AMD

#

Also I do not know anything about CUDA

pale latch
#

Yeah. AMD doesn't have your back. They screwed over every AMD customer that was using ZLuda for CUDA compatibilty this week

novel raft
#

I don't know what either of those are tbh

wet grotto
#

How can I generate more then one controlnet in one shoot? script?

desert dagger
desert dagger
wet grotto
#

I will show

pale latch
#

on comfyui i think you chain them or you use a node that inputs a few of them

#

"stacker" or something

#

webui has multiple nets built into the extension

fervent thunder
#

you could wire up like 40 nodes and then collapse it down into one node

#

that's the easiest way to get a single fancy node

pale latch
#

is that why i can't figure out many workflows? geeze i really gotta hit the books again

cedar salmon
#

right click and convert to nodes

pale latch
fervent thunder
#

a lot of workflows are just completely nonesensical

#

same with a lot of custom nodes

#

and there's not really a "thing" to work out

#

they are just perturbing or ablating the model

pale latch
#

yeah i don't mine too deep into figureing somethings out, but sometimes i wonder wtf is going on and try to follow at a glance

fervent thunder
#

there are some other bad trends
-using a ton of layout nodes like busses, to the point where it just made it worse
-chaining like 4 upscalers at the end (why does the image have to be 16k or 32k?)
-using a custom node pack which is normally named after themself, where its barely different from normal nodes

pale latch
fervent thunder
#

in some ways this is what invoke is

#

a more curated comfy

pale latch
#

yeah. invoke is more like a complete prudct and comfyui is mechano

trail lion
#

this whole fascination over flux, after 1 day using it, ok it's deserved

#

having said that, the community aspect is what made SD great, that still remains to be seen

fluid atlas
#

hi guys

fervent thunder
trail lion
#

locally, true, but it's so easy to get with cloud resources

#

seriously if the only barrier is like $1/hr, if that's a barrier than I guess it's not really that important

fervent thunder
#

seems like the vast majority have a resistance to cloud

#

a ton of SD 1.5 loras were trained by people with just 6GB VRAM

#

I don't think this will be repeatable for modern models

pearl minnow
vestal zealot
#

is ther eany ai that can make short videos or sum

#

like sora

civic cradle
#

But I do have a question

#

If it takes me 30 mins to generate a single 512x512 image on sd 1.5

#

How long do u think it will take me per image if I want to install flux?

#

F1 schnell to be exact

fervent thunder
#

not sure

sudden ruin
heady vector
#

Hi all, i have a very personal request. Yesterday our neigbour died and there is almost no picture available of him. Is there any chance to create a frontal image from a sideview? Maybe with openpose? Is anyone able to help me with this. Need it for the funeral.

ornate ridge
coral condor
#

i applied for a personal license and when i submitted my contacts, i was told that i should get an email with next steps, but I still don't get it. How long to wait for that email?

fervent thunder
# heady vector Hi all, i have a very personal request. Yesterday our neigbour died and there is...

Instant ID for the IP adapter side
Open pose would be fine for control nets but could also add canny and depth at low strength and stopping part of the way through the sigmas
Training even a one image lora would help a lot
If you are able to get the above working a little bit then you can keep doing generations until you collect a few samples
and then train a second Lora that adds these as artificial training data

heady vector
trail lion
#

how many images you're creating is irrelevant, how many you have to work with is the issue. that approach is likely the best one

nocturne pewter
#

A little while ago some kind of universal controlnet model was released, it covered most of the controlnet types within a single model. Can someone remember what it was called?

fervent thunder
nocturne pewter
#

Thanks! @fervent thunder

fervent thunder
#

its got a node to set the type

onyx oyster
#

I just swtiched from automatic1111 to comfyui. The amount of control is mindboggling. Though i'm still learning the new UI it seems amazing

fervent thunder
#

yeah its amazing

onyx oyster
#

Also, for some reason it's faster. I don't understand enough to know why. THough I haven't figured out inpainting yet. Any resources you guys got on inpainting in comfy?

lilac tusk
#

Does anyone recommend learning resources: books, blogs, videos, papers, to help me understand generative image tools better in a comprehensive way? I have a bit of matrix algebra and machine learning background, I've trained CNNs in tensorflow and other classifiers from scratch, worked with Word2Vec, done some segmentation stuff etc, but I'm having trouble finding resources about diffusion models and practical techniques to make images that aren't either "how to do X in comfyui" with little explanation of concepts on the one hand, or deep scholarly articles that are extremely specialized and assume a lot of foundational knowledge that I want to have but don't have yet.

My goal is learn enough background to use comfyui with various models to make controlled images in a way I can fully understand what I'm doing and how the tools work and what parameters are important, and diagnose where things go wrong. Going for correct conceptual models more then mathematical rigor though not allergic to math if it's applicable to know for practical use. Thanks!

bleak matrix
#

Good morning, everyone! How are we all today?

honest stirrup
#

Hi Guys!
3090TI 64gb ram here,
recommend Pytouch and Cuda version?

honest stirrup
novel raft
neon citrus
#

What is the sentiment on SD3 and commercialized models versus open source?

bleak matrix
last bay
#

does anyone know what is SDXL 1.0 ?

#

how much its different from SD 1.5

desert dagger
desert dagger
last bay
last bay
#

SDXL def look better

quartz siren
# last bay SDXL def look better

yes it should(for the base models). Its unet is basically a lot larger(so better image quality) and has 2 text encoders for better prompt following compared to sd1.5 which has 1.

last bay
novel raft
desert dagger
# novel raft I have no idea

then it's not possible to tell you what the interface should look like. first you need to pick an interface you want to use

severe sundial
#

Hi everyone! New to SD. What are you guys seeing as the pros and cons of SD3? I see a lot of comps bt it and Flux... Ik SD3 isn't fully ready yet, but just overall for where things stand / the direction SD3 is trying to achieve

fervent thunder
#

I think SD3 is better

#

cos Flux is slightly overtrained on a midjourney/instagram/aesthetics look

old blade
#

have you tried applying LoRAs to it? You might be quite surprised. I do agree the instagram look is strong.

low moon
#

How tremendous are we on a scale 1 - 10?

bleak matrix
sterile raven
# severe sundial Hi everyone! New to SD. What are you guys seeing as the pros and cons of SD3? I ...

Stable Diffusion's fate is sealed. Don't waste your time with it. With the adoption of Flux on X and the communities rapid advancment in Flux its obvious Flux is the future. Less censorship, better quality, legacy tooling like comfyui still work with it, it can inpaint, you can train loras (which people still cant figure out for sd3, which also hasnt been released, we just have the shitty broken version you see all the memes about), its open source and can run on commercial hardware locally. The list goes on

low moon
#

Basically yeah but they cna always redeem themselves

fervent thunder
#

I think its important to not conflate a company with its models

low moon
#

hypemad happemad <--- without the original visionary/idea man mastermind coolness maskot it will be hard tho. the unifying vision and awesomeness is gone

#

replaced by human idiots who are less human than the AI they train

#

all they see is money and money is dead

fervent thunder
#

they got a lot of investment recently and still have a lot of expertise
they will stick around long enough to train another model

low moon
#

SD3.5

#

they usually excell at uneven numbers

fervent thunder
#

funnily enough it will be uneven

#

its SD 3.1

low moon
#

yeah basically

#

I bet theyre gatherign so many images of women in grass for training data

#

"this time we have to make sure we get great tremendous women laying in grass renditions"

fervent thunder
#

on one of the websites that compares models

#

it has a whole category for that now

low moon
#

"woman laying in grass" is the new basic test benchmark.

fervent thunder
#

reddit was very unfair because
the best SDXL models, like Zavy or Hello
also struggle with it
they just don't at 1024*1024

#

but if you are pushing them outside their comfort zone of resolution they will struggle

#

its clearly a prompt that is hard for all models

#

noise injection into self attention to slightly perturb the model will also make it hard

#

but very stable prompts still work

rain crest
#

Anyone know why increasing Batch count to over 1 gives me green & crazy images, on A1111

wispy osprey
#

Guys, do you know somewhere which people can deploy their SD models for free?

last bay
wispy osprey
pale latch
#

hugging face

novel raft
pale latch
#

civit is the most used. shakker is civit direct competitor. hugging face is just a place to store models for resaerch and academic purposes with seemingly no commercial ulterior motives

#

hugging face is like github and civit is like nexusmods

desert dagger
median swift
#

I've gotten some requests to convert some of my artwork to NFT and getting offered what I find is pretty insane amounts for them.
Are these kind of solicitors actually real or are they always scam? Would converting to NFT hurt me in any way and what would be a safe way to proceed?

pale latch
#

real businesses can prove they're legit in basically ever situation

desert dagger
compact swan
#

hi guys, what's the best way to access stable diffusion on my phone? I need something with generous generation credits or ad-supported, just nothing I have to pay for please

also, any model is fine but SD1.5 would be best because that's what I have locally on my PC

#

I don't have access to my PC right now but I feel like generating some stuff

trail lion
# lilac tusk Does anyone recommend learning resources: books, blogs, videos, papers, to help ...

that's a lot, ok...well things are going to be divided as far as resources go. there's no one-stop shop for all that, and doubtfully ever will be. This stuff is just evolving too fast. how to use the tool, well you're probably going to find that youtube is a decent help here for walk-through type stuff. how to prompt... what most of us do is look at the recommendations of the model authors, look at some existing style references (such as can be found in the a1111 style extensions for example. there are also some great tips for controlling lighting and angles that can be found in a google search. some of it is trial and error too I'm afraid

sage reef
dapper oxide
#

whats the best free image to video right now?

fervent thunder
#

any way to "generate forever" but with some added downtime to cooldown?

pale latch
#

it'll slow down when it needs to

novel raft
#

So how exactly do the different UI access models and Loras? Do I have to download all of them to my PC?

#

Man all of this is difficult

fervent thunder
#

we ever getting a flux channel here?

#

or does anyone know of a flux discord plz dm me the invite

verbal osprey
#

send me the invite too, thanks

frail sonnet
fervent thunder
#

me finks they mad cuz flux is made by SD opposition

#

me finks this sub finna become obsolete

#

me finks 'tis a good time to start a flux discord 😏

frail sonnet
#

Someone locally asked about computer help for their windows7 system.......
I don't feel like mine are so obsolete now

fervent thunder
#

yuck. I cant stand slow computers

#

had to help my grandma deworm her snail PC

#

boy do I feel more thankful for my 13th gen than ever

torn jewel
#

Anyone interested in partnering to build a not too difficult resume coding project?

nova lily
#

Isnt there some sort extension or feature that swaps out/rotates a prompt every time?

surreal frigate
pale latch
nova lily
dawn oasis
#

Why is Forger so much better than regular Auto1111?

#

Renders much faster, results are much better

fervent thunder
#

Forge is made by the guy who made Control Net and Omost

#

its just well made

#

I've been going through his diffusers code for Omost, its really interesting

#

he writes in a very understandable way

dawn oasis
#

It legit fixed all my issues with rendering

fervent thunder
#

I don't personally use it because I like either node based (comfy/invoke) or CLI (diffusers/pytorch/JAX)
but out of the A1111 clones I think Forge looks very good

pale latch
fervent thunder
#

ah yeah

#

I can never remember how many ooo's

pale latch
#

more the merrioooooor

fervent thunder
#

there's SDnext/Vladdiffusion too but I don't know about that one much

#

not sure if it is still competitive with Forge

dawn oasis
#

I don't even need to user ADetailer for the eyes anymore

#

Just for good measure though...

last bay
#

anyone got a generally used negative prompts to use for every image generation?

fervent thunder
#

yeah its best to never use a negative prompt

last bay
little lava
#

1

fervent thunder
timber lake
#

how to make images now

fervent thunder
#

do you have a graphics card?

warm junco
# last bay really?

No. Only SD3 and flux dont need a negative prompt.
For SDXL it depends on the model.
But for 1.5 and pony models negative prompts are essential

fervent thunder
#

oh Pony and 1.5 are different yeah

#

some 1.5 checkpoints need variants of "worst quality" in the negative or something because of the NovelAI origins

#

I've forgotten exactly what the strings were but it was something like that

river sparrow
#

worst quality, bad quality, etc

fervent thunder
#

yeah

#

but you only want to do that if the model specifically requires it

#

SD3, Flux, SDXL HelloWorld etc all look worse with that negative

#

as CS1o said though it depends on the model for SDXL as people have trained weird stuff in

last bay
fervent thunder
#

SDXL has the unfortunate thing where you kinda have to learn each checkpoint as if it was a separate model

#

cos some have been finetuned to work better with different CFG levels and aspect ratios etc

#

and a bunch of trigger words have been trained for the positive and the negative

last bay
#

yea i noticed that

fervent thunder
#

the other problem is they bake in a certain look to their checkpoint

#

so if you want a different look you have to change checkpoint

#

I wish they weren't mostly trained with these issues, but since we get them for free I am still grateful for all the free fine tunes

subtle owl
#

Hi all. Does anyone run a local install on a laptop?

#

I bought a new rog laptop with an Intel and a 4050rtx but all generation programs use the Intel card.

frail sonnet
fervent thunder
#

Cartoon, anime, drawing, watermark
I promise this will make the image worse on average
if its a model that doesn't specifically need it

warm junco
fervent thunder
#

I used to do negatives like that too

subtle owl
#

is there a preffered model/ui combo for laptops?

white parrot
#

how can i train a lora

sudden tiger
#

Hello, I don't understand how to get DF_Text_Box.
Anyone have a solution?

warm junco
warm junco
#

Online by using a service

white parrot
white parrot
warm junco