#💬｜general-chat

1 messages · Page 165 of 1

fervent thunder
#

Thats not not what i meant. That barely has any influence. I want to change style of image but keep it mostly the same as original.

#

that sounds a lot like IP adapter

#

and yes flux has one

plain raptor
#

press L1 to kiss

hard swift
#

yo

split wedge
#

hey everyone, does anyone use Streamdiffusion here?

craggy cape
#

Hey

hardy nexus
#

did sd3.1 come yet ?

fervent thunder
#

not sure its ever coming

craggy cape
#

Which ui should i start with? I see 4 on github and Im kinda confused.

#

I mean like automatic111 comfyui

#

what even is the difference

sudden ruin
craggy cape
#

Hmmm

#

Okay thanjs'

#

what is the difference beetween the zluda and directml?

sudden ruin
unborn hedge
#

hi guys, i just got sdxl to work somehow, was very confusing but i got it working now! on my local automatic1111

craggy cape
fervent thunder
#

legalities

#

from the GPU manufacturers

sudden ruin
craggy cape
#

dang

still glacier
ashen rune
jagged mason
jagged mason
sudden ruin
craggy cape
#

rocm5 or rocm6

split wedge
hybrid cobalt
#

When I launch auto1111 it says "Zluda works best with SD.Next. Please consider migrating." This isn't an issue, my auto has been working fine since we fixed those issues, I'm just curious as to how SD.Next runs better with Zluda.

fervent thunder
#

SD Next uses two back ends

#

one is based on A1111 and one is based on Diffusers

split wedge
#

question, is it possible to fuse lora with sd turbo?

fervent thunder
#

its difficult, sometimes

#

if you use turbo lora, that can work ok at least

split wedge
fervent thunder
#

there's a nice one for Leosam models

#

its on his civit

split wedge
#

I am trying to train a lora for my streamdiffusion setup running on sd-turbo

fervent thunder
#

streamdiffusion is an area I am not familiar with

split wedge
#

yeah I would have liked to use sdxl-turbo , but apparently it doesn't work with tensorrt, which is important for me since I need the frame rate to be high

#

but thanks for your help anyway

copper crystal
floral umbra
#

Is there a way to keep models in memory with flux? As for every change i do to prompts, it empties the vram, and has to load it all slowly back in

floral umbra
#

Unless they have worked out it's kinks to work on any RTX? Thonk

copper crystal
#

i know that file in particular wont. but it's demonstrating that it does.

floral umbra
#

Ah. I've used tensorrt a lot. Downside with them, you need a model for each resolution/resolution range, and that shit can stack up real quick

split wedge
copper crystal
hybrid cobalt
fervent thunder
#

ok so A1111 is one piece of software for making AI images using stable diffusion

#

A1111 also gets called WebUI sometimes

#

also because A1111 is so popular, sometimes people get confused and think it is the only GUI so you get people coming in and calling A1111 "stable diffusion" sometimes

#

anyway

#

Huggingface made another library called Diffusers

#

which is CLI (its code)

#

and one some GUIs use it as a backend such as Invoke

#

however what SD Next did is use both Diffusers and an A1111-style backend

#

so it has dual back end

#

there is also Comfy UI which has some other slightly different lineage, although its closer to A1111 than diffusers

#

as far as I can tell Comfy UI is mostly based on K-diffusion which is based on the Karras 2022 paper commonly referred to as EDM

#

which is the same paper where both the karras scheduler and the sampler known as "Heunpp2" come from

#

this long story is why you use a K-sampler node in comfy

glass lily
#

Thank you!

dark hawk
warm junco
glass lily
#

@warm junco It reset my stuff to default lol

pulsar cargo
#

Hi, I'm new to the server, my name is Christian.

warm junco
glass lily
#

I put prompts into "txt2img/Prompt/value":

#

and there's nothing

#

And it also reset my Adetailer

#

And my resolution 😂

#

I don't get it

warm junco
#

Hmm strange

#

Can't test rn

glass lily
#

This is way too weird

#

I edited some stuff in txt2img, txt2img adetailer, img2img, I think img2img adetailer as well and that's about it

#

Idk what could be the issue

white parrot
#

can i upscale mutil image at once, or one by one?

rancid patio
#

Hello 👋

edgy notch
#

hiiiiiiiiii

nova lily
#

I can say after days of testing, USDU is is superior to Tiled Diffusion.
(unless im not using TD optimally)

trail lion
white parrot
#

what??

craggy plinth
#

hello

fervent crown
#

hey guys.

#

whats the best Stable diffusion model out rn?\

white parrot
#

hello everyone
what's hand negative stuffs most used overall

main junco
#

What is the best model or method to make faces similar but not identical, like siblings (brother/sister)? Also, how can I maintain the same style and posture?
Similar to face-swapping, but not exactly like that. I want to create something similar, like a brother or sister.

true totem
#

Anyone know if you can do the stable fast 3d stuff here via command line if so can I ask how?

fervent thunder
#

not sure

agile dove
#

hi! can someone help me understand how seeds and batch sizes work? Does every batch get one seed? I've noticed that the n-th number of image in a certain batch will stay the same as long as the seed doesn't change. Is there a way to isolate that one generation without having to generate the whole batch?

fervent thunder
#

it depends on the code base but in general a common way to implement it is for each generation in the batch after the initial one to have the seed implemented by one

agile dove
#

nty

#

u sound like ur trying to get banned

#

maybe miss me with that

desert dagger
fervent thunder
#

if you want to work with batches in more detail then the nodes Latent from Batch, Latent Selector, Flip Latent and MultiLatentComposite are nice

agile dove
#

i dont know what any of those words mean when you put them together like that, will it let me isolate single images from batches?

desert dagger
desert dagger
fervent thunder
#

you have to have some randomness somewhere for batching to make sense
the vast majority of the time this comes from the seed
but any source of randomness can work

agile dove
#

oh shit are we talking about comfy UI

fervent thunder
#

ye

agile dove
#

i'm still on Automatic1111

#

is that like, just straight up outdated

fervent thunder
#

ah ok I can't help with that

#

well some people seem to like it

desert dagger
#

batch is universal. you can't use the same seed with every image in a batch or they will all be the same

#

and auto1111 is going to save the images somewhere

agile dove
#

i might just be an idiot

#

i thought the seeds were per batch not per image

fervent thunder
#

its ok I thought that too at first

agile dove
#

no

#

u don't understand

desert dagger
fervent thunder
#

but it wouldn't work if there is no randomness it would just make the same image

agile dove
#

i found an image i liked 12th down in a batch

desert dagger
#

that's what they are for. so you can actually get the same image multiple times if you want to. that lets you tweak other settigns

agile dove
#

so i've been generating 16 pictures at a time trying to figure out the right clg for one image

fervent thunder
#

oh no

desert dagger
fervent thunder
#

in comfy you could get the seed but
I am worried A1111 might not have a way

agile dove
#

my question seemed so dumb that it didn't fit in the same caliber of

#

how to train loras and stuff

desert dagger
#

that's where the people that can help with auto1111 usually hang out

desert dagger
agile dove
#

i'm probably going to try comfyui

desert dagger
#

i just don't use auto1111 so i don't know the answer - except i can't imagine it doesn't have any way to specify a seed

agile dove
#

no it's embedded int he meta data

#

i literally just didn't notice the last digit is changing

fervent thunder
#

if I had to guess, putting your seed number for the start of the batch plus 15
will get you the seed of the 16th image in the batch

agile dove
#

you know that feeling when you realize you're jsut dumb

desert dagger
agile dove
#

but still kinda relieved

#

thanks a lot guys

desert dagger
#

welcome.

fervent thunder
#

👍

agile dove
#

you guys are using SDXL with comfyui right?

#

how's the generation time?

desert dagger
fervent thunder
#

I use SDXL but generation time depends on so many factors

#

my generation time has varied between 0.1 seconds and multiple hours

#

for SDXL

agile dove
#

but i'm guessing that's probably a nonissue since you didn't lead with that

#

i'm actually kinda psyched about flux

desert dagger
agile dove
#

i heard it's great for text gen

fervent thunder
#

I very recently did go back to SD 1.5

#

keeping all other variables the same, SD 1.5 is a decent multiple faster

#

I didn't keep track but something like 2-4x faster

desert dagger
#

SD 1.5 is a really good model, just requires some fancy prompting

agile dove
#

yeah i probably have a long way to go

#

i mean

#

i just figured out seeds worked today :/

desert dagger
#

🙂 you're doing fine.

fervent thunder
#

my favourite model is actually Kolors

#

which doesn't get as much attention these days

agile dove
#

what's good about it?

#

honestly i just want a good photoshop-esque GUI

#

can you imagine being able to inpaint controlnet stuff with the pentool?

main junco
#

Does fooocus support controlnet?

fervent thunder
#

Kolors has unusual training data

#

firstly its trained on a certain specific look

#

but also its a specialist chinese model

#

which gives it different results to the others

#

to prompt it with best results you actually have to translate your prompt into chinese first

fervent thunder
#

someone made this https://github.com/fenneishi/Fooocus-ControlNet-SDXL

desert dagger
main junco
#

Is it possible to change a face to make it look very similar, like a brother or sister, but not exactly the same while keeping the posture and everything else unchanged? I want to experiment with face obfuscation so that a person looks similar to others but can't be recognized by facial recognition software.
Has anyone done something similar?

fervent thunder
#

towards objects and people

#

which I kinda like

#

I use an SDXL model called LuanFangHua sometimes which has that too

agile dove
#

i feel like SD already has a strong asian bias

fervent thunder
#

some checkpoints do yeah

agile dove
#

anytime it doesn't know what i mean

#

it defaults to asian faces

fervent thunder
#

its hard to generalise across all SD checkpoints

#

some do, some don't

agile dove
#

that's fair

#

i don't use much outside dreamshaper and realistic vision

fervent thunder
#

would recommend CinenautXL, LuanFangHua, RealvisXL and Leosam's Hello World

desert dagger
#

yeah, most of the checkpoints do - because most of them are anime based

agile dove
#

LOL

fervent thunder
#

RealvisXL and realistic vision are the same guys

agile dove
#

i just checked it out

#

saying 乱芳华 has a slight asian bias is like saying

#

the fire was a little hot

fervent thunder
#

LOL

#

乱芳华 is definitely a specialist one yeah

#

and that's the watered down version that has been merged with real-vis

#

容华 is the original and is even stronger bias

agile dove
#

any stronger and it's going to start generating CCP propaganda unprompted

fervent thunder
#

its same with CinenautXL
its been merged with jugger

#

but the original CineVisionXL is stronger bias towards cinematic

#

but I don't use the originals

#

I use the ones that got merged with photo models

#

merging a fantasy/cinematic/Chinese model with a photo model
is the best look in my taste

agile dove
#

yeah it looks like something out of a jin yong novel

#

i dig it

#

wait, how does flux compare to SDXL in speed?

fervent thunder
#

also if you do multiple passes
you can choose the direction you want the image to go in
if the first pass ends and you want it more realistic, you can do second pass with jugger or realvis
or if you want it more cinematic, you can do second pass with kolors or helloworld

#

flux speed was originally slower

#

but now its complex because they made quants of flux

agile dove
#

i never thought about switching models in the middle of my workflow

fervent thunder
#

ye its rly fun

#

cos you can make realistic versions of cartoons, or cartoon versions of realistic
for example

agile dove
#

lol

fervent thunder
#

also you can use depth map
to only change the background

#

so you get realistic foreground but cinematic or cartoon background

agile dove
#

still testing out the different controlnets

#

depth map gives some wonky results

desert dagger
#

just add a geographic location to your prompt and you don't have to deal with the bias

agile dove
#

i just asked chatgpt if there is a way to generate a batch with different settings on the same seed

#

and it wrote a step by step guide on

#
  1. choosing a seed
#
  1. choosing my settings
#
  1. generate
#
  1. change my settings
#
  1. generate again
#

i've never felt so patronized in my life

fervent thunder
#

or even after a third

#

you also might have to add something like soft-edge or canny

#

because depth map alone can sometimes confuse the model

#

when there is something in the depth map and it can't quite work out what it is

#

because most of my images are in the setting of a palace or a mansion, when the model gets confused it tends to put a cloth or a blanket over the thing that confused it LOL

agile dove
#

i didn't think of it that way, usually when i start losing style/coherence i always start taking things away

#

like, less prompts, less CLG and etc

fervent thunder
#

it goes for the reverse too
with canny alone, what can happen is it thinks a foreground line is a background line, or the opposite

cloud lantern
fervent thunder
#

Is flux fine tuning possible yet?

#

ye

fervent thunder
#

its on diffusers

round dome
#

hello

tribal crown
#

Is there any way to turn an img2video via one click in comfyui?
in other words how easy it is?

rich umbra
#

gm

abstract quarry
#

if you talk about full finetuning: although that's possible with enough vram, it seems to be not necessary. Training lora is more efficient and sufficient

fervent thunder
abstract quarry
#

there are other parameter efficient finetuning techniques like lokr that can learn complex things more efficiently

fervent thunder
#

gets a bit squiffy trying to stuff lots of concepts into one lora

abstract quarry
#

that full finetuning is necessary was always a myth

#

in theory a full rank lora is equivalent to a full finetuning

#

so increasing lora rank brings you closer to finetuning

#

with loha or lokr you can get somehow best of both words, having a low rank model but being still able to approximate full finetuning

fervent thunder
#

Rank its not a problem, also i have lots of data, its more about finding the right configs

#

Im renting the compute so vram is no problem either

#

I will do some research on loha and lokr, atm im using ai-toolkit

#

if you have enough $ and time you could make a full checkpoint if you want

fervent thunder
#

are you sure

#

seems very high

tribal crown
#

anyuone can let me know or is it not doable?

fervent thunder
#

via one single click?

#

no

#

but in a short space of time yeah

fervent thunder
# fervent thunder seems very high

Actually 100k is very very low value. Flux was trained on 2.3 billion images, its 12b parameter model, the training cost is likely in the millions. Requiring months + hundres of gpus to train.

That is without factoring in storage costs and experimentations.

#

checkpoint doesn't mean base model

#

Oh

#

so to take the example of SDXL
the base model is SDXL
but a checkpoint might be "Dreamshaper" or "Juggernaut"
does that make more sense?

#

Yeah

pine fiber
#

did you guys see that doom paper lol

#

they trained sd 1.4

tribal crown
#

i have comfyui and flux installed

fervent thunder
#

you've got a few options
AnimateDiff or SVD are ok

tribal crown
#

can i get a consistent face and body via animateDiff?

#

and also where can i install animatediff on comfyui? can i install directly via comfyui manager?

fervent thunder
#

you can get a bit of consistency yeah but its not perfect

#

personally I don't use the manager I use terminal

#

but it may well be in the manager yeah

tribal crown
#

how do i install animatediff via comfyui manager?

fervent thunder
#

not sure if it is in there

#

but you could search around

#

generally manager is not for models its for nodes

#

but they do include a few models there

tribal crown
#

so basically what I would have to do is create an image through flux and then restart comfyui with animatediff?

fervent thunder
fervent thunder
fervent thunder
fervent thunder
#

you can do it in pytorch but its not rly possible to teach all of pytorch over discord

brave yarrow
#

is here can make picture with words?

fervent thunder
#

what is the point of merging a lora into the model? What benefits come from it?

#

not muhc

#

not much benefit

tribal crown
#

do you need animatediff for animateddiff- evolved?

trim finch
#

do u know how to fix it?

tribal crown
#

hey i am just asking

#

fix what?

trim finch
#

wait a sec

river sparrow
#

Btw, you typically do not want 20 loras though...

trim finch
#

TypeError: 'NoneType' object is not iterable

abstract quarry
river sparrow
#

Which now gives me the stupid idea to throw as many loras into my merged model

abstract quarry
#

I just remember that it's just a bad idea for flux.

fervent thunder
abstract quarry
#

why? There are many loras out there training on new domains

#

including nsfw stuff

fervent thunder
#

Learning new variation of what it already knows 60% its different than learning nsfw which it knows nothing about. Thats why the loras are not that good

abstract quarry
#

they are not worse than sdxl finetunes

fervent thunder
quartz siren
fervent thunder
#

I've had a lot of issues with SDXL loras
one of the issues I get is secondary characters being in the training data of the lora
for example C3PO being in someone's R2D2 lora

abstract quarry
#

you assume that low quality comes from loras, which is not the case. To train fully new concepts you have to mix them into a training set of millions of images and train with high batch size. Nobody is doing that. It doesn't matter if you use Lora or full fine-tuning to achieve that

fervent thunder
#

the really big projects like pony were in the millions of images ye

#

its a lot

#

I do wonder, and will test it, if maybe instead you create some nsfw lora with general basic concepts, it will be bad, decent at best. Then you merge that to the model and repeat the process

#

sounds like the mangled model someone on here made

#

it worked well

#

they merged something like 2,000 loras

#

Wow

#

I need to meet this guy

#

if you search mangled

#

it should show up

tribal crown
#

how do i use animatediff

fervent thunder
#

Will do

tribal crown
#

with a photo like this

#

ah

#

cant put photos here

#

where can i ?

fervent thunder
tribal crown
#

can you check my post please? would do appreciate it

topaz shard
#

h@dunno cole

#

hello

ornate flame
#

With flux, is there a way to remove the background blur depth of field bokeh effect from photos without negative prompts? I find the effect annoying, and the current methods for negative prompts slow down speed by a huge amount. There is a lora for this, but it seems to negatively impact details.

hard swift
#

sup

desert dagger
ornate flame
#

yeah it seems so deeply built into the model

desert dagger
# ornate flame yeah it seems so deeply built into the model

it is. it probably trained on way too many artsy photos with blurry backgrounds. to be fair, that's really what you want, it makes your subject stand out. you might try creating the subject in one run, just creating the background in a second run, then compositing the images

ornate flame
desert dagger
ornate flame
#

I've never done inpainting before, but maybe soon. I'm getting crisp backgrounds with an sdxl-based model much faster than flux generates it's blur, so I'm probably going to wait for good flux finetunes.

lean laurel
ornate flame
#

I probably tried that one, the blur is still there, even in some of the preview images. I tried a few loras

neon ridge
#

how does this work again?

desert dagger
white parrot
#

i have some problem about making good background, it usually give me unknown things btw
anyone can help?

floral umbra
#

When people make loras, do you just have a closeup of whatever you want added? Like just picture of a nose? Or do you include the whole face?

trim magnet
floral umbra
trim magnet
#

then yes a bunch of closeup pics of the snout works and then maybe a few of the cat without the snot as a regularization imgs

white parrot
#

i think space/star background is hard for AI

#

thank you sir

fervent thunder
#

Quick question, can you use ComfyUI to help you create images with 2 different characters in it without it mixing their traits and stuff? Especially with different character loras

floral umbra
#

Hanging onto that one as i couldn't for the life of me get 3 characters to wear their normal getup without colors being all wrong lol

fervent thunder
#

Ahh I see

#

I heard Flux doesn't work for SDXL?

floral umbra
trim magnet
#

idk about that srry i only use kohya or the civitai trainer

floral umbra
low moon
#

🥺 diffusionhand

unborn hedge
#

ai art do be hard tbh

trim magnet
#

no

civic cradle
#

Guess who fixed his drive

Turns out the cable is the fault and my drive is fine

#

I'm ready to continue to try to boot stable diffusion 1.5 from an external drive

civic cradle
low moon
fleet condor
#

Is there a node in ComfyUI that allows to expand a mask in a specific direction, such as downward or to the left?

sly eagle
#

for example, I want to make a lineart of a character or a pose and the model make the pose

fervent thunder
#

even 1 image

#

can do that

fervent thunder
#

Character + specific poses (that you have images of)

civic cradle
civic cradle
fervent thunder
#

ye there's no limit

#

just throw it all in

covert kestrel
#

Currently using A1111-Forge due to laptop is severely outdated. I'm currently upscaling an image using Tiling Resample method, and I've seen the "Seams fix" option down below the img2img. A lot of people said not using it is the best option. Is it true?

#

My laptop is a mid-range 2017 Acer Aspire E Series with a Core i7 U-series CPU, a 2GB VRAM NVIDIA GPU, 12GB of slower DDR4 RAM and a SATA3 SSD.

trail lion
#

you could try passing it through an sdxl refiner + uscale workflow...probably any solution will 'alter details'

covert kestrel
#

I'm mostly working with SD1.5 due to obvious hardware limitations. Is SDXL even feasible for a system that ancient and weak?
Though I'm currently holding out upgrade until Q3 next year when NVIDIA announces 5000 series and the price stabilised.

fervent thunder
#

you don't actually have to make images fast

#

I've waited an hour before

#

if you are willing to adjust to patient and slow style then its fine

#

I actually think it can even be a benefit sometimes to go slowly as you spend more time trying to make the next image good

trail lion
#

I just got a 3090 to replace my old AMD card, but I took a chance on a refurbished one

#

so far so good, no issues

#

the other thing to keep in mind, depending on just how outdated your system is, these newer cards need a decent amount of power, so you'll need a power supply and a pci bus to drive that performance, which means a nicer board. if you're doing that you'll need newer ram and might as well get a newer gen CPU, so basically you're doing everything

#

the reason I had an AMD card was because I made those decisions a couple years ago to invest on everything but saved money on the GPU which is the most expensive component by far, so I struggled with a slow GPU for a few years and finally upgraded that just recently, but at least I had already done the other stuff

night violet
#

Hello!
i am new at discord

unborn hedge
#

is that normal btw, to have to make 100's or 10's of images until the seed gets you something that looks good?

fervent thunder
unborn hedge
#

well i cant use flux tbh so im sticking to SDXL for now until it gets super super optimized

fervent thunder
#

You need upgrade.

unborn hedge
#

im not a billionaire lol

fervent thunder
#

Then become one

calm egret
#

If I want to make a virtual background for a camera that is 1280x720, would it be better to use normal stable diffusion or XL?

covert kestrel
quartz siren
trail lion
#

really, I'll have to check out that lora

quartz siren
#

yeah here is the link: https://huggingface.co/ByteDance/Hyper-SD, you can use 8steps with 2-3.5 guidance scale and it will be most of the time better quality then flux.1 dev. It's somewhat similar to flux.1 schnell or the dev-schnell merges but much much better.

last bay
#

does anyone know what's the difference between euler a simple to exceptional

trail lion
#

euler a is a sampler, specifically an ancestral sampler, the simple and exceptional are schedulers

last bay
trail lion
#

speed and quality are at play between the different schedulers...I'm not sure if you meant practically or technically

last bay
#

so the exceptional gets you better quality ?

trail lion
#

that's the idea, at the sacrifice of speed....but I havent done a ton of practical testing on that

copper crystal
#

i'm not sure sd3 has the knowledge depth to always be better than flux as well.

quartz siren
copper crystal
#

oh keen!

white parrot
#

is front prompt is more important than later prompt

pallid tide
#

why do all my outputs just look like a cave wall?

trail lion
#

CFG 1 for an SD model?

#

try like 7

abstract quarry
last bay
#

exponential

abstract quarry
#

it's about the timestep schedule. How much time is spend for composition and how much time is spent on fine details

last bay
abstract quarry
#

I think the karras scheduler (both exponential as well as karras) spent more time on fine details

last bay
#

dpm++ 2m ?

abstract quarry
#

yes

#

for Flux you should just use Euler and normal (=predefined schedule)

last bay
#

oh ok

pallid tide
abstract quarry
#

For photos you need lower CFG

trail lion
#

maybe try euler a vs euler and bump up the steps to like 35, you already said you've tried other models

abstract quarry
#

also you need custom models, SDXL base is not very good for that

trail lion
#

I saw realistic vision in his first image

abstract quarry
#

realvis is probably the best model for photorealism

trail lion
#

I still say start at cfg 7, you can lower it from there, but it'll listen to your prompt less

pallid tide
#

still looks ass , ill send pin in general with imgs

abstract quarry
#

prompt it differently

#

don't use 8k

#

also you don't need "photorealism"

#

use "photo"

#

or better:

#

"photo of a girl in the forest, 35mm photograph "

#

you can also try to just write "girl in a forest", because realvis should already make the image photo-like even without prompting for it

#

also

#

use width and height of 1024

#

not 512

pallid tide
#

it looks better but still not as realistic as they should be

abstract quarry
#

SDXL is not a 512 model

#

you have to use SDXLTurbo for 512, or use SD 1.5

pallid tide
#

90% sure its not an xl model

trail lion
#

oh, it's 1.5? 1.5 models are pretty inferior

#

for photorealistic at least

pallid tide
#

have they gotten worse than they were previously? i understand they wouldnt be better, but this is worse than real vis 3 used to be

#

i havent really done much with sd since then, so im not up to date at all

abstract quarry
#

if your hardware is good enough, use the best model available (Flux) or SDXL (if your hardware is a bit older)

pallid tide
trail lion
#

there's a realvisxl I have a v40 not sure if that's the newest, and there's juggernaut and a slew of others for XL

pallid tide
#

i just tried out real vis 3 and it looks as good as it used to, i think its just the new model and the way you have to promp it

trail lion
#

if you're not doing people, there's a dreamshaper

elder pivot
#

Hello, does anyone know where I should contact for questions about stability.ai?

trail lion
#

dont they have a contact on their page?

#

I know that this discord is not the right place

#

yah there's a giant button on their page that says "contact us"

#

and if they are unresponsive, it's because skeleton crew 😄

covert kestrel
#

I noticed on top of my Forge instance, I have two extra options: Swap Method and Swap Location. What are these for and what are the differences?

#

On Swap Method I have:

  • Queue (block method)
  • Async (asynchronous method)
    On Swap Location I have:
  • CPU
  • Shared
fervent thunder
trail lion
#

so there's an ipadapter for flux now, and it actually works. so exciting (it's been out for over a week, but it's news to me)

fervent thunder
#

ye it came fast

copper crystal
#

the xlabs one? people been telling me it's really bad and doesn't work, and other people been telling me they're just using it wrong and don'tknow what they're talking about. i haven't had a chance to get it working yet. i used forge to get started

fervent thunder
#

its not great

#

works a bit though

copper crystal
#

i saw some examples where they plugged in black and yellow striped theme image and got the same idea out of the gens. not sure what parameters they used. like i said i haven't dug in too hard on it.

#

some image types might work better than others according to their dataset

low moon
#

thats why non local generators are useless for anything other than casual gambling. u need many many iteierations so your credits will be eaten fast (which they count on) + the censorship is also a big issues with anythign non local

unborn hedge
trail lion
#

what I've notice, even if you only look at the last hour of conversation, you can get into endless debates on what is good...so I wont go there, but I was able to produce an image with ipadapter, and it did create something that looked like the source image

#

I'm using swarm/comfy, so the gentleman above using forge, I cant comment on that functionality

heady salmon
#

Hey all
Sorry if this is a stupid question, I'm just starting to explore AI stuff
I've been using DALL-E for inspiration and reference pics for a worldbuilding project I'm doing, but it would be nice to have something I could run locally
I generally like the results I get when I ask for a "hatched illustrated" style (very medieval-y, gets the vibes I'm looking for)
Guess my main question is, is there a 'best' model for this kind of thing? Most of the Stable Diffusion stuff I've seen is geared towards either hyperrealism or anime. Is it a matter of a specific model or tool, or of training an existing one, or both, or something else?

trail lion
#

historically speaking, auto1111 and forge lag comfy by some weeks usually for most things

#

that of course may change going forward, because comfy is no longer part of stability

opaque idol
#

Can someone help me getting to know what model was used to do a image ?

trail lion
#

that's possible, but only if the metadata is still there in the image

opaque idol
#

how can i check it ?

trail lion
#

what GUI are you using?

opaque idol
#

forge

trail lion
#

in forge, there's a pnginfo extension, drop it in there and it tells you everything used to generate it (if it's not been stripped out)

opaque idol
#

let me check...

opaque idol
trail lion
#

then there's no way to really know, sorry

opaque idol
#

and if i send u the image?

trail lion
#

it wont matter, if the metadata is gone, it's anyone's guess

opaque idol
#

ffs

granite badger
#

hello

#

can any of you help me?

i want to get similar images like the one on the screengrab

#

https://www.amazon.de/Fantastische-Fabelwesen-Stressabbau-Entspannung-Fantasie-Kreaturen/dp/B0CN5B8WTG/ref=sr_1_1?crid=3IBODT2J8X6H6&dib=eyJ2IjoiMSJ9.-3XggVW3uObjvvXQqObf-g-EWf_V6QDcBkrHerEySuY2P3W0J8JG92mAOXoFt2DWOwZHT1w0m6M4IrDxhUwXVi523Affpx6n5y5TI3Pal5iMGXUuSJEje7x1BSRxDuAhRJqcESyU0awWBpc07xA90cucn7Z_uETG34wev0if1-ON4ICntYnPnlLPGVH6WUk532dqEr89fXftuzS4TrhIrYMCKNik-WVzuMj3aU2Vvr8.d_Vd1P3m4memC-Dd8Agtfsyxu8CgD6J3vjQdJ--SaDo&dib_tag=se&keywords=fabelwesen+malbuch&qid=1724956770&sprefix=Fabelwesen+%2Caps%2C126&sr=8-1

#

I don't know how to upload a screen shot

#

I mean something similar to what this coloring book shows

trail lion
#

just dont post anything against the discord policy, like nudes, etc

granite badger
#

ok will try

glass lily
#

@warm junco Hey what do you think I could do to fix the ui-config.json issue I have? I edited some prompts and resolutions in txt2img and img2img in the text file, and it just reset everything to default, no default prompts, no Adetailer prompts and parameter settings (only the default ones).

floral umbra
#

How many steps do you guys do per image for x512, x768 and x1024?

glass lily
#

@warm junco Hey I'll leave this here if you get on and still feel like helping me:

When in A1111 I go to **Settings **➡️ **Other **➡️ Defaults and click View Changes, I get this in my CMD:

#

I'm missing perms

#

To send the txt file

#

But yeah

#

I get Traceback

#

venv, modules, and Python in AppData

quartz siren
# floral umbra How many steps do you guys do per image for x512, x768 and x1024?

depends on what model you are using, for flux.1 dev, around 30 steps for 1024 is usually good quality. The lower res, the lower steps you can do but I would keep it at 20.

Also, I would just recommend using hyper flux since its usually better quality then flux dev and just requires 8 steps for all res so a lot faster then it too.

For sdxl, kolors, sd3, the best amount is usually also 30ish steps but they will not work well below 1024 res.

warm junco
glass lily
glass lily
#

(same result)

floral umbra
abstract quarry
#

what?

#

I think that was in very old times

#

even for sd 1.5 30 steps were enough

floral umbra
#

Gotcha.

brazen orbit
#

Why is stable diffusion so bad

#

Why isn't it so easy available online..

tulip yarrow
#

How long does it take to train a base model for SD1.5, SDXL, SD3, and Flux (respectively)

#

Wait, several million dollars?

desert dagger
#

yes. those models are billions of parameters. that's a lot of gpu time

tulip yarrow
#

...nevermind

desert dagger
#

datacenter gpu time

tulip yarrow
#

Then how were those non-merge model guys able to train it

desert dagger
tulip yarrow
#

Oh... OH

desert dagger
#

those just update small amounts of the base model's weights for specific information

tulip yarrow
#

I meant like checkpoint, not an actual base model

#

I just call them base models

#

Like base model for SD1.5, SDXL, etc

desert dagger
#

okay well, depends on how specific you want your checkpoint or lora to be. you can get away with only 30 images for a lora if you're really specific and it doesn't take very long

tulip yarrow
#

I remember reading somewhere non-merge checkpoints take millions of images to train

desert dagger
tulip yarrow
#

Another anime model in a distinct style

#

Even though I know Loras can do that... I just want to have the experience of training a checkpoint

desert dagger
desert dagger
tulip yarrow
#

Yeah besides I don't have a computer with GPU yet

#

I use google colab to train my Loras 😂

desert dagger
iron swallow
#

what pc do i need to run stable diffision without issues? I want to create high res images, be able to make a model that keeps the person looking the same over diff images and positions

desert dagger
iron swallow
#

would that work

#

as my budget is low

tulip yarrow
#

According to a friend, technically you still can run stable diffusion with only CPU but the image will take days to create and then it will just crash.

#

So... you need a GPU

desert dagger
desert dagger
tulip yarrow
#

And meanwhile the cheapest one i can find is for $600 before taxes it's an RTX 2050 laptop

iron swallow
#

what about those mini pc's where u can allocate the ram to the iGPU

#

basically u got one with 64gb ram

#

then allocate 16gb ram to the Igpu

desert dagger
desert dagger
fervent thunder
#

the way I do things is to build the workflow at home, couple of tests with CPU

#

then upload to cloud GPUs to run fast

#

if I had to use CPU all the time what I would do personally is look into 1 step, or maybe 2-3 step hyper or LCM workflows for SD 1.5

#

👀

compact swan
#

where do I post 1.5 images?

low moon
#

On your wall. XD

desert dagger
compact swan
floral umbra
#

Shit.. appears training 1024 res images with flux eats juuust a bit over 24GB video memory. Or 23GB plus 0.1-0.2 swap storage. So I might limit my Lora making for it to 768 or 512

iron swallow
#

so 12gb vram wouldnt be nearly enough

#

to do proper shit?

#

with gtx 3060

pseudo pumice
#

sup chat

abstract quarry
#

you can train with MUCH less

#

I can event train on batch size 3 with 24gb

#

batch size 1 would be around 16gb I think, but there are ways to shrink it down even more

cloud lantern
abstract quarry
#

you should definitely use quantisation. As you use quantisation in inference, too, this is not a quality problem

#

and don't use a adamw optimizer but something more lightweight like adamw_16bit or adamw8bit (or adafactor, although this one often gives worse results)

rich umbra
#

sup

dawn mulch
#

Anyone aware of a workflow that lets you specify an area on an image cut it out and feed it to sampler to regenerate (with optional upscale or user can add it), and place it back in the correct position of the original image with the original scale? Same idea as inpainting but done differently.

abstract quarry
#

I mean, that's ADetailer from impact-pack

#

unfortunately, they chosed to solve this kind of simple problem in the most complicated way possible.

dawn mulch
#

I might have used ADetailer before i'm not sure. But I'd like to breakdown the process for various reasons one being that it might give the user more control the other is that it might breakup the processing. I tried FaceDetailer or face fix whatever it's called and it uses a lot of processing compared to if I just manually broke down the process then my computer would have no issue performing the same exact process, one step at a time.

abstract quarry
#

you can breakup as much as you want. They have hundreds of nodes for each single step

#

it's just totally unnecessary complicated and proper documentation is missing

dawn mulch
#

I am somewhat new to comfyUI but definitely have used it, but I am not super familiar with it or its nodes and I trying to figure out what node does what or is best for what is a complete trial and error that can take forever. So I was wondering if anyone had already broken down or done this process. I am sure that someone who is familiar with many of the nodes would be able to set it up quite easily. But yes the main thing is just more user control and customization and breaking up the process so it's not so processor heavy.

abstract quarry
#

I did this for SDXL, but not for Flux yet

dawn mulch
#

I'm still using mainly SD 1.5 since I have only 6gb VRAM. Occasionally use SDXL but it runs out of memory if i try to do anything a bit complex beyond something like UltimateSDUpscale. Like adding ipadapter or whatever. Soemtimes it generates most of the time it runs out of memeory.

warm junco
#

If your using Auto1111, go into Settings and enable FP8 mode for sdxl.
With that it uses less vram

quaint galleon
#

hi, i need heko with something

#

help*

slate stirrup
#

ребята

#

есть русскоговрящие?

#

anyone speaks russian?

peak orchid
#

Did you guys see that RunwayML deleted all their Stable Diffusion 1.5 repos on HuggingFace and Github?

desert dagger
fervent thunder
#

text 2 video locally, amazing

fervent thunder
#

we have civitai backup.

peak orchid
fervent thunder
#

yeah something is up.

#

i hope this happen only cuz stability ai want the files.

hasty hornet
fervent thunder
#

maybe they part away ?

#

looks like SAI dmca the models 2 years ago lol ... its taking effect now!

#

its funny how sd1.5 was a leaked model XD

desert dagger
river sparrow
#

I think its mostly a shame if public available resources are removed in general, with some exceptions of course

fervent thunder
#

Hey everyone! 👋

I'm a beginner in the field of data analysis and I'm actively seeking an internship to kickstart my career. I'm eager to learn and grow in this exciting field.

If any of you know of any internship opportunities or have advice on how to get started, I'd be grateful for your help! 😊

#

intersting fake social media app that uses ai.

floral umbra
outer cairn
#

Hey everyone,
I wanted to share my latest work titled "Enhancing Conditional Image Generation with Explainable Latent Space Manipulation". It explores a novel approach to improving fidelity and contextual alignment in text-to-image synthesis by integrating diffusion models with explainable latent space manipulation.
If you find this topic interesting or have any thoughts or questions about the work, feel free to reach out to me. I'd love to discuss it further!

You can check out the paper here: https://arxiv.org/abs/2408.16232
Github Implementation: https://github.com/kshitij79/CS-7476-Improvements-in-Diffusion-Model

fervent thunder
# abstract quarry I mean, that's ADetailer from impact-pack

ye I am so glad other people also feel this way
I never really even got past understanding "what is a SEG" and why do we need it
from the Python, a SEG is an abstraction over the following:
"cropped_image", "cropped_mask", "crop_region", "bbox", "control_net_wrapper", "confidence", "label"
so its 7 things that we mostly have existing nodes for already

#

what I can't work out is if this specific usage of SEG has actually been used before, e.g. in a paper or model somewhere
or if it really was made up for Impact Pack

abstract quarry
#

the use case is usually that you run a segmentation algorithm beforehand. Many segmentation algorithms also output bounding boxes.

#

they are then the starting point for the detailing

compact swan
#

should I switch to SDXL from SD1.5?

#

I only have a 3060 Ti

#

and I'm most worried about long generation times + model storage size

white parrot
#

what's best lora and prompt for fixing hand
anyone can help?

desert dagger
warm junco
white parrot
compact swan
#

or do I just add this stuff to the .bat

warm junco
desert dagger
white parrot
#

ty sir

compact swan
#

does adding that command install a medium vram version of sdxl next time I open up auto1111?

warm junco
#

They dont install sdxl

compact swan
warm junco
#

You still need models

compact swan
#

yeah

warm junco
#

These commands make sdxl run faster on GPUs with 8gb or less.
Xformers is a speed boost in general (reducing vram usage a lot) and always recommended to use.
--medvram-sdxl loads sdxl models splitted into the vram and reducing the vram usage to generate faster

#

--no-half-vae is for VAE compatibility

floral umbra
#

In ai-toolkit, is flowmatch the only recommended sampler for flux? And which optimizers are there,and what's recommended for flux?

floral umbra
feral pike
#

Who has 3060 or 3060 ti gpu?
What speed are u getting on sdxl and flux models

white parrot
#

can i upscale mutil image at once?

winged peak
#

اعطني صورة شمس

#

generate شمس

#

generate/ شمس

compact swan
#

even with xformers and med vram

feral pike
#

how slow?

compact swan
# feral pike how slow?

I tried to generate a batch of 4 images at 1024x1024 and I think it finished the first image but after that the time left kept going up

#

it started at 5 minutes and kept going up and up to 10

#

it made my PC so slow that even the interrupt option wasn't working

#

I had to close the tab and terminate the command prompt from task manager

feral pike
#

how many iterations per second u got?

feral pike
compact swan
#

yeah

compact swan
feral pike
#

which UI did u use

compact swan
#

auto1111

feral pike
#

did u try fooocus?

compact swan
#

nope

feral pike
#

thinking about buying a GPU for SD

compact swan
#

too broke for that

#

I think I'll go back to dreamshaper or other 1.5 checkpoints

feral pike
#

my 16GB vram GPU aint working

#

with SD

#

AMDead moment

feral pike
compact swan
#

JuggernautXL specifically

feral pike
#

🤔 and it wasnt working well?

real pewter
#

one question, can stable diffusion work without the need for a super powerful computer using only cpu?

feral pike
#

on the AI cpus

real pewter
#

I'm relatively new to this, so I don't understand anything.

compact swan
compact swan
#

you basically NEED a GPU because video game math happens to line up with AI math

feral pike
#

how much ram do u have?

#

because the freezing issue happens when all ram is used up i think

feral pike
#

i had this too , then i upgraded to 64GB ram from 32GB and it was fixed

real pewter
feral pike
#

nah i am asking kraken

compact swan
#

I only have 16gb

feral pike
#

hmm

#

u have that same issue as me then

compact swan
#

yeah it's probably low RAM

feral pike
#

everything freezes

compact swan
#

but jeez, 64 gb needed?

feral pike
#

yes it fixed it

#

for me

compact swan
#

I think I might go back to running SD 1.5

feral pike
#

but it wasnt running XMP

#

so i had to switch back to 32 GB

compact swan
#

I only have 16 gb ram and can't upgrade

feral pike
#

sadge

floral umbra
#

5 min on steam deck's 8CU's gpu

edgy light
#

how can i generate a video with custom prompts?

#

using the API

warm junco
compact swan
outer shard
#

i'm thinking of getting a new laptop, and i noticed this one from best buy. However it doesn't seem to have a gpu/vram, and i'm not an knowledgable about pcs and stuff. Can anyone tell me if how well it might run with ponyxl or run at all?

#

ASUS - Zenbook S 16 16” 3K OLED Touch Screen Laptop, AMD Ryzen AI 9 365 - 24GB Memory - 1TB SSD - Scandinavian White

warm junco
compact swan
#

seems to be much faster

#

ty again

warm junco
trail lion
# outer shard i'm thinking of getting a new laptop, and i noticed this one from best buy. Howe...

it's difficult to put into words just how unhappy you would be with that choice. It's an amd gpu. Here's the specs: https://www.asus.com/us/laptops/for-home/zenbook/asus-zenbook-s-16-um5606/techspec/ It has an 890M GPU which is a 4Gb vram card. In essence if you could get anything at all to work it would take a millenia to do it. For more fun, see this review of that 890M https://www.tomshardware.com/pc-components/gpus/amd-latest-integrated-graphics-perform-like-an-eight-year-old-nvidia-midrange-gpu

trail lion
outer shard
#

Hm I see I see

compact swan
#

another question, can I add custom hires fix models to auto1111? where?

trail lion
#

that's actually 2 questions

shy leaf
#

Como genero imagenes?

trail lion
compact swan
#

ty

#

I feel silly for asking so many questions

#

SDXL/1.5 checkpoint suggestions for generating spacecraft/rockets/aerospace technology/vehicles?

fervent thunder
outer shard
#

actually how well would it run on a macbook m3 pro?

fervent thunder
#

very poorly

#

its unfortunate that doing deep learning on apple silicon gets suggested all over the internet

#

unless you are making extremely secret images you can just use cloud at a verified datacenter and there is a very high chance your data will be fully private anyway

#

otherwise, even a very low end GPU, such as 8GB VRAM, would be a better choice than apple silicon

outer shard
#

i see

#

thanks for letting me know

fervent thunder
#

if you do decide to go with apple, its worth getting good with the speed-ups like turbo, lightning, hyper, LCM etc

#

SD 1.5 hyper at 1, 2 or 3 steps can make ok images

#

and that would actually be fast enough to be tolerable TBH

outer shard
#

hm not too concered with running sd on it honestly, i'm using a 4070 for my pc

#

but i was hoping that i could run sd when i'm away from my pc for long periods of time

fervent thunder
#

there's some personal choice cos not everyone cares about speed

#

here, and on reddit, there are some people very happy with their very slow setups

delicate tiger
#

in terms of “value for money”, which one is a better pick: rtx 3060 or 4070 super?

fervent thunder
#

rtx 3060

low moon
#

😮

pseudo pumice
#

sup

rich umbra
#

gm

unkempt hatch
#

what fabric is the simpson's couch made from?

#

not leather, right? What would you call it in a prompt?

fervent thunder
#

remember its not about finding the right word
its about finding what was in the captions

unkempt hatch
fervent thunder
#

same

#

this is one of the areas where I disagree with most model creators the most

#

I think they should open source extensive amounts of their data set

#

doesn't have to be anywhere near enough to copy the model so they don't have to be scared of doing that

#

but it would allow us to get a better idea of what prompts the model likes

#

the other big area where I disagree is that I would have made image diversity a larger priority, even if that damaged image quality a little

#

also I would have preferred almost all existing checkpoints and loras to have a looser fit, so we can stack more loras

#

loras could be like lego blocks where you build up an image from like 10 loras that each add a little bit

#

but that's not rly what Civit loras are like, they just give you a fixed finished image

iron swallow
#

rtx 3060 12gb good enough to create photo realistic images of humans?

fervent thunder
#

yeah definitely

#

you want at least 6-8GB so 12GB is definitely fine

#

the minimum is like 2GB

granite sedge
#

hey im trying to install automatic 1111 but when i try to run it it says error no checkpoints found
where do i get the checkpoint I need?

fervent thunder
#

civit ai is common

#

sometimes huggingface

#

sometimes github

signal quarry
#

huggingface

edgy light
#

hi, can someone help me?

dawn heron
#

I just did the math and apparently I'd need 282,641,285.12 strands of DNA to hold my entire comfyui folder

dawn heron
#

Back up of all family photos from 2012 to 2019 = about 9,983,283,200 bits, so need 99,832,832 strands of DNA to hold entire data of folder
School video from last year = about 484,380,672 bits, so need 4,843,806.72 strands of DNA to hold entire data of video file
Entire math class summer packet = about 1,044,480 bytes, so need 10,444.8 strands of DNA to hold entire data of pdf file

edgy light
#

how can i generate a video with custom prompts, using the API??

boreal venture
#

Hey guys!

Starting a challenge today: I’m earning 1€ and will increase it by 1% daily. By year’s end, I’m aiming for around 37€ a day.
It’s day one, so I’m looking for small tasks to hit that 1€ goal. If you need help with anything or know of any gigs, hit me up. Thanks!

mighty loom
#

yello bible thumpers

#

Jk i only thump the AI teachings

#

Anyway guys if anyone wants to dm or ciscuss AI or code related ideas, speculations or just imagination based, impossible becoming possible coding or AI developing I'm always game!

low moon
#

So I've been out of the loop for about 2 weeks which = 2 years in AI land. What's the latest bests Comfyui locally runnable models and video. I hear some new video thing is out?

barren pollen
#

hiya

#

when did we switch to payed creation?

low moon
#

there is still free stuff...

quartz siren
sly eagle
#

It can work for animation too?

low moon
floral umbra
#

Do people mostly use forge these days, as well as for flux?

high flint
#

I've seen a lot of hype on the Flux model lately. How good is it actually?

fervent thunder
#

it depends on how you judge it

#

the overall image quality is very good, aside from certain specific styles, which sadly includes realism

#

its distilled which has quite a few side effects, not all of which are even fully known

high flint
#

So, photorealism's out then?

#

(as in, other models can perform that better and easier)?

fervent thunder
#

other models can perform photorealism better and easier yeah

iron swallow
#

can u also make it so u can generate same human face and everything in diff positions and locations?

high flint
#

inpainting/controlnet is what Id look into for that?

#

Anyway, looks like digital painting style is what flux shines in?

fervent thunder
#

some of the distillation issues with flux are a lack of image diversity and loss of ability to work well with negatives

#

digital painting yeah or especially midjourney cinematic style

high flint
#

I mainly use my workflows for the following, Photorealism, digital painting, and anime, different models for each but, looks like it could be worth trying out flux for the digital painting stuff.

#

It's been a long while since I've touched any 1.5 model though to be honest

#

mainly been using SDXL/Pony

wise harness
#

Hey does anyone know how I can get to the image bot

fervent thunder
#

yeah I use SD 1.5 less now

high flint
#

what things if any do you use 1.5 for nowdays?

fervent thunder
#

the base model is really chaotic

#

so you can use it to start off an image

#

and then refine with SDXL

high flint
#

Yeah... I remember.... it would generate a monstrosity with 100 arms and other nonsense from time to time.

floral umbra
#

To train loras with A.I-toolkit for SD 1.5, can i use any model to make the lora of? Or only barebones 1.5 from hugging?

sly eagle
#

is there's a ai for animation that is free?

floral umbra
#

Though, as far as i know, only available models can only make 16 frames worth, then it needs to blend 4 frames, generate more, but those more isn't identically continuing, but blends into something else

upper turtle
#

stable studio or comfyui

next haven
#

tell me, I installed
flux1-dev-bnb-nf4 but all the pictures are blurry. The image is visible, but it's not clear at all. What else have I not installed?

#

Forge with CUDA 12.1 + Pytorch 2.3.1 also installed

#

4070TI Super

low moon
#

loooool im gettign Pond5 watermakrs in my cogvideos XD

#

ultimately its not a bad video generator probably beats svd... needs img2vid tho XD

fervent thunder
#

yeah img2vid for the bigger cog video model would be amazing

hardy sedge
#

Is topaz Photo ai the best upscaler?

fervent thunder
#

I think this upscaler is the best https://openmodeldb.info/models/4x-NomosWebPhoto-atd

barren mountain
#

what's the best option to capture/transfer hair likeness? like if I have an image of someone with unique hair and want to generate more images with similar hair

barren mountain
fervent thunder
#

the main face specialist one is https://openmodeldb.info/models/4x-FaceUpDAT

#

it does do a lot better with eyebrows and hair than similar models

#

these two big HATs are good too for people https://openmodeldb.info/models/4x-Nomos8kHAT-L-otf https://openmodeldb.info/models/4x-Nomos8kSCHAT-L

#

SUPIR is for super degraded images yeah

#

it was slightly strange that Supir got a strong reputation as a general upscaler

#

models trained on heavy degradations smooth things over too much (by design)

reef aspen
#

Hi - anyone using SDXL 1.0 ?

cerulean oyster
#

I'm a comfy user and want to learn more can anyone help me

#

Oh and im new here

quartz siren
round dome
#

hi

fervent thunder
serene forum
#

Hi,
I'm looking for ComfyUI tutorials for advanced architecture and interior design.
Could anyone help with some relevant resources?

wet grotto
#

anyone having problem with colab? my sd stop generate images after 15 min or less, and I have to run colab again... It's happening by almost 2 weeks, idk what is this 😣

trail lion
#

the free one or pro?

wet grotto
trail lion
#

Id think pro should work, imo colab is pretty crappy. you'd be better off on something like vast which still uses jupyter and cheaper and gives you more choices over the runtime environment

wet grotto
wet grotto
trail lion
#

choose a template, then you choose your runtime environment, once you've done it one time, it's easy

#

most of what I use vast for is training, but it can certainly do more

wet grotto
trail lion
#

kind of blows my mind how many people use that old model

tacit storm
#

hey

desert dagger
trail lion
#

yah, for it's time

#

and it used less resources I guess

#

whenever I go back to it though, I'm horrified

#

more abominations, more issues with drawing things in the distance due to the small latent space, and the prompt comprehesion is the worst of all the models

fervent thunder
#

well drawing things in the distance is a funny one

#

cos the thing about SD 1.5 is it wasn't trained on a lot of bokeh

#

I remember when SDXL base came out, I used to hang around in slightly different AI communities back then
and they were super unhappy with the bokeh

trail lion
#

I dont really mean bokeh, but I see where you're going.

fervent thunder
#

I know what you mean yeah- accurate fine details

#

what I would say is it depends on the resolution

trail lion
#

for sure flux seems to prefer photgraphic and bokeh

fervent thunder
#

SD 1.5 is much better at high resolution than SDXL
because Hi-diffusion works better with SD 1.5

#

for example

trail lion
#

I'm guessing a finetune will fix that, just like the flurry of amazing finetunes made the base XL model all but unused

fervent thunder
#

there's at least 4 big releases coming up this year

#

SD 3.1, the OMI and the next generations of Pixart and Auraflow which will have the modern VAEs

#

would also expect something from China e.g. Lumina team, Kolors team or that Huan one

#

I partly think its likely people will switch from flux

trail lion
#

if they do, it'll have to be because something is way better, because flux is very trainable and that's usually all the community wants

fervent thunder
#

really disagree

#

it does depend on what people mean by trainable

trail lion
#

well, as a way to define it, I would say that training on 1.5 and xl frequently let to 3 headed monsters, extra limbs, and burned out very quickly

ocean wasp
#

Hi 👋
Does someone know how to fix anime full body pictures in ComfyUI?

trail lion
#

sure, you can go into weeds on training params and such, but just saying, so far it seems the flux model is pretty tolerant and hard to overtrain

#

18k steps later ,you still get pretty solid images, in my experience

fervent thunder
#

ok so here you are referring to precision
but that's not the issue with flux, in that area it is the strongest model, the problem is recall and conditioning

#

it can make very high quality images, but it cannot make a wide variety of images

#

and it cannot handle a wide range of conditioning vectors before getting off-manifold issues (CFG burn)

#

when people say "Flux can't use negatives" its a bit of a misnomer
Flux can use negatives it just burns the image though, by flying off the manifold

trail lion
#

the thing with the negatives is a fair point and I do miss them at times, or I work around it by feeding it an image

desert dagger
fervent thunder
#

yeah that's a good tip

#

you can get a little bit of negative guidance

#

on reddit some mad lads got up to cfg 6 by chaining a bunch of anti cooking nodes (rescale/tonemap/threshold/auto CFG etc)

#

I actually had to do this with SDXL lightning today, its a similar issue, just not as bad as with flux

past furnace
oak latch
#

can i pay someone to make me a flux lora

fringe pagoda
#

Wondering there are githubs or repositories extending SD 3 to video editing or generation?

tulip yarrow
#

You think anyone made a Lora based off of Super Smash Bros' Master / Crazy Hand?

#

because you know how AI and hands work

#

or rather, don't work 😂

tribal sequoia
#

I want to upgrade from my intel graphics card so that SD doesn’t take ages to do anything. Any recommendations for a card that has a good price to performance ratio?

desert dagger
desert dagger
solemn ermine
#

hi

high flint
#

in ComfyUI, does lora loader have to be near the start of a workflow? I'm hoping to split up a generic prompt 10 times into loras and see exactly how much impact each lora has.

desert dagger
dawn mulch
tribal sequoia
#

8+ GB vram, got it

high flint
desert dagger
tribal sequoia
#

what does 8GB look like? Longer time to output an image?

little lava
#

1

fervent thunder
#

I've played around on most VRAM sizes and 8GB does not feel great

#

its doable but you get out of memory errors a lot

#

once you start adding extra stuff like control net or upscaling

eager obsidian
#

does anyone have a link to a google collab working and supported fork notebook of "deforum stable diffusion" please?

desert dagger
eager obsidian
#

I was hoping a clever person out there would have an updated version they are supporting

desert dagger
eager obsidian
#

not as it did....

#

"⚠️ NOTICE: This project is no longer maintained. ⚠️ This repository is no longer actively maintained or updated. Users are advised to find alternative solutions or fork the project if they wish to continue development."

desert dagger
eager obsidian
balmy hemlock
#

Hello, anyone know how to use adetailer + soft inpainting in img2img tab of a1111

tribal sequoia
fervent thunder
#

its a funny idea but yeah it sadly would not work

#

it just can't fit stuff in the box its not a matter of speed

balmy hemlock
#

is anyone know how use soft inpainting with Adetailer ? 😢

golden socket
#

damn all that installation just to realize my 1070ti is too old to generate images 😂

warm junco
azure drum
#

Hey, guys. I have a question. Can I, for example, generate an NSFW image? Doesn't this violate the policy of using Stable Diffusion 3 Medium?

fervent thunder
#

depends on the new license

fallow pine
#

anybody got this version of forge ui? Version: f0.0.17v1.8.0rc-latest-276-g29be1da7 I'm looking for this one please share me if you have it or tell me where I can find it

balmy hemlock
#

hi, is anyone know how use soft inpainting with Adetailer?

inner plaza
#

guys, best background remover with an API thats cheap?

covert bridge
#

hi there..i am running flux in forge, and i was hoping someone could explain what this means in the cmd prompt as model runs

[Unload] Trying to free 18359.89 MB for cuda:0 with 0 models keep loaded ...
[Unload] Current free memory is 8456.80 MB ...
[Unload] Unload model JointTextEncoder
[Memory Management] Current Free GPU Memory: 8959.56 MB
[Memory Management] Required Model Memory: 6246.84 MB
[Memory Management] Required Inference Memory: 10239.00 MB
[Memory Management] Estimated Remaining GPU Memory: -7526.28 MB

i have 32gb ram, plus 12gb on my 3080RTX.

native sundial
#

is just the way Forge makes sure your pc doesn't run out of memory

#

i do recommend grabbing the NF4 version of Flux as that one can run on 3GB VRAM even and not 16 GB at minimum

round dome
#

ok

copper crystal
#

100% looks like a phishing hub scam situation. i wouldn't start spamming your "young site" since it looks wholly illegitimate

#

oh got it's unsecured http too. this guy is just farming IP addresses most likely.

#

anything sent to this page is sent unencrypted and will be plain text to local network admins and any processes running at your ISP.

#

or if you use any proxy server, they'll see everything too

#

the only time you see http is when some inexperienced person is trying to launch a scam off and can't figure out how to get a signed domain

hoary ember
#

Error occurred when executing PipelineLoader:

Error no file named config.json found in directory D:\ComfyUi\ComfyUI\ComfyUI_windows_portable\ComfyUI\models\IDM-VTON.

File "D:\ComfyUi\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 151, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUi\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 81, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUi\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 74, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUi\ComfyUI\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-IDM-VTON\src\nodes\pipeline_loader.py", line 46, in load_pipeline
vae = AutoencoderKL.from_pretrained(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUi\ComfyUI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\huggingface_hub\utils_validators.py", line 118, in _inner_fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUi\ComfyUI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\diffusers\models\modeling_utils.py", line 567, in from_pretrained
config, unused_kwargs, commit_hash = cls.load_config(
^^^^^^^^^^^^^^^^
File "D:\ComfyUi\ComfyUI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\huggingface_hub\utils_validators.py", line 118, in _inner_fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "D:\ComfyUi\ComfyUI\ComfyUI_windows_portable\python_embeded\Lib\site-packages\diffusers\configuration_utils.py", line 374, in load_config
raise EnvironmentError(

#

guys having this error can anyone tell me how to solve i am trying really hard and going to loose my mind now

#

can anyone tell me how can i solve this

fervent thunder
#

when using stable diffusion, is it better to use prompts structured as sentences or prompts as individual characteristics separated by commas?

abstract quarry
#

when using SD 1.5 you can go for attributes separated by comma. It's not that it was trained on that - it's just that the CLIP-L text encoder used by SD 1.5 is soo bad in text understanding, it will only listen to certain words anyways without getting the complete sentence

#

SD 2.1 and SDXL use better text encoders, but they are still CLIP-based, so they still have very limited text understanding. They benefit more by natural captions but you still can use comma separated lists

#

Flux and PixArt are using T5, which is a powerful llm that really understands your prompt. It will benefit a lot by natural language prompts. Also, both models are likely trained on natural language captions, so its better to stick with that

trim magnet
#

the 1.5 anime models and its derived merges were actually trained on comma keywords because they contain the nai dataset and that was trained on booru tags

#

some XL models were also trained on booru tags like Kohaku XL,Animage and Pony

abstract quarry
#

even Flux was trained in booru tags

#

but a) will this bias the output to anime and b) will the model have better prompt coherence with natural language

trim magnet
#

yea on models with t5 is better to just use natural language

amber bear
#

anyone got a current guide on running stable diffusion on windows + AMD?

#

tried using lshqqytiger's fork of automatic1111 and am continuously running into issues

jovial stream
#

Could someone help me understand how i would make an anmation or video using ai?

amber bear
hasty hornet
#

do we have a somewhat decent text=>3d or img=> 3d yet?thinking

white parrot
#

how to know what i installed in ComfyUI

desert dagger
desert dagger
white parrot
#

i'm seeing new stuffs on comfyui called "flux"

#

and i dont know if i installed it or not

desert dagger
white parrot
trail lion
#

but even if not, it kind of does it already, by embedding the workflows in the images

#

so you just pick the image that you want to recreate and drop it in comfy

#

so what I do is have a subfolder in my output folder for saves, if I like something I put it in there so I can reuse that workflow

fervent thunder
trail lion
#

I think they are saying they want to save old prompts as text files

white parrot
trail lion
#

SD is stable diffusion, that's not a GUI, maybe you refer to automatic1111

white parrot
#

idk but i think yeah...

fervent thunder
#

yes they mean automatic1111 probably

white parrot
#

i'm just installed comfyUI and know that it can use same model as SD/auto1111 so idk if it can do like what i asked too?

trail lion
#

in any case, if you get used to the workflow in the images that's actually easier to search later

#

visually

#

and swarm also lets you save workflows separately as kind of like a favorites

white parrot
#

the IT display look hard to me at first place, but i'll try to learn

fervent thunder
#

I think they are saying save prompts
you can do this with a node like Save Text File from was-node-suite

trail lion
#

try swarm then, it's like a middle ground that still uses comfy

white parrot
trail lion
#

generate a few images with different prompts and drop them back onto comfy, you'll see what I mean

stark stag
#

Hi guys!

quartz cliff
#

Hi everyone

balmy hemlock
#

hello, anyone confortable with comfyUI + mask nodes ?

silk condor
#

Just bought an 3060 with 12 GB for my seconday PC ( host pc ) so that I can setup my AI and Stable Diffusion on and run it locally on my network 😄

#

I am excited 😄

past furnace
#

automatic 1111 just broke suddeny. was producing images okay, all of a sudden its all buildings? any checkpoint/vae I use, all buildings no matter what prompt

silk condor
past furnace