#💬｜general-chat

1 messages · Page 158 of 1

dusk scaffold
#

The problem is that, as well as creating a mediocre product, they hide the good work and the rare gems, creating a false sense of scarcity. They will fail because they have a mediocre product and only invest in marketing, but the small works will never gain notoriety because they don't have enough money for marketing. I'm generalizing the way things are in many sectors.

fervent thunder
#

its spam pls report it

#

dropshipping is a common spam topic

#

not sure why but it is

fervent thunder
#

why dosent stable diffusion use a consistent voltage
im bouncing between 780mv and 850mv
maybe the it/s would be faster if it was not unloading and reloading

trail lion
#

Stable diffusion doesn't use voltage at all, that would be your GPU

fervent thunder
#

okay my gpu is running stable diffusion

#

which is needing voltage to generate an image

fervent thunder
#

oh interesting, wow

#

im just watching it now, the voltage stabalized randomly

#

after 40 SVD 1.1 generations

grave veldt
#

Can someone help me ?

fervent thunder
#

with what

grave veldt
#

can you help ?

lusty beacon
#

What is the best model from juggernaut and dreamshaper

grave veldt
#

idk man

#

i just wanna be helped 😭

lusty beacon
grave veldt
lusty beacon
#

Why is it so hard to get good pictures locally in compare to leonardo even it's based on sd

grave veldt
#

@lusty beacon Pinokio is asking me to put ''--skip-torch-cuda-test'' where do I put it?

lusty beacon
grave veldt
#

i put it in webui user

#

in app

warm junco
grave veldt
warm junco
grave veldt
#

yes

warm junco
#

with "add python to path" checked?

grave veldt
#

yes too

warm junco
wise ginkgo
#

3080 ti vs 4070?

keen kraken
#

hi im getting this error when i try to upscale on forge, TypeError: 'NoneType' object is not iterable

craggy kindle
#

Can anyone help? i keep getting this annoying ass error in Forge. Happens randomly, happens especially whenever I switch models. But also whenever I ty and generate randomly, sometimes it renders fine, other times it just says "waiting 1/1" forwver and I check the console and it says this "CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1. Compile with TORCH_USE_CUDA_DSA to enable device-side assertions."

#

happened in Comfyui as well

#

doesnt matter the resolution I do either

#

Tried adjusting my pagefile too with no difference other than it crashing even more often

#

si I just set it to auto

#

had it set to 16gb-24gb as someone suggested, which made generations undoable

#

I wish these error messages wwre more helpful, like i have no idea what it wants me to do

craggy kindle
#

idk, although ive had that same error before, forgot what I did

trail lion
keen kraken
craggy kindle
#

mhm, idk

#

I dont upscale cause of my low vram

#

i used to use hires fix but I dont see the point anymore

#

oof donr you just love reading a forum on how to fix the issue and thy just tell you oh just do this debug thing and youre good without explaining how to do said thing. then you have the last commwnd asking to have it explained, which was a year ago

#

or you just find an empty forum post with nocomments. Ughh fuck

ruby drift
#

I'm confused as to how install new models. For example, when looking at https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0, what do I do to install it?
Do I only take the sd_xl_base_1.0.safetensors file and be happy with it, or must I bring the schedulers, encoders, etc., in their respective folders in my ComfyUI models folder?

#

I feel like I only need the safetensors files, but I'm not 100% certain.

cedar salmon
ruby drift
#

Great!

#

I had it installed on my old machine and the performance was... yeah...
I just installed it on my regular PC and it's night and day! (gtx970 vs 4070ti)

cedar salmon
#

let the memes begin

loud solar
wise egretBOT
#

Someone clicked the button!, please don't do it again 😂

wise egretBOT
wise egretBOT
wise egretBOT
wise egretBOT
wise egretBOT
wise egretBOT
ruby drift
#

We'll that crap app is an easy "block" target.

wise egretBOT
wise egretBOT
ruby drift
wise egretBOT
loud solar
ruby drift
#

Ok. The reason why I took the base one is I've seen an example of how to use that and they basically were using the base one, then the 2nd one on top of the result to refine it.

loud solar
ruby drift
#

Oh, so even for SDXL I don't need to do these two steps?

loud solar
ruby drift
#

Alright then. Thanks! I'll try that when I get in front of my PC later.

#

(I should open the firewall so I can do it from my phone 😛)

loud solar
#

I'm trying to stay uptodate but it's going into to many different directions and technologies ... but it's fun!

ruby drift
#

Hehe. It seems that way

keen kraken
#

anyone know how to use the hi res fix

#

like which upscaler to use

desert dagger
keen kraken
#

im getting used to forge and just making basic prompts with the character loras i have

desert dagger
fervent thunder
#

is this A1111 workflow? in that case real-ersgan is fine

#

its still good

#

if this if comfy or diffusers or something like that
this is a recent list:

ultracompact:      153.56 fps (0.0065 seconds)
compact:            82.92 fps (0.0121 seconds)
span:               60.46 fps (0.0165 seconds)
realcugan:          34.58 fps (0.0289 seconds)
esrgan_lite:         6.54 fps (0.1530 seconds)
omnisr:              4.89 fps (0.2043 seconds)
plksr:               2.65 fps (0.3770 seconds)
realplksr:           2.21 fps (0.4522 seconds)
esrgan:              1.97 fps (0.5083 seconds)
swinir_s:            1.07 fps (0.9375 seconds)
atd_light:           1.05 fps (0.9536 seconds)
srformer_light:      1.05 fps (0.9545 seconds)
swinir_m:            0.69 fps (1.4509 seconds)
hat_s:               0.44 fps (2.2763 seconds)
swinir_l:            0.39 fps (2.5610 seconds)
srformer:            0.27 fps (3.6405 seconds)
atd:                 0.27 fps (3.7223 seconds)
dat_2:               0.27 fps (3.7284 seconds)
hat_m:               0.23 fps (4.3972 seconds)
hat_l:               0.23 fps (4.4004 seconds)```
valid fulcrum
#

Hi everyone 👋🏼, looking for some AI image experts to join a soon to be launched virtual companion app. We're building an AI companion app to cure loneliness and help people grow and feel heard. Our app allows users to customize their AI friends and we're very close to launching. Our goal is to have hyper realistic images of AIs including our items. We're using SD 1.5 as a base model. Looking forward to meeting people who want to join part-time, as freelancers or core team members. DMs are open! 🚀

fervent thunder
#

lower on the list is higher quality but slower

#

higher on the list is lower quality but faster

#

you can see ersgan right in the middle for context

keen kraken
desert dagger
fervent thunder
#

ersgan is 1 step 🙂

keen kraken
#

i keep getting errors when i try to upscale from a generated image, so i have to run it with the hi res setting

fervent thunder
#

if you mean for the second diffusion pass, at least 20 steps

desert dagger
keen kraken
#

i started using forge because easy diffusion takes a long time

fervent thunder
#

thing with upscaling is there are so many options

#

and you can chain 2-3 together

#

if you get comfy you can access some of the fancier tiled upcales like mcboaty

#

or you can go for dedicated models like SUPIR or DiffBir

#

or any of the 15 or so options in the table I posted

keen kraken
#

whenever i try to upscale from the generated image i get this error

#

TypeError: 'NoneType' object is not iterable

fervent thunder
#

would suggest moving away from forge

#

I think comfy ui or diffusers are the two really good methods

#

I am including swarm in comfy

keen kraken
#

i like how fast forge is because easy diffusion would take up to 5 mins with just a 15 step image, but one thing is i dont like how the lora work, on easy diffusion they worked like drop downs that you can add

#

with forge it just uses the whole lora:1 thing and i dont really like that

fervent thunder
#

not too sure about forge

keen kraken
#

what ui do you recommend

fervent thunder
#

comfyui definitely

#

if you like code then diffusers

#

but even then comfy is nice

#

diffusers is more of a long term thing

keen kraken
#

how does the lora adding work

fervent thunder
#

you can put a node

#

that has a box to click and add lora

#

then you connect up the nodes

keen kraken
#

comfyui sounds weird

trim magnet
#

all of the ui's add loras the same way lora:1 is just that on easy diffusion you dont see that on prompt but is still there same with comfy

keen kraken
#

ill give comfy a try

desert dagger
fervent thunder
#

comfy has something like 100x more tools than any of the others

#

I wish diffusers had more tools but it is far behind

#

diffusers has the best code layout

keen kraken
#

how fast is comfy

fervent thunder
loud solar
keen kraken
#

how do i work multiple loras into a workflow, ive seed something about stackers

valid fulcrum
#

Hi everyone 👋🏼, looking for some AI image experts to join a soon to be launched virtual companion app. We're building a virtual companion app to cure loneliness and help people grow and feel heard. Our app allows users to customize their AI friends and we're very close to launching. Our goal is to have hyper realistic images of AIs including our items. We're using SD 1.5 as a base model. Looking forward to meeting people who want to join part-time, as freelancers or core team members. DMs are open! 🚀

desert dagger
zenith valley
#

cure loneliness

hoary kiln
#

🗿

pastel surge
#

I still havent tested again though

autumn oyster
#

where do i learn how to make good prompts to get really high quality, realistic pictures?

shadow oyster
#

Hello everyone

opal hedge
#

Has there been news of the 8B model's release window?

fervent thunder
#

💀

nova lily
#

How can you make XYZ save the images individually with & without out the banner? CHEERS
EDIT
include sub grids is the option

eternal violet
#

ultrapixel results are amazing
anyone else playing around with it?

nova lily
eternal violet
nova lily
#

oh im not comfy gang

nova lily
eternal violet
#

not sure honestly, believe it's a highres model
i havent used a1111 in over a year so not sure

amber bloom
noble ledge
#

Hello everyone 🙂

I'm thinking of ordering an AMD mini PC with a Ryzen 7 7840HS (https://www.gmktec.com/products/amd-ryzen-7-7840hs-mini-pc-nucbox-k6?variant=7530f28e-cce6-4e22-ac92-e7999375a6be) and I was wondering if this APU can handle basic LLM (Olama) and StableDiffusion (A1111) models at "decent" speeds (response times after prompt).

I looked everywhere for LLMs and SDF benchmarks for AMD's APUs on-line but I cannot seem to find any.

All I could find was setup and usage guides.

Does any one have any benchmarks on AMD's APUs for LLMs and/or SDF -or- knows of any places on-line that I could look for such info?

fervent thunder
fervent thunder
fervent thunder
#

would strongly recommend only using an Nvidia GPU for machine learning

#

if you don't have the budget to get one at home then

#

you can design your comfy workflow on your home PC and

#

rent a cloud GPU for $0.20 per hour to run it

warm junco
low moon
#

who cares anywya, let them fix the hands this time before release, there a new model weekly anywya sigma hunyuan auraflow...

fervent thunder
fervent thunder
#

do certain checkpoints produce better results with SVD ?

#

instead of 25 frames should I just try 14 frames twice

frigid hull
#

Hi. How much video memory do I need if I want to deploy it on my own computer?

fervent thunder
warm junco
craggy kindle
#

how do I use SD Couple?

wet grotto
oblique jay
#

What exactly is Pony for SDXL and why do I see so many checkpoints referencing it over anything else?

#

Are there specific use-cases for it?

fervent thunder
#

loras that specify pony need pony

#

usually

keen vault
oblique jay
#

I see, that makes sense.

#

I was wondering why I was seeing it for many things.

#

Lots of LoRAs have Pony/SDXL specifications, though I've noticed the Pony ones are mainly NSFW lol

fervent thunder
#

does anyone have instagram they share their art? ill follow u

#

also any tips to improve SVD results?

#

are certain images generated with certain checkpoints going to make better motion SVD

keen vault
#

Could anyone share dataset for pony xl style? I need some practive cuz my first style lora was a dud.

keen kraken
#

hi guys, i want to ask why some sampling methods just output absolute nonsense colours

fervent thunder
#

or not enough steps, or too much steps

keen kraken
#

1024x768 is to high

#

and 50 steps

fervent thunder
#

for some older checkpoints, yes

keen kraken
#

DPM++ 3M SDE Karras

fervent thunder
#

or its just not what the checkpoint was trained on, need to view documents of the checkpoint

keen kraken
#

thats what i used

fervent thunder
#

usually the checkpoint page will recomend a size

keen kraken
#

hmm i should probably check that

#

i just use Eular A because its really consistent

low moon
#

So Ultrapixel is here

fervent thunder
#

they also usually include example images with prompts, ext

keen kraken
#

decided to try other stuff

fervent thunder
#

is it gonna get added to the main comfy?

low moon
#

🤔

fervent thunder
fervent thunder
low moon
#

idk all i know is its takes 10 gb of hard drive space and 10 gb of vram

#

so

fervent thunder
#

i got 7900xtx

pale latch
low moon
#

yes

#

that is correct

fervent thunder
low moon
#

Funny how AI is becoming more humanlike by the day while people become more robot like

#

lol

fervent thunder
keen kraken
low moon
#

obviously

fervent thunder
#

the more time I spend with AI the less horrific it looks lol

low moon
#

ts improving

fervent thunder
pale latch
fervent thunder
#

the image generation bots are better than the text genration ones thats for sure 😛

fervent thunder
low moon
#

well at least u apologize thast a big plus

keen kraken
low moon
#

i grew up before online life became prevalent, i imagine someone born into this time when most interactions are thru screens raher than face to face will have a different mindset and approach to coimmunication

fervent thunder
low moon
#

lol

fervent thunder
low moon
#

i got internet'd in 2000

pale latch
#

I had a neat idea inspired in a dream state last night. it's probably been thought of already. Game AI right? so it was really good for a bit with games like FEAR , where the enemy squad would actually plan out strategies and execute them on you. Or Black & White, where you could teach an AI creature how to behave in the world. These were heuristics and decision trees more or less. Powerful stuff but it's not something that's trained, it's more engineered and crafted to purpose.

In fear , the levels would have to be constructed in a way that these decision trees could operate on. In B&W it was crafted specifically to that game's world design and structure. These are powerful systems but they take a high level of skill and research to engineer. Those kind of software engineers largely left game development and went into other fields where crunch culture was not as prevalient thoguh, and implementing crafted heuristics into games because expensive and unweildy, hense we have shit AI in modern games still.

So what if we train a model to craft these systems. I think game ai needs a boost and all this generative research should be in games by now. Modern game AI sucks so much

junior urchin
pale latch
#

i've seen llms in games. it always sucks and is just an llm saying what it thinks you really want to hear

#

my favorite has been the reverse turing test

#

also, i've never liked the idea of using a cloud based AI that requires an always online connection, for a game. the killer game ai is going to be baked to purpose and running locally

#

also, companies will likely want a system that protects the ai model that runs locally. they don't wanna drop a full llm into a game folder for hackers to repurpose and make loras for. that's why i'm thinking studios could use a trained model internally, that creates classic heuristical decision tree ai's instead. then they publish those systems.

craggy kindle
#

Icant get regional prompting to work properly wth loras, tried the same format the tutorial specified, added BREAK etc and no luck ugh

#

wish therr was a prompt template for eag thing

#

would make it make sooo much more sense in my head

#

instead of trying to teach you by using weird words only ai mathmaticians know what they mran without explaining how to get thre. Like oh just add a /function to the command and bam. But it never tells you what you to do add /function to command if you catch my drift. Just assumig that you know what theyre taoking about. really irks me. bleh. Tried SD couple too with no luck

pale latch
#

let me load up an see what i can do

craggy kindle
#

oof it says to use latent mode, thats helpfu, but then they fail to mention how you would write a prompt with two seperate characters, where to place the BREAK etc

#

just "do this" and done

#

"Using Multiple Character LoRAs with Regional Prompter
Characters often blend when using multiple LoRAs representing different characters in the same image. Regional Prompter allows for distinct character representation in the same image.

However, there are important considerations:

Use Latent mode, not Attention mode

If your generated images are degraded it may be because your LoRAs are corrupting the resulting image. There are some suggested fixes in the documentation. I've found lowering the weights of the LoRA and using ADetailer at a higher weight to fix the faces works well"

#

I wish less tutorials were like this ughh

#

Has a bit of good info, but fails to address any examples or comparisons in how to do it properly, what to do add what not to do

pale latch
#

theres also mask modes and prompt modes to creating regions which is kind of neat

#

Those need BREAK between region prompts, but that's easy. From left to right usually

craggy kindle
#

isnt addcol doig the break tho?

pale latch
#

yeah you can do one or the other. i just find it easier to use ADDCOL ADDROW instead

#

my regional extension is crashing the ui . probably have to update things. i can't help today

nova lily
#

I hate inpainting, because I have a horrible experience with it. Is it just that im expecting miracles? I understand that egregious errors cannot be salvaged and that not all inpaint checkpoints are made perfectly.

pale latch
#

i think a lot of people want prompt to image, and maybe an upscale. and thats fine. inpainting is a process that can be really powerful and used in many ways. it's fine to want to do things without it though.

it's trickier and can have more fail cases. using a proper inpainting model or a controlnet or something , soft inpainting, all sorts of tools in the box, it can have great results. Also bad results.

hard flax
sleek scarab
#

What is the local (uses pc power and free) BEST AI TOOLS?

fervent thunder
#

ok I think certain checkpoints are desgined for animations

#

like for SVD

#

gonna experiment with this I see in my grand collection of checkpoints theres at least 12 with the work "animation" in it

#

I never inpaint I just use detailer nodes

#

which are basically inpainting for you

desert dagger
fervent thunder
#

is there a prompt to create a "pop up image" or would I need a lora

#

oor I guess just use gimp and 50% transparency 😛

fervent thunder
#

god people hate AI art so much 😛

#

but deep down inside I know that they are the same people who were screaming that calculators wont be around forever

#

AI is just another tool

#

it can be trained on your own work, and made to do things outside the bounds of imagination

#

it takes an artist to use stable diffusion, its not just like they way bing makes it seem

#

no! your ideas are not valid unless you spent 20 years and 100k in art school to be able to know how to do a specific unique style! only then could you possibly comprehend the ability to imagine something!

#

is it "Cheating" that we print the bible on text and its not hand written by a scribe

#

The dao says that water is like truth, and water goes into the low places, where people hate it.

floral umbra
sleek scarab
sleek stream
#

is comfyui still the fastest ai image generator?

fervent thunder
#

depends on what ur trying to do

sleek stream
#

well just in general

#

im deciding if i should use a1111 oder comfyui

fervent thunder
#

ja vohl

brazen hull
#

Hi

#

Could you please advise on how to create a video featuring an AI-avatar? If I only provide images and a written speech, is it possible for Stable Diffusion to transform that into a video, including talking, dynamic gestures and movements?

desert dagger
desert dagger
pale latch
#

the new expression transfers and audio to expression systems. thats how to do it now

#

i need ot play around with those

native cairn
oak latch
#

can i pay someone to help me wit controlnet openpose hand gestures this shit confusing ash and i cant be bothered wit it

languid ice
#

how do i move stable diffusion into another drive>

fervent thunder
languid ice
#

but idk how to run loRA's

#

do i js put it in the models fgolder

#

then loRA?

fervent thunder
#

it doesn't matter where

#

what matters is what you say in extra_model_paths.yaml

fervent thunder
void nexus
#

生成

inland temple
#

how to install sd3

fervent thunder
languid ice
#

wait so i downlod models

#

and loRA's too

#

uhhh im lowkey confused 💀

fervent thunder
split hornet
#

supp guys im new here 👋

fervent thunder
#

hi

crimson veldt
#

agony ai replacing us
waow ai masked as "advanced feature" is an alternative way to draw

fervent thunder
#

its just like vaping, crypto, ext

#

just the flavor of the day to hate

#

once the nvdia stock bubble crashes and people realize chat bots arent going to replace doctors in 5 years they will find something new to hate

#

like electric vehicles

#

2024 is the year to hate ai

crimson veldt
#

like its been 2 years i have not yet seen a good ai generated story and people think its somehow going to take "creative" jobs....

fervent thunder
#

i have not yet seen a good ai generated story true

narrow kernel
#

With the models with much higher context sizes, you will. Think Mistral just put out a 10 or 11b model that can handle 128k context sizes. 99.99% of the stories you've read are derivative and contrived. Where there are patterns and formulas, AI can easily replicate. Model context sizes are really all that's holding it back.

fervent thunder
#

all of the jobs AI will replace have already been exported to overseas call centers

bleak matrix
#

Good morning, everyone! How are we all today?

static vessel
#

hi

old condor
#

Anyone has the unstable diffusion invite?

sour canopy
#

hi, AI world

fervent thunder
#

s .

#

🥰

old condor
#

Bro what ☠️

#

Meanwhile someone sent me a invite in dm:

#

You just joined to be annoying the fuck 🤣

#

Thanks whoever banned that guy :0

wet grotto
#

dont exist a small sdxl model?

#

smaller then 6gb?

trail lion
#

Sdxl is one size, 6gb vram may limit you to 1.5 model... Maybe check in the tech support room

wet grotto
#

ok thanks

pale latch
#

you can load sdxl in fp8 and get some memory savings. not all cards have support for this though and sometimes loras aren't compatible

fervent thunder
short kernel
#

I downloaded the fp8 model of SD3 and the portable version of comfy, and I also moved the SD3 model into the checkpoints folder, but when I clicked the eueue prompt in comfy, Reconnecting... or TypeError: Failed to fetch was displayed.

craggy kindle
#

oof I cant find a single tutorial kn yotuhbe about reuonal prompter and usig multiple character loras that Ivan understand

#

renderig a pic rn using adcol or whatever commands like someone suggested

#

and in latent mode

#

prvious pics just had judyhopps face on bothcharacters even with the weigbt lowered ugh

warm junco
verbal osprey
#

Has a better version of SD3 came out yet or are we still stuck with the bad version?

wet grotto
#

this base model sdxl turbo is very fast, it lost quality compared with sdxl default?

low moon
#

a bit

#

is it true that just throwing in key words is better than explaining things as if to another person... like (dog, couch, room, sofa), is better than (a dog sitting on a sofa in a room) idk, often i get better results just throwing in "keywords"

still glacier
#

depends of the model, for sd1.5 yes, but with sdxl and sd3 not necessarily.

pearl ermine
fervent thunder
#

T5, Clip G and Clip L each have different styles

#

for prompts

#

do you guys think creating ui design for mobile apps using stable diffusion is possible/viable

pale latch
#

start with design and then fill it with prompted generations to flesh it out instead. straight text to a purposed design? naw. That's ugly

fervent thunder
pale latch
#

yeha. diffusion models are bad ui designers

fervent thunder
#

do you know anything at least "okay"? even if it's outside of stable diffusion context

#

I don't need the best

#

just use material design?

#

it won't offend anyone

fervent thunder
#

never mind

trim magnet
desert dagger
fervent thunder
desert dagger
fervent thunder
desert dagger
fervent thunder
desert dagger
fervent thunder
desert dagger
fervent thunder
#

here is a link to material design

#

https://m3.material.io/

#

its basically Google's UI

#

its open source you can use it

#

there's lots of community-made collections of material design style components too
like this one:
https://github.com/Templarian/MaterialDesign

#

its got a whole github category
https://github.com/material-components

ocean jacinth
#

I'm fairly new to stable diffusion, would anyone be able to answer some quick questions for me regarding some image generation i am trying to do

split hornet
#

when testnet

fleet raft
#

Can I use Stable Diffusion Input Image and Inpainting to transform my store design ?

So my plan is to transform my father's store into something luxury and cosy vibey as a surprise and present, but the prompting is not working. So I was thinking is it possible if I insert my father's shop images and give Stable Diffusion some prompts and tell me to customise some part of it ? Is it possible ? I am using Fooocus btw because it's easy to use

vast fiber
#

Hello everyone, I am a newbie in this field. I'm currently using the Stable Diffusion web UI to create images, but I couldn't find the sampler 'DPM++ 2M SDE Karras' in the sampler list. How do I install this sampler?

quartz elk
#

noob here, hello there

lethal iris
#

Can I use Stable Diffusion Input Image and Inpainting to transform my store design ?

So my plan is to transform my father's store into something luxury and cosy vibey as a surprise and present, but the prompting is not working. So I was thinking is it possible if I insert my father's shop images and give Stable Diffusion some prompts and tell me to customise some part of it ? Is it possible ? I am using Fooocus btw because it's easy to use

loud solar
wet grotto
#

what is the difference between resizy by a image on I2I and use Extras?

wet grotto
#

where can I find CoAdapters?

tough tree
#

auto1111 has prompts from file or textbox how would i get that on comfyui?

floral umbra
#

Hoi, do you guys know of a small gradio "sorting app" for loras that can better help sorting loras and lycoris and the like in proper folders like "people, characters, clothes, costumes" and so on? Like adding "categories" in the program which is making folders, and program shows civitai info of it with pics, and you tick what they are in the GUI?

fervent thunder
#

you don't want a GUI for that

#

use a text classifier model and then a shell script

#

and add a config file where you can write the config

pine fiber
fervent thunder
#

HI folks, kind of a philosophical / practical question here -- I have a feeling that I am missing out, or not properly making use of AI -- I use it to answer a lot of questions on a variety of topics, and I also use it to help me as I am working on programing projects b/c I dont speak python etc yet -- so I tell it what I want and it spits out code at me. What Else should I be doing to utilize ai properly?

river sparrow
#

Learn python yourself

#

that's one of the best suggestion I can give.

#

Image generating AI is easy, you can see what you get. Code generation is a whole other story, cuz you cannot review the code in a single gaze. You either need to know lots of testing and debugging methods, which mostly require you to understand at least some code and Best Practices for coding.

fervent thunder
#

Working on learning python -- very much so

river sparrow
next bear
#

I'm unintentionally learning javascript, C and python from linux

river sparrow
#

most learning is more effective when there is some kind of intention

#

surely you must have some kind of intention, otherwise why would you be learning it?!

next bear
#

I wanted to learn how to linux from arch linux

#

Which included the language I just mention

river sparrow
#

what do you mean by "learn how to linux" ?

next bear
#

I nuked my windows yesterday after 6 months of linux

#

and 6months of google searching how to do things

#

It worked out very well

#

I'm now proud to use arch, btw

river sparrow
#

yeah, linux has a nice learning curves. Easy start, lots of info available, lots of depth to dive in.

next bear
#

The wiki is super nice

#

Fish shows the error and how to fix it

#

Best shell hands down

river sparrow
#

Yeah, I think I just recently (somewhere last 2 years) switched to fish

next bear
#

I need to port over my dotfiles from my other setup.
There's so many functions on the fish config.

fervent thunder
#

Foot is a good terminal

next bear
#

I like kitty

river sparrow
#

im using urxvt as terminal

#

Slackware + Openbox + urxvt + fish

#

no start menu or task bars for me, just some keyboard shortcuts and the rest from a terminal window

next bear
#

I'm trying to remember how to make sddm show on an external monitor instead of a built in..

tropic frost
#

okay.... i know i will probably sound like a total noob on this, but can someone explain to me what 'workflow' does for stable diffusion?
is it just the overall steps people take when they go generate their images, or is it something else?

next bear
#

I use conda to make the environment for s.d. 1st

river sparrow
#

workflow is a comfy ui thing. You can use all kinds of nodes, and connect them to eachother. The collection of connected nodes in Comfy is called a Workflow.

next bear
#

Why not just use basic autmatic 1111

river sparrow
#

ComfyUI workflow can be stored in the generated images, so you can drag-and-drop and image onto the comfy canvas to load the settings (workflow)

#

Cuz I like tinkering and building new stuff

#

And I worship the Noodly Spaghetti Monster

next bear
#

Speaking of. I just did something fun for anonymous email

pale latch
#

i dont' like using conda and just use the venv command. conda environments always seem to break and harm themselves.

#

i hate dealing with python and it'll be so great once a non python library for this stuff gets traction

tropic frost
pale latch
#

dependency hell is bad UX

river sparrow
#

I've not tested&validated, but its claimed that Comfy is faster than A1111.

#

Forge might be just as fast

#

yeah, python really is dependency hell. But even then I prefer a single installation - much duplication of data is just stupid if you ask me

next bear
#

guerilla mail (dispossable) > @duck.com(route mail hidden) > @outlook
.com (send 2fa to dispossable) > @gmail.com

now I just need to figure out which sms recieving works

fervent thunder
#

speed is kinda not relevant cos
the others can't do anything close to a big comfy workflow

floral umbra
pale latch
#

those claims are really hardware dependent. my 4080 doesn't get many gains on the different UIs. But then we get into situations like sd3, where for some reason a1111's implementation sucks and can't unload the t5 tenc before generation begins. comfy is faster there on account of not filling the available vram

#

if you've got 2000 loras, delete 90% of them.

fervent thunder
#

I don't use loras at all

#

not sure if I should

river sparrow
#

yeah, time for lora clean up

pale latch
#

they can be fun to play with. like little plug ins to a model

river sparrow
#

yeah, I'm happy with my custom 3d fractal loras

fervent thunder
#

I sort of feel like
training a full checkpoint where you also train the text encoders is a better way to go
but its expensive so I totally understand why people don't just do that every time

pale latch
#

for sd3, as i understand things, since the transformer blocks have a text network built into them that runs parallel to the image network, training the tencs there are pointless. I've done loras where i train the first two tencs but my gpu can't manage training the t5. Either way, the resulting lora is different but seemingly not improved

fervent thunder
#

I am kinda waiting for the final finished SD3 8B release

#

and for them to confirm that that is the final release

#

cos SD3 2B has a weird latent space in some ways

#

not sure how many SD3 gens I have done but something like 1,000
and it acts kinda weird in my experience

pale latch
#

i think 2b has potential yet. I love it's efficiency. I don't think i'll be able to train loras or run 8b even.

#

from the sounds of it, to fit 8b into 16gb you have to quantize it, which it doesn't like to do and lowers it's quality substantially

fervent thunder
#

yeah

#

I have to keep in mind my perspective is different

#

cos I use cloud

pale latch
fervent thunder
#

it also means I am more tolerant to slow generations than most people
cos I spin up additional instances while I am waiting

#

that's why I like stupid slow samplers and upscalers but other people don't

#

the weird latent space gets on my nerves TBH

pale latch
#

i like to iterate quickly and take pleasure knowing it's my pc beast pal that's doing the thinking and making the art

#

different strokes

fervent thunder
#

I started doing img to img from SD3 to jugger or dreamshaper
and then doing the rest in SDXL

#

the reason I like it slow is
I got interested in upscaling photos before I got interesting in image generation

#

and in the upscaling world a 40 minute wait is normal

#

but it turns out I was wrong about stable diffusion sampling, the fancy ODE stuff just doesn't work well

#

DPM++ 2M is fine

pale latch
#

i think dpm are a family of ode solvers but this is way out of my knowledge domain. i need to do some good sampler studying. i have a surface level knowledge

fervent thunder
#

DPM are still ODE solvers yeah

river sparrow
#

I need to train my 3d fractal on it

fervent thunder
#

but they are not traditional ones like Heun

#

DPM solves the linear part using algebra
and then the non-linear part using numerical methods

#

euler and heun etc are super old

river sparrow
#

euler rules

fervent thunder
#

I also only have a surface level knowledge of sampling

#

but I feel a bit better as I read like 100 papers this month on it

pale latch
#

im eager to learn again. for years i've been an energy drink addict. like 3-4 a day was normal. so i quit those couple years ago and the past year has been a huge struggle to do anything that requires focus or thought. turns out i'm ADHD and the doctor gave me ritalin . we're trying it for a couple months anyways. now focusing works again and i gotta plan out a learning strategy since it isn't hopeless anymore. hurray drugs!

fervent thunder
#

haha I went down the same path, from lots of coffee to medication

pale latch
#

used to be a smoker too but i quit that and thats when the heavy caffiene ingestion began

fervent thunder
#

I smoked only 5 times only

pale latch
#

well, ramped up. i've always sucked back coffee

river sparrow
#

hmm, i still drink coffee. and smoke some calming stuff in the evenings.

fervent thunder
#

remember that if ritalin doesn't work there are still others to try

#

its a bit genetic

pale latch
#

yeah totally. i'm open to exploring shit. ready for a new learning adventure

fervent thunder
#

my main other tip is just to exercise loads
even if its just walking

#

it doesn't always work but sometimes when medicine is wearing off, a workout can revive it

#

exercise is unreliable though

wild tulip
fervent thunder
#

just as a bit of advice
you need a bit more wording on your websites
to make it more trustworthy
like have an About Us section and describe the site

#

oh never mind you have a larger site https://tunehq.ai/tune-studio

#

thanks

primal shuttle
pale latch
fervent thunder
#

haha yeah sunscreen is good

pale latch
#

nothing disfiguring just skin issues that annoy over the years. plus a cool song where the guy just speaks as the song goes on

fervent thunder
#

lol yeah I happen to know the song

pale latch
#

i'm doing a lot of diet consolidating too and trying to cut sugar where its excessive and eat less processed foods. i guess all these healthy changes i've been making these past few years have derailed the existing coping mechanisms my body had for adhd. long undiagnosed cause i had coping mechanisms.

sorry to rant bout it but ive alwys thought this stuff should be more talked about, now i'm in the shit and i feel compelled to talk about it.

fervent thunder
#

it kinda makes sense yeah that if you had coping mechanisms and change derails them then it would make it harder
even if the change was healthy

pale latch
#

didn't realize my bad habits were self medicating lol

#

~~~ topic transition special effects wowolulwolwulwulouwlu ~~~

the new pixart model that expands the parameters to 900m. it's a refiner that adds 300m more parameters, like the sdxl refiner. the first stage for low frequency compositional structure to the image and the second stage for high frequency detailing. its a neat idea but i've always thought it didn't work for a couple reasons. mostly because it was sort of just a poor implementation of the concept with sdxl, but also becasue the community had no idea about what they should do to refine the models separately. so they just refined sdxl to do the high frequency detailing and nobody bothered with the refiner at all. Without the tools to understand both the weight sets, the extra parameters that sdxl had from the refiner were rendered moot.

interested in seeing how this new implementation of the concept with pixart 900m will evolve. https://huggingface.co/dataautogpt3/PixArt-Sigma-900M

#

the ensemble of experts i think is the moniker

craggy kindle
#

any cool loras for Pony?

fervent thunder
#

I kinda feel like pixart started out so small that

#

whilst the 900m project is great

#

its still way too small

craggy kindle
#

like cool effects, bodysuits, mecha, android, cybernetic enhancement loras?

fervent thunder
#

sorry I don't use lora

fervent thunder
#

1.8B (1,800M)

#

he doubled it lol

#

the OP is the guy who makes Simpletuner

pale latch
#

yeah thats crazy. they doubled it again lol

pale latch
# craggy kindle any cool loras for Pony?

lol i just went to civit, filtered "highest rated loras lycoris and doras for pony model" ... oh god. i thought i could pick out a couple cool ones from the top 50 but oh god. oh god!

#

i dont know why i'm so surprised tbh. its pony

gilded drum
#

Why Stable Diffusion become bad?

outer crescent
#

bad quality of images or the initial harsh SD3 license?

gilded drum
#

SD 1.5, I never using that big SD in my local machine

pale latch
#

The community is really toxic towards stabilty lately is what he's perceiving. Civit banning them, people quitting and forming new aliances to compete with stability, memes about sean parker, all sorts. Stability are not bad guys, there's just a lot of guerilla marketing against them. A lot of over reaction about the SD3 license i believe was intentionally dramatic to be subversive

#

like you know how soccer players pretend that they're fragile so the other team gets punished?

gilded drum
#

Sean parker.. oh, from Facebook movie

outer crescent
pale latch
#

He's bringing sexy back. Yuh

gilded drum
#

Justin Timberlake

pale latch
#

oh , you didn't mean any of that and you meant sd15 is now running poorly on your gpu?

gilded drum
#

It's follow my mind. If i expect it not work, the the result will fail.

pale latch
gilded drum
#

W-O-K-E in my language are translated as Awaken

pale latch
#

"The Social Dillema" covers all that problematic stuff

gilded drum
#

not that pride thing

outer crescent
#

i....haven't a clue what on earth you are angry about whatsoever. the model size? disney movies? prompt interpretation from text to image?

pale latch
#

woke in america when in context isn't about lgbt pride nonsense. it's about being aware of societal issues that are larger than your immediate situation, yet are still influential on people's lives. There's a lot of intentional misunderstandings being spouted off about it. Trying to confuse the message.

Calling it "that pride thing" is immediately pretty sus

gilded drum
pale latch
#

and to be clear, when i say lgbt pride nonsense, i mean the nonsense is those who are offended by it happening

gilded drum
pale latch
#

was going to try to help you but ugh. gGUHh. pretty sus all a suddent.

gilded drum
#

Yes. I am shaman, so, little out of the reality world

#

shaman, in real terms. Like American shaman

#

Still finding new candidate for next shaman maybe. I don't know.

outer crescent
#

who is against "the pride thing." so a shaming shaman. politics and podcast based opinions aside, what on earth has any of this to do with stable diffusion text to image models?

low moon
#

🍿

gilded drum
#

The text inside the picture or prompt is the problem (from my shaman side), from code side, yes, it's not work as intended as the code. I count maybe roughly 10 updates before the last one are the one that is works

#

Never mind, just that what i want to say.

pale latch
#

the way i understood shaman training is it's a lot of butthole stuff while they are exploring the depths of their being. like real shaman stuff

gilded drum
#

Oh, yeah. Can you provide specific code that explain all image that generated are says "money" or "do it" without say that explicitly? meaning, image that generated are valuable and protected for private use in the next update? I am sure the developer know what I mean by this.

outer crescent
#

(icey btw, i was mocking this in dms, not just sending you gibberish :D)

gilded drum
#

Like PHP, there is $ in PHP. Something like that

outer crescent
#

shaman, my brain is broke with you, you seem like a crank, yoda english aside. luck best you to, wing right job nut

fervent thunder
#

yoda english asideLMAO

gilded drum
pale latch
#

civit selling nfts now. profile cosmetics that are limited issue

gilded drum
#

llama. That means OLD in my language. As a muslim, shaman, blogger, programmer, fansuber, report and documentation, and hard worker and very respect guy, this event makes me headache

pale latch
#

llama means a donkey with a tall neck in my language. they spit on you too. i think ultimately , the name comes from it being a llm which sorta sounds like llama if you read it

gilded drum
#

I am really trying to give back people privacy. If It just for me and some specific people, no problem. But it's worldwide.

pale latch
#

that replicate node is just a node that accesses the replicate api i think. its not actually 405B llama3 running in comfyui lol. cool node though if you want to subscribe to replicate

gilded drum
fervent thunder
#

yeah I don't rly want to sub to replicate

#

I only like groq rly

gilded drum
#

Dev, whatever you code and your open source business, I am just user in this context

pale latch
#

have you tried 405B yet? it just came out.

#

i've heard a lot of hype over the months for it

gilded drum
#

something that impossible is possible for me

#

bye guys, thanks for the chat

floral umbra
#

Is it possible to train lycoris with a workflow in comfyui? Or is kohya needed for that? Or are there better "lora/lycoris" alternatives with easier tools available today

outer crescent
#

you can make output for lycoris i guess

#

but dont know about a direct pipeline as part of a workflow. do you mean make images, train images at the same time? ive no idea what kind of compute that would use but i dont think it would be effecient

floral umbra
#

Making them

#

Wait.. does this mean rc2 isn't compatible just because it's a hotfix? HAhaa

20:48:09-222092 INFO     Kohya_ss GUI version: v24.1.4
20:48:09-230099 INFO     Python version is 3.10.0rc2 (tags/v3.10.0rc2:839d789, Sep  7 2021, 18:51:45) [MSC v.1929 64 bit
                         (AMD64)]
20:48:09-231100 ERROR    The current version of python (sys.version_info(major=3, minor=10, micro=0,
                         releaselevel='candidate', serial=2)) is not appropriate to run Kohya_ss GUI
desert dagger
pale latch
#

theres a comfyui node that does lora training but it was updated last 6 months ago

floral umbra
pale latch
floral umbra
#

Aye, but kohya is throwing a fit atm, it's incompatible with anything it seems

21:08:40-282941 INFO     Kohya_ss GUI version: v24.1.4
21:08:40-290948 INFO     Python version is 3.10.0 (tags/v3.10.0:b494f59, Oct  4 2021, 19:00:18) [MSC v.1929 64 bit
                         (AMD64)]
21:08:40-292951 ERROR    The current version of python (sys.version_info(major=3, minor=10, micro=0,
                         releaselevel='final', serial=0)) is not appropriate to run Kohya_ss GUI
21:08:40-293951 ERROR    The python version needs to be greater or equal to 3.10.9 and less than 3.11.0

Just installed 3.10

#

I'm blind.. 3.10.9 needed

outer crescent
#

21:08:40-293951 ERROR The python version needs to be greater or equal to 3.10.9 and less than 3.11.0 thats the silliest error. Must be 180lbs or bigger. but no bigger than 180.5lbs. im guessing its autogenned from dependancies 🙂

fervent thunder
#

oh wow

#

someone added an implicit variable step second order solver

#

if my understanding is correct this should be the best

#

they also added Strong Stability Preserving Runge-Kutta

#

but I am not sure about that one

#

https://github.com/wootwootwootwoot/ComfyUI-RK-Sampler this is the node

low moon
#

Today's popcorn segment was very quirky and fast.

fervent thunder
#

the server is kinda chill now

low moon
#

The storm has passed.

#

so can anyone here who has tried everythign rate the current models: hunyuan/pixart sigma/auraflow/kolors

fervent thunder
#
  1. kolors
#
  1. auraflow
#
  1. pixart sigma
#
  1. hunyuan
low moon
#

interetsing i onyl tries sgima and kolors, kolors is better and very fast

#

tried

fervent thunder
#

kolors looks fantastic to me

low moon
#

its def what i was expecting sd3 to be

fervent thunder
#

SD3 image quality is way better due to VAE though

low moon
#

yes sd3 would be incrdible if ti worked

lavish ermine
#

hi

#

guys

#

how are you guys doing

low moon
#

chilling

lavish ermine
#

will anybody want to join my cool server

loud solar
lavish ermine
#

it cool

#

my being active and chatting you can get your own role

loud solar
#

I'm so cool I don't need it 😄

lavish ermine
#

erm wtf

loud solar
# lavish ermine erm wtf

Don't feel bad ... I am here without any role ... I know you have to do your stuff ... but it sounds a bit like SPAM ....

sand flax
fervent thunder
#

models with better VAE are coming

worn salmon
desert dagger
low moon
desert dagger
low moon
#

theoretically tho one could pass a crap sd3 image over to sdxl withe vae decode encode it and do minimal denoise on it and inpaint and send it back and regenerate and so on over and over again eventually u could fix all the mutations

desert dagger
low moon
#

i actually like it its just so udnercooked

desert dagger
low moon
#

yeah emads skill

#

XD

desert dagger
low moon
#

if the models is good then why did they apologize and said theyll fix it

pale latch
low moon
graceful pike
desert dagger
desert dagger
pale latch
graceful pike
low moon
desert dagger
graceful pike
#

but yea im out again, nothing new, not a single finetune on SD3, still banned from civit, going great 👍🏼

pale latch
low moon
#

I do remember tho the bit about "the last model youll likely need"

desert dagger
desert dagger
graceful pike
#

me? no why should i be lmao?

#

SD3 is tho, sooo

desert dagger
low moon
#

aw leave wizard alone, every company has fanvbois

#

pats him on the head

#

its cute

desert dagger
graceful pike
#

brother in here harassing without end and constantly screaming reported

pale latch
#

#🆕｜sd3 message really once you know how to use sd3 you can do whatever your mind can imagine

desert dagger
low moon
#

supposedly?

graceful pike
#

Oh yea they opened it a few hours ago, great. I checked yesterday and it was

low moon
#

ITS UNBANNED?

pale latch
#

tbf it's still not part of their generation services and there are only a few finetunes that aren't really complete. but yeah alex seems a little bitter about sd3. holding onto some resentment

pale latch
low moon
#

boobs unchained!? nudity refiner!?

#

the flood gates have opened

pale latch
low moon
#

now im hopeful

desert dagger
low moon
#

thanks but its not my lora

pale latch
#

last i checked if you filtered sd3 and loras, i was still top of the results. king of the sd3 loras. ballin on the top

desert dagger
pale latch
#

winner by default but i'll take it

low moon
#

this is good news anyway, the improvements will be very fast moving forward

pale latch
#

aw yeah it didn't last long. balls get knocked off top ez. theres a girl lora at #1 now

low moon
#

i liked it too :<

shadow oyster
#

Who have a1111?

desert dagger
shadow oyster
acoustic dagger
#

Can you install Stable Diffusion on Debian?

fervent thunder
low moon
nova lily
#

POV youve been using an inpainting model for the last hour 🤡

graceful bramble
#

What is the best way to merge several loras at individual weights? I'm using several style loras at very low weights to generate a specific look and I'd like to make things easier.

eager shard
#

is there a way using ultimate sd upscale to set a custom width and height like the base upscaler or no? cause i really really hate useless upscale by

rancid willow
#

hi

warm junco
#

Best is to use a square resolution

eager shard
#

was doing 4k wallpapers

narrow kernel
#

The image isn't going to turn to pudding by doing a slight upscale or downscale with lazanco

eager shard
eager shard
narrow kernel
#

Bro that's already 4k... There's zero need to go higher

eager shard
#

i know? thats why i said slap in 3840x2160

#

ultimate is just upscale by u cant do 3840x2160

narrow kernel
#

because it's not in the right multiples of 16/32/64/128 or w/e multiples it has to be in

#

but ultimate can definitely take some 1080p image and do a 4x upscale just fine

eager shard
#

it'd be hype if ultimate ever makes it to stable cascade

#

idk if youd need it though u can just make it pop out 4k immediately

narrow kernel
#

none of the common models are trained for spitting out a straight 4k image without tiling. at highest, like 2k. sdxl/sd3 can usually handle up to around 1.5MP before breaking

#

(1280x1280)

#

some finetunes can handle around 1536x1536 in some situations

eager shard
#

yeah i got a couple 1536x1536

#

i like cascade unfortunate that no one touched it

narrow kernel
#

it's not worth it, make the base image in the 1024^2 to 1280^2 range and ultimate upscale it. there's a new sdxl union promax controlnet that does tiling really well with ultimate. you can also use the older sd1.5 tile controlnets as well for it. there's also supir for upscaling as well, but you need to make sure you have the vram for it

eager shard
#

i could get a better image just batch sizing a bunch of 4k images with cascade than ultimate upscale sdxl

fervent thunder
#

you can gen at 4k with hidiffusion

eager shard
#

ooo ill have to look it up

narrow kernel
fervent thunder
#

its the best thing for keeping SD 1.5 relevant

eager shard
#

is it better to use 1.5 with hidiffusion or just plain sdxl models

fervent thunder
#

if you combine hidiffusion with adaptive token dictionary and supir
you can easily get 8k from SD 1.5

#

it works better with SD 1.5 but it works ok with SDXL too

eager shard
#

so 1.5 hidiffusion makes better 1024 x 1024 than sdxl?

fervent thunder
#

deepshrink is another node that can get you a big boost

#

deepshrink will get you at least 2k on its own

#

1.5 and sdxl are completely different looks so I don't want to say that one is better than the other

#

the top 1.5 models are still very relevant

narrow kernel
#

telling you though the new union promax controlnet is a beast for tiling. i tested out making an 8k image the other night and had to do some crazy filtering in photoshop to find the seams

eager shard
#

i havent used 1.5 for a year so i have zero memory of anything for it

narrow kernel
fervent thunder
#

do you know ultrapixel?

#

that's the best high res thing

#

out of anything

eager shard
#

ultra pixel? no

#

ik ldsr

fervent thunder
#

ultrapixel is based on cascade

#

but way higher res

eager shard
#

idk if i need more than 4k lol

narrow kernel
#

you dont

eager shard
#

is there a point in going higher if u only got a 4k monitor?

fervent thunder
#

yeah

#

8k is gonna be standard soon

narrow kernel
#

it wont be

eager shard
#

ehhhhh idk about that

fervent thunder
#

I already prefer 8k

eager shard
#

8k seems only useful for tvs

fervent thunder
#

there are 5k/6k monitors as well

narrow kernel
#

there's a thing called pixel pitch and our eyes can only see so many per sq cm based on distance. the only way an 8k monitor would be an upgrade over a 4k is if it were 2x as big, but then youd have to sit twice as far away to see it all

#

its a dumb placebo sales marketing gimmick

eager shard
#

i can see 5k being like the absolute best thing for a 34" monitor but i doubt youd be able to tell a difference more than 5k

fervent thunder
#

its not a gimmick, you can check in a calculator that combines PPD and FOV

narrow kernel
#

shit like this

#

statistically, you wear glasses. you already have antialiasing built in

eager shard
#

so is ultrapixel just a better cascade? or is the base ultrapixel model bad

fervent thunder
#

I don't really like to call models better or worse
cos they are more like different flavours

eager shard
#

cause from what ive done so far cascade makes the best backgrounds but cannot do people at low compression

#

yeah but ultrapixel is based off cascade

fervent thunder
#

ultrapixel doesn't have more training subjects than cascade

#

if that's what you mean

eager shard
#

rip

fervent thunder
#

but its 6k and cascade is like 1k LOL

eager shard
#

i was pumping out 4k images w cascade

fervent thunder
#

its not just about monitors
if you ever want to make prints one day then prints can go up to 32k

#

for a 300 DPI poster

#

like in the photo editing world 8k-32k are not really considered rare exotic resolutions

eager shard
#

ehhhhhh the only prints ive ever done are for shirts

#

i got one of those press thingies for shirt at a thriftstore for $120

fervent thunder
#

this sort of print you have to go to print shop for

#

the models are strongest at 1024x1024 though, and any upscaling of any kind will lower quality in some way

#

exception is latent or tiled

#

but they have other issues

eager shard
#

ive only used latent ever ngl

fervent thunder
#

if you just want one easy way
use adaptive token dictionary

#

its the best non-exotic upscaler

eager shard
#

except for now cause im using ultimate

#

ultimate is tile, no?

narrow kernel
fervent thunder
#

this one:
https://openmodeldb.info/models/4x-RealWebPhoto-v3-atd
or this one if you need a bit faster:
https://openmodeldb.info/models/4x-RealWebPhoto-v4-dat2

fervent thunder
narrow kernel
#

and they are just using some shitty method to upscale the image to that size. at best, they're just using topaz to ai upscale it.

fervent thunder
#

you supply the image so its on your end rly

narrow kernel
#

ahh, well nobody is taking pictures at 32k... so same idea, they upscale some image to that resolution with topaz(in best case scenario)

#

oh jesus, yeah, that's a gigapixel lol

fervent thunder
#

they could tile photos

#

panorama

narrow kernel
#

true, but on a per image basis, absolutely not

fervent thunder
#

yeah there is no sensor like that

#

not sure what the current biggest sensor is

#

I don't rly keep up with the tech any more

#

but I know they hit 150MP at some point

narrow kernel
#

even still, at the hardware level, it's just doing some gimmicky bullshit

fervent thunder
#

maybe yeah

#

like the quality per pixel could be poor

narrow kernel
#

right

fervent thunder
#

at least with Sony their little 12MP A7S series had best noise handling for a long time

#

I used to be really into this stuff but I don't keep up with the gear and the tech any more

#

I am happy with an old nikon

narrow kernel
#

well i love my EOS rebel t7, does the job really well

fervent thunder
#

yeah low end DSLRs are fine I think

narrow kernel
#

they are, i've used a ton of cameras in my life. even some that cost 10s of grand(not mine, just got to use them for photoshoots)

#

the secret is to just get good at post like you're already going to have to do with expensive ass cameras anyways. any advantage they had is going to be removed by post lol

fervent thunder
#

yeah pretty much

#

I wonder if AI will make photo editing way better

#

at some point

narrow kernel
#

though lens quality DOES matter. i fing hate chromatic aberration in cheap lenses

fervent thunder
#

yeah I don't like the kit zoom lenses

#

on Nikon we have some nice value ones like 50mm 1.8g

narrow kernel
#

yeah some primes are the best

#

or if you get a zoom lens, get one with a small zoom, like 50-100

#

not some all in one 50-400 lol

#

unless it costs 200k

fervent thunder
#

oh yeah the giant zoom range is a red flag

narrow kernel
#

i did a ton of photography during college and i'm an engineer lol

#

go fing figure

fresh pine
#

hi everyone

desert dagger
nocturne hatch
#

where do i find clip skip??

desert dagger
nocturne hatch
#

idk i just got here 😂 😂

desert dagger
nocturne hatch
#

auto1111

desert dagger
willow cave
fervent thunder
#

twinkle.. twinkle.. stay inside.. big randy did not die...

loud solar
#

Also some deletes in general-with-images needed

#

THX

glass yarrow
#

has anyone successfully generated paper notes with text?

untold latch
#

Has Stability AI disclose any new project since the fiasco of SD3? Or they have given up since then?

low moon
#

Emadcoin anyone? XD

desert dagger
stiff ridge
desert dagger
stiff ridge
desert dagger
stiff ridge
#

ty

fervent thunder
#

but we have opensora 1.2

#

its 67GB VRAM though

#

but you can rent H100 for $2.50 per hour

#

just pls don't expect actual sora level output cos OpenAI is far ahead LOL

desert dagger
fervent thunder
#

oh I meant zero shot

#

maybe with vid to vid yeah

#

there's probably some animate diff workflow I guess too

desert dagger
fervent thunder
#

I stopped doing video stuff when SORA got announced

#

but since SORA isn't here maybe I should restart

#

on a podcast I heard a leak about how much VRAM SORA uses
one of the better closed-source Chinese SORA clones needs over a dozen H100s
so actual SORA must be like 20-50 H100s or something crazy like that

low moon
#

i paused video attempts too

#

Well get there, well get there...

desert dagger
wet grotto
#

hi guys, exist some sd3.0 checkpoint good for architecutre? I don't test any 3.0 yet, I wanna experiment some one

desert dagger
wet grotto
#

nowwww

wet grotto
#

I'm seraching for but I don't find it

desert dagger
wet grotto
silk briar
#

hi guys

#

im new in this

small silo
#

this is a pretty stupid question, but I know I have stable diffusion installed on my pc, but I've forgotten where I've installed it. Any ideas on how to find it? Thanks

narrow kernel
fathom marten
#

Stable Video 4D, wtf 😂

#

Wen 18D ?

narrow kernel
#

it's a model that does 3d stuff and animates, so 4d isn't a bad name for it

sage ravine
#

can i rename lora file inside folder or it will not work then?

narrow kernel
sage ravine
#

automatic 1111

#

i am talking about renaming in lora folder

#

lik from g_oku to something like goku

#

it will work?

chrome walrus
#

I need some help I am generating image base on text but in image character image will me mine

#

is it posible ?

trim halo
#

Anyone knows how to inpaint an image to an existing image? For example i have an empty room and i want to inpaint my own chair into the room? All i know is inpainting is txt2img

rich pewter
#

(Masterpiece), (Best quality), (Ultra HD), (Super detail), (Whole body :1.2), 1 girl, Chibi, cute, smile, flowers, outdoors, holding the camera, sitting on the roof looking out into the distance, with mountains in the background, amber, warm yellow, sunset, artistic sense, Quadratic style, white clothes,

narrow kernel
trim halo
narrow kernel
# trim halo thanks will try!

Yeah, I know it feels like a lazy response, but it's not something that can be quickly explained. There are a ton of different ways to do it and they all vary based on which app you're using for it.

floral umbra
#

When training loras for example ears, mouths, eyes individually, is it best to have a pic of the whole face? Or just need images that just has the whole mouth in the frame?

low prawn
#

What was the first diffused image??

#

Like ever..

heavy notch
low moon
low prawn
#

lol I thought the exact the same thing 😂

pale latch
fervent thunder
#

hi

#

guys i can't use automatic1111 locly because i hve python 3.12 instead of 3.10 and i can't downgrade how did you solve that problem

fervent thunder
abstract linden
#

I have a question, what exactly is Pony? It's clearly an important part of the SD scene, but is it like a kind of model?

#

I also notice many Pony prompts use something like; "score_9, score_8, score_7"
What is that?

eternal ledge
warm junco
abstract linden
eternal ledge
warm junco
pine fiber
#

GANs were invented in 2014

abstract linden
#

Hmm okay. So in essence, they are meant for general improvement based on already trained assets? They sound like vital, if obligatory prompt details for quality... source however I have seen less. A couple times. What does the "source" keyword typically look like in practice?

#

In terms of use in prompts

warm junco
abstract linden
#

Should one only use one of those at a time? Would it make sense to try something like "source_splashArt"?

#

Also tyvm for the explanations! This is all extreamly valuable data

warm junco
abstract linden
warm junco
abstract linden
warm junco
abstract linden
#

That's crazy! Still, great to know! Thanks for the data

#

I've learned so much in the past few days, I am actually starting to create art now!

#

At least, something I can be proud of hahaha

low prawn
pale latch
pine fiber
#

it was like a classifier model that they found made interesting images when they ran it in reverse

#

and tweaked the output neurons

#

for example it classifies a cat, then you change the "horse" neuron in the output layer and it becomes more horse like

#

that was the idea, but you can see from the results it had a kind of hallucinagenic look

pale latch
#

Right. That's still wild times though

fervent thunder
#

this site is good for compare models

#

but I cannot recognise all of them

#

extrarealisticxl_2s

#

hghd_play_enh_hd

#

does anyone know what these are?

pliant bobcat
#

hello i need help .how can i use sd

#

i coundt use it in discord

fervent thunder
#

do you want to use it in discord

#

or download it ?

#

if you want to use it in discord then you can go here #artisan-faq

#

if you want to download it then you can go to civit or huggingface

#

and search for the model that you want

pliant bobcat
#

oh thanks i download discord and i can use midjourney ,but sd i dont no

#

so what is civit and huggingface

full lark
#

This model is underrated, It produces better results than sdxl most of the time, and its close to sd3 , sometimes beating it (of course sd3 has better prompt comprehension)

pliant bobcat
#

iam from china so i use it is so difficult

#

something is limit

full lark
#

You can also generate images on civitai using some models if you want to test them

pliant bobcat
#

ok bro ,i try it ,see you soon

fervent thunder
pliant bobcat
#

so sd and civial couldnt use it in the private letter?

#

just in the a12345?

fervent thunder
#

I don't understand

pliant bobcat
#

i want use it in personal channel

fervent thunder
#

oh I see

pliant bobcat
#

just like midjourney in discord

#

you know that ?

fervent thunder
#

they said they will add the ability to do that

#

but not yet

pliant bobcat
#

oh i see thank you bro

full lark
#

But what you said its true, there are a lot of models out there now, I still haven´t tested kolors, hunyuan and auraflow

fervent thunder
#

I don't think its worth using Unet any more because they will be gone going forward
the only exception is Kolors
because Kolors is so fabulously fine tuned that it kinda makes up for it, for now

#

we are in a really weird transition period though
the whole market looks much better in a few months

#

once the current plans have been done by the various people

full lark
#

Any idea if its hard to install kolors?

fervent thunder
#

there's a comfy node

full lark
#

Aaa oki

fervent thunder
#

if you look on L2 discord they use it a lot there

#

they have some workflows made

#

also check out Ultrapixel if you haven't already

#

Ultrapixel is most high resolution model ever, it can do 6k

frail sonnet
#

My old Dell precision 3420 workstation with 16gb ram still works. Apparently I'd only get around $200 for it though. Is there anything I can do with it, to either help out my newer laptop, or alongside it?

I'm dreaming about llm aren't I?

Or should I just sell it and forget about it?

desert dagger
frail sonnet
desert dagger
frail sonnet
#

It even has Nvidia vram!
Nvidia NVS 310, 1GB, 2 DP
lol

desert dagger
fervent thunder
#

I'm an API user rather than artisan, but artisan is really cool

#

its like GPT 4 Dalle chat but it knows what control net is

stuck sun
#

can someone with a better functioning brain than me tell me what power connector i need for a 4090

dawn mulch
#

@frail sonnet You don't have a computer with more than 1GB of Nvidia VRAM? I was once in that situation and it was quite horrible. Couldnt do much AI except CPU for SD. Which took 5 min to over one hour to generate one image.

wet grotto
frail sonnet
dawn mulch
#

I was $200 short of a 8GB GPU so I grabbed the 6GB.

nova lily
#

What does raising the strength of a double prompt do, or potentially? like this (prompt, prompt:1)

shadow oyster
desert dagger
#

interesting observation - kling is using kolors for it's image generator

lilac flume
narrow kernel
#

Nvm found a couple articles mentioning it

narrow kernel
#

Smells like review fraud...

near stump
#

Hello, everyone, does anyone know how to use stability.ai to realize the operation of combining multiple images into one image, the image-to-image on the document seems to only support uploading one image

deft warren
#

I am trying to update from comfyui manager. But doesn't work

fervent thunder
#

I don't want to get banned so I won't keep trying

fair imp
#

hello gyz

#

is this a tech server

wet grotto
full lark
#

Anyone saw this paper?

#

"AI models collapse when trained on recursively generated data"

desert dagger
desert dagger
fervent thunder
#

Kolors license allows for zero commercial use sadly

desert dagger
fervent thunder
#

they likely paid for a separate enterprise license

desert dagger
#

be funny if all the different chinese companies were just different subsidaries of Byte Dnce

fervent thunder
#

might even be true lol

median swift
#

Currently doing a 2 hour render. Imagine the pain if it fails..

fervent thunder
#

lol

#

yeah its rough with long ones

#

my longest was 40-60 minutes

median swift
#

4000x4000