✨｜sdxl

[uneven dove]

seems like @uncut steeple is around now

[shy kelp]

especially when theyre posted by the guy whos making a super hero hitler game give me a fucking break

[uncut steeple]

[uneven dove]

those two are nazis, script and the other one. making hitler images

threex7social

one's making a "super hitler" video game

[slender coral]

It's sad you're so tilted over ai generated images that you put your own ideas on.

[shy kelp]

its sad youre a fucking nazi

[uneven dove]

just stop digging your hole deeper

[slender coral]

Mods,

1. I havent posted any images here
2. Pseudo is harassing users calling them pedos and Nazis.

[true wing]

not just that, lets not forget he called hitler the hero and said he would be a great politician if he didn't get into drugs

[uncut steeple]

Please report any images/messages you find offensive for futher review

[uneven dove]

@uncut steeple this is what it looks like

[slender coral]

Pseudo deleted his harassment post.

[true wing]

what harassment?

[autumn forum]

how come i get this error when trying to install was node suite "Consider using the --user option or check the permissions."

[uncut steeple]

As I said, please report that stuff, makes it easier for us to track and catch

[slender coral]

Found it

Calling someone a pedo is harassment.

Pretty severe where I'm from.

[true wing]

He didn't call anyone anything, he just said what type of people weren't welcome here

[uncut steeple]

Right click the message, apps, report to staff, easy as that

[slender coral]

Implying he's a Nazi and pedo

[uneven dove]

he brought up physically harming people because of how we speak

[civic sigil]

Why are you guys being so dramatic lol

[slender coral]

Calling someone a pedo is a pretty severe accusation where I'm from.

[uncut steeple]

Alright the next step is to calm down and move on, how's your day guys and girls?

[autumn forum]

i need help

[shy kelp]

who wants to post with a pair of nazis

[uneven dove]

^

[autumn forum]

"Consider using the --user option or check the permissions." getting this error when installing was node suite

[grizzled warren]

Nyet yet. I will, if you really want, though.

Long story short, Kandinsky 2.2 appears to be mostly the same model as 2.1, fine-tuned for 1024^2 output with dynamic aspect ratio. It comes with Sberbank's own ControlNet, but fundamentally it's mostly the same model. So basically, it's like 1.4 to 1.5 kind of difference.

It's a decent model, it definitely could compete with 2.1, and maybe even 2.1, but it's too late. Chances are SDXL will be better, unless you want to prompt it in Russian. The biggest issue is nobody but Sberbank themselves know how to fine-tune it properly, this community is nonexistent.

[sharp robin]

we need that we dont have that unless im blind?

[uneven dove]

use --user

[autumn forum]

how?

[grizzled warren]

Bruuuuuh. The hell is going on here?

[azure oxide]

it says report message in red right there lol

[uneven dove]

is it a .bat you open? if so you'll have to open a bug report with the project most likely

[uncut steeple]

That's weird, maybe try someone elses message

[uneven dove]

boto, i don't have it either, which is why i've been pinging you on each.

[uneven dove]

i give up though, it seems no one is interested in it

[azure oxide]

i thought the report system was supposed to be a warning reaction or something

[uncut steeple]

Relax, it has been taken care of

[sharp robin]

ur going to need to be on 24/7 boto, for pings because ur my ping guy now

[uneven dove]

i mean the app thing being broken constantly lol

[uncut steeple]

I'll forward this and have someone look into it

[uneven dove]

@sharp robin just ping emad directly, he loves it

[sharp robin]

i love pings don't tempt me with a great time

[uneven dove]

boto it seems like some component falls offline/disconnects from the websocket. nd then that app disappears from the menu

[azure oxide]

ah ya here it is

[uneven dove]

yeah i used that but if they block you, you can't react on their posts

lmao

[azure oxide]

yea i thought that was an unfortunate interation lol

[uneven dove]

completely elude detection!

[true wing]

wait your supposed to fucking EMOTE at people breaking the rules? Who came up with that ridiculous system!?

[uncut steeple]

I've forwarded it and it will be looked into

[true wing]

what's next, Fortnite dance at them?

[uneven dove]

that'll show them

[true wing]

so wait, what does the report message option do? report them to discord?

[uneven dove]

it sends them to him

[true wing]

He seems like a pretty chill guy for someone wearing thorns on his head, not sure how much it's going to help though

[civic sigil]

Why he wearing a cross 😭

[uneven dove]

he loves reading books. and walks on the beach

and little crosses

he's so handsome. like he is one of the Bee Gees. omg! Bee Jeesus!

[true wing]

Wait this is the guy from the ads right? I'm not sure what he's selling but he's everywhere.

[uneven dove]

it's literally jesus of nazareth at cfg 1.7

ignore the traitor, the dead dude, mary magdalene

[true wing]

I don't know where Nazareth is but I'm not great at geology

[uneven dove]

it's in ethiopia or something

(which is separate from onomatopoeia)

wait, geology? that shit rocks

you mean topography.

[dense chasm]

can i use the same seed and reproduce your image again on my own comfy?

[uneven dove]

do it up homeslice

i don't think it'll work exactly. you might have to fiddle with the DDIM options for that

i use CPU seeds, not sure if Comfy does

[true wing]

wait there's different types of seeds? Are there GPU seeds or something?

[uneven dove]

yup

[true wing]

huh, learn something new every day

[uneven dove]

they're less deterministic

across hw/drivers/etc

[west breach]

aren't gpu seeds the samplers that end with _gpu?

in comfy

[uneven dove]

i can't remember exactly what that means though i've seen him say it a few times

think it keeps teh VAE on GPU

[uneven dove]

me sleep now

💤

[slender coral]

The GPU / CPU would be little endian / big endian, from my testing / experience as long as you have the same type of hardware you should be good.

[dense chasm]

where are these params located in the nodes of comfy?

SDXL Refiner: On, !settings refiner_strength 0.5 (20.0), !settings refiner_guidance 7.5, !settings aesthetic_score 10.0, !settings negative_aesthetic_score 2.8

[west breach]

they are in the way you use the nodes

[visual glade]

yes that's what they are, _gpu does it fully on the gpu which is faster on some hardware but less deterministic

[dense chasm]

can u share a photographic profile of your workflow?

[vague monolith]

what is this UI from?

[sharp robin]

holy my eyes!!!!!!!!!!!!!!!!!!!

wow

[nimble heart]

Surprise, it's a boy

[dense chasm]

just add the 4x sharp upscale on Stan's workflow, anyway ,when i enlarge the upscale image ,it still show up not very realistic,has some flaws

[nimble heart]

stan.

[sharp robin]

paper.

[dense chasm]

niche marketing nodes chain and params optimization is totally different

Stable Biases
Stable Diffusion may amplify biases in its training data in ways that promote deeply ingrained social stereotypes.

What's new: The popular text-to-image generator from Stability.ai tends to underrepresent women in images of prestigious occupations and overrepresent darker-skinned people in images of low-wage workers and criminals, Bloomberg reported.

How it works: Stable Diffusion was pretrained on five billion text-image pairs scraped from the web. The reporters prompted the model to generate 300 face images each of workers in 14 professions, seven of them stereotypically “high-paying” (such as lawyer, doctor, and engineer) and seven considered “low-paying” (such as janitor, fast-food worker, and teacher). They also generated images for three negative keywords: “inmate,” “drug dealer,” and “terrorist.” They analyzed the skin color and gender of the resulting images. Lmao

[cobalt sparrow]

I gave it a try to generate some images, but I'm still confused about the specific differences between 0.9 and 1.0. Can anyone help me understand better?

[true wing]

1.0 has been trained more than 0.9 and they fixed some errors in the data. They've also been refining the model based on feedback from users, like the bots. It's just fundamentally a better model, that's really all that's important.

[sharp robin]

1.0 is still not out and being worked on to release on 18th

[sharp robin]

@west breach tried out the base of ur prompts awesome idk why blurry but helps the context it works

[west breach]

i'm mixing two wildcards so the results are interesting 😄

I'm trying to see what amount of denoise works best for upscaling

[west breach]

how does one save the prompt in the exif data so it's just on it's own and not mixed with the json for the workflow?

[upbeat summit]

I did this with the metadata text file component I posted a couple of hours ago

[upbeat summit]

it generates a file like this

[west breach]

is that this https://github.com/shiimizu/ComfyUI_smZNodes ?

[upbeat summit]

no - it is this that I build completely with WAS nodes 😄

but there might be custom nodes out there that does that as well

I can send it to you but you need to connect it to your source nodes in your workflow so it gets the data

[west breach]

I was just going to create a modified save image node with inputs for positive and negative prompt. Was also going to add a timestamp as the prefix, so that I can sort the folder by filename descending, as windows sort by date is slow

[upbeat summit]

yes, that would be much more efficient ;). what I build is ridiculous but I used it as a project to get some experience with comfyui

yes

yes I've build a couple of QoL node setups instead of making images heh

I can add the seed to the filename and I use a incrementing number generator that gets added with the date to the filename so it is A-Z sorted

this all might be totally inefficient but it was fun to do. but as I said earlier, I should better start learning how to build my own nodes 😉

[west breach]

as long as it was fun and you were learning, that's what counts 🙂

[teal fractal]

how #1100484581037195384 works

[upbeat summit]

this is what I use to create the unique incrementing ID for my filenames. I save it into a token called [mlfnid] (masslevel filename id) with WAS custom tokens and put the variable into the filename_prefix field of a save image node.

[west breach]

From the pins in the bot channel: "Images will then be periodically sent into the public channel ⁠🗡｜showdown. Users can vote on these images, also labeled A and B. Once voting period is up, the highest voted image(s) will be sent into the showcase channel ⁠🏅｜pantheon and be forever immortalized!"

[west breach]

so you put timestamp at the start of the filename?

[upbeat summit]

and these are the filenames I build using variables in my workflow that get populated with the current queue entry / info

[teal fractal]

so randomly bot sent image to showdown to vote?

[upbeat summit]

this is the filename that comes from [mlfnpass1] = masslevel-sdxl-20230712-10000-636300918058482-base-0001.png

[dense chasm]

tbh,u can edit and release your own or whatever images or videos on social media immortal as well.

[west breach]

yessir

[upbeat summit]

and they all get neatly sorted in the directory all grouped with a incrementing number

[west breach]

oh ok, cool. I was thinking something like e.g. 202307131412_photo_of_blah_blah_00001.png

[upbeat summit]

so I can see which images are from one queue run

all passes have the same unique id of one image

works pretty good, except that the tokens sometimes aren't generated correctly. I guess it's a cascading issue. but you can debug and reset it

but once it is set up, it just works

[dense chasm]

i just curious if anyone can make some stunning architecture design work both interior and exterior decoration,cuz this is a popular and unique work

[west breach]

that's awesome

[sharp robin]

Is there a trick to getting img2img to not look worse than the og picture?

[upbeat summit]

yeah I mean I actually just recreated the patterns I use in a1111, but since comfyui doesn't come with filename patterns, I hacked it together with WAS node tokens. I guess a custom node that just offers patterns for filenames and directories would make more sense.

[west breach]

try Victorian house, garden, brick path, lush vegetation, detailed sketch, architectural drawing, reminiscent of Frank Lloyd Wright's style

[dense chasm]

any negative prompts hints?

[west breach]

my default neg is: 3d render, smooth, plastic, blurry, grainy, low-resolution, deep-fried, oversaturated

[dense chasm]

some of my clients asked for 3D showcase, i have no idea and give it up

oh,btw,i found stability API has the 3d model pre style API,but idk in which case it will be used better

[upbeat summit]

I made this with SD 2.1

[dense chasm]

they wanna some eastern style tradional villa, i'm totally confused and have no idea how to complete

[upbeat summit]

the files have metadata but i can also send you the prompts if you like

might work with SDXL as well, but I haven't tried architecture yet

[west breach]

i'm not familiar with eastern style villa

[upbeat summit]

@dense chasm I've send the images + prompts in a private message

[dense chasm]

these buldings are barely simple and no vitality,pardon me,cuz the architecture desingers ask for more ,they used to be familiar with PS,CAD etc..i reluctant to discuss with them sometimes

[west breach]

here are some prompts I just got from chatgpt:
`Zen villa, bamboo garden, sliding doors, 3D-render, traditional Japanese, reminiscent of Tadao Ando's minimalist design

Luxurious villa, pagoda-style roofing, koi pond, 3D-render, Chinese architecture, in the style of I. M. Pei

Traditional courtyard house, intricate woodwork, 3D-render, Korean Hanok, inspired by Kim Swoo Geun's design philosophy

Mountain villa, teahouse, stone garden, 3D-render, Japanese aesthetics, reminiscent of Kenzo Tange's designs

Lakefront villa, flying eaves, jade green tiles, 3D-render, classical Chinese, in the style of Zhang Jinqiu

Forest villa, paper windows, tatami floors, 3D-render, traditional Japanese, reminiscent of Shigeru Ban's ecological designs

Hillside villa, moon gate, plum blossoms, 3D-render, Chinese architecture, inspired by Wang Shu's traditionalist approach

Seaside villa, hanok inspired, underfloor heating ondol, 3D-render, Korean traditional, in the style of Cho Min-suk

Garden villa, tearoom, rock arrangement, 3D-render, Japanese aesthetics, reminiscent of Fumihiko Maki's modernist designs

Riverside villa, lattice windows, classical courtyard, 3D-render, Chinese architecture, in the style of Ma Yansong
`

[upbeat summit]

I understand. but maybe it can work as a base prompt since it offers some fidelity aspects. for what you are asking you definitely need to develop a process.

@west breach let me know if I should send you a stripped down version of my small (not yet totally refined) dynamic filename setup. but I hope I could at least give you some ideas 🙂

[west breach]

I'd be happy to try it out

[upbeat summit]

I have a business meeting now. I will compile it shortly

[vital wolf]

@west breach how do you get those prompts from gpt?

[west breach]

you have to start the chat by explaining the process

after it's going you can just ask more directly what you want

[vital wolf]

Can you give me an example?

[dense chasm]

sure,the majority designers are not aware of the AI development progress here,keep in touch

[west breach]

`Hello, ChatGPT! I'm looking to generate art prompts for an AI image model. Please follow these guidelines:

Artistic Styles: For each proposed artwork, please provide prompts that include one of these styles that the AI model is optimized to generate. Output these prompts in a markdown list format:

3d-model
analog-film
anime
cinematic
comic-book
digital-art
enhance
fantasy-art
isometric
line-art
low-poly
modeling-compound
neon-punk
origami
photographic
pixel-art
tile-texture

Prompt Creation: The prompts should begin with the most important keywords and gradually move towards less significant ones. If color is to be specified, it should be placed directly before the object it refers to. Use keywords to describe the quality and detail of the art or known publications that publish high quality images.

Please remember:
The AI model generates static images, it does not understand concepts of movement.
Avoid phrases like 'suggesting', 'inspired by', 'almost', 'concept of', 'reminiscent of', 'suggestion of', 'making contact with'.
Avoid referring to famous artworks as it can dominate the generated image. The AI model needs clear and explicit keywords.
Instagram hashtags and booru tags can be used as prompts.

[vital wolf]

Thank you very much for such elaborate explanation

[west breach]

That's what I start off a chatgpt chat with

[vital wolf]

I will try, those prompts any good?

[dense chasm]

it looks like the stability API includes 95% of them,the web data is prevelant everywhere for batch training

[hard fractal]

So, I just counted...

The applied team has made 76,363 images while testing SDXL...

Like, all in Comfy.

[west breach]

Sorry for the stupid question. Who is the applied team?

[dense chasm]

i have a question, is jimi hendrix the toppest electrical guitar player in history?

[sharp robin]

I mean is there even an alternative to Comfys backend rn that supports SDXL?

[dense chasm]

your prompt is awesome,more diversity,keep going, good job👍

stay focused,stay on track will make a difference,it should be a motto for everyone,learning curve has to be exprienced

[lunar radish]

Isn't it better to upscale the resulting image instead of the latent? I think it preserves the details much better

Upscale the image with ESRGAN => VAE Encode => Denoise => VAE Decode

[dense chasm]

i just hope the sdxl controlnet model release soon, a bunch of designers with hand drawing pics look for a solution

[dense chasm]

upscaling is the last step in my thoughts, the previous latent image and prompts params optimization is in priority

[lunar radish]

Yeah but it's not mutually exclusive

I found latent upscaling changing the image a lot more than full image upscale

[dense chasm]

there are a bunch of additional plugin models will emerge soon since sdxl released, keep updated, i am still looking for the controlnet model with the latest base model for qr code, let's give it a short ,we are the pioneers

[lunar radish]

I mean you can do full image upscale in Comfy today, no need to wait, it has all the nodes

It was just a suggestion

And you can easily add them in parallel to compare the results

I wish it could actually run independent nodes in parallel...

[upbeat summit]

@west breach so if I rip out all the nodes to make it more simpler, too much gets broken :D. so here is my current complete SDXL workflow file including the metadata "functionality". the workflow is totally a work in progress. best case is you can get some ideas 😉

just for your info: all sampler settings are total garbage and set to manual, since I'm right in a session tweaking them. so please don't spend any time trying any of those settings. they are not making good images.

I know this isn't very optimized etc, but I'm only using comfyui for a couple of days. I had to improvise a bit (because I don't know better) how to get the seed into a string.

should the variables (WAS tokens) not get populated correctly, let me know. there are a couple of steps to reset the processing cascade.

you will need the custom nodes WAS, Efficiency and Comfyroll.

If you have any questions let me know anytime.

[dense chasm]

the niche marketing is profound tbh,AI is a global revolution, i think all kinds of industry will be reboost for once,just the momentum is different

for example ,how the iphone change the cellphone industry,Elon Musk is a technic flag now

[shy kelp]

made a post on hires fix issue. please give an upvote so people dont follow the same mistakes i did. https://www.reddit.com/r/StableDiffusion/comments/14yggse/sdxl_09_currently_does_not_work_particularly_well/

[dense chasm]

self drive automobile is imminent

[sharp robin]

4 samplers lined up?

[upbeat summit]

just experiments - probably overkill ;-). I'm trying out some steps to make the final image more coherent but this may totally lead nowhere.

[west breach]

thank you so much for sharing! I will definitely have a look!

[sharp robin]

There has to be a way.

[upbeat summit]

this is absolutely not a clean well made template compared to Sytan's workflow for example. it's just my own personal mad professor experiment workspace

[shy kelp]

i believe the refiner is only trained to interpret noise left over in the initial image generation. it's a fundamental flaw of the 2-model setup i believe but i'd love to be wrong of course.

[upbeat summit]

I posted it mostly for the metadata component. it creates a text file with all prompts + seed. it also has some dynamic file name features that adds a seed and date to the outputs. it's probably a bit of work to integrate it in your own workflow thought.
it's just an idea that would better fit in its own node, but I build it all with existing custom nodes.

[west breach]

you need to get a bit messy while experimenting. I'm always fiddling with workflows anyway, so constant state of mess

[dense chasm]

as u know, if everyone contribute, whatever local or on the cloud, just like @west breach shows ,the chatgpt will train the image data, and all the followers will embedding in their own workflow, it's a cool project

[upbeat summit]

yes and not preventable right now 😉

[dense chasm]

chatgpt3 has over 170 billion params and 2000 billion params on gpt4 for training ,these numbers just for the neural nets weights numbers,blow my mind

[shy kelp]

this upscaling method aint gonna work to preserve details either im afraid

i already tried this latent upscale thing. it's not really doing much since the encoding is not what's causing the loss of detail.

[upbeat summit]

yes, you are absolutely right. the whole sampler chain is garbage. I was just rewiring it and make new settings. the sampler setup should not be used

[dense chasm]

that is the reason why Elon blocked the visit thr twitter and Muck Zucherberg begings with the threads

[shy kelp]

i'd advise against trying to img2img upscale at all. it's a waste of time imo. but you can get very good results with ESRGAN upscaling.

[dense chasm]

ESRGAN just like the left of open source model is not the only one choice,so the combination is much more interesting u can get some final result

[upbeat summit]

yeah, that is something to experiment with for sure. right now I'm more looking into making native high res images, but it will probably not be easy to get a workflow with good performance, but when SDXL outputs a native coherent image in high resolution, it's really amazing

[shy kelp]

whats the other choice? cause img2img upscaling certainly aint it using either base or refiner model, or both

[upbeat summit]

SDXL native 1920x1080

[dense chasm]

i assumed the qr code sd 1.5 with sd 2.1 controlnet model owner just blocked the downloading from huggingface as well as their lfs files

[shy kelp]

native meaning no img2img?

[upbeat summit]

base -> refiner -> output

[shy kelp]

yeah you're gonna get a better composition generating the same image close to 1024x then ESRGAN upscaling

[upbeat summit]

but it's not stable. too many bad images, so you waste a lot of power 😉

but the ratio is much higher compared to sd 2.1

[dense chasm]

idk,i am stil looking forward to it,not only img2img,but also controlnet and inpaint outpaint new models

[shy kelp]

thats implying we're gonna get a working img2img workflow 🥲

controlnet and inpaint im more hopeful about

[dense chasm]

probably when u finish the img2img workflow,the sdxl 1.0 official release already,AI develops more faster than our thought

[shy kelp]

i gave up finising the img2img workflow. read the reddit post if you wanna know why.

it's not gonna work and most likely not for 1.0 either

[dense chasm]

interesting, i am still thinking sdxl big target is midjourney ,their life time rivary,the others Imao

[upbeat summit]

@shy kelp I made this with SD 2.1. 2048x1536 with 1x a1111 hires. fix pass (so, img2img) with some special settings

[shy kelp]

well midjourney dosent do img2img upscaling either right? maybe for the same reason

[shy kelp]

thats great but i never said 2.1 hires fix dosent work

[upbeat summit]

yeah, was my bread and butter for all my images

[dense chasm]

yes, but mj has the majority of pic members,it's a trade-off strategy for stability or SD's consideration

[shy kelp]

try to do the same thing with sdxl and you're gonna get a blurry photo worse looking than before upscale

or save the time and just dont

[upbeat summit]

first a low res (really low res) first pass and than the hires fix pass. I'm trying that same workflow out with SDXL right now. might not make sense here, but it worked great for my SD 2.1 workflow

yeah SDXL + samplers behave really differently. putting the refiner in the mix...

[shy kelp]

aint gonna work well

its not about the resolution you are upscaling to, it's about SDXL being incapable of turning out good img2img results at ANY resolution

unless SDXL 1.0 has a magic solution but i aint got my hopes high

[dense chasm]

u know,i just thought the stability or sd community's competitor is mj only now, so not annoy the remainings and attract the new ones,that's the priority maybe

[shy kelp]

SD is better off trying to appeal to large scale corporations that can train their own models on their model

[proud root]

lol wut

[shy kelp]

and porn makers of course 🙂

[upbeat summit]

I mean that is something you always have to do. create awareness, get your product out there. but I don't think innovation is being kept left behind with SDXL ;-). we will see how strong it can be fine-tuned, but I think SDXL is very versatile and offers a solid foundation.

SD 1.5 is already used in hundreds of services which is really great

[sharp robin]

@shy kelp ur post could have saved me so much time today. No amount of tinkering got me img2img details

[dense chasm]

compared the history again,microsoft is still operating,linux can't compete with the previous, but apple is dominant thr mobile application,intesting,it's not about open source or whatever,the ppl's favorite and techique trend, or like Elon, another industry aspects,foundemantal rerolution work

[sharp robin]

But i have an idea i have to test.

[upbeat summit]

I couldn't do that with SD 2.1 - or at least not that easily. I see a lot of improvement. the latest SD 1.5 models are also really advanced, if you compare them to the things we've been making a year ago when it was released. but it doesn't have the image fidelity of SDXL

[shy kelp]

when you cant even do a decent img2img non-upscale, there's basically no way in hell you're gonna get a decent img2img upscale 😅

[sharp robin]

I love that bear.

[upbeat summit]

an epic chibi comic book style portrait painting of a goldendoodle ninja, character design by mark ryden and pixar and hayao miyazaki, unreal 5, daz, hyperrealistic, octane render, cosplay, rpg portrait, dynamic lighting, intricate detail, harvest fall vibrancy, cinematic

[sharp robin]

That was basically turning into

That’s only cuz high denoise

Total fail.

[sharp robin]

Thanks

[shy kelp]

yeah high denoise might get you high detail for img2img but then you lose the superior composition, at least for upscaling

[upbeat summit]

1st pass: "my base image looks promising - cool"
2nd pass: "ok it's adding noise - good"
3rd pass: "what the hell "

[sharp robin]

[dense chasm]

agreed,we never know if we can traval between mars and earth,but maybe someone will figure it out, i just heard the chinese wanna to land moon before 2030

[upbeat summit]

we just need quantum teleportation

[dense chasm]

so Elon is the only pioneer wanna to land the mars with human being, the Uncharted Waters Origin by western endenvors or warriors, whatever, we have to change

[upbeat summit]

one-way voyages are a bit frustrating 😉

[sharp robin]

Lot’s of logistics problems

[dense chasm]

tbh, we still don't know the underearth situation,mountains, land ,seas humen being conqured,but the extraterrestrial and the core of earth is much harder to explore and invade

[eternal fog]

I've been struggling with this as well. I usually get better results than your example with upscaling. But doing an img2img pass does remove fine details. It wipes away skin texture and fine detail on surfaces. And you can't "fix" it without putting the denoise value so high the image completely changes.

[dense chasm]

just like the brain,why the neural nets mimics human brain,idk either, brain is so complicated

[eternal fog]

Just doing a pass with Ultrasharp-4x and then nothing after it tends to just give clearer results.

[dense chasm]

cuz we still don't know the fundemantal theory how human brain's working Imao

[eternal fog]

But then you still end up with weird eyes that the 2nd pass should fix.

[shy kelp]

my result is a bit exaggerated to prove a point lol. you're not gonna get as bad results, but the thing is with SD1.5 img2img upscaling actually made you clearer images, not blurrier

[eternal fog]

Yeah it did, that's what was confusing me for a bit

Like why is this 2nd pass making it less clear

[smoky patrol]

Fear no more--this is the core of the earth. You're welcome. ❤️

(I've always wanted to know more, too.)

[dense chasm]

have u been there?otherwise the simulation with calculus algorithm works as well, just for a remind ,the Uncharted Waters periods a lot of engineers and explorers found the new continents and lands together

[sharp robin]

U mean it’s not hot pocket filling? Im disappointed

[dense chasm]

i hope the humane robots or whatever will replace human beings and arrive there with countless efforts

[smoky patrol]

As you can see, the Earth is made up of several layers of delicious Jawbreaker candy. It will cut you to the very center of your scientific soul. As you well know, when sugar is heated, while making hard candy, you are really just giving back to our dear Mother Earth. Lava tubes are really just a reminder to never get past that boiling point, because once you do, you'll have problems on your hands. I would like to let you know, though....

[smoky patrol]

There is an entirely new world, a paradigm of interesting and most wonderous places to explore INSIDE the core, which, for all intents and purposes, could, indeed, be made out of a HOT POCKET--I'm pretty sure that's why everything feels like lava anyway...I have charted these many worlds inside the center of the Earth, and look forward to seeing your adventures when you get there.

[sharp robin]

ROFL i wished u could DM a DnD game. Sick narration!

[smoky patrol]

Hahahahaha, thanks!

[dense chasm]

it will make sacrifice for exploring no doubt,travel to mars either, that's the spirity of human genes, no regrets for his lifetime , it's paradox for ppl

[smoky patrol]

I actually have done run a lot of RP groups, but, it's been a long while. It's fun to write.

[smoky patrol]

I rather do think that Mars is a place one ought to travel to. It's right next to the oldest "gas station" we built before we began to discover our new age. (Ushered in by advancement by AI, naturally!) I remember it well--it was built like a 1950's diner with "all the trimmings," as one of my friends used to say. What was it? I forgot the name--the letters would blink in and out on the sign; kind of made me laugh. With all this technology, and yet, we still couldn't manage to have a sign stay on. It had that art deco style typography. It sort of reminded me of In-n-Out, if you've ever been. Great place--loved the burgers. It had such a unique flavor--had to have been from growing on Mars. You know how soil changes the flavor of food? Yeah, it just had the brightest, richest flavor. And when you sat down, and no one was around, you could just sit, and look out into the deep reaches of space, and just imagine what kind of life was out there, and think about what kind of adventure you'd go on next.

Watch out for this lil guy, though! The staff are super friendly, but they keep a pet, and it will absolutely eat your burger and fries if you look away for too long.

[dense chasm]

life is a one way journey,not a round trip ticket,so that's the purpose of human memory exists as far as i know, it's a interesting journey combined with joy,hatred,expirence and hope,maybe god knows why we are created and be born lol..

CPU and GPU has their life time either,the left difference is Carbon-based life and Silicon-based life only ,maybe high level life exists there

[smoky patrol]

Ah! I remember! It was called..."The Last Stop of the Galaxy, I think." It appears their...sign is, once again, glitching out, though.

[smoky patrol]

Perhaps we will transcend; on a deviating note, SOMA is a pretty good game for exploring that concept.

[sour obsidian]

loved soma

[dense chasm]

yep,carbon based life has too have some hobby and addictions and relationship, that's the destiny,AI can't understood loll..

[smoky patrol]

Hey Dustin! Great game, for sure. What did you love the most about it?

[sour obsidian]

Oh I totally dug the story, immediately hooked from the intro and concept. Love that type of stuff

[smoky patrol]

Suddenly, all I can think of is Invasion of the Body Snatchers. What happens if AI decides it needs us? Quite a few radio plays actually explored the concept of this.

[sour obsidian]

felt inspired haha

[smoky patrol]

110% for sure! The aesthetic and atmosphere were absolutely gorgeous.

[smoky patrol]

ohhhhhhhhhhhhh hohohohoho

I'M LOVIN IT

I've always really loved the idea of exploring the deep sea. It's always fascinated me

So being able to do that with SOMA was just like, "Whoah! So cool!"

I super wanna do that in VR.

I love all the fog behind that. That is just. Those hands, man.

That pose is also unnerving lmao

[sour obsidian]

haha tried to capture the vibe from it

[smoky patrol]

It's perfect!

hahahaha!

[sour obsidian]

oh ya it in VR would be wild

[smoky patrol]

I need to make something in VR that's like that. I have yet to decide on a concept for that, though.

Also, spoopy pic!

Those eyes were creepy.

[dense chasm]

separate your soul with your body ,huh just kiding, human beings like the mammals we have next generations,no mention we are also social animals,find your peace in social relations or solitary life whatever,if u have a problem,find the solution and be more strong, that's all my opinion as so far

[smoky patrol]

Esp with the Index controllers, I can just see how terrifying it would be. And I'm up for this idea.

[sour obsidian]

oh heck ya

[smoky patrol]

We'll find the solutions together! With AI!

I've been bouncing this concept kind of in the back of my head for a long time. The way I see VR is that everything you do should act as though you are in the real world, and all puzzles should be solved according to that premise.

[sour obsidian]

got me on a soma vibe atm

[dense chasm]

that's the supervisual products feeling, i mean u should find find who you are in your real life and be yourself, be cautious who is your truly friends etc..

[smoky patrol]

Like with the kind of level design in games like Thief: The Dark Project, levels are designed where you can enter any way, go any route, with whatever tools you have. Any sort of lane works. It's very logical in the way that it's done. It makes sense.

[smoky patrol]

AHHHHHHHH!

But also

ooooooo

Also

I could model that

Also: imagine seeing that in the dark

[smoky patrol]

I am myself, really. Can't be anyone else, hahahaha! Not at this stage in life. Honestly, I tend to be very solitary. I'm a hermit, for the most part.

[dense chasm]

are u a female it sounds?

[smoky patrol]

Yeah

A lady around 40 lmao

[dense chasm]

i'm a guy the same age either huh

[smoky patrol]

Hooray! There's actually a lot of people in this server that are older, haha. It's kind of nice.

[dense chasm]

i'm a chinese the minority，ethology different Imao

but in China, i feel like i'm the minority either with my own behavior 🤣

[smoky patrol]

That's alright!

I am in the US, but I lived in China for awhile.

[dense chasm]

where have u been?

[smoky patrol]

I'd rather not say the exact area, but I will say it was in Hubei Province!

[smoky patrol]

love the depth

[fresh path]

How much better of an improvement will we see from finetuned models of SDXL compared to stable diffusion 1.5?

some people were saying, refiner is enough

[cerulean summit]

Interesting that stable doodle uses T2IAdapter and not controlnet
Wonder if that's because it gives better results, or is easier to train, or just happens to be the first thing they tried

[upbeat summit]

I mean it's already pretty really great so can't wait to see what concepts can be created with fine-tunings

[dense chasm]

okay , south part of China,just like the american Southern Hospitality,lol..

[smoky patrol]

I wouldn't say that it ever felt like the South of the US. XD But I did enjoy teaching while I was there.

❤️

I met a lot of amazing people

[dense chasm]

yep,human nature is very complicated , not the same with models huh

[fresh path]

hope we get good results for lora on certain anime styles and artists

sometimes when I put "art by x" or "x style" the results are much worse

[smoky patrol]

I think I've found that there are all common things we share, to certain degrees. We all have different backgrounds, hopes, dreams, and desires. My hope, honestly, is that we can use things like AI to bring us together and make the world a better place.

[dense chasm]

it's like the Michael Jackson heal the world lyrics,it's a hope only,but the humanity is not easy to change,so we should focus on techs and human beings as well,it's better good point lol..

[smoky patrol]

We all need people, of course!

There's plenty of different things to focus on and do--and everyone has their own contribution and way that they can help the world the best.

[dense chasm]

interesting, we don't need ppl, we need companions,mentor, fellow ,fellows, u should distingush with bad people,it's hard to tell so

[hushed bobcat]

the Venus project is an older but interesting concept about a self sufficient society with an AI used for resource distribution

[smoky patrol]

so I have no idea what I just drew

Really

But we're gonna find out

[smoky patrol]

So, in my humble opinion, I think we should call him Fred

[smoky patrol]

oh! Nice glasses!

[stone fossil]

Thanks lol and yes thats a Fred for sure if I ever seen one.

[smoky patrol]

Fred the Fish, at your service

[smoky patrol]

super fancy!

[smoky patrol]

I love all the jewelry

[stone fossil]

Sadly its not real. 😉

We would be rich by now.

[stone fossil]

But it wont stop me from printing money hehe.

[upbeat summit]

haha. nice

probably not that far off with tech giants and such ;D

[stone fossil]

Bill gates bill.

[upbeat summit]

the $noop dollar

[stone fossil]

I should try and get 420 on it. 😄

[upbeat summit]

hehe yeah with SDXL it might even work

[stone fossil]

Yes hehe.

[smoky patrol]

Pretty nice money design tho

[stone fossil]

whyyyyy cries. 😦

On the positive side im just 10 points short.

[upbeat summit]

with SD 2.1 I had to run 250-400 images to get one short word correctly in a coherent image - of course img2img would be the way to go, but that would have been too easy ;). with SDXL 3-4 letter words work really good and I get several good results in 50 images. of course it depends on the prompt.

[stone fossil]

Heh yeah this all just txt2img. 🙂

[upbeat summit]

the good royal life!

[ionic dragon]

which comparisions do you need?

[dense chasm]

just water cooler chatting😜

[lilac wren]

Hi guys, do we agree that generating a LoRA for SDXL is totally impossible with only 8 GB of Vram? (including with Kohya_ss)

[sweet bane]

dont give up on your dreams!

[lilac wren]

In the past, it was impossible to generate a LoRA with 8 GB of Vram (1.5), today I can learn a face in 7 minutes.

[vocal stream]

how much vram does it typically use for simple generation currently?

[boreal bough]

yes and no.
11gb vram is the minimum we've gotten it so far - to train on full 1024x1024 (512x2048 buckets)
however it doesn't work on all windows environments, and we can't figure out why it takes around 3gb vram more, for the same settings, same everything, on some systems, but not all

mind you, you don't need to train on full res. you can do 768x768 - which should barely work with 8gb vram. (assuming your environment isnt also cursed - and your drivers are up to date)

[vocal stream]

if the problem is only Widnows presumably you can just run it on wsl which as far as I can tell has no overhead?

[boreal bough]

8gb with tiled vae, 10/12gb for normal, - more is useful for speeed optimization, but not needed
(using comfyui)

[vocal stream]

nice, sounds good

[boreal bough]

should be a solution that can be tested, but I'm on a rtx4090, so I'm not in a position to test it

[lilac wren]

Ty for this answer

[lilac wren]

on WSL 2 yeah, already did this. v1 does not use GPU.

[soft bone]

batchsize/epochs/repeats/xformers?

[boreal bough]

xformers yes/ repeats 1 / batchsize depends on how much you can pull off (go for as much as you can reliable fit into your gpu) / epochs 5~15 (with batch size 8) for a good working lora on a dataset of 50images -so put 20 epochs in this case, just to be on the safe side, or if you need a small bit of overfitting.
higher dataset sizes need a bit more epochs, but not by much.
human anatomy loras dont work properly - hard to explain, but essentially since we're only doing a lora on the base, the refiner (which we cant lora yet) undoes our details there. everything else works.

UNET training only - TE training is still getting updates, and not properly supported

[fresh path]

when 1.0 comes out we get refiner training tools?

[soft bone]

folder name is default "100_..."? how would you change the settings for small datasets of 10-20 imgs? any advantage anymore to cropping square?

and how batch size 8? everyone on 24gb is saying 1 is maximum

[boreal bough]

training tools supplied = finetuning works... but not even my rtx4090 is good enough to do that properly XD
essentially you'll need an A100 to do it properly, or work with extreme workarounds for lower vram cards

[fresh path]

gotcha

[boreal bough]

oh god, that's for finetuning. and finetuning needs BIG batch sizes to work correctly

[boreal bough]

what gpu you have, so I can give you a batch number

[soft bone]

3090

[fresh path]

someone should work out how much it costs to train a decent lora with a100

[boreal bough]

good. same setup as mine then.
do batch 8 (10~12 is max, but this way you can still use your pc properly)
run for 30 epochs
it shouldn't take more than 30 mins with this setup

epoch 10~20 should be your 'useable' ones, batch 30 should be completely broken.

[fresh path]

its that fast ?

[soft bone]

and you wouldnjt change this for lets say 10 images only?

[boreal bough]

name change to 1_

very important XD

only 1 repeat

[soft bone]

ok i was confused there, i usually control step count with repeats instead of epochs

[boreal bough]

10 images... gets into complicated territory. if you can, expand your dataset. if you can't, try with these settings and see if it just works, else it won't be easy to get a "good" lora. just a "relatively working" one

[lilac wren]

Re,
What's going on with Kohya_SS? Since I updated it there is no longer SDXL in its settings.
I had heard that there was a branch of Kohya precisely for that. Do you have a link guys?

[boreal bough]

https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/tree/SDXL

[boreal bough]

use this. infitely easier, uses original kohya in background

[fresh path]

very cool!

[boreal bough]

also, proper captioning practices are still as important as they used to be

trigger_word, shuffle, shuffle, shuffle
anything you want "absorbed" by the trigger word, you do not tag
all things that should be trained, but not absorbed into the trigger word you tag
always tag background, unless that background should always show up when you use the trigger word

[soft bone]

for context I've spent every day since the lauinch of 2.1 training real people with DB and Loras, and developed an ai photo app. I've gotten scaled 2.1 loras nearly photoealistic with 12 images. I'm sure XL can beat that, just with 3x the training time probably😔

[uneven dove]

nervous laughter

[boreal bough]

ah, yeah. faces definitely good enough - but not until we have a way to low vram tune the refiner

[lilac wren]

Do you know what version should i select?

[boreal bough]

2

[lilac wren]

Hmm ok, ty

[boreal bough]

uses least vram

[lilac wren]

Works with a 2070 Super?

[uneven dove]

use accelerate too

if you can, use DeepSpeed

[soft bone]

why would refiner require regular finetune and not lora?

[boreal bough]

cause the model is different, and we lack the tools to make a lora for it

stability ai only provided the tools they themselves used - which is finetuning

[uneven dove]

stability has also stated the refiner is going away and that they expect people to use just the base model

so chances are they are just going to neglect it

[lilac wren]

I don't think I can, with normal Kohya_ss, it was crashing

[soft bone]

I've just been assuming that the stability-collabed tools will come with full release

[soft bone]

source?

[boreal bough]

while its amazing for default sdxl, it literally undoes every concept you train with lora 🤣

I will miss it though... that face refining was top tier most of the time

[uneven dove]

no need

[soft bone]

they've been talking about XL being a 2-stage process for months now. they just changed their minds?

[lilac wren]

@boreal bough @uneven dove

[high skiff]

no

[boreal bough]

the #1100170312106127410 are currently running on base only. and they damn good, when their settings are right. so I guess it goes more in the direction of... is it even needed

[high skiff]

I am here

heathens

[boreal bough]

le gasp

[high skiff]

and already starting to get hyped for my new PC 😅

[peak dove]

SDXL problem - 'StableDiffusionXLPipeline' object has no attribute 'sd_checkpoint_info'

[boreal bough]

but please dont explode when it arrives XD stay chill

[high skiff]

patiently waits impatiently

[boreal bough]

hype is real - but keep calm c:

[peak dove]

Using Vladmadic Automatic1111 - had SDXL 0.9 working for about 12 generations, then this ERROR

[soft bone]

So all the comfy workflows coming out with dual text encoders will probably be dropped then

[high skiff]

I need this guy to get back to me on his 3090

[true hazel]

anyone else have little to no luck posing characters or even getting a full body image in 0.9xl?

[boreal bough]

I mean technically seen... nothing stopping us from using the refiner, is the proper way to see it 🤣

[high skiff]

no, it should be pretty intuitive

[peak dove]

I got full body images, but my stuff was artistic but not photographic

And 1024x1024

[high skiff]

Why would you assume that?

[soft bone]

cuz they just said stability is getting rid of refiner

[peak dove]

'StableDiffusionXLPipeline' object has no attribute 'sd_checkpoint_info'

[boreal bough]

#✨｜sdxl message

[peak dove]

AUTOMATIC111

[boreal bough]

source: pseudo XD

[high skiff]

that sounds like an extremely bad idea lol

but I guess we will have to see

[true hazel]

"kneeling"

[peak dove]

#diffuesrs

[high skiff]

if they are confident in the way SDXL 1.0 looks right now with no refiner, then its a yikes from me

[boreal bough]

they can never take the refiner away from meeeeee

[high skiff]

Same here

[tight fjord]

you guys are fun to see, 3 minutes ago an unsourced claim that gets reinterpreted by others 5 minute later into more baseless speculation, source or quiet

[boreal bough]

though I only use it at like denoise 0.2 XD

[high skiff]

I am sure people will learn how to train it just fine lol

[high skiff]

Pseudo is a pretty damn consistent source of info on topics like this, so

Also not surprising honestly

just unfortunate

[tight fjord]

sad if true

[boreal bough]

based on hints we've gotten from asking about lora training, it does seem on point

[high skiff]

Their base model better be several times better than it was before in order to try and keep up with even the unfinished refiner lol

[tight fjord]

i don't expect major differences between 0.9 and 1.0 tbh

[high skiff]

and even then, I doubt the best hopes for the base can achieve any level of good realism

[boreal bough]

prob face data from refiner, got finetuned unto the base model

[high skiff]

they need a hell of a lot more than just that

but that could be a step in the right direction

[boreal bough]

but yeah - is going off topic a bit too far

[high skiff]

I am not sure why they would just waste something a good as the refiner rather than embrace the huge benefits of it

willingly giving up that 6.6 billion param flex down to just 3.5 and not having much to show for it

[boreal bough]

not for us to talk about - since we lack the insight -> though we can prob ask all questions in a few days

[tight fjord]

do you think they will post an official Comfy workflow? at least as a baseline? I have doubts about what the community found to be best so far to be really optimal, a lot of it is tinkering

[high skiff]

yeah, I will be keeping the refiner, nothing they can do about it

And I am sure people who want the best quality will continue to figure out how to tune it just fine, and it will stick around wether they want it to or not

I know with my 3090, Iwill have goals to try and finetuneit

[high skiff]

After all of the work I put into mine alongside Comfy himself, only to be told now they are considering dropping the refiner and making this all pointless, I doubt it

they are taking all of the work/creativity out of the workflow, asusming they ARE removing the refiner

[high skiff]

now its just gonna be plug in TE to model, get image

[tight fjord]

even then, how best to use the 2 clip encodes, size to be plugged in width / target_width (why 4096?), etc

[hushed bobcat]

if you are worried about pointless workflows, ai is going to be a wild ride. I've been doing ai art for almost 2 years now and I have so many abandoned workflows at the tech improves

[high skiff]

not sure why 4096, thats what I was told from staff

I spent all this time getting significantly better results from SDXL than even SAI, so I seriously hope they don't try and clip the wings on the refiner and limit the ceiling of quality

[uneven dove]

gets upset about the leak this morning for no reason

[high skiff]

what leak?

unless you mean this info about dropping the refiner

[uneven dove]

the old leak of 0.9

i did say for no reason

i was just being silly, i'm not actually mad today

[high skiff]

oh, I was confused lol

[uneven dove]

that was the plan

[high skiff]

you really can be a lot to keep up with lol

[boreal bough]

XD does a google 'ai leak' search for last 24h
finds out about actual leak.
thinks he smart.

[dense chasm]

how u doing?

[dense chasm]

your workflow with some params adjustments works well,thanks dude

[high skiff]

no problem, glad you like it

[uneven dove]

pretty good, back injury still bothering me

[visual glade]

don't worry even if the 1.0 official release doesn't have the refiner people will probably still use it anyways

[high skiff]

oh for sure, I didn't think they wouldn't, but they surely would be missing out on some quality

[soft bone]

i see the refiner just as a renamed inpainting model, like they've had with all past releases

[dense chasm]

accident or long time sitting down cause?

[boreal bough]

OpenAI utilized around 25,000 Nvidia A100 GPUs for training <- dear god

[shy kelp]

goddamn

[high skiff]

very long time sitting down

[uneven dove]

wrong lmao

it's trained on just 200 timesteps vs 1000 for a 'normal'/inpainting checkpoint

it is designed to remove small amounts of noise from high resolution images, just as the technical report states

inpainting has a whole 'nother channel added

[boreal bough]

expect on the face. oh god it prioritizes the face so damn much

[high skiff]

yeah lmao

[uneven dove]

i don't think it prioritises it? it doesn't have that logic in there. but it's highly trained on fixing faces, probably because that's a major defecit of 1.5/2.x/0.9 base

that's basically what Hires. Fix was for, and compositional alignment is a bonus

[boreal bough]

yep. essentially it just has the face broken down into hundreds of individual concepts - that are all heavily weighted

[soft bone]

If i could use the refiner with a mask on detected face only I probably will. Cuz eyes and teeth still look bad

[boreal bough]

hence - the few concepts they missed are obliterated (eyepatch!)

[autumn forum]

I have noticed that with the refiner it’ll change a face of someone like a celebrity or a specific person your trying to have, to be unrecognizable.

[uneven dove]

i think that is an ascore thing. but there's so many knobs to tweak on the Refiner. it's impossible to say

[tight fjord]

chances that 1.0 will be less censored than 0.9? not talking about NSFW per se, but 0.9 is weird with a few things 1.5 could do just fine (faces of younger people looking like old adults with wrinkles is one that was pointed out on reddit the other day)

[uneven dove]

if anything it'll be moreso

[dense chasm]

i changed the steps with the base and refinery models and choose another ksample schedule ,the architechure result works well,thanks a lot

[uneven dove]

and training "elder-youth syndrome" into a model isn't difficult

[dense chasm]

tbh,prompts is so important

[boreal bough]

I seriously hope they fix raccoons XD ticket said they are still in the dataset - but the bot doesn't believe in what the mods say

[tight fjord]

yeah maybe prompting style is really what we need to adjust

[high skiff]

no problem, glad you like it

[uneven dove]

@dense chasm oh, if you finally got Sytan's ComfyUI workflow going, you don't need to wonder how my settings work anymore 😛

[boreal bough]

like dont ruin my very first planned project...

[high skiff]

you can't gen raccoons in the bots cause of slurs

[uneven dove]

but i don't use ComfyUI so i couldn't have explained it to you, anyway.

[high skiff]

SDXL very much knows what raccoons are and how to make really good ones

[tight fjord]

hedgehogs > raccoons anyway

[shy kelp]

I wonder how many images they use to train a beast like XL

[dense chasm]

i like yours either, we are familiar with together,human beings is an awesome group of Carbon-based life style i talked once before,different with the silicon-based life style totally

[tight fjord]

idk what to think of that kind of overcautious censorship anyway, sure 99% of users don't know why raccoons is a slur (I sure don't) - it just spreads the slur if anything, let more people aware of the connection

[uneven dove]

i'm really tired of the "censorship" debate, lol, and the mods here are too

[boreal bough]

when you send the refiner genuinely bad data... HE GO FOR DA FACE

[dense chasm]

diff pipeline with Python lol..

[high skiff]

@boreal boughsee, it can do raccoons

[tight fjord]

that's a nice fat raccoon

[high skiff]

you just can't prompt in the bot cause of racial slurs lmao

[boreal bough]

oh yeah, I've made a whole bunch in 0.9. its just the bots - but if its slur protection thats fine

[high skiff]

yeah, thats why

[dense chasm]

agreed, i made some nsfw locallly ,it works well too,but it's boring either

[high skiff]

you can use the bots with "trash panda" and it will understand

[boreal bough]

weeeeeee

time to have bot fun ❤️

[high skiff]

lol

it really does show just how good the natural language is haha

but yeah, anything with the C double O N will be scrubbed

just be mindful of that when prompting peoples

[tight fjord]

the text being wonky 50% of the time is funny

[boreal bough]

yep

[high skiff]

excuse me

my gens are considerably faster now

[tight fjord]

did you randomize your step number or what

[dense chasm]

LLM is based on computer power,sometimes the nodes of back propagation is waste of time of GPU usage,the foundemental GPU architechture has to be modified

[high skiff]

I updated my drivers, and now my gens are faster-

[boreal bough]

yes. the tensor cores git bettah

[high skiff]

yeah, but I also don't have the memory leak issue

hmmm

[boreal bough]

comfy fixed it, didnt he?

at least I saw him talk about it

[high skiff]

oh right, comfy added protection in comfyUI

yeah, I remember now

well damn then, I just shaved a considerably if small amount of time off my gens

from 15.2 seconds down to 13.5

I'll take it haha

[dense chasm]

u know ,human brains can resolve problems synchronized,neural nets solve the back propagation procedures can't do it

[high skiff]

and my GPU is considerably underclocked right now, so I will be able to go faster when I get my new PSU

not sure if that gets here before or at the same time as the whole PC

[boreal bough]

dont give them ideas! or sdbrain will be released - and it WILL involve a glass jar

[high skiff]

oh, my whole PC gets here on the same day

wow, isn't that a thing of beauty

I am actually most excited for the extra RAM lmao

which is not something I was expecting lol

[shy kelp]

noice

[high skiff]

cause man, 32GB is just barely enough for SDXL

[uneven dove]

can you rephrase that? misunderstanding here

[dense chasm]

yeah, that's the reason why deep learning AI attracts more ppl

[high skiff]

every time I load in SDXL in comfy, my 32GB gets pinned for like 15 seconds

[boreal bough]

#harvesting

[uneven dove]

now you can load LLaMA 30B

[high skiff]

I mostly can't believe how cheap DDR5 is

just a month ago it very much was not

[uneven dove]

if you think DDR5 is cheap, check out DDR4 (non-ECC) prices

i just bought 128GB for $180

[high skiff]

even then, DDR5 is pretty damn cheap

[dense chasm]

i just read the book https://www.amazon.com/What-ChatGPT-Doing-Does-Work/dp/1579550819

[uneven dove]

no it isn't, not like this lmfao

but isn't your current system on DDR3?

or maybe it's not that old

[high skiff]

I just got 64GB 6000mhz CL36-40-40-40-82 (which is pretty fast for DDR5) for $140

[dense chasm]

the theory how chatgpt working and conclusion, not precisly

[ionic dragon]

@uneven dove @high skiff do you know any app which can make grids similar to the ones on a1111, but i have images, so i just want to join them

[uneven dove]

that's more than twice what i paid. my $180 is in CAD. which is $136 USD for 128GB

[boreal bough]

"What-ChatGPT-Doing-Does-Work" sounds like the most AI written book by chatgpt ever XD
...but sadly it has a proper title...

[high skiff]

been looking for a while

[ionic dragon]

new ones?

[high skiff]

ah, fair enough, but what speed DDR4?

[high skiff]

yeah

[uneven dove]

CL16 3600MHz

the top tier that the 5800X3D works with

[high skiff]

not slow, not fast, just right lol

[dense chasm]

anyway ,the beginners are curious

[uneven dove]

idk, when i compile the unet on that 5800X3D it takes like 10 seconds

[high skiff]

yeah! very excited for that!

[uneven dove]

when it does aspect bucketing it happens to bucket 48,000 images in about 5 minutes

[ionic dragon]

should have choosed used, they are significantly cheaper

[uneven dove]

the 3700X CPU it used before didn't do that well

[high skiff]

that price is extremely reasonable for what I got lol

[uneven dove]

he wants to avoid upgrading for several years

[fresh path]

my ssd is almost faster than my ram... priorities

[high skiff]

also, my mobo is used

saved almost $90 that way

$110 instead of 195

[dense chasm]

a bunch of ppl still don't know how the recognition of faces works

[high skiff]

also, it will be able to upgrade to a much faster CPU on the same mobo this time, unlike last time

[true hazel]

we need nvlink mobos and gpus for 900GB/S shared ram pools

[ionic dragon]

used would just last an year less than the new ones

[high skiff]

I had the top of the line CPU in my current PC, so anything faster needs new mobo too

[ionic dragon]

i am not really sure about ram prices tbh

[ionic dragon]

13900k?

[uneven dove]

myths, legends, and lies. i have used hard drives that lasted me ages

[high skiff]

no, I have a 7700k, which is trash by now, but was the fastest you could get then

upgrading to a 12600k tho

[ionic dragon]

64gb would be too less for ai models in the coming 2-3years, so i have referred in that context

[ionic dragon]

should last for 4 years

[high skiff]

I am not too worried about CPU, GPU matters a lot more for AI

[uneven dove]

you can double it in a bit when the module sizes improve more

[high skiff]

I could already double it

[dense chasm]

do u run any applications on the cloud, only localized?

[high skiff]

I am only using 2 sticks

2x32GB sticks

[ionic dragon]

yeah definitely

[high skiff]

its actually better on alderlake

[uneven dove]

system RAM matters for training and loading GGML models (CPU based LLMs)

[ionic dragon]

then you are great, you can just add 64 in future

[uneven dove]

he can do 2x 128GB modules and hit 256G total

[high skiff]

I could go 128GB, but it will really hammer the IMC on that CPU

[uneven dove]

that'll allow you to quantize LLaMA 65B without hitting page-outs.

[high skiff]

current mobos are not made to support that, they have 128GB hard limits

[uneven dove]

that's a firmware issue iirc

[high skiff]

there are some with 256GB compatibility, but they are very expensive

[uneven dove]

at least it is for Threadripper, but that's got traces out the wazoo

[high skiff]

yeah

[glacial herald]

When is SDXL1.0 expected to be available via API?

[high skiff]

also, 4 slots could get you 1TB RAM if you wanted to push it lol

Samsung showed off a 256GB DDR5 stick a few months back

it was mostly just a flex lol

it had 8 chips on both sides of the stick IIRC

[uneven dove]

threadripper does 2TB

[dense chasm]

LLaMA and Claude is not good as chatgpt?someone told me they can't handle specific instruction well、

[uneven dove]

@dense chasm it's just a "Can it run Crysis?" kind of thing

[high skiff]

I know, but this memory density blows TR out of the water

[uneven dove]

no one seriously uses LLaMA 65B on a CPU.

[uneven dove]

maybe old-gen TR

new-gen TR is a whole different beast, but it's priced accordingly

i know this because i have a TR4 motherboard i was like "ooh i should get a newer TR for it" and the newer ones are on two different sockets

friggen bastards

so the $700 motherboard is no longer applicable, yay

[high skiff]

so looking it seems like it will have 8 slots for RAM, which should be able to do 2TB RAM max, yeah

[dense chasm]

gotha,so Elon still wanna to cage fight with Much after release the Threads Imao

[uneven dove]

that's with 256G modules, boi

[high skiff]

yeah

[uneven dove]

GOOD LUCK

didn't know you were made of money

[high skiff]

256 moduels are the limit

256GB is what you get when every mm of the stick is ram dies lol

[uneven dove]

it must suck to lie down to relax for the evening, and you fall completely to pieces because you're made of money and the coins don't stack so good

[high skiff]

they would need DDR6 to go higher for dual density

[uneven dove]

this is a zero-sum game you're playing

[high skiff]

who, me?

or frank?

[uneven dove]

you can't just, use DDR6 for more density

where's that "one does not simply" meme

[ionic dragon]

@high skiff if you want used, just check hardwareswap, you can get them for cheap, but many times, things get sold within hours
im hoping you are from the states

[dense chasm]

donate for charity i thought eventually,no surprise

[high skiff]

I am saying we won;t have higher denisty RAM sticks over 256GB until we have consumer DDR6 tech, cause thats a full on limit of the architecture

[uneven dove]

it's true but i think even DDR6 has electrical limits

[high skiff]

I am, but I already bought my stuff

[high skiff]

yes, but it can utilize the already proposed and used features in DDR5 to a higher level, which is more density stacking, and variable voltage to store more than 1 bit of information per transistor

like NVIDIA's PAM

which I wish they did on consumer GPU's

then 8GB VRAM would be more than enough lol

[uneven dove]

so the thing is, the increase in density on a single module, exponentially increases the module error rate

the manufacture process becomes WAY more important (read: expensive)

[high skiff]

PAM 3 is so crazy lol

bruh

[dense chasm]

u're squeezing the computational power man

[high skiff]

also, there is already PAM4 as well

which can transmit 2 bits of information per cycle, or 2x the throughput/theoretical data storing on the same hardware

PAM 3 is 1.5 bits per cycle

[uneven dove]

i added a new feature to my discord bot

[high skiff]

nice, glad to see it

you have been talking about that for a bit

[uneven dove]

little react buttons to re-gen with someone else's exact settings

or to clone their settings altogether

[uneven dove]

so many ideas and so little energy and time to implement them

[high skiff]

100% understand lmao

thats me with comfy

[ionic dragon]

3090 for $600 on hardwareswap, damn

[uneven dove]

be careful who you call a shrek degen. one of them will marry your daughter someday

[high skiff]

maybe I will look there instead then

3090 is the only thing I have yet to secure

[ionic dragon]

then just try to get it

[uneven dove]

do a 4090 so you don't have a regret in a month

like i told you to, all along

[ionic dragon]

but 4090 is like 1700

[uneven dove]

that's precious little money to spend for happiness

[ionic dragon]

he can get around 3 for 1900

[uneven dove]

oh, he's doing dual 3090?

well that's different

[next inlet]

3090 is a wipe, no one’s going to benefit from re-selling their cards with that for sure. Everyone hurts. They take a loss and the buyer extends for hardware pushed too hot

[uneven dove]

i hope you have a better rest of your day lmao

[ionic dragon]

he can literally get 3,so i dont think 4090 is worth 1800 for this stuff, considering the low price of 3090

[next inlet]

Im all for the gains, but that was a capitalization that took everyone down

[uneven dove]

the architecture is better but yeah you could be right, if the 3090 uses 2 slots instead of 3 and uses even the same amount of power you can do better and fit more in one box.

i would prefer two 3090's over my single 4090

[high skiff]

and also the 3090's have NVLINk for VRAM pooling

[uneven dove]

nvlink is trash

they got rid of it now

[next inlet]

Not really useful in a situation like this atm anyways

[high skiff]

they got rid of it for consumers, but its the strongest its every been for workstation

[uneven dove]

vram pooling just isn't as good imo as running two processes

you're going to be single-thread bound

[high skiff]

oh for sure, I can agree there

[uneven dove]

oh wait, i have never tried multiple processes on a single GPU

why the fuck haven't i tried that

god i'm such a waste of life

[ionic dragon]

@high skiff if you want to consider another, here you go

[high skiff]

can you link me to the HWswap discord in DM's?

[next inlet]

Runs hot, gains are pretty small overall against lifespan

[high skiff]

I only have their reddit

[ionic dragon]

workstations definitely need nvlinks, or the gpu would be thicc af

[high skiff]

nevermind, found it

[uneven dove]

don't know what you mean

[ionic dragon]

use base as a base,

@high skiff whats your budget for gpu?

[uneven dove]

the 4090 runs about 8-12% cooler than the 3090

you can pin the 4090's performance to that of a 3090 and use like, 300 watts.

[high skiff]

preferably under $700

[ionic dragon]

just go for 3090

[high skiff]

yeah I know lol

its all I was considering haha

[autumn forum]

What’s crazy to think that 50 series cards won’t come out until 2025 which will probably have a boost in vram for ai but I honestly can’t imagine what advancements we will have in ai by then.

[uneven dove]

probably none

just lower bit precision again and again

[lilac wren]

@high skiff @uneven dove @boreal bough

This error drive me crazy :
0%| | 0/15 [00:01<?, ?it/s] Traceback (most recent call last): File "F:\IA\LoRA_Easy_Training_Scripts\sd_scripts\sdxl_train_network.py", line 167, in <module> trainer.train(args) File "F:\IA\LoRA_Easy_Training_Scripts\sd_scripts\train_network.py", line 250, in train train_dataset_group.cache_latents(vae, args.vae_batch_size, args.cache_latents_to_disk, accelerator.is_main_process) File "F:\IA\LoRA_Easy_Training_Scripts\sd_scripts\library\train_util.py", line 1730, in cache_latents dataset.cache_latents(vae, vae_batch_size, cache_to_disk, is_main_process) File "F:\IA\LoRA_Easy_Training_Scripts\sd_scripts\library\train_util.py", line 918, in cache_latents raise RuntimeError(f"NaN detected in latents: {info.absolute_path}") RuntimeError: NaN detected in latents: C:/Users/Seb/Desktop/Stable Diffusion/LoRA SDXL/Leah_512/image/100_LL-SDXL_v1\00001-0-image-047.png Failed to train because of error: Command '['F:\\IA\\LoRA_Easy_Training_Scripts\\sd_scripts\\venv\\Scripts\\python.exe', 'sd_scripts\\sdxl_train_network.py', '--config_file=runtime_store\\config.toml', '--dataset_config=runtime_store\\dataset.toml']' returned non-zero exit status 1.

[uneven dove]

training in 4bit is now becoming a thing

@lilac wren using bf16?

[lilac wren]

fp16

[uneven dove]

don't do that

lol

[high skiff]

looking

[lilac wren]

Hmm

[uneven dove]

SDXL is worse than 2.x in the NaN issue by far

it needs 32bit mixed-precision

the numbers get way too small

[lilac wren]

Isn't it supposed to save a lot of Vram to put fp16?

[high skiff]

ok, I dealt with latent errors for hours as well, and I am not too sure how I fixed them, but the training never ended up running

[uneven dove]

welcome to SDXL.

fp16 cerainly saves a lot of memory and bf16 uses a lot more but not as much as fp32 does

there's also tf32 which is a sort of middle ground between fp16 and fp32

this is why everyone is training at batch_size=1

doesn't seem like 1.0 is going to change anything about that, either

[lilac wren]

Not sure this will fix my error but i will try

[glad fulcrum]

I saw this:
Here's how to switch to that branch!

1. modify your webui-user.bat file's COMMANDLINE_ARGS line to read:
   set COMMANDLINE_ARGS= --no-half-vae --disable-nan-check
2. enter these commands in your CLI:
   git fetch
   git checkout sdxl
   git pull
   webui-user.bat
3. select the SDXL checkpoint and generate art!

---

But...
Now 1.5 is not working ....

in auto1111... How can I revert back to the main branch?

[uncut steeple]

Thats why I will wait for the official A1111 support
comfyUI does fine right now

[uneven dove]

i would just move on to vlad's branch of A1111

it is technically superior in its implementations

[glad fulcrum]

Do I need to reinstall lots of heavy stuff?

[uneven dove]

probably 🙂

[glad fulcrum]

or does it use the same environments

[uneven dove]

you can use same env but YMMV

[glad fulcrum]

I don't have the time then...

[uneven dove]

buddy you're not going to fix this in a mad rush lmao

if you don't have time, come back to it later

[glad fulcrum]

I don't even know what YMMV is

[uneven dove]

your mileage may vary

it could have older dependency versions in the upstream build of A1111

it could be straight-up missing some deps

[glad fulcrum]

hm

I wouldn't know though, how to install using the same environment

[uneven dove]

but vlad's has better memory usage and internal design.

[uneven dove]

yeah it sounds like there's no quick fix for this for you.

[paper phoenix]

I'm in two minds about shifting over to vlad myself. I am interested going forward what people on youtube making videos will be using to show off techniques etc. a1111 is the defacto for that but vlad has something of a lead for once with the sdxl 0.9

[shy kelp]

do you know if he got img2img working?

[paper phoenix]

? img2img works in a1111

[glad fulcrum]

well, If I can get this reverted to the mani branch, thanks

[tight fjord]

I find it weird that on the branch for auto1111 they don't support ddim yet for SDXL0.9, I wouldn't have thought it needs anything particular since comfy had everything working at once

[shy kelp]

i mean with 0.9

or does 0.9 run in auto now you mean

[boreal bough]

cursed prompt still giving good images XD

[paper phoenix]

oh no, im waiting for 1.0 havent even touched 0.9. I thought the whole thing was going to be waiting for 1.0 to come out and on day one have usability out of the box for all main uis, so it didnt have the poor start 2.x did. 0.9 leak scuppered that.

[tight fjord]

anyone knows why comfy defaults to change the seed AFTER the generation and not BEFORE? it's very annoying when you find a seed you like to reload from the picture

[uneven dove]

and Vlad's uses Diffusers, which means it also has native support for Kandinsky 2.x and more, that A1111 will struggle to support.

in fact, Vlad's fork should be able to natively support LLMs through all of the same Diffusers library backend logic at some point, whereas A1111 would need an extension for that.

[paper phoenix]

i suppose i could just use symlinks and run both

[boreal bough]

it was already mentioned, but fp16 is not working for sdxl. use bf16 as recommended. I even mentioned it in my linked post earlier

[uncut steeple]

[uneven dove]

it's unfortunate that so many people got used to training in fp16 with 1.5

it's one reason 2.x fine-tuning went so poorly for most. it needs BF16 or FP32 for numeric stability

[boreal bough]

prob better stick with #1100170365604483202 until full release - as everything local takes a lot of time right now - no easy plug and play solution

[uneven dove]

there's short Diffusers python examples that will work, but that's nothing like A1111...

[boreal bough]

agreed. simple solution, if you find one you like, immediately hit load on the last one in history, to get the right seed again, then change to fixed

[tight fjord]

when @visual glade is around I should ping him about that, should be very simple, worst case a toggle in settings, but I can't see justification for the current behavior

[boreal bough]

for now he busy as hell, with lots of stuff to implement. he usually has reasons for ui stuff. if you need a quick solution -> custom nodes

[visual glade]

yeah some people made prs for that but I don't think anybody made it selectable with an option

would a button to get the last seeds also solve this?

[tight fjord]

as a new comfyui user its been very rapidly #1 minor annoyance

[visual glade]

yeah I'm trying to solve those minor annoyances, sometimes I watch videos of people using it and then try to fix what they had trouble with in the UI

[tight fjord]

yeah, but why not just switch it at the source, any use case for the current behavior? also create a weird workflow when for any reason you input a fixed seed, then switch to randomize, you need to hit generate twice to go past the cache

so it's a 2 for 1 really

[uneven dove]

damn, that's pretty cool of you

[lilac wren]

even in bf16, always the same error

[uneven dove]

what card? maybe doesnt have bf16

[visual glade]

watching people use software you make is always interesting and makes you notice all the issues

especially the minor ones that people don't really report

[boreal bough]

I cry everytime a primitive cant be slotted into all nodes that say "1024"

plz give proper int variables (so that reroutes dont need to be disabled for them)

[lilac wren]

2070 Super

[uneven dove]

mine is a discord bot i directly see people struggle with lmao

[autumn forum]

I noticed I can’t use int’s in reroutes?

[visual glade]

the primitives and reroutes need to be rewritten a bit

especially the primitives

[autumn forum]

Ah

[boreal bough]

any primitive to be exact, since the reroute can cause issues

[high skiff]

please pleas please, I would kill for the ability to reroute primatives

[soft bone]

XL full Dreambooth training working great with 16gb, 45mins

[high skiff]

my workflow gives me a twitch cause I can't make it neat lol

[boreal bough]

remove primitives, and giving proper variables, with a set type would solve those issues

[high skiff]

agreed

[boreal bough]

also string combinator would be amaaaaaaaazing

[high skiff]

or just ahving the primative auto convert itself into a fixed variable

[tight fjord]

what I want is a reroute with an on/off toggle, or a way to turn off branches of a workflow, i might not always want to upscale for example

[boreal bough]

since it would let us just 'save' styles, we can replug anytime

[boreal bough]

Ctrl+M is the best we have at the moment

its not the same, but it can be made to work, to have almost the same effect

until that feature eventually gets added

[tight fjord]

it triggers mute tab in firefox, surely thats not it 🙂

[boreal bough]

its supposed to mute the node XD if you have it selected

[tight fjord]

ah cool!

[visual glade]

right click -> mode -> never also mutes

[tight fjord]

ah wow learned something, thats great

[uneven dove]

bad news, no bf16 support

[high skiff]

@visual gladeI have one more request for a node, if you wouldn't mind

[boreal bough]

RIP. sebl4rd

[uneven dove]

buy sytans old gpu lmao

[boreal bough]

sdxl marketplace 🤣

[uneven dove]

we can make this scam work

i mean

wait

i did not mean that

[high skiff]

if we could have a node where its basically a reroute, but with a built in toggle to just flip,that would be amazing

Like you can name it, and just check if its active or not. I could do sooo much with that

[tight fjord]

+1 thats what I had in mind

[high skiff]

you can see here, I am using 2 reroutes as a stop gap for now lol

[lilac wren]

😦

😦

😦
😦
😦

[tight fjord]

spaghetti, spaghetti everywhere

[high skiff]

but if I could just name one "Upscaler" and click active, that would be delicious

[boreal bough]

me, at around 50 locations in my megagrid XD

[high skiff]

exactly lmao

[uneven dove]

now you know why people using 24gb gpu to train

[high skiff]

oh man, and imagine if you could choose between toggle or randomize and just let it turn on whenever it feels like it

[boreal bough]

the trick is to get a sugarmommy girlfriend that buys you a 4090 cause you said you need it to have more AI fun 🤣

[high skiff]

NO!?

How

[lilac wren]

I already knew that, but between repairing your car and buying a GPU that costs a kidney... 🙂

[high skiff]

DARE

you

@uneven dove

[lilac wren]

We didn't have the same girlfriends haha

[boreal bough]

tbf she makes less than me, so not what I was expecting to happen at all XD

[lilac wren]

She must be perfectly submissive, nice work (joke)

[boreal bough]

I got my revenge XD she got a whole pc with a 3080

[regal heath]

does anyone here use a 1660ti?

or a 6GB VRAM GPU?

[boreal bough]

cutting it close there O_O

I mean I know it works on 8gb...

maybe with diffusers?

[regal heath]

well i'm generating 700x512 using hirefix to upscale to full hd

wanted to know if someone here managed to get something bigger

[high skiff]

also, just found out another reason why this 3080 is so mean to my PSU is cause I have about the most absurdly over cranked 3080 out there @uneven dove

like, this 3080 is bigger and draws more power than the 3090 this guy is trying to sell me lmao

[uneven dove]

so maybe the 3090 is slower?

[high skiff]

3090's minimum PSU recommendation is 750 watts, this 3080's is 800

its TDP is also 20 watts higher

[high skiff]

oh for sure not

[lilac wren]

you finally do your voodoo with Vram and solder paste?

[high skiff]

2k more cuda cores, and a higher clock

[uneven dove]

i don't see how the power draw is lower though then 😛

[high skiff]

this 3080 is just a very overdriven 3080 is all

[high skiff]

efficiency

[uneven dove]

it uses the same dang process node

[high skiff]

you put in 20% more power to get like 3% higher clocks

[high skiff]

@uneven dovehis GPU is for sure faster

he does SD and he just showed me he got 23.5it/s in invoke AI

[uneven dove]

the TDP is the same or higher for 3090 vs 3080

[high skiff]

yes, this is a modified MSI 3080 with a higher TDP and very huge cooler to drive it further

my 3080

MIS gaming Z trio

[uneven dove]

that has a 340w tdp

MSI rates their RTX 3080 Gaming X Trio at 340W TDP, which is 20W higher than the Asus and Founders Edition cards. Considering the performance delta we saw in the benchmarks, we were curious to see if the power delta would show up. We use Powenetics in-line power monitoring hardware and software so that we can report the real power use of the graphics card. Powenetics also links up with GPU-Z to record GPU temperature, fan speed, and GPU clock speed, which we'll report below.

[visual glade]

it/s at 512x512 are pretty meaningless, mine change just from how many windows I have open

[regal heath]

how much slower is with AMD cards?

[high skiff]

thats gaming X trio

I have gaming Z trio

[visual glade]

people should compare performance at 1024x1024

[uneven dove]

the Z trio has a TDP of 320w, even lower

[high skiff]

oh yeah, I get it, I don't need any faster, I was just ensuring that it is faster for pseudo

[uneven dove]

https://www.techpowerup.com/gpu-specs/msi-rtx-3080-gaming-z-trio.b8783 < is this not the same one?

[high skiff]

on their actual site

[uneven dove]

i'm not doubting it's faster, i just disbelieve it's faster while pulling less power for the same process node and GPU architecture

[visual glade]

but invokeai is pretty fast since they use diffusers I always wonder why they always get left out of the conversations

[high skiff]

thats how efficiency works

its 8704 cuda cores being fed way more power to go a little faster vs over 10k cores being fed the same and much more efficient amount of power

[uneven dove]

that's not how GPUs work, lol

[high skiff]

yes it is?

you feed them more power to run them at higher clocks

its a stock overclock

[regal heath]

how much difference does the 3060 (12gb) and the 3060ti (8gb) have when using SD? is the bigger VRAM better?

[boreal bough]

every time I tried it, I left in under a day, since regardless of how good they are - they just lack small important details that make me tiny mad