🆕｜sd3

[noble coyote]

Waaay way back I used a great "royalty-free" music composer named "Juke Deck"
I think 10 Cent bought it - and buried it!!!?

[bitter hearth]

this was fun making with voice and lip sync with DD Lightning checkpoint

the sound effects are nice its super deep

some real bass

[kind siren]

Tips for setting this better to get more interesting results?

[kind siren]

Very nice

[bitter hearth]

yes the background noise was 3 layers of different sounds

[bitter hearth]

60 steps
CFG 3
dpmpp_3m_SDE
beta scheduler

I clonded the voice with coqui TTS and the a voice over guy that does all the big movies called Red Pepper

I'm afraid best results require more steps and its hard to get around that

if you want to squeeze more out of low steps, UniPC 15 steps beta scheduler is ok

again with low CFG, often 3

this the cloned voiced of Red Pepper

DPM++ 2M at 20 steps is good though

but switch to karras or beta

these voices are nice yeah

and thats the finished cloned

coqui TTS I should try this

[kind siren]

Thanks guys.

Here's my current setup:

[bitter hearth]

yeah that should be good

IDK about the denoise number though

that's situational

[bitter hearth]

its awesome you only need a 5 sec voice clip to clone with it to

just be aware that the required % denoise might change, when other settings change

[kind siren]

What exactly does the denoise thing do?

[bitter hearth]

this is my cloned Morgan Freeman sounds just like him

denoise is for image to image

if you are doing text to image then you want to leave it at default

[kind siren]

After these changes now my image looks soulless like this:

Maybe Im doing smth wrong?

[bitter hearth]

for image to image, the more denoise, the more the image will change

my text to Morgan

what do you have going into the latent input?

if its create empty latent then you need denoise = 1

this looks a lot like you denoised an empty latent half way

[kind siren]

What do you mean with latent input?

[mortal mesa]

they are trying image2image

[kind siren]

what is the "latent input"?

[bitter hearth]

Morgan Freeman tells you about how he is addicted to Dream Diffusion Checkpoints lol

into the Ksampler is a pink noodle

whats the pink noodle going to

its plugging into latent_image

I need to know if its going to a VAE encode node or to a create latent node

[bitter hearth]

ooh this is img to img

okay

[kind siren]

yeah im doing img2img

[bitter hearth]

you may have to try different denoise values

cos this is kinda the hard part

predicting which value will work

[kind siren]

question: When I load an image... and I keep changing it again and again with prompts.

It works based on the latest image it generated?

Or does it always start from scratch from the original loaded img

[bitter hearth]

why is your steps set at 60 bro?

[bitter hearth]

lol i told him that

[kind siren]

yeah

[bitter hearth]

you can go back to 20 if you want but I like more

ouch

[kind siren]

Going back to my quesh::

[bitter hearth]

is that aflux model he is using ?

[kind siren]

Does it alter always the original loaded image

or does it keep changing based on the images it starts to change

does my question make sense?

[bitter hearth]

no I don't quite get it

I can explain though

if the original input image is always the same, then it will only change if the seed changes

unless you are using a sampler with "SDE, a or ancestral" in the name

thats the most optomised way

[kind siren]

Ok

So I load an image

then I tell prompt: change it this way

now it generates a changed image

now I tell it now change this thing.

Will it make the changes based on the change image, or will it make the new changes based on the original img

[bitter hearth]

steps between 20 and 30

[kind siren]

Will try this

[bitter hearth]

oh I see
it will be based on the original image still

keep your schudular as simple and not normal

if you want it to change the new image
you have to load that new image and then put it into vae encode

[kind siren]

I have to learn how to do this

[kind siren]

Ok, thanks!

[bitter hearth]

its in the screen shot up there

[kind siren]

Isn't there a search feature for this?

[bitter hearth]

for what?

double click for search

oh i see what you meant , thought you was searching for somthing particular

[kind siren]

For searching for the notes. But I learned double clicking works thanks to NeonNinja

[sacred jewel]

in a short two weeks time, there are over 200 LoRAs for Flux on CivitAI alone... INSANE!! (may be more... I gave up scrolling lol)

[bitter hearth]

200 lora is a lot
given its not that cheap to train

200 undertrained waste of space loras , made to earn buzz

[sacred jewel]

Double click on an empty space on the background canvas.

[noble coyote]

My favourite Flux LoRAs - rbcharc0al, Blue Future, Pop Art, Art Nouveau, Impressionism (landscapes only), Metallic, Vintage Anime ...

[sage burrow]

Well this looks useful

[noble coyote]

Prompt injection

[kind siren]

Current Setup so far:

I added three new nodes

now I have to connect them to everything else!

[noble coyote]

Does Load Image automatically replenish at the start of a new Queue?

[kind siren]

How can I connect these three new nodes to everything else? @bitter hearth

[bitter hearth]

Thats all Civitais doing , Making a rod for there own back as encouraging users to make merges and loras that they have no clue about. Yes there are developers on there making them correctly . but 70% of them are all undertrained and not doing as they are intended. Its a buzz race as they beleive they will earn money from it.... Which I can from personal experience on Civitai tell you , You cant. Only people on civitai making any real return are the big hitters with checkpoints promoted on the site generarator such as juggernaut and such like. I was the highest grossing buzz creator that wasnt on the site generator with 7 months of working on there. In that 7 months I accumulated 65k Buzz. thats translates to a cash price of $35 .....

[sage burrow]

So how do you properly merge models?

[bitter hearth]

And when i won the Civitai community Checkpoint favorite, Civitai hated the fact they had to pay me 100k buzz. 2 days later Civitai banned my account and took down all my 32 models and and removed my 2000 followers. And took mback there 100k buzz... So no I dont have any time for that site

[bitter hearth]

model merging is a 30 second job on auto1111 or forge. Its no skill, no work, but civitai will pay buzz so there site is now flooded with useless loras and horrendous merged checkpoints

I would never tip and merged checkpoints buzz on civitai, I highly tipped buzz to the trained checkpoints tho

[kind siren]

Any tips @bitter hearth ?

[bitter hearth]

tips on what my man?

[kind siren]

Like, I created these three new sections:

So how do I connect this to everything else

[bitter hearth]

do you want the complete workflow ready to mate ?

[kind siren]

yeah!!!!!

that'd be amazing

[bitter hearth]

tap the download on the right , save it to your pc, then clear your comfy work sheet, then hit load. and choose this file where you saved it to . ok

[kind siren]

ok lemme see

Great!

I have it now! 🙂

[bitter hearth]

once loaded you will see this workflow in your page

[kind siren]

Im getting this error:

[bitter hearth]

make sure the nodes are set as mine are to get started then you can adjust them as you learn

right that is telling you the vae and such are not saved to the correct folders

[noble coyote]

Dice has named his Diffusion Models different to yours - just choose one which works

[sage burrow]

The model merge I did yesterday, was that an actual proper merge, or?

[torn wharf]

I'll report csam on civit and they'll only take down one image. They won't delete the person's account or any of the other images they make.

I know people deleting their accounts there because of how much csam they allow

[bitter hearth]

clip name 1 shows as null. that should say T5xxxl

[kind siren]

Here's what I have:

[sacred jewel]

BRO!!! You've been killing it... nice generations...

(I am not jealous 👀 )

😛

[kind siren]

Last time I tried this, I got the error

[noble coyote]

Wow!

[bitter hearth]

correct now

[kind siren]

so what do I have to change this time to make it work?

[kind siren]

Last time with this setup I pressed on the queue prompt and got the error I sent!

[bitter hearth]

thats correct now mate

Dice i think he is using sd3

hit Queue prompt

[kind siren]

I get this error when I use it:

yeah I have SD3!

[bitter hearth]

he is showing as using flux hyper

[torn wharf]

Make sure you select the right files. Those are still referring to his file names and locations

[kind siren]

remember: my knowledge about how this all works is ZERO! haha

[bitter hearth]

Alfonso check which model you are using... they work differently

[kind siren]

I dont have t5xll that Im aware of

[bitter hearth]

Your Vae should be in the Vae folder, Your clip_L and the T5xxl should be in the Clip_vision folder, And my checkpoint flux needs to be in the Unet folder

[kind siren]

haha this is like Chinese to me

[kind siren]

I have this model:

sd3_medium_incl_clips.safetensors

[mortal mesa]

you need to reselect then as yours might be in a different place or name than the ones he included in that workflow

[bitter hearth]

t5 encoder wont work with sd3 models, or do they?

[mortal mesa]

sure it does

[bitter hearth]

im questioning cause i spent very little time with sd3 lol

yes they do

ahh ok

[kind siren]

Ok, I dont have anything in my unet or vae folders.

[noble coyote]

D/load files needed from here ...https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main

[kind siren]

ohhh ok ok !

[noble coyote]

https://comfyanonymous.github.io/ComfyUI_examples/flux/

[kind siren]

Ok wait a sec! first I will download the files 🙂

[kind siren]

Which of these should I install and where should I put them?

I dont have stableSwarm I have ComfyUI !

[bitter hearth]

download that t5xxl and a few up youll see clip_L also download that , save both to to your clip_vision folder

[errant dust]

you mean Clip

[noble coyote]

D/load all of them and place in Clip folder

[bitter hearth]

?

[errant dust]

They go into the Clip folder

[bitter hearth]

dont need the 16fp

[noble coyote]

Clip Folder is inside the Models Folder ...

[bitter hearth]

i save mine to clip vision

[kind siren]

Ok Im downloading this: t5xxl_fp8_e4m3fn.safetensors

Into clip vision folder

[noble coyote]

I save across Clip/Clip Vision ...

[errant dust]

My Clip Vision is empty

[noble coyote]

eau keigh

[kind siren]

Ive downlaoded it here!

[bitter hearth]

not fully sure of the difference but made sence to add them to clip_vision and they both show up

[noble coyote]

Arpillera for folk crafted colour

[kind siren]

nex t step is to download this: clip_l.safetensors to the clip folder??

[errant dust]

Yes

you need both the T5 and the clip_l

[bitter hearth]

I renamed the Vae to flux Vae after adding to the folder save confussion

[errant dust]

Same

[kind siren]

Ok here they are:

now what?

[marsh rock]

How do i use sd on mobile?

[bitter hearth]

thats the vae you need to download and save to Vae Folder

[errant dust]

Only with an online service

or remotely connected to your PC

[kind siren]

Link?

[errant dust]

https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main/vae

[bitter hearth]

https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main/vae

[noble coyote]

She has a stranggge right leg!!!

[marsh rock]

Can you give any links

[bitter hearth]

i used the schell vae

[errant dust]

it is the same

[bitter hearth]

not sure never checked lol

[noble coyote]

U can change the name to ae.sft

[bitter hearth]

change it to what ever you like

[kind siren]

it says file isnt available!

[bitter hearth]

https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main/vae

hit that

[errant dust]

I literally just downloaded it

[noble coyote]

Hector+ZsaZsa.sft

[kind siren]

OK I GOT IT NOW

[bitter hearth]

download that

[kind siren]

I have my vae files

and my clip files ready

[bitter hearth]

cool do you have the flux hyper model downloaded ?

[kind siren]

Nope!

[bitter hearth]

https://tensor.art/models/759856135286068673/FLUX-HYPER-TRAINED-DREAM-DIFFUSION-BY-DICE-V-1

down load that and save it to the Unet folder

[kind siren]

k just a sec!

creating account...

[errant dust]

or the offiical one at https://huggingface.co/Comfy-Org/flux1-dev/blob/main/flux1-dev-fp8.safetensors

[bitter hearth]

if you ever download any other flux model checkpoints always save them to unet ok

[kind siren]

ok!

Downloading into unet

What would be the next step after this?

[bitter hearth]

but you dont need to download all the flux model stuff as you now have all the parts required

[errant dust]

if you have Comfy and a workflow in place

just start her up and test

see if anything 'breaks'

[kind siren]

Ok, waiting for download to fiinsh

after this I just try the worlfklow u gave me, Dice?

[bitter hearth]

you simply refresh your comfy and the fields in the nodes will auto fill with the downloads you just did

[noble coyote]

Silly Old Hector

[kind siren]

ok I will try soon

[errant dust]

Any should be fine. Be suire to refersh though as some names mayu not match

[kind siren]

When this finishes downloading I'll let y'all know

5 more mins

[errant dust]

if you see a node that says 'null' just click on it to what the the names of files it found and choose the right one

[kind siren]

beautiful

[noble coyote]

Art New Veau!

[bitter hearth]

when you do your first generation it will take a min to start as its running the model, 2nd generation will be much faster

[kind siren]

ok, great!

[bitter hearth]

see you in 20 mins lol

[kind siren]

So with this I'll be able to load an img and change it right? with a prompt I mean

[kind siren]

hahahah

[noble coyote]

Yes, it reads and processes the prompt first time round

[bitter hearth]

not with that workflow but ill give you another workflow for img2img

[errant dust]

well, img 2 img is a bit different

[kind siren]

and with your workflow, Dice, can I also generate an image based on prompts only without using any loaded images?

[errant dust]

and requires a special workflow

[kind siren]

Ok great!!!!!!!!

[bitter hearth]

yes the workflow I gave you is txt2img to get you strted

[kind siren]

ohh hok

great!!!!!!!

I'd just like an img2img and a normal prompt based workflow!

[noble coyote]

I can never get Xflux IPAdapter working.
It always stall and hangs at X-Flux Sampler!!!

[kind siren]

It looks like u gave me the normal prompt workflow, right?

[bitter hearth]

if you want to save this while you wait to the same folder as you saved the first work flow. This workflow is the img2img for flux

[kind siren]

grrr8888 !!

[bitter hearth]

yes thats the prompt already loaded in the workflow , test that as its a text prompt. afterwards just add what ever you like

[kind siren]

Ok when Its done I will test first the normal prompt workflow

Download FINISHED!

[bitter hearth]

that was quick

[sage burrow]

Mage.space has a free plan (and paid also if you want to get fancy). Yu can also search huggingface for flux. Or you can use Glif (it's presently free)

[kind siren]

Ok so. Im there

im about to TEST this!

First error appeared:

this is just from pressing the queue prompt button

[bitter hearth]

change the vae title in that node

[noble coyote]

Where do you get these nodes svp?

[errant dust]

Can you share your workflow for that ControlNet you showed yesterday with the Car frame and natural landscape overlay?

[bitter hearth]

tap it and youll see the title of the one you downloaded

[errant dust]

Do you have Comfy Manager?

[kind siren]

So like this?

[bitter hearth]

sorry bro what workflow is that ?

[noble coyote]

Yes but it does not appear in Manager

[kind siren]

Now I get a different error

[bitter hearth]

perfect

[noble coyote]

#🆕｜sd3 message

[kind siren]

new error!

[errant dust]

even when you asked it to look for mising nodes?

[noble coyote]

Yes

[craggy crest]

those are your models it can't find.

[bitter hearth]

slide that red hightlighted node down

[kind siren]

Ok it's more down now

I moved it down

[errant dust]

WHere did you place your files? CLip or Clip Vision?

[craggy crest]

it's looking for the T5xxl encoder and the clip_l encoder

[kind siren]

clip vision

[errant dust]

Try Clip

[kind siren]

ok, will try

[bitter hearth]

refresh you page or clear the page and reload the workflow

[kind siren]

should I try moving it to "clip" first?

Like subtleone suggested?

[bitter hearth]

it will auto fill with a reload of workflow

[bitter hearth]

no

clear your work space totally

then load the workflow again

[kind siren]

Ok, Ive reloaded the entire thing again

Ive done it with: run_nvidia_gpu.bat

[bitter hearth]

now are the nodes all filled in correctly

[kind siren]

now it has loaded

[bitter hearth]

they will auto fill

[kind siren]

I go tthe same error!

[errant dust]

press F5

(browser refresh

[mortal mesa]

click the t5 and clip l box in the red node and select them again, if they dont show up ,ove from clipvision to clip folder

[kind siren]

When I click on the little white arrows I get "undefined"

[mortal mesa]

after browser refres

[bitter hearth]

kagi said the right thing

[errant dust]

right now it is using the saved workflow, meaning it is showing the values used to make the workflow, even if thge files are diffetrent

[craggy crest]

it can't find your encoders. click on each of them and see if they're even in the list. if not, make sure you put them in the right folders

[mortal mesa]

refresh browser first to make sure

[errant dust]

F5 will have it check to see if those files are even there

[kind siren]

browser refre3shed

[errant dust]

what does the DualCLip node show?

[bitter hearth]

ok copy paste both your t5xxl and the clip-l to the clip folder

[errant dust]

Yes, means it must be in Clip

[kind siren]

so move them to "clip"?

ok wait!

[bitter hearth]

try that yes

[kind siren]

reloading server again! And it's finished!

[bitter hearth]

mine maybe working different to yours as my comfy is the backend of swarm so mine had to go in clip-vision

[errant dust]

hit refresh of course

[craggy crest]

they go in /comfyUI/models/clip_vision

[kind siren]

ok now it can find thgem

[errant dust]

and then if you still see Null, which is normal, click on the nulll

[kind siren]

it can find the clip files!

[errant dust]

and it will show what files it sees in the folder

[kind siren]

IT'S WORKING NOW!!!!

[errant dust]

This last advice is useful because it applies even to new checkpoints, or LoRAs you add

[kind siren]

It's working fellas!!!!

Thanks everybody 🥳 🪅

So now waiting for image to be generated

[bitter hearth]

think ill make a youtube video of how to. As a lot of people dont know how to do this . think it would be helpfull?

[kind siren]

OF COURSE!

U should totally do a video about this

it would benefit many

[errant dust]

As long as it mentions.... (I WON'T say I told you so, I won't!) CLip

[kind siren]

how long is this gonna take tho?

[errant dust]

😛

[errant dust]

that will depend on your hardware

[kind siren]

I see

well my laptop is like from... 2021 or 2022

hehe

[bitter hearth]

Its not a simple process for new users . And i think a lot of new user are put off using flux as it seem to complicted for new people to SD

[errant dust]

FOr me, it can take 3 mins, for some with RTX4090s, it takes seconds

[kind siren]

I have GeForce RTX!

[errant dust]

(I sport a laptop with a laptop 4060

[kind siren]

Yes... This all seems so alien for newbies

[errant dust]

Yeah... believe me that is a VERY broad definition

[kind siren]

Now it's in the "SampleCustomAdvanced" step

quesiton. HOW DO I SAVE this configuration

so I dont have to do this ever again.. Like I just open it up and it works

[errant dust]

the first RTX was the 20xx series and could stoop to as 'slow' as an RTX2060 and that was in ... 2018 if memory serves

[bitter hearth]

hit save in the generation box

[noble coyote]

D/loading Florence node direct from Huggingface https://huggingface.co/microsoft/Florence-2-large-ft/tree/main?clone=true

[errant dust]

THen the fastest was a 2080ti with 11GB VRAm. today the king is a RTX4090 with 24GB Vram

and we are talking just the consumer GPUs

[kind siren]

hey... so everytime i use this like.. .if i use it tomorrow. The first generation will also take a lot of time?

like everytime I restart my laptop

the time for generating the first image will be long always?

[errant dust]

Both. It will open the last configuration, but let's say you lost or erased it. Just drag and drop the last image you made onto the browser with COmfy open and it will magically load the complete configuration, including prompt, it used to generate that image

[bitter hearth]

mine renders in around 24 seconds

[errant dust]

And we hate you for it

[bitter hearth]

😆

[errant dust]

lololol

[bitter hearth]

if it wasnt my job id be on a potato pc waiting 5 mins lol

[noble coyote]

Yes, the text encoder at t5 gets special attention, a deep reading and understanding of the text prompt takes some time first time around
Each time you change any models, or change words in the prompt - then there is a t5 delay again ...

[noble coyote]

welcome TO hell?

[kind siren]

wow

[errant dust]

Again, the question is what exact GPU you are using

[kind siren]

welll for me it's like this so far

[noble coyote]

I see there is a TensorRT for AuraFlow ...

[errant dust]

Show the CMD line screen

[bitter hearth]

you can always download the T5xxl fp16 version but id suggest having Vram of over 16GB otherwise you could hand draw the image quicker lol

[kind siren]

a long ass time

[errant dust]

it will show more precisely how fast or slow it is going

[kind siren]

How do I do that

[bitter hearth]

first render is the longest bud

[errant dust]

ok, when you open Comfy, it also opend a black screen

[kind siren]

ahhh yeah

the python thing?

[errant dust]

outside the browser

yes

it will show a lot more info

[bitter hearth]

first render is loading the model mate

[errant dust]

on the time it is taking and expecting

[errant dust]

oooooffff

[kind siren]

But it will be like that every time I use it? Or simply this very FIRST render ever will be the only ocassion it will take that long

like only the very very very first time ever I mean

[errant dust]

The prognosis is not good. Do you know what exact GPU you have?

RTX... what?

[kind siren]

wait lemme see

[bitter hearth]

example my first render takes 1.26 mins, 2nd render 26 seconds

[errant dust]

I only enjoy that if I don't change th prompt

[bitter hearth]

as the model has to load up first

[errant dust]

meaning a speedup

BUT, bear in mind I am VRAM starved

[kind siren]

ummm so how do I check my rtx thing?

[errant dust]

so there is no way it is ever keeping it all in memory all the time

[kind siren]

it's not in About this PC

it's at 40% now

[noble coyote]

Each time you change the prompt - or if u change a model - the t5 delay will come again.
If u make no changes after the first t5 delay - then it will not slow down the second time around etc

[errant dust]

Ok, you have WIndows, right? In the search bar at the bottom, type sys

it will suggest System Infomration. CLick that

[bitter hearth]

if you guys have really low vram i have a little fix for you. simply download this app, its called RamMap https://www.filehorse.com/download-rammap/

clears all ram cache in 2 seconds

[errant dust]

Yeah, my challenge is fitting a 16GB file into 8GB of VRAM

[bitter hearth]

frees up around 35GB of ram cache well certainly for me

[kind siren]

doesnt mention anything about rtx

[mortal mesa]

run T5 in CPU its not as slow as you'd think, save the VRAM for the model, for us potato users

[errant dust]

Look, if that app can magically change my 16GB Ram machine into 35, I am jumping all over it

[bitter hearth]

watch this short vid of rammap youll love it. I sware by it guys https://youtu.be/XSWIFbQVvsY?feature=shared

its only 2 mins long but will help 70% of you out

[errant dust]

Look, clearing cache is not going to change how much RAM I have installed

[bitter hearth]

no but will speed up your gens

[kind siren]

Does having a lot of stuff on my laptop make it slower at generating images?

if i had more free space would it be faster?

[bitter hearth]

i use that app each hour im working

[kind siren]

it's at 60% now!

[errant dust]

Try this instead: Open the Task Manager

[kind siren]

ok

[errant dust]

Windows 10?

[kind siren]

11

[bitter hearth]

if you want to clear all your pc of cache do Run box, type in run box Temp hit enter . Delete everything from that file. Do the same process but in the run box type %temp% and delete all files and again type Prefetch and then delete all file

[errant dust]

on the left, click that icon you see

[errant dust]

As you can see it lists both my GPUs (this being a laptop

[bitter hearth]

those files are all used and just glog your ram

[errant dust]

ok, now click on GPU 0

[kind siren]

it's nvidia geforce rtx 3060

[errant dust]

How many GB of VRAM is that?

[bitter hearth]

the next cache clear is a big one... Open CMD type in cmd line ipconfig /FlushDns hit enter

[mortal mesa]

maybe install Norton

[kind siren]

no idea

[mortal mesa]

no no dont lol

[bitter hearth]

norton isnt worth the script it was made on. |Norton informs microsoft of your every movent

[errant dust]

@kind siren Ok, try this: go to this link and download the model into the Unet folder

https://huggingface.co/Comfy-Org/flux1-dev/blob/main/flux1-dev-fp8.safetensors

[kind siren]

k wait!

[errant dust]

Use it instead as a litmus test

[kind siren]

wait wait

[errant dust]

no rush

[kind siren]

ill wait for the img to finish!

[bitter hearth]

dude he dont need another flux

[kind siren]

it's at 80%

[bitter hearth]

that will slow him down more

[errant dust]

No it won't

[bitter hearth]

his rtx 3060 will have around 20 GB Vram

[errant dust]

What???

lol

[bitter hearth]

3060 have a vram of 20GB

3090 have 24GB Vram

[errant dust]

sigh

[noble coyote]

Where can I get comfyui-tensorop? It doesn't come up in ComfyUI Manager?

[bitter hearth]

plus his pc ram is already quite full

[bitter hearth]

hugg or git

[kind siren]

88% !!!

[errant dust]

Ok, let's start with a few realities. a 3060 has EIGHT GB of Vram

[kind siren]

92!

[errant dust]

second the size of his file has zero to do with his RAM

[bitter hearth]

pc ram mate

88%

[errant dust]

yes

Zero

he can download 20 files and it won't affect his Ram

[bitter hearth]

he will want to keep his renders so with a low ram and multiple models he will fill that in no time

[errant dust]

There is no correlation between the storage space and his system RAM

[bitter hearth]

6GB to 10GB download per model

[errant dust]

it is completely unimportant

[kind siren]

image FINISHED!

[bitter hearth]

of course it is

[kind siren]

Ok so... How do I make this work faster?

this toook to much time for it to be fun

too***

[errant dust]

Of course it isn't. If I download one terabyte of data tonight, guess how much that will impact my RAM usage. Can you say zero?

[noble coyote]

I have d/loaded Florence 2 - but it still fails to load in ComfyUI - not available via manager?

Are u using kijai's Florence2 node

[kind siren]

tbh if everytime I wanna use this my first render is gonna take so much

i prefer an alternative!

like a simpler setup I mean

or something faster!

[bitter hearth]

sdd 1 ter. download comfy ui, download forge, download auto1111, save 2000 renders, multiple models , mutiple extensions. without including an=ything downloaded outside of ai stuff. the 1 TB SSD will be filled in a matter of months

[errant dust]

Alfonso, try the file I suggested. It may be as slow, but it is a known quantity for whcih one can more easily give you a baseline

[kind siren]

hmmmm

[errant dust]

Sidenote: did you have a lot of tabs or programs open?

[kind siren]

I had 2 extra tabs

ummm

any alternatives that wont be as long?

[errant dust]

It is hard to know. I have a laptop 4060, which will be faster than you, but not by a lot

and defintely not the amount of time it took you

[kind siren]

yeah but with a simpler setup it generated an image in less than a minute!

[bitter hearth]

what rtx 3060 do you have ?

[kind siren]

how do I check?

[errant dust]

You already showed, we don't need to know

he showed the screenshot

[bitter hearth]

i didnt see that

[noble coyote]

I swapped-out my 1Tb SSD for a 2Tb SSD.

Bad idea, I really need 4Tb!!!

[sacred jewel]

OH YES... this is the way. I went from 4TB to 8TB SSD because I did not like to swap files in and out to / from my NAS... so I just have all the junk in one drive 😛

[bitter hearth]

I swapped out my SSD 4 TB and planted a 8TB SSD and a external 20TB

[bitter hearth]

snap lol

[kind siren]

and a few programs opened like spotify

but not many

[noble coyote]

I'm stuck with this Florence w/f 😦

[bitter hearth]

do you know how much cache spotify uses. its crazy delete you spotify cache in the loacl files

[bitter hearth]

do you want the florance models?

or the nodes

[noble coyote]

I've got about 30Tb external, spread over 9 drives

[bitter hearth]

9 drives lol

i have C and D thats it lol

[noble coyote]

The ComfyUI Manager doesn' have these two - so I manually d/loaded Florence2Model - but cannoit find the tensorops node?

[bitter hearth]

partitioned C and D is all ai

[noble coyote]

10Tb, 8Tb, 4Tb, 3 x 2Tb ...

[sacred jewel]

Synology NAS FTW...

[sage burrow]

Also reboot your system right before long/difficult renders (if on windows)

[bitter hearth]

you tried https://github.com/kijai/ComfyUI-Florence2

https://github.com/kijai/ComfyUI-Florence2/blob/main/nodes.py

https://huggingface.co/microsoft/Florence-2-large

[sacred jewel]

get a NAS... your life will be soooooo much better 😛

RAID10 or Synology RAID or build your own using ZFS or whatever... I will never again manage individual JBODs

[kind siren]

Ok so let's talk about solutions

For this to work faster,

Do I have to try a simpler workflow?

Do I have to try a different model?

What things can I do

[bitter hearth]

https://huggingface.co/HuggingFaceM4/Florence-2-DocVQA

[kind siren]

Is this what I should use?

[noble coyote]

I may well do that 🙂

[bitter hearth]

no that was for Torcello

[noble coyote]

Sheesh, this is proving difficult!

[bitter hearth]

you think thats difficult try the morph img2vid workflow lol

that had me 2nd guessing for hours lol

[sacred jewel]

I have one of these with a pair of external 5 slot extenders

https://www.synology.com/en-us/products/DS1821+

20TB drives in each bay... but I also have an extensive bootleg collection of movies and music 😛

[bitter hearth]

thats a sweet drop mate

[noble coyote]

I've conquered that, gratefully!

[bitter hearth]

i like that reminds me of the extension infinte zoom

the next challenge is to get the video to move to the beat of the song. I use Parsqe , this video is my how to using deforum to instruct Parseq https://youtu.be/vKwmdeeDdQA?feature=shared

[sacred jewel]

I have a feeling that Forge/A1111 handle this LoRA differently that Comfy. my output is very noisy compared to their examples which are very smooth.

[bitter hearth]

I need to try a lora on forge. I havent made any loras for flux as yet

I made a lora for SD3 was a text and font lora as sd3 wasnt all it was made out to be with the texts

[sacred jewel]

I have not used forge or a1111 in months. And I do like the output of them... they are different enough in a positive way.

[bitter hearth]

given the choice and the way i made a lot of my SDXL models ive always favered forge. As i love the turbo samplers, I recoded auto1111 to run the same samplers . but i acnt work out how to add them to comfy

[sacred jewel]

Seems to improve the noise issue if I lower guidance to 1.0 from my standard 3.5

[bitter hearth]

if your on forge cfg scale 1 and distilled CFG scale 4

see the cfg scales in that for example

[sacred jewel]

Does forge work with full UNET Flux Dev and T5 16?

I don;t want to even play with all the 100 versions of FLux GGUF NF4, 8bit etc .etc . ... so confusing to me. 😛

[bitter hearth]

yes if you see the top row in my screen shot i can add or take out the vae, T5

[hexed dirge]

https://civitai.com/models/679352

[bitter hearth]

do you really need a lora to make a black and white render lol

[hexed dirge]

well try getting that horror style

[bitter hearth]

ok ill try now

[hexed dirge]

fine

[hexed dirge]

mmm really ? do you think it's the same???

[bitter hearth]

thats with your prompt without any alterATION

[bitter hearth]

An horror zombie in a cemetery. He's sitting on a tombstone looking at the camera with pain.

now ill alterate it

[hexed dirge]

ah ok

not use controlnet

just prompt

btw lora is just a shortcut

[bitter hearth]

yes ill show you once done

ok this is with only the wording black and white, next will be with the style

[hexed dirge]

well it's good but so different

[sage burrow]

https://www.deviantart.com/occulthorror 😄

[bitter hearth]

getting there

[sage burrow]

So how do I create Shnell loras?

Y'all not prompting "by Dore'" enough 😉

[cunning lintel]

🤡

[bitter hearth]

probably the closest im getting 4 renders to get this result

[sacred jewel]

I got this... very nice... should it be always BW or will it allow color?

[bitter hearth]

i madse this prompt 3D papercut shadow box of A highly detailed drawing pf horror zombie in a cemetery. He's sitting on a tombstone looking at the camera with pain, long wet hair. trench coat, Black and white , dark shadows, dark background . layered, dimensional, depth, silhouette, shadow, papercut, handmade, high contrast

[bitter hearth]

You can probably lower the strength of the lora enough for that
Well since you got color, maybe increase the strength? Cx

[bitter hearth]

close enough and didnt spend anytime making a flux lora . Right prompt on flux will get the result you want. took me 3 goes. anther 3 and it would be basically the same

Although i do like your art style @hexed dirge

I truely belive with flux you dont need a lora unless its a custom charater lora. SDXL models love a lora . But flux not so much

[sacred jewel]

I love it but only getting color...

Let's crank it to 100 😛

[errant dust]

Could you share the workflow that made this?

[noble coyote]

In which folder does florence 2 model go?

[bitter hearth]

you have asked for the only prompt and workflow id never share as its something im currently working on. But when its released ill share all made. Like i do with all my models , ill always share prompts and my workflows , just not on my current projects bro, sorry

[bitter hearth]

custom nodes

[hexed dirge]

this is with lora I think

[sacred jewel]

Yup and love it

exactly the kind of crazy shit I love to generate 😛

[bitter hearth]

anyone got a really weird lora i can try for flux. somthing a bit different

[noble coyote]

LoRAs:- using Wraith B&W (0.8) with Blue Future (0.8) and Novel Illustration V1 (0.8), then BF (0.5) and NI (0.5), finally BF (0.3) and NI (0.3)

Wraith B&W overwrites all other LoRAs (**NB)

[hexed dirge]

yes because of the rank

[noble coyote]

OK

I'm removing the B&W keyword to see if the other 2 LoRAs can reassert themselves ...

[hexed dirge]

blue is rank 2 while wraith is 8 or 16

[noble coyote]

OK, so the lower ranked LoRAs will always be struggling ...

[hexed dirge]

yes

[bitter hearth]

I amde a solo color lora for SDXL a while back, its nice the effects it gave https://www.shakker.ai/modelinfo/d9d2b5065c564cc0b48d1cb5f7a1ea45?from=personal_page

I just gave up making loras when i finshed making SD3 likeness checkpoint on a SDXL base

https://tensor.art/models/751438781070140461/SD3-LIKENESS-Dream-Diffusion-By-Dice-v1

released that 2 months before SD3 was dropped

[errant dust]

I didn't mean to put you on the spot. That's cool.

[bitter hearth]

And it still renders better than sd3 lol

[noble coyote]

How do you get SD3 out of SDXL at all?

[bitter hearth]

was a likeness its a SDXL turbo base but with the same capabilities as what was said SD3 would do

turbo base model 100% trained

check out the prompts and renders in this gallery https://www.shakker.ai/modelinfo/0ab32c16310a4b6bb52a065e09cb5017?from=personal_page

they were the prompts that were advertised by stability ai for SD3 before they released it

I said i think the turbo base sdxl model trained correctly would be better than SD3. They said no way. So I did it

That model went on to win Civitai Community vote for best checkpoint

The text and adhereance was never matched by any SDXL checkpoint

I wont be told that something cant be done... When i know it can . So had to prove them wrong

[bitter hearth]

soon as i drop it ill let you know tho bud

[noble coyote]

Ah! At 0.5, the Wraith B&W lets the other LoRAs thru ... no B&W keyword in the prompt - Blue Future and Novel Illustration LoRAs at 0.8

B&W Wraith at 0.8, 0.5, 0.3

[bitter hearth]

those are a few SD3 likeness renders

still no SDXL checkpoint can render that image on the likeness model

feel free to try here is the prompt four wine bottles lined up on a table. from left to right, they are numbered "4" then "3" then "1" then "2". from left to right, they are red, blue, green, and orange. the background is a rose field

flux obviusly walks it , but no SDXL checkpoint can

The funniest thing was that prompt was advertised that SD3 would do before they released it. And SD3 even struggles to get that right

[sage burrow]

suuuuurrrreeeee 😄

[bitter hearth]

all created with a SDXL properly trained model

[noble coyote]

Try Metallic

[bitter hearth]

I know your loras lol

[bitter hearth]

what do mean bro?

[noble coyote]

Metallic LoRA on Civitai

[bitter hearth]

for flux?

[noble coyote]

Yes

[bitter hearth]

got a link to any other website other than civitai?

[bitter hearth]

I have a matallic lora i made for SDXL

[noble coyote]

Sorry, Metallic at Civitai is only where I know it from! 😦

[bitter hearth]

https://www.shakker.ai/modelinfo/0b2c14f776c14341aa6442bdd642d73e?from=personal_page

[bitter hearth]

I dont use civitai , I leave that to buzz mining bots lol

[bitter hearth]

thats the one

flux taking it like a walk in the park

[bitter hearth]

here is a great example of stability ai advertising SD3 text adherence

here is SDXL Turbo SD3 Likeness with there prompt

even there own advert couldnt quite get stable correct

Its remarkable that a company as big and wellknow as |Stabilty ai cant 100% train there own work.

Its like why download SD3 when it came out with hardly any ui to run it on, Huge GB to download, seperate dependancies, when I fully trained SDXL does the exact same thing with a 1 click download , runs everywhere and 6GB download. Just amazes me

[hexed dirge]

oh so you're using without B&W trigger

[noble coyote]

To see how it affects the overall look

[hexed dirge]

yes it works

with good style

[errant dust]

Is this a LoRA?

I really like the colors

[sacred jewel]

Yes. BW LoRA from @hexed dirge

[sacred jewel]

Yeah I didn't see a trigger in metadata nor in civitai. Should I be?

[hexed dirge]

the trigger is B&W. I'm going to change metadata now

changed

btw with B&W it comes out black and white , without in color but keeping the style

[errant dust]

But it is in color

[sacred jewel]

See post above yours

[dry wave]

but you trained it on b&w images? That's a cool effect

[errant dust]

huh

[hexed dirge]

b&W only dataset

[hexed dirge]

also you can specify a color and maybe that image comes out like this :

also using the trigger, but telling the color

red roses

purple dress

red eyes and so on

[bitter hearth]

check out the scrolling top gallery of my lora bro https://www.shakker.ai/modelinfo/d9d2b5065c564cc0b48d1cb5f7a1ea45?from=personal_page

That was made 2 years ago buddy

[hexed dirge]

well you made for purpose

this was just unexpected

I used only B&W train dataset

and my purpose is B&W images

but it also give this effect writing a color

[bitter hearth]

I used only B and W data set to

this is the lora propmt for example a rose, (Black and white image, Blue Accent)

[hexed dirge]

the strange ai ways

[bitter hearth]

A glass sphere sculpture, concealed inside the sphere is a Ninja Warrior, Fantasy, in the dark, detailed image, 8k high quality detailed, the moon, shaped sphere, amazing wallpaper, digital painting highly detailed, 8k UHD detailed oil painting, beautiful art UHD, focus on full glass sphere, bokeh, background Modifiers: extremely detailed Award winning photography, fantasy studio lighting, photorealistic very attractive beautiful imperial colours ultra detailed 3D, (Extremely Intricate), UHD, lora:Black__White_One_Accent_Color_-_By_Dice:1, (Black and white image, Red Accent)

old hat bro

I think loras if you can be assed to make them should be something new something not seen before. Something to spark a users insparation

[hexed dirge]

Yes but they are unpredictable until you try them

[bitter hearth]

what do you mean

[hexed dirge]

i.e. I trained this lora for SDXL but it comes out very bad. for flux it's great. Other loras that were very good with SXDL works not so well with flux. Using same dataset

[bitter hearth]

if you use civitai lora creator then yes very unpridictble. If you know what your doing and how to control the creation of your lora,, then they are from unpredictble

where did you train your lora?

[hexed dirge]

I use my parameter on Civitai and also with kohja on runpod

[bitter hearth]

so you used civitai ?

thats why bro

[hexed dirge]

for this one yes

[bitter hearth]

so when you use civitai lora maker, i dont want this to sound patranising but do you know what each setting does exactly when you create in there maker?

[hexed dirge]

yes as for kohja

But, as told, also knowing parameters you can't know how lora will work

[bitter hearth]

can i ask how many images you used and captions and pixel size bro?

[hexed dirge]

it depends on lora (because I use only images for what I've the rights of use). Sometime 50 other 20

[bitter hearth]

like what would you caption this image in your lora creation

[hexed dirge]

I use kohja for caption with manual review

[bitter hearth]

civitai maker wouldnt give a monkies if you had perm on not for the images

[hexed dirge]

I know but I prefer using "legal" items

[bitter hearth]

to caption that image and trigger lora would be (black and white with red accent)

you dont need the image item

the checkpoint creates the image

lora just needs what it simply requires

[hexed dirge]

do you train lora for Flux?

[bitter hearth]

so say 10 trigger words each the same but change the red for green, blue ETC ETC

[bitter hearth]

I trained loras for every model in SD

includeing SD3

[desert garnet]

you are not seasoned like him bro

[hexed dirge]

oh wow sd3

[bitter hearth]

SD3 is no different , simple basics are still the same

https://tensor.art/models/751649208697700869/SD3-Text-Fonts-v1

[hexed dirge]

where do you train? Local?

[bitter hearth]

yes local, That was just a test lora to see if it would actually improve SD3 text gen. i have 68 loras created and all for SDXL

[hexed dirge]

4090?

[bitter hearth]

trained 14 full Checkpoints including SDXL, SD3 and flux, Hyuan DiT

2 yep lol

[desert garnet]

i heard he connected 2 4090s with nvlink first person on the world to do this

[hexed dirge]

well right. I'm with 4070. SD was easy training, but starting from SDXL too much time

[bitter hearth]

wtf did you hear that?

[desert garnet]

he used the new connector called ASSLINK

[bitter hearth]

thats the 3090 i built it with, its shown multiple times in the video

whats your point?

when i posted the video if you had read under it me a guy were discussing when i changed the GPU out

[hexed dirge]

I'm with 4070 so for flux Civitai or runpod and it costs a lot

[bitter hearth]

yes there not cheap

[hexed dirge]

12 GB

2200 buz for every lora

or 2000 with standard parameters

[bitter hearth]

is that how much there charging now

used to be 500

and a half a day wait

[hexed dirge]

for sdxl started with 500 but after some time the go up till 900

[bitter hearth]

ive been banned for around 3 months

[hexed dirge]

why?

[bitter hearth]

so those increases must of happened in that time

[bitter hearth]

very very long storey bro

[hexed dirge]

right

[desert garnet]

they hate seasoning

[bitter hearth]

They also hate developing there own website correctly https://sitereport.netcraft.com/?url=https%3A%2F%2Fcivitai.com

The users accounts on civitai are being data sort

[bitter hearth]

here is my profile page on civitai, They cant remove it as i put a block on it. also look at all the free stuff i added moving animated everything https://civitai.com/user/DiceAiDevelopment
if you tap my profile image that even moves
They cant even remove me from there own website. Just cant log in to that account as they banned the email address i used

al,l there shop for decorations is vunrable you get that all for free

10,000 buzz for a crappy picture frame. I took all there picture frames for free, and then made it a lora lol

mugging users off is there game

[desert garnet]

their store looks like an nft marketplace

[sage burrow]

why on earth would people pay for those decorations? Do they give more page findability or something?

[bitter hearth]

websites were selling that cheap nonsence in the 80s. then they intro that as there big new tthing for 2024, Spend the money on there old dated online genertor . Thing hasnt changed in years cant even resize an image ..... PANTS

[sage burrow]

However actual cash made from civitai isn't possible it seems

[bitter hearth]

to cash out in civitai you need a 1000 fiollows and then they take 25% back as a platform charge

100,000 buzz exchanges to 38 dollars US

[sage burrow]

uhm.....

you could make more selling loras on etsy or something

[desert garnet]

thats way too high should be at like 5 USD

[bitter hearth]

since moving all my models to shakker.ai in the last 3 months they have paid me out $1000 cash

[desert garnet]

how does it work there,do u upload and pay u in buzz for each download or what

[bitter hearth]

I strongly suggest getting on shakker as its rapidly becoming a awesome site and you want your models noticed as it grows as if your new creator to civitai youll just get lost in the untrained loras and useless merges as they pay for that old tosh

[bitter hearth]

civitai or shakker

[desert garnet]

shakker

[bitter hearth]

shakker you add your models . they pay you cash bonus to accout . Example just recently they had a campaign called shake the world . every checkpoint you added got 100 bucks roughly

then they had a sd3 image comp first prize 2000 bucks cash

now they are running the current comp which is aldo cash , the theme is movie posters with flux

into a creator to the site you get $10 bucks

always doing some cash incentive and they pay out as soon as you request it from your accoun

[desert garnet]

that sounds better than civit,having to wait til u get 100k buzz is a pain

[bitter hearth]

https://docs.google.com/document/d/1edY-MT2ggbnqKziw-Ej5Ee6dk4W58qFbWuIofIegDhQ/edit?pli=1

thats the flux comp. win a RTX 4090

https://docs.google.com/document/d/1Gc4uT2agzPs7xwR-ZcTPcL47AJoaBiPs1B85RM9R4P0/edit

thats the poster flux challenge

[sage burrow]

hmmmmmm 😄

[bitter hearth]

ka ching

[sage burrow]

I don't do humans though for my images/models, and you just know one with women is going to win...

[bitter hearth]

you could do semi hooman lol

[sage burrow]

THe world probably needs my horror sea creatures photography model anyways though 😉

[bitter hearth]

if your on less than a RTX 4090 theres a free one up for grabs for simply doing what you enjoy doing. Win Win

[sage burrow]

for checkpoints, darn

[bitter hearth]

models

not just real people

read there are loads of dofferent catagories

https://docs.google.com/document/d/1edY-MT2ggbnqKziw-Ej5Ee6dk4W58qFbWuIofIegDhQ/edit

and a 1000 bucks for community fave . I won that just last week for SD3 image challenge.

so i know and can assure you they pay you play

and if you join come to my profile and comment on a model so i can see your there and ill get you some cash to start with https://www.shakker.ai/userpage/8b0d2aadaa2a4f2592cbb367c329ea51/publish

just say hi dice you asked me to come from discord , sometyhing like that and ill get you free $$ for joining

[torn wharf]

i dunno. civit guys seem like lawrys guys

[bitter hearth]

would you like to know how i did it ? Will fix any confussion you have around ir then

[torn wharf]

pretty cool to nvlink two ada cards . neatoh

thats like, what? 48gb? beefy big gpus tho need a big box haha

[bitter hearth]

first of all ...... Make sure your motherboard has enough PCIe slots to accommodate two GPUs. Ideally, you'll want to use PCIe 4.0 or 5.0 slots for the best performance.
Ensure that your power supply can handle the power requirements of two RTX 4090s. You'll likely need a power supply with a minimum of 1000 watts, but it's recommended to go higher if possible.

[torn wharf]

||phalus joke||

two 4090's? i have a 1000w gpu and pushy the edge with one 4080

also 6 drives but thats beside the point

[bitter hearth]

To combine the VRAM of your two RTX 4090 GPUs, you'll need to use a feature called NVLink. NVLink allows multiple GPUs to communicate with each other at high speeds and share data. To use NVLink, you'll need the following:
Two RTX 4090 GPUs that support NVLink.
An NVLink bridge connector. This is a physical bridge that connects the two GPUs together.
NVIDIA's NVLink driver. This is a software driver that enables NVLink functionality on your system.

[torn wharf]

you talking TWO top end SOTA beefiest power chugging GPUS an you're only going to bring 1000w to the party? no go bigger. my immediate thought. that's asking for problems

no 4090s hav nvlink. i thought you modded one.

[desert garnet]

brb ill try to get one

[torn wharf]

you can't buy a 4090 that have the connector. thats what i was going for when i was like "yo thats neat"

and the silicon have the features turned off too

[desert garnet]

jensen huang edition

[bitter hearth]

mines over 1000 w but when i looked into how i was finding a 1000w would carry it

[desert garnet]

maybe its the jensen Juan edition

[torn wharf]

oh you just gotta have a dad whose frens with jens?

yeah i might know a dad. i'll look into it

[bitter hearth]

with NVLink enabled and two RTX 4090 GPUs connected, you will effectively have 48GB of VRAM available to your system. This is because the NVLink interface allows the two GPUs to share data and resources, including VRAM. So when your applications or games require more VRAM, they will be able to access the VRAM of both GPUs, effectively giving you 48GB of VRAM to work with. However, it's important to note that not all applications or games will take full advantage of NVLink, so your mileage may vary.

[torn wharf]

you just skipped over a whole part dice

bro back up

you just pasting gpt come on dude. don't hallucinate

lets be honest

[bitter hearth]

sorry bro was typing like a mad man lol

[torn wharf]

no you werent

you talkin about nvlinking 4090s is just pure gpt hallucinations. none of this works like anything you said bruh.

i just led you on cause i know you'd decimate your credibility

[bitter hearth]

I have a mic and keyboard set up to text long boring sentances rather than type

[torn wharf]

if you so lawry'd up you'd be typing at the speed of thought

[bitter hearth]

dude go look here this is even where i got the bridge connetors

[torn wharf]

buh you even vim?

[bitter hearth]

NVLink bridge connector for your RTX 4090 GPUs:
NVIDIA's website: NVIDIA sells NVLink bridges directly on their website, and you can find them in the "Accessories" section.

You can also find NVLink bridges on Amazon,

go look for yourself

[torn wharf]

https://www.youtube.com/watch?v=hzPpWInAiOg

[desert garnet]

only the 3090 has it 😔 ,thats why 3090 is a beast

[torn wharf]

thats why 3090 is still a beast its tru

[bitter hearth]

i cant see what that says bud

[torn wharf]

aw 😦

my whole point is defeated

i'm not worried about the future at all. people who bullshit with language models as aids will always be obvious. especially more so than had they bullshitted themselves. at least a bullshitter knows what domains they can play in

[bitter hearth]

https://youtu.be/zvws7zZ9TCI?feature=shared

[torn wharf]

you can get 2 in one pc. he doesn't have them on nvlink though

[bitter hearth]

easier to watch how than explain to how

[torn wharf]

no its easier to explain

i just did

[bitter hearth]

i have but your not getting what im saying

you need the bridge connector not just the NVLink

[torn wharf]

yeah this guy uses the two cards in tandem, but they aren't operating like one

rendering. classic use case

[bitter hearth]

he combines to get the 48GB

[torn wharf]

i'm not even seasoned i just know this

timestap that part. it'll be simple

you DO know how to timestamp link a youtube right?

the lack of expertise is palpable. i can't pretend its not dude

[bitter hearth]

watch from 3.24 mins in till the end

[torn wharf]

i've already watched it all in 2x speed

[bitter hearth]

now you know how even he says 1000w power sup would be ok to

mines 1200w

[torn wharf]

yeh but he knows. he knows at full tilt at the same time, you want good power

[bitter hearth]

yes

[torn wharf]

i respect the guy's build. i'm not shittin on him. he got a beast workstation for rendering and editing. its sweet

but its not nvlinked

[bitter hearth]

the point were discussing is you said you cant

[torn wharf]

his use case wouldn't need it. blender doesn't need real time frames or large memory swaps

[bitter hearth]

you can

[torn wharf]

i know what your point is. you've yet to demonstrate how you did it

[bitter hearth]

dude it can im sitting here typing to you on my rig

[torn wharf]

doubt

[bitter hearth]

doubt what?

[torn wharf]

that you got anything that qualifies as a rig

here i got a diss track for this moment holld up

https://youtu.be/QYlY5J9d4dg?t=65

[bitter hearth]

making a video as proff

[desert garnet]

if u got it working and its sharing vram it should say this on your task manager

if it doesnt say link its not working

and you are just using two separate gpus

[bitter hearth]

uploading video now

there

included our chat as proof for you

that internal monitor shows rtx 4090 GPU 1st card

and the itel i9

[craggy crest]

you're gonna get nowhere with these guys, trying to prove anything to them

[bitter hearth]

think there just trolling

[mortal mesa]

could just show the NVLink /shrug

[desert garnet]

no one is trolling here everyone knows that u cant share vram on 2x 4090

[bitter hearth]

no you mean you dont

[craggy crest]

word of advice - drop the conversation, they're not interested and if they're not already trolling, or just arguing to argue, they soon will be

[desert garnet]

you would be the first person in the world to do it should contact steven from gamernexus he could feature you on one of his videos

[torn wharf]

thats a fischer price toy

[bitter hearth]

inbox me . I have a little something for you. And from there on ill call you RAT boy lol

[torn wharf]

here comes this guy thinking that he's going to defend 2 4090's in nvlink being a thing

[bitter hearth]

haters gonna hate, Go play on your pototo pc you ill educated little boy

[desert garnet]

i mean maybe he soldered two gpu together would be first person in the world to do so

[torn wharf]

probably going to retire from the generative arts "community". this is it. theres better and theres worse and well, the worse is pretty bad. these are established people in the community behaving this way. like snotty nosed children. it's shameful. money don't equal skill.

good luck out there boys and girls

[bitter hearth]

since being in this discord server i've seen nothing good or helpful from you to anyone in here. Do you even train checkpoints or are you just a merge boy

[craggy crest]

i'm not defending anything, just letting him know the hostile envroment anyone that tries to tell you guys anything winds up in

before he 1. wastes his time and 2. hands you troll fuel

[torn wharf]

mods here don't care and will allow peeps like these two pretending to dogpile, to maintain as regulars here. shit's all fucked up when the administrators of a community side with the worst people. Hope things go well for stability still but this is just shit shameful and disappointing

[desert garnet]

the one saying that you can nvlink 2x 4090s is the real troll here,imagine someone coming here for buying advice and actually listens to him,he would end up wasting money

[craggy crest]

you'd best be happy that the mods don't care or you'd have been banned a long time ago

[bitter hearth]

does your mom turn your internet off soon . if not ill roll over and ask her to

[torn wharf]

i woudln't want to associate with this server if i felt like i had real expertise, i'll tell you that much

it would dilute me

[craggy crest]

then why are you here? or did you just admit you don't have any real expertise?

[bitter hearth]

https://tenor.com/view/the-door-is-open-you-may-leave-gif-15357849

[mortal mesa]

a perfect couple thoes 2

[desert garnet]

why dont you come here and do it yourself

[sacred jewel]

How did you get it to work tho? I asked yesterday but I don't know if you answered

https://www.techpowerup.com/299107/jensen-confirms-nvlink-support-in-ada-lovelace-is-gone

[desert garnet]

he has the 4090 jensen Juan version

[bitter hearth]

lol another basement keyboard worrior proven wrong. shame shame i know your name lol

[craggy crest]

https://www.nvidia.com/en-us/data-center/nvlink/ you use this

[desert garnet]

indeed another basement keyboard warrior proven wrong

[craggy crest]

steven is now going to state that nvidia doesn't know what they are doing and that you can't do that, even though they publish the software to do it

(do you honestly think that data center server racks don't share their VRAM?)

[desert garnet]

crystalwizard is now going to state you can use an nvlink on a 4090 like the clown he is

[craggy crest]

go to the link, bright boy, and read the information

[bitter hearth]

SD3 CEO lol you wouldnt know your ass from training a SD3 checkpoint. but you continue to pretend you do bro

[craggy crest]

or go tell nvidia they are clueless

[desert garnet]

not everyone can be a seasoned ai developer like yourself bro

[mortal mesa]

its factual the name callers always have the weakest argument

[craggy crest]

they're also the 11 year old kids pretending to be old enough to be on discord

[toxic bone]

nvlink was removed from the 40 series so they wouldnt cut into the enterprise market. everyone knows this i thought

[desert garnet]

i mean if these two clowns had an argument they would just put a screenshot of the nvlink on a 4090

[toxic bone]

why would people by the a100 or whatever.. if 2 4090s would do it

[bitter hearth]

they can , just peopkle like your goodself dont listen to advice. and for that reason you will always be the troll

[craggy crest]

and they're now being hauled into court for doing that

[toxic bone]

its a smart business call by nvidia though i don't know why y'all are mad bout it. so the 4090's can't do nvlink . stfw

[desert garnet]

if your advice is telling ppl that a 4090 can be nvlinked then it means you are full of shit

[mortal mesa]

i think one person is trying to double down while the other backpedaling, its funny

[toxic bone]

i think that got settled before it went anywhere.

[craggy crest]

would you like some more popcorn?

[bitter hearth]

dude i just showed you how, and again your not listening.... YOU NEED THE BRIDGE CONNECTOR NOT JUST THE LINK

[craggy crest]

1. he never listens 2. all you're doing is giving him fuel to continue to strut and argue.

[toxic bone]

huh?

[toxic bone]

huh?

[mortal mesa]

backpedaling

[sacred jewel]

I am well aware NVLink is alive and well... just not on 4090 as far as I know.

[toxic bone]

the whole 40 series dropped it. this is long known

[desert garnet]

you can always post a screenshot of your gpu linked on task manager best way to prove the haters

[toxic bone]

k well .. i just came for pics but y'all were aruging . i thought it was something else.

[desert garnet]

yea post the pics

[mortal mesa]

listen hes doing something thats like the same thing, that comes next

[bitter hearth]

the 4090 NV has also a version that isnt stock . go goole it

[craggy crest]

steven never listens. he just looks for excuses to argue, and when he finds one, he latches onto it and goes full speed ahead. and everything someone says just gives him more fuel to keep arguing with. he'll even twist what was said if he needs to

[toxic bone]

steven has a point tbf

[craggy crest]

perhaps, but the point is lost in his desire to have an arguement

[toxic bone]

huh?

[desert garnet]

isnt that what you are doing now

[toxic bone]

ok. i'm in the wrong place.

[craggy crest]

stick around a few months and watch him. popcorn's in the cupboard

[desert garnet]

the point is not lost,is just a fact that u cant nvlink a 4090

[craggy crest]

(he says to someone tha thas nvlink running with a 4090)

[bitter hearth]

ive been here 2 days and he has already bored the life outta me

[craggy crest]

YOU can't. other people can

[desert garnet]

yes that someone is not you tho

[craggy crest]

dude, i never said i was. now, it's time for you to get ready for your nap

[mortal mesa]

why you so nasty

[craggy crest]

he sounds grouchy

[desert garnet]

prob a redittor

[mortal mesa]

weak argument i suppose

[craggy crest]

i've been part of this discord longer than you have

[desert garnet]

thats prob your biggest achievement in life

[craggy crest]

then i have beaten your accomplishments quite well

[desert garnet]

if not the only one

[desert garnet]

you wouldnt know that

[bitter hearth]

stop arguing and make peace