✨｜sdxl

[fleet harness]

I trust him

[visual glade]

yeah frontend needs a bunch of work like a rewrite of the primitive nodes

[hardy cipher]

this is how I imagine the scene

[upbeat summit]

so we got the team ready to support comfy!

[fleet harness]

hacker guy is def a good prompt

[west breach]

comfy, could we get the history in descending order?

[hard fractal]

Let's chat next week about your wishlist!

[fleet harness]

time to work it seems

[hard fractal]

you got the whole company ready to support comfy 🤗

[upbeat summit]

I can only plan features and do some UI/UX design but I'm not good at coding 😄

[hard fractal]

My favorite little trick:

Right click a node, turn it into an input

Sure

We know that

But then, you can just double click the little green dot

And it'll make whatever node is best

[fleet harness]

[hard fractal]

Also, using shift to move multiple nodes.

[visual glade]

yeah lets chat about that and more

[hard fractal]

Also, holding down spacebar to move around.

What else...

[fleet harness]

found that one early from having used Adobe products for so long

[hard fractal]

command-enter to make a prompt go

[fleet harness]

lol

[heady vale]

a Crtl+Z undo would be nice

[visual glade]

right click the load image node -> open in mask editor

[hard fractal]

Oh yeah!

Great for inpainting

[urban fjord]

One more colour down, 16 777 211 left to go. Even without the LoRA it seems like SDLX is a lot better at red backgrounds than blue or green and this was trained a bit differently.

[visual glade]

some people gave a shot at implementing this but apparently litegraph needs to be improved a bit for it to work properly

[hard fractal]

who designed litegraph?

[alpine pine]

doesnt ctrl-z somewhat work tho?

[heady vale]

not when you delete nodes

[alpine pine]

ahh

[west breach]

I've accidentally deleted nodes before when I tapped the backspace key after I thought I clicked into a text field. So maybe only having the delete key mapped to delete a node?

[visual glade]

https://github.com/jagenjo/litegraph.js

[ionic dragon]

@hard fractal Do these 3rd party controlnets work with sdxl?
https://huggingface.co/TencentARC

[hard fractal]

No

[urban fjord]

The best would be to replace litegraph with something else.

[upbeat summit]

holding shift to move multiple nodes took some time to getting used to. because I would have expected to select multiple things with shift (like in file managers, photoshop etc) and than the selected objects can be moved with the left mouse button.

I understand it makes sense to have a precaution to not accidentally move stuff, but it is different to the "standard" behavior that is common in most applications.

[visual glade]

I picked litegraph because it was the simplest way to get a UI on comfy

[crisp owl]

Since we're on the topic of requests, could there be a rewrite which allows connected flows to not throw redwall errors if nothing is output?
So that way you could then have logic gates switch between one or the other based off a boolean, or some other future method?

Image for example of an image output switch node I wrote, but then realized the fundamental issue of NoneType errors

[visual glade]

but it's not perfect

[hard fractal]

I'll reach out. Maybe we can get him to do some things for us

[heady vale]

I learnt yesterday you can shift click multiple nodes and copy/paste into another workflow

[hardy cipher]

decent

[vital ermine]

@comfy I took that link and downloaded the picture to have its workspace but I got this when I generated.

[hard fractal]

That's fine

[west breach]

ctrl + shift + v also pastes the node with the same links as the one you copied

[heady vale]

sweet

[hard fractal]

Oh yeah that's the biggest one

[visual glade]

yeah I leave some debug info for myself, you can ignore it most of the time

[vital ermine]

alrighty

[west breach]

snap!

[urban fjord]

Unless someone mentioned it already, you can control drag to select more than one node.

[vital ermine]

It made this

[upbeat summit]

too fast for me! 😄

[hardy cipher]

I just want to know how to do 90 degree angles and straight lines

[urban fjord]

Use SwarmUI and have all the nodes hidden from view or use Auto1111 and don't worry about nodes.

[hardy cipher]

I need to try that. I keep meaning to

[upbeat summit]

I can also really recommend the ComfyUI Quality-of-Life extension package by @zealous horizon. Great helpers!
https://github.com/failfa-st/failfast-comfyui-extensions

feels like they should be native 😉

[hardy cipher]

I might have to nuke my install and start over. so many errors. it's my fault for not being mindful of what I was doing. but might be easier to just start over at this point. still works, but random errors

[vital ermine]

I am getting overwhelmed now with all of these extensions as I am not the type to have a lot of extensions only the ones I need. @vast narwhal

[upbeat summit]

just work with what comfyui comes with for a couple of days

[hoary saddle]

any stable prompt ideas for letters/words?

[upbeat summit]

than you can add features if you want or need them

[hardy cipher]

yeah, just wait a while

[fleet harness]

add cheese after you know how much cheese is on it

[vital ermine]

I need that mouse inverter 100%

[upbeat summit]

those pixel graphics look nice 👌

[vital ermine]

need a way to do a prompt where I just <lora> so I can have as many as I want not something that needs an ultrawidescreen 8k monitor that is 70 inches.

[fleet harness]

https://tenor.com/view/e-click-big-screen-wide-screen-gif-15481828

checking nodes

irl

[vital ermine]

That is how I feel

[west breach]

I made a node that supports lora:lora_name, but it's made for SDXL support only due to the conditioners

[hardy cipher]

I mean, you don't have to put them all horizontal. you can loop back, you can stack the loras vertically

you can you use that efficiency lora loader

you can minimize them

[fleet harness]

yeh, its modular and you can draw a face with them if you like

[upbeat summit]

you can put everything tightly together but navigation isn't really a problem with the canvas of comfyui. it's quite fast

[vital ermine]

Yeah, I will try that for sure and see, but what humble is talking about would be fantastic for all versions of SD.

[fleet harness]

I made my pixel art workflow arranged sort of like an old tv lol

[vital ermine]

Auto is practically dead now with XL and mem leaks that require over 64 gigs of ram and still wants more

[hardy cipher]

nice

[upbeat summit]

pics or it didn't happen

[hardy cipher]

whoa

[fleet harness]

i mean. you know

[hardy cipher]

that's nice

[fleet harness]

within reason

making that cheese

you know

[upbeat summit]

https://tenor.com/view/feel-me-think-about-it-meme-gif-7715402

arrange the nodes so it fits the current prompt / project you are working on

[fleet harness]

yes!

each prompt is so different

treat them like little personalities lol

[vital ermine]

Wouldn't it be cool if nodes autosized?

[fleet harness]

excellent

[heady vale]

is there a shortcut to convert input to widget and vice versa?

[upbeat summit]

expressive - cheese from space

[fleet harness]

the moon.... ITS CHEESE

[vital ermine]

In space noone can have your bree.

[upbeat summit]

https://tenor.com/view/wallace-and-gromit-cheese-benedict-gif-3476368

[fleet harness]

van life

[vast narwhal]

Trust me, I used to hate comfy ui, no my work flow looks like this

[hardy cipher]

that van is legit

[fleet harness]

I hated it for like 5 mintutes lol

[hardy cipher]

I never really did. just didn't know what I was doing at first

[fleet harness]

https://tenor.com/view/look-up-scared-willem-dafoe-gif-17761333

[upbeat summit]

The Daflon Machine

[hardy cipher]

still don't know tbh

[vital ermine]

Dude, I am not a Borg. 😦

[upbeat summit]

don't get sucked into those LHC machines. start simple and build from there - than it's way less overwhelming.

[hardy cipher]

I didn't realize for a couple days that I could load the images into comfy and get the workflow

so I was really perplexed by the tutorials

just pictures and vague descriptions of things

[fleet harness]

my main looks like a mess right now lol. I dont have fancy line straighteners yet.

[vital ermine]

I have seen stuff that looks like schematics. I mean I can't even follow all the routes.

[fleet harness]

right? like a CPU board or something

[hardy cipher]

I use way too many image filters. I've spent so much time doing random things with photoshop

[vital ermine]

Exactly

[hardy cipher]

I have to really hold back, lol

[fleet harness]

really its just time. after a while you realize you can merge projects etc

[upbeat summit]

yes, they look like schematics.

[fleet harness]

after a while you get more XP and lvl up your cheese bot

basically

[upbeat summit]

neither can I. you just take it apart if you want a feature. build your own. take some pointers from Sytan's workflow. it's well organized.

[fleet harness]

or become a lizard wizard... maybe join a band...

[hardy cipher]

whoa, nice

deep fried

[upbeat summit]

your settings are whack

[vital ermine]

Not sure why I get that as I am using exponential

[hardy cipher]

don't use exponential?

[fleet harness]

[upbeat summit]

I mean it could be an art form. it's almost like stained glass

[hardy cipher]

I'll see if I can load it. I'm curious about what's going on

[nocturne dove]

That's generally just not enough steps with the sampler and/or bad sampler choice. I have only used karras and normal.

[upbeat summit]

exponential does strange things sometimes

[fleet harness]

project, recue the prompt

[vital ermine]

Tried karras too and it is hit or miss

[visual glade]

make sure your base and refiner use the same scheduler

[vocal rapids]

awesome Pencil Art B&W Lora

[vital ermine]

karras

[fleet harness]

karrap

[upbeat summit]

what are you total steps?

[vital ermine]

Oh? I am using that image workflow you linked me to earlier

[upbeat summit]

there's a wrong setting

[fleet harness]

somethings afoot

[upbeat summit]

I suspect your refiner start steps are set incorrectly

[vital ermine]

I have no idea I am using one comfy sent me to

[upbeat summit]

pixel cheez robots? count me in

[vital ermine]

one has noise added the refiner doesn't and base is euler refiner is dpm

[fleet harness]

I mean, why not?! best cheese robot wins

[upbeat summit]

can you screenshot your settings? steps, cfg, both samplers

[vital ermine]

I bet this fixes it from what comfy said

[nocturne dove]

Yeah it could be start/end steps are wrong, different samplers chosen for refiner/base. Different schedulers on base/refiner can do that too but usually not so extreme. Most of the settings can cause similar looking problems if they are set badly.

[vital ermine]

this is so big my answer is not easily

[hardy cipher]

are you stripping the metadata from the images, general? I can't seem to load them for some reason. then looked and didn't see anything in the one I downloaded

[upbeat summit]

just screenshot your whole workflow so it's readable 😄 we figure it out

[fleet harness]

you'll get it. you could just snag anyones png and load up a new workflow and see how its set up etc.

[vital ermine]

this should fix it

[upbeat summit]

ahaa

[vital ermine]

imma save this so as not to forget

[fleet harness]

see not bad at all

[vital ermine]

where is the seed at or is the seed a noise seed and refiner has no seed?

[nocturne dove]

The refiner has a seed but I've seen some workflows that set it to a constant since it's not changing the look to much anyway.

[upbeat summit]

both need a noise seed. they probably use the same. which is fine

[vital ermine]

first is base second is refiner

[nocturne dove]

In that screenshot the refiner is set to a fixed seed and the base is set to randomize it.

But I guess if you don't add noise does it do anythign?

[vital ermine]

How since I do not see a seed node?

[nocturne dove]

It says "noise seed" and has a number. Under that you can set it to "fixed", "randomize", "increment", etc.

In the ksampler node.

[vital ermine]

I meant it isn't connected so how is refiner using the same one?

[nocturne dove]

It isn't. But I don't see how it would matter if it was or not.

[vital ermine]

I thought someone said they need to be the same seed?

[nocturne dove]

I don't think that is true.

[vagrant wasp]

what is XL version ?

[hardy cipher]

I keep my seed number constant when I want to refine the same image.

[vital ermine]

I see the refiner starts at step 20 and base is still going so both go to 25?

[fleet harness]

YO..your chEEse Sir..rr. r

[hardy cipher]

and I also just run the same number for all the samplers now

well unless I want different numbers for some reason

[nocturne dove]

In fact, with add_noise off, is the noise seed actually doing anything on thge refiner at all?

[vital ermine]

¯_(ツ)_/¯

[hardy cipher]

yeah, I have no idea, lol

[visual glade]

the seed only does something when add noise is on

[hardy cipher]

what's the default if the sampler doesn't have the add noise optoin?

[heady vale]

you can double click the workspace and type Seed to search. then connect the seed into the noise seed inputs, if you want it the same

[vital ermine]

I did not know refiner and base are running at the same time. base to 25 steps and refiner kicks in at 20 or am I reading this wrong?

[hardy cipher]

it's by the percentage I believe

so 80 percent at 25 steps

5 and 20

[crisp owl]

@visual glade is something like this possible to implement? Or would it require too much to fundamentally change in the way ComfyUI is written?

[vagrant wasp]

where to ask noob questions ?

[visual glade]

that would only need a change in the frontend

[hardy cipher]

general, what are you trying to do?

[vital ermine]

everything but seed was in there

[vital ermine]

figure all this out

[heady vale]

yours doesnt have it hmm. not sure maybe its a custom node?

[hardy cipher]

iif you want to run all the samplers off the same seed you should be able to right clcik and change seed to input

then drag the node out from the seed input and drop it into a random empty area

and it'll show you all your input node options

[vast narwhal]

Do you think this is possible, comfy? https://github.com/comfyanonymous/ComfyUI/issues/1051

[crisp owl]

That's essentially what I tried to do

[vital ermine]

I think these node boxes do not allow for that as these are KSampler (Advanced)

[crisp owl]

but with available nodes and a bit of a custom one I did

[vagrant wasp]

how to install 4x upscaler for AUTOMATIC1111/stable-diffusion-webui

[visual glade]

yes it shouldn't be difficult to implement in the frontend if someone decides to do it

[visual glade]

it has to be in the frontend, it's not something that can be implemented in the backend only

[hardy cipher]

I guess you didn't see what I just said, general. not sure what else to tell you

[heady vale]

save the upscaler PTH file into
stable-diffusion-webui\models\ESRGAN

[vagrant wasp]

did still does not appear

restarted too

[heady vale]

restart A1111

it should be in img2img page

or extras

[fleet harness]

looks like classic cinema

[vital ermine]

I feel naked as I did in 1.5 when I retried it as I rely upon all my models and loras, embeddings, and hypernetworks.

[heady vale]

which upscaler did you save?

[vast narwhal]

Ty

[vagrant wasp]

4x upltrasharp upscaler

[hardy cipher]

this satisfies me

[heady vale]

works on mine

[vagrant wasp]

can u send me the file

4x-UltraSharp.pth

[vagrant wasp]

thanks

[tribal jackal]

I will never tire of this

[crisp owl]

Would you know about which files would contain that portion of logic? I'd be interested at taking a look

[tribal jackal]

The hardest part is not spamming. I'll paws for a bit.

[crisp owl]

dogs are omitted from being considered spamming

[upbeat summit]

SDXL image

Runway Gen-2

[tribal jackal]

Really excited to see them progress

[crisp owl]

That's one healthy looking doggo

[tribal jackal]

I told you the hard part was not spamming lol

He's my best boy

[crisp owl]

Got 3 best puppers myself

[tribal jackal]

He's got a sis to grow up with, and they've got an ornery old bitch on her last legs to tiptoe around

[high skiff]

new 3090 gets here tomorrow

Excited to get back into LoRA's again

[tribal jackal]

I got mine down from 1.7GB to 19MB!

Still can't believe it

[high skiff]

I have seen some absurdly efficient LoRA's

hell, even more efficient than 1.5

[crisp owl]

My best bud is tipping almost 11 now, sad to see him get older, but he's still bounding around. Just never think they will get there at that age until they do

[high skiff]

didn't think it would be possible

[tribal jackal]

a 6x prompt LoRA sampler I just built, for when you do get back in

[high skiff]

Our old dog we had when I was younger made it to 17.5, and he was super happy and active until the very near end

[vital ermine]

How do I lock a seed in comfy because each gen is drastically changing?

[tribal jackal]

Oldest girl's about 15 now

[visual glade]

https://github.com/comfyanonymous/ComfyUI/blob/master/web/extensions/core/rerouteNode.js
this is the code for the reroute node

[spring fulcrum]

Does that have the whole workflow or is that just a peice of it?

[hardy cipher]

I already told you how it can be done, lol

[crisp owl]

That's awesome, I would love for him to get there, so long as he stays healthy. He's had joint cancer once removed from one of his hind legs, so hopefully it doesn't come back, then he should be good

[vital ermine]

As I showed I never could find it.

[hardy cipher]

because you didn't do what I said

[crisp owl]

CHeers, I'll take a look through, thanks

[tribal jackal]

Pick your model, your lora, and click queue

[peak dove]

<---- looking at an nVidia 4090 ... 🙂

[high skiff]

pftttt

just get a used 3090, wayyy better bang for your buck :p

[spring fulcrum]

Where do you get loras for sdxl 1.0? are any out yet?

[tribal jackal]

I made one

[high skiff]

lots of people are making their own, myself included

[spring fulcrum]

How do you make one?

I would love to do that

[tribal jackal]

Set aside a day and check youtube for tutorials

[tribal jackal]

Kohya_SS or Auto1111 are the two primary UI's to do it with I think

[autumn forum]

your making me wanna do ones for my cats

[tribal jackal]

https://tenor.com/view/dewit-do-it-emperor-palpatine-star-wars-count-dooku-gif-15162627

[spring fulcrum]

Can you send me the link for kohya_ss... it looks like a lot of people have their own version of this

[autumn forum]

now if i could just keep my cat still enough to take decent pictures lmao

[tribal jackal]

https://www.youtube.com/watch?v=AY6DMBCIZ3A the tutorial I followed

Your first model won't be perfect. Searching this server for keywords can quickly find answers to questions that may have been asked at least once before.

[boreal bough]

#🔧｜finetune message
with this is it should be easier - but captioning is still on you

[high skiff]

secourses

[vital ermine]

Where? Not there, and no right click to add it either.

[hardy cipher]

literally says conver noise seed to input

[spring fulcrum]

Wow just watched part of the video... It looks like training on a 4090 with a lot of images could take hours... possibly days to train one LoRA.. Is that correct?

[hardy cipher]

*convert

[high skiff]

^

convert noise seed to input

[vital ermine]

I don't even know wth a noise seed is

[high skiff]

noise

the seed

the image seed

[crisp owl]

that's what a seed is

[high skiff]

the seed of the noise

[vital ermine]

remember we don't have that in auto

[high skiff]

yes you do?

lol

[vital ermine]

I only had the seed

[crisp owl]

seed = noise

[high skiff]

thats what a seed is...

its the same thing

auto just doesn't lable it right

it is infact, the seed for the noise

[vital ermine]

I effing asked that way earlier and itr fell on deaf ears. I thought that is what it was but helkl

[crisp owl]

sometimes things get lost in an open chatroom

doens't hurt to ask again

[tribal jackal]

@spring fulcrum If you don't have a lora, you could just delete that node and run something like Dreamshaper XL (since it doesn't need a refiner) alone with the workflow

[heady vale]

I tried to help but your version doesnt have a seed node for some reason

[vital ermine]

Yeah, this is weird like that

[high skiff]

since when was there a seed node?

[vital ermine]

Must be why I am getting different 100% each time

[high skiff]

just set the seed to be static

[heady vale]

maybe its a custome node. not sure

[high skiff]

hmm

[hardy cipher]

but it does have a seed node. it's call noise seed as it produces random noise

[crisp owl]

change control_after_generate to fixed

[high skiff]

You can make that with a primitive

[hardy cipher]

change it to input and there will be an input for seed, or noise seed

[candid walrus]

@boreal bough i'm prepping to train a Lora for a style (on camera flash photography) i misplaced the command line you place in here a day ago and can't seem to go back and find it, do you still have that handy?

[hardy cipher]

then drag from that into empty space and you will see input options

[vale eagle]

for captioning, I tried this project https://github.com/jbmiller10/CaptionFusionator which used blip2, wd14, flamigo to caption image and summarize with llm. The result is pretty promising.

[shy kelp]

anyone know how to optimize sd?? i just want my images to generate faster. im running a gtx 1660 super with an overclock and my current start setting are --xformers --no-half --medvram. im just generating a single picture at a time but it takes upwards of 2 and a half minutes for a single picture to be done with these settings:

[uncut steeple]

That's just a primitive node connected into the seed

[vital ermine]

Heun? whoa

[vagrant wasp]

w1 = torch.bmm(q[:, i:end], k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
torch.cuda.OutOfMemoryError: CUDA out of memory.

fix ?

[high skiff]

don't use Heun, that is one of the worst samplers for SDXL

its slow, and inefficient. Almost any other sampler would be much better

[midnight shuttle]

For SDXL use 1024 native, not upscale. And use --lowvram

[heady vale]

yeah was just trying to work out why I can search for that by typing Seed, yet General didnt have it

[shy kelp]

what is sdxl?

[midnight shuttle]

The model that is the topic of this channel.

[candid walrus]

you might try Euler a and do like 20 steps

[hardy cipher]

alright, I'm not going to try explaining anything else, lol

[uncut steeple]

The latest stable diffusion model

[shy kelp]

alright ill try this

[glad grove]

2 mins for a 1024 img on 1.5 ?

[crisp owl]

Does this help ya?

[boreal bough]

#🔧｜finetune message

[candid walrus]

i've loaded that, but i don't see where the 8/1 settings are, and i don't know if you're just using the dreambooth tab and then extracting lora with the 'tools

i assume the Lora tab would be were i'd kick of the training

[shy kelp]

euler a cut my time in like half but it looks different then Heun, i was only using it because it seemed to give the best results

[crisp owl]

anything with a (ancestral) will continunally give you different results the more steps you have

[candid walrus]

yeah they are going to look different

oh derp i loaded the json into dreambooth tab hence being confused

[shy kelp]

is there any sampler that will look like Heun but perform like euler or other good samplers

[crisp owl]

relative rendering time compared to each other

[candid walrus]

would you use class images on a style lora then like dreambooth character

[vital ermine]

it broke when I did that but no big deal

[crisp owl]

broke as in?

[vital ermine]

I hit que and it finishes in 0 secs

*queue

[fleet harness]

https://tenor.com/view/rocket-jump-rocket-jump-the-rocket-jump-gif-10356223

DPM adaptive be like

[vital ermine]

Now I made it back to a widget broke still, so reload time

[crisp owl]

It does make some nice images in 1.5 at least though lol

[candid walrus]

i've got 2 aspect ratios 832/1216 and 1216/832, i'm assuming that i just put them into the same folder 1_triggerword

[crisp owl]

SOmething else is breaking then, not th eprimitive

[fleet harness]

https://tenor.com/view/klipist-streaming-short-film-tea-for-two-john-charris-gif-20642476

[boreal bough]

sketch lora grid. last column is 'intended use'. every odd column is without lora.
it was trained to be mostly used on humans - but for the fun of it I added pretty random scenarios XD

[vital ermine]

Yep, and I save the workspace only now I can''t find where it saved it

[fleet harness]

probably too much BF4

.>

[midnight shuttle]

What is the best way in Comfy to pass the same width and height to the SDXL Clip encoders for both main model and refiner, and also empty latent image node. Seems it needs the same values in 10-12 places.

[vital ermine]

it did not save it in the comfyui folder

[boreal bough]

'intended use'

[heady vale]

in your usual download folder for your browser

[hardy cipher]

this is one way to do it

[midnight shuttle]

I am using the nodes that have 2 prompts each. Each main model node needs width and height twice, and 2 nodes for positive and negative. Seems to run out of connections from the primitive.

And also need to pass the same value to the empty latent image. But primitives seem to only connect to 8 items max.

[vital ermine]

:? I never would have thought that as most save it into their own folder.

[crisp owl]

connections shouldn't be an issue

[midnight shuttle]

OK then it is a problem on my system. But that shows me the answer. Thank you.

[hardy cipher]

sorcery

[vital ermine]

By accident I just found that and thank you. 1 down a few more issues to go until I feel comfy in this.

[midnight shuttle]

I was only able to get it to connect to 8 locations and then no more connections would be created.

[crisp owl]

Add a reroute then inbetween one of them and just push off that instead

oh nvm doesn't like reroute

[vital ermine]

workspace loaded and works again

can I not use the same primitive between the two ksamplers?

[midnight shuttle]

I don't have it up right now but I'll try just rebuilding the workflow from scratch now that it is confirmed there is no limit to the outbound connections from a primitive.

[autumn forum]

sdxl is so good that half the loras i think of making sdxl can already do it quite well lmao

[vital ermine]

Same

[crisp owl]

You can

[vital ermine]

Working now so before I dunno what broke

[midnight shuttle]

I think SDXL will require a lot less finetuning for many things.

[crisp owl]

happens lol

[vital ermine]

Oh, I already see a bunch of new things it needs help with

Over baked Asians is one

[midnight shuttle]

Less finetuning than 1.5/2.1 I mean.

Not no finetuning,. but less.

[boreal bough]

absolutely. while there are issues, they are few inbetween. most things can be prompted for - albeit weirdly prompt gated

[hardy cipher]

turn your cfg down

[boreal bough]

like anime images are totally doable right now. but oh god the prompting gets weird 🤣

[vital ermine]

I am not sure a full on FT would be needed but a DB I can see

[vital ermine]

cfg 1

[earnest anvil]

amazing. is this with base sdxl?

[hardy cipher]

are you skipping steps?

[vital ermine]

I can skip steps?

this is 25 total steps with 20 to base and 5 to refiner

[hardy cipher]

well you can still have it skip the last fwe steps of the render if you'd like it helps with that sometimes. and if you'd like to know what connects to what, like I said a few times before, just drag from one of the inputs or outputs with your mouse down, then let go in empty space, and a little menu will pop up

then click search and you will see all possible nodes that will connect with that particular input or output

[vital ermine]

Yep, that is how I added the primitive

[hardy cipher]

if you'd like to know things like that

I actually don't exactly know what a primitive is. I just experimented until things worked the way I wanted

well some things

[vital ermine]

Seems to be just an integer

[midnight shuttle]

Primitive is basically a variable.

Will take on whatever type you connect it to. Integer, string, pick-list.

[hardy cipher]

gotcha. that makes sense

[vital ermine]

Now I have a static image I can work with

[hardy cipher]

once you get over the hump it shouldn't be bad

but I still miss some things from a111 like cfg scheduling

and noise offset

might be a way to do those things in comfy, but haven't figured that out yet

I guess with primitives I could schedule cfg

[vagrant wasp]

sometimes it runs only when i hit enter in the cmd console

[vagrant wasp]

otherwise its stuck

[hardy cipher]

gai, I had that same issue with a1111. don't have an explanation as to why

[vital ermine]

yeah, this hump isn't that bad but missing QoL stuff though these extension I will look at that might help that

[heady vale]

if you click and drag your mouse in the cmd window it will pause it. enter unpauses it

[hardy cipher]

,also, there's no reason you can't use both. but for me personally my resources are kind of on the low end, at least my video card. so isn't really worthwhile to run both at the same time

and these sdxl models take about 5 minutes to load, lol

[fleet harness]

gotta load fast

[vital ermine]

No, I am over, BACK, in ComfyUI because A1111 is totally screwed now when XL hit

[hardy cipher]

it was kind of a mess already. no hate on the people working on it. but just got bloated and disjointed

[vital ermine]

I mean the mem leak from switching from base to refiner and back I had a 100gig page file and I have 48gigs of ram

[hardy cipher]

and didn't really progress as it should

[vital ermine]

Yeah

Too many cooks spoiled the broth

[hardy cipher]

I appreciate it for all it taught me though. I didn't even know what stable diffusion was until around november of last year I believe

[vital ermine]

Alright I like the non gpu sde karras

[hardy cipher]

so it was good for beginner to low/medium level stuff

[vital ermine]

It has become so broke now I am not even sure it qualifies for that.

[hardy cipher]

what I like about comfy is it makes the process much more visible and intuitive, even if it's more complex and tedious at time

[vital ermine]

I just updated to the latest as I was using feb release still and omg, it became so sluggish and it was horridly slow.

[hardy cipher]

and I like to get weird with workflows, so it's perfect for me

[vital ermine]

night and day as if I was back on a core2duo

[hardy cipher]

I had to nuke it and reinstall about ever 2-3 months

[full idol]

Do you mind sharing the json file that you used for training?

[crisp owl]

It's forced me to learn more about what's actually happening, instead of just hitting big orange button to give me a picture

[vital ermine]

I am wondering if I can train XL on colab? I do 2.1 but if I do I can't use comfy with it, not yet at least.

[hardy cipher]

yes, so many of the concepts were just vague ideas to me before recently

[vital ermine]

I always loved nodes as they are pure power but I rather have the DaVinci Resolve interface/UI than a Blender type.

[hardy cipher]

I haven't used either, but it's ideal for my brain

[vital ermine]

Resolve is damn cool how they set nodes up. Top notch

You have a simple interface and can go deeper if you wish and most things are drag and drop connections so way more automatic.

can do manual if you wish just the grunt work is almost gone

[hardy cipher]

ahh,that's right up my alley. I've spent many hours running filter after filter,blending layers, etc, in photoshop

to be able to quantify and save the flow is quite nice

[vital ermine]

layers is akin to to Adobe AE and PS Nodes is Resolve and now comfy

[midnight shuttle]

The main thing I don't like about Comfy is that it's hard to iterate on something already produced. Yes I know it can be done but so can using SD fully from command line.

[vital ermine]

InvokeAI if they ever fix their current issues

[vital ermine]

what do you mean?

[midnight shuttle]

Generate a batch, click one, send to inpaint. Much easier in A1111. Can be done in Comfy. Just smoother in A1111.

[vital ermine]

Oh, yes

[hardy cipher]

I just miss using absurdly high cfg numbers. with scheduling, throttling, mimic cfg, etc, I could use mid 20s cfg values and it'd come out fine

[midnight shuttle]

Same with settings like "inpaint only masked". Can be done in Comfy. Easier in A1111.

[dire trail]

Can anyone answer me why clipdrop's sdxl is far superior than anything Ive been able to produce. Ive tried sdxl in tensor art with same prompts but it comes nowhere as close as the ones in sdxl. Has anyone been able to match the same results as clipdrop's if you have tried? Im always addicted to clipdrop's superior quality, sure the anatomy is not the best but realism is unmatched.

[vital ermine]

I don't batch much because my 1060 sucks now, and is dying, but I know what you mean precisely

do you think you said clipdrop enough?

[hardy cipher]

no

[vital ermine]

heheheh, I smell advert

[heady vale]

are you using styles?

[dire trail]

no, i found no style gives excellent results, prompts need to be detailed tho for that realism look

[vital ermine]

I still haven't figured out how to do styles in this

[hardy cipher]

don't even worry about it yet

[heady vale]

I barely understand it yet also, but I think its part of conditioning for G and L prompts etc, using separate prompt boxes

advanced stuff for me so far

[dire trail]

just look at the quality, in secs with no hassle in clipdrop.

[random bay]

2070S a1111 most recent update is taking over 10 mins per image all of a sudden

[vital ermine]

So it isn't done via the refiner?

[high skiff]

you can easily match and suprass that level of quality in comfy in also just a little bit of time

[random bay]

how am I getting 21it/s but it takes 10 mins for an image... before I avgd 3it/s and it took 1-2mins

[woeful patio]

right? you just have to know how to put a prompt together lol

[hardy cipher]

I can put prompts together, bud

not saying they're good per se, but I can make them

[midnight shuttle]

Look at the unit more closely. It might not be "it/s".

[random bay]

o shit u right

22.90s/it

I dont understand why its so slow now tho

[midnight shuttle]

Either using CPU or using system RAM due to VRAM overflow. What model and resolution? What interface?

[hardy cipher]

wow, 22.9 seconds, lolol

[random bay]

a1111, sdxl, 1024x1024
i was using this model the day it dropped and it only took me about 1-2 minutes per generation with these set

idid update a1111 today but idk how it could be such a drastic change

[midnight shuttle]

Very strange. What GPU?

[hardy cipher]

rendering 8192x9192

[random bay]

2070S 8b

[midnight shuttle]

Try setting --medvram or --lowvram

[random bay]

set COMMANDLINE_ARGS= --xformers --autolaunch --upcast-sampling --disable-nan-check --no-half-vae --medvram

these are what i have set already

lol

[glad grove]

dont use auto1111 for sdxl,its too buggy right now

[dire trail]

quality generations on clipdrop 😎

[midnight shuttle]

Remove --disable-nan-check because it will make fixing other problems harder later. Why use --upcast-sampling? That sounds like it will increase VRAM usage.

[crisp owl]

have you tried --opt-sdp-attention instead of xformers?

[random bay]

does sd next (vlad) work? cant stand the nodes in the other one

I will try your guys' suggestions rn

[vital ermine]

I am still not grasping refiner as I tell it moon, outside, etc... and it doesn't do it

[hardy cipher]

isn't the refiner more about details than objects?

details as in style

tbh I haven't really fully figured it out myself

[vital ermine]

the prompts I am seeing done no but you would think so

[random bay]

set COMMANDLINE_ARGS= --opt-sdp-attention --autolaunch --no-half-vae --medvram

trying this rn

[heady vale]

there is comfybox. basically similar to comfyui but with the same A1111 UI menu style

[hardy cipher]

I don't even believe most of the stuff I read about xl

it's just people making things up

[vital ermine]

I could not get that to work as he says to run run.bat and nope.

[hardy cipher]

two different articles or posts will say diametrically opposed things

[vital ermine]

no run.bat in his repo AND the command he says to start comfy says it was not compiled for cuda

[crisp owl]

Here's the papers
https://arxiv.org/abs/2307.01952

[midnight shuttle]

One person makes a video full of broscience and 1000 copycats repeat it as truth.

[vital ermine]

aitrapeneur was bad about that

all his settings were dog turds

[hardy cipher]

appreciate it. but my issue is most of what I find is either clickbait nonsense or is beyond my current level of understanding

wouldn't hurt to look over the paper though

[crisp owl]

It gets pretty weedy reading through the papers, but it's the actual project info, so it's the truth

most articles out there help with basic understandings, but that's about it

[shy kelp]

is there any settings i can change to make rendering faster

[hardy cipher]

indeed. worth looking at. take away some of the ambiguity. and I guess I know more than I did when I first started trying to read htose things

I just want to know why sdxl decided to superimpose images into this? lol

[crisp owl]

Rendering time will be directly related to several things. Your processing power, settings such as utilizing RAM or CPU as well as VRAM, image size, extensions adding onto the process, sampler used, etc
Not really just an easy answer

[hardy cipher]

or just paste like it's in paint or something. very strange

[vital ermine]

Now I think I am starting to get this

[hardy cipher]

nice

that looks good

[vital ermine]

I broke my normal prompt into subject for base the rest of the descriptors in refiner

[hardy cipher]

yeah, I sort of got the hang of that. but then I still don't know how to use the 3 prompts for positive and negative

how dou you even split the negative into 3 distinct group?

[vital ermine]

3 prompts?

[hardy cipher]

yes, pos_g, pos_l, pos_r

[crisp owl]

just negative to everything

[vital ermine]

I have not even experienced g/l yet

[hardy cipher]

it's in some of the node packs

[vital ermine]

I bet so

above my pay grade still

[hardy cipher]

I don't think it's necessary. but who knows. I haven't gotten great results with it

[vital ermine]

I'll really get into this when my 7900XTX is purchased at the end of the month

[crisp owl]

I've been experimenting with different combinations
Still extensively testing lol

[random bay]

60sec/it

[woeful patio]

does anyone here have a link to @high skiff 's workflow?

[vital ermine]

60 seconds per it?

[random bay]

yes

[vital ermine]

wth?

[uncut steeple]

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[high skiff]

I was about to link it lol

[random bay]

worse now

[high skiff]

beat me at my own game haha

[woeful patio]

thanks babe 💋

[vital ermine]

I know I know the answer to that

[high skiff]

MmMmMmMmMm

[vital ermine]

you are paging out and it will get worse and worse

[hardy cipher]

has anyone combined an unclip model with an interrogator? I'm curious how that would go

[uncut steeple]

what is it?

[lilac wren]

Hey General, yes, i think too

[high skiff]

My new 3090

[uncut steeple]

Ah neat

[vital ermine]

I know so because I have a 1060 and got that all the time until I had extra pc ram as it was paging to the pagefile. god that was bad

[hardy cipher]

these images keep getting more and more deepfried and it's really grinding my gears

[vital ermine]

this is why 7900XTX with 24 gigs will be nice

[lilac wren]

RocM 🤮

[vital ermine]

oh, no rocm is on windows now with hip support

[lilac wren]

Yeah, and with every new part of code you will wait 2 monthes

[vital ermine]

what ticks me off is not the rocm rather pytorch and tensorflow seemingly paid off to drag their heels on full spport

[hardy cipher]

they are a hassle

[vital ermine]

I am happy to wait a lifetime over nvidia ever again.

done with Jensen, now if Jensen died and a new ceo arrives I would give Nvidia a go again, or if they return to sanity

They play their malovelent game I go somewhere else with glee. EVEN if a tad slower although it really isn't.

[glad grove]

we happy to announce our new CEO : Jensen Jr.

[vital ermine]

abort

post birth abortion. Hey, I am 35. So? poof

Anyway I need a card as mine is dying on me and no way will I buy used and end up like Sytan or pay current 4k prices

[heady vale]

Ive seen people doing gens with AMD cards, what about training, like you want to do?

[vital ermine]

yes, it trains but iffy because of pytorch and tensorflow although with hips it just uses cuda

[lilac wren]

buying used fragile cards like a 3090 is not a good idea yes

[vital ermine]

I have not had hips tested as it is far too new

[vital ermine]

I know this the 4090 is bottlenecked even with SD. Amazing

[crisp owl]

Elenore

[hardy cipher]

might want to grab an a100

that'd do the trick

[vital ermine]

I look at it like this. I need a card and I expect poster gamercon the MSRP to drop 10% on the 7900XTX, maybe 15%. ANd it trains faster than Colab and gens 35-40it/s on 2.1 on SDXL no idea

[uncut steeple]

Love it

[crisp owl]

cost aside, (for sell value), I'd still take the stang over any "supercar"

[vital ermine]

colab is horrid, btw, as the T4 is 1.0x s/it at BS1 for most training. That is about the speed my 1060 would do if it had 16GB of ram

[woeful patio]

@high skiff is there a reason why your template is only set at 25 steps?

[vital ermine]

I think the whole hips thing is something we download as I saw it was around 2gb

[high skiff]

Thats the optimal threshold where additional steps don't seem to benefit quality much. In some instances, it can degrade quality

[boreal bough]

unless you're making hands/feet, it's the ideal performance/quality

[woeful patio]

ah got it , Is that in all of sdxl? i always used around 45 steps in 1.5

[boreal bough]

purposely reduced for all of sdxl. even if you do base only, no refiner, you'd still do 25 steps for most situations

[high skiff]

thats my findings for the sampler/scheduler I use, which has yielded the best results for my workflow

[vital ermine]

Sytan has always been a 25 or less

[uncut steeple]

Yeah steps depend on the sampler as well

[woeful patio]

Ah that is really good to know, probably saves on time too

[high skiff]

for 1.5, I used 15

Anything higher was kinda pointless for DDIM

but SDXL does need 25, any less and it can get kinda wonky for DDIM

[boreal bough]

50 is the theoretical 'ideal' - but quality improvement is very minor, for double the generation time

[vital ermine]

Auto1111 no workie for ddim in XL

[high skiff]

not surprising, Auto is even more of a shit show for SDXL haha

[vital ermine]

it is, I know

[high skiff]

only thing I have it for is inpainting, as inpainting is very convoluted in Comfy

[woeful patio]

@high skiff Anyway, truly an amazing workflow, it looks so much better then the other templates i've tried

[high skiff]

Glad you like it. Dozens of hours of testing adn 1000's of images went into getting it how it ended up.

I still have some improvements I will put out on a 1.1 update when I recover from the toll that the 1.0 crunch took on me 😅

[ionic dragon]

i hope your testing becomes faster when you get your 3090

[high skiff]

It should ge much faster

over 2x the speed compared to my 3080

[high skiff]

tho the benefit of the 3090 is for training LoRA's

lmfao

@fleet harnessHow did you even get that on your profile? lol

[high skiff]

straight people

[ionic dragon]

i think you can use dual gpu in stableswarm

[high skiff]

I have no interest in using stable swarm, and I don't have case room for 2 GPU's

[fleet harness]

https://tenor.com/view/hack-the-planet-gif-15624885

[woeful patio]

upgraded to a 8gb 3070 thinking it would be great, really regretting that about now lol

[high skiff]

what did you come from?

[woeful patio]

1060 so it was a real upgrade

[high skiff]

oh for sure

[ionic dragon]

for how much didd you buy it for?

[high skiff]

3090 is only so much faster than the 3080 cause it can run --gpu-only

[jolly creek]

Your work is truly appreciated, love using your workflow, thanks!
Finally decided to try comfyui over A1111 and love it

[woeful patio]

350 euros so not to bad

[high skiff]

That is the best thing ever to hear

So glad to get people into proper SDXL control and away from that dumpsterfire 1111 lol

[uncut steeple]

Thats less than what I paid for a 3050 almost 2 years ago

[woeful patio]

It was second hand tho, but still as good as new

[uncut steeple]

That reminds me I still need to sell of my 3050 and the two old screens

[vital ermine]

The one thing that came with this gen of AMD cards that I don't think anyone is taking advantage of is the AI accelerators. Only 7k cards have them.

[high skiff]

RocM uses the Ai accelerator cores

[ionic dragon]

rob NVIDIA

[woeful patio]

you know nvidia is putting low amount of vram in there on purpose so you have to upgrade again later

[high skiff]

for sure

[vital ermine]

very good as the 7900XTX has 192 of them I think it was

[trim orbit]

when i used rocm on arch it was great

i did a lot with the 8gb i had

[vital ermine]

rocm is on windows now as is hips

[uncut steeple]

Thats why I got a 4090 now, to be settled for a while

[vital ermine]

only thing lacking is pytorch and tensorflow support for 7k cards

[ionic dragon]

for a while = 6months

[alpine pine]

thats a solid while

[uncut steeple]

I have 2, so maybe 12 months

[buoyant axle]

@high skiff Was trying ur workflow yesterday as well, ty for ur work. So far i did understand everything in there except one thing. You are feeding the clip text encoders with 2048 for width and height while the empty latent image is 1024x1024. I dont really know how the text encoders work, so thats a bit confusing for me. So, can u tell why u are using 2048 for the text encs?

[alpine pine]

A larger text encoder size can potentially improve image quality as it searches for higher resolution CLIP images. But if you scroll up, there was this whole discussion on what is a good text enc width/height. Nobody seems to have a definitive answer, so it is worth experimenting on.

[ionic dragon]

hopefully if you use stablwswarm

[trim orbit]

i've got the thirst already. zombie for that silicon

[heady vale]

If I remember you have a 4080. hows that going with XL?

[buoyant axle]

i see, ty. I think i have to do a bit more research on how the text enc work, because i always thought the width/height has to match with latent image size to give proper results. my bad

[trim orbit]

very well!

[heady vale]

can you use --gpu-only?

[trim orbit]

probably not

[high skiff]

--gpu-only will max out a 4080 pretty hard

when I had my 3090, GPU only used about 16.8GB VRAM when idle

[alpine pine]

A solid bet is your image res, but experimenting can lead to maybe better or maybe worse results

[soft zealot]

generating a 2048x2560 image such as this on a 1080ti takes approx 250 seconds,

Sure its slower than a newer card but its hardly an eternity (and used 1080tis are cheap)

[high skiff]

since it keeps the base, refiner, and TE's all in VRAM

[alpine pine]

you can even use seperate positive and negative encoder resolutions

[heady vale]

ah ok for 4090 is basically it then

[vital ermine]

that is 2 extensions down

grabbed the manager first

I really like this manager

I just grabbed this but why, as I forgot now? https://github.com/LucianoCirino/efficiency-nodes-comfyui

[uncut steeple]

One is at home and one in my personal rig so... yeah

[woeful patio]

you have to disable the refiner if you're gonna use loras right?

[uncut steeple]

Not in my workflow

[soft zealot]

no

[woeful patio]

where workflow

[soft zealot]

who workflow?

[vital ermine]

Damn, all the old names are back

[ionic dragon]

tyou are using one as a door stopper?

[uncut steeple]

havent released the version with lora yet

[uncut steeple]

No im doing SD in the office and gaming at home

[soft zealot]

just put a LoAD LORA node in between the Base Model & the Base Conditioners & First PassSampler

Dont put one intot the refiner step

[ionic dragon]

SD in office?
lol
you work full time working with comfyui

[vital ermine]

my workflow has no conditioners it is just load model directly into the sampler

[uncut steeple]

Not full time and not only with comfy but im doing AI stuff as well yes

[ionic dragon]

cool

[livid cradle]

hey guys, anyone here interested in a Product Photography App using SDXL?? Check out out upcoming project: https://creativio.io

[uncut steeple]

I also do 3D, video and photo editing as well as layouts for advertisments, catalouges and stuff

[soft zealot]

so your model doesnt have these sort of step in then to condition the plain text input using CLIP??

[uncut steeple]

Thats why AI is very handy

[soft zealot]

fairy snuff 🙂

[vital ermine]

I just touched XL today so not really gone ro that l/g stuff yet

Well, hmmm

[soft zealot]

so on yours you would add a LORA Loader saround here , anything the "Model" goes to connect to the Modesl out of the LORA and same for the CLIP

Then connect Model Loader to the LORA LOader

[vital ermine]

Yeah. I am still trying to grasp refiner

[woeful patio]

what is the difference between strecgth model and strength clip?

[crisp owl]

When I changed strength clip, it did nothing at all. Just the strength model did. So I just leave cliip at 1 and bother with the model number

[fleet harness]

Conan O'Brian

[peak dove]

I'm looking at upgrading my RTX2070 to a 4090

[full idol]

Overtrained lora?

[ionic dragon]

get 2x3090, thats better

[peak dove]

OK tell me why? 🙂

[ionic dragon]

if you use stableswarm its better, you can use multiple gpus

24gb+24gb vram

[woeful patio]

has anyone succeeded in training loras on 8gb vram?

[vital ermine]

100% XL base

[peak dove]

So it'll need a completely rebuilt PC?

[ionic dragon]

you can just get a gpu case

and a motherboard, if your existing doesnt have 6 pin pcie slots

[peak dove]

Eau keigh

[ionic dragon]

you can get 3 pin 3090s

[uncut steeple]

Sounds like a decent upgrade to me

[vocal rapids]

pencil sketch is interesting ^_^

[peak dove]

L8r

[rustic garnet]

how strong the weights from text encoder are scaled (clip) and how strong the weights of the unet are scaled (model)

[ionic dragon]

are you from usa?

[wicked frigate]

clip is textencoder, model is unet. A lot of loras train only the unet, so the clip part does nothing for those

[ionic dragon]

models train the text-encoder too?

[charred beacon]

Spicy italian meatball

[wicked frigate]

they can yes, they don't have to

in SDv1 it was pretty commonly found to be helpful to train it a bit. In SDXL the textenc is way more powerful so it's less needed

[ionic dragon]

so if we want to add specific keywords into the model, we should mess up with the text encoder?

[wicked frigate]

adding new keywords is a lil weird to do, generally Textual Inversion is the preferred approach on that

training the text encoder is more to affect how it reads inputs

if you're eg training an anime model over the base, you might want to train the textenc so it learns to interpret the anime-style tag-list input better

[willow laurel]

hi

https://tenor.com/view/funny-smile-best-laugh-thats-funny-gif-8282901

[trim totem]

Does anyone have a colab script for running sdxl in automatic 1111? I'm a beginner and not very good in coding.

[floral island]

love the dude on the right tho

[languid iris]

He’s really jazzed just to be there

[ionic gulch]

what's the best practice for comfiyui to make a part of the workflow optional? is the only way to use two redirect nodes and cut the line between those two nodes?

[rustic garnet]

what is a good prompt to make a dark photo? I have the feeling SDXL still has it's brightness issue. Cannot make a photo at night without lights

[charred beacon]

I struggled with night time too

Try starfield, to make the night sky

[floral island]

didn't they add the lora for that?

[rustic garnet]

in general SDXL 1.0 makes images much brighter than SDXL 0.9 :/

[floral island]

hands are still the stuff of nightmares!

[somber hill]

sdxl myself :d lora

[vocal rapids]

lots of fun

[uncut steeple]

Updated Workflow but still WIP, now with a switch to enable LORA loading, wildcard support with Impact node and a unique output name for images using the date and time together

[languid schooner]

It created some strange decal on the car, but...a Lamborghini Huracan in Italy, came out okay. lol

[royal fern]

did anyone try the SergeZT's controlnets for sdxl?

[south frigate]

Is there any list of plugins that work with SDXL?

[west breach]

If you install the manager that has a list you can browse https://github.com/ltdrdata/ComfyUI-Manager

[barren bluff]

He is definitely entertained

[trim totem]

Does anyone have a good colab script for running sdxl in automatic 1111? I'm a beginner and not very good in coding.

[lilac wren]

A1111 already run sdxl

Just update

[floral island]

song lyrics win again?!

[uncut steeple]

We back to this?

[floral island]

well, can you argue with the results?

[uncut steeple]

[lilac wren]

Very nice

[copper willow]

Can you point me in the direction of those nodes?

I'm using wildcard files that have loras in them, so using the normal lora loaders is a no-go.

[median isle]

I hope the stability team @'s the entire server when ControlNet drops

[molten gull]

hi

[median isle]

I've gotten my workflow into a that I feel good with, now all I need is just some dynamic poses and we're golden

[molten gull]

do more of you guys use the LoRA_Easy_Training_Scripts ?

[fierce hollow]

I use the difficult training scripts, it builds character

[dense chasm]

sdxl LoRA training result is not bad

[lone cradle]

giant dwarf

[urban breach]

stunning

those loras you made?

[hollow halo]

Yeah, I was trying to train a LoRA on horror stuff but it ended up learning a high contrast style that does dark pretty well. Not mad about it, it was a happy little accident

[urban breach]

really amazing stuff

[hollow halo]

With just a little inpainting I might use this a desktop wallpaper for a bit

[woeful patio]

is there any news on controlnet?

[hollow halo]

This one might be ready for an initial release here soon

[upbeat summit]

this is the latest: #✨｜sdxl message

[woeful patio]

cool, guess I'll have to wait a little bit

[hollow halo]

Might be a little gory for some, so marked it as a Spoiler, just a little blood and skulls

[urban breach]

I'm not an artist, but this has gotta be some kind of contender for something imo 😍

[woeful patio]

why do i get something like this when use the prompt painting of?

[upbeat summit]

hey it's you 🙂

[woeful patio]

great job man, Your loras look a lot better then mine for some reason. Thanks for all the tutorials

[somber hill]

yep

[somber hill]

thanks a lot

i am making a new tutorial for 12 gb vram

[upbeat summit]

nice 🙂 thanks for all your work!

[upbeat summit]

firehound!

[hollow halo]

Have you trained any style LoRA's or just LoRA's on your face? I've found they require quite different settings, though for likeness have only used my dogs to test

[somber hill]

i was on vacation

so only face right now

next i am planning dreambooth sdxl hopefully

[hollow halo]

Looking forward to that, haven't been able to get dreambooth working properly yet, but have had little time to do so

[trim totem]

@somber hill I'm just watching your video on Lora Training. Do you have by any chance a shared colab script to do it?

[hollow halo]

Testing random Lexica prompts with my LoRA

[woeful patio]

is it me or does sdxl not follow the prompts very well, like you say one thing and sd just does something else without elaborating

[hollow halo]

It does seem to like to do that

[woeful patio]

for example the prompt here was: labradoodle dog, doomsday, nuke, explosion | centered| key visual| intricate| highly detailed| breathtaking beauty| precise lineart| vibrant| comprehensive cinematic| Carne Griffiths| Conrad Roset

that does not look like a dog to me but idk

[hollow halo]

Lol yeah I see no labra doodle dog there

[west breach]

some words just overpower everything

[upbeat summit]

yeah precise lineart and the artists are probably very strong tokens here

[woeful patio]

there should be a way of putting emphasis on certain words,

[upbeat summit]

you can with (weighting:1.0). in SDXL it makes especially sense to de-emphasize tokens that you want less of

[hollow halo]

Select the word, hold control and use up or down arrrow keys to add or remove weight

[upbeat summit]

see this for some examples how weighting works in ComfyUI:
https://blenderneko.github.io/ComfyUI-docs/Interface/Textprompts/?h=weighting#up-and-down-weighting

[woeful patio]

holy shit this changes everything

[upbeat summit]

also remember that a prompt is being processed from the beginning to the end. so a token at the beginning has more strength than a token that comes after it. with weighting you can change the balance, but it's also a good thing to remember.

[vocal stream]

that's only true for one of the text encoders while the weighing affects both I believe (though not completely sure)

[tribal jackal]

My Dog LoRA is up!
https://civitai.com/models/121064?modelVersionId=131716
https://www.reddit.com/r/StableDiffusion/comments/15g7obv/bestboynido_lora_now_available_under_19mb/

[cloud sparrow]

whats the best way to run sdxl locally? iv heard sd webui is not recommended/buggy?

[upbeat summit]

congrats to the release! now Nido is eternalized in the latent space 🙂

[median isle]

Hey I made a LoRA with your video as reference! Cheers man, I had great results!

[tribal jackal]

Same! That LoRA I just posted in fact lol

[thin nova]

Any info on open sourcing the discord dream bot like emad briefly mentioned?

[thorny frost]

What does the | (vertical bar) means in the prompts?

[polar epoch]

Lora's exists for a reason.. lol. I can't with these attempts of luffy lol

[polar epoch]

Adding "realistic photograph" made it way worse

[molten gull]

it's a generic thing @polar epoch

[wet nacelle]

https://tenor.com/view/jerma-jerma985-laughing-hysterically-chin-greedy-grinner-gif-19552901

[molten gull]

people who do realistic photographs dont tag their photos with "realistic photograph of a duck"

[polar epoch]

Damnit SD, monkey is his name, not his species

[molten gull]

so the ai doesnt really associate "realistic photograph" with a realistic photograph

[polar epoch]

there we go lol. Luffy D, exclude species to not screw up poor strawhat lol

[thorny frost]

'Realistic' is more a late 19th century painting style

[boreal bough]

"realistic" defines an artstyle. so ironically enough, 'realistic' pushes it in the direction of artwork 🤣

[thorny frost]

Gustave Courbet for instance

[polar epoch]

Hey, i want my 3d luffy realistic, but unreal ok?

Now, lets take naruto luffy here and give him a actual photograph look :P

[tough wave]

有人知道如何实现图生图吗？

[dense chasm]

with refine model,it's actually img2img

[uncut steeple]

Heyo the server language is English, if I understand your question correctly img2img might be your solution

[tough wave]

I want to use a photograph as a model, but how can the original photo be turned into a model?

How to implement img2img, if the first img is an old high school photo?

[wicked dagger]

any advice for getting good text out of SDXL? My results are so so

A professional photograph of "Ken's Pizza" restaurant at early evening. Show the exterior with large windows, and feature a neon sign spelling "Ken's Pizza." Include some urban street elements for a lively scene.

[soft zealot]

well that's as close to a progress bar as I'll get peobably (just under the preview images)
*
NB save anyone asking workflow embedded in this image, dragndrop into COmfyUI*

[trim totem]

does anyone know what the difference is here?

[meager canopy]

The embedded VAE version in each

[soft zealot]

ones got the sdxl1 vae baked in the other has the 0.9 vae baked in which fi8xes issues apparently for some occasions

[trim totem]

So Comfy UI from Searge I don't need the embedded vae version?

[soft zealot]

you can use either

your choice

noones holding a gun to your head 🙂

[trim totem]

haha

ok

thanks

[soft zealot]

personally I use the 0.9 vae sperately which is another option

or I might switch

depends on how bad the sunspot activity is

or the rain

[indigo vine]

pharaoh poker face

[wraith tide]

Comfyui or webui is recommended for deploying SDXL

?

[soft zealot]

yes

[wraith tide]

It is recommended to use comfyui or webui to deploy SDXL. Which of these two is better?

[meager canopy]

That depends on personal taste.

I prefer ComfyUI, but it's not for everyone.

[urban fjord]

I recommend Auto1111, others recommend ComfyUI, so use what you prefer.

[soft zealot]

there are multiple ways of using the SDXL model, ComfyUI,Automatic 1111, InvokeAI etc etc

Pick one any one, there is no right or worng answer

Its personal preference

That said I'm now a COmfyUI convert, its worth the learning curve

[meager canopy]

I'd recommend SD.next over A1111 though 😄

[soft zealot]

technically COmfyUI is a web UI

So is Autmatic 1111, so is Invoke, So is Vladmatic.....................

its like not all vacuum cleaners are hoovers or not all ball pont pens are Biros

[full idol]

Holy smokas

Sketch style lora in SDXL is nutttyy

[wraith tide]

Autmatic 1111 are running into issues

[uncut steeple]

#🤝｜tech-support

[strange mist]

refiner is bad for artwork, imo 🤔

this one's without the refiner

[molten gull]

i cant really find a lot good use in the refiner at any place

[strange mist]

I think it's good for photographic images.

[molten gull]

not if you throw in any LORA

[strange mist]

haven't tried any LORA with SDXL yet 😄

[molten gull]

you should, it's awesome

and easy

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

installs like a piece of cake, too

[strange mist]

I would. once I get a hang of the comfyui workflows 😅 won't be training any though.
so far, I've been stealing ppl's workflows and trying them out

[lilac wren]

Excepted if you don't understand half of settings

[soft zealot]

yeah dont "steal" mine its setup as a daily driver layout with all the workings out buried away#✨｜sdxl message

[molten gull]

i could run you through the settings if you like to

@lilac wren

[lilac wren]

Please, why not, I'm at the swimming pool with my children, they are having fun and I'm pissing myself off.

[meager canopy]

What did you use to embed the flow into the screenshot?

[soft zealot]

black magic lol

[soft zealot]

https://github.com/wyrde/wyrde-comfyui-workflows/tree/main/misc

[meager canopy]

Thansk 🙂

[prime depot]

Does anyone know if stable diffusion xl uses v-prediction or epsilon-prediction?

Help!

[soft zealot]

both |I think#

I seem to recall that the base uses one and the refiner the other but I could be wrong

or just plain confused lol

[prime depot]

Where did you see that? I want to know what method base use. Thanks!

[soft zealot]

some YT video hidden away in my history lol

I think its in this one

https://www.youtube.com/watch?v=Wk88rK9Zh1Y

[prime depot]

Wow, thank you so much!

[south frigate]

hi, i have question - thats normal I have only 2-3 it/s when rendering with SDXL? its SD 1.5 I have like 9-15 it/s... GF 4080.

[wraith tide]

What is the difference between these two basic models?

[pallid path]

oh that's just the vae

[pallid path]

it's a smaller model that you should use as a "VAE" in the options if you're in A1111

or load and plug in the VAE in Comfyui

[naive goblet]

0.9 Vae is baked in

[pallid path]

that one is required if you don't want weird colouring around edges close up

oh now it's baked in?

nicee

so that's just an improved model, not a separate vae

get that one

[wraith tide]

The size of the model ending in vae is the same as that of the model without vae

I still don't understand the difference between the two.

[wet nacelle]

@wraith tide

[royal fern]

trying to squeeze as many details as possible out of SDXL... a little creepy but the details are impressive

[thick mist]

What kind of images do you guys use for LoRa training on a concept?

i have only done a celebrity but i want to try a concept now

[royal fern]

I need more vram 😦

[soft zealot]

they have different VAEs baked in. Just after launch of SDXL1 some people were rpeorting odd artifcats so initial advice was to DL & use the 0/9 VAE, thats now available as a baked in option