✨｜sdxl

[crisp owl]

yeah it'll bork things for sure

[nimble heart]

ive seen nvidia users report that a lot. try restarting the server and re-running the same seed see if it changes or is coincidence

[peak dove]

SDXL from Distillery @upper fractal

Free SDXL generator in Alpha Test Phase

[nimble heart]

man why are light paths so easy but basic anatomy is usually so hard

the specular on surfaces is always pristine lmao

[uncut fiber]

there were image of x-ray hand which got bone where 6th finger is appearing. I think it will be better with anatomy.

[pallid path]

god I love the TEKN series

_ _ 🅱️
TEKEN

[uncut fiber]

M.I.B.

[shy kelp]

Someone knows what the commands are to enable animations in python?

[pure crystal]

math based with https://3b1b.github.io/manim/getting_started/example_scenes.html

[shy kelp]

If loras are already included in a sdxl model, do i still need to write "Lora example" in the prompt?

how does it even work?

[short prairie]

With noise injection between the samplers.

[weary yacht]

what are some of your guys' favorite models for SDXL?

[uncut fiber]

Jugernaut v3 or 5

[weary yacht]

what does juggernaut do over the base?

[uncut fiber]

Not exactly sure, but i like it in general. For general purpose. I think base isnt at all bad, just using jugernaut, and for hiding pictures in pictures or text in pictures, because only for sd, i am using i think realistic vision SD in CN in A1111.

[weary yacht]

i might give it a try and test out some stuff with it

[uncut fiber]

Curious if this is related to Soul's avatar or not 🙂

[cyan crown]

SDXL standard.

[versed merlin]

a surreal airbrush illustration of vibrant flowers coming to life in a dreamlike world. Let your creativity bloom with vivid colors and imaginative forms.

[uncut fiber]

@versed merlin #1100170312106127410 - #1101178553900478464

[uncut fiber]

Scary should i censor it? I think not. The first one got his left side issue 🙂

[crisp owl]

Nah, my 2yo kid says, "elk man"

Green face guy he said, "green elk man, silly red eyes"

[pallid path]

Jackson the Michael

[half cedar]

@ivory blaze

Beetlejackson

[hoary saddle]

comfy will be great for creating game assets, pulling directly from python to call an image directly into the game

just use variables and arg values to supply the text to the nodes

[cyan crown]

last one

[slender coral]

Can someone explain the differences:

[hasty smelt]

Hey guys, does anyone have a good prompt to share? I want to try creating some simple vintage black and white illustrations like the example. If you can help me out, I'd be happy.

[wooden pagoda]

Anyone got a prompt that creates a clean visual representation of a neural network? Or else a Lora or some embeddings? I'm lost and can't create one in SD (Automatic1111 webUI),.. any help appreciated, thanks. I tried Img2img and it just copies the existing one I got from the interwebz,.. hardly original or accurate.

[half cedar]

https://archive.ph/JJiih use python

[half ivy]

seek

hello team

i am here

[vital ermine]

sad xl died, but rumor is 3.0 might hit shortly. I hope only a rumor.

[crisp owl]

huh?

[heady vale]

anyone played around with the new SamplerCustom node and know how to set a denoise value with it?

[nimble heart]

died?

[vital ermine]

Yep

users went back to 1.5 and just look at the state of all discords people gave SAI the middle fing, mostly, to SDXL.

Olvio had a vid about this recently and people just aren't using SDXL.

[nimble heart]

idk 90% of SD social media posts that gain traction are just those animated things which I dont think has tooling for XL yet

so i wouldnt write off XL just because some tools arent updated yet

[vital ermine]

Oh, I have written it off and it died faster than 2.1 did and I stayed with 2.1 despite that.

[nimble heart]

1.5 can do most things 2.1 can but XL is different enough there's things only it can do

like 4k denoising

idk i havent had any reason to use my 1.5 tunes from XL unless there's some extremely specific lora or conditioning needed

[vital ermine]

Not gonna go into it but the people aren't having what XL gives. They are firmly happy with 1.5 and as one person wrote about 3.0 "If SAI doesn't rein in the resource requirements of 3.0 it will be still born".

[nimble heart]

most video games dont even run on 8 gigs of vram anymore nevermind SD lol

if people wanna use their 6 gig cards they'll always have 1.5 but idk why other people with 8-12+ cards would ever use it over XL outside tooling

[vital ermine]

Just telling you the way the public feels and thinks. When there is one that does 1536x1536 natively and only reuires 4-8 gigs yeah, sai needs to get it together.

[crisp owl]

The unhappy are always the loudest

[nimble heart]

yea i only post SD stuff if i make new nodes or something I dont post random renders

[vital ermine]

tools suck for it ngl. As a trainer none of us are happy with the BS SAI does. pump and dump is for ex hedge fund managers not for stuff like this.

[nimble heart]

last actual picture I posted was just demoing the colored latents

[nimble heart]

was 1.5 any different? it was just a plain model uploaded and everyone built their own tools around it.

[vital ermine]

I expect no less from people here as it is akin to walking into a dnc office and saying you are voting Trump.

fact is I have a 4090 so all moot to me but it sure sucks when we lack the tools to train XL because SAI just dump the mdel and ran off to work on 3.0

[nimble heart]

not even a 4090 just like an RX 6800 will get you 90% of the way there with most inference tools today.

[vital ermine]

1.5 tools for XL is not the way and this being so different with two tes and noise offset being used on base (a total BS move) it really hurts. eventually the tools will catch up but which first new tools or 3.0 before they can be even made for XL?

[nimble heart]

i wonder if they might rush 3.0 specifically because the tools arent catching up

drop the refiner improve the offset noise etc

make it easier to work with

but i mean idk. in the end we're not their customer base, their base is private tunes for enterprise iirc. so totally different world

[crisp owl]

Yup, end goal is some hefty contracts from corporations

[nimble heart]

wonder if you could 8bit quantize an image model and not have it get super fucked up

like how the llama people do it

the quantizing game is pretty advanced now with exllama2 and stuff. quants usually outperform everything else gig-per-gig of vram

so 8gig people could run a quanted version

i mean for 3.0. XL already runs on 8 gigs iirc

if you use xformers

[vital ermine]

I dunno but the german one has potential and a lot of trainers are watching it closely. Onetrainer is making a training tool for it so that will help immensely.

[nimble heart]

i have no idea what you're talking about

[vital ermine]

that is the one that is faster than SD and does larger gens

[nimble heart]

faster and larger but is it better?

or is it like dalle-mini at 2 megapixels

[vital ermine]

wurstchen 3

when the trainers get to it they will ravage it and make it better as they did with 1.5

I have a dataset that is pissing me off

refuses to train right and just blew up on me in 300 steps using adafactor

[nimble heart]

that's already in diffusers it seems. could try it out yourself and repot back

[vital ermine]

btw, my testing in comfyui shows the node leaks 😦

no, 3 is still being tweaked

2 is

I am on their discord

[nimble heart]

Ah

[nimble heart]

disable the smart mem

[vital ermine]

try it out you will see the noise and stuff but they are working on it just a couple of guys

smart mem?

[nimble heart]

--disable-smart-memory

[vital ermine]

let me try that

[nimble heart]

fixed the vram slowly crawling up until HIP died for me

[vital ermine]

yep, not in there

[nimble heart]

?

[vital ermine]

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --use-pytorch-cross-attention --gpu-only --dont-upcast-attention --disable-xformers

[nimble heart]

what about it

[vital ermine]

you ? there is what I was talking about

[nimble heart]

yea try adding --disable-smart-memory to your command

[vital ermine]

adding it now, thanks

[nimble heart]

Sytan said it helped with multi-model stuff too

for making grids

[vital ermine]

not sure why it was leaking because it should be plain on left and lora on the right but the lora was affecting the plain

they aren't even connected

[nimble heart]

ohhhhh I thought you meant memory leak like number go up

[vital ermine]

no

[nimble heart]

that still might help tbh since it'll unload more aggressively

[vital ermine]

the node leaks the lora

[heady vale]

try the new comfy with pytorch 2.1 it doesnt use xformers https://github.com/comfyanonymous/ComfyUI/releases

[nimble heart]

or just enable the pytorch attention via cli since that's what it's doing anyways I assume

[vital ermine]

12.1

my comfy still on 11.8

I think.

lol, I forgot this is windows and he doesn't use a venv in windows

[nimble heart]

ay you have a 4090 you should try the new brainfloat unet

get that infinity+1 batch size

[vital ermine]

when I train I use bf16

what does that really do for 3k and 4k cards doing unet in bf16?

[nimble heart]

how is it leaking lora? like you ctrl-b bypass it and it's still there?

[vital ermine]

I use the comfyroll node that has the on and off switch. I turn it on get pic X for left side. Turn it on get a different pic for left side.

left side doesn't even see my lora it isn't connected

[nimble heart]

idk tbh I suppose the unet autocasts to fp16 by default so it'd maybe give slightly better precision?

[nimble heart]

y tho

just ctrl-b the builtin node

[vital ermine]

I like the side by side

when I train I need that info or back to auto I go

[nimble heart]

if you have to click the switch anyways...

[vital ermine]

I don't, that is the point

[nimble heart]

im not understanding I guess. when you say switch i'm thinking bool toggle

[vital ermine]

left side is always without lora and right is always with (or make sure turn it off and should be the same) only with the lora anywhere the base left side is not the same

[nimble heart]

is it running in the same batch?

[vital ermine]

nope

[nimble heart]

you're gonna have to screenshot it cause i dont know what the fuck is going on

[vital ermine]

left side right side.

[nimble heart]

where even is the lora then

[crisp owl]

From the workflow I posted the other day showing two image results from the same prompt box, one with Lora and one without lora, I do not get any of that happening in my flow.

Something is messed up in yours if that's happening

[vital ermine]

okies, cool beans and good to know. Thanks

[nimble heart]

try without the custom node

im 90% sure it's that switch you're talking about. report it as a bug

[crisp owl]

I use that node without issues

[nimble heart]

if it's an auto switch or something comfy caches the model value at the tail end of the node so it'll re-use the lora'd version unless something inputting into the lora node itself changes

it also doesnt clone the cache for children nodes, it just passes a pointer

so the children have to make sure to clone and whatnot if they need to

or else you can pollute the upstream cache iirc

[vital ermine]

I used that --bf switch

not sure what facerestore is doing in comfy as I never downloaded it but I guess it doesn't like it

I disabled it

[nimble heart]

do you just have the comfyui equivalent of Skyrim with 200 sex mods

now suddenly Serana is a floating head and arms and you have to run those BSA compatibility checker tools to find every mod that touches her character cell

[vital ermine]

no, I try to keep it as clean as I can until some goddamn image I get has unknown nodes then I have to go grab them for its workflow

[nimble heart]

when I open someone's canvas and it's 90% red I just give up

not worth

i dont even have the manager installed I download my nodes by hand like the good ole days of 9/11

[vital ermine]

I just noticed something

when I ran the workflow it started the lora then the left side then the right side. It should have done the left side, then load the lora and start the right side

[nimble heart]

post the full flowchart

and I can look at it like a guy that opens the hood of his car and stares at it before calling a mechanic

[crisp owl]

It'll process nodes in order

[vital ermine]

left side should not have changed

if I tell it off or none for lora it goes back to the right pic which is the first of those two.

it does not do it with all loras only some loras

making me have to go back to automatic1111

here is a prime example

no idea why that is happening but not on all loras

wth?

ctrl-b made even a different image

anyway, now you see what I am talking about

I wonder if because I only load the one model node and split off model to lora and to ksampler?

nope

not a big deal I always have automatic1111 to test my trainings with. has scripting too

[nimble heart]

all I can say is i've never had the prompt/lora bleeding ever in auto or comfy so I cant relate

the main problems are always just memory use which comfy without smart mem does best at from my findings

are you on Xformers or SDP? Could try SDP see if its still bad. I read someone saying xformers can make the problem worse but idk he might just be on dope.

[vital ermine]

I gave up on comfyui for this and realize it can't do what I need it to do. Might just be on my side but, in the end, that is all I care about is on my side so if I can't get it to work with what I am doing I just go back to automatic1111. As I said it isn't a big deal.

[woeful patio]

Is there a limit I should look out for when training a lora for sdxl

like a max amount of images

[vital ermine]

no hard limits but it depends on what you are training

I just trained on 75 images for a lora. Most times I like 20.

[woeful patio]

so 200 images would be to much or redundant?

[noble shoal]

I just trained one with 1.750. But i have no idea 😅

Next round i try 4.000 and see if there is a difference

[uncut fiber]

@woeful patio what is your gpu btw?

[woeful patio]

I use google collab to train my Lora’s, locally I use a 8gb 3070

[uncut fiber]

o.k. i have same gpu and no chance to training loras. So if google collab is working for you then o.k. 👍

[woeful patio]

Nah in my experience you need a minimum of 15gb of vram, I bought my gpu right before sdxl released too

[noble shoal]

12gb is fine for training. I am using a 4070

[uncut fiber]

it isnt so rare

[uncut fiber]

nice color like real life one

[noble shoal]

Thank you, i am currently working on a lora for that.

[half ivy]

wth

[heady vale]

havent seen a polar bear-tiger? 😆

[half ivy]

oh how it makes sense now

[uncut fiber]

random prompt generator.

[delicate kelp]

I accidentally created an American Eagle magazine ad generator.... lol

[woeful patio]

what is the best auto captioning for loras right now? is it still blip?

[delicate kelp]

GT-V

GPT-V*

It's how I'm doing all my captioning now. Gonna be even better when the api is out.

[woeful patio]

through bing or are you paying the 20 bucks?

[delicate kelp]

I added a tutorial for it over on SCG-Playground

I pay, but I use ChatGPT for a lot more than SD, though it saves me soooo much time I would consider paying for it all the same. It literally saves me hours per training

[woeful patio]

yeah I should really pay for it too

[woeful patio]

do you have a link for this, it sounds like it might be usefull

[delicate kelp]

Have you been on my discord yet? If not, I'll shoot you an invite for my server, instructions are a community post

[woeful patio]

nah shoot man

[half ivy]

pls sers
any appreciation goes a long way, Auto, XL, just flips last %

^^will delete

[crisp owl]

properly selected vae?

[half cedar]

150 steps too high, Euler a Sampler, pick another, wrong vae, 1.5 weight on ctrl net too high, cfg scale 2.5 too low

[half ivy]

my usual settings do the same, resorting to old faithfuls

K so must be the VAE thing cause i ive never touched it
THX

[crisp owl]

I don't use Auto anymore and I know these options slightly changed so I'm not 100%, but check here

I think that checkbox is new, if it's selected, looks like that's when it uses the manually selected vae

[half ivy]

thank you ser, will look into it

[delicate kelp]

That's a VAE issue. Likely using xl VAE on 1.5 or vice versa.

Hail to the king baby!

[crisp owl]

Willie still looking good as always

[uncut fiber]

worst on spiderman are fingers

[ionic dragon]

@uncut fiber i trained an llm lora to generate sd prompts, wanna try it?

[uncut fiber]

Thank you but i dont know how to install llm at all 🙂

[lusty wolf]

Hello all...

[indigo arch]

guys i never used sd xl; which ui should i use? the node one is good?

[crisp owl]

the node one (comfyui) is more friendly with resources

[lusty wolf]

Crazy quality... must be a massive workflow and GPU?

[ionic dragon]

just install oobabooga

@upbeat summit @crisp owl do you guys also want it?

[upbeat summit]

Thank you 🙂

I'm using ProtoVision by @delicate kelp as a model here which is a big part of the fidelity and coherence.

base resolution: 1344x768 - so SDXL in-spec

base sampler -> latent upscale (hires fix) -> pixel upscale -> post-processing (film grain)

rest is prompting. so nothing really crazy.
realistic cinematic film still from behind, perched high on a city skyscraper, Miles Morales wearing a hoodie and his mask and is illuminated by the twinkling city lights below. his figure is highlighted by the cool, diffuse light of the moon, close up action hero pose shallow depth of field

here are all 4 stages of the image gen (base, latent upscale, pixel upscale, post)

[upbeat summit]

yeah - sounds existing! would love to try it out. but have to do some work first - but I can check it out later!

[crisp owl]

I can take a peep yeah, would likely be tonight I fiddle with

kinda a pia to get, but finally got a good decent depiction of Odin's sacrifice of his eye

[uncut fiber]

@ionic dragon oobabooga is extension or where i found it? Going to check github 🙂

o.k. i found it, will install tomorrow. I hope i will manage it. Thank you!

[ionic dragon]

yea, please dm so i can send the link

[crisp owl]

That looks fun

[sullen dagger]

What Loras and models can I use to make my pictures as good as those on Pixiv? Or at least close to it?

[crisp owl]

you can use base model with no lora and get good images

[cyan crown]

but like this one?

or this?

[uncut fiber]

@ionic dragon i have only 8GB VRAM will it works?

[ionic dragon]

no

[uncut fiber]

i knew it, but will install oobabooga anyway I mean i expected it

[sullen dagger]

For example but also the lewder ones

[cyan crown]

Well.......don't know

those ones are with standard model

[lusty wolf]

I think I'm getting closer. Not bad for a rookie... trying to follow what you said...

[ionic gulch]

reactor (result, source, face)

[clever verge]

@upbeat summit Have you shared your workflow somewhere?

[peak dove]

Distillery SDXL @upper fractal, free in Alpha test

[nimble heart]

@vital ermine
AMD 7900 cards are officially supported for pytorch and other now.
https://community.amd.com/t5/ai/amd-extends-support-for-pytorch-machine-learning-development-on/ba-p/637756

their "official pytorch" support is rocm 5.7+ so still nightly

[shy kelp]

opinions about the sapphire 7900 xt that's on sale for $400 on amazon?

keepa history for context

[crisp owl]

userbenchmark says it's "hugely better" than my 2060S.
That's all my input lol

[shy kelp]

I have not found that site to be useful so I stick with passmark

[crisp owl]

tbh I haven't really checked much on those sites in general, but that's good to know 👍🏻

I'm still waiting on any official info regarding the 5k series cards before i determine what I'm gonna do to upgrade from mine

[shy kelp]

passmark's score is pretty linear and they have some nice shopping tools. it's always my first stop on a build

can SDLX combine images? For exa: I generate a clothe with a certain drawing and I want a character with that clothe

[cyan crown]

https://www.youtube.com/watch?v=Gz_nG2t0zA8

my last video. Next should be Van Gogh

[nimble heart]

if it's a sapphire OEM those are usually good. if it's just the reference model with a saphire sticker I hear they arent the best

but I mean $400

I paid $1000 for my XTX

check the seller cause 400 seems way too low. the 7800 XT is more than that

[nocturne dove]

I guess you have choices depending on what you need:

(1) Use img2img where you mask out the person but not the clothes and generate a new person but leave the clothes alone.

(2) Train a lora for your clothes and then you can easily make different people where it by using the keyword.

(3) Use ipadapter with the source image being your clothes and sdxl will often not modify them too much.

Here's a super quick example of (3). I just downloaded an image from the internet (left image) and then used the apadpater+ node in comfyui with a different character name. It messed up the face in the sdxl image but you can see how similar it kept the dress. There is also a weight for idapter then controls how strong the source image will be.

[nimble heart]

seller with a single review and highly abnormal prices

also the AI generated chinese name

used to get around blacklist

[shy kelp]

dall-e is making some great stuff, but it always go wrong with poses, colors, details and/or background

[hasty smelt]

Hi, which prompt did you use for this? I've been trying something similar but without success. Could you help me? Thank you.

[south horizon]

I cheated, it's dalle3. 😐

[upbeat summit]

this looks great! nice work

[vital ermine]

@nimble heart yeah, I told him and he said fantastic as it may now solve the issues he has getting it to work right in Linux. Thanks for the info for him.

[nimble heart]

I gained about 0.25 it/s on sd 1.5 with the latest nightly over stable torch

smol steps

[vital ermine]

comfyui no longer works right in Linux on his amd as it will say it received the prompt and wait for 20minutes but no issues in windows.

[upbeat summit]

I share my workflow with almost all my images that I made for other fine-tuners on Civitai: https://civitai.com/user/masslevel. It's a very simplified one, but it does the job. It doesn't always have to be complicated or complex to make interesting images.

I got a couple of other workflows like the one I used for the Spider-Man images, but they use a lot of nodes that I customized to my needs. But if you have any questions, let me know anytime!

[nimble heart]

apparently someone hacked flash attention to barely work for inference and got 30 it/s

[nimble heart]

comfy works great for me.

[vital ermine]

Don't know, don't care just saying he isn't having it work for him. Not sure why but oh, well.

[vital ermine]

I wish I got that but 7it/s is my max for XL

[nimble heart]

make the venv manually instead of using the script

I always do for cloned torch projects and I never have any issues

[vital ermine]

His problem, not mine as I have my own set of problems with this to try and figure out.

[nimble heart]

other than that idk, pytorch statically links the rocm libs and compiler so it should be distro package agnostic. skill issue I guess

[vital ermine]

I suspect my 6ish it/s is due to being gen 3 x16 so half the bandwidth but gen 4 is not going to make me get x2 it/s

[nimble heart]

im still at like 3.14 or something for XL lol

[vital ermine]

was that 30it/s for 1.5 or something?

[nimble heart]

yea

[vital ermine]

ahhh. well I get fast on that but nothing like that

[nimble heart]

if you take 5 hours to hack flash attention to work apparently the XTX hits 30 it/s on 1.5 512

[vital ermine]

if I were on gen 4 I suspect 25-30

[nimble heart]

havent personally verified cause im not C++ fluent

yea I peak at 3.17 it/s after warmup

ffor XL

[vital ermine]

I peak from 6-7

[nimble heart]

using sub quad attention

[vital ermine]

I am xformers

or sdp

[nimble heart]

SDP I only get 3.13 and the VAE bombs vram even harder

[vital ermine]

I love it when people are saying their 4090 gets 20+ it/s on XL

[nimble heart]

comfyui literally spits warnings when using SDP whithout flash attention

pretty cool used to not do that

[vital ermine]

YET, sdp was supposed to be brand agnostic/neutral

[nimble heart]

it is

in that it works

but yea it uses I think the doggettx or something attention as a fallback if flash isnt available

[vital ermine]

yeah, the old orgiinal one

[nimble heart]

amd alerady has flash for all the MI gpus so i guess we just wait for rdna 3 support....

[vital ermine]

slow now but was a godsend when it hit

[nimble heart]

i remember when I first got xformers working in auto for 1.4 and it was like brain explosion

like 50% speed and half memory

[vital ermine]

well, I am waiting for Xformers new version to arrive it is faster and less mem again for amp/ada cards.

Uses TC I guess

[nimble heart]

yea uses flash attention 2.0

meanwhile I just want flash 1...

afaik there's nothing stopping the rdna 3 cards from running it besides software

i forgot if it was gemm/wmma or what but whatever hwaccel flash uses has rdna 3 kernels in rocm

wonder if fused attention would work since its just implemented in triton afaik

[cunning locust]

does anyone have sucess training in XL?

[vital ermine]

yes, but not exactly easy

the tools are being update because research is showing to turn off one of the TEs not both but for now all we can do is turn both off.

[cunning locust]

I want to train a model so it paints like my artstyle and I can work faster in my projects, it is possible?

[cunning locust]

how do you turn them off?

[vital ermine]

for lora?

--network_train_unet_only for lora

[cunning locust]

For both Textual inversion and lora

[cunning locust]

not sure how to input that code

it´s on a notebook?

[vital ermine]

I use it locally and haven't done a TI since 2.0 so can't help there. I use Kohya_ss.

[cunning locust]

no prob

[glass notch]

is it possible to have a separate bat file that automatically loads sdXL_v10 as the model, sdxl refiner as the refiner tab, and the sdxl VAE as VAE and sets image parameters at 1344*768?

[weary yacht]

I'm about to try some outdoor and architectural stuff with Juggernaut XL, which I guess it's supposed to be good for

[half cedar]

Found a creature, gill man creature, horror

[weary yacht]

I'm trying to create something of a demon throne room

[narrow cradle]

Hi guys, I'm new here, and I use DiffusionBee & InvokeAI that can't/doesn't allow SDXL. Where are you guys using these SDXL models? what interface, or program please?

[vital ermine]

is there a node or something so we can colour code the noodles to better trace them out?

[crisp owl]

Love it

[weary yacht]

I have another one I did with SDXL

by comparison, this was done with Bing

[pure crystal]

https://civitai.com/models/136070?modelVersionId=155532 where do these go in comfy? (what dir)

[uncut fiber]

it seems it is running on my cpu, will it make some difference?

[ionic dragon]

Cpu is very slow, like really slow

[uncut fiber]

so probably on gpu, but according to msi afterburner, it is in normal memory running about speed 6tokens/sec

i cant read english faster 😄

o.k. it uses gpu but very low, and VRAM usage is say 2,5GB

[half ivy]

Will delete, does this look right XD Or does it have to be SDXL VAE?

[crisp owl]

sdxl vae

[half ivy]

argh XD

[upbeat hound]

i get error on comfy ui when i use this node, it said theres something wrong in the face detailer part
I just wanna know if its only me or its actually the workflow so if anybody have time, plis try it for me T-T

[half ivy]

So heres the thing, when i set the VAE to SDXL VAE, then press "apply settings", it says "0 settings changed", and if i reload UI, the vae is set to the old one i had o0
any ideas would be appreciated much

this is for Auto

or must one change it to XL VAE ere time he/she/they/them loads the UI

[dusk cape]

I love the look of XL images, but how do you quickly iterate on prompts when they take like 4x as long per image?

Does anyone have fast settings to suggest that don't degrade the image too much? e.g. 512x512 is faster than 1280x768... but the images look nothing alike and are often half baked

[crisp owl]

Unsure exactly on the workings of A1111 anymore, but I know I made my VAE options quick options so they appeared at my top bar. (user interface menu I believe is there you set those)

[nimble heart]

watch the live preview and cancel if its gonna be bad

[indigo carbon]

I always build torch and xformers from source because I'm actually crazy, but if little speed ups like this is what you want I'd say that's the right thing to do

[nimble heart]

torch from source is misery

i have 24 threads and it still takes hours to compile

[indigo carbon]

yeah, that's why I mentioned I'm insane

[half ivy]

argh just wont switch to SDXL VAE XD
keep getting "AttributeError: 'NoneType' object has no attribute 'pop'"
I deleted repositories and VENV to no avail

[nimble heart]

gonna spend more time recompiling than you'll gain in speed

only reason I'd ever do it is if I need different build flags from nightly like for a new ROCm version

[indigo carbon]

also idk how the pre-built stable torch packages are built with flash attention on windows

[nimble heart]

¯_(ツ)_/¯

I was just looking at that seeing if I can bullshit my way into making it work on AMD lol

[indigo carbon]

flash attention isn't rawly compatible with windows, they must have some kind of workaround when they're building the stable releases

[nimble heart]

given this
https://github.com/ROCmSoftwarePlatform/composable_kernel/pull/943
it might be possible now?

but im not smart enough to implement it

[indigo carbon]

exLLaMaV1 is faster than V2 if you don't have flash attention, which is really dumb

[nimble heart]

2 is faster for me still

maybe linux or amd thing

main benefit of v2 is the new quant format anyways which isnt compatible on v1

[indigo carbon]

maybe, my environment is fully optimized and exLLaMaV1 is about 60% faster than V2

[tropic turret]

I did something I guess

[nimble heart]

v2 has native rocm v1 doesnt

[indigo carbon]

I'm talking 40T/s vs 14T/s

[nimble heart]

i get 40 t/s on v2 lol

though every time I turn it on it yells at me for not having flash

[indigo carbon]

idk, V2 isn't as refined or flexible as V1 for NVIDIA users due to flash attention being the main speed up for NVIDIA

[nimble heart]

does v1 use xformers for you or something?

[indigo carbon]

yeah, V1 uses Xformers instead of flash attention, and it's way faster than V2 for me

[nimble heart]

cause xformers actually works on windows lol

so if i get 40 T/S without flash wtf will i get with

[indigo carbon]

I mean, Xformers IS flash attention, just a refined version of it..

[dusk cape]

No, the point is to generate a batch of like ~50 pictures trying different things, go do some chores and come back to figure out what worked before doing a higher resolution version of that image

[nimble heart]

at that point XL should be fast enough

V2 for me

[dusk cape]

"the problem you haven't isn't a problem I have so just stop having it" - thanks

[indigo carbon]

yeah, I get half of that with V2

[nimble heart]

if you're going to do chores how is it too slow?

[indigo carbon]

that's identical to what I get on V1 though

[dusk cape]

Because obviously I don't use your computer

[nimble heart]

you haven't given any context besides "slower"

are we talking like 10 minutes per image or something?

[dusk cape]

Generating ~78 images took about ~40 minutes for me

[nimble heart]

not too bad for 1024 tbh what sample count is that

[dusk cape]

This is my setting

[nimble heart]

if you're just iterating prompt I'd drop it to like 30 or less

[crisp owl]

You could do some tricky stuff, but it would would require doing something like using the advanced ksampler, setting it to your preferred amount of steps you usually do, but then setting it to end at say 10 steps, instead of the full say 40 steps it's ultimately aiming to achieve.
Then just decode that latent and save.

You won't have a full complete image, but you can set 50 of those to complete much faster than waiting for the full amount of steps

[nimble heart]

is the refiner on?

[dusk cape]

It is off

[nimble heart]

I'd try DPM++ 2M Karras @ 25 samples. should be faster and probably converge a bit better

DPM2a is one of those double samplers so it renders at half speed, which means you're at equivalent 32 samples

[indigo carbon]

yeah, that's also what I personally use on ComfyUI

[nimble heart]

if you have VRAM to spare bumping up the batch size to 4 or so will give you like 10% more speed too

[dusk cape]

Ah, maybe I selected the wrong one earlier or forgot what setting I was using

[nimble heart]

70 images is also super overkill for just seeing how a prompt does IMO. I usually do like 10 ish for each variant.

10 at the most

[dusk cape]

I like to make grids

so I can see how different words effect each other in combination

[nimble heart]

you're on what was it a 3060 ti?

[indigo carbon]

4070ti

[dusk cape]

example of my workflow (SD 1.0)

[nimble heart]

damn that seems low. thought those had more tensor cores

[indigo carbon]

40T/s is low on 13b with exLLaMa?

well, idk then

[nimble heart]

i was referring to the 14 on V2

3090 gets about 37 or 38 I think

on windows

[indigo carbon]

that's just flash attention not being compatible with EXLv2

on windows*

[nimble heart]

i thought the 4070 ti was supposed to be comparable to a 3090 but with half vram

is 14 with context 100% loaded?

[indigo carbon]

it is, when it's entirely allocated

[nimble heart]

so like 3800 context and 200 new tokens or something

[indigo carbon]

with DeepSpeed I get closer to exLLaMa 1

but not as good as if I would use flash attention

[nimble heart]

flash helps most with full context

[nimble heart]

these numbers are empty context

just inference

[indigo carbon]

no, with just 1 message

[nimble heart]

ah

hm

yea idk then.

[indigo carbon]

I hate NVIDIA, man

they always find new ways to screw with the users

[nimble heart]

for the price of a 4070 ti you can buy a 7900 XT

[indigo carbon]

if the 5000 series will be as cucked as the 4000 series I WILL find them

[nimble heart]

meanwhile the XTX is still cheaper than the 4080

[uncut fiber]

4070ti got small bus, or is it different card?

[nimble heart]

does it make that big of a difference with LLMs?

I thought the cache helped out enough

yea 4070 ti is the 192 bus boy

[indigo carbon]

AMD is miles better than NVIDIA with hardware, I'd love to see the day NVIDIA gets completely destroyed by AMD, we've had enough

[supple knot]

are the vram components expensive ?

[uncut fiber]

i think ppl using llm using as well image generation

[indigo carbon]

no, NVIDIA saves nothing by cucking them, it's just a method of forcing you to buy a 4090 or a 4080TI

the rest of the 4000 series has 128bit VRAM

[nimble heart]

im show of my XTX in simracing stuff sometimes. Playing VR with 100% maxed out graphics only using 40% of my GPU lol.
all the kids in the discord server think I have an overclocked 4090 or something

for games the new amd cards definitely fuck

but the ML software is still in omega catchup mode

[indigo carbon]

I'd argue if AMD would have software at the same level as CUDA the fastest AI inference is likely AMD

I hope that'll be the case one day

[uncut fiber]

4080 has 256

[nimble heart]

yea its a super weird config. AMD cut costs on VRAM by using gddr6 non-X and still giving it a fat bus, cache, and clocks.

[uncut fiber]

@nimble heart simracers? Great! What sim you prefer?

[nimble heart]

ah Im not into iracing or anything. I have assetto corsa but I mostly play beamng.

im omega casual. have a $250 wheel and no shifter.

with a homemade desk mount

[uncut fiber]

assetto corsa isnt bad, only that manager is must.

[nimble heart]

I fucking hate the interface on that game. paying for the manager just to make the game usable is insane.

I dont play multiplayer really anyways so beamng has lots more to do

only thing that hurts with beam is the AI racers are garbage.

[indigo carbon]

@nimble heart any idea?

[uncut fiber]

no it is possible with free version @nimble heart

[indigo carbon]

I was testing stuff again and now it didn't even load

[nimble heart]

yea it goes from miserable to just unfun lol

[uncut fiber]

no, graphics is phenomenal 🙂

[nimble heart]

if you have exllama2 installed manually to repositories/, move it and update text gen

exl2 is bundled with the deps now

[indigo carbon]

I'll just reinstall the UI

[uncut fiber]

i am playing most AMS2 because easiest make multiplayer with friends, no rocket science 🙂

what model would you suggest average user with 3070 8GB?
Using llama-2-7b-chat.Q4_K_M.gguf

[nimble heart]

and the graphics mods are paid for the good versions too lol

[uncut fiber]

it is not. It is all posible with free version

[nimble heart]

idk beam looks fine, has good VR/wheel support, and has lots to do so im fine with it. Not interested in buying a $1k rig and getting competitive

[uncut fiber]

yes i know it

[indigo carbon]

@nimble heart which torch version does oobabooga work best with?

[nimble heart]

new versions default to 2.1

as of like a week ago

[indigo carbon]

and cu118, nice

that's what I use for ComfyUI

[uncut fiber]

you all are using 13b?

[nimble heart]

yea

my model is based on llama2 so i also have 4k context, but I dont measure with it filled.

[uncut fiber]

i have only 8GB but can probably try it, but i think i am lost

[nimble heart]

I just wanna hook up my wheel toss on my rift and drive around for a while. Setting up SRP with the AI cars and visual mods and content manager and all that shit is way too much effort. spent enough time modding Skyrim I've had my fill.

[woeful patio]

is it recommended to train a realistic lora on a realistic model like realvis or just use the sdxl base model?

[uncut fiber]

yes just that free thing is easy to setup, and it is like night and day. But one has to trust developers. I understand you! @nimble heart

[nimble heart]

• beam aint look half bad anymore if you havent seen it in a few years.

[nimble heart]

13B needs to be quanted down to 4bit to work on 8 gigs

so if you're benchmarking speed that'll perform way different than 8bit

[uncut fiber]

o.k. so i will stay with what i have 🙂

you made it 🙂

[nimble heart]

i made what?

[indigo carbon]

I just figured out how PyTorch are building the packages with flash attention

It's possible to allow Conda to build flash attention from source

[uncut fiber]

i mean you won the lap in video

[nimble heart]

wonder if you could to that for exllama

[nimble heart]

ah yea. I gave them a gnarly head start since the car is OP.

[uncut fiber]

yes

[indigo carbon]

Because exLLaMa V2 isn't conjoined with flash attention, if I build flash attention from source it should work with exLLaMa 2

[nimble heart]

hypothetically the GT4 looking car should be faster but the AI corners like grandma

[nimble heart]

yea what I was thinking

it checks for it at runtime

[indigo carbon]

After I'll finish doing that I can truly compare exLLaMa 2 and 1

I think Oobabooga should be shipped with a pre-built venv so everyone could use everything right off the bat

[nimble heart]

if you add another lap even a 4cyl sedan with no aero handily beats their AI. biggest peeve tbh but not enough to redownload all 80 gigs of asseto corsa lol
https://youtu.be/WE2Y5gXYBhA

[indigo carbon]

ugh.

[nimble heart]

oof. what flags does torch use

you might need the cuda SDK too

[indigo carbon]

maybe flash attention expects CUDA 12.1?

[nimble heart]

11.6 and above

https://github.com/Dao-AILab/flash-attention#installation-and-features

also
https://github.com/Dao-AILab/flash-attention/issues/595

[indigo carbon]

Then wtf is that error?

[nimble heart]

¯_(ツ)_/¯

dont do CUDA

[indigo carbon]

I'll try a different torch version

I remember TRT also having issues with stuff like this

[nimble heart]

12.1 might literally be too new

[indigo carbon]

I have to install the other torch version with Conda, right? It will make sure it's all compatible

[nimble heart]

dont do Conda either

good luck

[indigo carbon]

Why not? That's how Oobabooga says it should be done

[nimble heart]

tf you mean why not? virtualenv is a lot easier to set up manually

[indigo carbon]

nope, same stupidity

[nimble heart]

just virtualenv virtualenv; overlay use virtualenv/bin/activate.nu; pip install -Ur requirements_amd.txt

boom ooba works

repeat for comfyui, auto1111, anything

what'd it say while compiling the flash extension?

ninja spit out anything funky?

[nimble heart]

https://github.com/Dao-AILab/flash-attention/issues/604#issuecomment-1763226621
guy seems like he knows what he's doing. got it working on mixed 3060 + 4070

[indigo carbon]

I'll try to uninstall flash attention, it's really funky rn

I'll test if even normal EXL2 works

[icy brook]

anyone good at controlnet here? I have used it a lot for my datasets but never with body positioning. can someone say how I do that?

[supple knot]

with Contrlnet OpenPose

[nimble heart]

oh General a sorta tracker PR for some major rocm flash attention stuff
https://github.com/ROCmSoftwarePlatform/composable_kernel/pull/810
navi 21 and 31 pass whatever they're testing it seems. They dont even test 5000 series cards lol.

495 commits ahead of develop oofie

[vital ermine]

sheesh

Well, I hope they can pull it off and my friend told me last night he finally got Linux to run right with comfy.

[nimble heart]

nice. 19.5 it/s on 1.5 and 3.17 it/s on XL is what I get for reference. I have the latent2rgb live previews enabled so that might reduce it a few %

[vital ermine]

yeah, it will, it will

[nimble heart]

okay 495 commits sounds scary but 65,000 added lines sounds scarier

[vital ermine]

I wouldn't want to track them

[simple agate]

is anyone using sdxl base on automatic1111 ?

[vital ermine]

using it right now for the aboves

[simple agate]

how is the results ?

[vital ermine]

comfy is better

[nimble heart]

im a bit out of my water here but it almost seems like rocm straight didnt have the architecture in place that newer flash attentions needed so they have to build up the entire stack from scratch with all the shit they were missing until they can finally compile the flash attention tests

[vital ermine]

I just can only take comfy for so long

[simple agate]

mine is pretty bad

can u share a photo of results

[vital ermine]

I did above

[uncut fiber]

what resolution @simple agate

[simple agate]

this one is mines 😄

[nimble heart]

bru

[vital ermine]

oh, damn I know that issue

[uncut fiber]

512x512? SDXL need 1024x1024

[vital ermine]

yep

[nimble heart]

512, also maybe wrong VAE?

compounding together

[vital ermine]

you cannot, not even in comfy, use 512

[nimble heart]

XL Can do 512 it just looks soupy, not outright fried

[simple agate]

oh ok lemme try

ru using the base 0.9 ? 13gb version of the model ?

[vital ermine]

LOL, no

1.0

[uncut fiber]

13b is llama 🙂

[simple agate]

with refiner right ?

[vital ermine]

most times no but above yes

[nimble heart]

XL @ 512 for "photo of a locomotive. bad yes but not even close to that XBOX red ring of death looking picture

[uncut fiber]

size of model is about 7GB refiner 6 so yes

[nimble heart]

XL 512 + the 1.5 vae looks closer

so double check that ig

actually I can probably just look at the meta

damn its a copy paste

[icy brook]

oh ofc 🙂

thanks

[nimble heart]

the untrimmed XL 0.9 download was indeed 13 gigabytes iirc

[uncut fiber]

i think sdxl+refiner is 13GB still i think

[nimble heart]

ah maybe

[simple agate]

now im taking error about vae 32 bit something ..

[vital ermine]

yes, --no-half-vae

[simple agate]

thanks gonna try

[vital ermine]

k

[nimble heart]

does auto still default to 16 vae?

[simple agate]

yup

[nimble heart]

bru

even sd 1.5 has issues on fp16 sometimes

[vital ermine]

@nimble heart He told me it was his linux install as he installed fresh but something didn't work then installed it again and all good now.

[indigo carbon]

@nimble heart yo?

[nimble heart]

that with flash or just exl2

[indigo carbon]

flash

[nimble heart]

damn

speed?

[indigo carbon]

it's so bad

[nimble heart]

bru

[vital ermine]

at 1024x1024 it's 2.34 it/s consistently

[nimble heart]

what quant is that? you out of vram?

[vital ermine]

what are your settings because that is what he is getting?

[nimble heart]

I have one that should work on 12 gig cards

[nimble heart]

comfyui or auto

auto I cant get above like 2.5

[vital ermine]

comfyui

[indigo carbon]

VRAM is over 90% full, it's 4bit. when a 4000 NVIDIA card that isn't 4080 or higher get's above that precentage it ceases

[nimble heart]

        --output-directory /tmp/comfyui
        --port 7776
        --dont-upcast-attention
        --preview-method latent2rgb
        --disable-smart-memory

ah

the not upcasting gives you like +50%

[nimble heart]

that seems crazy. from my measurements a 5bit model should fit...

[indigo carbon]

another instance of NVIDIA being soooooo good

[nimble heart]

I have some quants if you wanna try them
https://huggingface.co/Beinsezii/MythoMax-L2-13b-EXL2

those measurements are without flash attention so you should hypothetically use even less VRAM

[indigo carbon]

Output generated in 54.21 seconds (0.02 tokens/s, 1 tokens, context 65, seed 792262846)

nah.

[nimble heart]

yea something aint right

[indigo carbon]

I would actually die of old age before it generates a sentance

I'm not even kidding

[nimble heart]

is it falling to software

[indigo carbon]

no errors whatsoever, it's just NVIDIA's card is too smart for this

[vital ermine]

@nimble heart Thanks he put those in and is now at 3.19 it/s

[nimble heart]

basically exactly what I get with the default settings too

[vital ermine]

yeah, he said they were rock steady

[nimble heart]

nice. .02 higher than me lol. I have my card in silent mode instead of OC mode though

if he wants to try getting flash attention working he could possibly push that up to like 5+ it/s if the other numbers were to be believed

[indigo carbon]

you saw what that did to my speeds..

[nimble heart]

not sure. those guys on the flash attn github got it working on a 4070 so start copy pasting I guess

[nimble heart]

he's on linux with ROCm totally different

• yours isnt compiled correctly clearly

[indigo carbon]

okay, it's now 35T/s, way better. it just throws random words though

great random word generator imo

[nimble heart]

nvidia doesnt like you lmao

i havent had that happen since i tried running an opt model with BnB

[indigo carbon]

that's mythomax-l2

[nimble heart]

that's super fucked up my quant looks nothing like that

[indigo carbon]

if I can recall; that model isn't a random word generator, is it? is it because I used a GPTQ model?

[nimble heart]

mythomax L2

its a chat model

[nimble heart]

try one of my EXL2 quants I know they dont die lol

hypothetically with FA2 the 5bit would probably work with 4k or close to context

[indigo carbon]

I'll try a version I compiled a while ago

ok, it's back to 0 speed

[nimble heart]

exllama2 can do 4bit gptq or mixed bit exl2

[nimble heart]

out of curiosity try reducing the exllama2 shape from 4096 to 2048

[indigo carbon]

it's already on 2048

should I do 4096?

[nimble heart]

damn wtf

i mean if its a 4bit you should be able to

but it definitely wont run better I wouldnt think lol

try it

leeroy jenkins

also maybe try the non-HF version

[indigo carbon]

I was already on that one

[nimble heart]

the _HF version uses different samplers

[indigo carbon]

should I try the HF one?

[nimble heart]

¯_(ツ)_/¯

yolo

_HF 4096 boogie

[icy brook]

does it work well for sdxl yet?

[indigo carbon]

is this fine?

alright, now down to 9t/s, it does this with those settings

I give up, man

exLLaMa 2 is a failure, V1 is faster and it doesn't turn 13B llms into weird gpt-2 offsprings

[nimble heart]

skill issue

IDK how you managed to break it

I've converted like 3 other windows users to Exllama 2 and you're the first with it being slower

Nvidia heard you talking shit and blacklisted your card

[indigo carbon]

I'll just stick to exLLaMa 1 then, it was fine

[indigo carbon]

good, because I meant 100% of it

they suck

[nimble heart]

if you're using gptq there's really no difference anyways. they basically use the same kernels

[half ivy]

To whomever helped me recently, THANK YOU'S!!!!! We did it, all sdxl babbbyyyy, Base & Ref, CN, VAE SSSSSOOOOOOOOO HAPPY its fixed WWWWWOOOOOOHHHOOOO

[nimble heart]

you have to use the exl2 specific quant to see the huge benefits

[indigo carbon]

I did, it turned into a slower version of GPT-2

[nimble heart]

bru

IDK then

[uncut fiber]

@nimble heart what is best preset? Simple?

[nimble heart]

my 13b llama2 based EXL2 quanta outperform 30b gptq llama 1

[half ivy]

that summin a saffa would say

[nimble heart]

on ooba? simple is the best overall, but others will be better at chat specifically

[uncut fiber]

o.k. thank you

[nimble heart]

there's a big spreadsheet on the repo

with ELO scores and everything

[uncut fiber]

o.k. going to check it

[nimble heart]

I lied it's not on the main repo it's here
https://github.com/oobabooga/oobabooga.github.io/blob/main/arena/results.md

tldr depending on the model Midnight Enigma, Yara, or Shortwave are theoretically the best at chat

for mythomax I think midnight is clearly the best. it hates Yara for sure.

[uncut fiber]

Thank you @nimble heart

[nimble heart]

don't know who that is

I probably picked it up from my best friend's wife

[uncut fiber]

You
which star is closest to the earth
AI
The star that is closest to Earth is Proxima Centauri, a small red dwarf located about 4.24 light-years away (approximately 26.8 trillion miles or 43.4 trillion kilometers). Proxima Centauri is part of the Alpha Centauri star system, which also includes the brighter and more massive stars Alpha Centauri A and B 🌠
LOL!!!

[nimble heart]

did it add the star emoji itself

[uncut fiber]

yes

[nimble heart]

that's funny. what model

[uncut fiber]

llama 2 7b chatt

closest star to earth is sun that is bigger fun 😄

[nimble heart]

what's the quant level

[uncut fiber]

4bit?

[nimble heart]

gptq?

gptq is usually 4

[uncut fiber]

no mmnt i think llma.cpp

[nimble heart]

ah. for llama.cpp you could easily do 8bit and just offload the last like 6 layers onto your cpu

you could probably do 13B 4bit entirely on GPU if you set up flash.

[uncut fiber]

i am happy with what i got. That question is bit tricky

[peak dove]

SDXL from Distillery @upper fractal

My IPAdapter setup keeps crashing when I go to 1024x1024 from 512x512!!!

[half ivy]

https://tenor.com/view/thank-you-thanks-gif-26335477

@crisp owl @delicate kelp

[upbeat summit]

have fun 🙂

[half cedar]

#✨｜sdxl message right here 🤭

[glass notch]

in Auto I would recommend selecting the VAE selector (I think it's called sd_vae or something) as one of your quick settings. Then you can monitor it easily as it should appear at the top of your screen in the generating screens

[half ivy]

And @half cedar 🥰

[half ivy]

Yea got that alpha and did it, SD VAE 👊

[real vine]

what is the latest sdxl model and refiner ?

[crisp owl]

No other official SAI models, just base and refiner.
And as far as I've seen, no community based refiners.
But there's always new community models trained and released on places like Civitai

[cyan crown]

SDXL has almost everything inside. Working with prompt and just a lora like xl_more_Art you can get endless styles

[icy brook]

this makes me think of MJ v3

🙂

[cyan crown]

True!! I like so much that style. Btw this is SDXL

[icy brook]

yeah 🙂

[noble shoal]

Yesterday i have trained a lora but accidentally used the wrong image folder. With that lora i can generate images with a resolution as low as 392x392 without getting complete garbage. 11 it/s for the win. lol

[azure oxide]

whoahhh official nvda extension on a1111 on tensorRT
https://github.com/NVIDIA/Stable-Diffusion-WebUI-TensorRT

very juicy

[deep dagger]

Now we just need a port for comfy

[ruby schooner]

Is SDXL ever going to be supported on AMD cards?

[azure oxide]

i think its pretty neat that a large company actually devoted dev hours in creating the extension lmao

[cyan crown]

did you try it?

[azure oxide]

not yet

[uncut fiber]

isnt it response that amd bought something related?

[cyan crown]

ok I'm going to install it now

[azure oxide]

damn they even got someone to do technical writing on customer support in how to install a1111 lmao https://nvidia.custhelp.com/app/answers/detail/a_id/5487

[cyan crown]

Error caught was: No module named 'triton'

RuntimeError: Couldn't install protobuf.
*** Command: "D:\SDXL\stable-diffusion-webui\venv\Scripts\python.exe" -m pip install protobuf==3.20.2 --prefer-binary
*** Error code: 1

[uncut fiber]

go over it and it will work, just not sure it is faster, having 7it/s at 768x768

[cyan crown]

I get error

entry point for ?destroyTensorDescriptorEx@ops@cudnn

and also

ERROR:root:Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument mat1 in method wrapper_CUDA_addmm)

opened an issue

[uncut fiber]

me as well 🙂

it is weird they are talking about generating button and button is exporting... i hope everything is o.k.

[cyan crown]

uninstalled

[crisp owl]

Odin (a bit pudgy) riding Sleipnir

[iron coyote]

what simpling method, simplig steps, cfg scale is used here?

[crisp owl]

This was using a preconditioning method.
3 steps to start, latent pushed to the main doing 20 steps, latent pushed to refiner for 5 steps, latent pushed to upscale for another 10 steps.
cfg 7.
dpmpp_3m_sde_gpu/exponential

[narrow cradle]

Oh, for the love of Sleipnir, how did you prompt the 8-legged horse? That has been my quest sinve midjourney v4

[crisp owl]

final outcome was actually several back and forth processes between my main workflow and my img2img workflow.
So not necessarily purely prompted, though I did manage to get several 6/7 legged horses just with prompting, I just couldn't get the complete 8 without a bit of additional work

[nimble heart]

that's the one Loki gave birth to right?

[dusk cape]

Question on Refiner.

When I use it, it has to unload the XL model and load the Refiner model. This creates a lot of churn... Is there some way I can keep both loaded? Or maybe batch all of the XL work, then refine later?

[crisp owl]

Yup he is said to have given birth to Sleipnir

[crisp owl]

A1111 or Comfy?

[dusk cape]

I'm using A1111

But I'm open to recommendations

[crisp owl]

There should be an option in A1111 which lets you choose how many models to keep loaded, I believe it uses RAM to boot it quicker

[dusk cape]

Oh I see, I can look for the setting

[nimble heart]

wonder what drives a man to shapeshift into a mare then bear the child of some demon stallion

gotta be the wildest midlife crisis

[crisp owl]

Loki has an interesting life 😆

[dusk cape]

Base model only: 19 seconds
Add Refiner: 34 seconds
Untick the option setting: 31 seconds

Hmm it's still really slow to use the refiner. I guess I'll go without the refiner until I'm ready to make a final image

[timber plover]

Is there a way to create animations with SDXL?

[real vine]

lora:xl_more_art-full_v1:1 what does this lora usage mean ? use this file with 1 weight parameter ?

[tender timber]

There was, not sure if it’s still there. Seems like it’s not worth it. https://www.reddit.com/r/StableDiffusion/comments/15bbq1p/a1111_performance_tip_for_sdxl_disable/

[cyan crown]

I use with 0.8 and it works great. It enhance images someway. It isn't a tyle (because you can specify styles in prompt) it just enhance image composition

[crisp owl]

I haven't used A1111 since SDXL came out, so not sure what's really changed there, but at least with my card it did help when I would switch between a couple checkpoints often to reduce the loading time. Not sure what's changed since though, may work differently as there's been a lot of changes since

Containing the tales of Fenrir

[dim hare]

Is SDXL The best model for fine-tuning? Are there any other SD version that are better for this? Or maybe better model? I am looking to tune it for pixel art style, so I was wondeirng if SDXL was the best choice to do LoRa for

[cyan crown]

yes it is

[dim hare]

Nice

[cyan crown]

btw you can have pixel art also without finetuning

[crisp owl]

This is NOT Tyr losing his right hand to Fenrir, but man they sure do look like good buds

[undone vine]

I’m fairly new to the ComfyUI scene but out of curiosity, is it the SDXL models that causes the Mac silicon chip to run so slowly or is it purely down to Comfy’s poor optimisation for MacOS?

[crisp owl]

just to make sure, did you follow the section in the github where he describes the install process for mac users?

[lusty wolf]

Hello guys... and good night... Hope all is well with everyone!

[indigo carbon]

how do I make it more visible?

[indigo carbon]

inverting the post processed input somehow improved it

[supple knot]

https://huggingface.co/hotshotco/Hotshot-XL there this I haven't tried it because the demo I saw it was max 24gb of gpu vram

here the stream of demo I saw https://www.youtube.com/watch?v=vqxQY6Cu5X8

[upbeat hound]

whats the different in text l, text g in comfy ui?

[pure crystal]

two different CLIP models in SDXL, valid to mirror your prompt to both. not much guidance, except g might be considered overall image and L is about the subject, YMMV

[vagrant valve]

just how did you do this

[rustic shadow]

RealVisXL is good at capturing phsychedelic realms!

actual params are closeup of ancient entity of the universe, octane render, raytracing, psychedelic visuals, (non-euclidean geometry:1.2), iridescence, DMT visuals, LSD, ayahuasca visuals, vibrant metallics, reflections, fluorescent highlights, 4D, psilocybin, DNA helix, magic mushrooms, depth, crisp details, 8k, <lora:sdxl_offset_example_v10:0.8>

[vagrant valve]

Im really getting into creating psychedelic visuals with stable diffusion. Would you send me the link for RealVisXL?

[rustic shadow]

https://civitai.com/models/139562/realvisxl-v20

[vagrant valve]

I appreciate it! I’ll let you know if I create anything worthwhile

[rustic shadow]

doooo iiiitttt

[vagrant valve]

Fantastic! I’ll mess around with it tonight

[rustic shadow]

I have seen the universe. I AM THE UNIVERSE.

[rustic shadow]

big boi boa

[crisp owl]

Trying to get a good Jormungandr, not a whole lot of norse stuff in models

[rustic shadow]

ooooh that would be badass. Had to look him up haha

definitely needs more beard

[crisp owl]

Yeah this wasn't really close at all to what I want, but thought it was a nice pic regardless haha

[rustic shadow]

ya nice one, this stuff is too much fun

sticky

[upbeat hound]

anything for fixing hands in sdxl?

[woeful patio]

anyone know why my eyes keep coming out looking like this?

[vital ermine]

Why so small of an image?

XL only likes large images

[woeful patio]

its not the size that matters, its how you use it

[vital ermine]

wrong with XL

as to your eye problem welcome to XL base where it loves snake eyes, wonky eyes, and blind eyes.

[fierce hollow]

Has anybody managed to get the new comfy ait repo working? For a supposedly better foundation the patch fails and there are missing deps

[frank falcon]

Hi I'm still pretty new to SD with A1111, can anyone give some tips for making photorealistic images less noisy/blurry? I'm trying to create 3440x1440 wallpapers and am currently using just the txt2img tab with a base res of 1376x576 with R-ESRGAN4x 2.5x upscale. How do I know if the issue is noise in the base image from prompt, or suboptimal upscaler/config?
sorry if asking in wrong place.
"a beautiful picture of the ocean, the sunset of the sea, the sparkling light blue sea water, the stars shining, the beach, many golden twinkling crystal covering the beach, the sunset, the beautiful coloruful clouds
Steps: 40, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 335770400, Size: 1376x576, Model hash: 31e35c80fc, Model: sd_xl_base_1.0, Denoising strength: 0.7, Hires upscale: 2.5, Hires upscaler: R-ESRGAN 4x+, Version: v1.6.0"

[upbeat hound]

i think the blurriness is from upscaling far to big of a scale, and i think you should instead use img to img just denoise it about 0.4 then use the model you used on the base image, just go up about 1.5-2 times, rinse and repeat for a couple then it should be better

[frank falcon]

isn't 4x the default upscale amount, so not 100% sure what you mean there. will try that also tho thanks

[nimble heart]

for model-based upscalers if you use any value below 4 it just scales to 4 then downscales back to what you originally picked

so if you do 2x it scales to 4x then shrinks by half

[vital ermine]

which is so redundant, but I am unsure why there are not more 2x models (there are some).

[nimble heart]

probably cause esrgan was 4x and 99% of them are just tunes of that

easier to just downscale a 4x'd model then retrain the whole thing around 2x

esrgan's how old now we need a new arch

we're still in the OPT/GPT2 days of upscalers we need the Llama equivalent.

[vital ermine]

Yep, they are all ancient now.

[peak dove]

Free SDXL in Alpha test, Distillery@Discord

Anybody know IPAdapter at all?

[noble shoal]

DAT (dual aggregation Transformer) Upscalers are pretty good in my opinion.

[uncut fiber]

https://openmodeldb.info/
there are some
You can as well render at much higher resolution like 1720x720 like it was said, and then just resize 2x with some native 2x resizers.

[tropic turret]

🙂

[peak dove]

Why does my IPAdapter crash at 1024x1024? It works fine at 512x512!

[glass notch]

What 2x upscalers does the community recommend? Cuz I've been using x4 for x2 images without understanding the implications lel 😛

[tropic turret]

Invasion of the... lips with eyes?

[uncut fiber]

SwinIR? @glass notch Or you can test what you like from
https://openmodeldb.info/

[pallid path]

yeah higher resolutions or highres fix fixes that

you could also just use face enhancement (like gfpgan or codeformer)

For up close shots where the image is stable and sharp, don't use it, it might even make it blurrier (at least the last time I used it for close ups)

[hoary saddle]

Anyone know of there's a different way to pass timestamps with prompts instead of using a batch scheduler for animation frames? The webclient method for the API refuses to pass it due to the formatting of the scheduler with the extra " and : separating frame numbers and prompt text

[icy brook]

combining ink, sketching and watercolor in one LoRA

[lusty wolf]

Knock, knock... who's there?!!

[vast swan]

A modern good looking portfolio website design ui ux

[crisp owl]

Nothing special, just thought it was kinda pretty

[hasty smelt]

Hi everyone! I'm looking for information on the system requirements for using AUTOMATIC1111 with SDXL. I'm currently using CmfyUI, but I'm finding the node-based system a bit confusing. That's why I'm considering trying AUTOMATIC1111.
My system is as follows: RTX2070 Super FTW3 8GB + 64GB of RAM + M.2 7300MB + 5700X.

Does this system support AUTOMATIC1111 + SDXL?

[boreal bough]

when your lora works, but the sticker prompt fails you XD

[viscid warren]

I think thats good enough

I read RTX2060 and above is already good

what do you need 64gb ram for btw its overkill xD

Are you doing coding calculations ?

[crisp owl]

64 is amazing to have

[noble shoal]

I have created a Lora that can only produce elevation heightmaps....of terrains. Into the archive.

[viscid warren]

Sure but why much better than 32gb ddr5 for normal useage

[glass notch]

I can use SDXL (painfully but it works) with 6GB VRAM 16 GB RAM so you' ll be fine. I am currently switching to comfy as per another user's advice as there is a performance cost with automatic1111 apparently.

[vital ermine]

I did the same when I had my 1060.

[hasty smelt]

SDXL with CompyUI can consume more than 32GB of RAM during Vae pass.
I had only 16GB of RAM, I was getting blue screens during Vae pass, so I upgraded to 64, 2x32.

[hasty smelt]

Of course, I think you were the one who helped me make up my mind. I was in doubt between upgrading from 16GB to 32 or 64. Upgrading to 64GB was the best thing I did. Now, even during the Vae pass, I can work with other apps at the same time without suffering stuttering due to lack of RAM.

[viscid warren]

Sounds good

What does Vae pass mean btw?

And does dreamshaper with comfy UI also need 64gb?