✨｜sdxl

[lyric wigeon]

Hello! Anyone know of any resources for running SDXL autotrain dreambooth or similar on 12GB VRAM?

[midnight shuttle]

Fresh install will require some time because pytorch CUDA is 2.5 GB download.

[midnight shuttle]

Good luck. I do not think this is possible.

[hearty ginkgo]

what workflows are you guys using/found the best for sdxl 1?

[glad fulcrum]

waiting for a Lora google Colab for SDXL

[orchid oar]

Yea, i mean could at least try. And there was not link or something in the command

[midnight shuttle]

Too many parameters for backpropagation. Cannot fit in 12 GB.

[trim orbit]

i'm home now. what's happened today? anything worth recapping?

[orchid oar]

dont think so

[crystal bone]

still giving me a bunch of errors, different ones this time

i'll just nuke it all 🤓

[trim orbit]

aw that mean no controlnet release. fuck i'll check again tommorrow

[lyric wigeon]

Is that true for 512x512 training as well?

[orchid oar]

Yea sadly

[orchid oar]

Yes at this point just fresh install

Sorry about that

[crystal bone]

no problem, is it best to just delete the folder

[midnight shuttle]

I haven't tested but I think it will not work. Even inference with SDXL on A1111 requires almost 16 GB. Training will require a lot more due to backpropagation.

[crystal bone]

or should i do some other things

i have to say i dont remember much of all the steps I took 😁

[native moon]

hahahh look at the tail

[orchid oar]

Goodnight Guys

[winged bear]

redownloading the model fixed it. got a picture of a cat

[craggy basalt]

works on MacBook Pro 2019

[agile quarry]

https://github.com/Linaqruf/kohya-trainer look in the files it's not listed in the readme yet

[nimble heart]

if that's using my principled node make sure it's updated. I just pushed some changes to help i2i an hour or so ago

or maybe wait cause I'm re-writing the high res fix portion to better support pixel scaling

and be less goofy

[hoary saddle]

oh, i didnt submit to github, just added it to local copy

works great btw, loving it!

[nimble heart]

ik but when you update a node it can break the workflows attached to it depending what changed

so just warning

[hoary saddle]

ah, i didn't know that

[nimble heart]

mostly if you change the number or kind of parameters

which I did

and am going to do again

cause I'm evil

[hoary saddle]

gotcha

[nimble heart]

what's your experiences with the node so far

[hoary saddle]

excellent, getting great results from the input image so far

custom principled node is top notch

[nimble heart]

if your version is new enough you could try setting refiner_amount to 1.0 and denoise to like 0.1 so it just adds some refiner details without doing anything else

it can resize for you but looks like you already did that

[hoary saddle]

<--- still a node noob, but yeah i'll play around with that, wasn't sure of other ways to make sure all input images are same size before getting processed, those 4k images hurt

so used to the restful API of a1111 and making all my web calls directly to it via python, this is all new but loving the save to API button, copy/paste to my scripts and send output directly to the enduser

[worthy bear]

has anyone had much success with inpaint in sdxl?

[high skiff]

How are you fine people doing on this insert locally applicable time of day here?

[vast narwhal]

Original and 4096x4096 Upscaled in a full SDXL 1.0 process

[glad fulcrum]

thanks. Now I need a tutorial XD

[high skiff]

I just got home from shopping with my mom, I'm eating dinner, and I'm trying to keep my mind off of the person who sold me this shitty 3090 trying to currently screw me out of the refund process, so, did I miss anything?

[trim orbit]

while scrolling through different communities today i came across more mentions of secret watermarks being in sdxl generations in order to track them and people generating. it's all so dumb and i find it to be such a hilarious conspiracy theory. like, we're all out here begging people to post with metadata

[nimble heart]

im rewriting like half my node cause of you gj

[high skiff]

OwO?

In what way?

Don't get too comfortable, I have more information to release on my 1.0 documentation lol

[nimble heart]

redoing the upscale to fully support pixel upscaling instead of it jut being a rekerjiggered latent upscaling

[high skiff]

Oh my God, I've never seen that word before, and it looks extremely problematic lnfao

[nimble heart]

right now if you upscale the first pass is always just base no refiner

but I found it looks way better if you do the first pass using partial diffusion before pixel upscaling

latent it doesn't matter since its so noisy anyways

[nimble heart]

it's a fun word, you should use it.

right up there with discombobulated

[trim orbit]

i've spelled it recajiggered before, but i say it all the time. i'm always rekerjiggering stuff

don't see it often in text now that i think of it

[high skiff]

lets just say, it reads as very... concerning lol

[nimble heart]

concerning how?

what's on your mind?

[high skiff]

I am a very careless reader ok

it has all but one letter lmao

[nimble heart]

I have actually no idea what you're talking about

pls enlighten

[high skiff]

[nimble heart]

I'm a sheltered child

[trim orbit]

https://tenor.com/view/park-stan-jeopardy-southpark-gif-22579301

[high skiff]

SAME LMAO

[dense chasm]

does diffusers not has lora training for sdxl 1.0 yet?

[midnight shuttle]

There is a very serious racial slur in English that contains many of the same letters. It is such a serious word that anyone who says it will be fired from any job instantly. It is one ofthe worst words in the language.

[trim orbit]

straight to jail

[high skiff]

||erjig||

[midnight shuttle]

YOU'RE FIRED!

[trim orbit]

d8ahazard is building a ui thing that uses diffusers and has training far as i know

[high skiff]

[nimble heart]

if you're talking about the pewdiepie word "rekerjiggered" is so many letters away from it idk how it's problematic

[hoary saddle]

speaking of racist, AI has a hard time picking up queues of skin tone and translating it

[native moon]

thats not racist

[hoary saddle]

joke

[native moon]

[trim orbit]

i've watched maybe 10min total of pewdiepie. funny phenomenon that guy. but yeah, rejiggin is way older. think it goes back to ww2

[hoary saddle]

i use fstrings to inject race so the user can choose

[thin edge]

need a lil help with the speed issue in sd 1.5

can someone help?

ik this channel is for sdxl

i put it already in #💬｜general-chat

[nimble heart]

ik it's an old ass word I've just never seen it make the news faster than when a youtuber said it

[thin edge]

apologies. #1072238304042438758

[trim orbit]

thingamajig is another gooder

[pure hazel]

How is the refiner used in Auto's, as an IMG2IMG pass?

[trim orbit]

yeah a whole manual second pass

[dense chasm]

what's the link?diffusers is come in handy,comfyUI sometimes process the workflow slowly

[trim orbit]

sure. last i looked it was in rough shape. like unfinished rough. i'll dig it up.

[pure hazel]

Hmm what does this mean? AttributeError: module 'lora' has no attribute 'lora_Linear_forward'

if my pictures are coming out black in 1111, what could that be with SDXL?

[dense chasm]

btw,i learned a lot from diffusers ,high noise denoise step run first ,low noise denoice next

[trim orbit]

don't know man. i went and looked for it on d8hazards github. there's no evidence it ever existed. stable-diffusion-plus. i still have it installed

[high skiff]

denoice haha

https://tenor.com/view/noice-nice-click-gif-8843762

[pure hazel]

Does opt-spd-attention cause black output?

[high skiff]

no, its better than xformers now

black output is cause auto has a botched implementation of SDXL

[pure hazel]

hmm why do my images turn out black :/ I can see them forming through the noise, but then black at the end.

[high skiff]

its the VAE

[midnight shuttle]

Use --no-half-vae command line option.

[high skiff]

I think there is a way to fix it, but it seems to only be an issue in auto ATM

[south igloo]

funny seems you can accidentally just use the refiner model and get decent outputs 😛

[midnight shuttle]

And remove --disable-nan-check. You want NaN check. If you are getting NaN something is wrong you need to fix. Don't just hide the error.

[trim orbit]

the gains for spd are better on older cards than newer. there's such a narrow margin on all 3 on my 4080 that i don't bother. sdp with attention is the slowest by a hair

[high skiff]

yeah, its its own model lol

[south igloo]

forgot to change from refiner back to base after img2img

[queen ether]

Goodbye automatic 1111. It crashes everytime I try to load sdxl

[high skiff]

yeah, auto has a really bad implementation of SDXL ATM

[trim orbit]

delete the venv and rebuild

[south igloo]

also i threw the refiner at a 2048x2048 highres fix and it worked fine

[vocal rapids]

what's your workflow to upscale to 4096

[ionic dragon]

@high skiff Joe, said that 1.0 base is batter than 0.9(base+refiner) is that true?

[high skiff]

I would say in some cases

[south igloo]

how much vram did 4096 use??? 2046 used 24gb

[high skiff]

the 1.0 is a lot better than the 0.9 base, but I would not say its consistently better than 0.9 base and refiner

[high skiff]

ok, you must be doing something weird

4096 used 8.7GB VRAM for me

if in Auto, that explains why

[vast narwhal]

Highres fix 2x and Ultimate Upscale 2x

[high skiff]

I really can't recommend against auto enough ATM. Its just overall a worse SDXL experience

VAE issues, no mixed diffusion support, bad VRAM optimizations, slower gen times

[vocal rapids]

thx

[midnight shuttle]

I think it will be hard to implement good SDXL workflow in A1111. I looked at the code and it's what you could call spaghetti code.

[high skiff]

for sure

[midnight shuttle]

It was never made for a major workflow update.

[high skiff]

I have no idea how auto plans to fix it

if he even does

[trim orbit]

i'm happy to use auto still in a lot of cases. i jump back and forth between ui's a lot lately. AUTO still has tons of great extensions i'm familiar with and they work really well with sdxl

[midnight shuttle]

If I had no job I would launch a Kickstarter to fix it. But I have a job. 😆

[trim orbit]

you can probably get most extension's functionality into comfy ui, but just doing roop was a greulling week of research and digging around for the multiple nodes i needed to install to get roop and codeformer working

that was a simple one

[high skiff]

I have no job, and I have no idea how to fix it lol

[viral lily]

Can I ask a question? How does SDXL 1.0 implement selecting different styles here? Are different models trained for different styles?

[midnight shuttle]

I think I could need 3-5 days just to understand the code.

[trim orbit]

i think this is the last pass for auto. voldy will either make something new or fall off.

[dense chasm]

currently there are only 1.0base and refiner models in diffusers,but once u begin to run,more powerful than comfyUI

[high skiff]

diffusers is cool

[maiden sage]

hey so I had SDXL working in A1111 earlier today but now the model is failing to load? was there an update I missed or something

[high skiff]

and pseudoterminalx and I worked together to implement my workflow in diffusers

Which ended up being taken on by the lead dev as their new pipeline in general

[trim orbit]

i hear this a lot about diffusers. nobody has ever explained to me HOW diffusers are more powerful

as i figured it, it was like directx vs opengl. two libraries trying to achieve the same end goal

[pure hazel]

Woah doing a high-res fix on an SDXL blowsup my ram lol

[marble hamlet]

Is this server currently running sdxl 1.0?

[maiden sage]

pretty sure in the talk earlier they recommended not to use hi-res fix with sdxl

[viral lily]

How does SDXL 1.0 implement selecting different styles here? Are different models trained for different styles?

[trim orbit]

d8hazards UI, which i guess he nuked because there's literally NOTHING anywhere. even his discord server is just wiped clean of any mention of it, i used that and diffusers on it sucked. they needed to convert every model i wanted to try for like 20min

and seemingly to no benefit. less worked

[marble hamlet]

.

[trim orbit]

i dont' see the point in diffusers. how does it benefit me. the end user.

[marble hamlet]

Yoo it's meat loaf

[trim orbit]

in death, members of project mayhem do have a name, and his name is Bob

[viral lily]

👀

[marble hamlet]

HIS NAME IS ROBERT PAULSON, HIS NAME IS ROBERT PAULSON

[trim orbit]

they're just text on the end of the prompt far as i know

[worthy bear]

how long would it take to train a control net model for SDXL?

[high skiff]

upscale is working pretty good

[pure hazel]

Which upscale?

[midnight shuttle]

Training time for any model depends on data set size and number of epochs.

[trim orbit]

https://github.com/twri/sdxl_prompt_styler/tree/main here's a node to use them locally

[high skiff]

look at them details

[high skiff]

workflow I am mixing into my SDXL 1.0 workflow release

[trim orbit]

about 3.. no 5

[viral lily]

thanks！

[heady vale]

Im also playing with UltimateSDupscale in comfyUI. First attempt

[trim orbit]

but for realisies, i don't know how long it'll take. i just know the original author made v1 entirely on one 3090. here's the manpages on trainign https://github.com/lllyasviel/ControlNet/blob/main/docs/train.md

[dense chasm]

can u share with me?i'm learning diffusers right now

[high skiff]

I don't have any info on it unfortunately, that was pseudos work

I provided how it worked, and he made it work in diffusers

[orchid oar]

https://www.reddit.com/r/StableDiffusion/comments/15afvnb/sdxl_various_styles_keywords/

These are the Styles for sdxl

[marble hamlet]

someone pls tell, Is this server currently running sdxl 1.0?

[raw prairie]

Yes

[raw prairie]

The bots are running 1.0

[south igloo]

anyone worked out how to fix the long limbs on people thing?

[modest mulch]

hi, please help i am new to stable diffusion and whatever image i generate it comes like cartoonish non realistic ugly image. how to get realistic image any settings please helps.

[raw prairie]

Set a 1024×1024 res

Not 512×512

[marble hamlet]

Tnx

[elder rose]

Played with it a bit, it has potential for sure(if you like making grids)

[dense chasm]

where is he?a couple of days not see him in this channel

[high skiff]

He left the server. He doesn't want to be here anymore

[inner maple]

How could he leave all of us?

[high skiff]

He has many reasons to

*had

Stuff I'd wish not to get into, personally

But he's left for good reason, and he won't be back

[maiden sage]

When I load SDXL model in A1111 I get:

changing setting sd_model_checkpoint to SDXL\sd_xl_base_1.0.safetensors [31e35c80fc]: RuntimeError
Traceback (most recent call last):
  File "C:\Users\xxxx\Documents\code\stable-diffusion-webui\modules\shared.py", line 605, in set
    self.data_labels[key].onchange()
  File "C:\Users\xxxx\Documents\code\stable-diffusion-webui\modules\call_queue.py", line 13, in f
    res = func(*args, **kwargs)

etc...

I didn't have this issue earlier, anyone else encounter this?

[inner maple]

On another note, I am having issues turning one photo into something else. Any tips on that? For example if I want to take a pic of me and say "me as an elf" how would I do that? I'm using a1111 and I try to keep the diffuser low but it either changes the image too much or not at all. I guess I could inpaint everything but my face, but what if I want my face to change slightly?

[pure hazel]

What am I missing, both in comfy ui and in 1111, using the refiner gives me identical results no matter what denoising im using.

In fact it's less detailed, often a little blurrier.

I've had this issue since 0.9 leak

Comfy ui def has a way faster workflow without having to switch to and from refiner each time.

[lusty raptor]

sounds like you're using it as img2img? Have you checked out other peoples workflows?

[visual glade]

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

are you using these?

[lusty raptor]

Best bet is to train your face as a LoRA. Selective inpainting is a chore, especially if you wanted to make your cheekbones more elfy or something.

[steady chasm]

In a1111, under the img2img tab, import your photo and in the prompt instead of "me as an elf" describe the desired image - describe "an elf with dark brown hair, hazel eyes, and square glasses" but obviously describing yourself instead of that
As well, make sure you enable restore faces and turn the denoising strength low, but bump it up if not much is changing
Also, like throttlekitty said, a LoRA of your face is a definite solution but it will take training photos of high quality and training time etc

[steady chasm]

Without a lora you'll get elves that look kinda like you, but with a lora you'll get you as an elf for sure

[heady vale]

anyone else have an issue with comfyui where the save image preview is a cached image from a previous gen?

[dense chasm]

holyshit,he is cool,sometimes we can discuss techniques together

[hollow halo]

yeah

[vast narwhal]

Hey comfy. Do you remember this node? It have a notable influence in SDXL generation, often positive. I couldn't find info about it anymore

[visual glade]

https://github.com/comfyanonymous/ComfyUI_experiments

[high skiff]

@visual gladeHow would I go about submitting my final 1.0 workflow to you?

I am still working on it, and it will have some third party upscale nodes, but I can remove the upscale part to make it comparible with your wiki

[gloomy barn]

is there a way to change the output dir for Comfy?

[visual glade]

oh just give me the json or an image with the metadata and I'll do a quick check and add it to the example page

[high skiff]

awesome, goal is to have i tout within the next couple days

Is there any way I could link it with a link to my github for the other offshoots of it?

[high skiff]

Cause I will have a dedicated upscale workflow built off of it as well

[dense chasm]

diffusers is faster than comfy,i prefer to dig in

[visual glade]

sure I can add a link

[high skiff]

awesome, thanks Comfy!

[heady vale]

Ive seen some people with custom save image nodes with a bunch of parameters to change, but I cant seem to locate it even after installing a bunch of scripts

[hollow halo]

I inject a second prompt halfway through my workflow and only hit the issue if I change one prompt but not the other

[glad fulcrum]

is this how to load the Lora?

[visual glade]

--output-directory

[floral tartan]

draw a tiger

[high skiff]

tigor

[visual glade]

yes

[vast narwhal]

Ty.

[glad fulcrum]

Does using the lora cause the generation to be waaay slower? or is it something at my end?

[dense chasm]

diffusers can use the refiner model instead of inpaint nodes,more convinient and custermised

[high skiff]

In comfy we trust

[visual glade]

it needs to apply the lora

but it shouldn't make it much slower

[high skiff]

check your VRAM

SDXL LoRA's seem to hit VRAM massively

[visual glade]

don't use --highvram or --gpu-only when using loras

[glad fulcrum]

11,6/12 my VRAM

[high skiff]

yeah, one LoRA with --GPU-only with my 3090 nearly maxed my VRAM at 21.8GB

[hearty ginkgo]

Comfy do you know why --lowvram dosnt won't on amd cards?

[glad fulcrum]

oh that might be it. why though? if I don't use highVram, my generations are slower as well, because the model loads every time

[high skiff]

the LoRA loads a ton more into VRAM, which makes it pool into share system RAM, which is much slower

[visual glade]

yeah lora loading is optimized for the regular vram option

didn't optimize it for the other options yet

[glad fulcrum]

so it will be optimized in the future? ok

[high skiff]

glad to hear the is potential for that <3

[visual glade]

it should work

[gloomy barn]

where does that go?

[soft bone]

network dim doesnt effect training time does it?

[visual glade]

command line option

[high skiff]

I didn't see any difference

I tried DIM 8 and 128

but I was not specifically looking

[soft bone]

SEcourses has been using 256 network dim, and 1 alpha. I wonderrr

[high skiff]

secourses

[hearty ginkgo]

It didn't I tried to get it to work on my friends amd and when I do --normalvram it says out of vram and when I told him to use --low vram it says Cuda not enabled he has a 12gb vram card

[dense chasm]

diffusers loads GPU VRAM highly as well,i have to set the base model pipe image size under 768*768,otherwise VARM usually exceeds 15GB

[soft bone]

idk his lora results are still the best ive seen and that was weeks ago on 0.9

for faces

[heady vale]

edit the bat file and add it to the end like I have

[visual glade]

comfyui works on free colab with the workflows on here: https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

1024x1024 with base + refiner

[high skiff]

I did a subject LoRA test, but it wasn't a human, so maybe its different

[gloomy barn]

I am looking at it

[high skiff]

I don't care much for him in general

[soft bone]

i said the same about aitrepreneur and got booed 😦

[high skiff]

oh man, screw Aitrpreneur

genuinely one of the most trash Ai youtubers out there

almost all his info is hot dog water

[soft bone]

clickbait

[void parrot]

Hey @high skiff , did you ever figure out what was making your generations look noisy and grainy?

[high skiff]

not just that, but flat out wrong too

[high skiff]

hmm... I don't remember having that issue

when was that?

[trim orbit]

youtube has generally just straight sucked since the thumbs down button was killed and everything became a race for traffic grabbing

[soft bone]

yep exactly. giving one-size-fits-all params that he made for one specific dataset and use case, literally saying "these are the best settings"

[void parrot]

I thought it was you, but maybe someone else. Right when SDXL dropped. My images are generating with a lot of noise, like they're not fully processed or something

[gloomy barn]

That causes an error

[midnight shuttle]

Clickbait is always a problem for popular topics. As is the problem of people who don't know what they're talking about spreading misinformation by repeating what they heard on other videos.

[high skiff]

bruh, his settings aren't even good for his own usecase lmao

[heady vale]

paste what you changed

[high skiff]

sorry, i have been so busy since launch, i don't remember much 😅

[nocturne dove]

That guy makes a lot of videos and stuff but doesn't seem to understand much (secourses). He makes so many nonsense videos of "tests" where nothing is properly tested. But other people don't make new videos to properly explain newer lora options/types,

[soft bone]

he does skimp out on the testing process but tbf his videos are already 45 minutes

[nocturne dove]

I've seen him threatened with bans in other discords for bothering the devs with too many dumb questions.

[soft bone]

of mostly useless info

[trim orbit]

they're mostly all just producing pulp filled videos

farming traffic and ad dolla dollas

[midnight shuttle]

Most 45 minute videos could be replaced with 3 paragraphs and 2 screenshots. 😆

[trim orbit]

poopular channels can make 6 figures a year

[gloomy barn]

From this .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build

To this .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --output-directory D:\SD Images\Comfy

[soft bone]

someone make a gpt script to summarize the main points of ai yt videos

[heady vale]

does that folder exist on your machine?

[gloomy barn]

yes

[trim orbit]

fr

[heady vale]

Im not sure then. works perfect for me

[trim orbit]

recipes are my big golden goose example here. you search for a recipe website, you're goiong to find some bullshit junk that's irrelevant and just spamming you with dumb content. you ask gpt for a recipe and it tailors it to you perfectly for exactly how you asked

[gloomy barn]

I guess it didn't like the space in the dir name

[soft bone]

those blogs are straight up reader-abuse

[midnight shuttle]

Need to put quotes around file names that contain spaces.

[soft bone]

or news articles that lock half the article behind a paywall while you're reading

[hoary saddle]

pretty consistent results @nimble heart

[trim orbit]

traffic farming is just a plague on the internet. enabled by marketing dollars

[gloomy barn]

ah

[soft bone]

still not as bad as the microtransactions in gaming

[void parrot]

Does 512x512 work with SDXL? I get comically bad results, that look like horrendous quality. Trying to see if I have some troubleshooting to do.

With Auto1111

[midnight shuttle]

No. Use one of the supported resolutions.

[hoary saddle]

if i had a transparent PNG and i wanted to overlay it onto the resulting image, is that something ComfyUI can do?

[void parrot]

That's good to know. So minimum of 1024x1024?

[midnight shuttle]

1024 x 1024
1152 x 896
896 x 1152
1216 x 832
832 x 1216
1344 x 768
768 x 1344
1536 x 640
640 x 1536

[gloomy barn]

Now it's going to the right dir. Thanks.

At last I can have my revenge on the Jedi

[midnight shuttle]

Use one of those.

[trim orbit]

naw. it's about the same. same problem. corps needing their profit margins to always grow

[void parrot]

Oh nice! So it's not just all square anymore?

[trim orbit]

but this is a wholly nother topic now

[midnight shuttle]

Right. Using one of those will give best results.

[void parrot]

Amazing. Thank you!

[high skiff]

updates to the upscale workflow are showing proper deformity repair

the weirdest changes are making the biggest differences lmao

more to come

[soft bone]

Something I was surprised to see from SEcourses training was the use of regularization images, using real photos no less.. He swears it helps realism. I haven't used reg images for small datasets since DB on 1.5

[heady vale]

nice. comfyui also has ctrl+enter shortcut like A1111 to start a gen

[high skiff]

I train LoRA's with reg data

using reg data from the direct model you are training is much more beneficial from my tests

[high skiff]

yeah! found that WAY too late lol

[soft bone]

the way he put it was "it helps if the images are better than what the model can already produce". I've never heard that, and I've found better results using generated images as well.

[high skiff]

best part of my upscale workflow

Only takes about 2.5x as much time as a normal gen

so for me, from unloaded TE's to 2048x upscale, it takes about 35 seconds on my 3080

[high skiff]

yeah, that claim makes nos ense

reg data is so the model can see what it has to change better

it also helps with style consistency across models

cause it can see better what parts were from the base model, and what was applied as an underlying concept

[dense chasm]

it's awesome when GPU VRAM is been occupied highly under diffusers with following codes:pipe.to("cuda")
pipe.enable_model_cpu_offload()
refiner.to("cuda")
refiner.enable_model_cpu_offload()

[wicked frigate]

When you load a custom workflow it should provide every input as an option in the other menu - that is, assuming it didn't break? The firstmost sampler is automatically redirected to the default sampler param and any other sampler inputs are given their own KSampler group. At worst it probaby registers the secondaries as text inputs rather than dropdowns (need to fix it to recognize the valid input options)

[high skiff]

with normalization, you can train an anime guy into an anime model, then take that LoRA and put it on a relaism model, and it should be able to generate a fairly realistic version of said anime guy cause it knows what the underlying concepts were that it had to process through its modifiers and weights

[midnight shuttle]

Scalpers are selling 4060 Ti 16GB for $600 and people are buying. 🤮

[polar epoch]

Man, takes time to finetune whacky gens lol

[high skiff]

if its an anime model, it doesn't ahve to learn a new anime style, it just needs to learn what human qualities are being put through its predefined anime parameters

Which allows it to trigger those human weights the same in a realism model and so on

[soft bone]

then wouldnt it be wise if you were training on background-removed images to also use background-removed reg images so it knows better that you're only wanting to change the "person" data?

[high skiff]

now that I have no idea

I usually just tag for lack of background

[soft bone]

i just do white bg and tag "white background"

but i dont use reg

[dense chasm]

a bunch of customers wanna controlnet or 3D architechture experiences,can u dig in and do some research work on this topic?thx

[high skiff]

I am not edcuated enough for that kind of stuff unfortunately

I am not a coder, I am just really good at usning the tools we do have efficiently

[dense chasm]

so where is the hell of pesudo

[slow marsh]

Hey everyone,

I have a quick question that I'd love your help with. We've been using Stable Diffusion XL on both Stability Developer platform and DreemStudio, with the same model (SD XL) and input settings for both. However, we're getting quite different results, and we're curious about the version of Stable Diffusion XL used on the Stability.ai developer platform. The results from that platform have been impressive, and we'd like to know if we can access the same version for other purposes.

We get good results from Stability developer platform so we want to have that instead of DreamStudio one. If anyone knows which SD version is used on the Stability Developer platform and whether it's open-source, I'd really appreciate your insight. I've attached some details and outputs from the Stability Developer platform where we got good results.

Thanks in advance for your help, and I'm looking forward to hearing from you!

[high skiff]

Thats all he said in response

[dense chasm]

i miss him too lol..

[midnight shuttle]

Your screenshots are mostly too small to read. But already I see you are using different number of steps (50 vs 30). That will cause different output.

[raw prairie]

who is pseudoterminalx?

[dense chasm]

a python expert

[high skiff]

My counterpart, who is just as mad at coding as I am at screwing with things

[raw prairie]

like rick and morty

[high skiff]

kinda haha

hes a python master, I am a fuckery master lol

I think of how to do something, he makes it work lol

[raw prairie]

interesting

[high skiff]

I made my mixed diffusion pipeline for comfy, he adpoted it into diffusers main branch

[dense chasm]

why he block my friend adding request?

[high skiff]

I think he has them all off by default

[slow marsh]

Actually I tested them previously, with the same input setting, this screenshot is just for me to show you all the SD version haha

[high skiff]

lmfao

i am just the messenger lol

[midnight shuttle]

You tested with same seed, no style, etc?

[dense chasm]

tell him i'm digging in diffusers right now, i wanna to discuss with him

[slow marsh]

Yep correct

[sullen linden]

so I think the answer is no, but the easynegative negative embedding won't work on xl right?

[earnest anvil]

guys, in comfyUI, are supporting terms weighted less than Linguistic Positive?

[midnight shuttle]

Also the same sampler? I don't think you can even choose sampler on DreamStudio, so how do you know which sampler is used? You might need to ask Stability.ai support. Why do you need exact replication?

[dense chasm]

thx buddy

[high skiff]

no problem

[slow marsh]

Oh yah, that's a great question!

Thanks man!

[dense chasm]

why u expose you are a chinese,be careful about the national security,Imao

[old mulch]

Some SJW

[slow marsh]

lmao, i'm asian bro

HAHAH

[old mulch]

Do you guys prefer dpm 2m karras++ or euler a? For sdxl

[rustic surge]

hello man,where are you from

[midnight shuttle]

Euler for people or less detailed scenes, DPM samplers for intricate details. But others may say different.

[shy kelp]

My initial consensus over the last couple of days is that SDXL 1.0 is going to be an AMAZING model to work with......in about 4 months time.

[dense chasm]

just no kidding,i have been dealing with local police,FBI or CIA is far more beyond your imagination

[old mulch]

Ye its slow af atm getting 44 sec gens on only 20 steps 64gb ram

Beautiful, but slow

[midnight shuttle]

How much VRAM? That is what matters.

[old mulch]

This except i got 64gb core gpu

[midnight shuttle]

Oh, M2 unifed memory. I don't know much about that.

But it's not too bad I think. I get about 25 seconds for Euler 20 steps.

[old mulch]

😦 nice

[midnight shuttle]

It would be faster if the VAE didn't overflow into system RAM. A1111 problems...

[rustic surge]

Currently Comfy UI has more friendly support for SDXL, with faster running speed, especially on devices with less VRAM.
But Automatic1111 has more features and supports more extensions, with more flexible operations.
For high-end GPUs, Automatic1111 can run SDXL normally and quickly. But for low-end GPUs, Comfy UI performs better.
Comfy UI has very good backend optimization, it can run SDXL even with small VRAM. Automatic1111 consumes more memory.
When SDXL was just released, Automatic1111 needed various command line parameter tuning to run it. Comfy UI can be used directly.
Automatic1111 needs to first generate the initial image with txt2img, then refine it with the refiner model using img2img. The operation is more complex.
Comfy UI can complete generating and refining the image in one step. More simple and straightforward.
Some users developed extensions to integrate Comfy UI into Automatic1111, combining the advantages of both.
For users pursuing performance, currently Comfy UI is a better choice. For those pursuing flexible control, Automatic1111 can be used.

This is a summary made by Claude2 after reading through the chat records in this channel.

[midnight shuttle]

I'm impressed. I didn't realize that was AI generated.

[shy kelp]

Yes. plus I think refined custom models will truly be able to get 1.0 to do amazing things. But that's going to take time...and a few A100s

[midnight shuttle]

I thought you typed that.

[sullen linden]

rate my dreegon

hmm why is my background trash

[shy kelp]

Yes, whoever is the brilliant nerd behind a1111, he needs to do some more refinement. It's going to take time for multiple elements and custom models, working together to overtake sd1.5 but it's happening.

[dense chasm]

dragon is evil or good?

[shy kelp]

dragon is an anti-hero

[sharp robin]

my dragon is a little deformed, but cool too!

[shy kelp]

he's still dangerous but sometimes he does the honorable thing

[midnight shuttle]

I've looked at the code and I think it will be difficult to change it to work with SDXL. So maybe it will take some time.

[shy kelp]

@sullen linden lol sorry man, @sharp robin has got your dragon beat!

[sullen linden]

is hires not working right... that's the second time the background has seemed real blurry

[old mulch]

How did u get the chat logs?

[rustic surge]

copy

[dense chasm]

game of throne,empire symbolic,dilemma

[old mulch]

Oof thats painful

[midnight shuttle]

Pretty sure Discord API also lets you get channel history.

[rustic surge]

yes

[shy kelp]

anybody remembers 2002's "Reign of Fire" ?

[dense chasm]

big data time,trace everwhere lol..

[rustic surge]

So how to use Discord's API

[old mulch]

U prob need a key for this chat

So prob useless

[midnight shuttle]

https://discord.com/developers/docs/intro

But I've never used it so I don't know if it can do that. I would have to read the docs.

[rustic surge]

This is very important for learning SD

[midnight shuttle]

No need for sarcasm. I will leave.

[rustic surge]

Because I have to look at your chat history one by one every time and learn from it

[old mulch]

I dont think he was being sarcastic lol

[urban fjord]

Dragon pirate pope, it's hard to mix consept at times with SDXL but luckily after many attempts it worked.

[rustic surge]

I'm not sarcastic

[midnight shuttle]

OK I misunderstood.

But I must leave anyway.

[rustic surge]

why！！

[midnight shuttle]

I cannot spend all time in Discord. But perhaps I will return in the future.

[fast vector]

hey does, anyone have adetailer worfklow for ComfyUI?

[sullen linden]

yeah hires.fix definitely is destroying my images in a111

[heady vale]

dont use it for XL. I dont think its been changed to work with XL

[sullen linden]

yeah, just trying to make sense of the note for dreamshaper. dude said No need for refiner. Just do highres fix (upscale+i2i) ...

[polar epoch]

Any idea when 1.0 full unpruned will release? As I desire custom models and Lora's lol

[shy kelp]

I get the feeling the entire interface for A1111 will have to be adjusted to accommodate SDXL 1.0

[polar epoch]

Butchers all my memory, but comfyUI and esrgam 4x upscaler model works quite good

[shy kelp]

@polar epoch It's a waiting game. It's gonna be a few months before civitai has vast amounts of custom models and loras for sdxl 1.0

[polar epoch]

So the models we do have are all trainable? Thought the pruned ones were ridden of training data

[boreal heart]

10GB vram gonna be the first threshold though

[sullen linden]

thanks, seems like esrgan_4x isn't completely mutilating my backgrounds, you're right

[candid walrus]

they mean that you generate your image with sdxl, then take the image back into image to image and upscale it with a 1.5 model you like

[hearty ginkgo]

What workflow are you using?

[calm shell]

gow fo I cgange the memory alloc in A1111?

[sharp robin]

it's currently a mess i wouldnt even be able to explain

[calm shell]

oops helps to turn on the lights

[still dove]

this is ComfyUI, right?

[sharp robin]

yes sir

[calm shell]

How do I fix memory (vram) alloc in A1111?

[hearty ginkgo]

Use comfy

[calm shell]

I do use comfy.

but I like A1111 as well

[sharp robin]

idk this is all i run on auto1111, maybe u could try --medvram

set COMMANDLINE_ARGS= --xformers --opt-split-attention
set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.9,max_split_size_mb:512

[urban fjord]

Not really, all it needs is an update to high-res fix to support using the refiner in the same workflow.

[hearty ginkgo]

How much vram do u have total?

[hearty ginkgo]

Is this your workflow for comfy?

[amber vale]

Is there Some kind of roop thing in comfy ui?

[sharp robin]

yes

well worflow in progress

[sullen linden]

which is my best fake

[sharp robin]

3

[calm shell]

My poor 1660 super is vram starved I'm afraid

[sharp robin]

can it not run SDXL?

[azure oxide]

this is adorable considering the history of you two. man i remember this convo like it was yesterday, time really does fly #🏞｜general-with-images message

[hearty ginkgo]

Auto11111 isn't good for low vram

[calm shell]

it can but needs the low vram stuff

[sharp robin]

u can try ArtroomAi as last resort for SDXL more barebones but should get u genning

[candid walrus]

i use comfyui to just gen from the base model, then i take those images into a1111 and do an image to image with my 1.5 models i like. so far i think i'm getting more control and better details

[calm shell]

Did this with Comfy

[autumn forum]

currently working on a movie still lora, left no lora, right lora, turning out pretty well!

[calm shell]

I have artroom, I'm trying it next

Then Invoke

[sullen linden]

blonde looks like the lawyer from walking dead

[still dove]

where do we use lora? after base or after refiner?

[autumn forum]

in the base

[hearty ginkgo]

Invoke is stepping up there game in my opinion haven't tried it since the major hyper of 1.5 2.0 stable diffusion came out tho

[still dove]

ty

[calm shell]

yes but you still have to cut and paste directories, no selectors

In invokeAI

I like the way it handles Lora though

[wet raven]

well, my ability to use SDXL at 1024x1024 on Automatic1111 was short lived.
even at 920x920 i get "out of memory"

"torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 20.00 MiB"

[dense chasm]

try the following code if you know some basic python code::pipe.to("cuda")
pipe.enable_model_cpu_offload()
refiner.to("cuda")
refiner.enable_model_cpu_offload()

[sharp robin]

1080ti?

[dense chasm]

with vae:pipe.vae.enable_tiling()

[sharp robin]

i can runn it on 1080ti no problem

[sharp robin]

#✨｜sdxl message

if that dont work add --medvram

[candid walrus]

anyone struggling on a1111 with memory, it's worth a try with to download the stand alone comfyui and check this post out to see if you can play with it https://www.reddit.com/r/StableDiffusion/comments/14sacvt/how_to_use_sdxl_locally_with_comfyui_how_to/

[sharp robin]

damn the cigar fused with the skin, but impressive details

[candid walrus]

that's nice

[bold osprey]

seems to be a pattern in my experiments so far

[wet raven]

I've done that perviously, but not with the 1.0, so i'm trying it now

[shy kelp]

Unfortunately it seems like hands in the SDXL 1.0 base is worse then finetuned 1.5 models. Has anyone figured out how to make hands generate better or do we need to just wait for the finetunes?

[still dove]

I am with 1080ti 11GBvram

facing no problem at all

[bold osprey]

hands lora would be a start

[sharp robin]

Same here 1080ti no issues

[wet raven]

#✨｜sdxl message
my settings now:

@echo off

set PYTHON="D:\ai\a1\venv\Scripts\python.exe"
set GIT=
REM set VENV_DIR=D:\ai\a1\venv
set VENV_DIR=
set COMMANDLINE_ARGS=--api --no-half-vae --disable-nan-check --xformers --opt-split-attention
set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.9,max_split_size_mb:512
call webui.bat

[shy kelp]

Is there a hands Lora for SDXL out now?

[candid walrus]

while the base image from sdxl isn't much that i desire on the left, it has some data that ends up getting what i like out of my 1.5 models with image to image. Enough for me to spend time generating some starting points then hand pick them to get some nice results

[bold osprey]

still need to figure out inpainting with comfy

[short marsh]

What’s everyone using for their 1.0 LoRa training?

[candid walrus]

sdxl just has a better vocabulary/life experience

[wet raven]

@still dove and @sharp robin what is the max resolution you can do?

it's letting me do 1024 x 1024 again, but i think if i try 1280x1280 it messes it up, i'll see though

[hard fractal]

🅰️ 🅰️ 🅰️ 🅰️

🅱️ 🅱️ 🅱️ 🅱️

or

[sullen linden]

3b got that groot energy

[autumn forum]

hey joe! im working on a movie still lora and its turning out well! left is lora, right is base.

[candid walrus]

i haven't gotten into training yet, i would probably rent a 24gb card and use the guide on their example files with their 'diffusers' like i have with 1.5, just because i never seemed to get the same good results as fast using things like a1111 or kohya_ss so far (as i tried them locally with lower vram)

[still dove]

I am doing 768x1344 right now

[wet raven]

yep, after trying 1280x1280, i can't do 1024x1024 now

[sharp robin]

B by a mile. It’s subjects are more ‘complete’ more going on that seems to make sense. A is like busts of heads.

[wet raven]

the VRAM gets stuck at full, once i try too high of a resolution

[sullen linden]

I like a2 the most but that's just because it has a sense of scale from the little d ude

[candid walrus]

yeah so far i try to keep my ratios the same as what you'd get with 1024x1024 so like 64 * 16=1024, so if you want to change aspect you can do like 64 * 14=896 and for the other side you would take the 2 difference you removed from 16 and then do like 64 * 18=1152... 896x1152 is the same as 1024x1024 pixel density

[still dove]

may be you are using your gpu on something else as well

[wet raven]

no, i've even tried closing everything, discord, etc and only 1 browser with 1 tab for stable diffusion

[candid walrus]

right, also you can also try to restart the bat file if somehow the vram hung (assuming you started the interface with a bat file)

[chrome flicker]

Hi guys, I find that sdxl 1.0 do img2img to get varition img compared to source is diffuicult. I modify the denoise in the base sampler. But small value cause little change and bigger value got unrelevant image compared to source. How to get variation image using img2img in sdxl?

[candid walrus]

when i do the image 2 image from sdxl, i tend to have a denoise of .5

[still dove]

but why do you want 1280 output? you can upscale your 1024 render upto 4x

[wet raven]

testing purposes

[candid walrus]

but depending on things i'll move it, also i'll use canny controlnet with varying strengths too on what i'm trying to get out of the image

[still dove]

right

[rotund bear]

does anyone have Collab Version Link?

[chrome flicker]

0.5 I got very close to input.

[candid walrus]

on some models i'll like the detail better at high 1280 vs even 1024, but its rare i wont get it at 1024

[rotund bear]

we want collab version link

collab version link can save our lives

[candid walrus]

yeah i wouldn't really do image to image to get that much a variation from what i've been able to achieve

[chrome flicker]

sdxl can still not do img2img like midjourney that with an input then get very creativate output related to the input

now results shows that very close to input or very different

[candid walrus]

i think you could possibly try the reference_only control net but i'm still not thinking you'd get that same midjouney variation button

so like txt to image with the reference_only controlnet

[rotund bear]

sdxl is improving day by day

[chrome flicker]

yeah, reference only is support in sdxl?

[wet raven]

Stable Diffusion 1.5
at 1280x1280 no upscale

[sharp robin]

That’s actually crazy

[wet raven]

Stable Diffusion 1.5
at 1280x1280 no upscale

[still dove]

you mean no sdxl?

[wet raven]

no, it's 1.5,
the Photon SD 1.5 model works well on very high resolution

[still dove]

photon is impressive

[wet raven]

however, if you want to make people, you are highly likely to get the repeating bodies

[nimble heart]

that big Principled node update I was warning you about is live. Should do text2img with pixel upscale much better since the low res pass can use the refiner

[cunning peak]

https://youtu.be/EVFrmOxirk0
Txt2Img with SDXL 1.0 Base and Img2Img Enhancing with SDXL Refiner using Automatic1111 (My first attempt at a Tutorial/Walkthrough)

[sharp robin]

U don’t feed it a latent img?

[nimble heart]

Latent is optional

if you don't feed it one it creates an appropriately sized one automatically

the VAE is also optional, it's just needed for pixel scaling

[wet raven]

" --medvram" is slower right?

[nimble heart]

kinda

[wet raven]

well, --medvram let me make a 1280x1280

[nimble heart]

Yea

it should probably be the default

all it does is let things unload when they're not being used

[wet raven]

SDXL 1280x1280 silly picture, most of it is empty space

[nimble heart]

make your prompt longer

[wet raven]

yes, it is unloading the VRAM like it used to

"hd 4k, intricate detail, masterpiece, award winning macro minnow, photo from a Fujifilm XT3 with ZEISS lens of"
haha, thats the positive prompt i used

[nimble heart]

ah, macro is probably making it all blur

[wet raven]

oops, doesn't help much when you use the same seed...

[still dove]

same prompt

[wet raven]

i've done lighting effects in stable diffusion

[still dove]

I want this workflow, how can I get it?

[wet raven]

SDXL 1280x1280

[still dove]

how are you doing 1280x1280 on sdxl now?

[nimble heart]

From my git
https://github.com/Beinsezii/bsz-cui-extras

[wet raven]

you can set the resolution in automatic1111

i'm going to see how high i can push it

[nimble heart]

I see 2048^2 a lot

[wet raven]

1400x1400
i'll try 2048x2048

I got a popup warning saying "you are attempting to do 2048x2048 on SDXL which is against the law and you will now be arrested" (not really)

13 seconds per iteration

[nimble heart]

what in god's name is that negative prompt @wet raven

[wet raven]

haha, i just reuse and copy/paste stuff

[wraith apex]

Is automatic1111 now default with sdxl? If i do git pull, or do i need to download xl.model separate?

[candid walrus]

i'm not sure, i haven't looked into anything control net with sdxl

[wet raven]

sometimes i accidentally have stuff in the negative that prevents me from doing what i'm trying

[wet raven]

it ALMOST made the 2048x2048 but aborted.
this is the last preview i got of the image before it deleted it

[still dove]

how can I install it in comfyUI?

[fleet harness]

you got grandma and grandpa in that neg

[wet raven]

i got that from someone else, i thought it was funny so i added it

[nimble heart]

put the files from my github's "custom_nodes" folder into your ComfyUI install's "custom nodes" folder.
the workflows you can just load by dragging them onto the canvas while comfy is open

[still dove]

ty

[fleet harness]

I was dying reading it lol nice find

[wet raven]

"several" "large group" are ones i was doing to try to do a high res of 1 person, but may not be good for most things

[nimble heart]

2048^2 for 20+20 steps
2.25 seconds per iteration for me

kinda wanna try 3k but the decode already takes like 60 seconds @ 2k

[wet raven]

20/20 [04:30<00:00, 13.50s/it]

[nimble heart]

not that bad for a 5 year old card

my 1070 really struggled

[wet raven]

SDXL: 64x64, 1 step

[wintry dragon]

Can anyone point me to reference where it is explaind why text2image checkpoint work for inpainting as well for sdxl?

[midnight shuttle]

It doesn't. If you inpaint only masked you get the same problem as using 1.5 or 2.1 non-inpainting model.

[wet raven]

beautiful

[fleet harness]

I love how well SDXL can do realistic painting as if it were captured in print

[calm shell]

had to upgrade my nvidia drivers, and python 3.10.7

[midnight shuttle]

Python 3.10.11 works fine and has security fixes. Guides that say to use 3.10.6 are outdated. But do not use 3.11 on Windows due to Pytorch limitation.

[nimble heart]

What's wrong with 3.11?

[midnight shuttle]

Pytorch has some parts that aren't supported on 3.11 on Windows. I forgot the details.

[wet raven]

so it looks like, when stable diffusion "crashes" from running out of VRAM, it doesn't seem to know how to reset itself without restarting the entire .bat file

vram gets stuck in "used"

[midnight shuttle]

There is probably something that doesn't get unloaded. Best to restart at that point.

[heady vale]

in A1111 go to settings >actions > unload checkpoint, then reload checkpoint buttons

[wet raven]

tried

well, then again, it let me do 1280x1024

but the VRAM was maxed the whole time at 11gb

whereas before, it was at around 7-8gb

oh, i think part of it is because i tried switching to full preview

i went back down and it uses 9.7gb

[zealous viper]

anyone know if that basic workflow in Comfy for inpainting works with XL if you just attach the refiner to the output before the VAE decoder?

SUPER new to comfy

[marsh sapphire]

Is it me or is SDXL considerably slower that every other model? Normally on 1024x1024 I get 2s/it with my 4070 now with SDXL I am getting 30s/it. Anyone else's experiencing this?

[midnight shuttle]

Using A1111 or Comfy?

[marsh sapphire]

A1111

[midnight shuttle]

Due to VRAM. SDXL needs 16 GB VRAM to run at full speed on A1111 right now.

When all VRAM is used it swaps to system RAM and performance plummets.

[marsh sapphire]

Wow, that expands it

[zealous viper]

as someone with 8gb vram, when i switched to comfy full render process is 30 seconds instead of 5 mins

worth the headache to learn to me

[marsh sapphire]

Guess il be learning comfi today

[nimble heart]

just use latent preview always. You get used to the way it looks after a while

[marsh sapphire]

Can I import my created Syles to comfy?

[zealous viper]

prob but i dont know the workflow to have "saved prompts" other then to save a workflow file for each prompt ... prob not the right way to do it

[midnight shuttle]

I don't know why A1111 uses so much more VRAM. It seems like it should be the same since both use the same technology.

[sly jay]

it kinda is worth it. I still feel much less powerful with comfy even if I can do a gen in 30 seconds vs 11 mins lol

[wintry dragon]

then how are people using sdxl checkpoints for inpaint?

[wet raven]

png TAESD Combined looks nice at 50%+

[midnight shuttle]

They are using "inpaint whole picture" instead of "inpaint masked".

But masked works better for many workflows.

[nimble heart]

even taesd has a performance hit on SDXL. I just use the cheap one.

[calm shell]

A1111, slow but it worked GTX 1660 S, 640x640

[wet raven]

TAESD on left, right before finished
Final Output on right

the cheap one does speed up generating though

i need to feed my minnows

[zealous viper]

my look on it is that i just need more time to get the workflow into muscle memory, that and I think after a few days when the guys who are super good at programing workflows has their stuff released to get the "fancy tools"

[sly jay]

any tips for the former? I feel like how I felt when i first started using a1111 and had no idea what I was doing, but now I cant even use styles for an early "i win" button lol

My issue mainly(idk if this is because of comfy or sdxl) was that it would follow my prompt and look awful, or not follow my prompt and look amazing

[midnight shuttle]

Comfy is a "cool" interface because it is very complex and hard to understand. But I hope A1111 or SD.next gets good SDXL support because I like those interfaces better for my workflow.

[rigid laurel]

Comfybox isn't that hard to understand

[sly jay]

for example "women wearing in croptop"

[midnight shuttle]

Make sure you use a supported resolution. And use simple prompts. Don't copy/paste long sequences of magic words.

[zealous viper]

XL to me needs ALOT less hand holding on prompting ... i've just been writing them up as I go. all "styles" in A1111 is just a copy paste of a previous prompt

[midnight shuttle]

"women" is plural and you don't "wear in" something.

[rigid laurel]

it gets confused by grammatical mistakes easily, I found this out earlier. Also being more verbose with your descriptions can help.
"A photograph of a woman wearing a crop top"

[midnight shuttle]

Try to write it as "a photograph of a woman wearing a crop top" instead.

[rigid laurel]

better yet, "A photograph of a young woman with black hair wearing a black crop top while posing"

[cyan crown]

https://youtu.be/29hKqm92Rvo

[rigid laurel]

you don't need to add 20 different tags like you did with 1.5, but it still benefits from specificity

[sly jay]

i think this is what it was, but with out "beautiful face" the face looked bad, but with it it was cropped to just the face

"A photo of a beautiful women wearing a black crop top, beautiful face, detailed face"

[midnight shuttle]

That's because you asked for "face". Just generate again with a different seed. You don't need magic words like beautiful, masterpiece, best quality.

[rigid laurel]

50 steps with 2S a
mfer wha

[midnight shuttle]

The more you say face the more it will focus on the face only.

[rigid laurel]

yeah, something like "cowboy shot" might give better results to tell it that you don't JUST want a picture of the face

[sly jay]

better again without the specified face stuff her face looks deformed. "A photograph of a women wearing a black crop top"

[midnight shuttle]

So use Codeformer to restore the face.

Some custom models were badly trained by using tags based on quality (masterpiece, normal quality, bad quality). Instead they should have just discarded the bad quality training data. That is where the habit of magic words came from.

[sly jay]

cowboy shot creates cowboy hat lol

[sly jay]

makes sense

[rigid laurel]

I see no probelm with this

[midnight shuttle]

Also the word is "woman" not "women". Women is more than one woman.

[tall lion]

Try the same on ComfyUI and you will see a huge difference. How much of VRAM do you have ?

[marsh sapphire]

8GB

[sly jay]

generally don't like using codeformers as it doesnt work well with any facial obstructions and doesnt workw ell with non realistic models

[rigid laurel]

I got 2it/s with SD 2.1 and I'm getting 1.5s/it with SDXL

[sly jay]

trying this, my dyslexia may be the culprit

[midnight shuttle]

It's only for use with photorealistic. Other faces aren't supposed to be realistic.

[marsh sapphire]

How?

[rigid laurel]

yeah CLIP isn't programmed to handle language errors very well. dyslexia will absolutely confuse it, but you should be fine if you just paste your prompts into some online grammar checker to catch small things like that.

[rigid laurel]

I'm using Comfybox. Never had to do anything to fix speed.

[sly jay]

okay will try

[marsh sapphire]

Oh I used a1111 up til now. I am now installing comfy ui

[sly jay]

even with grammar fix its mixed results

[rigid laurel]

a1111 and SD.Next technically have "support" for SDXL, but for some reason they have speed issues. Hopefully whatever Comfy is doing gets ported soon.
Also, Comfybox is easier to use than comfyui by itself

[midnight shuttle]

Codeformer on? Sometimes you need to try multiple times. Perfect waifus are not guaranteed.

[midnight shuttle]

The reason is simple. VRAM usage is less on Comfy and is more on the other two.

[rigid laurel]

huh

do you know why?

[midnight shuttle]

No. Looking at the A1111 code gives me a headache.

Maybe they will fix it someday.

[rigid laurel]

A cowboy shot of a black female model wearing a black crop top and posing for the camera

[rigid laurel]

well that could be a reason unto itself 😂

[zealous viper]

my prompting still needs work, but man XL as a base model really is nice

[rigid laurel]

ikr

can't wait until the people who made deliberate and whatnot get to fine-tuning it

[sly jay]

is specifying things like type of shot, skin colour and posing for the camera helpful? I know SDXL likes simple but still unsure on how simple

[rigid laurel]

I want to make my own fine-tune, but I don't have a supercomputer 😂

[sly jay]

looks great!

[sly jay]

grow a 4090

[midnight shuttle]

Rent an A100 for $1.79/hour.

[zealous viper]

yeah, will be nice to have a mecha lora and or model at some point, but this is head and shoulders above what I could do before

[rigid laurel]

with old SD, you had to make paragraphs of "quality tags" like ("Masterpiece, best quality, great lighting, smooth skin, perfect beautiful eyes, five fingers, photorealistic lighting"). The whole "SDXL likes it simple" thing is that you don't need all that bullshit. It still likes it when you're specific. You don't need to, but the less guesswork you make it do, the more likely any given photo is to not suck.

[midnight shuttle]

Base model wasn't trained on those "quality tags". Those are from badly trained custom anime models mostly.

[rigid laurel]

You have my attention.
Do you know of any guides on how to train checkpoints on SDXL, given the intended use of refiners and whatnot in the pipeline? I've been trying to find a way, but despite the hype around SDXL being easier to train, I haven't found a set method for doing so.

[sly jay]

its still kohya no?

[midnight shuttle]

No I don't. I haven't tried to train SDXL yet and I only trained hypernets on original SD.

[sly jay]

oh no thats just loras

[rigid laurel]

yeah, but using those quality tags with models merged with that anime shit generally gave better results than other models with more normal tagging.

[zealous viper]

give it a couple days, let the code heavy guys sink their teeth into and im sure some tuts will be out on youtube over the next week

[midnight shuttle]

I don't understand why they say it is "easier" to train. Training is training. You give it images and captions, set the hyperparameters and off you go. Same for any NN.

[rigid laurel]

I believe the idea is that it responds better to training. Needing fewer source images to be able to represent a new concept without making it look like shit, etc

[uncut steeple]

Somewhat implemented the Ascore into my workflow too now, atleast I hope I did it right

[sly jay]

is there anything else you are doing workflow wise? Even with all of that it isn't listening to me

[midnight shuttle]

What are you using for sampler, steps, resolution?

[zealous viper]

prob because it has a higher baseline for quality due to what I'm assuming is culling and recollection of higher quality assets in the training of the base model, so you have to give it less training because your not having to remove as much "useless" data in the NN

[sly jay]

dpmpp_2m with karras, 25+5refine, 1024 x 1024, cfg 6.5

[zealous viper]

not sure though, just what my logic runs threw

[midnight shuttle]

Those sound like good settings. Maybe reduce CFG to 4 since it's a person. Also can tru Euler sampler instead.

[rigid laurel]

I'm just using euler a with 25 steps, cfg 8, 512x512 for the base image. Then after making a batch of 6, I pick one I like and upscale it by 1.5x at denoising 0.5, 20 steps

like so

those fingers are made by lord satan himself, but everything else is pretty good

[sly jay]

512 for sdxl? my experience it loses it shit at anything below 1024

[midnight shuttle]

512x512 with SDXL? Normally that looks bad.

[rigid laurel]

lol, both of ya at once

I just do it for basic composition

[winged bear]

anyone know how i can choose what gpu to use in comfy ai. I know how to do it in stable diffusion but im not seeing how to pick between 2 diffrent gpus in comfy

[rigid laurel]

I've found it does better with a low-res source image than doing a 1024x1024 image from scratch. Like I said, I upscale it during the 2nd step. It gets to 1024.

[dusk mica]

How many steps do you guys use?

[midnight shuttle]

20 to 50 depending on the sampler and model.

[rigid laurel]

25 for euler a, 30 for heun, 20 for everything else

[sly jay]

same prompt at 4, 6.5, and 8 cfg. All don't really listen, but I think 8 looks best and at least makes the croptop part black

[midnight shuttle]

I don't think you have restore faces turned on though.

[sly jay]

I do not currently, more worried about it generally not listening to me

[rigid laurel]

yeah, I don't have restore face on either

[sly jay]

also don't know how to use codeformers in comfy ui lmao

[rigid laurel]

(same tbh)

[midnight shuttle]

I don't think you can as far as I know.

[rigid laurel]

there's nodes for it

I just don't know how to use them right

[midnight shuttle]

It seems to be "listening" to you. But you might be expecting too much from a very new technology.

[rigid laurel]

they might not work with sdxl yet

[sly jay]

euler looks way better(6.5 cfg) but still doesn't listen. She is a cowboy, doesn't have a black croptop, and does not have freckles

[sly jay]

fair

[rigid laurel]

I mean when I do it it listens to me fine, so presumably there's a way to improve the technique

[peak dove]

Boys Gone Fishing style of Norman Rockwell

[sly jay]

generally though I kinda was hoping SDXL looked meh, thats what refinements for, but followed my prompt better. It seems the opposite is true

[midnight shuttle]

I see a few freckles. And I think you should not use the word "cowboy shot". Maybe say "full shot" or "full body" or just set resolution to rectangular.

[rigid laurel]

why does 'cowboy shot' never give me cowboy hats when I use it? 🤔

[midnight shuttle]

I've never even heard that term before.

[sly jay]

I live in america so it only does that

[rigid laurel]

I'm in america too

[sly jay]

me either I just copyed what the other person said lol

[rigid laurel]

fuck, I'm in cowboy country. CLIP shouldn't be influenced by region

[rigid laurel]

a cinematic term for a type of shot generally featuring the main character from mid-thigh to top of head, popularized by spaghetti westerns

[sly jay]

i think this is the key, at least for me. No idea how it was fine for you @rigid laurel

[midnight shuttle]

But was SDXL trained on that term? It seems rather unusual.

[rigid laurel]

you can use terms like "action shot" or "still from a {producer} movie" so it does seem to know cinematic terms

[sly jay]

this plus eulur work way better. Dpmpp_2m karrass is no longer very good for me

euler vs dpmpp

[midnight shuttle]

Same seed?

[sly jay]

think so?

[rigid laurel]

euler a, heun, and dpm sd 2m karras have been best for me

[midnight shuttle]

It is only a valid comparison if you use the same seed.

[sly jay]

yes same seed just double checked

940318754379450

;p

[rigid laurel]

changing the schedulers affects a lot. Different samplers do best with different schedulers. I got fuck all with dpmpp samplers before I realized I need to change it

[upper storm]

Do you know if you should just train it with the exact same repo and steps an what not as the "old" ones? Just use the joepenna repo and get going? There are probably some changes like maybe not crop images, use higher resolutions etc.?

[queen ether]

I ended up uninstalling automatic 1111 and then reinstalled it and I still have the same issues

[sly jay]

do you think euler a is better than euler regular?

[upper storm]

Really want to get my face in that model 😄

[rigid laurel]

that's what I was asking ._.

[sly jay]

yeah, but even with karrasss dpmpp_2m was my go to for 1.5 and now it is kinda ass

[rigid laurel]

yes. This has been the case literally every time for me.

[upper storm]

Ah sorry, didn't realize who I was replying to xD

[rigid laurel]

Maybe there's a way to make euler classic good, but I've never seen anyone do it.

[soft zealot]

Question::

Why after the modeli s loaded (in Comfy) do I always get the below "Missing" message ?

(nb its not new happened with 0.9 as well just forgot to ask)

making attention of type 'vanilla-xformers' with 512 in_channels
building MemoryEfficientAttnBlock with 512 in_channels...
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
making attention of type 'vanilla-xformers' with 512 in_channels
building MemoryEfficientAttnBlock with 512 in_channels...
missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'}

[nocturne dove]

It won't help you with things like the number of steps for refiner vs base, but you can read this if you want information. https://stable-diffusion-art.com/samplers/

[woeful patio]

why do generated images come out looking like shit?

[sly jay]

euler vs euler a

think euler regular looks better 🤷‍♀️

[rigid laurel]

man we've come a long way that people are saying this looks like shit >_>

[midnight shuttle]

Because SDXL is not mainly intended for generating images of people. It says right in the readme that it is not good at that.

[sly jay]

impaint or restore faces

composition and lighting is good besides that

adetailer is great, think it should still work for sdxl

[woeful patio]

true, I'll look into it thanks

[sly jay]

(adetailer automatically impaints and regens faces)

[soft zealot]

first question (as this isnt a OG output image)

What image size are you generating at?

Which UI are you using?

Whats your prompts?
How many Steps?

Sampler & Scheduler being used?

[bold wing]

When will there be an inpainting model?

[nocturne dove]

The trade-off for sdxl for me has been that if the base leaves too much noise than the refiner will modify things too much and things like likenesses (from lora) are ruined. But if the base leaves not enough noise, than the refiner will either leave things a little distorted because it needed another step or two, but if you give it another step than it cleans things up too much and they look too smooth and unrealistic.

[woeful patio]

using 1024x 1024 on 45 steps cfg on 12

[sly jay]

cfg 12 is quite high

[nocturne dove]

The actual sampler doesn't matter too much other than leaving the "right" amount of noise imo.

[midnight shuttle]

Soon I hope.

[soft zealot]

turn the cfg down , I generally use betweem 4.5 & 7

[sly jay]

4-8 is a good range, i generally start around 6.5 and adjust from there

[wet raven]

some models like it as low as 3

[boreal heart]

higher than 8 usually gives me a burnt cookie

[rigid laurel]

man with SDXL I've been able to go down to 2. Never was able to do that with SD1.5-2.1

[sly jay]

this is actually really impressive for 12 cfg

[rigid laurel]

good for artsy shit

[sly jay]

hmm I'll give that a shot sounds cool

[woeful patio]

Yeah I think that might have been part of the problem thanks for the help

[sly jay]

any idea what refiner cfg vs regular sdxl cfg should be? Both the same or mismatch?

[nimble heart]

XL kinda likes higher CFGs I found

in 1.5 I defaulted to like 6, in XL I use 8 and even 10 seems to do well

[sly jay]

I think XL is just crazy and likes whatever it feels like on a day to day lol

[rigid laurel]

If you manage to get photorealism with cfg2 I'll kiss you, but it's great if you want an impressionist oil painting of a dragon made of crystal falling out of the clouds

[nimble heart]

Why CGG 2 specifically?

cg

ffxdf

[woeful patio]

I found that in sd 1.5 restore faces wasn't necessary anymore, should we use it again in xl?

[nimble heart]

'ccf

cfg

[rigid laurel]

the lower it goes, the more creative it gets. But also the more incoherant. 2 is the lowest I've been able to go without getting pixel spaghetti.

[midnight shuttle]

Base 1.5 definitely needed restore faces. You might be thinking of a model trained on a lot of human photos.

[woeful patio]

yeah you're right, base 1.5 was unusable for me

[rigid laurel]

yeah 1.5 needed it, but deliberate1&2 didn't

[sly jay]

didn't quite get the crystals but it looks pretty neat

[rigid laurel]

yeah well it never listens on low CFG, so no surprise there 😛

[sly jay]

fuck if you asked me if this was ai IDK if i could tell you without knowing, looks super cool

[rigid laurel]

low CFG is basically only good for "show me what you can do" generations, but it's cool to see what SDXL comes up with. It can be real imaginative sometimes.

[sly jay]

yeah i am fucking around with it now, its super cool

[rigid laurel]

this about what you were looking for?

[marsh sapphire]

@rigid laurel switching to comfy worked, from 30s/it to 1s/it

[rigid laurel]

no idea how it's better, but it is! 🤷
glad it worked for you

[delicate grotto]

30x speed?

[rigid laurel]

uh

what is the question

[sly jay]

oh yeah that looks pretty good! I think I kinda got there with my gens above but I like the composition of this one generally

[delicate grotto]

...

[sly jay]

A1111 has issues with vram below 16gb i believe

[rigid laurel]

😂 right, I meant that I don't know why it's faster

[sly jay]

hopefully will be patched

[rigid laurel]

yeah

I like a1111 more, tbh

[sly jay]

same

cfg 2 kinda go crazy

[delicate grotto]

A1111 is much better in terms of usabilty, but comfy is soooopp much more efficient

[rigid laurel]

I've used a1111 with a card below 10GB VRAM, and with a card above 20. it kinda always fucky 😆

[sly jay]

I hope either comfy gets more usable or a1111 gets faster

cuz rn I am just coping a workflow and hoping for the best lol

[woeful patio]

do ya'll recommend using the refiner?

[sly jay]

yes

[sly jay]

I mean you don't need it, but i've personally found the gens look better

[delicate grotto]

Bit more than half the time pics look better with the refiner

[rigid laurel]

I've found using the base model with the "detail LORA" on civitai just almost as good as the refiner

and it's more memory-efficient

[sly jay]

If i am using a1111 I just use the base, if I am using comfy may as well use the refiner

[soft zealot]

would you use crude oil to fuel your car ?

Probably not, you would use a refined substance such as diesel or petrol (unless you have an EV)

[woeful patio]

it takes so long to switch models in a1111 tho 😦

[sly jay]

in a1111 don't bother

[soft zealot]

dont use A1111 then. Seemples

[sly jay]

lol

[woeful patio]

Comfy just looks like rocket science to me lol

[sly jay]

a1111 is great honestly, if you can actually get fast gens with it then dont bother with comfy

[rigid laurel]

comfybox puts an a1111-style gui on comfy

it's what I use

ain't here for rocket science

[sly jay]

I haven't heard of comfy box

[woeful patio]

ooh that sounds interesting, I'll take a look at that thanks

[sly jay]

Will give that a go

[dusty turtle]

XL into 1.5

[sly jay]

I love how weird cfg 2 can be

[soft zealot]

comfy will give you a better understanding of how the whole process works.

Im starting toformthe opinion that A1111 is fo rthe 12 o'clock flashers (and yes I started out on A1111 but I've seen the light, hallelujah!!!)

[woeful patio]

amen

[zealous viper]

I forgot just how much changing aspect ratio can effect the quality of a character / subject.

[sly jay]

wow that looks great

[woeful patio]

it just stretches stuff out when i up the height of the canvas

[dusty turtle]

comfy is so much easier to use for complex workflows if you know what the model is doing

[sly jay]

big if

[rigid laurel]

joe posted something before about how the model works best at certain specific aspect ratios

anyone have a screenshot of that?

wish they'd pin it

[heady vale]

#✨｜sdxl message

[rigid laurel]

🙏 bless

[zealous viper]

yeah its in the comfy basic workflow in an info box so i dont have to look it up everytime which is nice

[soft zealot]

it wasnt so much that it works better at specific ARs but rather what Pixel setting give what aspect ratios

Thsi is what I use in my workflow as a guideline to myself

[candid pond]

Hey guys! I've got 40 hours for a LORA training with 20 images yesterday

Will share my config and can you guys guide me how I can lower that

https://gist.github.com/pixelass/446c3cafde6bd81d4f2de01f8ce9b8ee

[woeful patio]

anyone know what this means?

[rigid laurel]

clever!

[dusty turtle]

Idk for people that only lightly use SD for fun might be pretty rough to get used to

[candid pond]

I've got a 3060 with 12GB's of Vram

[rigid laurel]

could mean a few things. what GPU you running?

[little gazelle]

Has anyone had any luck using Kohya on Google colab?

[woeful patio]

3070 8gb

[candid pond]

Waiting for LastBen to optimize it

He has one on Runpod but it doesnt have local pricing so it's pretty expensive in Türkiye

[little gazelle]

Is that LoRa training?

[sly jay]

can you autouse the refiner on comfybox?

[candid pond]

Yep, the one on Runpod is for LORA's

[little gazelle]

Interesting they've gone with doing RunPod first over Colab

[rigid laurel]

odd error to have then.
The easiest way is to just launch ComfyUI in the background and then run comfybox without the cors-header flag

[candid pond]

For sure. I have no idea why he did the RunPod first.

[rigid laurel]

You either have to set that up yourself or download one of the workflows on Civitai to have it do it automatically
but you can, yes

[candid pond]

Please someone for the love of god help me optimize Kohya 😄

[rigid laurel]

not my area, sorry.

[vocal rapids]

high rex. fix 2046*2046 , pretty decent result

[rigid laurel]

most people who know enough about optimization to help you are probably waist-deep in an ide, not on discord :p

[rigid laurel]

that is clean. what cfg you use for that?

[sly jay]

download workflow from citvai? searching comfybox comes up with nothing

[vocal rapids]

CFG 8

sampler DPMPP_SDE

[rigid laurel]

tyvm

[sly jay]

also, comfybox is amazing I feel so much more at home lol

[finite moth]

if someone's workflow has no batch size option, is there a way to add one yourself?

[uncut steeple]

sure

[rigid laurel]

yeah, they just called it a "comfy" workflow, not a "comfybox" workflow. Trying to help you find it 😩

[woeful patio]

How do I add the refiner in the comfybox ui?

[little gazelle]

Looking at working on a config for today to see if I can get it working in Colab. Managed to get an output yesterday but it had NaNs when trying to test out the LoRa

[uncut steeple]

In the empty latent image node

[sly jay]

ahhhh okay

[candid pond]

Oh

You can use the fixed vae to fix it

Did you use fp16

[little gazelle]

Yep I did

[nimble heart]

yea just change the "batch size" on the latent image

[rigid laurel]

If you use Searge's SDXL nodes, there's an SDXL node that you can just plug the refiner into

[finite moth]

there's no empty latent image node but im investigating how to add one

[candid pond]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

Try this maybe, it's optimized for fp16 training and inference

[nimble heart]

double click type "latent" add node

[sly jay]

comfy has the workflows but there as pngs not as actually workflow files :/

[rigid laurel]

imma just upload mine. uno momento. faster than searching.

[finite moth]

POG

i legit forgot u can just search nodes

thx

[candid pond]

Can you also please share your config so I can try it locally?

I've got 70s/it yesterday with my settings

[nimble heart]

what kinda cursed worfklow doesn't have a latent image by default in the first place?

[rigid laurel]

it probably does

just hidden in a subgraph

[nimble heart]

you can't really hide nodes in comfy

unless you use custome ones

[rigid laurel]

subgraphs

[nimble heart]

yea idk what a subgraph is in comfyui context

[finite moth]

yeah i think its hidden somewhere

[rigid laurel]

you can select a group of nodes and turn them into a single node to save space

then double click that node to look 'inside' it and edit the contents

[nimble heart]

?

where tf is that

[soft zealot]

nested nodes on git

[finite moth]

i am using some custom nodes so that probably explains why i cant find the empty latent image node

[nimble heart]

Oh it's an addon?

lame

[rigid laurel]

is it?

huh

subgraphs

because organization is sexy

[woeful patio]

is there restore faces within comfy?

[sly jay]

in theory yes but i have no idea where to do it or how to find it

[nimble heart]

why make a subgraph when you can write a custom node with literally every option in it

[soft zealot]

this is bothe the advantage and disadvantage of comfy, its flexibility and customisation.

Great if you're a natural tinkerer not so great if you're a 12 oclock flasher

[rigid laurel]

this is faster
plus you can copy nodes in and out of the subgraph, so you can decide to move things in and out without having to translate them to or from code

[nimble heart]

yes, but

fat node

[rigid laurel]

@sly jay @woeful patio On today's episode of "it's ugly but it works", here's a simple workflow that uses SDXL and lets you just stick in the refiner easily

[rigid laurel]

png for anyone who thinks that json files are scary

[little gazelle]

Will do, have you also been getting this error Kohya WARNING The following values were not passed to and then it just freezes?

[soft zealot]

other reasons why images posted here may not have embeedded workflow

1. the user has copied from a browser window rather than uploaded the generated image from the outputs folder directly.
2. the user has opted not to embed the workflow in the saved image (WAS Nodes suite option)

[candid pond]

Didnt have that

Sorry

[sly jay]

thanks! getting this and it refuses to gen :/

[sly jay]

I am getting unreal engine blueprint flashbacks

[rigid laurel]

it's pretty similar :/

[rigid laurel]

I downloaded Searge's extensions. Nodes in that workflow are from his stuff.

https://civitai.com/models/111463/searge-sdxl-nodes-and-workflow-for-comfyui