✨｜sdxl

[vital ermine]

I am still having an issue knowing which epoch is the best. How does one tell?

[ivory blaze]

batman can tell you

he likes them all

[ivory blaze]

Is more, well done and full imo

[ivory blaze]

is much cuter

[heady vale]

the one thats more versatile

[ivory blaze]

so depends on your want

[ivory blaze]

that is batman!

epoch batman

so batman, how do you like the other epoch?

"I am BATMAN I LIKE THEM BOTH!"

[vital ermine]

Yeah, epocj 10 or 15 though I did do 17

[vital ermine]

here is batman 17 epochs

[tepid surge]

so cute!

[hardy cipher]

is epochalyptic a word?

[ivory blaze]

I wish he had higher contrast

on the latter,

but they're both decent

[vital ermine]

I am not using refiner at all

[tepid surge]

epileptic? @hardy cipher

[tepid surge]

like seizures

[vital ermine]

I added highly detailed and high contrast

[ivory blaze]

His contrast coloration is correct, but he looking "soft'

[ivory blaze]

now he's sick

[vital ermine]

yeah, I am not sure how to fix that

[ivory blaze]

in a positive manner of ""sick"

[tepid surge]

dynamic lighting's good too for like rim lighting and good shadows

[vital ermine]

oh, yeah

[ivory blaze]

he washed his cape with his whites without colorsafe...

[ivory blaze]

is good

[vital ermine]

ugh, dynamic lighting

[ivory blaze]

except for the 27 pack he has

ouch

[tepid surge]

i also started throwing (shadows) on too if i want a darker image but still have to test it more

[ivory blaze]

my eyes

he should look like this,

[vital ermine]

oops, I had a typo

it wasn't using my locon

Now that looks a bit bad ass to me

[vital ermine]

epoch 15

[ivory blaze]

he looking kinda like he wears a fake muscle inflatable

[vital ermine]

epoch 10

[tepid surge]

"empty, mindless, blank..."

[tepid surge]

mantid's pull through

[vital ermine]

Reason I don't release anything is because I can't tell which one

[ivory blaze]

they all good

[vital ermine]

well

[half ivy]

everytime i blink theres something new o0\

[vital ermine]

WOW

[tepid surge]

fr fr, I've been eatting all the stuffs catching up being away from image gen for like a year, now i here:

[vital ermine]

0.8 weight and this is 1.0

[tepid surge]

wait what was that guys name that makes those shitty action movies xD

[crisp owl]

James Cameron

[ivory blaze]

1 year?

[tepid surge]

idk about a year but it's def been a while, ie several months

[ivory blaze]

oh was gonna say, last time you were here, we thought this was amazing

oops

[ivory blaze]

hold on

wtrf

okay... @tepid surge Last time you were here, we all thought this was amazing

[tepid surge]

oh yee the spooky dark, i was all like "okay my binge is done" and then i'm shoved further into the stuff casue there's no other lights on cx

it is hard to argue with this:

[ivory blaze]

yes, thisi is . good

this is sharkbucks coffee

may favorite flav

[ivory blaze]

i need to try with SDXL

[tepid surge]

snoop dogg's such an icon cx

[vital ermine]

yep

[ivory blaze]

and then of course, the year old aspic and body pillow we all know and love

AI image gen has come a long way since these days

[ivory blaze]

yesterday we had that, today we have this:

[tepid surge]

the gpu forges are hot

https://tenor.com/view/forge-blacksmiths-creating-a-weapon-gif-16530724

[ivory blaze]

we don't need no stinking sharks,

[tepid surge]

https://tenor.com/view/ocean-shark-gif-20676612

nibblies xnx

[shy kelp]

lol

[tepid surge]

when your fish barks for the first time cx

[shy kelp]

lol

[autumn forum]

ayyyy i managed to get ip-adapt and ReVision working together all into one image with prompting as well.. now to add other things

[tepid surge]

new series?

cx

[shy kelp]

how long gcolab provides per day?

i mean gpu thing

[vital ermine]

2h if you use it every day

[ivory blaze]

not grimace from mcdonald's but thank you

[shy kelp]

how do you achieve this art style?

[vital ermine]

A lora I trained

[shy kelp]

wow

[tepid surge]

when you trip so hard the mountain starts attacking cx

[shy kelp]

amazing

[crisp owl]

And Grimace starts to look like this
#✨｜sdxl message

[ivory blaze]

he has a skin condition, be nice

[tepid surge]

i've got an idea!

[tepid surge]

okay here is goes....

[tepid surge]

"say my name"

[shy kelp]

amputated

[tepid surge]

Mr. Flash... you killed my dog

[tepid surge]

how do i get this thing to work

[shy kelp]

are you sure bout that?

[tepid surge]

absolutely

[pure crystal]

sushi date

[dense chasm]

one of my short video visitors number at one of the top platform except tiktok in China is up to almost 300k,originally i wanna to increase the other AI videos page view traffic,it's a surprise

[tepid surge]

man emerges from vending machine

[pure crystal]

the flying kick in the background is great lol ^

[vital ermine]

🙂

[tepid surge]

wow that's a lotta views cx

[hard fractal]

https://tenor.com/view/ninjala-jane-hacker-hacking-computer-gif-20337624

[crisp owl]

https://tenor.com/view/leet-bbtf-safe-moon-blockbusters-tech-gif-27506827

[shy kelp]

kiss my hand

[vital ermine]

My other lora

[tepid surge]

can do cap

[shy kelp]

what did he got??? a dog?

[tepid surge]

dont you question my dog owning habits! i'm gonna!... kiss your hand!

---

WHAAAAAAAAT IS HAPPPPENING xDDDDD

[glad grove]

new steven seagal movie where he fights king kong

[tepid surge]

okay, here we go... you and me sd xl

[glad grove]

u cant win

[shy kelp]

HAHAAHAHHA

[tepid surge]

I tried as hard as I could not to flash...

[spring fulcrum]

LEET ftw

[tepid surge]

(i've shoe horned myself into this cx) You'll have to take this as a hand kiss barry...

[shy kelp]

HAHAHAHA

[hardy cipher]

https://tenor.com/view/cats-surprised-shocked-oh-no-gif-22538739

[shy kelp]

real

[tepid surge]

https://tenor.com/view/thats-impossible-thor-impossible-ragnarok-impossible-impossible-not-possible-gif-19896965

Flash! seasson 8 cx

[shy kelp]

[shy kelp]

keanu face crushed XD

[tepid surge]

cool art style xox

[raw prairie]

Sus

[vital ermine]

Not judging, lol

[dense chasm]

https://huggingface.co/datasets/HuggingFaceM4/OBELICS huggingface bust a gut

Content warnings
Despite our efforts in filtering, OBELICS contains a small proportion of documents that are not suitable for all audiences. For instance, while navigating the interactive map, you might find the cluster named "Sex" which predominantly contains descriptions of pornographic movies along with pornographic images. Other clusters would contain advertising for sex workers or reports of violent shootings. In our experience, these documents represent a small proportion of all the documents.

[tepid surge]

---

i wish i could be 10,000 years old

[crisp owl]

same, just..fix my back and knees please

[tepid surge]

"and i know where i might be told" 🎶

[tepid surge]

i'm 100% sure there are aliens out there that make art like this admiring us

[tepid surge]

I just realized your username cx

[vagrant acorn]

I have a rtx 2060 super 8gb, I tried using sdxl on my pc, and the images have infinite load. Do I have some wrong configuration or my video card is not good?

[crisp owl]

I run the same card, it works fine

[tepid surge]

try to find how to set the --medram flag for your ui (though now that i think about it mine still runs, though slowly, when it spills over to the shared vram-ram shared memory)

[vagrant acorn]

I downloaded the webui.zip, installed it, and just put the base model and the refiner,

[crisp owl]

War, you using A1111 or COmfy?

[vagrant acorn]

--xformers too ?

a1111

[crisp owl]

Yeah you'll want --medvram and either --xformers or --opt-sdp-attention

[tepid surge]

stayvoon seeghoul

[vagrant acorn]

@astras @crisp owl ok thanks

[tepid surge]

"i may have given the wrong info but i'm happy to help"

"as a wise man once said..."

[shy kelp]

guest this animal (wrong answer only)

[tepid surge]

the letter H xwx

[crisp owl]

Now I might watch that live action

[shy kelp]

watch on the right channel

[ivory blaze]

cutests grape tictac ever

[vagrant acorn]

Now I'm managing to generate the images, a doubt, doesn't --medvram decrease the image quality?

[tepid surge]

nah, idk exactly but I think it moves some of the data to computer memory so it'll take longer to generate

there's a --lowvram flag too if there's ever a need, and you can get confyUI to run sdxl on a 3gb card with that flag if you're that tight

[vagrant acorn]

I tested creating the image in my template / in google colab

the quality looks the same

[crisp owl]

Medvram does not change quality whatsoever

[crisp owl]

You can even set Comfyui to do CPU only if you really want to torture yourself

[crisp owl]

If you always hit that when it hits that vae decoder/encoder, you can just manually put in a tiled vae there to save you a few seconds of it detecting for you.
That's what I've done

[shy kelp]

oooo i seee

[vagrant acorn]

is there any better sampling method in sdxl?

[crisp owl]

kinda comes to personal preference and everyone will have their own they say works best for them.

I personally still like DPMpp_2m with Karras scheduler

Some have gone back to using DDIM a lot more

[tepid surge]

10 mins later xD

[crisp owl]

Or more lol

[tepid surge]

at that point i'd go and squeeze out all the free daily credits from generation website cx

[crisp owl]

I'm usually doing work, or training, so I just click a few batches to process and then get to work, then examine my beauties (or grotesque creations that still need more work)

[proper cargo]

does anyone know why when I want to download an image I get it in webb format and not in png format?

[tepid surge]

the preview maybe?

should be in an output folder if automatic1111

or an addon

[tepid surge]

moood

[soft zealot]

and a lot deends on the effect you're after as well.

I mean I switch between different one sa loot although I do like Euler_a

[crisp owl]

Yeah you do use that one a lot haha

[tepid surge]

man AITemplate go zoom

10 sec off my generations (was 27-30 now ~16 for some of the samplers, eur-a went 16->10)

[crisp owl]

hoping and praying for 2k series support soon

[tepid surge]

you could just try it no? i was reading it was only tested on the high up cards

works on 3060

[crisp owl]

Oh it poops on any attempt, not supported yet, needs to be compiled I was told

[tepid surge]

oooh i see

[tepid surge]

you missed all my seagal gens cx

[glad grove]

sensei 🙇‍♂️

[shy kelp]

bro is thinking

[uncut fiber]

Comixwave films <- this tag is worth to play with 🙂

[tepid surge]

i feel like i just watched a movie and was changed by it really beautifully xwx

[shy kelp]

my guy flying

[tepid maple]

on multiple people scenes sdxl tends to make lots of twins why ?

it feels does this more than 1.5

[tepid surge]

"hmm yes, hmm, hmm yes, business business"

[tepid maple]

haha yep

yep business making everyone wear suit

another annoying thing

it is very powerful word

tooo powerfull

[tepid surge]

how to fit in at the office:
wear a suit

[tepid maple]

it has weight more than it shoud lbe dont know

weird

[peak dove]

Using Underwater LoRA

What does the sdxl_ offset_example_LoRA do at all?

[tepid maple]

better

I dont wanna go weird with prompt 😄

it puts men behind whatever reason

...

[glad grove]

real

[tepid maple]

he is not sitting

cant be real

[thorny frost]

Is that of any use for simple mortals as us?

[glad grove]

even more real

[shy kelp]

how do you get old like pic

[dense chasm]

unique and have your own trail,whatever,being existence,from eastern idiom,life journey is memorable

[peak dove]

Minimum Spec B4 setup - Brain The Size Of THREE Planets!!! - 😄

[shy kelp]

close up

[uncut fiber]

are you using standard resolution? Anyway twins will be twins.

[tepid surge]

"you know what you want"

😭 😂

[molten gull]

why did the logo change ? do we have a new version ?

[strange mist]

it's sd's birthday 🥳

[molten gull]

ah, party time 🙂

it's been a long year 🙂

[heady vale]

you mean the 1? Its the Ist anniversary

crazy how much has happened in just a year

everything moves real fast these days

[molten gull]

thats what i meant, doesnt feel like SD beeing around for just one year

[indigo carbon]

resolved batch size for AIT, just pull requested for changes

[heady vale]

probably wont be that long where images on the internet are generated on your PC as you open them and not downloaded

[molten gull]

i wonder what's going to happen with art

[rustic garnet]

art didn't die due to photography, so it won't die due to AI 🤷‍♂️

[molten gull]

i dont know if you can compare that

[rustic garnet]

I'm sure you can

[heady vale]

yeah same with printers taking over painters and people who hand wrote things

[peak dove]

"Who will need the internet when each Mac/PC becomes its own AI Factory shaping our world, our art, our music, our weather, our news, our TV and Radio etc etc right on our very own desktops?!" Yuval Noah Harari, where are you now?!

[shy kelp]

which one

[peak dove]

It'll promote net-neutrality, as it releases pressure on overloading bandwidth! It is an audacity untold that as I pay for increased speed from my ISP, it facilitates even more and more advertising bandwidth to our machines "which we are paying for!!!!!"

[heady vale]

thats where ad blockers are heaven sent

[tepid surge]

p-yoyg

[shy kelp]

i think the best samplers for me is ddim

[peak dove]

ComfyUI SDXL using 512x512 embeddings (made using A1111)

[west breach]

made this 4k image on my 3060 6gb mobile in 252 seconds (4 minutes) with my new custom tile sampler

[native knot]

Nice. Now do a human.

[native knot]

Looks good. I don't see tile borders.

[west breach]

Yeah I'm pretty happy with it. Took lots of attempts to get rid of the seams 😄

[shy eagle]

yo what are some good XL models that can do realism, and celebrities

[thorny frost]

Pretty nice, is it available somewhere?

[icy brook]

Aether Ghost v1.1 coming up

Even better full body transparency

[west breach]

sure 🙂 you can get it through mikey nodes in the manager or at github https://github.com/bash-j/mikey_nodes

[shy eagle]

🗣️

[thorny frost]

Great. Ty. Will try it asap. I love big crazy compositions. Was trying to do an image to image from the garden of earthling delights by the Bosch from the huge 8000x6000 pic a available in Wikipedia

[zinc cargo]

hey everyone 🙂
did you see this error in comfy before?

[indigo carbon]

published it =]
https://civitai.com/models/133818?modelVersionId=147337

[uncut steeple]

thats a lot of steps

Also somehow I dont see any difference running AIT on my 4090

[indigo carbon]

tf, do you have the node installed correctly? other 4090 users like @upbeat summit report over double the speed..

if you have a 4090 you should be getting ~14 seconds on average with that workflow. if that's not what you're getting- it's an issue on your end

AI generated food makes me sad, I'm still hungry =[

[icy brook]

Updated Aether Ghost to 1.1 rn for even better transparency!
https://civitai.com/models/133412/aether-ghost-lora-for-sdxl

[peak dove]

ComfyUI SDXL using several 512x512 embeddings

[eternal fog]

It's not an error it's a warning and Comfy has said it can be safely ignored.

[thorny frost]

Get " TypeError: 'Image' object is not subscriptable" and not upscaled image

it is also very slow and takes all the memory of my 3090ti even if upscale is at 1

[coral bolt]

Draw robot

[wet nacelle]

hey bro. Us and the Holy Saints don't like that type of language here."

[uncut fiber]

@indigo carbon for me like AITemplate stops working today...
Those are my speeds.
1it/s Most time when AITemplate but doesnt work properly.
2it/s No AITemplate
3it/s When it works happy generation 🙂

[glad grove]

the other bro's face is meltin 😬

[wet nacelle]

the sheer amount of psychic power does that to a sinner.

[indigo carbon]

odd. try to reinstall the node, you might have an outdated version

[cyan thicket]

SD's chart is ugly now~

[wet nacelle]

?

[uncut fiber]

i made already clear install and new install of that nodes. Dont know what is going now. @indigo carbon
In your workflow i even have 2,5s/it 😦 But nice results 👍

[wet nacelle]

@uncut fiberAre you on the right Comfu UI version? Are you using standalone or portable?

[indigo carbon]

odd. what CUDA and cuDNN version do you have?

[indigo carbon]

^

[wet nacelle]

I had to switch over to portable for all of the new stuff recently

[uncut fiber]

Just saying it was working. I am on 7zip version wait a minute

[wet nacelle]

my ssd storage hates me though

[wet nacelle]

?. a 7zip version?

[indigo carbon]

I think the portable version might be using a different VENV layout

[wet nacelle]

Would that be bad here?

[uncut fiber]

\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable

[indigo carbon]

the standalone might use torch 2.0, AIT wants torch 2.0.1

[wet nacelle]

oh this one you mean?

[uncut fiber]

you know, until now it works well.
Yes

[wet nacelle]

Yeah that's the one I'm using too.

[uncut fiber]

i think not sure is there xformers 0.21?

[wet nacelle]

Alright. My help is done from here. I can not help.

[uncut fiber]

@wet nacelle thank you!

[wet nacelle]

When you open the CMD you'll see at the top "xformers"

[indigo carbon]

and for you it works correctly, right?

[wet nacelle]

It will let you know if they are installed and what version too

[uncut fiber]

xformers version: 0.0.21

[wet nacelle]

I mean, just stuff works. I don't know what you guys are doing here right now.

[indigo carbon]

the AIT workflow

[wet nacelle]

Oh wait. Your Workflow? Yes it worked for me.

[indigo carbon]

@wet nacelle do you have Xformers 0.20 or 0.21?

[heady vale]

worked for me OK

[wet nacelle]

Let me check.

[indigo carbon]

I think .21 isn't compatible

[wet nacelle]

Well it worked for me

[indigo carbon]

then .20?

I'm confused.. idk what's wrong with @uncut fiber 's setup..

what pytorch version are you using @uncut fiber ?

[uncut fiber]

o.k.
where to find it? Everything else is working

[indigo carbon]

python -c "import torch; print(torch.__version__)"

when venv active

[uncut fiber]

ok

1.12.1+cu113
seems too low numbers
What about update with python things?

[indigo carbon]

jesus

that's so outdated

hold up, I'll pull up the update command

[uncut fiber]

o.k. i reinstaled it, so i will update both

[indigo carbon]

wait, I'll give you the best one

[uncut fiber]

o.k. thank you. waiting

[indigo carbon]

@uncut fiber run pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 --force-reinstall while ComfyUI's VENV is active

[uncut fiber]

it freeze 🙂 going to check version

[wet nacelle]

@indigo carbonHey do you think it would be worth it to see if I can get in touch with the hyper realism model creators? My clip_vision workflow is doing great at outputting it plus coherence.

[indigo carbon]

on what screen? let it cook if it says it's installing.

[uncut fiber]

just installed it throught update and i have 1.18 but that report is from somewhere else. if only i know

[indigo carbon]

idk about CLIP vision.. it's supposed to be able to merge images, I tried to merge an image of food and an image of a car, it made a bear.. far from being seemless

[wet nacelle]

I don't use it for blending. I use it for styles. I grab photos because it's a literal photo. It then reads that photo and adds hyper realism to the output.

Blending sucks. What I'm saying is that you would train a model off of the outputs to make a hyper realism model.

[indigo carbon]

probably VENV. you need to run that command on a terminal with ComfyUI's VENV active. otherwise ComfyUI won't see the new version you just installed using the command

[indigo carbon]

hyper realism can be achieved by just changing the prompt. I don't see why CLIP vision is necessary atm

[shy kelp]

what kind of image do you use

[indigo carbon]

wait is that Jerma

[wet nacelle]

yeah

[indigo carbon]

@uncut fiber did the instruction I gave you resolve the issue?

[fierce compass]

sdxl. How do you guys solve hands (any easy solutions, with ComfyUI?)

[peak dove]

ComfyUI SDXL with one 512 x 512 embedding

[upbeat summit]

TI embed?

[peak dove]

TI Embeddings which I made using A1111

[indigo carbon]

didn't know SDXL already has textural inversions LOL

[upbeat summit]

nice - used only in clip_g?

[thorny frost]

I dont know why all workflows give me error in Comfy

[peak dove]

Yes, I am using the format embedding:EmbedName.pt in Linguistic Positive G

[forest geode]

In case somebody else face the same problem : I don't know if it is related but I updated sd.next + extension comfyui by accident this morning form the Automatic UI, since then comfyUI won't launch. I'm not the only one in this situation, I spent 3h trying to solve this but I'm not a python expert. I posted the last message on this but report, regarding the "ModuleNotFoundError: No module named 'lib_comfyui'" error (source : https://github.com/ModelSurge/sd-webui-comfyui/issues/149 )

[thorny frost]

Error occurred when executing KSampler:

'AITemplateModelWrapper' object has no attribute 'num_timesteps'

File "K:\ComfyUI_windows_portable\ComfyUI\execution.py", line 151, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
File "K:\ComfyUI_windows_portable\ComfyUI\execution.py", line 81, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
File "K:\ComfyUI_windows_portable\ComfyUI\execution.py", line 74, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
File "K:\ComfyUI_windows_portable\ComfyUI\nodes.py", line 1206, in sample
return common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=denoise)
File "K:\ComfyUI_windows_portable\ComfyUI\custom_nodes\AIT\AITemplate\AITemplate.py", line 175, in common_ksampler
samples = comfy.sample.sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image,
File "K:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Impact-Pack\modules\impact\hacky.py", line 9, in informative_sample
return original_sample(*args, **kwargs)
File "K:\ComfyUI_windows_portable\ComfyUI\custom_nodes\AIT\AITemplate\AITemplate.py", line 298, in sample
sampler.model_denoise = comfy.samplers.CFGNoisePredictor(model_wrapper)
File "K:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_smZNodes\smZNodes.py", line 705, in init
self.inner_model.num_timesteps = model.num_timesteps
File "K:\ComfyUI_windows_portable\python_embeded\lib\site-packages\torch\nn\modules\module.py", line 1614, in getattr
raise AttributeError("'{}' object has no attribute '{}'".format(

[upbeat summit]

very cool. I tried SD 1.5 TI embeds before on clip_g but since SDXL is obviously a different latent space it didn't really work

[peak dove]

It is hit and miss - some work, some not ...

[fierce compass]

selfie photo of a cute age 21 girl in Mykonos, showing a victory hand sign, rainbow hair color

[indigo carbon]

you aren't using the manager's preview. I addressed that it is required in the workflow's description

[uncut fiber]

i deleted VENV from A1111 and it seems is working now.

[indigo carbon]

interesting. does ComfyUI's VENV have torch2.0.1+cu118 now?

[uncut fiber]

not sure will have once run again and it will install it

[shy kelp]

i search for an empty background??

[soft zealot]

well the hair came out nicely 🙂

[fierce compass]

Yours came out a bit better, but there's still the middle finger missing. I was testing the "sdxl comfyui ultimate" workflow.

[soft zealot]

she has the same finger missing in the next one as well lol

wonder if I will get 3 out of 3 lol

[fierce compass]

ComfyUI Manager can't find the "Seed With Text" node, I wonder where that comes from

[native knot]

Just because why not...

[soft zealot]

look in the credits & notes of my workflow (to the RH side off screen) ithe link is in there

[native knot]

Or also, just get ComfyUI Manager and use it to install missing nodes anytime you run into them.

[thorny frost]

now it works ty. I had an old version of the manager and there was not preview option. Anyway there was something else. Had to make a fresh install of comfyui, install the new manager and download the missing nodes. Only then worked

[soft zealot]

got all her fingers this time , just a bit squiffy

[fierce compass]

I've used the Manager, it shows no missing nodes, but when reloading the workflow, it ComfyUI complains about missing the missing node

Maybe I need to fetch updates

[shy kelp]

how do you have that vintage looks

[soft zealot]

this section for Credits & notes

[native knot]

Helliid Slaw Süll

[fierce compass]

scary teeth in the closeup, same with her nails

[native knot]

Close.

Look out fellas...she's packing.

[thorny frost]

does the manager update all custom nodes when you click update?

[native knot]

No

[thorny frost]

😦

[native knot]

You click on Fetch Updates, then go into the Install Custom Nodes and Alternativesd of A1111 to update the things you need/want to update.

[thorny frost]

ok ty

[native knot]

Anyone have a workflow or an image with workflow embedded that uses IPAdapter to merge 2 images that they can share? I have the example using text, but I'm not sure the best method to do that with the 2 images.

[native knot]

Those sausage fingers!

[soft zealot]

added the rainbow hair to an earlier prompt 🙂

[native knot]

The earlier ones were already nice...but that's... delicious.

Can you run that at 1344x768?

[soft zealot]

just adds a little extra 🙂

[native knot]

Want to see that in the other aspect ratio.

[fierce compass]

nails showing on fingers

[soft zealot]

so a landscape rather than portrait

[shy kelp]

better

[native knot]

Sorry...meant the other way around, 768x1344

[fierce compass]

still some traces of nails 🙂 a stormtrooper never shows fingerprints, right?

[soft zealot]

added to Q

[shy kelp]

hahha

[native knot]

Times were tough for her...had to pick up a job with the Empire.

[ionic dragon]

@soft zealot how many secs does it take to generate a hires pic using your workflow?

lol

[native knot]

lol

[soft zealot]

with my 1080ti ? around 160-180 seconds (ish)

[soft zealot]

oooh eerr matron

[ionic dragon]

just 180 seconds?
is it the same compared to sytan's workflow?

[soft zealot]

well that got interpreted slightly differently (used same seed as previous)

[native knot]

Not hating it.

[soft zealot]

lol must be the first time anyone has ever said "just 180 seconds" and yes on my 1080ti the time is about the same

[native knot]

Right hand's knuckles have a bad look, but it's good.

Like...the jewelry is on the knuckle instead of between.

But otherwise, pretty nice.

[soft zealot]

@native knot 768x1280 is nicer

just waiting for the ping

different seed mind you gos i forgot to fix it

[vital wraith]

I find I don’t move far enough from 1152x896

[soft zealot]

I like these ones where the hands are just out of shiot lol

[shy kelp]

last memories

[native knot]

oh man... 🤣

[native knot]

I feel like some further modifications on the prompt could take this to over 9000

[soft zealot]

I may take a look later

[soft zealot]

presume you would like her made older than inher 40's lol

[native knot]

Thinking some additional accessories, a better background, that time of thing...

[soft zealot]

I like the plain background :p

dang its close

[thorny frost]

wow

[native knot]

This would be a bit better if the skin on the body was closer to the skin on the face, but still, pretty awesome.

[thorny frost]

was trying that precisely. It was worse at first. Upscaled added some noise and reremndered again to get some more skin texture there.

[soft zealot]

just changed "photo of submissive mature woman in her 40's " to "photo of submissive mature woman in her 20's "

[native knot]

I don't see much change...

There are snaps on the hip of the dress now, which is cool.

also, a bit of something lining the arm opening.

[soft zealot]

plus she looks a little younger lol

[native knot]

Maybe just a touch, but it's not tremendously noticeable. She ages well, aparently!

[native knot]

You're really digging that icecream sandwich thing, aren't you? 😄 😄

[indigo carbon]

idk man I'm feeling creative but I don't have ideas atm

[native knot]

Change the prompt up to include Pop Tarts.

See if you can get them to render right.

[soft zealot]

CHanged "photo of submissive mature woman in her 20's " to "photo of a girl in her 20's"

[native knot]

That definitely takes her age down too far from the original older woman. None of them are bad and I love what it did in each prompt with her dress, but I liked the facial structure of the original one with the multicolored hair the best.

[soft zealot]

i just like seeing what small shangesd make t the final picture using thesame seed

[indigo carbon]

civitai down LOL

what are these images

[native knot]

lol...the one I got:

That's great.

[soft zealot]

so instead of starting with "A fashion magazine cover ." this is "A glamour photoshoot for a fashion magazine cover ."

[native knot]

That's a tasty looking Pop Tart.

The images of the sweets are nice too.
🥁 🥁 💥

Civit's back up

[soft zealot]

its like a whores drawers

@native knot yup prompting for a more mature woman does look better 🙂

[native knot]

Add the terms "glossy bismuth lips"

[soft zealot]

wil try next, just popped her in a gothic church"

[native knot]

That's fitting for a setting

[soft zealot]

losdt the rainbow hair though 😦

[noble shoal]

Scraped a lot of the promt and added "Bear"

[soft zealot]

@native knot Gothic Church + glossy bismouth lips

Lets drop the Gothic Church a minute

[native knot]

Careful with the spelling on "bismuth".

[strong field]

masters of inpainting, i beckon, I have spent all morning trying to get a good workflow and it all looks like crap, anyone have any good examples?

[ionic dragon]

im on my laptop, so its slower than desktop 3080

[soft zealot]

@native knot nope going to stop at that I think

Lets do some other seeds out of curiousity

[native knot]

Nice pose there, but lost the hip angle, too. 😦

[soft zealot]

I still dont get the whole obession with everything needing to be done as quick as possible, just go watch a YT video whilst its working in the background

[native knot]

Mostly agree...I do think there's a point where it's unbearable, but like, who cares if it's a minute or so for a decent render. I also wonder how many people leave deeper sampling steps on before honing in on something worth rendering out higher, causing their own issues.

[uncut steeple]

I just downloaded the extension with the comfy manager

[soft zealot]

IMHO anything sub 5 minutes for a 2048x2048 (ish) sized generation is perfectly fine

[uncut steeple]

Maybe im missing a step

[uncut steeple]

got down to 50s for 2048²

[native knot]

2048^2 sure, but I wouldn't want to sit around for that long on something 1024^2 with SDXL's current performance.

[indigo carbon]

pytorch version?

[soft zealot]

well whoopdy doo for you.

(and thats my point, why the bloody obsession)

[uncut steeple]

dunno, any way to see it in comfy?

[soft zealot]

that frustrating moment when you rrealize you've queued up 10 generationd and forgtotten to unfreeze the seed

[ionic dragon]

no i am fine, i was justing checking with other cards so I can verify if my gpu is good and nothing's wrong, 2 days ago i was generating within 130sec

[indigo carbon]

@uncut steeple what pytorch version does your VENV have? AIT requires 2.0.1

[uncut steeple]

Its all up to date, thats for sure

[indigo carbon]

open venv in terminal and run: python -c "import torch; print(torch.__version__)"

[soft zealot]

what GPU does the laptop have?

[ionic dragon]

3080

[native knot]

Premiere night for District 9?

lol

[soft zealot]

I would have thought that would be faster than my 1080ti TBH, as in 2-3 times faster howver YMMV

[thorny frost]

wow. how did you do it?

[noble shoal]

Magic

[thorny frost]

controlnet

[ionic dragon]

yeah, it should be
thats what is concerning me

i'll due a 3d test

[indigo carbon]

also, @uncut steeple, CUDA 11.8 is best to get as much speed boost as possible with AIT. also if you haven't installed it, you need the latest cuDNN

[noble shoal]

nightvisionXL is kinda good at hands. Kinda

[thorny frost]

downloading...

[soft zealot]

I like all of his models, Dyna, Night & Proto Visions

[noble shoal]

For what i have seen now, the model is pretty usable for realistic stuff

[soft zealot]

these rainbow haired woman Ive been doing have been Dynavision

[indigo carbon]

idk, I tested most models, best one imo is copax_colorfulV2

[native knot]

A bit off-topic, but does anyone else feel like using LORAs is almost a bit like cheating sometimes? I personally like the challenge of trying to get a certain look without them. Not that I don't use them at times, but just prefer to try to squeeze the look I want out of the model.

[thorny frost]

better than newer versions?

[indigo carbon]

way better, the colorful version is so much better than the others imo

[soft zealot]

I personally rarely if ever use LORas although I do admit to using Styles a lot 🙂

[native knot]

But essentially, styles are still just prompts, right?

[soft zealot]

yup

[thorny frost]

few need for loras with XL models. the famous Kappa_Neuro understood it quickly

[native knot]

LORAs come in handy when you're doing something extreme or if you really need to nail down a specific person's look to a T.

But I just try not to use them.

[indigo carbon]

also, Xformers 0.0.20 is faster than 0.0.21, about 10% faster.

well, at least with AIT

[uncut steeple]

I'll check tomorrow, I'm on my way home now

[spark sinew]

when did you make a horror and colossal theme

[soft zealot]

dont want to overspam the same sort of thing repeatdly but here we go @native knot a few more variations

[native knot]

The last 2 are

[soft zealot]

????

[native knot]

The hair in the 1st one is great, though...wish that was as good on the others.

[soft zealot]

ah are you referring to the options in the style selectors in my workflow?

Must have been added to thelast update to https://github.com/twri/sdxl_prompt_styler

[native knot]

Omg...found this on Amazon...it's worth you guys seeing to put into your workflow:
https://www.amazon.com/dp/B0CCVRTKD9/

I just....

[soft zealot]

well that 's what you call being inclusive 🙂

[spark sinew]

i just love to see your unique ai artwork 🙂

[noble shoal]

Controlnet gave me this beauty

[native knot]

Who was making that spectral ghost model? They need in on that Amazon link I put up there.

😄

[peak dove]

I'm using an "UnderWater" LoRA - it goes very well with mermaids and Koi Carp 😄

[ionic dragon]

can you try with longer hair

[uncut fiber]

4 at once Spock's hello 😄

[soft zealot]

@native knot Is this rainbow enouh lol

[native knot]

That's pretty far on that side of the spectrum!

[uncut fiber]

Does anybody using in comfui {a|b|c|d} ?

[strong field]

all the timehere

[uncut fiber]

great 🙂

[peak dove]

a|b|c|d as in multiple prompt choices?

[noble shoal]

Caught him off guard eating a cactus

[uncut fiber]

@peak dove it is {he|she|it} is cooking {egg|bread|cake}
And it choosing randomly

@crisp owl sorry didnt see it

[crisp owl]

np, I already was in the github so grabbing the screenshot took a couple seconds

[strange mist]

some actually asked me if I knew how to do something like that 😬
I assumed they were trolling, so I didn't say anything.

[crisp owl]

to be consistent, would probably require a Lora

[strange mist]

most likely, yes 😓

[shy kelp]

do you know any good lora for sdxl

[crisp owl]

I rarely ever use Lora's (except for the noise offset which I use nearly 100% of the time).
So I've not really gone looking at them

[peak dove]

So that's what the sdxl offset LoRA does?!

[uncut fiber]

Dont know how to understand it

[stone fossil]

Noice lol. 🙂

[soft zealot]

@native knot a bit of ControlNet added

[native knot]

Nice! Is that what added the sleeves?

[shy kelp]

how long colab reset their free gpu?

[soft zealot]

not really,, this was the control image

[native knot]

It adhered to the position well.

[native knot]

Trying to add a refinement step into my current workflow and getting errors... grrrrr.....

[steady grove]

they don't publish these numbers else people will just automate their scripts around them

[hardy cipher]

you know, I hadn't considered that. but makes sense

[steady grove]

they're trying to prevent non academics from abusing their free system provided for training people to use ML. It'll probably be shut down or limited only to education institutions soon. Too many people abusing it.

[hardy cipher]

yeah. it's unfortunate. but any system like that will have people trying to exploit it to the max

I've had their 10 dollar per month subscription for a while. sometimes I won't use it for a while, but the credits stack, and it's there when I want to do high vram training or whatever

[steady grove]

i mgiht use runpod when i want todo more intense training. collab needs a browser window open the whole time i think

[hardy cipher]

xl does some seriously wacky things with mona lisa's hands. I know hands are an issue anyway, but it really takes it up a notch in this instance

[noble shoal]

I know, not everybody here likes pixelart, but here are a few samples of a 64x64px Lora i am working on, trained in 08:26 Minutes on 58 images.

[hardy cipher]

pixel art is great. just not always my go to approach

[upbeat summit]

this is beautiful! nice work

amazing clarity

[noble shoal]

Thanks, i am more surprised about the training time. Not that i have a high end card. Just a 4070

[shy kelp]

[noble shoal]

I use a homebrew pixel refiner node to clear it up, but i get consistent "nearly" pixel perfect results out of the box.

[upbeat summit]

the different layers and color shades are impressive

[thorny frost]

a bit chessy

some deers have serious neck issues too

[hard fractal]

A B C or D?

[hardy cipher]

A

just because of the mouth. seems a bit more natural than in the other 3

[lean grove]

How does restore faces work exactly? I found myself actually getting better faces when I turned it off and now I'm confused

[hard fractal]

here's another:

[crisp owl]

It's just an automask and a trained model. Essentially a quick inpainting

[noble shoal]

A on both

[lean grove]

Should I keep it on at all times then? When working with faces

[hardy cipher]

C seems like it's rendered better. but I guess that's personal preference

[crisp owl]

it depends.
Sometimes it's helpful, sometimes it's detrimental.
If you find it's not working well with your current prompt, that's entirely fine to turn it off for it.
And vice verse.

Each image takes a bit of fine tuning

[lean grove]

So always try both? Thank you

I found myself getting a better face when I'm not using it

[late laurel]

Hi guys im new here how dose things work here

[hardy cipher]

Settings

[crisp owl]

Pretty much

If you saw how many iterations I go through before deciding on a final flow for an image style.....lol

[lean grove]

I would too buty generating a single image alone takes so much time lol so cant really do that

[crisp owl]

Just discussing and sharing.
If you want bots #1100170312106127410 etc

[lean grove]

Which upscaler do you people recommend by the way?

[crisp owl]

a full process for my generations takes approx 150 seconds

[lean grove]

3-4 mins here 💀

I saw people recommending swinIR upscaler on reddit but that was like looong ago

[steady grove]

i think there's a meta in the community where people rely on upscalers. with sdxl especially they're often not needed. gan upscalers ALL leave artifacts

[crisp owl]

I like this upscaler
https://huggingface.co/uwg/upscaler/blob/main/ESRGAN/4x_NMKD-Siax_200k.pth

I used swin a lot in the past

[hardy cipher]

Different upscalers have different effects on the image.

Some better for realism, others are better for animation stuff

[lean grove]

Whats the difference berween SwinIR and Esrgan for example?

[steady grove]

I just upscale the latent and diffuse it again

[hardy cipher]

they work differently under the hood, there are websites that compare them.

[steady grove]

we have diffusion technology available to us. why still use gans?

[vital ermine]

https://civitai.com/models/133981?modelVersionId=147528

[lean grove]

Can you explain what that diffuson technology is please?

[steady grove]

stable diffusion

[steady grove]

its how we go from a 1024 x 1024 pixels of absolute noise generated randomly, to a very coherant and detailed photo

[lean grove]

And how am I supposed to be using it?

[upbeat summit]

oh it's your release. congrats! I'll check it out for sure

[steady grove]

you might need to go back to the beginning and learn the fundamentals, if this is confusing you.

[vital ermine]

Yeah, I chose epoch 15

[lean grove]

I'm mostly self taught going by trial and error just messing with settings so I'm not good with the terminology, thought stable diffusion was only the name of the program

[vital ermine]

LONG ass time since I released anything and the first on Civit

[hardy cipher]

that's mostly how I learned. might have got a head start on you though. I'd just suggest reading tutorials and stuff. or watching videos if that's your thing. it's just a lot

[lean grove]

Thanks for the tip

[hardy cipher]

as far as upscalers, use them however you want really. knowing the underlying algorithms won't necessarily help you learn how they work in application

[lean grove]

Should I just choose them randomly for now?

[hardy cipher]

so it's a lot of just seeing what comes out. at least for me

[hardy cipher]

I've seen some websites that give recommendations

some better for realism, like I was saying, and some are better for other things

some of them really bring out noise and fine details which isn't always preferable. other times it is

[thorny frost]

i think viewing the forming images during the generation will give you a idea about how stable diffuison works. It turns noise in things, guided by your prompt using the info contaned in the checkpoitns.

[lean grove]

Oh yep

it creates visuals from noise

thats how latent noise also works on inpaint

[vital ermine]

I'd live here

[noble shoal]

I am not going to live in there

[vital ermine]

I dunno, I might but yeah, at night I am sure I would be shitting myself

[wet nacelle]

Does anyone have a resource I could learn from to make my own SDXL model?

[vital ermine]

I am coming to a realization that a full blown model is really not as well needed on XL as previously. Eventually it will have so much info that they will not be needed at all and only a simple low rank will be needed.

[wet nacelle]

Well then what is a good way to train anything at all then? I just want my hyper realism style to be it's own file instead of the workflow I'm using to actually make the images.

[stone fossil]

Nope use Kohya and make a LORA instead.

[wet nacelle]

How would I do that. Is there any resource I can learn from?

[crisp owl]

@visual glade I'm at a loss here.

Three different controlnet Canny models:

1. "Control-lora-canny-rank128"
2. "Adapter-xl-canny"
3. "diffusion_pytorch_model.fp16".

2 and 3 run just fine in my flow here, which for testing I adapted to your base process where I added controlnet where conditioning feeds only into the BASE ksampler. Then after refiner I have an img2img upscale process (in gold bottom right).

So as I said, 2 and 3 run just fine, a tiny bit slower than no controlnet, but nearly the same.
However, if I use 1 (control-lora), 128 or 256, when it comes to the img2img upscale process, I suddenly get absolutely horrid s/it times.

Alternatively, if I take the refiner image and save that, then load it in it's own dedicated workflow, I get back to normal processing times.
Talking about a difference of 15 minutes for the inline upscale, to 79 seconds for the standalone upscale.
This inline horrid time again is only occuring when using the control-lora models, but not the others. Included screenshot of time for adapter-xl-canny for comparison also

[stone fossil]

I am sure there msut be a lot of online tuts, a mate of mine learned me how to train. 🙂

[wet nacelle]

What do you use?

[stone fossil]

Kohya.

[wet nacelle]

ah okay

[wet nacelle]

How much storage do you think I should have to start with? I have 25Gbs right now.

[stone fossil]

It does not matter storage wise just start with a small data set like say 75 images.

[visual glade]

how much ram and vram do you have?

[stone fossil]

And test.

[wet nacelle]

Alright

[crisp owl]

64gb RAM
2060S 8GB VRAM

[stone fossil]

Also you do not have to corp or resize them you can use any size, Kohya will use buckets for each diffreitn size. However keep in mind when training for XL I would atleasy use 1024x something images and best 1024x1024 and up even.

[vital ermine]

@wet nacelleI spent over six months learning on 2.1. It was hell and hard, time consuming (over 1k hours) work, but I never managed much on 2.1 beyond knowledge I carried over to XL. There are a lot of materials online just stay away from Aitrapeneur for learning and you will be fine. Take your time, and if they are explicitly geared towards XL all the better.

[wet nacelle]

@stone fossiland @vital ermineThanks for the info

[wet nacelle]

Holy shit! I too hate this guy!

[high skiff]

@visual glade Is there any info for using CFG rescaling in comfy UI?

[visual glade]

https://github.com/comfyanonymous/ComfyUI_experiments

there's a node here

[high skiff]

thank you!

[steady grove]

aitrenpreneur be like "This is the ONLY way you need to do it. ALL OTHER WAYS ARE WRONG!" and then really makes a big deal out of this one thing he learned and barely explains anything else

its kind of cult leadery

[stone fossil]

Id love a hit from that one. 😄

[indigo carbon]

@visual glade hey we solved the batch size issue. the fix was adding to the compile script from diffusers import UNet2DConditionModel, AutoencoderKL, StableDiffusionXLPipeline. was this the fix for you as well?

[visual glade]

I don't use diffusers so I never had that issue

[crisp owl]

@visual glade not sure if you caught my reply here

[vital ermine]

I don't hate him, personally, but I unsubbed him a long time ago when all he did was publish 100% wrong info, then others, such as Sytan, began to spot his idiocy. I just watched his XL training and only the info SAI told him was right and past that point he fell back into his same idiocy to lead people astray.

[native zephyr]

[uncut fiber]

mega detailed miniature diorama a soviet residential building, brutalism architecture, suburban, warm lights are on in the windows, man lies in the snow, dark night, fog, winter, blizzard, cozy and peaceful atmosphere, row of street lamps with warm orange light, several birches nearby, 1 9 9 0

Still think these houses are very idealised on place and time, but like photo.

[vital ermine]

Yeah, though I will take idealized over the crack house burnt to the ground Detroit style happening all over the USA.

[indigo carbon]

to learn stuff, I usually just read the papers published with progress. but that's probably because I'm a psychopath

[vital ermine]

Possibly. I prefer, in 2023, to get out of reading papers as I grew up when that was all there was to learn from.

[uncut fiber]

Those dogs playing reminds me something. But i think prompt is quite large. Tried I Feel Lucky

[vital ermine]

I am finding revision to be heavily limiting as it ignores any prompts I use

are there parallel nodes in ComfyUI?

[soft zealot]

Possibly??
https://en.wikipedia.org/wiki/Dogs_Playing_Poker

[uncut fiber]

yes i know 🙂 @soft zealot
I always liked those images. But will check prompt for authors. Thank you!

[soft zealot]

FYI I believe "I Feel Lucky" node does the same thing as the WAS Nodes Random Prompt thing and queries Lexica to bring back an existing prompt randomly.

[hardy cipher]

what's the github for it?

[uncut fiber]

Prompt: dogs wearing top hats in a smoky room playing poker, by edgar maxence and caravaggio and michael whelan and delacroix style, artistic, intricate drawing, light brazen, realistic fantasy, extremely detailed and beautiful aesthetic face, establishing shot, 8 k resolution, dramatic lighting

Yes trained on 80000, also dont see difference @soft zealot

[hardy cipher]

found it

[uncut fiber]

github, dynamic prompt
o.k. @hardy cipher

[uncut fiber]

Random is total random feel lucky is working and modified what is in prompt probably? Realy dont know

[native knot]

From > To

[strong field]

what do you use instead of diffusers?

[strong field]

make tornado in lamp

[strong field]

Control nets?

[native knot]

Nope... IPAdapter

It feels very much like what I had imagined Revision should have felt like.

Just personally.

Not like a knock at Revision.

[strong field]

kids seem to like it

[soft zealot]

yes it does 🙂

[native knot]

That's great.

[uncut fiber]

i think if not connection with artstation then tag trending on artstation is useless

[soft zealot]

well yes it probaly is given that unless a node is specifically configured to query an externakl database then AFAIK there is no comms to the outside world

[uncut fiber]

i think it is good for learning prompting. In future i will go only for random.

[ivory blaze]

okay so using certain styles, chibi for example, anthropomorphic anything works out

but, not other styles

any hints on how people have made this work?

Colored pencil Realism realism style

[uncut fiber]

i think it is just art by somebody and somebody and ...

[ivory blaze]

chibi style

[ivory blaze]

well I mean I can make it in any manner, any object, no one has drawn before. So it isn't per se. I think perhaps, anthropomorphism is understand in some styles, but not others, due to a second tiered inclusiveness within the art style

the only thing changed here was chibi / pencil realism

[soft zealot]

over 1k hrs?

Presuming 10 hours a day solid in front of your computer that's 100 days

Wow

[ivory blaze]

Aitrapeneur is who I leanred everything from... I you suggesting he taught me something incorrectly?!!?!?!?!

look at this quality!

you cannot find better

I mean come on, I can't see you doing better.

Don't knock AIPretnaueur ...

just kidding, he's fun to listen to when he's goofing with some stuff for the comedy value, but, yeah, don't learn from hi m

I like that one guy, nice little round fellow with glasses, oliv...something. I will have to look

He's great

[visual glade]

Hopefully it's fixed now

[crisp owl]

updating

[indigo carbon]

I genuinely hope this is satire

[ivory blaze]

satire? moi?

how could you even assume....

[indigo carbon]

jesus

[glad grove]

yes son?

[indigo carbon]

listen, I don't want to yuck your yum- AIPretnaueur is NOT a valid source of information.

[ivory blaze]

His voice is cute though, he is so excitable, how can he not be authoratative?