✨｜sdxl

[shy kelp]

i still dont get it

[high skiff]

especially with a prompt as painfully simple as "medium shot of a handsome Indian man shopping at a busy street market in Bangladesh "

what is there to not get?

[candid walrus]

yeah that's a good one, i mean they all are looking away, good job on the photographers direction 😛

in the background

[candid walrus]

the other one it's like randomly place (more real life) but a good photographer is going to have that nice touch like the angled heads in yours

[steady grove]

Why did one fail that?

[high skiff]

what do you mean?

[steady grove]

Both look like that prompt

[high skiff]

yes, but look at the mangled people in the background of the first

specifically this lol

[steady grove]

Where

[steady grove]

Just a fat man in a robe

[high skiff]

also, comjoined hands with a third arm

[alpine pine]

that a lora?

[candid walrus]

that's reality haha 😉

[steady grove]

Both images have mangled hands and the basket is 4D

[wet nacelle]

He got that Two 1st Graders Trying To Look Taller look.

[candid walrus]

i do think in that one i like your lora better

[high skiff]

here we go, I have a new example

[wet nacelle]

naw

[steady grove]

Don't get it. Not arguing.. I just don't understand

[candid walrus]

less to fix and more pleasing on a 'golden ratio' perspective

[wet nacelle]

(exterior cute bedroom:1.2) fisheye@alpine pine

[steady grove]

I think man with grey beard is more handsome

[high skiff]

maybe this one shows the coherency improvement a little better lol

[steady grove]

Nohomo

[high skiff]

medium shot of a pretty Nigerian woman sitting next to a fire pit

[shy kelp]

how are you training it to do that tho

[cyan crown]

Share your Lora!

[candid walrus]

something about that gen tho feels a bit 'pregnant' tho tbh, can't put my finger on it

[steady grove]

Same seed?

[high skiff]

if that doesn't show what the benefits are, I don't know what will

[shy kelp]

is it a negative lora ?

[high skiff]

same everything

its a positive LoRA

and its nowhere near done, just early tests

[steady grove]

Do you never get extra limbs with your Lora is the claim?

[candid walrus]

i would classify it as natural lighting photography

with portraits as it's focus

[steady grove]

Both look like good images but extra limbs is an architectural issue

[cyan crown]

Time to go to bed and to the sea for 2 weeks. Se you soon!!

[steady grove]

Fixing that means over fitting

[hardy cipher]

I don't think anyone could claim to never get extra limbs

[candid walrus]

it's not about them being good images

[high skiff]

base vs my LoRA

Close up photograph of an old man whos balding with a slight smile in black and white

[shy kelp]

how it feels using comfyUI

[candid walrus]

it's about the one being naturally lit

[high skiff]

that and much more

[steady grove]

Seems bodacious

[candid walrus]

in that last one, you're also pushing to a real photo and not hyperreal art

[wet nacelle]

@high skiffWhen do you think you'll go public?

[steady grove]

Training is fun and you can tune towards captions very well, but to insist you've improved the bases architectural shortcomings so early....

[high skiff]

I have about 90 images right now, I wanna get to at least 200, preferably 400, and they are tagged by hand, so a decent while

[candid walrus]

i don't think he's saying that

[high skiff]

just this 90 image one took about 3 hours to train

[candid walrus]

i think he's saying he's getting more out of his promt in a way that he can classify (that he is wanting)

[wet nacelle]

I get ya

[high skiff]

Its a LoRA, not a finetune

[shy kelp]

aren't you technically conditioning the clip to favor your way of thinking what a prompt should be

[wet nacelle]

So a month and or slightly longer?

[high skiff]

the whole point is better realism with less fuckinga round to find magic words to get SDXL base to look good

[high skiff]

oh no, likely much less than that haha

[wet nacelle]

Ah okay. rofl in the copter

[high skiff]

I won't say any specific timeline, as I have bad health issues that have been flaring up

but likely before a month for sure

here is one from my old 60 image dataset that was undertrained

looks decently like a real person, rather than glossy plastic

[steady grove]

One would think if all quality takes is a prompt, just save those magic prompts and add them as a style

[wet nacelle]

Health issues can be the worst. especially when you know you might need to put into surgery knowing that it's for the best.

[high skiff]

I think i could get it better, but I am happy with that example

[steady grove]

Staff at sai keep saying prompting is king

[high skiff]

its a hell of a lot more than just "a few words"

It changes massively from image to image

Staff at SAI say a lot of things.... lol

[steady grove]

Big bit about that in the super stage event. They really went into detail about the breadth of prompts available to save as styles.

[high skiff]

styles work when its a general one size fits all glove, but its not for SDXL realism

and I should know, I have gotten some incredible realism out of base SDXL, at my own peril of prompting

[wet nacelle]

Here is the upscale. Everyone looks like Jimmy.

[high skiff]

lol

this is base SDXL with an absolutely huge positibe and negative

[high skiff]

take away even a few words and it reverts back to plastic city

[candid walrus]

i think they purposefully left the base unfinished to help others train/lora it into what they want, hence the refiner, and them saying prompt is king, and to save them for being blamed for it being too good at bad things, and to save money

[high skiff]

vs my LoRA that listens better to the prompt in under a quarter the prompt

"a medium close up shot of a pretty 20 year old woman with blonde hair and red lips being lit from the side by dappled sunlight through blinds"

[high skiff]

I can agree with that, it seems reasonable

[candid walrus]

yeah and i don't mean unfinished with any bad connotation at all

in fact it does serve the public well

[high skiff]

"A tiger in a forest"

SDXL base vs my LoRA

[shy kelp]

what if you add tokens for photorealism

how does it compare

[midnight shuttle]

SDXL base does tend to favor matte paintings vs. photos.

[wet nacelle]

@high skiffHave you tried any word type prompts?

X holding a sign that says "enter for X"?

[shy kelp]

also how long does it take for you guys to make a 2k image

2048x2048

takes like 8 minutes for me 😭

[ashen oracle]

you mean pre-upscaling ?

[shy kelp]

i mean with upscaling

[high skiff]

Do you have an example of what you mean? Like can you provide a prompt example?

[strong field]

You asking because you want someone to give you a reason to upgrade?

[shy kelp]

no just curious how shit my pc is

bro i'll never be able to afford upgrading

i'm living off disability

neetbux

[high skiff]

uhhh... with my workflow... about 30 seconds on a 3090

[midnight shuttle]

What GPU do you have?

[candid walrus]

imo the lora is a little 'cold' in photo temp

[strong field]

20-30 seconds for me for 4k

[high skiff]

I kinda averaged it to neutral white, which I have not messed with prompting as of now

[candid walrus]

if it could know about that as well it'd be a plus, it's nothing photoshop can't fix but nice to have

[strong field]

Using AIT

[high skiff]

let me see how well I can prompt for color temp

[strong field]

I’ve pushed to 8k in under a minute

[shy kelp]

1080 ti

damn thats so fast

[shy kelp]

💀

[midnight shuttle]

Not the greatest but should be able to handle SD 1.5 512x512 and maybe even SDXL 1024 with optimizations and patience.

[shy kelp]

im running sdxl rn

in comfy

normal image takes a minute

[high skiff]

interesting, I can't seem to change the color balance

let me see if I can find a proper termp to help

[candid walrus]

its funny tho some dslr will sort of undershoot on auto color temp, so if we call it an error it errors on the right side

[midnight shuttle]

Sounds about right. I can get 20 sec/img on a 3060 but I have to use the FP16 VAE.

[candid walrus]

white balance 6000k

maybe something along those lines

would be camera speak

[high skiff]

i got it a little more warm

[candid walrus]

damn

that's perrrritty

[high skiff]

my dataset is very neutrally color balanced, so I am not sure how to combat that

I am sure I could add in color temp tags

[shy kelp]

add images with color temps and label them

just do 6000k, 7000k, etc

[candid walrus]

sort of like i said before cameras often favor the cooler side, so not sure there

i was just going to say that ^

[shy kelp]

my guy is missing a leg 💀

[candid walrus]

it'd be interesting if there was a tool that took the average colors to a group of images and see if they are favoring any way

[trim orbit]

he's half stepped in and half stepped out bruh. like on the edge of the liminal

[shy kelp]

shift is almost over

[glad grove]

we did it

[shy kelp]

yooo cant wait for this one to upscale

[high skiff]

FAT.

lol

[trim orbit]

it really catches alex's style

[shy kelp]

shmolies

crazy waimi fat

[glad grove]

shmoies magazine do be gettin crazy tbh

[shy kelp]

its a lora

[trim orbit]

do you guys ever click to drag comfy ui, but acidentally click a group title and drag it, then you move it back and it takes a bunch of other nodes with it?

i gotta use space more

[shy kelp]

i just realized my ability to read is as good as SDXLs ability to write

😂

[shy kelp]

yeah, i tend to pin things now

[soft zealot]

I would contend that the image on the far right is probably more realistic than most women post on IG/FB/X etc

[soft zealot]

this is why I like https://github.com/failfa-st/failfast-comfyui-extensions, you can pin all & freeze all from a TH click context menu

[trim orbit]

got

thanks

[shy kelp]

🤔

[trim orbit]

you're actually using an alex lora! nice. is there one?

[trim orbit]

i can dl

[indigo carbon]

I prompted "rich" and this guy is holding 10 bucks

[trim orbit]

https://civitai.com/models/118211/alex-grey-sdxl-10-art-syle-lora found'er

[glad fulcrum]

Is SDXL Lora training possible with gtx 1080 (8GB VRAM) ?

I know I can do it in 12 but my friends asks

[high skiff]

is that stan lee? lmao

[trim orbit]

the man in the back lol

[soft zealot]

I have a 1080ti with 11Gb vram and whilst I happily use it to generate images the thought of trying to train a LOTA on it has mostly passed me by.

[trim orbit]

he's got the biggest bills of them all

[indigo carbon]

he didn't pay taxes

[high skiff]

decided to give a little water gun gen a try with my realism LoRA lol

[strong field]

lol

[shy kelp]

machine pistol 5

[high skiff]

not as wet as I'd like

[upbeat summit]

nice

[high skiff]

Oh, I bet this thing does fire guns

[soft zealot]

its just a little moist

[indigo carbon]

this is how my workflow does waterguns

[upbeat summit]

it does

[strong field]

here's a wild ass 8k (over UHD) for yall that were asking

[high skiff]

I've gotta admit... thats a pretty damn good glock lmao

[strong field]

impressive

[high skiff]

real glock

tho, I didn't specify a specific model

[strong field]

dang even went for the logos

[strong field]

thats wild

[high skiff]

wait a sec-

woah

My LoRA seems to greatly improve celebrities

[strong field]

im waiting for your DROP

[high skiff]

its not gonna be any time soon

its not well trained

I need a much bigger dataset and more time

[strong field]

Yeah my favorite thing about AIT is the reduced temps and power usage

[high skiff]

but I mean

[strong field]

dang, thats really good

[high skiff]

trying others

[crisp owl]

oh what movie is that from?

[soft zealot]

the corporate catalogue look ;o)

[indigo carbon]

yeah, it's a breakthrough. I'm generating stuff like this in a few seconds

[indigo carbon]

I kinda like how my stuff follows that prompt better, but nice!

[zealous horizon]

Well, we generate LoRAs from a single image and OH EM GHEEEE
I really put 0 effort into the prompts on these
(LAST image in the field is the original)
SDXL is crazy

[soft zealot]

more than one to skin a cat

[strong field]

same

[high skiff]

what is AIT?

[strong field]

how you getting that beauty of a txt file winston/

[indigo carbon]

like, aesthetics are completely opinionable, but speed is factual. my AIT workflow allows me to generate this in a few seconds

[high skiff]

what is AIT

[indigo carbon]

tensorRT's big mama

[strong field]

https://github.com/FizzleDorf/AIT/tree/sdxl

[high skiff]

How does one set that up?

interestinngggg

[soft zealot]

All this down in the basement of my workflow

[high skiff]

I will have to try it sometime

[strong field]

https://github.com/facebookincubator/AITemplate
bit words on here that says things i dont understand

[indigo carbon]

I actually built a workflow designed specifically for AIT. this image took my 4070ti 12-ish seconds to make.

[high skiff]

very interesting

[upbeat summit]

AITemplate is a game changer

[strong field]

20 seconds to MUCHO HIGH RESOLUTION with 3090

[indigo carbon]

yeah, it's actually insane

[trim orbit]

i'm going to have to throw all i know about steps out the window i think

since everytime i try to not use your workflow, my images blow

[high skiff]

lol

[indigo carbon]

you can change as much stuff as you want with the workflow where it comes to prompts, ARs, etc.. but the sampling settings are crucial

[soft zealot]

oh (((^* , just realised its not putting all of it in there, thats a later today job, gone midnight off too bed

[trim orbit]

baulders gate 3 is out now an has some great visuals. been thinking bout that lately in the context of training xD

[soft zealot]

but then I realise, all it is is cos these 2 arent joined

thats better, all present and correct

[hoary saddle]

Good Lord man. That's a lot of prompt beef. 😂

[soft zealot]

I only typed in this bit lol

[hoary saddle]

Wow, that's pretty insane it added all that.

Linguistic styling templates?

[soft zealot]

I'm using @high skiff s terminology ;o)

[glad grove]

cant wait for SDXXXL to have base prompt L,base prompt G,Base prompt P ,Base prompt F,Negative support,Negative base,negative global,positive global,positive linguistic,positive L,positive S,positive plus,

[soft zealot]

You forgot the pre conditioning options

[heady vale]

Reading through what was discussed just earlier on AIT, I also did my own tests and found there is a slight degradation in coherency.
The top row is my normal workflow.

The bottom is the same setting and seeds but with AIT

would that be the CPU/GPU seed generation thing?

[midnight shuttle]

Did we ever get any confirmed proof from the code that CLIP-G is "linguistic prompt" and CLIP-L is "supporting terms" or was that determined by guesswork?

[upbeat summit]

first of all thanks for doing this test! I'm working on it and will share as soon as I have some results.

would you be able to share some specs about your test? what gpu, os and what sampler settings?

I was wondering the same if there are any trade-offs.

[heady vale]

WIn 10, 3080, dpmpp_sde_gpu

I tried various steps and found the same result

so not sure if its maybe my setup or something else. Was wondering if anyone had done something similar

The other thing I found is that AIT doesnt like a fixed seed for Base and Refiner. I had to use random seed for refiner or it gets all messed up

[upbeat summit]

there were problems with xformers too but a lot of stuff was optimized

[upbeat summit]

very interesting

[indigo carbon]

very likely. I built the AIT workflow with GPU ARG in mind.

[heady vale]

where can I change that?

[upbeat summit]

so my go-to sampler is dpmpp_sde_gpu karras and I noticed at least a 80%-100% speed increase. but dpmpp_2m karras is even faster and @indigo carbon thankfully shared the AIT optimized workflow and I have since then worked with dpmpp_2m and I think the results are great.

but I need to do A/B testing shortly

[heady vale]

cool. will be interesting to see what you get

[indigo carbon]

ComfyUI\nodes.py noise = torch.randn(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, generator=torch.manual_seed(seed), device="cpu") change this line to this: noise = torch.randn(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, generator=torch.cuda.manual_seed(seed), device=device)

[upbeat summit]

I don't have any clue how this technically works so I can only speculate if there are side effects we don't know yet - and maybe only with certain setups and settings

[indigo carbon]

this line tells ComfyUI how to generate an empty latent image from seed. default is CPU, but with GPU quality is somewhat higher IMO.

[vagrant acorn]

guys, I put a video to generate in the deforum, the video has 13 seconds, and it has been generating for two hours, a 920x500 video, is the delay time normal?

[heady vale]

I will try some sample test with that later today

[spring fulcrum]

@indigo carbon Can you send me the AIT optimized workflow or an image that uses it?

[indigo carbon]

dpmpp_2m with Karras enabled is very good with SDXL, especially with AIT

that's what's used in my workflow

[heady vale]

2m is quite good and much faster but I generally see better results with sde_gpu

the trade off is its slower

[upbeat summit]

yeah I'm glad I'm given it a chance again 🙂 thank you!

[indigo carbon]

my pleasure. I'm glad you guys enjoy the workflow I made =]

[strong field]

Lots of experimenting left to do

[autumn forum]

getting this error lol "Input type (torch.FloatTensor) and weight type (torch.cuda.FloatTensor) should be the same or input should be a MKLDNN tensor and weight is a dense tensor"

[strong field]

Been finding best consistency with dpmpp 2p with Kara’s

[indigo carbon]

screenshot? this might be the preview error

[autumn forum]

maybe let me see

[autumn forum]

it was lmao my bad

holy shit this is fast

[strong field]

Lol

[autumn forum]

i blinked and it was done

[indigo carbon]

if you want live preview use ComfyUI_manager's live preview, only one that works with AIT

[autumn forum]

i did

and 9 seconds for 67 steps wtf

[strong field]

That’s what we’ve been saying

[spring fulcrum]

Ok so I'm using the workflow but getting the same speeds.. Is there anything that I need to download or do to get AIT to work? I haven't been in the discord server in about a week.

[strong field]

https://github.com/facebookincubator/AITemplate
bit words on here that says things i dont understand

[indigo carbon]

that's odd.. does the console say something like: "module not found"?

[spring fulcrum]

no... everything is there. I just don't know what AIT is... Is it part of the workflow itself?

[indigo carbon]

the first generation with AIT is always slower. do you have the AIT nodes?

[spring fulcrum]

Not sure.... These both say disable?

---

[indigo carbon]

that's how it should be. I don't understand why doesn't it do that boost for you... what GPU are you using?

[strong field]

Got clone the link into custom folder

[spring fulcrum]

4090

[strong field]

Nvm looks like you got past that step

[indigo carbon]

nah, you should be getting atleast a 2x boost. this is strange

are you on linux maybe?

[spring fulcrum]

i get 8 its/s

windows

[indigo carbon]

very strange

what CUDA version?

[spring fulcrum]

2.something... I forgot how to check what version it is

[indigo carbon]

that doesn't make sense.. it should be either 11.x or 12.x

[indigo carbon]

11.8 is great

no issues with that

does your Comfy setup have any tweaks?

[strong field]

Are you running with command line args?

[indigo carbon]

@strong field you have a 3090 right? @spring fulcrum should get at the very least the same boost as you get

[strong field]

Yes 3090 here

[indigo carbon]

that's docker?

[spring fulcrum]

no just regular build

not on docker

[heady vale]

I searched in nodes.py but cant see that exact line to change. I only see this that is similar

    device = comfy.model_management.get_torch_device()
    latent_image = latent["samples"]

    if disable_noise:
        noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu")```

[strong field]

What’s the standalone build commands for?

[spring fulcrum]

It was in there when I installed it... idk

[indigo carbon]

oh, right, bat file.. your flags are weird. I don't use any flags, it might interfere with AIT

[strong field]

I do not use any modifiers either

[spring fulcrum]

what speeds do you get with this build? what it/s?

[indigo carbon]

I have no other explanation why it doesn't give you that boost if that's not it.. maybe your behind a few ComfyUI commits? @spring fulcrum

[indigo carbon]

I have a 4070TI, and it's at least 9.5it/s, sometimes over 10. overall generation time is 12-16 seconds

[spring fulcrum]

its all good

its not like its slow... I was just thinking it would be something crazy like 30 it/s

8 is still fine

[indigo carbon]

it should be WAY faster than 8 if your on a 4090.. @upbeat summit uses a 4090 and it's way faster than that

[spring fulcrum]

maybe I should just do a completely new install of ComfyUI

[indigo carbon]

try to do a few gens, the first one is always slower @spring fulcrum

[spring fulcrum]

I did like 10

[indigo carbon]

odd..

shouldn't be like that

likely a problem on your end. most likely due to Comfy flags or even you might be a few commits behind.

[heady vale]

have you tried
pip install -r requirements.txt in the comfy folder?

[indigo carbon]

also this. a 4090 should be way faster than a 4070ti with AIT, I don't understand if the problem isn't an inference issue

[indigo carbon]

also yes, this line

[spring fulcrum]

Did you guys use the portable comfyui with nightly?

[lilac raven]

Is there something to see the image metadata more consistently? For now, I'm just using strings

[indigo carbon]

no

are you on pytorch nightly @spring fulcrum ?

[spring fulcrum]

no but I was wondering if that might work better?

[indigo carbon]

no, it won't work at all

[indigo carbon]

it might actually be you pytorch version, you need pytorch2.0.1+cu118

if that's not it it's 100% the flags you are using, and maybe even a few commits behind

[heady vale]

this is what I get on my 3080

[indigo carbon]

that's way faster than without AIT on your 3080, right?

[spring fulcrum]

how do i check my pytorch and cu version?

[heady vale]

miles faster... its crazy fast

[indigo carbon]

i figured

@spring fulcrum go to the main directory of ComfyUI and run pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

if this is not the torch version you are using that might be the issue

[spring fulcrum]

it says already satisfied

[indigo carbon]

I don't understand what tf is wrong then. maybe things are running in the background?

[heady vale]

do pip list torch

its shows everything

[spring fulcrum]

thats a long list are you looking for anything specific?

[mellow tendon]

This SDXL Monster Lora is cool: https://civitai.com/models/86377
Makes some freaky S*** thought you would like it 🙂

[indigo carbon]

theres your issue =]

torch 2.0.0

wrong version

[spring fulcrum]

what is that command line to update that

[heady vale]

for reference Im also using nightly versions of torch and xformers

[indigo carbon]

this should also work

to update torch run pip3 install torch --index-url https://download.pytorch.org/whl/cu118

[spring fulcrum]

when I run that command it says its already satisfied

[indigo carbon]

force reinstall then, I forgot what was the force reinstall flag, let me check

[indigo carbon]

pip uninstall torch
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

[glad fulcrum]

has anyone tried these? it seems they are controlNet models for SDXL? https://huggingface.co/SargeZT/t2i-adapter-sdxl-multi/tree/main

[spring fulcrum]

it keeps using a cached version 2.0.1 is there a way to clear that cache?

[indigo carbon]

2.0.1 is what you need, on the list you sent a few moments ago it said 2.0

[spring fulcrum]

see the last line in the ss

[indigo carbon]

you need VENV for this I'm afraid

make a venv in ComfyUI dir and run comfy with the venv, it should resolve all this on it's own

[spring fulcrum]

idk how to do that. I never use conda

[indigo carbon]

not conda, python

go to ComfyUI main directory, run python -m venv venv

after you did that, add call "DIRECTORY\venv\scripts\activate.bat to your batch file and remove unnecessary flags @spring fulcrum

[hasty smelt]

Hello everyone,
I'd like to know if I can participate in the SDXL image contest on civitai.com using the samples generated by my bot here in the room. Unfortunately, I can't use the XL with my 2070 GPU as the model is too heavy for an 8-gig GPU. I've been getting great results using the bot, and it would be awesome to be able to use them in the contest. However, I'm unsure how to obtain the information that the bot uses in my outputs, such as seeds and other relevant details present in my images. This information is necessary for participating in the contest.
Thanks for your help!

[spring fulcrum]

like this?

[heady vale]

fairly sure the contest requires meta data parameters on the image which the bot doesnt attach, so that wont be possible

[spring fulcrum]

this is what it looks like when I start it up

I don't see any reference to venv. Is that normal?

[indigo carbon]

yes, it's active

[indigo carbon]

oh wait, no

[hasty smelt]

I understand, that's what I thought. The bot doesn't provide the essential metadata required to participate in the contest, which is a shame. A new GPU would be amazing.

[indigo carbon]

call venv needs to be before python main.py @spring fulcrum

[heady vale]

have you tried ComfyUI? 8gb is possible

[indigo carbon]

yes, that's good

[strong field]

Yeah that’s looking right @spring fulcrum

Waiting eagerly for the OH SHIT moment

[spring fulcrum]

?

[heady vale]

your system ram is barely enough, probably why its slow. Comfy will use 20gb of system ram

[strong field]

You should be able to do comfy with 8gb, maybe a tweaked workflow but possible nonetheless

[spring fulcrum]

The system ram is 128GB DDR5

[indigo carbon]

that's amazing

[heady vale]

oh yeah, I read the number wrong...lol

[indigo carbon]

you removed a zero..

[spring fulcrum]

silly numbers tricks are for pure fire lol

[heady vale]

I did

[hasty smelt]

I tried with automatic 1111 and my rendering became extremely slow, so I ended up uninstalling the XL model. I have never used ComfyUI, and I couldn't find any clear tutorials in Brazilian Portuguese to teach me how to use it.

[spring fulcrum]

Ok Ill be back later to mess this up further... I need to eat.. havent eaten anything since I woke up this morning.

I forgot to eat

[strong field]

@spring fulcrum mad lad that spec is fire

[indigo carbon]

yeah, easily some of the best specs

[indigo carbon]

that's why I'm confused why AIT didn't boost @spring fulcrum's performance by atleast 2x

[spring fulcrum]

Check dem specs... I'm going to go eat something.

[heady vale]

not often you see 128gb.. very nice

[indigo carbon]

yeah, this is some Linus Tech Tips stuff

[mellow tendon]

This SDXL ControlNet works with ComfyUI out of the box: https://huggingface.co/diffusers/controlnet-canny-sdxl-1.0
Only has Canny right now.

[glad fulcrum]

Yes I downloaded that one already, but there are other ones in the link I posted, but don't know if they are working really.

also they are way lighter-weight

[mellow tendon]

I read they work standalone but there is more to do to get them compatible with Automatic1111 or ComfyUI and the developer is working on InvokeAI intergartion first.

[indigo carbon]

@spring fulcrum I need to go. If you would still need help, @upbeat summit might help you if they're up for it, they also got a 4090 and the AIT workflow I built works for them flawlessly.

[hasty smelt]

I will try to install SDXL and ComfyUI. Do I need to remove my previous files and also automatic 1111 from my system to proceed with the XL installation?

[heady vale]

nah make a new folder and install comfy there. its separate

[hasty smelt]

thanks buddy

[trim orbit]

also, symlink the model folders aww baby

[heady vale]

@upbeat summit I changed the nodes.py file as Tdg8uU mentioned from "cpu" to "device" and Im seeing better results, for your reference. The faces are a lot better

[upbeat summit]

Very interesting. I should check that out as well. Thanks for letting me know!

that does look a lot better

[strong field]

Oh wow I haven’t done that yet but I will be when I get back to command center

[glad fulcrum]

what do you mean? what is this

[heady vale]

its related to AITemplate

mentioned a little bit back in the conversation on what to change

[upbeat summit]

you mean this #✨｜sdxl message, right?
currently in nodes.py
line 1154
noise = torch.zeros...
and change that to
noise = torch.randn...

[hardy cipher]

fancy lads

[heady vale]

yeah because the code was different. all I changed was the last part where it says CPU, to device

[strong field]

Oooooo

I’m down

[upbeat summit]

ah ok. just cpu = gpu

[heady vale]

not sure why Td had different code

yeah

[upbeat summit]

yeah I was wondering to replace the whole line

[heady vale]

so you have the same as Td?

[upbeat summit]

I have noise = torch.zeros

[heady vale]

yep same

Im not a coder so not sure what it is. but logically cpu to device would make sense

[strong field]

it makes sense, i wont pretend to know like i know what half those lines are for

[hardy cipher]

which cpu do I change? found 6 of them. just hit em all and see what happens?

Line  397:         latent = safetensors.torch.load_file(latent_path, device="cpu")
Line  857:     def __init__(self, device="cpu"):
Line 1154:         noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu")
Line 1268:             i = 255. * image.cpu().numpy()
Line 1327:             mask = torch.zeros((64,64), dtype=torch.float32, device="cpu")
Line 1374:             mask = torch.zeros((64,64), dtype=torch.float32, device="cpu")

[upbeat summit]

hmm... yeah I don't know how any of this science fiction technology works, but it's a revolution :D. I'm not quite sure why you get so much differences in quality though because of that change

[heady vale]

yours didnt?

[hardy cipher]

how do I force them all to run on my onboard video?

[upbeat summit]

doing it now and will compare seeds - if it's possible. the seed might be totally different

[strong field]

1154

[hardy cipher]

awesome, thanks1

[heady vale]

yeah it definitely has a different seed between AIT and non AIT. Other than that Im not sure what its doing. Maybe its possible that all the images were just bad seeds for the comparison?

[upbeat summit]

hmm... I don't think so. it could be an unstable prompt build

but if you are now doing good images all the time with the same prompt, it must have something to do with it

or you relaunched comfyui and it was something else

tech is complicated 😄

[heady vale]

lol it can be

[upbeat summit]

testing now

[strong field]

well it sure didnt break anything changing 1154 lol

faces are worlds better

impressive

but i gotta say this is hilarious

sdxl HILARIOUS with it

[heady vale]

cool, so its not just my imagination

[strong field]

yeah i was generating trash for faces earlier testing

going back to old prompts to compare, this is almost no prompt fu

i do think mine slowed down ever so slightly

from 6.9it/s to 6.44it/s

[vernal cloak]

Hey guys. Can anyone confirm to me whether or not this (Text:1.2), is the same as ((Text)) This?
Or do they have slightly different results when in use?

[strong field]

so... not a big deal

[strong field]

in theory or in practice hehe

i believe they are handled differently

[dense chasm]

https://civitai.com/models/126239?modelVersionId=137968 just upload character rola model on civitai and huggingface hosted inference API

[vernal cloak]

Like in practise. I have used both in the the past. But i have a feeling that ((())) These gave me different results to (:1.0) these.

[upbeat summit]

hmmm - I guess it must already work in gpu mode on my system?

really really minor difference in this one.
Windows 11, 4090, dpmpp_2m karras, 70 steps, cfg 8 (that's high I know, but it works with some prompts)

doing some more tests

1: device="cpu" / 2: device="gpu"

[strong field]

i like using the (:1.0), i have found better control

it likes to ignore my ((whatever))

[vernal cloak]

yh. thx. it had me wondering. lol

[strong field]

my faces are defnitely better

[heady vale]

I need to head out. I will be back a little later

[upbeat summit]

ttyl!

[upbeat summit]

yeah I totally believe you

[hardy cipher]

guys, why not change all the cpus to device? push the limits

[strong field]

lol

[upbeat summit]

https://tenor.com/view/smart-thinking-thoughts-think-ponder-gif-7713620

[strong field]

guess you could always just git pull if you break it

[hardy cipher]

I already made a copy of the file

plus you could just save it again

[urban fjord]

I made a new LoRA for perfect faces

[spring fulcrum]

@indigo carbon @upbeat summit @heady vale Ok so with the venv setup I'm getting like 12 it/s @upbeat summit does that sound about right or are you still getting higher numbers than that?

[trim orbit]

https://tenor.com/view/kanye-kanye-west-roblox-kayne-ratgifz-nervous-kanye-gif-14908890

[strong field]

im not liking refiner tho

[upbeat summit]

what's your gpu?

[strong field]

on a 4090 that seems about right?

[spring fulcrum]

4090

[upbeat summit]

that's awesome 😄

[strong field]

thats from 8 to 12, thats decent

[spring fulcrum]

@upbeat summit they said you have a 4090... How many it/s are you getting?

[hardy cipher]

those outfits probably cost 20k or something

I don't know if stable diffusion could outdo what kanye just wears in real life

[upbeat summit]

so with the Tdg8uU's unaltered workflow it takes ~17 - 18 seconds per generation

[urban fjord]

One image, 80 epochs, SDXL is quite good at learning.-

[spring fulcrum]

oh... mine is at 14 seconds per image.... I was just wondering what the it/s was while it was generating?

[upbeat summit]

ok I have not used his latest one

let me try

[hardy cipher]

so does changing that cpu to device do it for all the samplers, or only that particular one?

[strong field]

nah bruh, i think i got u

[upbeat summit]

beats me - probably all of them

[hardy cipher]

those are his pajamas

[trim orbit]

how did all the cool adidas sneakers turn into these ugoh things he marketed?

[hardy cipher]

narcissism

[trim orbit]

i started buying nikes cause they're still doing sneakers

[hardy cipher]

and various other mental illnesses

[trim orbit]

no but, people buy them. thats why the whole market shifted for the product

i heard that they sold more than air jordans

[hardy cipher]

who knows. I like the ridiculous old school nike yeezys

[trim orbit]

thats insane to me

[void stirrup]

Is there any documentations somewhere that talks about all prompt commands for SDXL 1.0?
Just wondering
I'm super new to this LOL

[hardy cipher]

the red octobers

[indigo carbon]

I'm back =]

[hardy cipher]

look like some 1998 k-mart charles barkleys or something

[indigo carbon]

@spring fulcrum did you fix it?

[hardy cipher]

but sell for 10k+

or get the reps from china for 150

[strong field]

this isnt official documentation but a solid start

https://docs.google.com/document/d/1BxdWqfBJ3QPggHnBCBx3QIkUpHnBWShd_s3zEt2dTLM/edit#heading=h.po8dv7lcq48s

[void stirrup]

@strong fieldThanks a lot!

[trim orbit]

converse used to be a kmart brand lol

[hardy cipher]

are british knifhts still around?

knights

[spring fulcrum]

I think so... I'm getting like 12 it/s now maybe it could still be better

[trim orbit]

i had the very first litemup shoes as a kid. the bk lights

frigggin so cool

[trim orbit]

i still cheer those kids on when i see them with their litemup heelies

[indigo carbon]

12it/s sounds right. I think that's the same as @upbeat summit 's speed?

[hardy cipher]

I remember an episode of cops where dude was trying to run from them with the light up shoes

[upbeat summit]

I ran a test with just this image and its embedded workflow by Tdg8uU

11.56 it/s on base model
10.75 it/s on refiner model
Prompt executed in 14.87 seconds
Windows 11 22H2, 4090 (driver 536.67), latest ComfyUI and AIT build (20230812) 

[strong field]

i think i found my new hobby

[indigo carbon]

@spring fulcrum this is what you should be seeing if you have set it up right

[spring fulcrum]

Seems like I am on point now

[indigo carbon]

great! it worked =]

[upbeat summit]

nice - have fun 😄

[hardy cipher]

I'm not sure I quite understand the civitai prompt submission contest. am I allowed to do anything other than just base/refiner submit? I mean, I want a 4090

[spring fulcrum]

I think my numbers might even be a little better... but not by much its really close

[strong field]

shit this too funny

[hardy cipher]

too soon?

[upbeat summit]

you need to share the workflow - which will be funny if you use wildcards also they probably just want a lot of workflows and new metadata

[spring fulcrum]

Its a ton better than the 5 it/s I was getting before doing this

[upbeat summit]

amazing tech 🙂

[strong field]

no doubt its about the data

[indigo carbon]

yeah, it's an AI breakthrough if I ever saw one

over 2x boost for you it seems

[spring fulcrum]

ya its really awesome... thank you for the help with that 🙂

[upbeat summit]

added some info to my post, just in case for later referencing

[indigo carbon]

huh, I'll add this to my documentation. interesting how much faster AIT gets with pytorch 2.0.1

[upbeat summit]

how do I update to pytorch 2.0.1 ?

[spring fulcrum]

I still think I'm on the wrong pytorch version btw... I think it was the change to venv that made the difference here

[strong field]

pip uninstall torch

[vast galleon]

Comfy I2I is now public if anyone would like to play around with it. Please let me know how it works for you.

https://github.com/ManglerFTW/ComfyI2I

[indigo carbon]

you're already on it, the reason they had issues is because they weren't on 2.0.1

[upbeat summit]

ah - it makes sense now. I started with a fresh comfyui setup for AIT

[hardy cipher]

what about bitsandbytes? I forget why I uninstalled it. I had reasons

seems to work fine without it

[strong field]

i dont have bits and bytes for comfy

wonder why you would need it?

[hardy cipher]

some nodes wanted it

but then it made something else angry

so just uninstalled it

[buoyant axle]

@indigo carbon used the aitemplate nodes with a custom workflow i am playing with. The flow is pretty much the same (without upscaling)...i just have a lot more primitive nodes to control other nodes input. the only thing that is a bit different is, that i am using ksampler adv instead of simple and i am using the sdxl clip encoders (for base and refiner). with this setup i "only" get ~6.5it/s (base) ~4.41it/s (refiner) instead of the ~12 with your workflow. i am wondering what exactly is making your workflow that faster now?! is it using the simple clip tenc for base? or using the simple ksampler? im confused ....

[upbeat summit]

since I'm testing a lot of fine-tunings has anybody done tests with X/Y plotting (efficiency nodes etc) in ComfyUI? Or is it still better to use a1111 for it?

[strong field]

yeah i do my testing with those

lots of setup time but once you get a workflow going its easy to spot what you like

[hardy cipher]

I think for major xy plotting just run it when you're doing something else

set it to run when you go to the store or something

[strong field]

yeah overnight for that

[hardy cipher]

I have my outputs folder synced to my google drive so I can watch the progress from afar

[strong field]

man ive been trying to get the --listen args to work on my home network so i can just watch from my phone but cant get it done

[hasty smelt]

I did it! I finished installing SDXL and also ComfyUI. I'm not sure what magic they did here, as the render time is much faster compared to automatic 1111. I just conducted my first test, and everything worked perfectly. I'm really excited to be able to use XL in a local host setup. Thank you for the suggestion; ComfyUI seems amazing. As someone coming from 3D, I'm accustomed to nodes, so this is great. Thanks to everyone for the help!

[hardy cipher]

hardly even worth loading comfy on your phone

[upbeat summit]

yeah. I mean x/y plotting is very important to find out if a fine-tuning does what you want. I need plots to help test stuff and variations of models.

so in a1111 it's super easy and I'm thinking about doing a setup in ComfyUI but I have not really checked it out yet and I'm not sure if it makes sense or just to switch to a1111 for the plots.

[hardy cipher]

unless your has some sort of capabilities that mine doesn't

[prime bison]

Is this yours by any chance? If so, would you be able to export to HTML? Google docs loads content after the page loads and loses scroll position 😦

[strong field]

not mine, the owner hangs around in ehre

its a WIP

[indigo carbon]

you don't need SDXL clip encode. the ones built into ComfyUI are way better. main difference is - the SDXL ones are made for multiple positives, which is a bad idea IMO. one positive prompt seems to make better, faster results.

[buoyant axle]

okay let met try that. my custom wf uses only one prompt as well. it is just attached to _g and _l...ill replace it and do another check

[hardy cipher]

as I understand it there are use cases to seperate the G and the L, but for most of us it's probably not beneficial

still, I do it, lol

[strong field]

can you do XY with AIT though? I havent found batch is supported yet

[hardy cipher]

what do you guys do for the refiner though? that's the magic question in my book

[upbeat summit]

would you say it's lot harder in ComfyUI to run x/y plots or are you getting all the options you need?

[hardy cipher]

I've been sending both prompts, combining them, and overwhelming the refiner, lol

[urban fjord]

You don't need to do batches with XY, you just create each image in sequence.

[hardy cipher]

maximim quality

[strong field]

a1111 is def easier to get up and going

but comfy just makes every little option available

i like to tinker 🤷‍♂️

[hardy cipher]

I legit cannot wrap my head around the comfy hate on reddit. it's like hating on photoshop or ableton or something

but if they were both free

[indigo carbon]

pass latent without VAEs in between. so much better and more efficient this way

[hardy cipher]

because you want to use ms paint

[strong field]

lol i thought it was batching this whole time

[hardy cipher]

don't mean to sound like a dumbo, but how do I do that? I'm not sure I have VAEs in between though

[upbeat summit]

are you using efficiency nodes to create the plots?

[strong field]

i started with the efficiency workflows on their page

[indigo carbon]

IMO, the best option for people that hate nodes is ComfyBox, which is just like A1111- but it's only a frontend

[upbeat summit]

yeah and their nodes I suppose

[urban fjord]

If you don't like people disliking ComfyUI, don't compare the alternatives to MS Paint please.

[strong field]

yes

[upbeat summit]

thanks! I will check it out

[strong field]

they are actually really really good haha

[buoyant axle]

@indigo carbon ah i think i found at least one factor, what makes it slower...i am using the sdxl clip tenc because i am passing in a custom widht height (2048x2048 --> taken from sytans workflow). If i replace the tenc node with a simple one, i cant pass in custom sizes anymore though ...

[crisp owl]

From what I understand, it's supposed to apply Comfyui Metadata, but it's a new feature on civitai, and works terribly inconsistent. I see some people's posts with nodes applied (Comfyui Metadata), but I haven't gotten it to work yet either.

[hardy cipher]

I don't dislike people disliking it. that's fine. it's their trash attitudes and entitlement that gets me. but I guess that's the world we're in

[hasty smelt]

My first render using SDXL and also ComfyUI. Thanks guys for the support.

[strong field]

XY doesnt want to load SDXL, hmm, interesting

@upbeat summit

heres one of my tests, but cant get it to work with sdxl for some reason

[upbeat summit]

good plot! ok I will try to set it up shortly

[strong field]

interestingly enough AIT works with SD1.5 checkpoints lol

time to go hunting

this is what i get with sdxl

[buoyant axle]

I think because sdxl is not supported yet https://github.com/LucianoCirino/efficiency-nodes-comfyui/issues/69

[strong field]

ayyy thanks for the info, i didnt bother to look and i was about to start googling pooled output lol

[upbeat summit]

this... hmm.. makes me sad now 😉

but I can run it without refiner?! maybe?

[strong field]

i tried without refiner lol

didnt work

[buoyant axle]

but for simple basic stuff like samplers/schedulers/etc. StableswarmUI has a grid generator. U also can bind your custom comfyui workflow to it....works pretty well for me and i don't need to switch to a1111

[upbeat summit]

ah, thanks for the tip!

[strong field]

thats a good idea, do they already have a workflow built or do you just drop in a comfy?

[upbeat summit]

I need plots and grids - I can't help it heh

[strong field]

lol

i think i know why it wont work with sdxl, but i gotta go diving into this monstrosity

[buoyant axle]

You can use comfyui as backend (self started or your own instance via api). by default it uses a built in comfyui workflow, BUT you can provide your own workflow. the ui will adapt the parameters of your workflow. it may take a bit to get used to it, but i like it more and more and it has the benefit, that you can use your own comfyui workflow

[glad fulcrum]

How do you install AIT and is it compatible with RTX 2060 ?

[strong field]

I think thats SM75 and I think AIT is SM80 only

[strong field]

yeah id definitely rather try this than spelunking in efficiency_nodes.py

[spark sinew]

Error occurred when executing MiDaS Depth Approximation:

MiDaS_Depth_Approx.midas_approx() missing 1 required positional argument: 'image'>>> What is wrong?

[buoyant axle]

yeah try it. at least when you want to have x/y grids it is very handy. u still can use basic comfyui for your standard stuff or switch between them whenever u like. i think thats what i like so much about it ^^

[glad grove]

kanye goin crazy with the new yeezy's

[native knot]

Someone asked him to put out some frogskins.

[trim orbit]

https://www.reddit.com/r/StableDiffusion/comments/15oeymz/depth_sdxl_controlnet_coming_soon_brace_yourselves/ i've been braced for hours

like, how much time do you need to call brace for?

[glad grove]

[native knot]

Same prompt, different seed. I can still see Seagal here. 🤣

[urban fjord]

Auto1111 doesn't support ddim for SDXL, just use one sampler and stick with it, there is no point really to test lots of different samplers.

[tender timber]

I remember now! I wanted to add the date time the filename, but your time string did not include the three letter month abbreviations. If you add %Y-%b-%d %H:%M:%S to 'time_format', in line 257 of your Python script will it allow for the abbreviations?

[strong field]

Yeah I agree once you do the tests then no need to do them again

Maybe some step plots but def not samplers and schedules

[glad grove]

this one better imo

[dense chasm]

huggingface inference API developed by Rust run fast

[urban fjord]

Best Seagal frog I could make

[native knot]

I can still see it.

[upbeat summit]

trying out some science fiction stuff using NightVision XL by @delicate kelp

[heady vale]

that armour suit is awesome

[upbeat summit]

is this you? 🙂

[glad grove]

yes

[civic sigil]

Can someone explain how you do the two step workflow with img2img? Its not working for me for some reason

[upbeat summit]

let's put this in the living room

[hardy cipher]

indeed. these are interesting

[heady vale]

beautiful work

[hardy cipher]

might need some refinement

[hardy cipher]

thanks!

strange hodgepodge of loras and styles

those hands, lol

[hardy cipher]

divine steven seagal

[glad grove]

It can't be bargained with, it can't be reasoned with. It doesn't feel pity or remorse or fear and it absolutely will not stop

[hardy cipher]

like the real steven seagal

no hostages, no surrender

[gloomy barn]

Except when they roll out the buffet.

[hardy cipher]

https://tenor.com/view/steven-seagal-stare-look-serious-intense-gif-5002551

[gloomy barn]

You could hear it in his whispery voice. "Remember, I go first."

[hardy cipher]

he fills 3 plates in one go

[gloomy barn]

"It's four or nothing!"

[pure crystal]

@upbeat summit solo detailed (ukj:1.4) (sci fi goblin:1.2), upper body portrait photography, (sitting anthropomorphic creature by Jim Henson, sci fi rogue, Mos Eisley cantina in Star Wars (1977) by Jim Henson, bokeh, cafe booth:1.3), (masterpiece, highest quality, extreme resolution character art:1.3), (photon mapping, physically based rendering, global illumination, area light, indirect lighting, transparency, reflection:1.2)

[hardy cipher]

I like it when people for real go nuts over the buffet. that might be a quality theme for some images

buffet wars

battle at the buffet

I guess I should run that face detailer after some of these. although not sure how much it'd really help

wish I could figure out why I can't load all the models with it

[hardy cipher]

lots of options with this one

[gloomy barn]

Darth Seagal holds the lightsaber by the blade.

[hardy cipher]

that hypertension

[gloomy barn]

But Darth Seagal, that buffet was supposed to be for everyone. By eating so much, you've altered the portions.

Darth Seagal: Pray I do not alter them further.

Redness isn't always hypertension

[hardy cipher]

that looks real

[gloomy barn]

Help us Obi Wan Seagal. You're our only hope.

Judge Seagal. "You know what. I am the law."

[hardy cipher]

I like to believe that everything we make here comes to be in another universe

and that there's a steven seagal universe somewhere

[native knot]

The Seagalverse.

[hardy cipher]

yes, like when john malkovich went into his own mind in being john malkovich

[hardy cipher]

they just say "Seagal" to each other over and over.

[glad grove]

theres an universe where seagal is our earth president

[hardy cipher]

because that's the only word they need

[gloomy barn]

Judge Jedi: "I am....the force!"

[gloomy barn]

Steven Seagal went in for the role of Kingpin. They told him he could get it if he lost 100 LBS.

Palpatine: "At last, we have our revenge on the....WTF!!!!!????"

[slender coral]

Anyone know why I would be getting shit results?

[glad grove]

top 10 chewbacca quotes

[pure crystal]

my favorite is "burrrUrrrrr!!!"

[upbeat summit]

Thanks for sharing it! I remixed the prompt a bit and used NightVision XL as a model

[gloomy barn]

Rowwrrrrrrrrrrrrrrrrrrrrrrr reeerrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrr grrrrrrrrrrrr!

[pure crystal]

gorgeous, care to show prompt? I am going to try that model next

[gloomy barn]

Steven Seabacca

https://www.youtube.com/watch?v=2GME8rI4RPk

[upbeat summit]

of course!
hyperrealism photo portrait of a solo detailed (ukj:1.4) (sci fi goblin:1.2), upper body portrait photography, (sitting anthropomorphic creature by Jim Henson, sci fi rogue, Mos Eisley cantina in Star Wars (1977) by Jim Henson, bokeh, cafe booth:1.3), , face symmetry, intricate accurate details, cinematic color grading, cinematic, artstation, 8K

NightVision XL makes it so realistic - it's crazy

[pure crystal]

(solo:1.3), detailed bald (ukj:1.4), space jockey, (wearing square black spectacles:1.3), (upper body portrait character photography:1.2), facing the camera, crossed arms, (alien sci fi horror film 1987, by Donato Giancola and Anna Liwanag, ornate, hyper realistic, bokeh, alien mothership interior:1.3), (masterpiece, highest quality, cinematic still:1.3), (photon mapping, physically based rendering, global illumination, area light, indirect lighting, transparency:1.2),

[gloomy barn]

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

[glad grove]

we can rebuild him we have the technology

[pure crystal]

@upbeat summit your prompt, base sdxl

[upbeat summit]

pretty cool 🙂 the prompt build is very powerful

[strong field]

what model, what sampler/scheduler, steps? what aspect ratio/resolution

[slender coral]

Would all those apply to how there's multiple shitty arms?

/ legs

I just want some cool dinos/

There's multiple eyes on this...

[pure crystal]

@upbeat summit how to get nightvision xl? I see it referenced in DynaVision XL description

[strong field]

hmmm, might be a prompt thing

[upbeat summit]

It's not yet released. I'm currently testing it for SoCalGuitarist

[slender coral]

😢

prehistoric Brachiosaurus, by Daniel F Gerhartz, greg rutkowski, asher brown durand, Johan grenier, Anna Dittman

[gloomy barn]

Out loud laugh. 4 eyes are you!

[heady vale]

not quite Seagal likeness + inflated lora

[strong field]

this is what im getting

[slender coral]

What's the prompts?

[gloomy barn]

If I wrote a script to generate an image a few hours after I typed the prompt, it wouldn't be a prompt anymore

[strong field]

not sure which workflow you using but this is what im getting with your prompt exacly

[slender coral]

Feet are fucked up but at least theres 4 😄

Will test.

[strong field]

lol

[gloomy barn]

SDXtinct

[strong field]

you might have good luck with the Searge workflow, lots of good optimizations there

[slender coral]

ffs

[dusk mica]

Maybe add " bad arms" to negative?

[slender coral]

Thanks will give all those a try,

[gloomy barn]

T-Rex: "Very funny!"

[crude sparrow]

cool

[slender coral]

So, it takes about 3x longer to generate, I can live with that, but it's throwing an error forget exactly what it was but it's not able to remove the "metatdata", any ideas?

[hardy cipher]

is it trying to disable metadata? not sure what that is

namespace I mean

[slender coral]

Some sort of meta data flag, I don't know enough to know if its a property in the workflow or not.

[hardy cipher]

well hard to say without knowing what node it is

or what you're using

seems like a node issue to me

[slender coral]

Searge workflow

[primal vault]

you should update ComfyUI to the latest version

[hardy cipher]

I like that you low key have pipes in your node pack and I'm just now realizing it

[primal vault]

yeah, makes for a cleaner workflow

[hardy cipher]

pipes are the future. now just need someone to come up with collapsable group nodes. nested nodes are alright, but really doesn't do it for me. it'd be cool to have them all in place but then minimized into a node until you need access

[slender coral]

Thank you that was it.

[hardy cipher]

can't really complain as I'm not the one making these things, and so don't want to sound needy or ungrateful

[slender coral]

FINAL question, f

[primal vault]

I think I saw an extension that adds the ability to group parts of your graph into a single node

[hardy cipher]

nested nodes? or something else?

[primal vault]

yeah

[slender coral]

For Searge wf, do I have to use an image / how could I not use it as a prompt?

[primal vault]

you only need those for img2img and inpainting, in txt2img they are ignored

the readme file explains it well

[slender coral]

Okay I assumed since it looks completely diff than the outputs I was getting:

[slender coral]

Yeah it's just been a long day on my side will read tomorrow.

[hardy cipher]

really just waiting on that advanced sampler now. although I haven't quite figured out how to set all the parameters in those

[primal vault]

which sampler?

[hardy cipher]

advanced sampler. how to set up the start from and stop steps, as well as add noise and return with leftover noise

I kind of just wing it and it works. but might be better approaches

[primal vault]

ah, I see. I just made my own sampler for SDXL to take care of it

[hardy cipher]

yeah. I normally use that. but then if I want to add a second round of sampling is it ever reasonable to start the base steps after the refiner steps in the previous sampler? so say first one goes 30 steps, 24 base/6 refiner. is it ever a good idea to start another round of base at 30 steps and then refiner at 54?

hopefully that makes sense

[primal vault]

Not sure, just experiment with it. I tried so many different variations I can't remember anymore which ones were best. I just took the setup that made the best images and turned it into a node.

[hardy cipher]

that's fair. I do a lot of weird experimentation. now just need to step it up and start experimenting with making some nodes myself

[vale eagle]

I made a sampler for that

loop over multiple times

[hardy cipher]

yours are pretty on point though, I must say. they're my go to default nodes

[hardy cipher]

nice! which ones are yours?

[vale eagle]

sdxlmixsampler. For introduction https://civitai.com/models/108594/sdxlmixsampler-comfyui-jason-node-workflow-included

[hardy cipher]

I might have that. haven't had the time to really check out all the things yet

I really wish I could figure out how to get this sort of control over cfg using comfy. one thing I miss from a1111

I've tried to vary cfg using a cosine function, but kind of hard to really do with the refiner. not sure how that would work. and then not exactly sure how mimic cfg works either

[primal vault]

you could replicate it with nodes, but it can get messy. I'd just implement that in a node

[hardy cipher]

would an external node be able to have the sampler's cfg change over the course of the steps? or would it need to be implemented within the sampler?

[spring fulcrum]

that is awesome.. Did you use a different model or lora for that?

[heady vale]

default XL and prompt

[spring fulcrum]

wow that is awesome

[hardy cipher]

those hands

well anyway, I realize there's a lot more to it than just the cfg. should probably learn how mimic cfg works before trying to make anything myself

[primal vault]

And I really like the results that I can get if I prompt in a non-sensical way, like feeding unrelated and completely different prompts into each of the 2 clip models

[hardy cipher]

yes. those are nice

I'm a big fan of putting unrelated things together

[primal vault]

I tried thing like putting a description of a fantasy scene in the main prompt, prompting for the end of the universe in the secondary prompt and mixing in some disney style cyberpunk as style prompt

[hardy cipher]

I've been using sdxl prompt styler. use the same base prompt but feed it through different stylers for the G and the L

[ionic dragon]

@hardy cipher do you have any lora settings suggestions, all my loras haven't been through

[hardy cipher]

been through? not quite sure what you mean

in general my approach is to only add one lora at a time and then see how it impacts the overall structure of things

default is to start with the set last layer node and then the noise offset lora, then build from there

normally more trial and error than anything else. some things just don't work together. other things that seem like they shouldn't work together do well together. and once you get a feel for how a particular lora changes things you can guage if it's overbearing in a particular setup

[ionic dragon]

i mean, it didnt work out

[ionic dragon]

you suggest fidling with the settings?

[hardy cipher]

yeah, just add one at a time and feel out how it interacts with your model and prompt. and then add another when you're satisfied with the first. some of them vibe well together, others don't. you just have to feel it out and no one right answer really

[ionic dragon]

oh ok, but i want my lora to be perfect so that it can generate the face even using the simplest settings

@hardy cipher why do you have so many loras

[hardy cipher]

added them one by one to see how they worked out

I wouldn't suggest that many if you're going for photorealistic

[soft zealot]

you willl probably have a number of other similar errors relating to other bits and bobs as not all the "enablers" are turnd on. Even says on the next lline "ourour will be ignored"

Should still have run ok though

[hardy cipher]

I think I've inadvertently been using a 1.5 lora to make a bunch of these images, lol. but it works somehow?

[soft zealot]

its either working or (and more likely IMHO at this point) just being ignnored

[hardy cipher]

well getting errors like this

ERROR diffusion_model.output_blocks.5.1.transformer_blocks.0.attn2.to_v.weight shape '[640, 2048]' is invalid for input of size 983040
ERROR diffusion_model.output_blocks.5.1.transformer_blocks.0.attn2.to_out.0.weight shape '[640, 640]' is invalid for input of size 1638400

I'll drop the weight to 0 and see if it changes the output

[soft zealot]

The COmfyRoll LORA Loader has an "odd" switch 🙂

[hardy cipher]

ugh, still getting that error either way. doesn't seem to be ruining the output, but curious what's causing it

[soft zealot]

If its not doing anything to the output then its a warnning not an error (even though it says error) lol

[heady vale]

maybe because some loras are trained on 640 resolution?

[hardy cipher]

makes sense. I guess I'll just figure it out through trial and error. not sure how else to do it

[heady vale]

yup. I loaded up a lora made for 2.1 and got a whole screen of the same diffusion errors

[hasty smelt]

Guys, this is my first day with SDXL and ComfyUI. Everything is new to me since I was used to the 1111 interface. Sharing my first promising result. I can hardly believe my 2070 Super is doing all of this!

[heady vale]

this was using a tarot card lora for 2.1

[heady vale]

nice one

[hardy cipher]

starting to wonder if it's the model itself

[hasty smelt]

Thank you for encouraging me to use ComfyUI buddy!

[hardy cipher]

bypassed all but the noise offset lora and same error

[heady vale]

hahaa

[hardy cipher]

nice

I'll run another model and see what's up. if not that it's some kind of crazy mystery I guess

not the models, not the lora. ugh. wtf? I guess I could just ignore the errors. but I don't like taking that approach

[heady vale]

try removing all lora loaders except 1

like delete them

[hardy cipher]

lol

still the same thing. ugh. well whatever, I'll figure it out

[crisp owl]

geez mr 14 pack

[hardy cipher]

look how he eats

[crisp owl]

valid point

[hardy cipher]

FECTH DATA from: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/custom-node-list.json

I need to fix this typo

[hasty smelt]

This is going to be a lot of fun. Friends, currently I only have 4x_NMKD-Siax_200K as an upscaler model, do you recommend anything else?

[crisp owl]

That's my preferred one

[hardy cipher]

I use ultrasharp normally. is that one better?

[hasty smelt]

I not tested ultrasharp

[hardy cipher]

goodness, so many node updates. that's probably my issue

[crisp owl]

I tested a ton a while ago, just personally liked the siax one the best

[heady vale]

NMKD Siax is usually better , yep

[glad grove]

i tested 4x nmkd siax 200k on anime and it works there too,its a great upscaler

[hardy cipher]

always forget I need to fecth updates in order to see them

[hasty smelt]

Thanks guys, I'll download them to test

[hardy cipher]

I haven't explored them much. prior to using comfy really just used a1111 defaults. esrgan 4x, anime6x, esrgan5billion, etc

[glad grove]

theres also one called "1x_NMKDDetoon_97500_G" it removes the 2D or anime like effect from some pics

[hardy cipher]

these crazy names

[hasty smelt]

thanks guys for hard support

[hardy cipher]

is there an "update all" option I'm missing with node packs? or do I just have to click each one every 30 seconds?

[hardy cipher]

https://tenor.com/view/high-five-secret-handshake-gif-24763988

[autumn forum]

Btw I can see in your image that you might be using the 1.0 vae, I would use the 0.9. It doesn’t have those weird artifacts when you zoom in

[hardy cipher]

some of us might like weird artifacts

[autumn forum]

Ur right my bad😂 I’m just sayin as a helpful tip is all

[hardy cipher]

alright, countless nodes updated, very exciting

I like how the manager says update so I click update, and then it says it needs updated just like before

[hasty smelt]

thanks buddy

[hardy cipher]

you guys ever go to huggingface.com instead of huggingface.co?

a real feast for the eyes, let me tell you

not even sure where to get the 0.9 vae

[heady vale]

#✨｜sdxl message

[hardy cipher]

gracias

[soft zealot]

So as a demo to the question that was been asked somewhere at some point.......... @high skiff I believe is doing "research" on that.

Does changing the target sizes in th Text Condirioners make a difference?

I give you 3examples using same seed/promppt etc which from Left to Right are
a) a factor of x1 (ie same h&w as latent space)
b) a factor of x2
c) a factor of x4

Now to m y eye all 3 are pleasing (although I do have a slight preferecne for b) simply because the beer in the glass is a nicer colour)

[vale eagle]

As my observation, higher value would give the model more freedom to generate the image. Incorrect value would scale the latent to the actual image.

[rustic garnet]

just redownload SDXL. They provide it now with the 0.9 vae

[soft zealot]

"incorrect value" hmmm that almost makes it feel like saying " this is wrong" (and I'm not syaing thats what you are saying, Im just having an early morning not had a bacon & egg sandwich yet ramble) /

ILike many things related to SD I don't think there is such a thing as anyone being inherently wrong.

Yes to my eye there is a difference but again (to my eye) all 3 are acceptable

[rustic garnet]

no need to download and configure a separate vae

[peak dove]

Anybody using this at all? https://youtu.be/H5103u5uRII

[vale eagle]

Sorry for the misleading term, "Incorrect value" means empty latent image with w1024h832 and the target width and heights set as w4096h4096. The ratio of target width/height doesnt match with the output image.

[soft zealot]

and to that end I have now added a "Conditioner Scale" input to my workflow (attached) which sits in between the Height & Width Inputs to the Text Conditioners to ensure all Aspect Ratios are kept aligned tegardless of which scaling factor I choose to select 🙂

[vale eagle]

You might want to test more on non sequre image like 4:3, 3:4, etc. Sometime exact same value works better than 2x, 4x

[hardy cipher]

when you follow bing's advice on fixing an error,

luckily I know LLMs are psychopathic make believers so backed up the file beforehand

tbf I'm no longer getting any of those errors, so I guess it technically did what I asked it to

[upbeat summit]

it does look totally usable

[high skiff]

golden doodle using my LoRA

same prompt and seed in base SDXL

[peak dove]

Werf!

[hollow halo]

What is your LoRA trained on? Looks very much like base sdxl

I don't mean the composition but most other aspects of it

[high skiff]

Its trained on realism, and more specifically portrait photographs

I have several different epochs of it which affect the images differently, from small changes to realistic lighting to huge changes to composition

Base SDXL vs my super light realism LoRA

that LoRA is only 1/10th of the actual final LoRA

[soft zealot]

1/1th ?

Isnt that like 100 % ??

;o)

[high skiff]

har har lol

fixed lol

[soft zealot]

you could just have said 10% in the first place...............

[high skiff]

10% makes it sound like its the final LoRA at strength .1

its only 1/10th the actual training

[soft zealot]

so its 10% trained?

hehehehe

[high skiff]

basically, yeah

[soft zealot]

thing is its far more consistent once you get say a quarter of the way in to say "25%" than to have to siwitch units because you can't ereally xpress 1/4 as x/10

percentages are so much easier to work with, same as celsius and mm 🙂

[high skiff]

yes, but it also interferes with LoRA strength, which is also extremely important

[hollow halo]

If you can't 1.0 a LoRA it was trained badly, thats my opinion

[high skiff]

if I say "my LoRA at 10%"

People are gonna assume its the lora at only 10% strength

[high skiff]

this can be the case sometimes, but not others

[soft zealot]

says the man who lives ina country where we pay for fuel by the litre but express fuel consumption in miles per gallon (thats proper gallons not the smaller American ones)

bur where beer & milk can still be sold in pints

[hollow halo]

Don't get me wrong there are uses for other strenghts but the ideal training should have the LoRA "working" at a strength of 1, regardless of if its the desired results.

[high skiff]

the final Epoch 30 of my LoRA runs good at about 0.75 strength, but it still has a very different look that is beneficial over the properly fitted previous versions at 1.0

So its not really a one size fits all

its insane how different every 3 epochs of my LoRA look

also, the validation images are fucking hilarious lol

gotta love validation images lol

so goofy lol

[ionic dragon]

which one?

[high skiff]

I like the second one more cause of the contrast, but I think they could both use some work for realsitic skin and stuff

[soft zealot]

second one on the RH has been upscaled

[high skiff]

yeah

[hollow halo]

Spoiler as its a zombie with blood on it and missing skin, so click at your own discretion

[high skiff]

pretty dope

[hollow halo]

I trained a lora on zombies and found out sdxl already knows them pretty well lol

moral is, check your subjects, styles and concepts before training

[soft zealot]

Ive gone down the corporate rabbit hole lol

[shy kelp]

Just found out the gligen doesn't work with sdxl 😦 anyone got any other alternatives which allow to place certain objects at specific places for sdxl?

[hollow halo]

Photoshop? As far as AI though, we don't have segmentation models yet I don't think

[shy kelp]

Sad.. too lazy to use manual work so just gonna use area composition and latent diffusion for now as alternatives

[soft zealot]

technically I guess the woman is next to the man

[ionic dragon]

i just found some lora named faces so used it

[clever verge]

@high skiff Would be very intresting to know how your dataset is captioned.

[high skiff]

I am still experimenting with it

I have plans for another set of changes for my next training

[mellow tendon]

Where is this node type (seed With text) from?
Install missing custom nodes does not find it.

[clever verge]

Even if it's not final still get a glimpse of direction

[high skiff]

its honestly just like bulk tags

[clever verge]

Wish you could run SDXL in reverse, like BLIP

[soft zealot]

Let me give you a clue!!

[molten gull]

can you somehow call a comfyUI json file via python ?

[soft zealot]

doubt it as its not a slef contained script, needs the core comfy etc to run

[molten gull]

there's this here:

https://github.com/comfyanonymous/ComfyUI/blob/master/script_examples/basic_api_example.py

but i cant get it to run

[west breach]

did you enable api?

[hoary saddle]

I stopped using that one and now just use the web hook method listed there