✨｜sdxl

[boreal bough]

but im pretty sure that ascore examples will be added to the guide book, released by sai on v1.0 release - since its quite important to understand

[high skiff]

I'm feeling Painting, Realism, Anime, Arts and Crafts, digital art, and drawings/lineart

[boreal bough]

there might be value in waiting out the last 6 days, to avoid double research - since it will be easier to stack on official sai documentation

[dense chasm]

does anyone knows how to make a stylish QR Code on comfy?i know someone test it successfully on automatic with controlent just like this one

[high skiff]

I am not too sure how much I trust their research, which is unfortunate

Their settings aren't exactly the most consumer friendly

and also they are testing these settings in the vacuum of very locked down prompting and a basic img2img style refiner workflow

[tribal minnow]

I did something weird, and improved my image. i put the refined image back through base model only (as latent with full steps), so i guess that is like an image to image ( seemed to increase all the details, and the lighting effect), gonna have to play around with that some more.

[high skiff]

The fact that Joe said that DDIM is the most disliked of all of the samplers for SDXL, when it was the best from my testing is interesting

no idea, joe just said that DDIM was the lowest rated

[civic sigil]

Right is after? I feel like left looks better

[high skiff]

exactly, and they aren't using my workflow

[uneven dove]

ppl upvoted deformities lol

[tribal minnow]

yes right is after, the coloring is off becuase I have a lighting prompt (so its amplified gonna try without the lighting)

[urban fjord]

To really judge the different sampler you need to judge them under optimal settings and not random settings. Aren't you and the bot using different schedulers too Sytan?

[nimble heart]

almost like euler/ddim have the lowest floor and the highest ceiling, and the bot doing random everything means they aren't presented at their best usage.

[quasi remnant]

dpm++ 2m normal is what i ended up on, the non-deterministic stuff like 2s a/sde left me with leftover noise.

no matter how many steps i gave the refiner

ddim is fast, but i get better results with 60 steps of dpm++ 2m normal split between base/refiner

i haven't tried exponential

i switch to karras on more anime/cartoon gens since that's better for that though

[upbeat summit]

got any hint at all how you are achieving this fidelity ?
are you using an additional model or just a very fine tuned processing workflow?

your image has amazing quality nonetheless - impressive. you perfected the fidelity of my prompt here lol

with SD 1 and SD 2 I used one workflow: generating an initial low-res image and used that as guidance for a hires fix upscale and this got me high fidelity and mostly coherent outputs.

[high skiff]

exactly my thought

he also said the best was SDE

[dense chasm]

does sdxl has any controlnet models yet?

[high skiff]

no, we have said no for like 3 days now

[uneven dove]

i was disappointed in that image lol

[upbeat summit]

well - I like it very much 😉

[uneven dove]

it's just SDXL bein' SDXL 😛

[quasi remnant]

what's weird is that NVIDIA is like a leader in AI and makes tons of AI programs themselves & stuff; you'd think they'd be testing their studio drivers across many different AI apps

[uneven dove]

i shall get you the exact settings

[upbeat summit]

that would be amazing

[nimble heart]

SDE is my favorite if you're not refining. The moment you have that 2-stage refine like in your workflow its not great and even pretty bad at photoreal

[quasi remnant]

yes.

For base, SDE is amazing.

If you add refiner in

you can not use SDE

[upbeat summit]

yeah I noticed this too

[high skiff]

yeah lmao

Its insane how big of an issue they let through

[uneven dove]

Seed: !seed 368218058, !guidance 7.5, !settings guidance_rescale 0.0, !steps 40, !settings strength 0.6
Model: SDXL Base + Refiner
SDXL Refiner: On, !settings refiner_strength 0.5 (20/20), !settings refiner_guidance 7.5, !settings aesthetic_score 10.0, !settings negative_aesthetic_score 2.8
Resolution: 1152x768

[visual glade]

it's image data in a torch.tensor

[uneven dove]

make sure you have a batch size of 4. he's frame 3 or so

[upbeat summit]

Thank you! So I guess this is not a comfyui workflow ;D

[quasi remnant]

yeaaaa; shortly after i uploaded my config i had to make edit notes about using normal & non SDE X_X; but at least lots of testing got me to a point where i love my results now lol

[nimble heart]

So I'm guessing the stats Joe talks about are only if you're not refining using the return noise method.

[quasi remnant]

yea;

it'd probably be too computationally expensive/take a lot more time for them to pass it through a refiner in a bot test too

[high skiff]

@uneven doveOk, so I just figured out why I am having errors now

I switched to uniPC to run some tests lol

[uneven dove]

no, he hates code review and changes interfaces a lot. it's not a stable backend for integrating into something like my discord bot. that level of flexibility makes no sense for a targeted use case imo

[high skiff]

it doesn't like such low steps lol

[quasi remnant]

first attempt

[uneven dove]

@high skiff i spent like 4 hours using Euler before i realised that's why everything looked like trash

[upbeat summit]

all the images I've been posting were dpmpp_sde: base, upscale, refiner. but I'm also just experimenting

[urban fjord]

The bot has been using refiner in the past, just not now as they are thinking about getting rid of the refiner.

[uneven dove]

i use DDIM

[quasi remnant]

i really wish SDE worked; lol, it really does do better bases

[upbeat summit]

gotcha! great output quality for sure

[quasi remnant]

like incredibly better bases

[uneven dove]

ddim_uniform today

[nimble heart]

also I can't even get dpm++ normal to work at all with return noise. Its a garbaled mess.

[quasi remnant]

wat?

it works in my flow

[nimble heart]

look how they massacred my boy

[quasi remnant]

what's the prompt?

[upbeat summit]

too far

[nimble heart]

happens to every image

[quasi remnant]

that's what SDE does to me

[nimble heart]

bruh

[quasi remnant]

dpm++ 2m normal works fine

[nimble heart]

dpm++ 2m karras works but dpm++ 2m normal gives drunken rave

[upbeat summit]

yeah I tweaked SDE like step by step. very finicky

[boreal bough]

check out the image grid custom node - you should be able to take that code directly, and only make minor changes

[tribal minnow]

works well enough to get some more details and change things. im sure there is a better way (also didnt need full steps 12 worked)

[high skiff]

looks like we are smart but not smart lol

[civic sigil]

What is this I literally just finished my image grid node the past 5 minutes 😭

[civic sigil]

Yeah I guess right looks better here

[boreal bough]

https://civitai.com/models/31126/imagesgrid-comfy-plugin-xy-plot

[civic sigil]

Ohh its for visualisation

[uneven dove]

@high skiff how did Birch-san squeak some credit for our work out of there lmfao

[tribal minnow]

those are both updated.

[high skiff]

Wait what? lol

[uneven dove]

he's listed after you and i because he did something similar on his own 2 days ago

[boreal bough]

also, will be outdated soon, as the dev of the infinite image grid, which is by far the best, will be making a plugin for comfy

[high skiff]

interesting lol

[quasi remnant]

i think this is sfw ._. this is what i got for that prompt for instance

[uneven dove]

it was weird but he has some demo code for k-diff on his repo so it's not... crazy to link there lmao

[quasi remnant]

no noise left

[high skiff]

I don't think that NSFW personally

[uneven dove]

safe for WHOSE work lmfao

like being on this discord is already a workplace violation and then that pic comes up, you can not explain that to a boss

[quasi remnant]

lol

[upbeat summit]

haha

[uneven dove]

so many times i've been in a meeting and Sytan messages me something obscene and i'm like

[upbeat summit]

"it has no noise"

[quasi remnant]

it's still some kinda armor i think

just weirldy placed decals

[boreal bough]

https://sd.mcmonkey.org/megagrid/
if you scroll to the bottom of this page, there's an example of what a 'good' grid looks like.
which we'll be having in comfy (eventually soon)

[uneven dove]

ngl it seems like she's holding her dead sibling lmao

[shy kelp]

bro i just popped a bonr that sh*t def aint sfw

[uneven dove]

"oh bethany we miss you already"

[tribal minnow]

lol, hopefully just sleeping

[uneven dove]

naw

[upbeat summit]

SDE gang, rise up!

[swift hamlet]

how i add this bot in my server

[urban fjord]

The bot is only for use on this server.

[uneven dove]

you have to sacrifice something important to you on a nearby altar during a full moon

[upbeat summit]

so many people reacted to my ima.... oh wait

[boreal bough]

I AM at work. and that DID just come up. but I'm at homeoffice and safe today XD

[urban fjord]

That looks quite nice if it is implemented how I hope it is implemented.

[quasi remnant]

iz beautiful.

[quasi remnant]

didn't you ask this yesterday? the bot isn't a public bot

[west breach]

area seems to be the best latent upscale method, you can denoise less without that weird scratchy grid being left in the final image. so the final image is closer to the original

[uneven dove]

they ask it 6 hours at a time

they're a bot asking how to get the other bot

they have a server full of just other bots

it's pretty wild man

[quasi remnant]

i believe it

[uneven dove]

don't believe everything you read, kiddo

i made that up

[upbeat summit]

lies

[swift hamlet]

everyone steal my ideas

[boreal bough]

based on unconfirmed rumors - it will be on the level of idiotproof XD so I'm at least assuming it wont be hard to use

[upbeat summit]

someone stole my prompt and made british comedy prompts

[swift hamlet]

any subscription need to private bot?

[uneven dove]

someone stole my prompt and made british comedy prompts

help me

[swift hamlet]

me too

[nimble heart]

yea idk. Only thing I can think of is I don't start at 0 since I'm doing a low res pass first. Maybe it's just broken for that purpose?

[ionic dragon]

You can use the api

[uneven dove]

your starting timestep t has to equal the end timestep of first stage T

[swift hamlet]

how can i use api key?

[boreal bough]

accept that prompts arent safe. change how you prompt to storywriting. tell comfy a story. comfy sigh. then you get the story of chaz, the 30 year old, and all his life achievements, and post mortem achievements

[ionic dragon]

https://clipdrop.co/apis

[high skiff]

ok, I need some help peoples

some crowd sourcing

let me get this ready first, just a sec

[urban fjord]

Just subscribe to me for $5 a month and you can get prompts that can run your business, change your life and achieve ethernal youth.

• Half of AI twitter.

[upbeat summit]

@uneven dove you really pushed the positive ascore with that british prompt

[high skiff]

Ok, got it

I will be doing 3 rounds of this, just to see if there are any top 3 trends

[urban fjord]

Top middle on image 2.

[high skiff]

Good pick, I seem to agree with it, which has me surprised haha

[upbeat summit]

bottom mid - color and fidelity wise

[high skiff]

On which side?

[quasi remnant]

bottom right on left image, bottom middle on 2nd image imo

from a purely fidelity standpoint

[upbeat summit]

3rd row, image no 2

[high skiff]

Ok

Very interesting so far

Very very very interesting infact lol

[boreal bough]

first image top left is the best

[upbeat summit]

top row too much contrast / hdr look

[boreal bough]

looks are subjective

[upbeat summit]

sure

just my opinion 🙂

[high skiff]

I know, just really going for aesthetic preferences right now and detail

Let me write these votes down

[upbeat summit]

we can try to be analytical or just going for taste

[high skiff]

@uneven dove @west breach mind chiming in?

[shy kelp]

1# top right 2# top middle

[upbeat summit]

beside the tree through the roof 😉 top left has a very nice mood for sure

[boreal bough]

detail - all but the left row on first image getting a failing mark, for multiple critical failures
aesthetics - 1st bottom right, top middle - 2nd image would have contenders, but they are blurry, so I cant endorse them, since clearly the settings went overkill
realism - only 1st image top left, and left middle pass
all in, top left wins (not due to being that much better, but simply by avoiding critical failures)

[high skiff]

I wish I had a way to better share these images with you all

so you could see all of them at 1024x

I would have to send 17 images into this channel, which would be hella spam

I mean, I guess I can, cause they grid together not

Ok, I'll do that

[rose prism]

'/

}:

[west breach]

left image, hmm.. changed my mind so many times haha

[high skiff]

Actually here, I will remove the critical failures to make my job easier

[urban fjord]

The current bot on this server? I feel the opposite. I guess things have stabilized.

[west breach]

i like the bottom left and bottom right. the bottom right would be great if the house was less smudgy, but the environment overall is great. the house in the bottom left is sharp and you can see the stonework in the chimney, but the trees and mountain look a little too sharp

[high skiff]

from those 12 I just sent, please pick your favorite for the prompt I just listed

feel free to pixel peep

[upbeat summit]

if I look at it if I would do a color grading project here:
top row images have great fidelity in content and image depth. but the contrast is pretty strong. they have a processed look to it. I would raise the black levels, reduce contrast and raise mid-tones a bit. they have a style - for me reminiscent of HDR or sharp clarity.

my favorite image is bottom row no. 2 / in the middle. it looks flatter in dynamic range compared to others, more natural to me and it would give more options to process it since it does not have an overly processed look.

[high skiff]

please ignore those ._.

here, I will try this all again, just a moment

[uneven dove]

what picture i see none

lol

[high skiff]

I am redoing it cause it got all messy

just a sec

[uneven dove]

logan paul looking afraid in a forest

[upbeat summit]

we are describing air

[ionic dragon]

Logan Paul after his marriage

[uneven dove]

my ascore is always 10

is that bad?

[high skiff]

ok, I lowered it down to just 9 options after eliminating all images with coherence issues

[upbeat summit]

I've done some ascore 10 - 12 images. but right now I'm mostly at 3-5

[high skiff]

I will be generating a new seed to compare this time

[upbeat summit]

from my tests the ascores are really depended on the prompt / style

[shy kelp]

thats not how i remember that scene

i remember he was laughing and gagging

[upbeat summit]

did it involve japan?

[uneven dove]

yup

can't seem to make a happy logan paul

[ionic dragon]

Why ascore used?

[shy kelp]

its higher so its better

[visual glade]

I put 1000 as my ascore because I want max aesthetic

[upbeat summit]

over 9000

[high skiff]

ok, I narrowed it down to 8 samplers that didn't completely fail in order to direct compare

Will send when they are done, cause I am curious to see

not gonna say which ones they are :p

[uneven dove]

lol can you ever do anything without gallons of hype leading up to the moment of impact

[high skiff]

No :p

[uneven dove]

[high skiff]

I need people to know whats going on so they are prepared to vote, thats why

I already did this 2 times and it was a mess lol

[upbeat summit]

I'm hyped

well kinda

[uneven dove]

well i like the middle option

[upbeat summit]

mostly tired

[uneven dove]

(it's always the middle options)

[high skiff]

IT CRASHED

:C

I have to redo it

[upbeat summit]

gotta run your tastes through the refiner first

[visual glade]

this is 1000 aethetic

[upbeat summit]

looks pretty good

try 0.1

[visual glade]

yeah this is default comfyui prompt using only the refiner with 1000 ascore

[upbeat summit]

I also thought it worked a bit like aesthetic embeddings... so 0 - 15 or something and incoherence starting at 12-15

[uneven dove]

Journey back to the eerie 1970s Soviet era with "MindFrame++," an imaginary old TV show depicting a world plagued by uncanny phenomena. Set within the confines of the enigmatic Phantom Zone, the show follows a morally ambiguous investigator as they navigate gripping tales of unnatural encounters and illusory realities. Atmospheric visuals and hauntingly nostalgic backgrounds

[high skiff]

ok, pick!

make sure to pixel peep

[visual glade]

it's easy to see which type of images it considers to be aesthetic

[upbeat summit]

how many?

[high skiff]

1

pick the one you think fits the prompt:

"A cinematic photograph of a rustic log cabin up in the snowy mountains surrounded by a dense snowy pine forest with fluffy clouds at sunset"

[uneven dove]

#4

[high skiff]

I will be doing this 3 times with drastically different realism subjects to see if there are any trends

[upbeat summit]

So do I look for the best prompt interpretation, best coherence, best look or the image with the best technical aspects? or just my subjective favorite image? 🙂

[uneven dove]

you look for the one that's least likely to bring Cthulu's wrath upon you

[high skiff]

which image are you the happiest with the result from, looking at artifacts, details, and prompt comprehension

[autumn forum]

im not a big fan of the over saturated ones thats for sure. 😬

[high skiff]

thats fair, in that case, which is your favorite?

[visual glade]

only the refiner + 1000 ascore

[uneven dove]

nice watermark broo

[high skiff]

i had more supporting tags and a big negative (likely messed with the cinematic part with the negative, so ignore that part)

alright, so we have a vote for 4 and 8

[autumn forum]

3rd to last

so 8

[high skiff]

so that 4 votes for number 8?

sorry, 8 lmao

seems like we have a lot of people agreeing on number 8

[upbeat summit]

4

[high skiff]

you vote 4?

[upbeat summit]

no

look at my circle

[high skiff]

ok, thought you changed it lol

[upbeat summit]

it's 3 if you count backwards

[high skiff]

ok, looks like most people are voting for number 8, or the one you circled

interesting

[upbeat summit]

I think it has the less over processed look, the trees look good, the windows are coherent etc

[high skiff]

it starts 1-10

in the order you click and hit next to

[upbeat summit]

wouldn't that be 7?

[high skiff]

he means backwards lol

[upbeat summit]

ah ok so it would be 2

that's why i made a circle

[high skiff]

you guys are trying so hard to mess with my voting

[half ivy]

yo

[high skiff]

ok, so it looks like number 8 won from this small selection

[half ivy]

where can i find some custom node layouts?

[high skiff]

gonna get another prompt going

lol

[high skiff]

you pick 5?

[west breach]

yes

[half ivy]

bruh

[high skiff]

ok, we have our votes in

votes are
#4-1
#5-1
#8-4

[upbeat summit]

you should really put that in a function

[high skiff]

Captn, with all due respect, if you don't stop trying to throw me off with numbers, I will block you until I am done with this

I am trying to collect data, not make jokes ._.

[upbeat summit]

you got the results right, sytan

[steady grove]

so, for training sdxl loras, i don't use captions?

[upbeat summit]

@high skiff if you want any other feedback on images, let me know

[high skiff]

I will be doing 2 more rounds

3 rounds on 3 very different realism subjects

They will not be in the same order, to avoid bias

woah, I had an error

@visual glade

One of my results just spit out a very old image

oh, it was a visual bug, that was weird

it spit out the result of something i prompted hours ago lol

[upbeat summit]

yeah I've seen that on the sampler using taesd high quality previews

[autumn forum]

abandoned cabin!

[high skiff]

ok, generating the second round

[upbeat summit]

sometimes the browser view just doesn't refresh immediately - at least on preview nodes. might be firefox related though

[high skiff]

no like, it went back over 100 gens 😅

it was weird lol

[upbeat summit]

video frame buffer

[high skiff]

ok, i have them

Ok, please cat your votes again, from 1-10 in order of first to last

the prompt is:

A cinematic portrait photograph of a young ginger woman with freckles and green eyes on the beach at noon wearing a black dress blowing in the wind

wow, one of those images fried her lol

[floral island]

fried ginger, tasty!

[visual glade]

were you using the preview node?

[high skiff]

the save image node

[royal fern]

am I hallucinating or sdxl can't do freckles?

[high skiff]

SD has never been able to do freckles right lol

even the best 1.5 finetunes

[boreal bough]

??

[high skiff]

thats why i prompted it

[floral island]

i feel this is pretty good already tho...

[high skiff]

ok, number 4

[high skiff]

number 2

[autumn forum]

number 4 yeah

[boreal bough]

I always got overkill freckles on my gens in 1.5 x_x

[floral island]

this.

[boreal bough]

might be finetune though

realistic vision definitely did them right

[floral island]

inspired me to do this 😄

[royal fern]

I have the opposite problem, I often need to (freckles:0.8)

[upbeat summit]

there are some close seconds, but no. 4 is overall the best image to me

[high skiff]

@west breach @uneven dove anything to add to the vote?

[west breach]

last one

[high skiff]

last one, ok

[urban fjord]

For your latest batch everything is too similar for me to really pick a winner, even if there is a couple of losers.

[high skiff]

I think we can all agree that number 7 is terrible

overbaked to hell

ok, let me do one last generation

ok, looks like 4 won this one

which follows a small trend

[boreal bough]

"Diodotos the AI guru, 30 years old, giving you prompting guides that will run your life business dreams and let you achieve eternal youth"

forgot to post this earlier XD

[lusty raptor]

can't decide, 3,4, 10. leaning toward ten

[autumn forum]

definitely looks like he would sell me a ai prompting course lol

[ionic dragon]

@visual glade what does scheduler do?

[high skiff]

I'll put it down as 10

[visual glade]

scheduler is what decides the timesteps values the sampler uses to sample

[ionic dragon]

i didnt get you

[high skiff]

ok, final one

please pick 1-10 in order of first to last for which image you think did the prompt best:

"A national geographic nature documentary photograph of a Corgi in a field of flowers at twilight with sand dunes off in the distance"

[boreal bough]

sdxl - "cute 27 year old woman, with (freckles:0.3)"
literally had to tone it down to 0.3

I think it's cause you dont spend enough time in the base model, and the refiner is extremely biased when it comes to face, as proven by the blindfold demolition

[high skiff]

none of them did the sand dunes it seems lol

[urban fjord]

4th dog for me (If I counted right)

[high skiff]

this one?

[high skiff]

@lusty raptor@autumn forum@west breach@upbeat summit@uneven dove what do you guys think?

[high skiff]

ok, gotcha

[upbeat summit]

there are a couple of good ones, but this one won for me

[autumn forum]

was leaning towards 10 as well

[urban fjord]

Yes. 5th had too strong of a "HDR" effect sadly.

[high skiff]

I am gonna add my picks in as well

[upbeat summit]

yes, please do

[autumn forum]

kinda interesting if these were not randomized as far as different settings, we were kinda all over the place!

[west breach]

hmmm... 4 i reckon

[lusty raptor]

9

[upbeat summit]

true. I tried looking for: color -> fidelity -> coherence

[nimble heart]

9, 10, 5, 4, rest.

[high skiff]

alright, I think I have the info

[autumn forum]

great! so how did we do? a+ 100%? 😛

[high skiff]

Winner is DDIM

with DPM_Fast in second, and DPMPP_SDE_GPU and UNI_PC tied for 3rd

[upbeat summit]

interesting

[autumn forum]

yes very interesting

[high skiff]

but I will say, DDIM only won off of its basically unanimous vote for the cabin

[lusty raptor]

no way that's ddim, noooo wayyy

[high skiff]

if you guys pick images, I will tell you what sampler they were

Also

Winner of Cabin: DDIM
Winner of Ginger Woman: DDMPP_SDE_GPU
Winner of Corgi: DPM_FAST

[upbeat summit]

corgi's are fast

[nimble heart]

are the samplers being changed in a 2 stage or 1 stage

[upbeat summit]

which scheduler did you use for DDIM? normal?

[urban fjord]

What was 4 and 5 for the last one?

[upbeat summit]

ddmpp_sde_gpu - my favorite 😄

[floral island]

i prefer ddim/unipc any day: they generally get transparency things better than dpm type samplers, to my experience

[upbeat summit]

haven't really done much with unipc. only read the paper a little

[urban fjord]

I prefer Euler a as I'm too lazy to switch the default in automatic and it does an okay job. But with sdxl I guess it becomes ddim.

[autumn forum]

mixed the corgi with the cabin on accident lol

[upbeat summit]

beautiful image

[autumn forum]

it turned out amazing somehow haha, yeah one of my favorite gens from sdxl so far,, edit: also thank you

[high skiff]

sorry, I am back

[upbeat summit]

welcome back

[high skiff]

all of them were normal, all of the mixes are too much to test on my hardware

also, I should note, the sampler that won the corgi prompt is also the one that made that horrible image of the ginger girl

Euler (2)
Euler A
DPM_3
DPM_Fast (4) (ginger woman fail) (Winner of Corgi)
DPMPP_2S_A (1)
DPMPP_SDE
DPMPP_SDE_GPU (3) (Winner of Woman)
DDIM (5) (Winner of Cabin)
UNI-PC (3)
UNI-PC_BH2 (3)

I will be doing more tests against SDE GPU and DDIM

[royal fern]

how does setting a higher resolution width/target_width impact the final result? Has anyone figured out? I understand they are used as position/size conditioning, but why set a higher res than the actual latent?

[high skiff]

Also Euler and DDIM make almost identical images, which was interesting to me

[high skiff]

it makes a minute difference, but devs and Joe Penna both recommend just setting to 4096x4096 and leaving it

It makes a minute difference from my testing, honestly not enough to fret about

[royal fern]

so basically x4 the actual latent

(in case it's not square)

[high skiff]

no just 4096x4096 always

no math needed

[floral island]

@molten heart shouldn't euler/ddim/unipc make near identical images, given enough steps?

[high skiff]

no changes for base res

[royal fern]

ah thanksgod

[high skiff]

they did for what I tested

They were pretty close

[floral island]

i thought those were deterministic

[high skiff]

and they only had 20 steps

Euler and DDIM were always extremely close

[floral island]

35 is my golden number for steps, usually gets out just a bit more detail imho

[high skiff]

one is Euler and one is DDIM

thats how close they are lol

[royal fern]

is sdxl better with ancestrals?

[high skiff]

no from the test I just did

[upbeat summit]

no. 2 is slightly sharper

[autumn forum]

which one is the right?

[high skiff]

right is DDIM, its also faster

Which is the sampler I chose to base my whole workflow around

[ionic dragon]

35(base)+35(refiner)?

[high skiff]

¯_(ツ)_/¯

[royal fern]

I generally use DPMPP_2M karras

[high skiff]

I am down to test this as well

[royal fern]

with XL I find euler needs a few more steps while DPMPP_2M converges faster

[high skiff]

13/7 split vs 26/14

20 steps vs 40

[floral island]

35 in sd 1.5. For SDXL i use sytan's solution: 22+13

[high skiff]

oh, I can test that

22/13 vs 13/7

see if there really is something there, cause who knows

[floral island]

so that's 22 steps of resolve in base, and then 13 steps in refine

[ionic dragon]

@high skiff i am testing this prompt with different samplers and schedulers, so which -ve prompt do you suggest? should I go with this?disfigured, kitsch, ugly, oversaturated, grain, low-res, Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, ugly, disgusting, poorly drawn, childish, mutilated, mangled, old, surreal, text, blurry, b&w, monochrome, conjoined twins, multiple heads, extra legs, extra arms, fashion photos (collage:1.25), meme, deformed, elongated, twisted, fingers, strabismus, heterochromia, closed eyes, blurred, watermark

[maiden matrix]

SDXL is more controllable than sd1.5 model?

[high skiff]

I am not giving any advice on negative prompts, I have done like 0 research into them

I am sure that will be fine

[ionic dragon]

ok

[high skiff]

Wayyy so

[floral island]

responds WAY better to prompts

[maiden matrix]

i designed a product based on a lots of sd1.5 model

an important problem is, u can't control details, like material, color etc.

[high skiff]

ok, gonna test 20/35 steps for my workflow with DDIM

[maiden matrix]

it seems SDXL can't do that too

[floral island]

@maiden matrix it's hard to control details, as it just mashes stuff together into what it thinks to be a coherent image. you can't force it to include specific details 😦

[maiden matrix]

so, SDXL running in discord aim to collecting choice data for feedback and finetuning, just like what Midjourney's way?

[floral island]

yeah, rlhf finetune

but unlike midjourney, release model after

[maiden matrix]

im wondering if we can do that by multi-contolnet or something else

we're working on it

[maiden matrix]

what's differences ?

[high skiff]

one is 20 steps, one is 35

[floral island]

you can't run midjourney on your own pc, you can run SDXL on your own pc

[maiden matrix]

our product here:https://www.fabrie.com/ai/copilot

[high skiff]

one took 19 seconds, the other took 29

[royal fern]

I do that by combining conditionings... like "a woman with a white blouse" + "red hat" + "blue scarf"

at least on 1.5

haven't tried on XL

[floral island]

oh, does that actually work?!

[autumn forum]

i honestly cant tell lol

[high skiff]

SDXL is so monumentally more intellegent than 1.5, to where I can only imagine how far we can get SDXL fi we train it as well as we did for 1.5

[royal fern]

with comfy, by conditioning combine node, yes

[floral island]

nr1 the 35 steps?

oh wow, my pow winning prompt is so good in SDXL 😮

[maiden matrix]

so result as u want? what do u think about output?

[high skiff]

looks decent

[royal fern]

how do you mean?

[maiden matrix]

i mean correct color and materials, in correct place, all same as your prompt for it

[maiden matrix]

red hat , blue scarf , white blouse,

[high skiff]

not that hard to get from SDXL

[royal fern]

just tried

[quasi remnant]

i have totally been ignoring the anime potential of SDXL but it can do some pretty good looking things in 0.9

[high skiff]

yeeeuuupppp

[proud root]

a little flat, lacks character

[high skiff]

so like most anime

:p

[quasi remnant]

i like the variety

when using an anime model you always get the same looks

sdxl mixes it up with each look

[autumn forum]

the anime does lack veriety on sdxl but with finetunes itll be chefs kiss

[proud root]

lol tru, but even amonst anime it's producing rough images, but supposedly sdxl is highly trainable

[quasi remnant]

yup

[upbeat summit]

I agree and funny that you say that - just made this 😉

[high skiff]

it is

[quasi remnant]

yea i'd love to see like wd or some other big anime model trained on it

[high skiff]

I have some friends that are training it

[quasi remnant]

niiiice

[royal fern]

you can set other characteristics (like material or whatever) of course not just color @maiden matrix

it works to a certain extent

[floral island]

i feel it's not quite there against the better models, but as a base? VERY good

but sd 1.5 anime models have been baking for nearly a year now?

[high skiff]

exceptionally good honestly

[quasi remnant]

i mean vs gyoza & such it's not there, yeah, but with anime models you often start to get the same look across nearly every gen

(this is gyozamixv5)

[high skiff]

SDXL unfinished base is better than even some of the best SD 1.5 finetunes at most things

[maiden matrix]

community is important

[high skiff]

very much so

SAI is nothing without us

[floral island]

if you can't change the default look, the model is overfit

that's what i specifically killed in my own model mix, so it could do other things

[high skiff]

Overfit models are the best for achiving the best quality for a single thing, but of course they take up a lot of space

[autumn forum]

pretty good anime i must say. just needs more life. definitely lacks something

[civic sigil]

Yeah not super stylized

Looks kind of uninspired

But those eyes are so much better than what 1.5 would give

[high skiff]

I think thats pretty common for most anime

[civic sigil]

That pic doesnt look like most anime to me

[floral island]

add sad, lifeless in your neg prompt

[civic sigil]

You mean AI or in general?

[floral island]

this prompt still a bit too challenging for SDXL

[high skiff]

I just mean uninspired and not very stylized

Anime is generally pretty similar when in animation, but their art styles outside of direct animation are where they shine in diversity, like in Manga, where they are able to dedicate more time to getting better details in a single image rather than thousands of underbaked frames

typically is a single plane with some extremely rudamentary lighting when in anime, but in manga, you can get much more depth and style

[civic sigil]

I mean the limitations of anime encourage more stylization and less detail, the pic he posted is the exact opposite

So I dont see why you would draw that comparison

[upbeat summit]

try cinematic anime in pos
and as negative:
3d render, smooth, plastic, blurry, grainy, low-resolution, deep-fried, oversaturated

[high skiff]

I am not saying in regards to his image, I just mean that most anime is pretty low detail, its usually other mediums of Japanese illustration that have more stylized looks and detail to them

of course there are exceptions to that

but most anime can be swapped between each other with not much clash because of how animation in general is

its a money cow, where other mediums like manga and web cartoons usually can have a little more spice and creativity put into them because of the far more attainable frame counts

[civic sigil]

Yeah the type of stylization is very similar within the medium but the amount of stylization itself is high

[high skiff]

yeah, I can agree with that

[civic sigil]

As it would have to be, you cant draw too much detail per frame especiall back in the days of cel animation

[autumn forum]

the refiner keeps makin her sad lmao had to add happy to pos btw the 2nd image is just base, the refiner reallllly doesnt do anime well

[high skiff]

yeah, thats what I was getting at more so

[high skiff]

this level of over complicated visual noise would never be viable in an Anime without being too visually disorienting, or way too expensive to make

I do think there should be model distiction between Manga and Anime styles, cause thats not an anime style, thats like a visual nodel/manga style

anime fundementally needs to be much more... reasonable haha

[maiden matrix]

thx!

[floral island]

yeah, it seems SDXL somehow has really low confidence in anime-styled images

hmmm.... lemme try smth stupid!

[royal fern]

I kinda like it...

[high skiff]

hmm... I am doing natve 1400x1400 gens in SDXL, and they aren't bad

[floral island]

they're not falling apart in coherency?

[high skiff]

ok, at 1536x, it does lol

[ionic dragon]

is there a way to make filename_prefix dynamic, like can we make sure that it automatically picks up the "value" of the node, like for example if the node name is sampler the value is "euler" then the prefix should be "euler"
can we do something like that?
like in excel we use =A2, so can we specify in the filename_prefix something like =node("sampler").value?

[high skiff]

ok yeah, ist inconsistent lol

kept generating, got some weird ones lol

[floral island]

meanwhile, i got something esher-like

[high skiff]

this 1400x image came out mostly fine tho lol

[floral island]

they seem to be hit or miss

it got the junji ito part right 😮

[quasi remnant]

i think 1280 should be fine, usually I can get like 25% over the dataset trained image size and still be fine for any given model

after that, you can sometimes gen well otherwise it'll be a total fail

[sharp robin]

the best thing ab this

[upbeat summit]

SDXL 1920x1080 native - acceptable image ratio is a lot better compared to SD 2.1 but still it costs a lot of power to get good ones ;-). an upscale workflow makes more sense most of the time.

[sharp robin]

is that we get to do it all over again on the 18th

i hope they figure out not using a refiner

[quasi remnant]

but by then they'll publish official workflows i think

then can work off of those, provided they're substantially different from the stuff we've been doing.

[sharp robin]

it all boils down to get model, put text get picture in the end

[floral island]

plus or minus a few (actually a lot) variables

[royal fern]

mh conditioning combine doesn't seem to work with XL...

[floral island]

seems to work for me though?

you do combine both for base and refine?

[royal fern]

no, only one of the two

[floral island]

although i feel average works better in terms of consistency

combine breaks it into abstract... things

[royal fern]

as soon as I add "blue hair" everything turns blue

[soft zealot]

That has shades of the Citroen Centipede from the early 70's

https://www.carthrottle.com/post/citroen-centipede-10-wheeled-tyre-testing-monstrosity/

[royal fern]

let me try with both

it looks a bit random 😄

oh well after 100 tries it got it right 😄 (PS: I love this 90s-mangaish style)

[violet estuary]

quite creative the XL thou.

[solid zodiac]

I wish Clip Drop had an option for Negative Prompt, then I would think the images I try to generate would have better results.

[royal fern]

only downside now I'm tiling with VAE decode... 😛

[soft zealot]

agagagaga

OK switched back to W10 and changed the mouse settings to exactly match W11 and its still doingthe crazy zoom thing.

Cursed!!

[shy kelp]

An important unofficial update (from me) about using HiRes fix with SDXL: https://civitai.com/models/81540/comfyui-advanced-upscaler-workflow-sdxl-09-compatible

[sharp robin]

a part of me was rooting for you

[shy kelp]

my setup still works great, just not for multi-model generation

ill still be using it with great result with SD1.5 models

[dense chasm]

anybody knows how to generate a QR code with comfyUI?

[hollow halo]

Yeah I just trained a whole LoRA on this style, so all cars now look long and boxy like that

[high skiff]

man i wish I had the hardware to train LoRA's for SDXL

[hollow halo]

My 3090 can just barely handle BS1 with checkpointing, we need more optimizations or people won't make nearly as many models/LoRA's

It's still early though

[high skiff]

woah, what did you do that used that much VRAM? I don't thinky ou have things right

BS 1 should work on a 12GB GPU

[hollow halo]

Well I sometimes OOM because of the text encoder, without it I have no issues

[high skiff]

I have a friend who is doing LoRA's as well and on his 3090 he said it took just 11.6GB VRAM

[hollow halo]

I haven't actually checked the memory usage tbh, just was getting OOM so dialed it way back

[high skiff]

@boreal bough got any info on this?

[dim steeple]

When is it supposed to come out?

1.0

[hollow halo]

I keep hearing the 18th from people but who actually knows

[dim steeple]

Thanks

[high skiff]

18th

Its been said directly by SAI

[hollow halo]

They've lied before

[high skiff]

They are testing the 1.0 candidates in the bots right now

[dim steeple]

So they’re not calling it 3.0?

[high skiff]

no, its SDXL 1.0

[hollow halo]

This is completely different, 3.0 will come later

[high skiff]

yeah, 3.0 will still happen, talked to some devs about it

[dim steeple]

How are they different?

[floral island]

cat with a cheese popsicle? (sdxl failed the actual prompt)

[dim steeple]

It being a new name I feel like implies it has a different use case than the other “branch”

[high skiff]

I had this convo months ago, so this may be outdated, but I was told that all of the best things of SDXL will be taken and improved upon even more while being put into a package with less params (so it runs just as easy as 1.5 and 2.1 on older hardware), including the new if not even better data set to train on

[dim steeple]

Hmm so XL is sort of experimental then?

Then 3.0 will render it passe

[high skiff]

so I was told that SDXL is kinda just a stepping point to 3.0, which should be even better, but easier to run

So the XL part seems like a branch for models that the more hardware capable users can run

[floral island]

i think XL is just "SD but with A LOT MORE parameters"

and parameters take up vram, simple as that

[sharp robin]

Yeah, I hope this Refiner thing goes away if they can build it into the model.

[high skiff]

it is, and a ton of new technologies

[high skiff]

tho they did massively reduce the amount of VRAM they take up, which is insane to me

[dim steeple]

Thoughts on SD catching up to midjourney?

[floral island]

yeah ok 🙂

[high skiff]

there is no catching up to midjourney, they are not comparable in any sense

[dim steeple]

Why are they not comparable?

[high skiff]

Midjourney is for people who want ease and decent results with little effort

SD is for people who want fredom and granularity, and incredible images with high amounts of effort, while also having tools and features that MJ is not compatible with

[sharp robin]

if u look at results it basically already did

[hollow halo]

Apple vs Android basically

[dim steeple]

Yeah but you can still compare the outputs

[sharp robin]

u just need to find correct prompts

[high skiff]

not even close

I mean, i guess I see why you say that, nevermind

[dim steeple]

I agrée, or mac vs windows

[high skiff]

One is overpriced, locked down, pretty meh but reliable

the other is free, for the people, and massively more expansive and controllable, so I do see the analog there

[hollow halo]

Yeah that was my point

[dim steeple]

Feels to me that the raw output of Midjourney will be always at least slightly ahead

[high skiff]

nah, I can already make images in SDXL pre-release that blow MJ out of the water lol

[dim steeple]

Also they are making midjourney more contrôlable

[hollow halo]

This goes back to an argument I had with a friend the other day about how the iPhone camera is better than my technologically superior camera, and the root of the argument was just that it's easier to point and click and get a good photo with iPhone compared to most Android phones

[dim steeple]

Hmmmm examples?

[high skiff]

let me know when you can train it on subject you want and use all of the controlnet features and such

[high skiff]

Try getting a 2048x pixel image out of MJ at this quality level in less than a minute lol

[dim steeple]

Prompt?

[high skiff]

should be able to figure it our yourself lol

SDXL has multiple text encoders, MJ has one

[dim steeple]

I was just thinking it would be interesting to use the same prompt

[high skiff]

MJ can't use the same prompt

[dim steeple]

You ran this in XL?

[high skiff]

yes

[dim steeple]

I am asking what the prompt was

[high skiff]

this one as well

[high skiff]

I am not sharing it

and this one

[dim steeple]

A little weird but okay

[high skiff]

I'll share half the prompt for that corgi tho

you can try that in MJ

[dim steeple]

I was just thinking it would be a good experiment to run the Same prompt

[high skiff]

A cinematic photograph of a corgi sitting in a field of flowers with mountains off in the background at sunset with a cloudy sky

there, thats the part that MJ will work with

I just gave you the same prompt for the corgi

[sharp robin]

how can i figure out output nodes?

like what number they are?

[high skiff]

in what context?

[sharp robin]

Failed to validate prompt for output 19:

• ImageScaleBy 113:
  • Required input is missing: image

[high skiff]

I am not sure what you are asking

[sharp robin]

idk what that means

[high skiff]

oh, I am not sure

[high skiff]

oh wait, it should be giving you a purple flash on the node that failed

[sour obsidian]

just your prompt

[high skiff]

from MJ or from SDXL?

[sour obsidian]

XL 1.0

[high skiff]

ah, it looks rough for sure

assuming no refiner?

[sharp robin]

it shows up in the cmd window but lets me run things, but when this happens it makes my images dumb

[sour obsidian]

just pure model, not sure where you are seeing roughness lol

[high skiff]

These are not bad honestly

looking good

[dim steeple]

Yeah, out of curiosity do you use MJ?

[high skiff]

ther is static, and weird artifacts, and the flowers are deformed, and the dog doesn't look real

[high skiff]

not the newer versions

I use the older versions before MJ started being shitty

[dim steeple]

The owl one is literally just: an owl sitting on a tree in the snow

[high skiff]

@sour obsidiannot sure if this is just from how you copied the image, but there are issues all over it, weird stair stepping and interlacing artifacts it seems

[dim steeple]

The newer versions are really good

[high skiff]

that owl one looks super rough, but the corgi ones are solid, I'll give you that for sure

although, for just the base model, that does look better than 0.9

[sour obsidian]

thats no refiner first output on the base lol

[high skiff]

yeah, for no refiner, it does look better than the current base

I am not sure about all of those weird artifacts tho

I have never seen that from SDXL before

at least on my monitor, the image looks cooked AF

[dim steeple]

MJ just might not be great at owls

[high skiff]

weird grid like patterns all over the place

[floral island]

i see it too, you're not imaginging things 🙂

[high skiff]

yeah, I see some parts of it that look great, but other parts are just really woky with them

[floral island]

same as with the novelai vae -> it has this grid/raster effect which you cannot unsee once you've seen it

[dim steeple]

What do you think of this one

[high skiff]

I think those look worse

the way it does the eyes is really offputting

let me look at them all again

Like, they look good, very good even, but they don't look real

[sour obsidian]

im screwing with sampling a good bit

[high skiff]

One of these looks much more like it was taken with a camera

[high skiff]

taht one looks a little better, but the artifacts are still just as there

not sure whats up with that

like what is all of this

[floral island]

perhaps put anti-aliasing in neg prompt? it might work

[high skiff]

Yeah, i am not sure, thats a weird look

definitely artifacts I have not seen from 0.9

also curious to know how that model runs with my workflow, as I know SAI is not using it

[sour obsidian]

im doing very odd things to sampling atm, modifying noise gen, possibly that

[high skiff]

yeah maybe, other than those really bad artifacts, that definitely looks a lot better than 0.9 base

[sharp robin]

i am dense

[floral island]

anyone need pillls?

[sharp robin]

@high skiff i figured it out, was feeding refiner model into a place that the base model should have been

all good now

this ruined 1/2 my day

[high skiff]

@sour obsidianReally excited to see how my non standard diffusion method works with SDXL 1.0

It does wonders for 0.9

[sour obsidian]

got it avail?

[high skiff]

for comfy, yes

[sour obsidian]

im using comfy

[high skiff]

splits the single image diffusion between base and refiner, and it adds a lot more realism to the gens

[digital isle]

I don't know if it's the right section but I wanted to ask you, is there a way to create two characters with different clothes in sdxl with comfyui? Because I try to put for example a woman nurse and a man with a leather jacket and generating the image, they both appear with a leather jacket

[high skiff]

and then when paired with my new fractional offset timestep tests, its an even bigger difference

[sharp robin]

to

[high skiff]

oh lord lol

[sharp robin]

to this

after i fix the mistak

fml

smh

ROFL

hahaha

[dim steeple]

After some prompting with MJ

[high skiff]

again, they look good and detailed, but not realistic

They look like fake real owls, if that makes sense

[sour obsidian]

alr swapping to yours

[high skiff]

like, a photograph that looks like that is not possible

[high skiff]

ok good, the issues are gone

well, the fine artifacts

now the flowers look all messed up 😅

IDK, nothing I could adapt my workflow to without the model itself lol

[sour obsidian]

hey flowers are never perfect

its natural flowers

haha

[high skiff]

mine definitely look a lot less messed up lol

[sour obsidian]

dont you have like a whole 2nd half of a prompt sitting there haha

[high skiff]

oh, yeah

cinematic, fujifilm, national geographic, nature, snowy mountains, flowers, colorful, overcast sky, sunset

[sour obsidian]

ah yea

im just rolling first part, no negative
A cinematic photograph of a corgi sitting in a field of flowers with mountains off in the background at sunset with a cloudy sky

[high skiff]

first prompt goes left, second part goes right

and the negative is

just a sec

Deformed, unrealistic, bad quality, grainy, noisy, plastic, hazy, low contrast

so the full prompt is:

Linguistic: A cinematic photograph of a corgi sitting in a field of flowers with mountains off in the background at sunset with a cloudy sky

Tags: cinematic, fujifilm, national geographic, nature, snowy mountains, flowers, colorful, overcast sky, sunset

Negative: Deformed, unrealistic, bad quality, grainy, noisy, plastic, hazy, low contrast

[sour obsidian]

model is way much happier now to do detail without negatives from 0.9 -> 1.0

[high skiff]

thats good to hear

[sour obsidian]

still, its mostly unexplored, im sure you will all figure out ways to 10x it

[high skiff]

the 0.9 base model is rough without the refiner

[sour obsidian]

yea that was main goal was to make base good enough to be on its own detail wise, was my least fav part of 0.9

[high skiff]

Lookingh forward to getting my hands on the next major release early haha

Jeo said to talk to him and he'd make it happen 😅

[floral island]

i mean, look at the mileage that the community got out of finetuning sd1.5...

[high skiff]

Also told me I was too late to this release to secure research hardware

but I'll be ready next time :p

who knows, maybe I will discover something big enough to have diffusers rewrite part of their library again

[sour obsidian]

totally insane, that model looked like sandpaper originally

[high skiff]

1.5 was hot dog water on release lmao

[sharp robin]

I remember 1.4

[floral island]

still so happy i still have some of my og 1.3 prompts 😄

[high skiff]

I remember running 1.0, but I don't think I have any of the leftover images anymore lol

I sent 3 in this server!

man they were rough haha

mmmm yes rainy tesla

[sharp robin]

One of my first gens. We have come far.

[high skiff]

that sounds like a horrible idea

people are gonna abuse the hell out of it lol

[sharp robin]

That’s a discord issue

[sour obsidian]

we are considering open sourcing the bot code at some point but very likely never going to allow straight additions since its a research tool and not a service

[floral island]

same prompt 'anime elf in forest'

[high skiff]

yeah, thats what I was thinking

[floral island]

yeah... sdxl mildly outperforms sd1.x OG

[high skiff]

just a bit

no no no lol

[floral island]

OG 1.3 vs sdxl

[high skiff]

thats base 1.3 vs SDXL 0.9 lol

I am gonna run some 1.5 vs SDXL comparisons right now lol

[dim steeple]

Idk @high skiff I got some more for you

[high skiff]

lets see

[floral island]

but, can i get an e30 from SDXL 😮

[high skiff]

that first one looks considerably better, the others all still have that fake look to them

especially the second and last

[dim steeple]

Yeah just needed to add disposable camera

[high skiff]

interesting idea

[sour obsidian]

oh imma try using that

[high skiff]

same here lol

[floral island]

wait... SDXL got beaten by SD1.3?!?!?!?!

and not by a small margin

[dim steeple]

Idk this one is pretty nice imo

Looks like a nat géo pic

[high skiff]

it is nice, and it does look good, but it does not look real still

[floral island]

i think sytan bases of image artifacts

[high skiff]

that and the fact that that image would not be captured on a real camera

[floral island]

i mean, inspecting closely you can see something is unnatural about how the pixels relate to each other

[high skiff]

my whole look is supposed to mimic real camera features

[sour obsidian]

"if its not real there is no deal"

[high skiff]

I'll do it again

[sour obsidian]

just add "photorealistic" a bunch of times, that always works

[high skiff]

One of these looks considerably more probably for being shot on a real camera than th eother

[sharp robin]

((((((((((Photorealistic)))))))))))))

[dim steeple]

How about these

[high skiff]

even with the slight deformities in my image, the whoel look and scene is far more probably

[floral island]

oh my ❤️

[high skiff]

they are just random groups of branches with no tree, its just a mess of details that don't really lead to anything

I will have to say that MJ takes a loss on the owl prompt, but it did the corgi considerably better

[floral island]

dude, i grew up driving those!

[high skiff]

owld card

[floral island]

i've had all e30 models except the m series

[dim steeple]

The wood here is not realistic

[high skiff]

time to compare 1.5 to SDXL lol

[dim steeple]

So MJ 1 - SDXL 1

[floral island]

you do not want to know how much i had to wax the red-paint e30 i had

that paint was prone to going pink

[high skiff]

extremely overpriced and locked down paid service 1 - Completely free, open source, and non completed tool for the masses 1

Thats how I view it at least lol

[dim steeple]

I’m not against you I use both

[floral island]

"completely free" -> you just need a decent gpu 😄

[high skiff]

yeah, I see merit to both stylistically

[dim steeple]

I am just talking about which one can output the best image

[high skiff]

for somebody like me, that was a prerequisite :p

[floral island]

pre-SD, i bought a 3080, thought it was ok

[high skiff]

I already did things that needed a better GPU

[floral island]

then i went SD and got bothered by the 10gb issue

[sharp robin]

Sobs in 1080ti

[floral island]

then LLM's came around, and i just said fuggit, imma buy a used 3090

[hollow halo]

I happened to buy a 3090 for rendering right before I heard of SD

[high skiff]

time for me to pit 1.5 vs SDXL lol

[high skiff]

I bought my 3060ti for rendering before as well

[dim steeple]

The best owl I’ve ever seen

[high skiff]

oh thats silly and I like it haha

[floral island]

ngl, getting output like that is quite a challenge at times!

[dim steeple]

badly drawn crayon drawing of an owl, drawn by a 3-year-old toddler

[high skiff]

IDK guys, I think 1.5 might have won this one

[floral island]

needs more epic!

[sour obsidian]

which is which, cant tell

[floral island]

kek

[high skiff]

right? lol

[high skiff]

damn, it got the kid lmao

[dim steeple]

Lol nice

[high skiff]

sure it can, just need to know what you are doing

well thats an ethical dilemma

[dim steeple]

@high skiff were tied what’s the next shot

I think the owl one was a tie though

[high skiff]

I am not really here to try and compare which is better, we both have very different experience with our tools

[shy kelp]

dall-e 2 seem to fall quite behind. no news from the devs?

hahaha i dont think so

[sharp robin]

An owl face on a bears body.

[dim steeple]

They’re both kinda bad tbh

[high skiff]

with all due respect, I think thats just user error lol

good prompting and a special workflow I have researched for dozens of hours for that specific purpose lol

[dense chasm]

scribble controlnet is much more realistic Imao

[shy kelp]

looks like such a good boy. the dall-e 2 looks like a bad bad dog

[high skiff]

1.5 or SDXL, the world may never know

[dim steeple]

I thought u only can use 1.0 using the bots here

[high skiff]

yeah, only for 1.0

[dim steeple]

Is the corgi .9?

[floral island]

obviously sdxl is missing more legs, clearly the inferior product

[high skiff]

lol

anybody have anything for realism they want me to compare between 1.5 base and SDXL? lol

[sour obsidian]

more legs, more better

[floral island]

dude, that's just not needed man

[sharp robin]

An owls head on a bears body in the jungle

[high skiff]

I will nuke you lol

[high skiff]

I am talking about real life realism, I am still experimenting with fantasy realism (I do have some working tho)

fantasy realism prompt I was messing with for a demigod of plants

SDXL doesn't know pirates, but I can try! haha

[sharp robin]

Fine give me a banana tree.

[high skiff]

get rwkt

mispelled on purpose

[dim steeple]

@high skiff i dont understand how you can prompt for anything other than a basic prompt in 1.0?

[high skiff]

I am not using 1.0

I am using 0.9

[dim steeple]

oh i see

[ionic dragon]

@high skiff sorry for asking similar questions.
how do to optimize/pick the right keywords for negative prompt?

[high skiff]

again, I have no method for negatives, just do shit and stick with what works good haha

[sharp robin]

Start w empty then add what u don’t want to see. The freedom SD gives is experimentation

There is no right or wrong way.

[high skiff]

yeah, i have a method for everything else, but that is just pure goofing lol

[dim steeple]

real life?

[high skiff]

the foliage on those look dope

[floral island]

sdxl fails hard on "origami owl"? i'm clearly doing something wrong

[ionic dragon]

thats tru, but currently i am generating multiple images with all sorts of samplers and schedulers

so should i continue without using -ve prompt?

[high skiff]

oh yeah, I have not messed with oragami yet lol

the text above literally says 1.5 lol

ok, thank goodness you are joking lol

[high skiff]

Yeah, I must say, those look amazing

[ionic dragon]

@high skiff do you have a list of prompts?
i am currently generating images with different combinations of samplers and schedulers

[high skiff]

no, i do not

[ionic dragon]

with different styles of images

[dim steeple]

@high skiff best one yet

[shy kelp]

how you upscaled these?

[high skiff]

They are generating them in midjourney

[shy kelp]

whats the max res for midjourney?

ive never seen a happier corgi 😂

[plain flame]

Which is which ? SDXL and 1.5

[high skiff]

Jack and jack talking to each other lol

[plain flame]

Ohk. So till current date SDXL seems to be the best of all the image generation tools. Right?

[high skiff]

ok, now you are going too far lol

[plain flame]

All caps JACK talking to single cap Jack to be precise 😛😆

[high skiff]

Jack, and his louder brother JACK

I don't have to look again, I know you don't know how to use SDXL too well lol

[dim steeple]

careful

[high skiff]

for the masses, yes

[dim steeple]

i disagree

[high skiff]

Yes, I know, its my prompt lol

[plain flame]

For pro ?

[high skiff]

MJ is very expensive, SDXL costs nothing, MJ is not for the masses, its for people who are willing to fork over dozens of dollars a month

[high skiff]

your milage may vary in that case

SD can do specific things far better than MJ, but MJ can do a lot

its just expensive

[dim steeple]

if there was a blind compeitition where average people rated an mj vs sdxl image, i would bet on mj winning

[dim steeple]

you can use mj for free

as someone said before you need a nice GPU for SD

same with mj

[plain flame]

Free? Like how?

[high skiff]

you can gen like 10 images with it before it tells you to pay

[dim steeple]

25 and then it starts at what 10 bucks

vs a 600$ GPU

[high skiff]

yeah, 10 bucks for like 400 images worth of generating

I go through well over a thousand a day lol

sometimes several thousand

[dim steeple]

no unlimited

[high skiff]

no, its not

[dim steeple]

well unlimited for 30

[high skiff]

you get 3.3 hours of GPU time for $10

[dim steeple]

yeah my b

[high skiff]

unlimited for *$30

[dim steeple]

idk 30 bucks for unlimited jobs

[high skiff]

and thats unlimited slow jobs

[dim steeple]

vs a graphics card seems at the very least the same

price wise

plus you can use mj on your phone

i know theres more control with sd and thats great

like i said i use both

i was just answering the guys question with my opinion

[plain flame]

Clipdrop SDXL demo gives 400 images free per day

[ionic dragon]

what the size of your outputs folder?

[dim steeple]

but the quality is way worse shhh

[high skiff]

oh man, my 1.5 folder is well over 100k

[plain flame]

Yes it gives low rez images and can't set any aspect ratios too

[dim steeple]

what am i doing getting into a mj vs sd debate on the sd server lol

[ionic dragon]

so 100GB

[high skiff]

even so, thats 400 images you pay nothing for per day

Which is 12,000 images a month for free

on MJ you pay $10 for 400 images for a whole month

[dim steeple]

or 30 for unlimited

[high skiff]

98.9GB to be exact haha

[ionic dragon]

is it $10 for 400 generations or 400 images?
as mj gives 4 images per gen

[high skiff]

unlimited slow gens that take a considerable amount of time to complete

[dim steeple]

if we are talking about the best image generator purely based on quality image it is stil MJ by a healthy lead IMO, plus 6.0 is slated for a few weeks with huge improvements all around

[high skiff]

400 generations in that case I believe

[dim steeple]

about 20 secs max

[high skiff]

Oh, its way faster now? They managed to cut their times by a full order of magnitude?

[dim steeple]

yeah it depends on how many cocurent ones you have but if its under 3 at a time yes

[ionic dragon]

so 1600, but sometimes a single gen doesn't give desired images

[dim steeple]

you can run 12 concurent jobs too which is nice

[high skiff]

exactly, and thats 1600 for the whole month for fast ones

[high skiff]

10, not 12, but yes

[dim steeple]

"but sometimes a single gen doesn't give desired images"?

sure theres pros and cons to the pricing etc. The one thing i can confidently say is that quality wise mj is a few months ahead of SDXL

[high skiff]

oh also, you have to pay $60 a month for other people to not be able to take what you generate, which I think is insane

[ionic dragon]

mj is only for random gens
sd is very customizable, you feel like a creator when generating through sd(especially comfy). you dont feel guilty to showcase it, you completely own it

[shy kelp]

from what I heard mj barely runs on an a100, doesnt xl run on 8gbs?

[high skiff]

XL can run on 6

hell, I think even some people said 4GB with aggressive offloading

[dim steeple]

SD is great, MJ is just better at generating images of high quality period

i've used both extensivly

[high skiff]

ok, now you are going too far, cause you are suggesting that no SD workflow will match MJ, which is just flat out false

[shy kelp]

they have gone back and forth a bit, who knows what will be in a month or even a week

[high skiff]

if you wanna rephrase that to "MJ generates better images on average for minimal work", then sure

[ionic dragon]

mj barely runs on an a100
i dont think thats true, as if thats the case they should have invested billions. its self funded i dont think the info is correct.

[shy kelp]

the devs literally said it in office hours

[dim steeple]

idk prove me wrong?

[dim steeple]

can you rephrase that

[high skiff]

All it takes is generating something MJ can't do, which in and of itself contradicts your statement lol

[ionic dragon]

for real?
then how are they profitable?
is there algo too inefficient?

[dim steeple]

do you mean will not

[high skiff]

They are profitable cause they charge people $60 to have access to generate images that the masses can't take

[shy kelp]

I mean the service costs a ton, a lot of ppl use it, and at scale its prob not too much to have that many avail but still. its like 10x requirements to run the thing

[high skiff]

MJ is good at making images, but that doesn't mean everything else thats wrong with them goes out the window

MJ makes money hand over fist

[dim steeple]

if SD could do what i need it to quality wise i would still be using it

I was forced to start using MJ

[high skiff]

hell, on Value.Ai you can rent a PC with 48 cores, 128GB RAM, and 8x3090's for like $2/h

[shy kelp]

I mean it just kinda makes up its own stuff all the time, very opinionated which honestly im not too big a fan of