✨｜sdxl

[shy kelp]

Finally made my first image with ComFyUI. What model the is equivalent to the bot here?

[stone fossil]

😉

[steel mist]

gdi hate it when you install new nodes and then comfy won't load anymore

[vital ermine]

welcome to hell

[steel mist]

nuked em from custom nodes and it's still bailing out. bleh.

[elder rose]

What's the error...

[steel mist]

I think one of the packages fucked around with the python install since I was not using venv for comfy

[elder rose]

Ah yeah protobufs

[vital ermine]

comfy doesn't use a venv on windows

reason I refuse to touch it in Linux as not using a virtual environment is an unpardonable sin for a dev.

[steel mist]

Is this fixable or should I just do a reinstall?

[stone fossil]

RD2D powah.

[elder rose]

You can try <path_to_embedded_python> -m pip install protobuf==3.20.3

[stone fossil]

Unleash the R2D2 power. 🙂

[steel mist]

phew -- that did it. Thanks!

[uncut fiber]

His indexfinger 😄

[vital ermine]

SD always has issues with stuff round behind someone

[glad grove]

we truly live in a society

[vital ermine]

lol

[hardy cipher]

he has the physique of a middle aged large body wwe wrestler

[glad grove]

perfect male body

[hardy cipher]

skin like a cooked hot dog

[wet nacelle]

@noble shoal

[noble shoal]

Art at it's fines 👌

[wet nacelle]

*finest

[glad grove]

real art

[noble shoal]

Try this in between your VAE Decode and save image node. Then you get a bit of the contrast back

[wet nacelle]

Thanks for the info.

[west breach]

when you accidentally run with a blank prompt

[glad grove]

my fav movie

[noble shoal]

Take 2? I didn't even saw the first one.

[hardy cipher]

this was the first one

[glad grove]

real steven seagal movie

[noble shoal]

Is that a toe on his hand? 😬

[vital ermine]

apparently Civit has broken NSFW filter

[noble shoal]

Oh boy, oh boy

[native knot]

Something is definitely afoot with that image.

[hardy cipher]

https://tenor.com/view/gotcha-wink-see-what-i-did-there-george-seinfeld-gif-4649116

[hardy cipher]

strange chest things happening there

this is probably the most powerful image I've ever made

[glad grove]

Sensei teachings are a pathway to many abilities some consider to be unnatural

[glad grove]

noodle fingers

new pet

[hardy cipher]

dream body

I would really like it if one day someone gathered up all these images and sent them to mr Seagal himself

it'd blow his mind, lol

[glad grove]

he'd prob break my fingers if i tried to do that 😔

[hardy cipher]

well he did kill a man

and doesn't seem too concerned about it

[glad grove]

his room looks like this,i dont wanna mess with him

[hardy cipher]

oh no

[glad grove]

those pink aliens look pregnant

[heady vale]

make a twitter account, post the images and @ him 😆

[hardy cipher]

isn't he a bounty hunter? he'll track me down

[low hill]

XL is wooooooooooooow!

[glad grove]

feedin my pet

[glad grove]

our newborn baby looks kinda weird

[native knot]

Grootcat.

[noble shoal]

Groot?

[vital ermine]

https://civitai.com/models/142995?modelVersionId=158622

I tried so damn hard to get that one trained in 2.1 but no dice after over a month of trying

[crisp owl]

2.1 had potential, but was missing the pizzazz

[vital ermine]

I loved 2.1 and never looked back but training it was a sob

It fought me like a bucking Bronco

[crisp owl]

I was heavy in 1.5, then right around when I started to play with 2.1 I started with Comfy and spent more of that time getting my 1.5 flow started, then SDXL came out

[vital ermine]

I have been here since 1.4 dropped and it became public. I did a lot on 1.5 but the minute 2.0 hit I never touched it again.

[native knot]

I've basically abandoned 1.5 at this point.

[vital ermine]

I did go back for aboutr a week a few months ago and I just can't stand how the models look for realism. Much better than they were but they all have that 1.5 look and feel

[native knot]

Exactly. You can completely tell when certain things are made in 1.5.

[vital ermine]

The one thing I so much missed from 1.5 is all those juicy camera control prompting words. For a photographer they were wonderful to have

[crisp owl]

Yeah 1.5 very much keeps a "look" to em

[vital ermine]

when the clip changed all those 400+ camera terms went POOF

[glad grove]

where was this 4k image made

[dim steeple]

Can someone please help me with this error?

modules.devices.NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

Using Auto1111

[native knot]

No idea...but what I'm saying is that some things really do have a "I was made in 1.5" look to them.

[vital ermine]

Could be a handful of things tbh. Start with which card do you have?

[dim steeple]

3080

[vital ermine]

was a lora/embedding used? Seen that if a wrong version was attempted as well as a wrong vae being used

[dim steeple]

a lora is being used yes

[glad grove]

try to remove the lora

[vital ermine]

yep

[dim steeple]

can I not use the lora then?

[vital ermine]

not if it is the issue

[dim steeple]

I was using it no problem yesterday same settings

just tested it without and same issue

Maybe restarted PC?

[vital ermine]

then your vae is it set to automatic or none or to a certain version?>

[dim steeple]

I am pretty sure these are the same settings that worked yesterday

[vital ermine]

make sure the vae is set right

[dim steeple]

it is automatic

[vital ermine]

switch it to none and see

[dim steeple]

okay testing

same error

I will try restarting everything

[vital ermine]

yep

I only ever got that from XL due to the vae

[dim steeple]

I got this error yesterday then restarted auto then it went away

I already tried this though

[vital ermine]

just realize auto is flaky as hell for xl

even 1.6

[dim steeple]

I am preparing to switch to comfy

[vital ermine]

it is a hard transistion but worth it

though a few things I miss

[dim steeple]

I am used to nodes because I work in 3d and blender

is that the main hard transition?

[vital ermine]

I despise blender and as far as nodes I use them in DaVinci Resolve and the nodes we have with the front end is about kindergarten level in comparison. Pretty rudimentary and clumsy.

My muscle memory still double clicks a link to remove it

[noble shoal]

100%. Fusion is the most confusing thing in Resolve imho

[vital ermine]

Fusion is a stand alone too and is better but still it can get hairy

I love how nodes will automatically connect and I need never go to them if I don't need the power but they are there if I do need them unlike in ComfyUI

[noble shoal]

Although i love node workflows and will never turn my back on Comfyui, i will try to avoid fusion as long as i can.

[vital ermine]

Hopefully someone takes up the challenge of making a new frontend

[glad grove]

me and my friend

[native knot]

What a handsome lad.

[lucid cradle]

can we genrate image here ??

[native knot]

#1072220168534642768

[shy kelp]

is there a comfyui node configuration which replicates the same settings as the discord bot?

[supple knot]

@hardy cipher This series reminds me of Contra on NES,the one Level 2 https://youtu.be/EH19efnxwwY?si=xgaAc0_lYHAJuZjj&t=100

[hardy cipher]

Yeah, using ipadapter with 8-bit/ 16-bit images and various other things. It's amazing how well a lot of them come out

[supple knot]

Can I use Ipadapter in Comfy ,I looking at it know ,But demo I found he doing it in A111

Oh its a cmd auto install

[hardy cipher]

Yeah. Just be patient with it because sometimes the parameters don't vibe well and it outputs a black box

[supple knot]

idk I like following a video

[hardy cipher]

It's such a new concept though, the tools will get better over time

Not sure about a video. I'm not really big on watching videos for such things. My approach is to connect things until they work

[supple knot]

What model does this work with ,both 1.5 and SDXL?

This github in Japanese is the right one?

[hardy cipher]

You should look in the manager if you use it. More straight forward

[supple knot]

Did you know Jean-Claude Van Damme is in new MK1 https://www.youtube.com/watch?v=JWcnLw5SXGI

[hardy cipher]

As far as tools

[supple knot]

ok, ill try

[hardy cipher]

I know which github you speak of though

[supple knot]

this is the video i found https://www.youtube.com/watch?v=GU9jBW6aQ20

[native knot]

So is Megan Fox

[hardy cipher]

Ugh. Those click bait titles and thumbnails. It might be informative. But I'm guessing it's 70 percent filler so they can fit more ads in

[glad grove]

our overlord

[glad grove]

he looks like elvis 🤣

[shy kelp]

I'd like to utilize Colab, but I'm frustrated with having to install everything each time. Is there a workaround for this issue?

[hardy cipher]

Not really. You can save models and things on Google Drive so you don't have to redownload them

[hardy cipher]

But colab doesn't save your python environment after you disconnect, so there's no way to avoid having to reinstall things

[shy kelp]

i mean the models

[prime bison]

I deleted everything except the Donald Trump ones, the rest were crap.

[half ivy]

I am here
What did i miss

[native knot]

Everything. Might as well pack it up and go home.

[half ivy]

1 week away and i legit feel like that

[hardy cipher]

uploaded new nodes. I hope the 3 people that end up trying them out find a use for them

hmm, I wonder how many people actually have downloaded the nodes I already had up

169 unique clones, woo

[half ivy]

what do teh said nodes do ser? was it the xy plot i saw o0

[hardy cipher]

I have a few up. but no xy plot

one that calculates aspect ratio stuff. a couple lora stack loaders, canny edge detector, encode and decode with vae model loader widget, and a couple nodes that modify conditioning pooled output values

random stuff I wanted to use myself and figured someone else might find a use for them

[half ivy]

i yet to get a working CN and i see you haz the edge detector..

[half ivy]

post processing blurry pics from XL, its a me thing

[hardy cipher]

the edge detector could use some fine tuning, but I just set it up so you can modify different parameters

[half ivy]

Its called a work flow 😂

[hardy cipher]

fixed the threshold values and have widgets to adjust other things

honestly it gives you more ability to finetune

but also if you do it wrong you'll get trash. you want the adaptive thresh around 5 or 6 normally, and gaussian blur ksize 3 minimum

[half ivy]

Got this one to work by changing a reddie principle, so making progress

[hardy cipher]

technically ksize and blocksize are only odd integer values, but since it appears to be impossible to have the widget only allow odd values I just set it to round up from even values

[half ivy]

Adaptive canny wts 😅

[hardy cipher]

it uses an algorithm to differentiate between foreground and background, at least theoretically

[half ivy]

by creating a blur around the edges cancelling out a background?

[hardy cipher]

not exactly

[half ivy]

can ignore if its too mcuh, just curio bro

[hardy cipher]

it just uses fancy math to make an educated guess on what's what. also, it's just part of a library I used.

I didn't write the algorithm

it's super fast though. like .05 seconds to run

in case you were trying to shave a couple hundredths of a second off your render time

[half ivy]

every bit counts 😂

[hardy cipher]

good lord

the conditioning multiplier is kind of fun though

capped it at 3x, could change that, but after about 3x things start getting very weird

might focus on the ipadapter stuff now, see if I can figure out any dumb stuff to put in a node for that

[crisp owl]

Borrowed your pose

[half ivy]

How rude!!

[crisp owl]

https://tenor.com/view/mine-now-steal-take-grab-mine-gif-13196166

[uncut fiber]

Stupid Decorative High quality image <- what a tag 😄

[half ivy]

wanna try make this room change to have this controlled in

[hardy cipher]

que?

[glad grove]

yes

[low hill]

You can try......

[half ivy]

but got this for now 😅 , will change up the control strengths i guess, maybe add another CN

[low hill]

I wanted to create a nice wallpaper for my monitors

[hardy cipher]

that's super clean

[low hill]

Thanks man!

It fit's nice on my Desktop! 😄

[low hill]

@hardy cipher

I need to make a large one, for 2 monitors!

[glad grove]

what model? this ai image look rly good

[low hill]

No that's my setup! IRL lol

Sorry to Confuse you!

[glad grove]

[hardy cipher]

very aesthetically pleasing musical things

[low hill]

It's my actual Home Studio

[crisp owl]

Yeah my desk is that clean also

[low hill]

I don't like desktop icons

[low hill]

I have a shortcut so I can popup all apps

[native knot]

There's always a bigger fish.

[low hill]

Don't feed the animals!

[glad grove]

speaking of pets

[native knot]

Evolution.

[low hill]

Small face!

Fit's the size

😄

[native knot]

The current workflow. I call it Soul Flow 2.0:

[low hill]

Wooooow

[native knot]

Most of it is currently bypassed as I don't need controlnets, upscaling, IPA, PPF, or masked 2nd pass at the moment. It's only got standard inputs, The Randomizer, 1 LoRA, and primary output activated.

Here's the piece that was in the screenshot...

[low hill]

Anyone? The material is great tho...but have no clue what I have created!

Alien Tech?

[native knot]

Oh, that's easy; it's a fluzendorphalikenthropicalutionatorisk.

[low hill]

Oh yeah.....off course!!!!!!!!!! Thanks man, I didn't think of that

[native knot]

Easy to forget.

🤣

[low hill]

It has a fan

[native knot]

Well of course...wouldn't want it to overheat.

[low hill]

probably rechargeable

[native knot]

Runs on solar.

[low hill]

Ah I c

You are well versed

[native knot]

Just don't plug it in the wrong way.

[low hill]

I also don't leave it lying around on my girlfriend's bedside table

You never know

[native knot]

Current shot of The Randomizer:

[low hill]

Will you release it on Civit?

[native knot]

The workflow?

[low hill]

Yeah

[native knot]

Probably not...it's really just for me to learn and try stuff.

[low hill]

I want to create Super Women too hehe

[native knot]

The prompts that The Randomizer makes are quite interesting sometimes. I've grown it each night.

She's a Superwoman with a sword.

[low hill]

Here is some of my work: https://civitai.com/user/pauluz

[native knot]

I wouldn't fsck around with her.

[low hill]

She will chop your head with Super Power

[native knot]

Nice...you can check out my galleries here:
https://lychee.soulctcher.net/

I highly recommend looking through Random Diffusions.

When I get up in the morning and get a chance to go through what it's making now, I'll upload another batch, but there's already over 1k images in that one album.

[low hill]

Is that a plex extension?

[native knot]

There's probably another 500 across the other albums.

No, not Plex.

[low hill]

What is it...?

[native knot]

It's called Lychee.

[low hill]

Is it cloud-based?

Or local

[native knot]

I've got a server.

[low hill]

Cool!

Online server?

paid?

Or Nas-like?

[native knot]

My own.

[low hill]

You must have a speedy up/dl

[native knot]

Proxmox VE running several LXCs and VMs.

Yeah...synchronous 1G.

[low hill]

Great stuff on your site...nice creations

[low hill]

I like this one

[low hill]

And a lot more

[native knot]

Gucci Gang

[low hill]

Do you sell your pics?

[native knot]

Nah...

[low hill]

I'm trying too

to sell my own

[native knot]

CC-BY-SA-4.0 license

[hardy cipher]

tbh I'm trying to sell soul's pics

[native knot]

haha

[low hill]

hahahahaha

[native knot]

BEANZ

[low hill]

A robot that eats! Nice

[hardy cipher]

beanz?

[native knot]

beanz

[hardy cipher]

I'm a bit of a beanz man myself

[native knot]

Sexiest Siamese twins right here, yo...

[low hill]

This one nice: https://lychee.soulctcher.net/gallery#MQ_puxaNJWQNQfuiZDHBny1s/IOU9mnZyB3Qzd1tPrUH0Bctv

[native knot]

Haven't gotten that bad of a diffusion in a while...

[low hill]

Hi-Res fix ignore!

[native knot]

lol

[low hill]

Anyone have this: https://civitai.com/models/4439/luisapmonster-maker

[low hill]

LOL

[native knot]

I should get to sleep.

[low hill]

Hand is sooo good

[hardy cipher]

beanz

[native knot]

BEANZ!

[glad grove]

dont u hate when your little brother interrups your workout routine

[low hill]

OMG! That bottocks

[hardy cipher]

yowzers

[low hill]

Creepy

[low hill]

hehehehehehe

I've been watching you!

[low hill]

I have seen those earlier

[low hill]

with and without clothes

[hardy cipher]

just their cousins

[uncut fiber]

arcimboldo @hardy cipher

[low hill]

hahahaha right! Acrimboldo

Giuseppe Arcimboldo

[hardy cipher]

indeed

[low hill]

Traffic controller?

Is out of control hahaha

[low hill]

Cooool!

[glad grove]

new baby comin out next week

[hardy cipher]

elvis seagal

[glad grove]

queen 💅

[hardy cipher]

whoa

[hardy cipher]

that last image I posted was just made. not quite sure how to feel about it

[low hill]

It will pass......

😄

They are still doing it

[low hill]

Go to see the dentist

Alien teachers

Class photo

Creating aliens/scenes was fun

[low hill]

Very nice model

[hardy cipher]

absolute unit

[supple knot]

that middle pictures size giant dude

[ionic dragon]

which depth map is better
Midas, Zoe or Leres?

[hardy cipher]

I'm a midas man

[supple knot]

up next star wars

[noble shoal]

Man, the class of 1980 looked really old. I think not all of them made it.

[crisp owl]

"And that's where I met your mother"

[supple knot]

they were all on the titanic

[worthy orbit]

trying some of that

[uncut fiber]

want to make Donald bubblegum

[hardy cipher]

https://tenor.com/view/chefs-kiss-french-chef-perfect-dish-excellent-food-perfection-gif-20341505

[hardy cipher]

I appreciate the hat

[supple knot]

Should have went to sleep ,whipping up a nightmare

[mossy canopy]

anyone has any upscale workflow for controllite blur?

[indigo carbon]

hey, is it possible to clean VRAM every time switching workflow?

[stone fossil]

Cats are cute. 🙂

[waxen cliff]

but that isn't a cat

it's a panda

[stone fossil]

Holy moly its not?

[waxen cliff]

look at it's black eyes clearly a panda

🤣

[stone fossil]

It purs so cute tho.

xD

[stone fossil]

Hes growing fast tho here he was only 2 here we are 30 min later and hes 7.

[stone fossil]

The down side is if he keeps going at this rate, he be dead in 2 hours.

[mossy canopy]

Finally sdxl controlnet LLite blur upscale workflow is ready..

[peak dove]

Here is a very interesting tutorial on Replicate and SDXL https://replicate.com/blog/fine-tune-sdxl

Using @cyan crown 's LoRA - African Tribal Dance

https://civitai.com/models/142352?modelVersionId=157839

[peak dove]

England Rugby Team (Aardman stylee) @cyan crown's LoRA

[peak dove]

@cyan crown's LoRA Barbie & Ken

[stone fossil]

Sometimes having more is rky less.

[bleak relic]

@stone fossil

[peak dove]

Ryan Gosling and Margot Robbie @cyan crown's LoRA

[uncut fiber]

trying to achieve printing effect, but cant. Any ideas?

[mossy canopy]

nice particle effect

[west breach]

photocopy. dot matrix?

[uncut fiber]

Thank you both, will try!!!

[peak dove]

Use newsprint style as part of prompt?

[uncut fiber]

will try! @peak dove

[peak dove]

newsprint style, half-tone

[mossy canopy]

DC One Million Superman Prime

[peak dove]

Onetime I had a colleague who made models for Superman movies - one was a house destroyed in a flood ...

[clever verge]

Pointillism style?

[peak dove]

This prompt - Interior of a space station cluttered with electronics and science experiments, Sci-fi, View of earth from the window, Atmosphere, Dramatic lighting, Epic composition, Low angle, Wide angle, by Miyazaki, Nausicaa Ghibli, Breath of The Wild

[uncut fiber]

@clever verge i resignet on it 🙂 I want something artificial. Not like pointilism style. I think it is simply limited by resolution

[clever verge]

You could go from that, it produces good dots. pointillism style {prompt} . composed entirely of small distinct dots of color.

[peak dove]

I will ask chatgpt to define the newsprint picture style ... Half-Tone Process: To address this challenge, newspapers use the half-tone process. This involves converting continuous-tone photographs into a series of dots of varying sizes. The darker areas of the image are represented by larger and denser dots, while lighter areas have smaller and sparser dots. This process simulates different shades of gray and creates the illusion of a continuous-tone image when viewed from a distance.

Grainy Appearance: The half-tone process, when applied to newsprint, results in a characteristic grainy or textured appearance in photographs. This graininess is a result of the dot pattern and the absorbent nature of the newsprint paper. The coarse texture adds a unique visual quality to newsprint photos.

[clever verge]

Halftone dots pattern?

[uncut fiber]

i tried this already. i will send some pictures, now i am somewhere else 🙂

[peak dove]

This is good - more pointillism than half-tone

[mossy canopy]

i need to see a pic to get the idea.. and this is pointilism

[clever verge]

Try this and see if it works:
{
"name": "halftone",
"prompt": "halftone pattern {prompt} . monochromatic, dots, screen printing, vintage, pop art, high contrast, dithering",
"negative_prompt": "gradient, full color, highly detailed, painting, illustration, realism, low contrast, smooth, textureless"
}

[peak dove]

Better, but more like an Epson printer than newsprint - but going in the right direction

[uncut fiber]

screen print tried as well. halftone pattern, and all i give up on it.
i am trying different, weird rounded eyes 🙂

It is ascii art included

[mossy canopy]

almost there but need black and white.. try zooming in

[noble shoal]

Like this effect?

[uncut fiber]

maybe, looks like it.

it is it

[noble shoal]

Ok, then a custom Node is probably the best way to achieve this.

[peak dove]

Best yet ...

[noble shoal]

Yeah, but only with Python Pillow

[uncut fiber]

or batch processing in krita, seems easiest for me 🙂

[stone fossil]

Full armor bullet proof.

[native knot]

None of you have the swagger of that guy.

[dapper dragon]

top dog

[mossy canopy]

a close up of a dog wearing a vest, featured on zbrush central, digital art, military carrier rig, official product photo, ripley scott, full body armor, 2 0 2 2, true realistic image, various posed, darpa, samurai, photo realistic”

[dapper dragon]

barbie girl style

Dog in the style of barbie girl by mattel, mattel barbie girl style

[native knot]

ICP's new look.

Now there's a mid-level videogame boss.

Two heads are better than one...and it would be some weird challenge to defeat two simultaneous strategies.

[shy kelp]

is there a node config in comfyui that replicates the same settings used by the discord bot?

[uncut fiber]

@shy kelp i doubt as if not wrong discord bot is still changed parameters and all.

[mossy canopy]

snowflakes macro shot

[uncut gull]

SDXL seems to me like some serious competition to artistic typography... Although this coming from an amateur who doesn't know the first thing about artistic typography.

[mossy canopy]

crystals macro shot

[native knot]

That kid in high school that "knew magic"...

[noble shoal]

@uncut fiber Are you using comfyui?

[native knot]

The deadly and dangerous skunkupine:

[peak dove]

Porcuskunk!!!

[twilit nexus]

how can i get my images to be more of platformer view

more something like this

https://img.itch.zone/aW1hZ2UvMTQ3NzQ3LzY5NDc5MS5wbmc=/original/JDDXju.png

everytime whatever prompts i give it always gives me isometric

[twilit nexus]

i want to get image to use them in my side scroller game so more like this is nice
https://img.itch.zone/aW1nLzU5MzkzMDIuanBn/original/Jn9LYr.jpg

[native knot]

Like this?

[mossy canopy]

a screen shot of a video game, pixel art, inspired by Alexis Grimou, trending on polycount, pixel art, anthropomorphic fox, birds and trees, 2 5 6 x 2 5 6 pixels, ancient wood environment, an anthro cat, loosely cropped, video game characters designs, art in the style of joshy sly, beautiful screenshot

[twilit nexus]

yes

[native knot]

Stuff like:

highly detailed, HDR, 4k, field of view```

[twilit nexus]

m generating objects

[mossy canopy]

do a side by side with my prompt

[native knot]

Well, you threw some other very stylistic words in there, so it'll probably end up being visually better. I was just giving them the quick route to a few terms that will point them in the right direction. 🙂

[twilit nexus]

what if i just want to generate things like chests only

or mountains only no backgrounds

[native knot]

But there's an example of your prompt, legion:

[mossy canopy]

i thought it would be worse..

[native knot]

lol

[native knot]

Just tell it the specific things you want and include a few visual descriptors for the item(s).

[mossy canopy]

whats wrong with you?!! u almost gave me heart attack.. just popped up out of nowhere..

[peak dove]

This prompt - beautiful woman with a purple mohawk, finely detailed inked line art, intricate complexity, rule of thirds, complementary colors, face by artgerm, dramatic lighting, smooth, denoise, unreal render

[noble shoal]

@uncut fiber https://civitai.com/models/143293?modelVersionId=158997

[native knot]

ooh...halftone

[peak dove]

TY d/loaded

[native knot]

halftone is perfect when you're rendering old comic book imagery

[noble shoal]

Try it out. Just put it together and couldn't play much with it.

[native knot]

Maybe later. 🙂

[native knot]

highly detailed, intricate```

[peak dove]

@cyan crown thanks for the great LoRA Shaun!!!

[native knot]

highly detailed, intricate, (solid black background:1.5)```

[twilit nexus]

do i have to add pixel art

[twilit nexus]

not a fan of tat style

i just like hand drawn things

[native knot]

You can do whatever you want. 🙂

I added it because you were talking about adding it to a game.

[wet nacelle]

sowwy

[twilit nexus]

it gave me this

i typed mountain

[native knot]

You should use a versatile model that's not only character focused.

[uncut fiber]

Thank you very much!!!

[twilit nexus]

what model ur using?

[native knot]

Currently, I'm using JuggernautXLv3

[uncut gull]

It really doesn't take much to get top-tier text graphics. Less than 4 hours for the large cover text, and not even an hour for these three spine variants.

[twilit nexus]

Link

[native knot]

https://civitai.com/models/133005/juggernaut-xl

It's highly versatile.

[uncut gull]

SiliconThaumaturgy just released his model comparison. 🙂 I've been using CrystalClear has been my best for ages, but I've downloaded his top recommendations and will test them later.
https://www.youtube.com/watch?v=Fr4-k9k9OMY

[peak dove]

D/loading now - thanks for the heads up!

[native knot]

CrystalClear is decent as well, but I find that I can get better results from Juggernaut

[uncut gull]

No, you definitely can't. I've tested it across more categories than ST has.
(Well, to clarify, technically individually good images can come out of Juggernaut, and individually bad images can come out of CrystalClear. So Juggernaut can outperform sometimes. CC just performs better more consistently.)

[peak dove]

These are good - but very pale?

[twilit nexus]

I am using sxz luma

https://civitai.com/models/25831?modelVersionId=127416

[vital ermine]

I guess you never figured out the issue with the lack of contrast? WHich program are you using?

[wet nacelle]

I actually like it a lot. It's all the Refiners fault too. I really like it though.

also lora

[vital ermine]

I know what is causing it as it just happened to a training of mine, and no, I don't like it so must retrain. glad it happened though.

[wet nacelle]

It's meant to look like crapy newpaper and it does it well with the refiner.

[native knot]

Considering the images I've been creating using a ton of different models, I can say with very good certainty that I definitely can get better overall results out of JuggernautXL v3. But this is art and subjective. That opinion is my own.

[twilit nexus]

As long I get good game art m good

[uncut gull]

Objectively: Does it include the subjects in your prompt? Are there glitches in the image (walls turning into sky, clothes fading into skin)? Do the hands have 4 fingers and 1 thumb? Do the eyes have pupils? etc. That's the kind of stuff we're talking about here. Style is totally subjective of course, these rankings are about usability and accuracy.

[native knot]

Well, first off, I wasn't talking about any "rankings". But yes, JuggernautXL's results do include the subjects in my prompt most of the time. There are occasions where I find something that a model can't render, but that's true for absolutely every model. As far as cohesiveness, even by that video's rankings, it sits in the same tier as CrystalClearXL, ranks better as resolution than it and at the same quality level.

And on top of that, they ranked v2, not v3. JuggernautXL v3 would likely do better than v2 on some of those tests they ran and it'll continue to get better.

[uncut gull]

I'm having Llama2 generate a list of prompts right now, so I can test the models he recommended and see what I get.

[native knot]

In any case, my comment was specifically about being able to get better results for what I want. 🙂

[leaden fable]

hello wich one should i take ? i installed the ones from huggingface but they all have those weird artifacts

[native knot]

You seem to have the one without the VAE fix.

You need 1.0 with the 0.9 VAE.

[floral island]

protovision all the way

love the versatility in there

[native knot]

Proto is good for sure

[floral island]

plus it deals with my bs prompts 😄

[native knot]

lol

[uncut gull]

But only valid if you test. 🙂 Otherwise, you're just pretending it's the best for what you want, without actually knowing.
(But, let's be real. I've got stuff to do, and I don't even know if you're using your own PC or how much HDD space / time you have. Do as you please, sir.)

[floral island]

silicon was right about the resolution though, 1280x1280 is just simply not feasable on protovision

you WILL get artifacts, a lot

then again, highres upscale, or controlnet bigger source images kinda fix that

although i've not ben able to do upscaling very well in A1111 using sdxl models

3090 oom's

[native knot]

Because it's my opinion, the only thing that matters is what I think, not what some objective test says. If I don't like the results of something, I switch models, settings, etc.

If I was trying to get some specific objective results to make a statement that was fact and not opinion, I'd certainly want to do what you're saying.

No, the difference is that I have tried the other models and do continue to use them.

[floral island]

this is what i mean, dealing with bullshit prompts. it gets me pretty images ❤️

[native knot]

That looks great.

[floral island]

the actual prompt: ```

[anime cell shading:0.2], [inspired by Naoko Takeuchi:0.2], BREAK I project a storm A whirlwind of gales
BREAK O fire that dwells in an unleashed heart. Soar up, put it on, now beyond the dawn Unleash the fire we started now If you soar and run tomorrow should change
BREAK for whom we. Follow this thought Cut through the darkness, under the sun until the day the light shines

Negative prompt: photo, realistic, fantasy art```

[vital ermine]

I can't get break to stop bleeding on XL

[native knot]

@uncut gull - No offense taken, btw. The objective here is what matters, and that's my point. If the objective were to do something specific, then your thoughts and approach would be more valid to the situation. But all I did was share my own opinion based on my own experiences. I've generated thousands of images in the past few days, most with JuggernautXL v3, but not all of them. I am definitely getting better results that are pleasing to me. Not everyone may feel the same, and that's ok. 🙂

[pallid topaz]

Sometimes my SDXL image changes at the end to grey like this in A1111. In the same batch the other pictures are fine. Someone can give me a hint?

[cyan crown]

prompt and model?

[uncut fiber]

@pallid topaz seems like difference between images. SImply weird

[vital ermine]

A new lora I am working on

[cyan crown]

good!

[uncut fiber]

how long it can take on 3070 to make lora for sdxl?

[cyan crown]

too much...I've a 4070 but use runpod

[uncut fiber]

o.k. rather not try it then 🙂

[vital ermine]

A 4090 on linux is 11-15m

double that for windows

[uncut fiber]

oh, so nor day probably

[vital ermine]

Oh, that is for 2400 steps

[uncut fiber]

Oh 8GB seems to be biggest issue

[vital ermine]

a 4k step took me about 20m on Linuc

17m actually

[cyan crown]

well I use a settings that 4090 takes 2 hours.....

[vital ermine]

Yeah, I struggle with 24GB

[vital ermine]

Jesus

[cyan crown]

wich Network Dim are u using?

[vital ermine]

64/64

[cyan crown]

I use 256

[vital ermine]

matters not for length of time taken

I can do 4/4 or 64/64 or 128/128 and same time

[cyan crown]

how?

[vital ermine]

just is

I was hoping by going smaller I could save time but nope

[cyan crown]

well there must be a settings if you 17 min and me 2hrs

on 4090

[vital ermine]

17m is 4k steps

[cyan crown]

...... 2hrs is 3000 steps

[vital ermine]

3k I did once and was about 15m

[vital ermine]

all times are Linux Ubuntu 22.04 in Windows 10 about 2x longer.

[cyan crown]

strange

can you share your josn ?

to compare and see difeerences

[vital ermine]

I bet I know where it is. scheduler?

[cyan crown]

adafactor

"lr_scheduler": "constant",
"lr_scheduler_args": "",
"lr_scheduler_num_cycles": "",
"lr_scheduler_power": "",
"lr_warmup": 0,

[vital ermine]

for DB that takes longer so about 27m in Linux 40-45 in windows. I use ADAFACTOR for it and Prodigy for all others but would use prodigy for DB only 24gb simply can't do it OOM

[cyan crown]

bf16

[vital ermine]

optimizer?

[cyan crown]

adafactor

[vital ermine]

that is a nasty one that never works for me with lora

I even tried for 6k steps and barely anything while in DB it works wonderfully well

[cyan crown]

"sample_sampler": "euler_a",

[vital ermine]

what BS are you using?

#1 turn off sample images. ditch it. #2 turn off logs. Ditch them

[cyan crown]

already turned off

[vital ermine]

then why did you post sample_sampler?

thought you meant you were using that, heh

[cyan crown]

no

[vital ermine]

good

[cyan crown]

BS 5

[vital ermine]

I played with BS for over a day back and forth and the best one for the fastest overall is BS8

I can do BS15/16 but too slow overall

less is way slow and more gets slower and slower

[cyan crown]

maybe it's that

i'll try 8 next time

thanks

[vital ermine]

BS1, for instance, is over an hr

[willow bane]

i used an expired credit card for free trial will it cut money from the new credit card if i did not cancel on time?

anyone knows?

[mossy canopy]

who played this game?

[mossy canopy]

cancel subscription

[cyan crown]

where......

[willow bane]

it is not subscription

it is trial for 7 daYS

[mossy canopy]

site?

[willow bane]

KAIBER

[cyan crown]

so you have a subscription even if there is a free trial

[willow bane]

so i should cancel that on 6th day

[cyan crown]

usually you cancel now on other site and you still have trial period

don't know there

[willow bane]

okk

i just want to test text to video ai thats why

i did not have a card

[cyan crown]

first 3 are very good, last 2 less

[vital ermine]

those last two were native so good

I hate this dataset since 2.0 as it always makes them with their arms up and I can't stop it

That pose

[cyan crown]

1 or 2 ?

[vital ermine]

It learns it so hard even though it shouldn't

2 has more detail

[cyan crown]

yes 1 is only upscaled with gigapixel AI. 2 is img2img 2x 0.5 noise

but wich do you prefer?

[vital ermine]

I am a technical junky so will always go for the one with more fine details.

Hell, the blaster ray is far superior in 2 than 1 alone

I ditched gigapixel as they abandoned it anyway.

never updated it etc... new thing for them is topaz ai

[cyan crown]

photo ai

yes I use for photography

[vital ermine]

I think the previous version of it was better too

photo ai but the latest one seems to have gone backwards for me

I use it on my training data whereas gigapixel would make it do weird things and codeformer was 100% DOA for training.

[graceful egret]

can anyone explain the difference between these 2 models? which should I be using?

[uncut fiber]

@cyan crown why that snow? Some sort of protection?

[cyan crown]

it's an unwanted effect

[uncut fiber]

o.k.

Nice turkey

[uncut fiber]

it reminds 50s US posters about modern kitchen and family 😄

[mossy canopy]

latent upscaler from city69?

with perlin pin

pin

[cyan crown]

no the lora training set has lot of them 🙂

[mossy canopy]

it had the similar artifacts when perlin noise didnt match in the workflow

[cyan crown]

in this case it is wanted, because Lora is for this kind of generations

so for "normal" pictures it can be a little messy

[mossy canopy]

ahh.. i see..

[mossy canopy]

best upscaling method in your opinion right now for comfy?

[cyan crown]

In comfy I use sarge workflow 4.1

in Auto1111 img2img

[mossy canopy]

this are simple workflows.. any advanced workflows with latent and noise?

[cyan crown]

sarge has noise

I set between 0.3 and 0.5 and it works great

[mossy canopy]

I know but they are not advanced noise.. afaik

[cyan crown]

and img2img of auto1111?

[uncut fiber]

speaking about upscaling, is there some offline personal comparison of images, as are those online with slider moving left and right? Asking offline because not sure how about for example some tv series or so.

[mossy canopy]

for video upscaling?

[uncut fiber]

yes, i need to know if worth.

[mossy canopy]

have you tried the video upscaler from topaz?

[cyan crown]

yes

[uncut fiber]

OT... sorry
i am trying comfui, with restoration and resizing models.

not worth of problems 😄

[mossy canopy]

with comfy you can use any upscaler models to upscale ur video..

no restrictions there

[uncut fiber]

i know, i am just trying restoration models as well

[quartz sequoia]

@marble dew @hard fractal so, its been awhile hasn't it, perhaps.. a bit long of awhile.. perhaps... some might say a lil bit too quiet.... so so so.. hows Sai txt2vid comin huh?

[half ivy]

whatv the actual

now auto has updated

[mossy canopy]

tried any colorization model?

[half ivy]

im behind on ere tang

[uncut fiber]

static picture and motion picture, it is very different, i mean to watch. picture vs video

@mossy canopy no have not. going check if any there

[mossy canopy]

https://openmodeldb.info/

[cyan crown]

any of you use sdxl for photo restoration?

[uncut fiber]

https://openmodeldb.info/?t=colorization

i know i am there

no 🙂 just models in comfy.

downloaded deoldify

also got my grandma and grandpa old photos to test it

[mossy canopy]

not sure how much consistency you'll get between frames

[uncut fiber]

oh this i will try for photos first

[vital ermine]

In windows. In linux about 9-10m

[native knot]

Peak internet picture here:

[hardy cipher]

Add narwhals and bacon and you'd have 2012 reddit

[uncut fiber]

pure evil

[vital ermine]

Linux

[vital ermine]

I think Windows may be a lost cause for training AI stuff 😦

[mossy canopy]

wsl?

[hardy cipher]

Well not too hard to use wsl or something

[vital ermine]

Dual boot so real linux with no sub layering involved

For DB I go from 1.x s/it to 1.x it/s. Jaw dropped

I wish I knew why. I mean 5-10% alright but 50-100%? something is wrong

[icy brook]

Middle one is best

[vital ermine]

Reminds me of a faded 1970s mag/weekly world news rag mag

[hardy cipher]

Large lad there

[vital ermine]

In all the right ways, apparently.

[icy brook]

Guinness Record photo - largest and smallest person in the world 1984.