🏞｜general-with-images

[novel nest]

The real robot, filled with dazzling feathers, has fierce god-ray eyes and vomits rainbows and galaxies, a little evil dragon, wicked to the extreme.

[jovial tiger]

mixtral local did this one, so there might be hope...

[south temple]

rendering 🙂

[south temple]

roodoge

[crisp stream]

what?? 😃

[south temple]

kangy

doggy

[crisp stream]

ah, sry, yes 😀

[south temple]

you know the diff between rife 47 and 49?

[crisp stream]

2

[hallow otter]

please generate logo on white background for Mental Health Union for Women

[south temple]

my dude is thinkin

[inner badger]

Promotional poster to protect marine life

[glossy island]

sd

[languid pebble]

You are funny as hell 😄

[dull temple]

sb

[reef wadi]

https://tenor.com/bUqMB.gif

[wispy nest]

SMH. "AI makes everything so easy"

[floral wind]

#🏞｜general-with-images Please help me generate some cute Chinese alligators playing in the dock side cartoon image ip

[clever oar]

can somebody know how install manualy image_encoder for ip adapter in a111

[timid vector]

Superwide angle photograph of (a wooden ornate antique animatronic funky musician machine with a mechanical guitar player automaton:1.15) at a carnival, in the style of zoltar automaton has a very uncanny valley look and porcelain skin, automaton is depicted wearing a vibrant steampunk costume including a headscarf adorned with rich patterns and beads, a dark emerald green jacket with golden trimmings, a gleaming golden belt, automaton's eyes are accentuated with punk makeup, suggesting depth and wisdom, the automaton gazes forward with an enigmatic expression, automaton's hands rest the steampunk guitar covered in an ornate gears and shiny bolts and screws raised, surrounding scene is that of a eerie carnival

[wise raft]

An astronaut standing in the middle of a mix color of flowers field

[quiet current]

daily/hourly reminder that bot is down, cf #1047610792226340935

[small cloak]

Help me generate a photo of James playing in Beijing Dongdan

[wispy nest]

Realistic studio shot of James playing I. Begjing city scape scene photo realistic full render complete

Add modifications as needed

James is too vague you need a name or tv show for reference that will just give you random images not consistent

[small cloak]

LeBron James

[nimble mason]

Here is the image you requested.

[nimble mason]

Here is the image you requested.

[small cloak]

Photos of LeBron James playing basketball in East Beijing

[nimble mason]

Here is the image you requested.

[nimble mason]

Here is the image you requested.

[nimble mason]

Here is the image you requested.

[clever oar]

Clownshark Batwing maybe you help me with ip adapter?)

i cant install it two days

[nimble mason]

comfyui?

[clever oar]

a111

[vapid crest]

any tips to avoid the upscale be so messy?

[nimble mason]

i use forge sometimes but afaik all you need to do is drop the models in the right folder, not sure though cuz i don't use it for ipadapter, only for convenience when generating images on my phone

[glacial oracle]

photograph, athlete walking in sport
ground on a empty ground, raining,
Film Grain, --ar 16:9

[clever oar]

i download it and put in right place and get file maybe corrupted

[languid pebble]

Looks like a wrong VAE?

[nimble mason]

dont use euler normal with 20 steps

try dpmpp ancestral, karras, 35

also take out almost all of the neg prompt, sdxl doesnt benefit from that and it can make images worse

cut cfg down to 6 or less

and delete the upscale at the end until the first part is fixed

its also possible your lora is causing problems, set strength to 0 until the image doesnt look burned, then bring th strength back up, start with 0.5, then 0.7, etc

you always want to turn off as much as possible, to simplify, then add one thing at a time until it looks fd, if youre troubleshooting

[vapid crest]

its a sd 1.5 model though

but i'll try it out

[nimble mason]

Oops you're right I got it mixed up with ghostxl

But yeah just simplify everything, if I have something that's fd and idk why, I turn off everything, remove the neg prompt, remove all quality descriptors, cut it to bare bones

[languid pebble]

Hero!

[nimble mason]

Then fix any settings issues, THEN add one big chunk of the prompt back in at a time, then star

Start working on putting stuff back in

[languid pebble]

Wrong quote but message is right. I always mess with Comfy 😄

[clever oar]

oh after 1000 attemps i install ip adapter

mix with open pose)

[languid pebble]

Yay! Congrats!

[clever oar]

THX!

big step for me)

[languid pebble]

I stopped trying ... did you use a special documentation?

[clever oar]

oh is need big memory

i use google and cs10 help me in tech support

[languid pebble]

4090 here ... should work 🙂

[clever oar]

I wanted to give up and forget about it, but suddenly I saw how the picture began to change in the preview

but now i have cuda out of memory)

i change picture

[languid pebble]

Maybe you can start your GUI with a --lowVRAM or --MedVRAM argument?

Didn't write it 100% correct ...

[clever oar]

i have it(

now its works but not do what i want

lost face

[languid pebble]

It's always a trial and error ... just don't give up. I'll try to install it, too 😄

[clever oar]

you do it in comfy?

[languid pebble]

Yes ... I will try ...

I think there's a special face module not for use commercial that causes a lot of problems

[clever oar]

i have special model for face

but it not do what show me slownshark

he show me cool hybrids

[languid pebble]

He is damn good!

[clever oar]

you do some goode stuff too)

[languid pebble]

But Comfy I only do to learn a bit more. Usually I use Forge ...

[clever oar]

maybe you install it in forge?

[languid pebble]

Not sure whether it already can do IP Adapter Plus ...

[clever oar]

if feel now empty -i try to install it two days 🙁

[languid pebble]

Sometimes it can be good to stop work on something and just do something complete different. I often get new ideas that way.

[clever oar]

nice way

i have promblem often i cant stop

lost all energy

and if i get what i want i dont feel happy)

[languid pebble]

Doesn't sound good ... well you did another step ... maybe time to reward yourself?

[languid pebble]

She's a bit slim but it kinda worked ...

[clever oar]

yes i need found solution for do this better)

[languid pebble]

Sure you will!

[clever oar]

for some reason the background cannot be changed using promt

maybe is only for body

[languid pebble]

Did you select to change whole picture?

[clever oar]

yes

[languid pebble]

I'll have to try by myself before I want to say something 😄

[clever oar]

smoking

[languid pebble]

Maybe the shadows in the original picture make her look thinner?

[clever oar]

maybe i use different resolution

[languid pebble]

A.I. just gave me an animation ... better than I thought but worse the same time 😄

At least it's walking down the steps ^^

[clever oar]

😃

drunk robot

[languid pebble]

In transform mode 😄

[wispy nest]

Face is breathing

[languid pebble]

At least I still can laugh ^^

[clever oar]

canny test

[languid pebble]

Using 1024*1024 output?

[clever oar]

512

[languid pebble]

Maybe input and output size should fit better?

It's like the face ... slim

But worked!

[languid pebble]

😉

[clever oar]

broken

[arctic laurel]

2024 Best Diffusion Award goes to you

[nimble mason]

Style blending with ipadapter plus is amazing

[nocturne oak]

Epic tweet.

[south temple]

New node available

[languid pebble]

IP Adapter Advanced sollte hier nun auch laufen 🙂

[clever oar]

clownshark

my ip adapter not give me big different

[vapid tendon]

I need help finding old prompts within SD}

[clever oar]

how use it for hybrid

[clever oar]

monkecrab

[arctic laurel]

isn’t that the drummer from the rolling stones

[wispy nest]

i call it a manonkitang

mononkitang*

[clever oar]

one extinct animal was somewhat similar

[nimble mason]

sweet, thxk for the heads up

[nimble mason]

glad to see it - it's been the one big thing a1111 has had that comfy's been missing afaik

perfect timing too - now we have all the tools for "prompting" with an image from the conditioning side, right after the new ipadapter plus dropped

prompt: high quality, detailed. input image, then output with: reference_attn, _addain, and _attn+adain

couple ipadapter outputs for comparison

[nimble mason]

composition weight style 😄

[clever oar]

umbrella)

[nimble mason]

ipadapter? controlnet reference? something else
?

[clever oar]

ipadapter +

promt man siting

[clever oar]

i like this adapter)

but this use all vram

[nimble mason]

whats all the settings on that image

haven't tried it in a1111

so curious what translates over

[nimble mason]

full preproc model name, and model name?

(mostly i'm wanting to be able to answer the "can i do this in a1111" questions)

[clever oar]

ip-adapter-plus_sd15.safetensors

and clipvision encoder

[nimble mason]

does it work for ya with the sdxl version or is that an OOM situation for you

[clever oar]

sd 1/5

no xl i think

[nimble mason]

how much vram you got

some fun results by using a combo of ipadapter and controlnet reference

[clever oar]

4 gb

[nimble mason]

i had ipadapter with composition weights control the steps up to 0.3, then 0.3-1.0 was controlnet ref

ahh dang hope you find an intact 4090 sticking out of a trash can sometime soon

[clever oar]

i want buy 3060

[nimble mason]

spend a bit more, get a 4070 super

much faster, and 12gb vram

[clever oar]

but i not have bit money

4070 expensive i thnik

ipadapter its best in controlnet)

i dont want use it before you post your pic

[nimble mason]

yeah they are a bit more, 500 or so vs 300 or so for the 306

3060

[clever oar]

how safe face with change pose

instant id?

[nimble mason]

yeah or faceid etc

best for preserving likeness is portraitID

but that does require closer shots

[clever oar]

huh lamp jelyfish0

[ripe pilot]

Did something similar a while back

[nimble mason]

ipadapter style blending

[clever oar]

i try it for inpaint

amazing

[nimble mason]

i bet it is, any examples you wanna share? 🙂

[clever oar]

i inpaint my real photo

[clever oar]

it was clear that the area was very different, like obvious photoshop

now is natural

[south temple]

My machine is rendering like a madhouse at the house

I put about 64 things in the que before work

Now Im trying to render out a cricket, when the cricket rubs its legs… colors go flying.

I guess… alien bug works too lol

[karmic flare]

39.057317, -77.124324 I think I saw one here

[nimble mason]

more mutants i see 😄

[clever oar]

it's not a mutant, it just twisted her 😃

[nimble mason]

you should post the actual output! 🙂

very cool seeing the input too though 🙂

[clever oar]

need sleep after 4 hours on work 🤭

[nimble mason]

you need to get up for work in 4 hours?

[nimble mason]

damn umbrellas @jovial tiger

[jovial tiger]

wow

[nimble mason]

seriously... this ipadapter... gonna one of these nodes tattooed on my forehead

[jovial tiger]

hah

sd really likes to have umbrellas as hats.

sometimes it gets is accidently right

[jovial tiger]

how cool would that be. some way to have an sd rendered image, where they're holding up a sign or have a tattoo on their head, but it's the image you've loaded into a node, perfectly inlaid.

[nimble mason]

using attention masks and timestepping with a second style adapter to frame a shot

nowwww we're talking.

i drop the second style adapter during the upscale to ensure the details are coherent

[hazy warren]

Oh cool, this looks like one of those artworks you see at hotels

[nimble mason]

what i'm really li king is this is providing some ability to subtly guide the composition

[south temple]

Enjoy!

Tonight’s theme…. “Frogs”

[pallid ruin]

@icy smelt

This is SD Forge

[nimble mason]

might wanna drop the .json too with the image 🙂

there's also a node that allows you to embed it in the workflow iamge

[south temple]

Ill drop it when I get home

[warm beacon]

https://tenor.com/bHizPCpOujV.gif

https://tenor.com/bnsLu.gif

[pallid ruin]

Wooow, that seems so difficult 😨

[hazy warren]

this is what ppls charts look like when they trade stocks

[formal dawn]

Awesome, is this channel dominated by you?

[formal dawn]

niu bi

[nimble mason]

at times

[formal dawn]

It’s hard to imagine what people will say when they know about this in the future. There is a group of people on the channel who answer people’s questions and help people who don’t know how to get what they want in their minds. somewhat great

[nimble mason]

@hazy warren

[hazy warren]

looks too human

hes normally way shorter

[pallid ruin]

Anyone got an idea of what small lora can I train? (15-30 images) 🤔

[hazy warren]

id say if you want to train anything worth it, needs to be atleast 500 imagess min

[nimble mason]

yeah, just saying, that was a prompt only, so you can easily do it

you don't need 500 images

i've trained some pretty good ones with 40 or 50 or so

quality over quantity

variety and captioning and quality over quantity

[hazy warren]

super sampling?

[nimble mason]

what are you referring to exactly

[hazy warren]

the thing where it can train on the same thing multiple times, I forgot what they call it

[nimble mason]

you don't need to do anything weird

loras just don't need tons of images

a couple dozen is all you need

[hazy warren]

oh I see

[nimble mason]

if you go overboard with the training too (too many steps) it'll usually start to look "Burned" and get too stiff

start only replicating the training data

[hazy warren]

I was interested in that "Train" tab, I might have to try it out some time.

[nimble mason]

don't use webui for that

personally i recommend OneTrainer

[hazy warren]

😮

[pallid ruin]

Last time i trained a lora I did it with mj images but I forgot to check them correctly...and ALL of them had blurry backgrounds 😭

this are the results like, well it could be worse

[hazy warren]

train it on ya cat

[hazy warren]

this a good tutorial?
https://www.youtube.com/watch?v=yPOadldf6bI

[south temple]

there ya go

oooo

I see

[nimble mason]

there's a good chance it is

i don't watch videos lol

i lack the patience

[hazy warren]

😮

[nimble mason]

yeah i haven't watched a single tutorial ever

so i have no idea

i can give you my training parameters though if you like

kinda like comfy... you can just import it and go

i think it's way easier to learn something starting with something that works

[hazy warren]

I like this.
do some gum tree's

[hazy warren]

yes plz

[hazy warren]

put wheres waldo in the background

[hazy warren]

now make them Kermit

[south temple]

I dont... have words

[south temple]

ok.... you win lol

[clever oar]

my new pc

[languid pebble]

Not an spple ^^

[dense tapir]

https://civitai.com/models/180637

[thin scaffold]

green grass

[dense tapir]

blue sky

[thin scaffold]

Switch colors

Fairy tale style

https://cdn.discordapp.com/attachments/1004159122335354970/1225007591029997599/R.9b688ef053db436e33f9d7d85e6de967.jpeg?ex=661f9039&is=660d1b39&hm=423324774d50dc6b6dea52bf4f145b2d79f9efc46e6dcde5df9c5274383d861d& Fairy tale style

https://cdn.discordapp.com/attachments/1004159122335354970/1225007591029997599/R.9b688ef053db436e33f9d7d85e6de967.jpeg?ex=661f9039&is=660d1b39&hm=423324774d50dc6b6dea52bf4f145b2d79f9efc46e6dcde5df9c5274383d861d& /Fairy tale style

[clever oar]

robot+pc image

[thin scaffold]

Make it a fairy-tale lawn picture

https://tenor.com/6AtA.gif

[lofty ivy]

I made my first Lora and wanted you guys to be the first to know. I was motivated by an Original Character that i made for my manga, so I made a Box Braid hairstyle specific lora with great results. check it: https://civitai.com/models/381076?modelVersionId=425404

already working on V2

[undone lotus]

/game_statsA series of display boards and cabinets are arranged within an industrial museum, rendered in the style of Cinema 4D to evoke a sense of historical depth and the passage of time. The materials featured include wood, stone, brick, and concrete, with a color palette dominated by pale amber, red, and touches of black. Exhibits on display encompass miniature industrial models, photographs, and historical documents, among other artifacts.--ar4:3

[languid pebble]

someone asked whether this would work ....

[placid cedar]

https://tenor.com/bcN9J.gif

[narrow widget]

#🏞｜general-with-images Brandon Jiang is a tall and handsome CEO with a commanding presence. He possesses deep, brown eyes that exude determination and confidence. Clad in impeccably tailored black suits, he emanates an aura of authority that is both intimidating and captivating. His long, slender fingers adorned with a sophisticated wristwatch reflect his attention to detail and refined taste. With a strong nose and a sturdy jawline, he carries himself with self-assurance and decisiveness. Despite his outwardly tough demeanor, Brandon harbors a gentle and compassionate heart, showing unwavering dedication to both his career and loved ones.

[frank tartan]

a cat looking sky

[junior sky]

Here is the image you requested

[nimble mason]

Here is the image you requested.

[clever oar]

Clownshark Batwing my ipadapter doesn't want to change clothes

monkey in sports suit

[nimble mason]

I like the soccer ball foot lol

[lofty ivy]

its pretty cool but the slight jitter makes it feel like the buildings are morphing abit...youd have better like generating the image then using another video editting software to pan thru

[gleaming dagger]

黎明前的宁静被打破，天空中的星星逐渐隐去，一轮红日缓缓升起，带来第一缕温暖的阳光。打工人的房间内，闹钟的铃声在静谧的空气中显得格外刺耳。他从混乱的梦境中惊醒，睡眼朦胧，心中充满了对即将开始的一天的期待与不安，最好的质量，8k

[crude badger]

#🏞｜general-with-images a space particle dispersed in a medium

[nimble mason]

Here is the image you requested.

[crude badger]

particles not a creature

[crude badger]

no sir particles

[crude badger]

yes yess this is grate

can u genrate a 16:9 ratio image with more particles and blured edges of the images

[nimble mason]

oh man i love that

deliciously weird and creepy

[junior sky]

Cthulhu Carwash. Awesome!

[sudden current]

@nimble mason I'm hoping to make something like the white image in the style of the sketch image

(Not including writing)

[nimble mason]

here's the output using the sketch as a style, and basically a blank prompt (just "high quality, detailed") with a generic negative ("umbrella, NSFW, low quality, bokeh, blurry, low detail, text, watermark")

there's a lot you can do to play with it to get closer adherence, and the text prompt does remain important

[sudden current]

Wdym by using the sketch as a style?

[nimble mason]

here's using the sketch as a style input, and the white image as a composition input

this is using ipadapter plus in comfyui

[sudden current]

I like how the style has transferred, but is there a way to get it to recreate the original more accurately?

and is this for SDXL?

[languid pebble]

I think upload stuff to discord doesn't make em better. To make this effect with another video editor would be the better idea. But I experiment with SVD and you never know what you will get 🙂

[nimble mason]

yeah, sdxl

[languid pebble]

SDXL 😄

[nimble mason]

yeah, that's possible if you use controlnet as well

input sketch, output, and style reference, in that order (for anyone wondering what you can do with ipadapter plus nodes in comfyui right now. afaik there's no style transfer ipadapter available in a1111/forge, but i'm wrong about that, lemme know!)

[clever oar]

canny+ipadapter test

[jovial tiger]

what svd settings are you using? at this point animatediff does a better job, although it also loses more of the original image that I loaded.

[languid pebble]

6FPS, 25 Frames, 135 Motion Bucket ... I optimize afterwards

[jovial tiger]

what's your augmentation level?

[languid pebble]

Still at 0 don't know what it does 🙂

[sudden current]

Ty for all the help, much appreciated

[jovial tiger]

so that worked out well. augmentation level is how much noise it'll add. so if you say 0 , which is what it's at for this pic, it'll usually just be some kind of pan or orbit rotation, the subject won't move.

[pallid ruin]

wow! nice

[languid pebble]

I never really know what it will do 🙂

[nimble mason]

big takeaway is... is that what you were looking for?

cuz if so, you can do it in comfyui, and the workflow is embedded in that image so you can just click and drag it into comfyui and do it with other images

[sudden current]

It's exactly what I was picturing

[sudden current]

That sounds ideal

[nimble mason]

good to know... yeah, pretty cool what you can do now

that style transfer ability is next level

[jovial tiger]

I'm willing to accept that it won't do much, but it takes a long time to render and not knowing what it'll do for you is frustrating. lots of wasted time on stuff that comes out bad for 1 good one.

[languid pebble]

80% acceptable, 30% good

acceptable includes good 🙂

[jovial tiger]

yeah... so much of this...

whereas the animated diff stuff comes out 80% good, but because it always has to go through an sd 1.5 model, the original image is changed a good amount. the models I used in sdxl aren't available in 1.5, even similarly

[languid pebble]

Augmention Level 0.5 without optimization ...

Augmention Level 2.0 without optimization ... 🙂

Looks like 0 is pretty good

[jovial tiger]

haha yeah

[languid pebble]

But sometimes I just wanna kick it for a small mistake ^^

[jovial tiger]

@pallid ruin @nimble mason ok fine this ipadapter thing is pretty neat.

[clever oar]

put your pic in adapter)

[languid pebble]

Nice, too!

[languid pebble]

🤣

Try with Trump 😄

[clever oar]

i use only pose biden

[languid pebble]

New Pixar ...

Frankenstein and the red haired lady ^^

[languid pebble]

Knock him out!

[clever oar]

how you generate biden ?

my very bad

[jovial tiger]

nah, biden is a lover not a fighter. 🙂

this ipadapter thing is pretty neat. adding a little bit of prompt can make actions happen between source images.

[clever oar]

his face very close , when i do that his face not him

[jovial tiger]

skeleton man is even friends with tentacle man

[jovial tiger]

did you call him out in the prompt? It wasn't close with ipadapter until I said it was biden.

[clever oar]

😄

[clever oar]

yes i call,but i have not full face

[languid pebble]

Looks like I've ordered my SDD at the right moment ...

[nimble mason]

that merge really does have great prompt comprehension. didn't get that output for wading through congress in anything else i just tested

[jovial tiger]

@nimble mason so this was pretty cool. took that 2 image ipadapter thing that was kind of doing 50/50 regional prompting with ipadapter and 2 images, but instead taking one and just plopping it into the middle of an existing image. so super sayan head on top of a wolf in sheep's clothing.

[nimble mason]

what's the 50/50 one you're referring to? was thhat one we were takling about the other day?

btw... the style transfer ipadapter with attention masking is very, very good with the otherwise weak and crappy qrcodemonsterXL controlnet

[jovial tiger]

@nimble mason by 50/50, it's that one that was going around off the internet where it was just a 50% mask, so you could add one subject next to another. I changed the mask so that it was a square in the middle of the image instead of half the image

[jovial tiger]

oooh that's cool.

i never had good luck with that stuff. island in shape of a text mask or whatever.

[languid pebble]

❤️ Taiwan

[steep shadow]

"snake"

I'm thinking something. Are you thinking what I'm thinking?

[jovial tiger]

Gotta love Taiwan, it's where all the nvidia hardware comes from.

[steep shadow]

Stately speaking

in some space perhaps

[languid pebble]

Thank god I already ordered my new SSd ... prices could explode ... and they are friendly people ...

[nimble mason]

btw, highly recommend using ipadapter advanced so you can toy with the other weights

i'm finding weak output to be pretty useful too now

[cyan shoal]

SEGMoE

[jovial tiger]

with the fully real model instead.

[clever oar]

spirited away

[cyan shoal]

@tired basin here's an image I made with segmoe lol

[languid pebble]

Watching Star Trek Picard at the moment 🙂

[nimble mason]

source image for that? nice one

[distant vortex]

Hmm. The Rock is Bravestarr 😄

[languid pebble]

One of these ... can't find original ... ^^

Here it is ...

[nimble mason]

thx, yeah that's a really cool one

[wispy nest]

ok these are great especially for craiyon

[languid pebble]

My workspace 😄

[cyan shoal]

@minor fractal thoughts?

This is kinda worrying

[wispy nest]

that’s serious water mark

[cyan shoal]

I mean if it can still pump out complex images such as this I won't be complaining

its still a massive improvement to me

[minor fractal]

hahahahah

oh i want to tell you the story but they're wrong

t5 is using it's full ability

it has 512 prompt afaik

[cyan shoal]

really? so it DOES have 512 context? no cap?

[minor fractal]

afaik? yes

[cyan shoal]

can you make sense of "because of bootstrapped clip compatibility"

[minor fractal]

no, because they arent making sense

[cyan shoal]

lmao

[minor fractal]

oh they mentioned that?

tldr: they got booted from sd3 for purely arguing that "sd3 needs to be dall3"

[cyan shoal]

LMFAO

well thank you I guess holy crap

[minor fractal]

they are the only person as of now that has been booted.

[cyan shoal]

💀

so now they are just straight up boycotting it and spreading misinfo or what

[minor fractal]

they also "made a mmdit implementation a year ago" that totally exists and wasnt made up

[minor fractal]

yeah, basically

[cyan shoal]

epic

[minor fractal]

tldr: "noooo!!!! sd3 not dalle3!!!!!"

[cyan shoal]

lol

thank you, you have once again revived my interest in SD3

also got 24GB so I can do highres with bigger batch sizes

and use controlnets or whatnot

[minor fractal]

also, the flaw they mentioned about "cogvlm changing screws to nails"?

it doesnt exist.

i just tried it

[cyan shoal]

LMAO

[minor fractal]

and it did make nails

[cyan shoal]

you tried cogvlm?

[minor fractal]

they trained on half cogvlm captions

[cyan shoal]

yeah and THAT TOO

[minor fractal]

so if it really does suck

it shouldnt do nails

but

[cyan shoal]

^^^^^ yesssss

[minor fractal]

it does

[cyan shoal]

its good that it's only half of the dataset

so pop cultural stuff stay which might not be recognized by cogvlm

[minor fractal]

they could do 100% but then we'd lose out on small prompting or direct name things

[cyan shoal]

I wonder if Ideogram did a higher ratio of auto captioned dataset

I just don't know why but the same prompt without the assistance of Magic Prompt makes nearly the same composition every time

[minor fractal]

could be

cogvlm and other img taggers are good

[cyan shoal]

yeah I saw the examples of cogvlm and its really advanced and high quality

isn't it like ~30B too?

oh 17B

CogVLM-17B and CogAgent-18B

[languid pebble]

That sounds gooooood

[cyan shoal]

yeah they used this to caption SD3's dataset

[languid pebble]

Still no beta for me 🙂

[cyan shoal]

I don't think anyone from the waitlist has been invited yet

[jovial tiger]

They have have added a few more people on twitter. I noticed at the beginning of the month that a few who have it now mentioned they're still waiting. And now they're posting sd3 pics

[cyan shoal]

I hope one day a huge pile of people get access and we all start engagement in this server and on the subreddit

[languid pebble]

Damn ... I've entered 20 after announcment ^^

20 Minutes

[cyan shoal]

I just wonder if the reception will be highly negative or neutral

[jovial tiger]

Well I think it's just people on the AI industry

[jovial tiger]

I'll be happy with it because it means we got our hands on it, and no matter what happens with stability after that, we still have the engine

And I like cascade, people will make models for it

Unlike cascade

[cyan shoal]

yeahh

I will be super positive about SD3

oh like I mean with the early access where we generate images only

then after we get SD3 Code and Weights then of course it'll be the end game

and no matter what happens with stability after that, we still have the engine
exactly this, we will be able to enjoy more complex images forever

and if we're missing something people will just make loras or massive finetunes

I just hope that 24GB will be enough to train loras

[languid pebble]

I hope I'll still have a benefit from my early training 🙂

[cyan shoal]

We just need to stop our current prompting habbits

its not like current style prompts don't work, it's just that they might not be as good

Superprompt or other instruct LLMs will help a lot

[digital hornet]

given that 24gb is realistically the limit on consumer GPUs and that Loras are the main way to extend the model, I'd expect it to be possible, just possibly slow

[cyan shoal]

well its not stopping the big boys from training with their a100s and etc

@jovial tiger SegMOE

[ripe scroll]

my first image! lol

[nimble mason]

here's my output

[hazy warren]

ooo, now do the Batman who laughs

😂

[hazy warren]

nah, not the right one 😦

[nimble mason]

alright now try a prompt that includes "batman laughing maniacally with a mouth that looks like a shark mouth with long pointy teeth"

[hazy warren]

he ment to look like this

[nimble mason]

f yeahhhh

you're gonna have so much fun with this

you'll prolly fail out of school and get fired as a result

😛

do you have comfyui manager installed?

[ripe scroll]

how did u do that one?

i dont i just have comfyui

[nimble mason]

do you see the manager button

[ripe scroll]

im 45 lol

yah i dont have it

[nimble mason]

ahh k lets get you that

https://github.com/ltdrdata/ComfyUI-Manager

this is critical

if you load a workflow from someone, you'll often not have the nodes they have

they'll be red

then you can just go into manager, click install missing custom nodes, and restart comfy and you'll be set

[ripe scroll]

i typed git clone https://github.com/ltdrdata/ComfyUI-Manager.git but it errored out

[nimble mason]

errored out? as in no connection?

[ripe scroll]

git is not recognized as the name of a cmdlet

[nimble mason]

oh you need git

[ripe scroll]

install git?

[nimble mason]

https://github.com/git-for-windows/git/releases/download/v2.44.0.windows.1/Git-2.44.0-64-bit.exe

yeah

the command git doesn't exist until you get that installed

won't be able to add any custom nodes without very clunkily downloading them manually from the repo without having git

[ripe scroll]

ok installing now all on default recommendations

[nimble mason]

sweet

[ripe scroll]

damn didnt work still

same error

or do i run everying in git cmd?

yah that worked

[nimble mason]

k you got it?

if it doesn't recognize it in the future, might need to add the path manually to cmd.exe

[ripe scroll]

yah now do i close and reopen comfy?

[nimble mason]

yeah

and then you'll have the manager

and i can drop some workflows here for you to load that'll get you some of the best custom nodes

[ripe scroll]

i closed and reopend but manager not there

[nimble mason]

hmmmmmm

where did you git clone

[ripe scroll]

https://github.com/ltdrdata/ComfyUI-Manager

[nimble mason]

are you using the portable comfy?

[ripe scroll]

yah

is ther a proper way to restart comfy?

[nimble mason]

just go to the console window and close it

thats all theres to it

[ripe scroll]

humm

i did this method

but i did it though git cmd

[nimble mason]

so if you open up your ComfyUI/custom_nodes folder, do you see it there

ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Manager

should be in there

[ripe scroll]

yah not there

[nimble mason]

k so it got cloned into the wrong folder prolly

i'd just follow the instructions directly now

[nimble mason]

launch a new cmd.exe window if you haven't

you'll have a random copy of comfy-ui manager somewhere on your hd just chillin but that's cool lol

[ripe scroll]

ok manager is now in the folder

[nimble mason]

k cool now it should work

[ripe scroll]

yah but it didnt

[nimble mason]

wtf

[nimble mason]

after closing comfyui completely, including the console cmd.exe window, it's not there?

[ripe scroll]

its working now!@

[nimble mason]

yay!!!

cool lemme get you a workflow

click on this, click open in browser

then drag into comfyui

might take a sec but lemme know if a workflow pops up

yes it's a mess lol

[ripe scroll]

buch of red everywhere lol

[nimble mason]

that's good!

click the manager button

[nimble mason]

install missing custom nodes, then restart comfyui

it will take a bit

but these are all really handy ones i use nonstop that you'll def want

and will save you lots of time hunting

[ripe scroll]

my eyes are hurting lol

i guess it takes time to restart

[nimble mason]

yeah, usually it downloads and installs stuff when it restarts if you just added nodes

good habit to get into... if you don't hear your comp buzzing when you run a workflow

[ripe scroll]

yah i see cmd putting along

[nimble mason]

check the console window... it's likely downloading a model or osmething you were missing

[ripe scroll]

man this is a lot of shit lol

[nimble mason]

lol

[ripe scroll]

this is way over my head lol

[nimble mason]

i have over 100 node packs lol

you're all good

[ripe scroll]

geez lol

[nimble mason]

the vast majority i don't use much

but these get a lot of use

[ripe scroll]

am I going to have to do this all over when version 3 comes out?

oh gawd

ok now what lmfao

[nimble mason]

hahahaha

now i clean up the workflow and send it to you again so you can actually understand wtf is going on kinda

[ripe scroll]

yes please lol

[nimble mason]

i just tossed you what i had up so you could get it downloading

[ripe scroll]

still cleaning it up? lol

[nimble mason]

yup almost there

ypou'll have fun with this

oh, you prolly don't have that sampler, but you can change that from res_momentumized to dpmpp_2m_ancestral or whatever too

but load some images in the load image nodes that you think are cool that have a style you want to imitate

and also try changing the weight type in the node on the top right

those were my inputs

[ripe scroll]

do i drag the img in?

wow!

[ripe scroll]

wait u want me to do this now?

[nimble mason]

yeah why not go for it 🙂

this is a brand new feature in comfy, one of the best yet imo

[ripe scroll]

node in top right i dont see weight type

[nimble mason]

so if this works you're at the cutting edge

[nimble mason]

top right of this

[ripe scroll]

mine still looks like crazy

[nimble mason]

oh that's the old one haha

drag this one into comfy

[ripe scroll]

ah much better!

alright buddy thanks for this. Tremendous help!

this is awesome

[nimble mason]

no prob!

[ripe scroll]

how long have u been using this?

[nimble mason]

a couple months

[ripe scroll]

damn seems way longer than that

[nimble mason]

discovered stable diffusion about 4 months ago

[ripe scroll]

i discovered it like a year ago but never messed with it until today

you can call yourself an expert in my books

[nimble mason]

yeah once you get over that first hump to be able to make shit that's actually really, really cool, it's worse than drugs lol

[ripe scroll]

lmfao

yah im in a crypto project called print and I like to make images for them and tweet them out on x

i guess i'll have to photoshop text in

i was hoping v3 would do that but not sure now

prior to this it was bing and chat gpt to make the images lol

[nimble mason]

ah cool

yeah text is iffy unless you use it pre-generated and build on it

this was the input

[ripe scroll]

did it make that or did u build on it?

[nimble mason]

used the black and white one as a control image

[ripe scroll]

oh shit thats awesome!

[wispy nest]

anyone knows what model this is, apparently its really popular in the community rn

lik everywhere

[ripe scroll]

where would i load the control image into?

[wispy nest]

on pinterest i just dont know the name uno idk if this is the worng place to post

[ripe scroll]

the first node?

[nimble mason]

oh god that one is a disaster

the cleaned up one isn't quite able to do that

here's a version of the cleaned up one that also upscales btw

[ripe scroll]

should i use this one now? lol

[jovial tiger]

that indeed looks neat.

[nimble mason]

yeah, the upscale can be a bit slow though

i generally recommend just leaving the upscale not hooked up... just disconnect one of the wires to the second ksampler

[ripe scroll]

i have a serparate upsscaler

[cyan shoal]

gotta love the pixar eyes

[nimble mason]

then if you really like an image, reduce the seed by 1, so it runs with the same seed again, reconnect the wire you disconnected, and run

[jovial tiger]

do you like sandwiches by any chance/

?

[ripe scroll]

so where do i load the control img?

[cyan shoal]

nails

sure...

[cyan shoal]

wow nice

did you see my segmoe images

its surprisingly decent for zero shot

[nimble mason]

with the cleaned up workflow?

[jovial tiger]

so I was out at the grocery store when i saw the segmoe stuff. takes bits and pieces from various models and renders the result in realtime?

[nimble mason]

this discord has added a few min to some of my grocery trips...

[cyan shoal]

I actually have no idea. All I know is that it's a Mixture of Experts...

[jovial tiger]

I've done a couple of model merges lately to end up at this artistic/dark/slightly pixarish view of everything. it's why everything I post looks like that.

[cyan shoal]

no idea what it means for Diffusion models exactly

[cyan shoal]

I think I could actually merge the darkarts and proteus model to make a 2x1 segmoe

but no pixar lora though, unless I merge it into one of the models before hand

[ripe scroll]

ok so the first node correct?

[jovial tiger]

so I did that merge, but then I found that the new version of paincreator's ai creator artistic v15 or some such is incredibly prompt following, even beats proteus which is a tall order. so now i just use that and andrea75c's cute 3d render lora merged in ata 0.5 with no clip from it for the final output. the difference in prompt following with complex stuff is incredible.

[nimble mason]

yeah, those two images above are the inputs

[jovial tiger]

so now i couple that with regional prompting, and voila.

[nimble mason]

you should be able to see where they go with the screencap of the workflow

so it's actually borrowing from both images

[cyan shoal]

can you link that model please

[ripe scroll]

ok i need to mess around with this youll probably see me tomorrow with a shit load of questions lol

again thank you so much man. I really appreciate you going out of your way to help me

[nimble mason]

https://civitai.com/user/pAInCREAT0R

https://civitai.com/models/219616/aicreat0rxl-artistic

https://civitai.com/models/227059/fullyrealxl

https://civitai.com/models/321687/dark-arts-images

all three of those are amazing

[cyan shoal]

thanks

[nimble mason]

no prob enjoy the shit outta it

just experiment like crazy

[cyan shoal]

so aicreator is the smartest?

[ripe scroll]

these images that we drop into it are they just images?

[nimble mason]

don't get caught up in trying to do things perfectly, cpu & gpu time are basically free for ya so just hammer out all kinds of crazy tests and see what happens

yeah

most formats work

[ripe scroll]

like how does it know to do these crazy node setups

[nimble mason]

comfyui?

[ripe scroll]

yah

[nimble mason]

it's just the programming

[ripe scroll]

so if i drop any image the node will change?

[nimble mason]

it's open source so if you wanna dig into that uncommented mass of code... 😄

[ripe scroll]

lol

[nimble mason]

oh, what is shown will, yeah

you can also ctl-c an image outside comfy

then left click on the background in comfy, hit ctl-v and you'll get a new load image node

[ripe scroll]

mindblown

[nimble mason]

if you click on the load image node, then hit ctl-v, it'll replace the image

can be a lot more convenient... recommend cruising google images or deviantart or whatever and just pasting stuff in

seeing what ya get

[ripe scroll]

if i like the node setup i just hit save then just click on the json file to load it?

[nimble mason]

save will save the workflow, yes

and every time you generate an image, if you're using the save image node, it embeds the wrofklow in it

so you never lose a workflow

you can recreate anything you make

[ripe scroll]

fire!

[nimble mason]

yup, best feature imo

[ripe scroll]

alright see yall later i gotta go. Thanks again!

[nimble mason]

i know myself, soul, festivalman (when using comfy, he uses a lot of a1111 with LLM bots... some fancy stuff), and others leave their workflows in their images here

later

[nimble mason]

composition, two style inputs, output

and fyi if any of your images show up in my inputs here, it means i think it's straight fire

[jovial tiger]

https://github.com/cubiq/ComfyUI_IPAdapter_plus/tree/main/examples

so i'm looking at this example directory. which one is just straight style transfer? I just want one image input and a prompt.

but i want just style, not composition.

[nimble mason]

idk but i'll get you one in a sec here

[jovial tiger]

hah ok that works too.

[nimble mason]

fuck i can't stand how much i love these new nodes lol

style transfer for txt2img?

[jovial tiger]

sure

[nimble mason]

input, output: "a car driving in a city"

honestly, i think the style transfer is the crown jewel of all of this

[jovial tiger]

i don't see any ip adapter nodes in this one? is it something else?

[nimble mason]

IPAdapter embeds?

look at the one with the car

def recommend using res_momentumized for that first stage

i get better style transfer

[jovial tiger]

yeah this is great stuff

[jovial tiger]

this was actually what i was originaly looking for when i hit the compoisition one.

[nimble mason]

gotcha

that's really cool ^^ but what really impresses me is how well it translates to different subject matter

[jovial tiger]

style transfer is nuts.

[nimble mason]

yuuuup

that's when i lost my shit

hm, no luck using this as a style input

cool output, but not engraving style

dpmpp_2m exponential... great stuff, just not engraving style

[wispy nest]

smackalishes dick and rice

[nimble mason]

with juggernautxl instead of fullyreal... wow does that look great

[wispy nest]

curly dick fries with ketchup

spring dancing dick jennie with a chili dick chenie

dickatronisanating sexology

[wispy nest]

telepathy dicks to a higher level of teleportation sex topic

decrepit dick chooser

the omega slopethen sex creatures

[jovial tiger]

@nimble mason something of note. in the one you sent, it did the render, then 0.5 denoise with 1.5 upscale. I added a 3rd and as you can see, it didn't actually become prompt adhering until the 3rd one which is weird.

[nimble mason]

that's weird... is that something that applies with other fringe prompts?

[nimble mason]

Here is the image you requested.

[jovial tiger]

gonna try a different style source and see if it still applies.

[nimble mason]

one thing i'm excited about with this is it frees us from needing to use any tokens to describe style

[jovial tiger]

nevermind, i screwed up the chain. fixing now

indeed

[jovial tiger]

so i switched everything back to dpm++2m karras again at 20 steps, x3. with 1.5x scaling each time. keeps the style and content of the image. r_momentum ends up with a wildly different and messed up image at the end compared to the first in the chain.

[nimble mason]

yes. res_momentumized makes huge changes

use exponential too if you want more stability when upscaling, it makes a big diff

if you want even more stability, unsample then sample with kadv with add noise off

[nocturne oak]

either version of restart can make big changes

[jovial tiger]

ok tried exponential with res_momentumized, now he's growing a face from his chest. 🙂

[nimble mason]

hahahaha

oh i mean with the others

res is hard to control

[jovial tiger]

it's the added noise at each step that's the issue.

[nimble mason]

you gotta dial that denoise waaaay down with that

i've found dpmpp_3m_gpu with exponential at 50% denoise is pretty good at not letting things get wacky

[jovial tiger]

I really need to be able to get those 2m and 3m_gpus working via api. they're the ones that don't for some weird reason.

[nimble mason]

that's lame

[jovial tiger]

res_momentumized works though

[nimble mason]

yeah dpmpp_3m_sde_gpu is way better at maintaining composition and tone than dpmpp_3m_sde

it's the best out of all of them with exponential that i've tried

i wasn't really good about keeping records but i think i tried all combinations at some point

[jovial tiger]

so that worked. res_momentumized with 0.35 denoise and exponential

[nimble mason]

awesome

[nocturne oak]

mmmm....sexy

[nimble mason]

these are settings that i've found work reasonably well in general

but it's something to throw in if you want 50% denoise after upscaling

[nimble mason]

so awesome lol

[jovial tiger]

midjourney has some really great styles at times. being able to drop them in on these is pretty great.

[nimble mason]

we should get a lil collection going somewhere

i've been trying to pick out images of my own that have led to particularly good outputs with the style weights

[jovial tiger]

yeah, it is more than just style btw with this.

one on the left is from midjourney, one on the right, there's no water in the prompt, but it's getting it from the style image.

[nimble mason]

yeah it does pick up just a bit from the composition

good and bad

[jovial tiger]

although I guess it's really just stuff that you didn't mention. it's still massively less than the composition workflow.

[shy berry]

In a bustling metropolis, skyscrapers towered into the sky, while crowds bustled through the streets and traffic flowed like a river. Suddenly, a colossal Godzilla emerged in the heart of the city, towering high into the clouds, its muscles bulging with power and authority. Its skin was covered in tough scales, its eyes gleaming with wildness, and its roar thundered through the air, deafening all who heard it. Godzilla's massive claws crushed through streets, buildings crumbled beneath its feet, and smoke billowed as flames erupted. Terrified citizens fled in panic as police cars and fire trucks rushed to the scene, attempting to stop the impending disaster. Yet Godzilla remained fearless and mighty, commanding everything around it like a sovereign ruler, demonstrating an unmatched aura of intimidation and strength.

This scene portrays Godzilla's ferocity and power as a monster, as well as its conflict and confrontation with the human world.

[jovial tiger]

this one it picked up that her eyes were closed, so now his are.

[nimble mason]

i bet if you combine multiple style embeds, images of the same style, with different settings... that will disappear

kinda like training a lora, need variety or it'll tend to carry over more from the training set in terms of composition

[jovial tiger]

that's an interesting thought

right

[nimble mason]

do you have a MJ sub?

would be interesting to make a lil set of some stuff of a given style to test that

even if it's just half a dozen images

[jovial tiger]

i do.

yeah just give me the prompts

[nimble mason]

hmmmmm. never used it so i'm not sure how it works with keeping a style consistent. is that part of the prompt like SD, or a special tag?

[nimble mason]

Here is the Sharkzilla you requested.

[jovial tiger]

@nimble mason https://www.midjourney.com/explore?tab=random

is this available to you? you can just grab the already generated ones for styles as well

[nimble mason]

ooo

[jovial tiger]

there's 10 random ones with style

[nimble mason]

hmm pretty diff though

i guess maybe the style options is something deeper than what i'm thinking of (watercolor vs. oil painting vs. photograph vs. monochrome photograph vs. engraving etc... and color palette, etc)

[jovial tiger]

i'm just going through old prompts with these style images, and it's giving wildly different stuff than the original.

[nimble mason]

now you're gonna end up having to incorporate ipadapter plus into your bot haha

it's just too good

[jovial tiger]

seriously who doesn't want more of this? this is with that blue-high tech shopping cart in that 10 set before