✨｜sdxl

[vague spoke]

generate a logo from the letters JO

[civic gyro]

Zoom from eye into the brainnnnn - https://twitter.com/HikariUchu/status/1764705840240746659

Hikari Uchu (@HikariUchu) on X

Dive from the gaze into the gateway of the mind🧠👁️ #MindVoyage #NeuralJourney #自然 #nature #ジェネラティブアート #generativeart #ビジュアルアート #visualart #アニメ愛 #animelove #変 #strange #ai #aiart @ai_artcommunity #ai_artcommunity #brain #axons #intricate #eye #journey…

Twitter · Mar 4, 2024, 5:33 PM

[copper lion]

Hi!
What is the difference between Base and refiner?

[worthy orbit]

you use base and you don't use refiner

[copper lion]

Why?

[worthy orbit]

it is a model that was never upgraded, most civitai models will tell you they don't need refiner

in my experience it is mostly messed up the generations adding crap

but it works in the final steps of the generation, you can set the steps it takes

[copper lion]

I see, thank you very much!

[copper lion]

I think Base is enough for a start.

[rustic garnet]

refiner is supposed to improve fine details on the image. Its trained on high quality images and only at the low-noise timesteps. Anyways, most custom trained models nowadays are better in doing fine details than using base+refiner, so you can ignore that

[covert barn]

[worthy orbit]

sorry for the answer I was trying to be edge. I did some good generations with the refiner but mostly it killed details. At low steps it just kind of messed up what was there, at more steps it killed details and replaced by some "high quality" texture, but it didn't know how to make real detail. So it gave some soft saturated kind of image. You can try it, or just download some of the highest rated SD XL model on civitaai

[lime tangle]

create a video from this image

[inland canyon]

https://tenor.com/bJmSN.gif

[timber terrace]

imagine a rat

[native knot]

Imagine thinking you could just give a command and art would appear without reading around the server.

[jovial smelt]

😆

[hasty smelt]

Guys, how to get download this file " DPM++ SDE Karras"? I not find this in Civitai.

I not have this

[eternal fog]

You do. dpmpp_sde

Then pick the karras scheduler

[hasty smelt]

I decided to install Forge for Stable Diffusion, I'm testing things out here. It's a little different for those who are used to ComfyUI Nodes

[hardy badger]

Human skull looking forward, drawing, black and white, white background, 2d, no cracks, flat

Any tips would be appreciated.

[uncut gull]

dpmpp_sde is the sampler. Karras is the scheduler.

[willow kiln]

Hi all!

I am learning to finetune a sd model via dreambooth. I just need a small dataset for getting through basics. If anyone has any resource where I can find small image datasets of the same object, pls share them

Ex: 25 images of the same animal/thing/place

[dreamy lake]

"Generate an evocative and poignant image of the legendary composer Karlheinz Stockhausen crafting his final musical note in his iconic studio. Capture the atmosphere of this momentous occasion, blending the essence of Stockhausen's avant-garde spirit with the emotional resonance of bidding farewell to a lifelong creative journey. Use your artistic prowess to convey the significance of this last note, encapsulating the fusion of innovation and reflection in the twilight of a brilliant career.""

[marsh ore]

Design a modern and minimalist coffee shop logo with the English letters ‘US’ as the central element.”
Color and Emo

[dreamy lake]

Becoming warmer

[pallid path]

all three are good kittens

[icy brook]

Some random images off Aetherverse XL.

[steady grove]

reasons like this i don't know why people hype comfy so hard as "betteR". A smidgeon faster on a few hardware configurations, but worse UX than a hacked together gradio

you have to do a ton of homework to use comfy. same with automatic but the interface is more intuitive and doesn't do shitty abbreviations like pp for ++

if you can even call that junk an abbr

[eternal fog]

It's an abbreviation used elsewhere and it's usually because you can't use special characters for some reason.

Comfy is fine, but it could do with better and clearer documentation, especially for making nodes.

I use it because it's nearly always got the latest stuff on it, either natively or via nodes. Auto is always fairly far behind. Maybe that will change with Forge, as that seems to have quicker development, but I've not used it or A1111 in ages.

[steady grove]

gets things first because the node architecture allows coders to do less work. i like that . gets tech out faster. now theres the unet patcher in forge. layer diffusion coming there first is a sign of the tides

comfy put himself in a position. can't copy the unet patcher without looking like he's copying forge. after accusing it of copying comfyui. don't expect it'll get a patcher architecture for nodes anytime soon

[copper kraken]

in the spirit of open source it would be nice if everyone just copied each other, was open about it, and encouraged it

then everyone wins

[tribal lantern]

Or maybe.... the author of the layer diffusion paper also made forge...... (and has a bit of a big ego so won't create the library for other UIs, but meh no one seems to do, stupid fragmentation, why can't everyone just agree on huggingface diffusers as a base)

[steady grove]

lllyasviel and LayerDiffusion accounts are the same ? Got evidence of that allegation?

I don't think there is a paper for layerdiffusion either

[steady grove]

Diffusers is the worst format i've ever dealt with . lots of issues. why agree on that format that already is one of the worst for compatibility?

even docx had more compatibility

if you ask me though, ego's a little deserved when you're the author of the original controlnet models.

I don't even think illy is the one with the ego. Clear to me it's comfydev

while ego may be deserved a little there too

[tribal lantern]

Well, there is. https://arxiv.org/abs/2402.17113. Lvmin Zhang = controlnet author (https://arxiv.org/abs/2302.05543) = forge dev

And no, you can't just rip the core code from another opensource project, then cast shade on said project and happily announce you don't use it 🤡

of course, license permits it as long as you atribute what goes on in someone head to say "not based on comfy" while actually reusing all comfy code, it's weird at best, malicious at worst

[steady grove]

comfydev created such an air of toxicity aroudn forge. so disapointing

stability should over rule him and recognize the contributions of those authors to the community loud and clear. shut everyone up and muzzle comfy on the issue

egos getting out of control and need to be leashed

[tribal lantern]

actually, had a brainfart, comfy being gpl, means any project using it must attribute the authors (comfy). so what happens in forge reusing the code like this is worse

[steady grove]

yeah you're blocked. clearly can't be reasonable. " it's gpl so you CANT copy it!" what?

[tribal lantern]

but meh, in my experience the whole ecosytem is a bit messy, let's just think of it as growing pains

[steady grove]

its so much worse than growing pains. It's a toxic direction being steam rolled by highly recognized representatives of Stability AI.

Leads to community members with clown hot taks

[tribal lantern]

it's a copyleft license... so yeah, you can't copy it without attribution and keeping the new code under the same license

[steady grove]

then @visual glade should take him to court. is he serious about it or just memeing? pretty sure licenses are attributed too

[copper kraken]

court for what? sue for what? waste of time and money for both parties

[steady grove]

guy here is claiming forge doesn't attribute licenses

[rustic garnet]

I have the feeling you enjoy this soap opera and "air of toxicity" way too much

[steady grove]

naw. i hate it. want to see comfy get it together and be more FOSS spirited

[visual glade]

forge is a hackjob that just took comfy code and slapped the a1111 interface on top

I wouldn't have any issues with that but he went and denied it

[steady grove]

Emad needs to muzzle you lol. You got your fanboy entourage going after people recommending forge on forums tehse days.

we're in a post "linus quit linux to work on diplomacy" world

[rustic garnet]

You talk about "air of toxicity" and just insult everyone. Maybe thinking about a muzzle yourself? Or just dial it down a notch

[steady grove]

Can't confront toxic topics without hurting some feels. Medicine shouldn't taste good. Poppins got it wrong.

[rustic garnet]

the problem is already that you think what you say is "medicine"

[steady grove]

Gaslighting. That's a cool way to approach it. Abusers protect abusers i guess. same ol story.

The future will continue as i predicted. The small niche of comfy users will continue "punching down" from their faux position of elitism, enabled by culture leaders.

[copper kraken]

🙄

[rustic garnet]

you are right, ego is the problem here 😉

maybe just switch the topic. The paper for SD3 came out today

seems like the T5 text encoder is optional. So you can use it for complex prompts, but you can also just use CLIP. Should help a lot, in particular you could leave out T5 for inpainting, upscaling and so on which should save a lot of time and vram

[copper kraken]

the modularity is a good idea

[steady grove]

i just hope the modules are named appropriately instead of the first stage being c and the last being a

[copper kraken]

now if you could still use a lora trained on a smaller model with a bigger one (just losing some ability to precisely reproduce the training set) we're in biz

i hope the smallest model is named Y, the middle size one is named X, and the smallest one named P

[steady grove]

parameter count would cause issue there potentially but we got xadapter tech now a days

[rustic garnet]

I would say that's unlikely, although not impossible. But I guess the loras won't be compatible

[steady grove]

i'm sure the numerous models will be impossible to understand how to load in an intuitive manner, but i can hope for a positive future

[polar hollow]

https://tenor.com/m4Ev996kw0E.gif

[torpid dune]

Hey which sdxl tool/model for placing a specific clothing product on a person/multiple differet people? As well as seeing it in different poses.
For example, given the (1) mockup image --> getting (2), (3), (4)

[shy kelp]

/jeune fille sur une balancoire

cinematic film still of a Beautiful Rococo Princess sitting on the Japanese metro, elaborate gown, massive curly red rococo hair, head and shoulders portrait, pretty eyes. Crowded metro with onlookers

[torpid dune]

?

[meager canopy]

#1047610792226340935

[coarse timber]

what？

[civic gyro]

To infinity and beyond! - https://twitter.com/HikariUchu/status/1765399437264998472

Hikari Uchu (@HikariUchu) on X

Embark on a journey with 'Starlight Harvesters: Humanity's Galactic Destiny.' 🌟🚀 #GalacticDestiny

#StarlightJourney＃アニメーション #animation ＃デジタルアート #digitalart #美人 #beauty #hikariuchu #visualart #vj #aigenerativeart #aiartcommunity ＃コミュニティ#未来的な…

Twitter · Mar 6, 2024, 3:30 PM

[elder plume]

did you promise to someone that this was possible?

[cyan crown]

https://civitai.com/models/336274

[upbeat summit]

that looks really great! nice work

[cyan crown]

thank you very much

https://civitai.com/models/336225

[upbeat summit]

you've been busy 🙂 I really enjoyed the work you've been posting over the last couple of months. Keep it up :]

[cyan crown]

doing in the night when family sleeps......

[small thicket]

的

[pallid peak]

办公桌子

[plain timber]

这满是脚气的鞋子是怎么做出来的？很牛的样子。

[quasi valve]

😆

[glass forge]

So will all SD1.5 and SDXL xheckpoints and promt method be totally obsolete in 3 months with SD3?

[sharp moss]

i need help

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.

what should i do to fix it

[craggy moat]

Hi, have prompt or Lora suggestions to limit the number of colors or shades? Eg. I want simple images, with as less color shades as possible in order to be able to easily vectorize

[meager canopy]

Why not just prompt for monochrome images?

[craggy moat]

that may be an option, however I would like to have color images, just somehow get ride of the shades/gradients

[meager canopy]

Muted pastel colour palette

Washed-out

[shy kelp]

Hey, sorry I'm new here,but ,Can I generate image in discord if yes where and if no,where should i ?

[meager canopy]

No #1047610792226340935

[native knot]

A breath of fresh air, though...someone actually asking before just lobbing random thirst prompts into the chat channels. Good on you.

[rustic shadow]

Most of us run it locally but there are lots of online generators! I don't know what they are though sorry

[vital ermine]

https://civitai.com/models/339571?modelVersionId=380226

[unkempt bobcat]

jesus and lucifer fighting , photorealistic

[meager canopy]

https://tenor.com/view/what-really-seriously-flying-super-man-gif-16080893

[copper kraken]

Here is the image you requested.

[meager canopy]

https://tenor.com/view/sarcastic-gif-20350005

[gloomy plover]

resting dog

[copper kraken]

Here is the image you requested.

[noble shoal]

Good Bot.

[gloomy lark]

On the next episode of Naruto....

[covert crane]

Hi there. Is stable-diffusion-xl-1024-v1-0 that I'm using with your dreamstudio/api equal to stabilityai/stable-diffusion-xl-base-1.0 from huggingface? I'm trying to get same results with your api and local build but images are different (same seed, prompt, negative prompt, cfg, checked different samplers).

[minor pecan]

iridocyclitis

[pallid path]

made a goofy ahh image a while ago

[glass forge]

Please make it so we can "lock" certain elements of a generation and reuse them... characters objects backgrounds...

[copper kraken]

Can't you do that already?

[glass forge]

how?

[copper kraken]

most directly, inpainting or outpainting, less directly, ipadapter, clipvision with cascade, or training a lora or embedding

[glass forge]

say you want a movie scene with 2 characters talking in a room. how do you lock each character so you cna generate them from different angles and how do you lock the room and the furniture in it so you can film it from different angles? coz if this was doable then you cna generate all this and put it in a motion video ai thing and hopefully promt up different areas like "lift hand" "walk here" etc and add lip-sync music and thats a movie in essence lol

more or less

small discrepancies could be ok , real movies have continuity flaws too...

maybe in 6 months ? a year? or maybe doable already idk with this AI stuff it moves so fast

you can generate up amazing things, almost anything but you cant make it move... no stories, just sad still images..

[cunning geyser]

Hey question

if im making a lora is it normal to only have 300 steps for a sdxl model with 100 images

and it taking 4 hours to do so

given its doing high quality over quantity

[copper kraken]

depends on your card...

re: speed

[upbeat summit]

testing Aetherverse XL Lightning by @icy brook

[icy brook]

Real nice. Very Patrick Nagel to my eyes.

[glass forge]

madness

[fossil turtle]

Does anyone know how to get A1111's break to work in comfyui

[vital ermine]

https://civitai.com/models/341251

[gloomy lark]

regional?

[native knot]

3 separate identified...no tricks.

Though Bats maybe laid on the lip gloss a bit

[gloomy lark]

hah that's impressive.

[native knot]

It's using the base Tempest checkpoint.

[gloomy lark]

https://huggingface.co/dataautogpt3/TempestV0.1

wow very cool, I realized I should be putting this here instead of cascade

[native knot]

Yeah

Since it's XL

[gloomy lark]

wow the clarity of this, since it can render at 1920x1088 natively is really impressive.

[copper kraken]

wow. incredible

[native knot]

Yeah...that first bat pic I put up above was pretty sick

[gloomy lark]

i'm amazed that it did it without subject detail bleeding

[copper kraken]

no kidding

[native knot]

Time to test something...

[gloomy lark]

yeah im running through my promtps

it looks like it doubles every now and then, but for the most part it's good

[native knot]

heh

[copper kraken]

can get some crazy shit by cutting down the early steps

this was ksamp adv steps = 20, run 0-5

then ksamp adv steps = 40, run 10-40

so fraction of denoising was the same but denoise rate per step went way up

[vital ermine]

I don't think puny man has a chance, but boy does he have hutzpah.

[jagged crown]

A Mercedes luxury sports car. Open top. Red

||A Mercedes luxury sports car. Open top. Red||

[jagged crown]

||A Mercedes luxury sports car. Open top. Red||

[jagged crown]

||A Mercedes luxury sports car. Open top. Red||

[jagged crown]

A Mercedes luxury sports car. Open top. Red

[jagged crown]

||A Mercedes luxury sports car. Open top. Red||

[jagged crown]

||A Mercedes luxury sports car. Open top. Red||

[jagged crown]

A Mercedes luxury sports car. Open top. Red

[jagged crown]

||A Mercedes luxury sports car. By the sea||

[jagged crown]

A Mercedes luxury sports car. By the sea

[jagged crown]

||A Mercedes luxury sports car. By the sea||

[gloomy lark]

lol

[gloomy lark]

Here is the image you requested:

[worthy orbit]

a ControlNet tile model for SD XL https://civitai.com/models/330313/tplanetsdxlcontrolnettilerealisticv1

still testing it

[native knot]

Bumblebee has sure grown up.

[gloomy lark]

he went back in time to defeat megacybershocktron

[gloomy lark]

really amazing. did you have to region prompt for the mirror?

[copper kraken]

this was differential diffusion

posted a few more in show n tell

the mirror was indeed in the prompt

[gloomy lark]

holy cow that joe87. those images where he's got rule of thirds going with the man and the dogs. I can't believe I haven't been sitting doing regional prompting 100% of the time. instead just sitting here getting angry at sd...

I could have been doing this the whole time. 🙂

[copper kraken]

lol

[gloomy lark]

seriously this darkimages model is better than most I've seen.

[copper kraken]

it's fn amazing

it's really versatile too

though i'll admit i'm hesistant when i hear versatile

[gloomy lark]

the high res nature and incredible small details. i keep finding stuff in the images.

[copper kraken]

that word to me screams jack of all trades too often, overtrained

but this one is flexible

[gloomy lark]

wow that's really cool.. i did learn 2 things tonight concerning this stuff. the 75 token limit still applies. I was just putting whole giant prompts in each section and it was dropping stuff. when i dialed it back, things reappeared. second, I was finally able to get the mecha to raise his fast and grab something by dedicating a whole column to just "closeup of mecha fist"

[copper kraken]

awesome!

yeah i kinda figured that'd be the case... i assume that applies even if the regions don't overlap

my guess is you'd need the two samplers node to get away with 150

(i often have these massive prompts that are way overflowing so i can cut and paste stuff back up at the top... my sloppy way of having crap ready to fire off)

[gloomy lark]

that's an interesting trick... having them rendered completely separately to get around the limit.

[copper kraken]

yeah, separate samplers

with differential diffusion you can go way over the limit

cuz you're using separate samplers over and over

now what i'm curious about too... this will have to be a tomorrow thing

is what if you use regional prompts WITH differential diffusion? will that allow differential to be even more coherent in terms of the final image cuz it knows what's supposed to be outside of its masked region? idk

darkarts + juggernaut merge

[gloomy lark]

right now I'm just using a1111 with regional prompter. creating rows and columns on the fly.

Although I'm blown away by your work, I don't think I can keep all those comfy nodes straight when it gets to that level of complication.

[gloomy lark]

hah man, who needs sd3

[copper kraken]

I'll prolly be doing that soon

Forge on my phone is a great way to ensure my SD induced insomnia can continue in the luxury of being horizontal instead of vertical lol

[gloomy lark]

yeah, i'm at the point where I need to make sure i CANT hit it from my phone. losing sleep (like right now) is the biggest issue.

[copper kraken]

Lol

[gloomy lark]

i'm looking through that show and tell thread. incredible stuff. none of that is ever shown on the stable diffusion subreddit. I made a post a couple months back on how regional prompter is the key to everything, but even then I had no real idea what it was capable of.

[copper kraken]

Guessing it didn't get much attention?

Subs always totally suck compared to affiliated discords IME

All the ppl who are hardcore about whatever it is will be on discord and so the best content always accumulates there

[gloomy lark]

I think the people on there are the general people, running off a 3060 or less, only willing to use a1111 in its base configuration with probably 1-2 models, if I can't do it with straight prompting, then that's just the end of it.

[copper kraken]

Yep

I'm not surprised but I'm still impressed how many ppl here have the 24gb vram

[gloomy lark]

when sd3 was announced, 90% of the conversation was "will it run on my machine from the 80's"

[copper kraken]

Lol yes

Not "should I set aside some of my beer money for a card in a few months"

[copper kraken]

A lot more life on this discord than most of the others I've hopped on

[gloomy lark]

hah soul is just posting pictures of his night activities as a way to tell everyone where he hid the body

[copper kraken]

much better merge

mohawk instead of juggernaut

mohawk can gcet pretty creative

[gloomy lark]

now every time i look through the samples section of a model, i look to see who probably regionally prompted. this was on the mohawk one

[copper kraken]

oh the latent noise lol

cascade has given me an eagle eye for that

kinda wish i didn't have that now

[orchid cobalt]

What is regional prompter?

[gloomy lark]

the one marked installed

[orchid cobalt]

That's something completely knew to me, so it allows to have control over composition

[copper kraken]

yep, it's amazing

differential diffusion is what i'm doing with these recent examples

[gloomy lark]

https://github.com/hako-mikan/sd-webui-regional-prompter

[copper kraken]

this was regional prompt

[gloomy lark]

but it's not that complicated.. you just start with adding 2 columns, then rows, then more columns, until you have something neat.

[noble shoal]

It has a certain vqgan+clip vibe with all the stuff thats happening, if you know what i mean.

[copper kraken]

more regional prompt

anddd back to diff diffusion

that's 80% darkarts 20% albedobaseXL2, just simple merge node

[orchid cobalt]

That's... insane

[gloomy lark]

now that ollama-webui has built in text to image, I can check the promtps before moving them into a1111 or comfy. and it lets me say "change it to a mall" or "add more beach balls" to shape it

[copper kraken]

That's fn awesome

[gloomy lark]

I think this city is sufficiently destroyed.

[orchid cobalt]

What set of 1111 bleeding edge plug-ins do you use?

[real heron]

a dog

[orchid cobalt]

Also, which SD model?

[gloomy lark]

regional prompter, reactor for face swap. sometimes controlnet. I've been using a lot of the clipvision stuff that batwing has been making workflows for in comfy

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

[orchid cobalt]

I see

[gloomy lark]

this is the new hotness for the day.

it has an really impressive amount of detail, and I like action shots, which it does better than many

tomorow it'll be something else. 🙂

[orchid cobalt]

Cool thanks

[copper kraken]

Here is the image you requested.

[gloomy lark]

Here is the image you requested

I'm sitting here trying to recreate this ideogram image of the cat in the plane. aaaaaand how it's going.

[copper kraken]

Lol figures

That's something I feel requires in painting when we're talking sdxl

Maybe fenrisxl would work

When it doesn't work, man... Some of those damn prompt fails have eaten up an entire evening

Then I look back and wonder why i even cared about that image and realize it was only because I couldn't do it

[gloomy lark]

exactly. i feel like i'm making progress.

it's funny though. when ideogram first came out, i put every thought and idea I had into it for the first few days. one and done, i'd put it in, and it would do it immediately. no countless generations to get it right. I realized I wasn't spending any time with this stuff anymore because I always got what I wanted immediately. Then I started realizing that the image quality is significantly below sdxl. and i eventually ended up back here.

which kind of goes against what I said a week or so ago.. that prompt adherence is king. it is, but I want it all.

[copper kraken]

Yeah, it is... Kinda

Ultimately there's shit like Photoshop and firefly

The ultimate I think is gonna be a bit of both

[real heron]

create a dot

[copper kraken]

Here is the image you requested.

[gloomy lark]

ok i'm not posting what it made

the badbot

[copper kraken]

Lol

[gloomy lark]

ok, gonna call it done. looks way better than the ideogram one anyway.

[real heron]

A Chinese woman wearing a sexy bikini,by the drive

[gloomy lark]

Here is the image you requested

I think I've determined that the badbot paired with that dark images model is a match made in heaven. the image it just made... jesus

It did this one perfectly, with amazing detail: Surrealist, darkly lit scene featuring a bloated, disheveled man with flushed face and bloodshot eyes, clutching an empty bottle in his clammy hand, surrounded by broken shards of glass and stale cigarette butts.

[copper kraken]

Man you got me curious now lol

Amazing holy crap haha

[gloomy lark]

yeah gonna delete that one, don't want to get canned before sd3. 🙂

[copper kraken]

Lol

[gloomy lark]

In this Impressionist-inspired depiction, a powerful Chinese woman radiates confidence and grace as she strolls down a bustling city drive, bathed in soft sunlight. Dressed elegantly in business casual attire, she embodies strength and self-assurance while cradling an endearing litter of kittens close to her chest. Her love for these feline companions shines through, highlighting the tender side of this formidable woman who seamlessly balances ambition with compassion.

an image even Google Gemini could make.

[orchid cobalt]

Self-assurance lmao

[copper kraken]

Here is the image you requested.

[gloomy lark]

hah the one on the right looks like one of Naruto's enemies

ok, i'll see ya guys later. ran out of steam

[copper kraken]

Nite!

[ruby zodiac]

https://tenor.com/tKgaYjwJq16.gif

[copper kraken]

@gloomy lark caroselloXL gamma is another very good checkpoint so far

[torpid dune]

? Do you have an answer or not, idk what you saying. I just asked a question

[icy girder]

The orange cat is wearing a bachelor's uniform, standing hopefully in the middle of the campus, holding a graduation certificate in his hand, with a joyful graduation ceremony in the background

https://tenor.com/bp3BK.gif

[meager canopy]

Using simple area prompt workflow

[static prawn]

any advice to get photorealistic stuff a lil more smooth

like it sometimes look to sharp for me

[gloomy lark]

use euler sampler.

[static prawn]

like here

[gloomy lark]

the SDE samplers are better with skin texture, the euler ones are worse.

so they'll be more rendered looking

[static prawn]

dunno its just tooo much dunno how to extactly describe it

oh i use 2m karras

[gloomy lark]

yeah, same line though.

[static prawn]

wait

sooo which one i should use?

[gloomy lark]

euler

for smoother.

less detail

[static prawn]

kk

i might try, but it only happens a few time

i like skin texture, but sometimes it looks oversharpen

[static prawn]

thats more my problem

[gloomy lark]

this is with euler

this is with euler

[static prawn]

i dont mean this, nono, i want fine skin details 😄

its more like some fotos get too sharp, too much contrast maybe

[gloomy lark]

oh, then use dpmpp_SDE if you want the best details. it's also the slowest sampler though, so people often go to the turbo models which take less steps of dpmpp_SDE

what CFG number are you using? if images are too high contrast or blown out, lowering the CFG can sometimes help.

[static prawn]

i feel like its hard for the AI if u have a light (ice for example) background

i get way better result if i prompt dark room or something like that

[gloomy lark]

different models have different CFG tolerance. some will get blown out at 10, some at 30

so if you drop the CFG to 5 that might help

[static prawn]

i tried several guess its my prompt atm

and in general always struggle when i prompt something with ice / snow

[gloomy lark]

porque no los dos. 🙂

[cyan crown]

https://civitai.com/models/341979

[prime mauve]

This could be super cool to translate to svd1.1 to add motion and speaking. Nice!!

[bleak totem]

Skilled archer, bow and quiver of arrows, standing in forest clearing, intense, detailed, high detail, portrait

Skilled archer, bow and quiver of arrows, standing in forest clearing, intense, detailed, high detail, portrait

/Skilled archer, bow and quiver of arrows, standing in forest clearing, intense, detailed, high detail, portrait

[copper kraken]

Here is the image you requested.

[bleak totem]

This is a basement of Qingdao Ocean University in 1950, The owner of this room is Professor Wen Shengchang, There are a large number of books and materials in the room, 16:9 horizontal composition, 4K resolution, wide shot, panorama, high details, 4K --stylize 0 --v 6.0 --ar 1:1

[boreal copper]

redraw the picture with 90 percent similar

[gloomy lark]

Here is the image you requested

[glass forge]

I smoked so much spice I learned how to fold space.

[upbeat summit]

nice stuff!

[vital ermine]

https://civitai.com/models/342998

[gloomy lark]

you don't want to mess with the boxing cat, he's just all fist

if I ask it to punch, it won't. i just have to say boxing match and just regen until the seed does what I want.

[copper kraken]

have you gotten it even with seed hunting?

seems like a good time for ipadapter maybe

[gloomy lark]

yeah.. the ability to do things where the subjects aren't upright, like an upside down face isn't possible in sdxl. I'm running into that. trying to get animals falling into a vortex, but it all just looks like they're dancing in the sky.

as you said, i'd have to controlnet it or something

[copper kraken]

Loras have issues with that and that's proof you're right imo

If you denoise an upside down face... Yikes

[gloomy lark]

dark arts images model + cute 3d render lora makes a good combo

[copper kraken]

lol great idea

that is such a great checkpoint

what i really like to is instead of tending toward soft images at times ilke most models

it has a tendency toward grittiness

[gloomy lark]

yeah the detail that comes out, even before upscaling is incredible

@copper kraken you see this one on their sample page? from clownshark, countless renders burst forth https://civitai.com/images/7420332

[copper kraken]

i'm really liking that other illustration checkpoint ig ot two nights ago

[gloomy lark]

3d render lora takes the edge off. 🙂 which illustration one?

[copper kraken]

https://civitai.com/models/253953/caroselloxl-illustrative-styles-mix?modelVersionId=380159

prolly the best one i've used yet

almost everything i've thrown at it has given really great stuff

[gloomy lark]

https://civitai.com/user/andreac75 this person posts on here as well, and they made the cute 3d render lora. they've also got some interesting illustrative ones

[copper kraken]

i really like the look of some of those... gonna have to check them out!

[gloomy lark]

dark arts vs. carousel. purple one is carousel

[copper kraken]

carosello?

[gloomy lark]

technically neon chemicals is in the prompt, which the purple is probably more adhering on

[copper kraken]

interesting how similar the outputs are there

[gloomy lark]

yeah carosello

[copper kraken]

i find the two to be very complementary

outstanding at two distinct styles

[gloomy lark]

so this is the carosello one with the cute 3d render lora, really brings in the qualities of her face and hands, without going overboard on the cutesy part of things.

[copper kraken]

it does, that's interesting

[supple hawk]

Paths in ancient forests, ancient symbols --auto --s2-imagine -

[copper kraken]

darkarts and carosello working together

[gloomy lark]

although you have to spice it up with something in the scene

[copper kraken]

the mask for diff diff for darkarts, the inverse used for carosello stage

[gloomy lark]

hah the roses are an interesting addition

I love seeing the engine trying to make sense of it. the mechanical tendrils into the snak

[copper kraken]

yep really interesting to watch

[copper kraken]

darkarts

[gloomy lark]

weren't those the enemies in the aquaman movies?

[copper kraken]

no iidea lol

that was a single stage of diff diff using this as the input image

[copper kraken]

ha nice

is that layer diffusion?

lack of a background...

[gloomy lark]

google images. 🙂 apparently none of these models knows black manta so SD wasn't helpful

[daring rain]

/prompt Tang Dynasty style icon design

[gloomy lark]

these are kind of nea

neat

[copper kraken]

Here is the image you requested.

[daring rain]

/prompt Tang Dynasty style icon design 2D

[gloomy lark]

yeah this bad image bot prompter thing definitely requires a cleanup in aisle 10

[copper kraken]

holy shit lol that's amazing

what moddel is that

[gloomy lark]

dark art images

that prompter and that model are made for each other

"Post-Apocalyptic Expressionist" depiction of a grotesque, deformed character covered in filth and grime, illuminated by a sickly green neon light amidst a desolate wasteland littered with debris. A scorched and rusty sign in the distance reads "Welcome to Tang Dynasty."

lol the prompt on that is drippin with sarcasm at the end

[copper kraken]

Here is the image you requested.

[gloomy lark]

hahahah

[gloomy lark]

in all seriousness, if you want icons or 2d art like icons, midjourney is really good at them, probably the best

do not taunt happy fun ball

[copper kraken]

that's amazing

omg

not boobies

[gloomy lark]

i dunno man, that one at the bottom got pretty close before a trapezoidal mask veered it away

[copper kraken]

verified, no nip slip!! lol

wild as f image

obv some issues in the lower middle part with some incomplete merging of the image but yeah

that thing at the center, the "head" was a giant shark bat

[gloomy lark]

so strangely enough, I put emad's tweet about there not being another major text to image model after sd3 into the prompt generator and it made this.

[copper kraken]

much better

wow

this shit is hopelessly addictive lol

[gloomy lark]

yeah and it's 3:30

am

[copper kraken]

i've always loved this kinda surrealistic art

yep...

i just noticed that

i blinked it and it went from 1 to 3:30

totally in the zone here

img2img

[gloomy lark]

hah yeah i saw that yesterday. the left looks like a north korean propaganda poster

[copper kraken]

lol 😄

that was i think my first generation with carosello

love at first sight

[gloomy lark]

good point, let's see what it does with this prompt

"Neo-Surrealist" depiction: A luminous, otherworldly creature emerges from a swirling vortex of digital colors, its ethereal form shimmering in the electric glow. A massive neural network sprawls behind it, adorned with intricate patterns and glowing nodes that seem to pulse in sync with the cosmic energy. In the distance, towering structures of abstract geometric shapes pierce through the surreal skyline, casting long shadows on the bustling cityscape below.

[copper kraken]

holy hell that is great

i really like the compositions it generates

def listens to prompts better than most

[gloomy lark]

this one is dark arts

[copper kraken]

i've found in genearl the anime models are pretty good there

[gloomy lark]

same seed

[copper kraken]

also spectacular^

[gloomy lark]

I just tried this on a couple other models and it doesn't make the "creature" part of it, so these 2 are definitely more prompt adhering.

[copper kraken]

yeah

what i'd like is a comfy node that would allow me to just dump a list of my models from the terminal and copy paste them in then gen the xy plot

no way i'm going to be bothered clicking through 100+ of them on the efficiency node one

[gloomy lark]

so these are ideogram and MJ for comparison. SD does a better job in a lot of cases.

[copper kraken]

no kidding

[gloomy lark]

yeah I can do that with my scripting, where I have an sdxl comyui api json, then I convert that to a powershell object, change the model name, convert it back to json and run it. one could pretty easily create an array of all the model names, and then just have it run it.

[copper kraken]

i'll see what dalle3 does

[gloomy lark]

I think there's going to be a LOT of similarity between models though, so there may not be a benefit other than ones that are very different from each other

[copper kraken]

so much for the great dalle3... waiting for one more turd

awful lol

generate 4 images with starting the batch, showing 1 and ending the batch and starting a new one with the next one and no text in between each batch and so on til all in all i see 4 images Topic: "Neo-Surrealist" depiction: A luminous, otherworldly creature emerges from a swirling vortex of digital colors, its ethereal form shimmering in the electric glow. A massive neural network sprawls behind it, adorned with intricate patterns and glowing nodes that seem to pulse in sync with the cosmic energy. In the distance, towering structures of abstract geometric shapes pierce through the surreal skyline, casting long shadows on the bustling cityscape below.

that was the prompt

the gap between quality of those outputs vs what you just posted in here is greater than sd15 vs sdxl

[gloomy lark]

these are dalle in highest quality mode

[copper kraken]

those are badddd

[gloomy lark]

they like the swirls.

[copper kraken]

yuck

lol

those are also awful imo

also poor prompt adherence

no sign of circuits in the first

[gloomy lark]

what they are, is censored. anything that might look scary etc isn't going to be shown on it.

[copper kraken]

vague grid stuff in the second

yeah true

[copper kraken]

what model?

[rose smelt]

Playing around w/ training/mixing - but I have a couple of LoRA on civitai that are the same concept (just could replicate exaclty): https://civitai.com/models/313770?modelVersionId=352057

[copper kraken]

sweet

really liked the look of taht one

[gloomy lark]

https://civitai.com/models/145221/sdxl-blow-it-up-explosion-artstyle

[copper kraken]

now THAT is my kinda lora. holy shit

in the words of the great @noble shoal , day is ruined now lol

[gloomy lark]

you can't be dropping this stuff at 3:42am my guy

volcanos explode good with that lora

[rose smelt]

Almost Nirvana at the end of the world

[rose smelt]

Cali......although clocks jump tonight...

[copper kraken]

that's quite the scene

[gloomy lark]

yeah dall-e has more to say about that one than the previous one

[copper kraken]

what's the prompt for that?

[gloomy lark]

mine was: anthropomorphic volcano island belching forth green smoke. Sea creature warriors surround the volcano in anger / it expanded it to: Imagine a scene where an anthropomorphic volcano that takes the shape of an island, spews forth green smoke. The volcano, dramatic and imposing, appears to have human characteristics, giving it a sense of sentience. A range of sea creature warriors of different descents such as an Asian serpentine sea dragon, a Middle-Eastern giant sea turtle equipped with armor, a Black mythical kraken with a multitude of tentacles, surround the volcano in an uproar. Their expressions filled with assertiveness, reflecting their combined force and anger against the erupting volcano.

I was watching the new aquaman movie so it's based on one shot from the movie

yeah the exploding lora is pretty neat.

[gloomy lark]

haha it's poster art for the second matrix movie

[rose smelt]

That's a killer image - if you ever post comments on civitai in reply would be cool if you did w/ that one.....dig it....

[copper kraken]

dling myself

[gloomy lark]

sure, it's now posted to your page

[copper kraken]

img2img

[gloomy lark]

lol that's awesome

no army would want to face off with that thing

this is with the dreamy lora added. that poor boy wants to add fuel to the fire.

[copper kraken]

img2img

early iteration from that lol

swapped some of the masks and got this instead

[gloomy lark]

that's quite the chin strap beard on that middle one.

[gloomy lark]

with and without dream vibes lora

hahah and now with explosion lora

he's having quite the blowout

[gloomy lark]

I'm loving those middle few. that dripping glass under the cork is really cool

[copper kraken]

yeah sometimes the midway through ones are the best

i make sure to save at every step

The best ones would be ones where you stopped at every step

Reconfigured the masks, rewrote all the regional prompts, then did the next step

Etc

What I've done here is kinda approximate something that works so I can industrialize art production lol

didn't even prompt for bat sharks... lol

[gloomy lark]

what's your ability to change resolutions? if you wanted to go widescreen, could you?

[copper kraken]

haven't tried but i bet it could be done

giving it a shot with a sloppy attempt

1280x768

[gloomy lark]

1152x696 is typical

[copper kraken]

i've done tiled upscales with these np

[gloomy lark]

sure or that

ok, after 4am, that has to be my cut off. I'll see ya next time. 🙂

[gloomy lark]

veeery nice

[copper kraken]

got one native render coming then i'm off to bed too

[gloomy lark]

that widescreen opens up more room for chaos

[copper kraken]

wild party of barbie clown freaks dancing in a zoo with nuclear bombs going off and german tanks storming a village of penguins next to the ocean at night with a burning comet impact

^^input image

that last one i just stretched one

[gloomy lark]

random question. do you know if there's a way to fuse a lora with a checkpoint? so i could just render against that new product and have it be at checkpoint speed instead of lora + checkpoint speed which is slower?

[copper kraken]

that is indeed a good q

i feel like i heard that could be done?

i know you can extract them...

[gloomy lark]

https://www.reddit.com/r/StableDiffusion/comments/11claq7/can_you_merge_a_lora_into_a_checkpoint/

yeah looks like you can.

[sudden lodge]

swarm has a gui for that

[gloomy lark]

ok neat, thanks

[noble shoal]

Are you using comfy?

Could work

[pine topaz]

/im

sefv

[shy kelp]

?

[jolly stirrup]

wild party of barbie clown freaks dancing in a zoo with nuclear bombs going off and german tanks storming a village of penguins next to the ocean at night with a burning comet impact

[mossy salmon]

joker

[meager canopy]

Her ya go

[native knot]

Bataman 👌

[noble shoal]

@native knot I have improved my workflow. Example is in #🏞｜general-with-images . I hope the users are satisfied.

[noble shoal]

I am very proud of my newly written "Save JPEG at 30% quality" node.

[meager canopy]

You know that already exists in the WAS pack?

[noble shoal]

Well, obviously not, i would say. 🤣

[meager canopy]

Sorry 🤷🏻‍♂️

This is handy too

[noble shoal]

Ok, now i have to say something bad about the WAS implementation to feel better. It's way to bloated.

[meager canopy]

It's "feature rich" 😄

[noble shoal]

@native knot There is a guy who has send a prompt in #🌠｜show-and-tell . I can't serve at the moment, because i am finetuning a 1.4 checkpoint (don't ask why)

[meager canopy]

https://tenor.com/view/good-rainy-morning-gif-13545928996254568915

[gloomy lark]

Here is the image you requested

That's actually a rather good prompt : Art Deco style scene, featuring garishly colorful Barbies dressed as wild, manic clowns, flailing about in a dismal zoo amidst deafening explosions from nuclear bombs and thunderous German tank rampages over a devastated penguin village by the moonlit ocean under a blazing comet impact's smoke plume.

[copper kraken]

Lol that person echoed the random prompt I punched in to generate a 1280x768 image for my img2img workflow

Kinda weird to have an echo lol with no other comment

The Bot Seekers are a strange lot

[gloomy lark]

indeed. i expect a new DM from them. that said, one of the ones I responded to the other day sent me a dm with "fatso"

Lykon keeps posting these portraits (he doesn't post literally anything more complicated, sigh) but yeah, that 16 channel vae is something special

it's funny, we've been chasing skin detail, but with that sd3 vae, even the ones he posted where it's a little blurry and there's no skin texture, still look way more like a photo than anything from sdxl

I believe, it is possible to say yes to more detail too many times.

[copper kraken]

Haha remember last night we were like whoa, wtf, how is it suddenly 3:30? ...daylight savings

[vital ermine]

https://civitai.com/models/343800?modelVersionId=384855

[copper kraken]

Can't see the workflow (jpg)

[copper kraken]

that's a diff style of color bleed than i'm used to seeing with sdxl

[pallid path]

⚠️ Light Zombie Ketchup Gore Alert ⚠️

Open Original and zoom in for the intricate details

[noble shoal]

Ohhh, yeah, that jpg is part of the workflow. I meant more like: You can see the product of that workflow in....
It's littered with self-written nodes that are available nowhere.

[copper kraken]

dang

in a month or two i'll have a lot of extra time on my hands and get into that

writing nodes, that is

there's so many things i want to clean up or change

countless nodes that desperately need that One Extra Feature (customsampler... my god, that needs an advanced version)

i think i have some idea of what kind of "fine" tuning it is you're up to.... 🤣

this is my message to the bots

[noble shoal]

😬 Oh yeah?

[copper kraken]

just curious, what kinda nodes did you add? wondering if we're seeing the same gaps

[noble shoal]

-A prompt style hat makes everything just worse.
-A 30% JPEG quality save node (got to hear later on that that already exists)
-I think an altered version of an Text on Image node.
Defiantly not the same gaps, i guess.

[copper kraken]

gotcha

yeah the lack of an advanced custom sampler drives me nuts... you can do really cool shit by changing the rho value for karras and polyexponential schedulers

the other thing i would really, really like to have is some kind of model merge node, even a simple one, with timestepping

[noble shoal]

I think we have that. Not with timestepping, whatever that is

[copper kraken]

^as in, first 0.3 fraction of steps it's 0.3 strength, next 0.3 it's 0.6 strength, etc.

starting vs ending strength

there is a tileddiffusion node that ports over the style from webui but i swear the tiled ksampler gives better results (at least without regional prompts)

[noble shoal]

You mean strength of the model?

[copper kraken]

the strength of one vs the other, yeah

[noble shoal]

Oooof. Ok, that's a usecase i personaly have no use for. But if you simply would like to merge them or merge loras into the models, you can just do that.

[copper kraken]

yeah, you can timestep by chaining advanced ksamplers

but then the interface starts to lag with too many nodes out

[native knot]

If you want to mess w/something interesting, do some model subtraction.

It's essentially a way to create a LoRA that you pull out of a trained model.

[copper kraken]

oh? go on 🙂

i haven't done that... nothing beyond "modelmergesimple"

[native knot]

You grab the trained model, subtract the base, and whalla! You've got a LoRA.

[copper kraken]

oh wow that's all there's to it? makes sense i s'pose

we do already have this: https://github.com/asagi4/comfyui-prompt-control which does allow timestepping with a lora... so that's really good to know

[native knot]

There's a few interesting use cases...so for what aimingfail is doing, he could re-merge the LoRA that is the net-result of something he didn't want back in at a weight of -1, pushing the model further into bad results.

[pallid path]

Is it effective in keeping characters and styles well?

[copper kraken]

merging within comfyui or outside of it?

[native knot]

You can do all this in comfy

And you can save out the models.

[copper kraken]

yeah, just wondering if there's any downside to using those nodes vs something else

[native knot]

It's all quite interesting.

[copper kraken]

very!

so merging with a weight of -1... which node were you using for that? subtract again?

[native knot]

Just search for subtract...you should see the merge subtract node in the list.

[copper kraken]

gotcha, yeah just checking that's what you meant 🙂

anything interesting you've seen with modelmergeadd?

[native knot]

That's the basic merge model, right? I've used it in the past. But ttn makes a much more interesting multi-model merge node.

[copper kraken]

that one i have but haven't used

Some of the conditioning nodes are interesting but I haven't fully elucidated their use

Conditioning multiply, concat vs combine... Etc

[noble shoal]

Oh they are interesting

[copper kraken]

i'd love to hear if you've found specific uses

[copper kraken]

but how's it diff from combine?

i've done stuff like that a bunch... but with combine

[noble shoal]

You have one conditioning from 0 to 30% of the time. And then it switches to the other

[copper kraken]

yeah, i mean

this works pretty well... just don't know why/where/when i'd use concat instead of combine

guess i really need to just start looking at the code

[noble shoal]

I don't know what the Impact Pack combine node does. I am talking vanilla nodes here

[copper kraken]

i think there's a vanilla combine too

the impact one is identical except just keeps adding inputs instead of being stuck with two

[noble shoal]

I think it just smashes it together

[copper kraken]

is order of operations important with concat, like how the order of tokens matters?

[noble shoal]

I am honest and don't even pretend. I have no clue. But back in the days, i made weird stuff with concat and setting time step ranges

Because you can just switch the whole prompt in the middle of it all

[copper kraken]

gotcha

order of op matters

[noble shoal]

By the way. I hope the guy can use your images for his scientific work. Good work there.

[copper kraken]

two pos prompts combined using the combine node: "a car" "a shark"

order didn't matter here

concat node: "a car" -> to, "a shark" -> from

swapped: "a car" -> from, "a shark" -> to

a single prompt "a car shark" "a car a shark" "[a car|a shark]" "a [car|shark]" etc don't come close to replicating the result of course

"a car AND a shark" in a1111

https://github.com/comfyanonymous/ComfyUI/issues/1403 ah here we are

a car
BREAK a shark

^^in a1111

someone was asking about break in a1111 being avail in comfy... gonna go get them their answer now

[copper kraken]

https://github.com/comfyanonymous/ComfyUI/issues/1403 conditioningconcat node

[limber kettle]

icon for video maker app

[limber kettle]

:.(

How can i use

[native knot]

#1047610792226340935

[limber kettle]

Thnkx

[copper kraken]

Here is the image you requested.

[limber kettle]

Can you help me generate icons

[native knot]

One other thing...if you either adjust your KSampler so that the sampler/sigmas widget(s) are inputs or if you use the SamplerCustom node, you can then apply the KarrasScheduler and PolyexponentialScheduler nodes to adjust the min/max for the sigma and the rho. 😉

[copper kraken]

conditioningaverage is pretty interesting but very finicky with the strength

[copper kraken]

i will be doing that for sure

actually just started looking at the code

not a python programmer, almost all my background is asm/C/C++... did a lot of RE and sploit stuff back in the day

but should be doable

[native knot]

If you pull down the ComfyUI Extra Samplers custom nodes, there'a a neat little SamplerCustomModelMixtureDuo node you should check out.

[copper kraken]

ooo, anything in particular you've done with this?

[native knot]

I just discovered it not too long ago...trying to set up some tests.

[copper kraken]

i've installed every sampler i could find

don't know wtf to do with a few of them yet

ksamplercycle is on that list

[native knot]

Check the node out:

[copper kraken]

slightly diff results but otherwise yeah that was easy

samplercustom however does replicate ksampler there

[native knot]

Ugh...keep getting a 2 device error trying to use this.

I wonder if it's hitting a vram limitation and then moving something to cpu. :/

[copper kraken]

how much vram you got?

[native knot]

10

[copper kraken]

drop a workflow here and i'll run it and let you know what i hit

[native knot]

It's a first gen 3080.

[copper kraken]

i've got a 4090

(and whether it runs or does whatever else)

[native knot]

Super basic just to test...

[copper kraken]

just installed the mikeynodes will test in a sec

did you modify your sampler like i was talknig about? the custom thing

[native knot]

I just retitled the nodes...nothing code-wise.

[copper kraken]

i just realized my naive as f hack change just causes it to default to euler/normal

Error occurred when executing SamplerCustomModelMixtureDuo:

'SDXL' object has no attribute 'model_options'

hit 15.6gb vram

[native knot]

:/

Was hoping to get something interesting...but not red text.

Damn.

[azure oxide]

Woahhh mikey's got his own nodes now?? Niceee

[copper kraken]

wait... did you mean that you got a ksampler advanced that accepted samplers/sigmas as output or are you referring to something else

[native knot]

You asked if I modified the sampler, I said no, I just re-titled it.

[copper kraken]

what effect did re-titling have

[azure oxide]

They sound cooler

[native knot]

It literally changes the title of the node. That's all.

[copper kraken]

ahh k just making sure there wasn't something i was overlooking

or misunderstanding

[native knot]

Can you explain this a bit more?

[copper kraken]

oh, i'm just looking to make samplercustom be able to behave like an advanced version where you can turn off add noise, and do something have it run steps 10-20 out of a total of 30 steps, etc

return with noise

that or make kadv able to take samplers/sigmas as inputs and actually pass them on correctly

[native knot]

Ah...yeah, I think the KSampler (Advanced) node's sampler and scheduler input will only accept primitives. @visual glade could really stand to adjust those to accept the KSamplerSelect and various Scheduler nodes.

[copper kraken]

yeah it would be really handy

espec with stuff like the res momentized sampler...

[livid basin]

I don't quite understand why they trumpet the reduced latent resolution if they increase the depth

Shouldn't the cost be proportional to channel * internal height * internal width

If width and height go down, but channel goes up, is there really a net savings

[copper kraken]

Yeah spent a bit looking through that code, gotta admit I prolly don't have time to fully wrap my head around it for a while yet

Not sure which approach would be easier - modifying the custom sampler or kadv

Guessing the former

@elder plume would this be something you might know about ?

[livid basin]

I guess it does kind of make sense if they use self attention, then the cost would scale to the square of latent height/width

[prime bison]

Dreamshaper XL

[cyan crown]

https://civitai.com/models/343700

[cosmic quiver]

What do you think guys?

[elder plume]

anything is possible it's always a question of what will be merged 🙂

[copper kraken]

Yeah

If I can figure it out I'll just do it as a custom node

[gloomy lark]

I'm gonna need another node to clean things up before VAE...

[gloomy lark]

I think this lora help me get to where I wanted it to be. 🙂

[glass forge]

Now put Will Smith there.

[prime bison]

I only prompted for hair made of noodles, and he doesn't have any

[gloomy lark]

Here is the image you requested

will smith goes through an impressive transformation for his next role as he uses method acting to become, the spaghetti.

[native knot]

The only roll he should ever play in again.

[glass forge]

😆

[gloomy lark]

in this candid shot, will smith goes in for the slap, but realizes he's much taller than chris rock

[glass forge]

Shit, what's this? MIB 3? I mean mib 5 coz there is a mib 3 qand 4 apparently? Idk i sotpped watching movies after Tranformers came out... more or less

Cinema died at that point XD maybe AI will revive it

[gloomy lark]

i forget if he was in the last one, it was chris hemsworth and the woman from westworld

[gloomy lark]

Last pic before bedtime, just wanted to extoll the virtues of image to image in SD. the bad quality one is from ideogram, which has amazing prompt adherence, but generally pretty low quality visuals. SD to the rescue (dark arts images checkpoint) with an amazing restoration with iterative 0.35 denoise upscaling.

[copper kraken]

got one for ya in case you need any help getting insomnia

that's with my dual regional prompting + diff diff workflow

decided to take a crack at my original simpler diff diff one again...

diff diff is king

also, used my mistral llm to enhance the prompts... def was a quick and easy way to get better results

alright now i'm really liking this

used conditioning concat for the third pass, instead of a new conditioning

[rustic shadow]

This neighbourhood was so nice before the demons and constant lightning 😦

[copper kraken]

some second passes...

[valid basalt]

https://tenor.com/bElDl.gif

https://tenor.com/bIHSJ.gif

[smoky patrol]

Love these!

[prime bison]

a sad cake in the toilet

[noble shoal]

A lot of chocolate sauce you got there.

[prime bison]

Sure, call it chocolate sauce if that makes you happy

IDK where I am

[peak dove]

Yumm!!! Where's my knife & fork?! 🥳

[crisp owl]

Some already have a knife hanging somewhere close to the toilet if you're lucky

[rustic shadow]

good ol poop knife

[copper kraken]

Here is the image you requested.

[copper kraken]

just crazy how good differential diffusion is

[molten breach]

What exactly is differential diffusion? I've seen a few references to it

[eternal fog]

https://differential-diffusion.github.io/

It allows you to use masks to edit small portions of images in a way that doesn't distrupt the things around it, if done properly.

[molten breach]

Nice, I'll check it out

[copper kraken]

the rocky mountains viewed from... directions

[native knot]

Are these from you diddling around with the regional prompting workflow again?

[eternal fog]

The differential diffusion does work pretty well most of the time

[copper kraken]

diff diff and rotating the image between multiple iterations

shit, it figured out it was upside down

i always wanted to work in an office building

[vital ermine]

https://civitai.com/models/346397

[meager canopy]

Big upscale on these 8 images below. Zoom in and enjoy the view! 🙂

[solar merlin]

been making a few environments and really loving nature compositions, anyone have any tips for upscaling strange perspective images like these?

[onyx tartan]

This was done by SVD?

[warm forge]

nope, both it and the sea turtle gif next to it are random tenor gifs lol

[alpine sun]

||gil||

[jagged cobalt]

19/2000 Logo with talent element

[copper kraken]

Here is the image you requested.

[wind hill]

The fractal art of a pentagram

[shut lantern]

can you let me know which model is it? Does we need to use ADetail to get this result?

[meager canopy]

It's my UltimateBlendXL v2 on civitai...which doesn't seem to be working. Just search my name on there, when it works.

It was a while ago since I made those, and I honestly can't remember if I used the face detailer or not.

[leaden drum]

prompt: In a room, there is a man wearing sunglasses, a gold necklace around his neck, and a big cigar in his mouth in front of the coffee table style: Anime aspect-ratio: 16:9 character: None

[rich smelt]

hairy snake

[minor pecan]

[leaden drum]

how to use?

Which one is the beginner's room? How do I send messages and draw pictures? Thanks!

[slim wren]

You either install it locally or use online services. Right now the bots are down, cf #1047610792226340935

[leaden drum]

@slim wren Oh! So it can't be used here? Thanks!

[willow dagger]

My only grievance is that it VAE encodes the whole image and thus also changes the "black" parts of the mask.

Did anyone create a nice Workflow to circumvent that?

[copper kraken]

yeah i've been doing that for a while

just use latentmaskcomposite to patch it back in

[pallid path]

@shy kelp DreamShaperXL lightnign with highresfix

6 step first + 4 step highres

[willow dagger]

Can you share a simple Workflow showing how you set it up?

[shy kelp]

Why did I see low tier god here

[pallid path]

lmao

[shy kelp]

it's the lightning bolt

[copper kraken]

sure

workflow embedded

[valid basalt]

a fish is playing game in the ocean

[gloomy lark]

Kind of neat. I've got a prompt that attempts to split an image up into 6 pieces and create a regional prompter compatible prompt automatically for me based on a given regular prompt. GPT4 works on it maybe 20% of the time. Claude 3 just did it perfectly on the first shot. The requested prompt: a cat with a red hat playing baseball with a dog in a green jacket.

[livid narwhal]

New here, is this a good place to ask for poses/etc?

I see all sorts of human ones, but I can't seem to find any animal poses for 4-legged friends and whatnot.

[copper kraken]

reason #1 million why comfyui is such a good thing to learn how to use... this image

tried like 15 times to inpaint the shadowy figure on the porch in forge/a1111 without luck

got it in the first shot with the first seed in comfy and i'm done

[gloomy lark]

Hah nice

[gloomy lark]

could always just region prompt it. 🙂 claude 3 is apparently awesome at it. i'm sdxl'ing one of the SD3 shots

[copper kraken]

I've had issues trying to region prompt something like that that is that small

Claude 3v

?

[gloomy lark]

This was the #SD3 shot that lykon posted

A black panther leaping, in the style of John Howe ADDCOL
A blue elf with white flowing hair, in the style of John Howe ADDCOL
Yellow eyes and a blue scarf , in the style of John Howe ADDROW

A black panther leaping, in the style of John Howe ADDCOL
Blue armor and a wielded sword, in the style of John Howe ADDCOL
A stone valley, in the style of John Howe

[copper kraken]

Ahh gotcha

Nice work

[gloomy lark]

the key is having vertical columns that are the same subject, so you can have horizontal granularity, but have a subject inhabit a whole 1/3rd of the left to right

I want you to create a terse text to image prompt. For that image, I want you to split the image up into 6 pieces; 2 horizontal rows of top and bottom, and 3 columns per row, left and middle and right. I want you to describe what is in each piece, starting from top left, then top middle, then top right. Then for the bottom row, bottom left, bottom middle, and bottom right of the image. Use the word ADDCOL to delimit columns and ADDROW to delimit rows. Don't add words denoting which image piece it is. Don't mention more than one subject per image piece. Put each image piece prompt on its own line. A subject is allowed to span vertical image pieces. Determine an appropriate artistic style for the overall image and mention ", in the style of " with that style at the end of each prompt line, but before the row or column delimiter. Please make a text to image prompt for: a black panther is leaping to the left of a blue elf who has white flowing hair, yellow eyes, a blue scarf, and armor while weilding a sword. backgorund is a stone valley, which they're both at the bottom of

gpt4 couldn't really handle this complex a prompt. but claude did it

[copper kraken]

Wow nice

I haven't tried Claude 3

[gloomy lark]

yeah it was just released a week ago and it's probably 20% better than gpt4, which is obviously a big deal since that was the king before this.

[copper kraken]

Can it count

Yeah gpt 4 is pretty great I use it daily

[gloomy lark]

good question, I don't know. 🙂

[copper kraken]

I'm wondering if it was able to handle that prompt because it understood the concept of a grid and could count

Gpt 4 def can't

[gloomy lark]

apparently transformers based models can't tell how many tokens an answer will be, or conform to that, so getting sub 75 token prompts from these generators is the tough part.