✨｜sdxl

[nimble heart]

even the Porsche eyes are asymmetric in SDXL :(

[bright valley]

That's pretty awesome! Personally I've never tried the qr stuff yet, or seen somebody use it with my lora before. It's a really cool idea, I'll certainly have to give it a try now

[eternal fog]

Has anyone been playing about with the attention masking in IPAdapter, it's pretty cool

Prompt : Blonde Woman wearing a long jacket standing in a Tokyo Street next to a billboard

Input was a screenshot of a view in Elden Ring.

Then for the masking that the IPAdapter would be applied to I used a very rough billboard shape to the left of the image.

[lusty moss]

IPAdapter has been giving me headaches lol. Looks cool though!

[eternal fog]

It doesn't always work, it depends on what you've masked and the prompt. For example I just masked the entire right hand side of the image.

That's apparently too much, because now it's basically just put a copy of the input image on the right hand side lol

[lusty moss]

Universe flea lol!

Same prompts, different seeds

[eternal fog]

Took your alien and applied the IPAdapter to only the left side of the image

[lusty moss]

They seem happy 😅 lol

[eternal fog]

I applied it on the top quarter and strange things happened

[upbeat summit]

so a quick method to do latent space mash-ups

[lusty moss]

Their child 15 years later lol

[eternal fog]

Yeah, if you use 2 IPadapters you can mash those together too. The dev of the extension did a pretty good example video. https://www.youtube.com/watch?v=vqG1VXKteQg

[lusty moss]

Trying to create a quick set of alien/space backgrounds using chatGTP and ComfyUI.

[upbeat summit]

could you show the input images?

[eternal fog]

I've come off my PC to go to bed. But that one was an image of Malenia from elden ring. And then I just used the mask to tell it only to use the adapter in a person sized area in the middle of the generation.

[upbeat summit]

very cool! thanks for posting those

[eternal fog]

It can do some pretty cool things. I did one earlier of a corridor as the prompt and then I masked down the centre and used an IPAdapter input of a flowerbed and it made a corridor with flowers on the floor down the centre

[steady grove]

metroid vibes

[crisp owl]

I keep wanting to use the lcm, but the drastic decrease in quality stops me each time. Tried several work arounds, but meh, not worth it.

[sharp robin]

did u lower cfg to like 1.7?

[crisp owl]

Oh yeah, settings are fine on it, changed em to what they recommend. But for realism focused stuff, it loses a lot.

[peak dove]

It works very well on more artistic output, I've been using it alongside other LoRAs and its quite sharply detailed. I was thinking FreeU might help, as it works well when a low number of steps are involved - it gives good high-frequency detail. I will try linking LCM to FreeU ...

[vital ermine]

I am not a big fan of lcm either and my fear is most models will be trained in it and we lose quality.

[vale eagle]

The lcm lora has quality drop but I want to try the lcm distill from an sdxl model

[nimble heart]

if you can run regular SD at even 20 steps in a somewhat timely matter it looks far better than LCM

[supple knot]

how do you grey out or turn on grey out parts in a workflow?

[crisp owl]

ctrl+m is mute

ctrl+b is bypass

[supple knot]

thanks

[nimble heart]

bypass is probably what you want because it allows connections to still work

mute's only really useful for output nodes

[supple knot]

your right bypass works better

[supple knot]

The bear a first view .Is it farting? ,no its flying

[vital ermine]

Probably got into a Taco Bell dumpster

[supple knot]

Earlier that day

[peak dove]

SDXL @Distillery

[ionic gulch]

two girlfriends in front of a landscape with a lake and cherry blossom trees (top/bottom masked ipa)

[zinc cargo]

can i get some advice about ipadatpor? i cant really the same person like everyone else does.. not even pets..

what's the trick?

[eternal fog]

What are you trying to do? It's pretty simple, you just use an input image for what you want to create and you can use prompts to change it about a bit.

There's a few different models, a regular one that does lighter changes, "Plus" models that follow the image more closely, and face models that are better at copying faces.

[zinc cargo]

end of the day, i want to take my brother in law's cat and put him in a traditional Thai clothings.

(dont judge me)

and i dont want to use a lora for this

[worthy orbit]

I tried, got this error https://huggingface.co/williamberman/sdxl_controlnet_inpainting/discussions/3

[fierce hollow]

yeah I was getting it too, guess it doesn't work

[steady grove]

goooddddd damnit. the first time i ever try uploading a model to civit and it pulls the ol crashed server on me. i spent an hour typing it all up and then posted , error 503. now the site has been trash load times and unworkable. so much for posting a lora today

the new investors want them to be more cash positive so they're cheaping out on their literal main function, hosting

if i knew how to do web stuff better, i think i'd try building an ai torrent bay. Hosts like civit were a flash in the pan, and now that investors are deep into it it's all going to get enshitified

[eternal fog]

If you want an exact copy of the cat you'll need a Lora, without you can get fairly close. Although if it has a unique pattern it might be difficult.

[crisp owl]

Civitai has been pretty bad the last day and a half, I'd wait another day for them to settle things

[crisp owl]

Some interesting outcomes trying to get an image of an eagle carrying a man 😆

[steady grove]

friend insisted i published this but i dont think it'll survive the week. Civit's been doing dmca takedowns of celebs https://civitai.com/models/201347?modelVersionId=226611

[native knot]

How about this one? 😉

[steady grove]

thats how LOTR should've gone

[native knot]

[crisp owl]

Ooh nice, what'd you prompt??

I was just bring ing one into photoshop to mess around with, but would prefer to have one gen'd lol

[native knot]

Giant eagle, talons holding tiny man warrior in the air

Neg: watermark, signature

Using JuggernautXL v6 with the add-detail LoRA cranked to 10.

[crisp owl]

jeez when did juggernaut release v6 😆 So much to keep up with haha

[native knot]

6 came out on October 25th. 🙂

[crisp owl]

Somehow i missed that 😆

[steady grove]

juggernaut was what i was using mostly until i got ahold of think diffusion. both are really great

[crisp owl]

I've used that model as well, good one yeah.
I use ZavyChroma a lot as well

[steady grove]

probelm with sdxl refinements is that they'll be different from each other, but also SO versatile still, so theres little reason to switch between them

those huge latent spaces take a while to explore

my friend just found out i trained that pornstar model without any sex or nude scenes and he's legitimately pissed now and calling me a schizo when i explained my reasons

think i am losing a friend today

i wanted to publish to civit and see if it wouldn't get taken down. Without nudes i think it is more fair use, so we'll see

also, porn made with AI is just dumb. it's so broken and demented.

[crisp owl]

porn for the sake of "porn" yeah, but for a more artistic approached nude photo, can make some beautiful images

[steady grove]

give it some of that heffner style

reminds me, i wanna do a classic pam ai from her home improvement years

[native knot]

Y'all talkin' 'bout porn and I'm over here just trying to get something that looks more like Falkor.

[crisp owl]

I've mostly been making images of norse tales

[native knot]

This is hilarious...it just cannot give me a proper Falkor. But look at the locks on this lovely beast:

[steady grove]

hows that differen?

[crisp owl]

Think I can work with this and get it right. (can't actually be in the talons)

[eternal fog]

Superiour Eagle

[crisp owl]

absolutely majestic

[native knot]

We haven't seen the Oracle in maaaany years.

[crisp owl]

reminds me of Final Fantasy 13

[native knot]

heh

[eternal fog]

I'm quite enjoying the IPAdapter Attention Masking

[upbeat summit]

nice! cool stuff

[eternal fog]

That was a prompt of a Spaceship interior and then IPAdapter of a monster and attention masking to put it in that position

You can do some cool stuff

Like this as the IPAdapter Image

The mask

Spaceship Exterior as the prompt

And if I put the mask up on the top, you get the clouds and sort of castle battlements on the ship

Or you can go a bit crazy with the masks and let it sort of work out what it wants to do

You can see that it's taken the castle bits for the middle bit of the mask, the clouds on the left and right. And then because the bottom on the image is water, it's done that a few times so the rest of the generation has decided there is water at the bottom

Nooooooooo kitty

[eternal fog]

Same Seed, Same IPAdapter Image and Settings, Different Prompts.

the 2nd lot I refined the mask which let me put the weight to 1 without it dragging in some random background stuff

[pure crystal]

https://x.com/MartinNebelong/status/1724919110830633328?s=20

[steady grove]

diffusions like a box of chocalotes

[native knot]

lol

[steady grove]

perhaps the greatest box

[native knot]

Our people have the largest box, some say.

[steady grove]

step 1 : you cut a hole in the box

[steady grove]

it doesn't seem like theres a whole in the box when i image prompt with this

[steady grove]

lol it's poorly applied ❤️

[native knot]

Yes

I lul'd.

[steady grove]

image prompts are pretty fun

its just blending people when you prompt multipel though

same ol

it doesn't seem to understand denver the last dinosaur, but soon as you describe him, there he is. therrres the little buddy!

[bright valley]

I'm glad IPAdapter came out, because Midjourney has had the feature for months and I felt like I was the only person who used it

And it's so fun to do

[vital ermine]

It can be fun but a bit too unstable for real usage

[bright valley]

I dunno what a real usage is but yeah it's fun to put donald trump and elon musk and whoever into stupid situatiuons

[vital ermine]

well, considering how fucked up this channel became for months on end withendless Segal shit that didn't even look like him, I guess.

[bright valley]

mitch mcconnel's going SSJ2

[vital ermine]

fuck mitch

mitch the bitch

[native knot]

lol

[vital ermine]

well, he is close to death so if it happens I hope before the next election. I mean he had a brain aneurysm in front of the world then says he is fine. My ass. gonna be Feinstein, or Robert Bryd and die in office afdter 1k years of being in it.

[bright valley]

i dont even remember the other image i used for this one

Didn't Feinstein croak already?

[vital ermine]

just did instead of leaving. She was so bad before she didn't even know where she was voting as they told her to vote.

[bright valley]

here they are combined with a sphynx cat

[vital ermine]

honestly, the entire global political system needs an enema and start over.

[bright valley]

dora the explorer

[bright valley]

pikachu mario and goku all kinda ways

Okay back to regular art

[ionic gulch]

ipadapter with masks:

[minor glacier]

Distilled LCM SDXL base model plus SDXL refiner.

[zealous horizon]

Because you make the real difference

[half cedar]

ick

[crisp owl]

Good thing a locally installed sdxl can exist 🙂

[half cedar]

Praise be

[late marsh]

pewpew

[near wraith]

hey guys quick question

for SDXL training

what model do I use? what's the best one? because there are different ones out there

could you send me hugging face link please?

[stone fossil]

I always train vs base XL.

[lilac raven]

Is there a model/LoRA to generate images in the style of a 3D render? The one I'm having the best luck, so far, is animeArtDiffusionXL_alpha3, even SDXL base is being inconsistent

[crisp owl]

(text added post generation)
Saw a meme with these cut and pasted onto actual fruit pics, decided it needed proper treatment

[native knot]

This Senate hearing will commence. Our guest, Ms. Thicc is here to testify.

[steady grove]

mad balls vibes

[hoary saddle]

is there a node that will recolor an image using the color palette from the input image?

the rendered images seem to change hues quite a bit from the original, dunno if that's possible to correct

[steady grove]

you could apply color matchiing to stuff in post

photoshop/gimp stuff

[hoary saddle]

hmm, i use comfy as an api source for a webapp where users add their input images, would need to automate into the workflow, still looking through all the custom nodes, gotta be something that reapplies the original palette to the new image

[steady grove]

https://github.com/ManglerFTW/ComfyI2I this has color transfer

is that the right understanding @vast galleon lol, i just connected the dots an thought "waaait a second he posts here"

[hoary saddle]

YES! exactly, ty Icey

[steady grove]

the true heros are the developers. glad to help you find it

[half ivy]

bruh 🔥

[vital ermine]

Bad kitteh is bad.

[half ivy]

cops are in purrrrrr'suit

[vital ermine]

I think I like this one better than the one before.

[vital ermine]

yep. 🙂

The winner is Super Hulk

[west breach]

https://civitai.com/models/197719?modelVersionId=227464 A new model I have put together where the output is focused on painting styles and not photorealism. It was trained on real paintings, no photos, no digital artwork. You don't need to prompt the painting style. Please check it out 🙂

[spark bear]

the piggybank and money roses look like Tom Martin paintings

[tepid maple]

guys why 1111 refusing to put the model on vram ?

this started to happen after I updated to the latest 1111

I just have these

is this new bug ?

[peak dove]

Dynavisionxl NightVisionxl Unstablediffusersxl Counterfeitxl

[uneven gazelle]

Hillarious ahahaha

[onyx compass]

Redraw according to this diagram

[native knot]

Darth Vader prom picture? lol

[crisp owl]

Bingo hahaha

[native knot]

That's great.

[crisp owl]

senior prom photoshoot of a young woman and Darth Vader in space

[native knot]

The space background gave off cheesy 90's photo background vibes.

[crisp owl]

Memories 😆

[native knot]

Would be even better if the photo had the little Jostens logo on the bottom in gold.

[crisp owl]

too hard to prompt for, just tried a couple and it keeps wanting to put a huge thing behind them lol
(well, probably could get it eventually but eh haha)

[native knot]

I wouldn't prompt it...just a quick photoshop would be better in that case.

[crisp owl]

Yeah I was just curious if it was hidden in the latent space

[crisp owl]

hold up. I had watermark, text as my negative prompt, forgot. lemme push a couple again lol

ah yes, that's better. upscale finishing

[cyan crown]

I'm posting these two images from DALLE 3 because I'm really amazed by the latest capabilities introduced by ChatGPT. Essentially, I created a GPT (a feature from November 2023) by feeding it the text of all the stories by Edgar Allan Poe. Then, I explained to it to use that text to derive the style in which to respond. In practice, it's as if I'm speaking with Poe (or rather, with the writer Poe). Next, I asked it to write a short story about a man who hears noises under the bed and to illustrate it. Well, it wrote a brief story in the style of Poe and these pictures are exactly in the style I would have expected. Incredible

[native knot]

@crisp owl ^

[native knot]

heh

[crisp owl]

Hahahahaha 🤣

[native knot]

Thought you might get a kick out of that one.

[lilac wren]

i did something like this long time ago

[vagrant valve]

oh... nice feet

[crisp owl]

@native knot Finally, FINALLY, got nearly exactly what I was trying for originally

[native knot]

lol...yeah. I just let SDXL do whatever it wanted with the image. All I cared about was getting the title to be correct. Once I had a correct title, I moved to the next movie. 😄

[native knot]

Noice!

[crisp owl]

it's not perfect, but hell of a lot better than I was going to end up using lol

[vagrant valve]

to be honest they are amazing. Using a LoRa?

[native knot]

Only the add-detail-xl one.

BTW, running that prompt gave me some horrid stuff...lol

Oh, this next one is great...

Holy cow

I like how Michael Rosenbaum is suddenly in The Matrix.

Top-right.

This is legit.

[native knot]

1980s Horror Film: Aladdin

[vital ermine]

https://civitai.com/models/203078

[native knot]

1970s Western: The Godfather

[native knot]

Noir detective film:

[vagrant valve]

These are great! You said you were only using sdxl and the add detail xl?

[native knot]

Yes

Bruce Lee as
The Godfather

Bruce Lee as
Robin Hood

Bruce Lee as
The Terminator

[violet sapphire]

Very good one.

[native knot]

Bruce Lee in
Jurassic Park

Bruce Lee in
American Psycho

[native knot]

lol

Bruce Lee in
Bad Santa

I don't know why I'm stuck on putting Bruce into so many other movies...he just works so well.

[vital ermine]

All those bells going off in your head after he hits you.

[native knot]

Bruce is going to take the Pepsi challenge.

[exotic obsidian]

1

[eternal fog]

New actor

IPAdapter face with Attention Masking is pretty good

[crisp owl]

https://tenor.com/view/futurama-george-takeis-head-oh-my-gif-21002337

[vital ermine]

My new loha

the above without it

[crisp owl]

adds more glossiness?

[vital ermine]

no, I really can't get a handle on what it is doing

depends on the prompt

loha is fun as you can allow it to run free so no idea what style it grabbed

I am prompt dead and lexica is of no use.

with and without

with is way more true to the lexica prompt I grabbed

[crisp owl]

Just realized you're saying loha, not lora. I don't even know what loha is lol

[vital ermine]

a much higher form of a lora

[crisp owl]

Oh, interesting. Applied same way by a lora node?

[vital ermine]

have a realistic prompt I can try?

might figure this out cause what I am seeing is nothing I trained on

lexica.art image

[crisp owl]

Uh, here's my last one I was just doing

Norse goddess young iduna in her beautiful magical garden with blossoming flowers, large trees, singing birds, small creeks flowing, norse gods and goddesses gather

[ionic dragon]

RAW photo, face portrait photo of beautiful 26 y.o woman, cute face, wearing black dress, happy face, hard shadows, cinematic shot, dramatic lighting

[crisp owl]

variations of this is what I get

[vital ermine]

LOL

wth?

qwerty one in a sec

norse one now

with and without

without and with this time

Well, I have no idea except I think it is blurrying the background

it does seem to add something else too

[crisp owl]

Seperating the subject from background more, so yeah, perhaps just more of that bokeh effect

background depth in essence, maybe

[vital ermine]

I rather like how it fixed the lighting too

there were no animals in my images meaning humans too.

this was mainly for effects and it seems to have latched onto something for sure

I really like it

whatever the hell it is doing, lol

I think you are right it separates the subject from the background

yeah

8k wallpaper of a beautiful anime adventurer girl wearing gold jewelry in the streets of a city in the Western Sahara, by artgerm, intricate detail, trending on artstation, 8k, fluid motion, stunning shading

It seems to make anime into realism

professional portrait of a alien race tatooed typ, the head with 3d bony growths under the skin on the head normal face full body side view the backdrop sea and clouds the sea is ocean_blue the male is natural colored , abstract beauty, approaching perfection, delicate face, dynamic, moonlight, highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by Carne Griffiths and Wadim Kashin

I hope when I retrain it for size it retains this

[crisp owl]

dayum

dat troll hair

[vital ermine]

Yep

I have no idea wth this loha is doing but it is fun

[nimble heart]

@vital ermine
https://github.com/ROCmSoftwarePlatform/pytorch/commit/56449c79a76e23764dd6eb7810831919cb811a23
ROCM adding itself to the unit tests for the Flash Attention backend in PyTorch's SDPA

i have an FA2 wheel built but it doesnt support the specific shuffled forward op needed for inference in oobabooga so I'm not sure if it works

[vital ermine]

oobabooga sucked 4 days of 12h days out of my life. I hate it and I wish there was something else like it but there really isn't.

[nimble heart]

4 days and 12h how

works pretty well for me tbh

i just use the amd requirements file instead of the cuda one

[vital ermine]

I am in windows and the day he stuck in that openai api I upgraded. Not even a downgrade helped so I just said fuck it. Most of the models I like (GPTQ and AWQ) others don't run so I will check back on the state of things every six months. I refuse to Linux or WSL just for this.

Everyone one I talked to said DON'T UPGRADE. Ahem, thanks I already had.

[nimble heart]

Ah, can't say there. I know llama.cpp has been having issues upstream and ooba's been having a helluva time changing versions trying to make it work

[vital ermine]

yeah, when it worked I liked it then it just didn't and became hell

[nimble heart]

but tbh I havent used GPTQ in months since Exllama2 does all the same shit at 10x the speed

just TheBloke doesnt make exl2 quants for some reason so you either have to make them yourself or spend a minute to find some decent ones on HF

[vital ermine]

Yeah, and I only use his

the fuck tons of models on HF I threw my hands up and I thought the SD community was dysfuctional it has nothing on those people

[nimble heart]

with a 4090 you can probably make your own in decent time

[vital ermine]

All I wanted was a nice captioner as I have no further need of one

[nimble heart]

I made like two dozen on my XTX

[vital ermine]

btw, QWEN took one scene from one movie and described it to the tee. My jaw dropped but worthless for captioning as you can't limit it at all.

[nimble heart]

is it consistent though

[vital ermine]

btw, on my 4090 they all ran like shit. Fastest I got was 10 to 18 tokens per second with most being around 7-8

If qween allowed me to temperature it or tell it not to exceed XX token, or words without it breaking completely I would use it

*Qwen

[nimble heart]

sometimes hypercreative language models will blow your mind and destroy your pernus but a few seeds later it's just hallucinating first grade science questions and wikipedia pages

[nimble heart]

the quantizer or inference?

[vital ermine]

sec

[nimble heart]

sec?

[vital ermine]

{'image': 'D:/CleverCaption/test/1/The.Book.of.Eli.2009.1080р.BluRayRemux.mkv_snapshot00.03.58[2023.10.19_20.29.23].png', 'text': 'Craft a succinct, vibrant depiction of the image, weaving complex syntax into a tapestry of detailed yet terse statements. Include one elaborate sentence for the focal subject, another for the backdrop, and a final one for the visual composition. Think carefully step by step.'}
The Book of Eli, a 2009 BluRay Remux, features a powerful scene where the titular character, played by Denzel Washington, battles a group of ruthless gangsters in a dusty, post-apocalyptic wasteland. The sun beats down on the dusty streets, casting long shadows across the barren landscape. The air is thick with the smell of gunpowder and the sound of bullets zipping through the air. In the midst of the chaos, Eli stands tall, his weathered face etched with determination and resilience. With a single swift motion, he dispatches a would-be attacker, his trusty shotgun at the ready. The backdrop of the wasteland serves as a stark reminder of the harsh realities of survival in a world gone mad. The visual composition of the scene is masterful, with each frame carefully crafted to convey the sense of danger and tension. The final shot, a long shot of Eli walking away from the scene, leaves the viewer with a sense of both awe and respect for the hero's strength and courage.

[nimble heart]

exllama2 has the fastest possible inference right now I believe

[vital ermine]

now tell it to no more than X token or words. 100% hallucinates

[nimble heart]

for a 13B I get like 42 or 43 T/s on fresh context

[vital ermine]

I never ever get that period end

stop

[nimble heart]

sometimes super aggressive exl2 quants need adjusted settings if that's what you're referring to

like no BOS

[vital ermine]

Qwen has no settings

[nimble heart]

Ah thought you were still on exl2

[vital ermine]

as to the rest of the models I am done with it. too slow to fuck with and if I have to have Linux to get them I can live without any of them

some were 0.8 to 2t a sec

[nimble heart]

all Llama-compatible models should work fine on windows with the built-in ooba backends

[vital ermine]

thanks, but no thanks.

[nimble heart]

there's also kobold.cpp which is a portable binary without any venv install bullshit

but I've never used it cause im on linux

it has llama.cpp and exl2 backends i believe

[vital ermine]

kobold and tavern is made more for story telling and rubbing one out to your waifu than anything else

I still do not getwhy bloke doesn't do exl

[nimble heart]

well not sure what to say lol. lot of backends have their own standalone inference script too you can use without spinning a server

[nimble heart]

no idea. there's like 2 or 3 people on HF that've uploaded a few dozen EXL2 quants each

[fierce dirge]

yooo guys can anyone help me with this? https://discordapp.com/channels/1002292111942635562/1072238304042438758/1175347960121397269

[vital ermine]

well, openflamingo went insane on me and blip/2 isn';t all that good

[vale eagle]

how about kosmos2?

[peak dove]

Anybody know how to marry ChatGPT4 as a prompt-helper directly into ComfyUI? It says my OpenAI API Key is invalid ... how do I re-validate at all?

https://www.reddit.com/r/StableDiffusion/comments/160vozo/use_chatgpt4_as_your_prompt_engineer_inside_of/?sort=new

This workflow ...

[vital ermine]

Yeah, I don't do anything but the free 3.5

[lapis gale]

playing with sdxl styles (again)

"lego set"

peace by pieces 🙂

[clever verge]

Anyone tested this? https://github.com/comfyanonymous/ComfyUI/commit/bd07ad1861949007139de7dd5c6bcdb77426919c

[lapis gale]

i'm using SDXL which does up to 1280x1280 without duplicate issues natively, faster than SD 1.5 with hires fix before... not my usecase 🙂

blues brothers!

[zinc cargo]

so that double head 😛

[zinc cargo]

wow

[zinc cargo]

lora?

[stone fossil]

Model new one.

Paradox 2.0, 6 days left.

It is trained on a lot of stuff, I just happen to love double exposure lol and it does that well also.

[zinc cargo]

it's not on civit yet?

[stone fossil]

No.

Paradox 1 is.

[peak dove]

I cannot update ComfyUI??!?

[shy kelp]

do anyone know how to put ipadapter link in this form for comfyui

[uncut steeple]

#🤝｜tech-support

[dapper dragon]

https://tenor.com/view/letter-kenny-good-enough-drink-gif-14433251

[lusty wolf]

Better?

[dapper dragon]

haha

it was joke with @stone fossil

[dapper dragon]

It rocks bro 🙂

https://tenor.com/view/fabulous-cat-cute-gif-15272529

[stone fossil]

Aye the output rocks, more images. 🙂

[lusty wolf]

Oh, you didn't specify... 😜

[lapis gale]

speaking about star wars and DJs... how about her?

had to change the hair for the headphones 🙂

oh well

oh YEAH!

[eternal fog]

Huh, you can use a couple IPAdapters along with a couple attention masks to manipulate outfits and faces. It works pretty well.

Had it setup with this image and mask on an IPAdapter Face

Then this image and mask with IPAdapter Plus

Result

Prompt was just generic "4k Desktop Background" which is why there's the hills in the background

However, ideally don't put the face in the "outfit" image, it occasionally decides to use it lmao

It's done a sort of male neck and jaw, female top half merge

[vital bane]

which is the lora, which is the real person ?

[hollow tide]

why does my VAE look like this?

is it not working correctly?

[eternal fog]

What VAE are you using, there's one that does that on purpose

[nocturne dove]

Just got an M3 Max macbook yesterday so I spent today getting some stuff set up. First sdxl images with some ink drawing Loras. The GPU is just ok in terms of speed but 128GB ram means I can do stupid stuff like run a batch size of 24 with sdxl (or more although increasing batch size is pointless in general) or I can run sdxl at the same time as a 70GB Q8 quantized 70B llama model.

[hoary saddle]

anyone get anything close to results the creator got on this lora?

https://civitai.com/models/183354?modelVersionId=205793

just using base model like most of the samples, same sampler etc

[crisp owl]

@noble shoal made it

[hoary saddle]

cant figure out why i dont get the awesomeness in the samples lol

😦

not even close

[crisp owl]

here's his intro images, then just a couple posts down when he released it.
#✨｜sdxl message

[hoary saddle]

thanks mate, maybe i can find some clues

[vital ermine]

money shot

[nocturne dove]

Yeah they're kinda good for some reason and that's not the goal 🙂

This is better (worse)

[hoary saddle]

lol, closer than mine pretty sure

if i use the civitai create, it comes out good (bad)

just not in my workflow hah

[nocturne dove]

Scarlett Witch. I added "derpy" to tge prompt. Don't know if it helped. lol

You're using the keywords right? "MSPaint drawing"

[hoary saddle]

interesting

the only difference in my workflow and yours is the clip set last layer node

that was the problem

this little bastard hah

[nocturne dove]

I was using A1111 because I'm on a mac at the moment and it seemed easier than getting comfyui up and running

So I have no idea actually what it's doing internally like I would on my PC with comfyui

[hoary saddle]

you're the man! ty

been struggling with this for over an hour like a doofus hah

[nocturne dove]

Lol. This lora is fun. Apparently this is "Thor, marvel movie"

[hoary saddle]

perfection!

[vital ermine]

https://civitai.com/models/205251

Civit is still a pita to upload to

go to page 2 of the upload process and 404 error.

[hoary saddle]

this model is F'n amazing @noble shoal

added a MSpaint font to my chat overlay node i made, match made in heaven

[noble shoal]

I see, i see. You guys having fun with MS Paint 😬 . Let me add a few notes to help you getting the ultimate garbage quality.
-Use the Base model or weight higher in realism models (i saw that you use jaggernautXL, right? @nocturne dove )
-Keep the Keyword in your sentence in front as if you would describe your image like: MSPaint drawing of a interior of plant store, city street shown outside window, sunlight coming in window, shelves full of many plants and colorful flowers
-Using the Keyword MSPaint portrait should give slightly better (so worse) results, even with non portrait scenes.

MSPaint portrait of ironman sitting on a toilet in a bathroom / Base model + No fancy prompt = Perfection 😅

[shadow oak]

Hey lads, trying to figure out where to hook up the seamless texture, where would I go about attaching it?

[noble shoal]

Ah, good old seamless. Take a look at this repo here to get true seamless textures: https://github.com/FlyingFireCo/tiled_ksampler

Here is a super barebone Workflow in this image

[shadow oak]

Thanks, your experience that's better than the WAS node?

[noble shoal]

In my experience, yes. It creates 100% Seamless pattern

[shadow oak]

You're a ledgend, thank you.

Do you drop that right before you save or before you upscale?

[noble shoal]

Just check out the Workflow. I have no idea how it handles upscaling, since it used a different VAE Decode node. You might also need to get rid of your refiner, but that unnecessary in most cases anyway. I have a custom tile stitching node somewhere where you can preview the "seamlessness". Let me check.

[shadow oak]

If you mean this workflow, it doesn't seem to have it embedded with the image.

[noble shoal]

Did you clicked the image and then viewed it in the browser? The discord "preview" doesn't has the meta data, i think

[shadow oak]

Yeah, it seems the default: https://media.discordapp.net/attachments/1089974139927920741/1175727133067333642/ComfyUI_15497_.png?ex=656c4838&is=6559d338&hm=496e4615bac939bc4feb464313ec2989971a0bf3163470cfebe72003ecf53eec&=&width=701&height=701

with the url params is stripping the encoded data, while https://media.discordapp.net/attachments/1089974139927920741/1175727133067333642/ComfyUI_15497_.png works, thanks again mate, appreciate it!

[noble shoal]

Drop this into your custom_nodes folder. It has a keylogger and will steal you bank account details, but it also lets you preview your tiles. 😅 jk

Here is the workflow for it

[shadow oak]

from POL import Image, ImageOps would be easy to drop a logger 😄

[shadow oak]

Thanks, removing the url params in discord allows the workflow embedding.

Thanks all!

[noble shoal]

You might change this line in the code: CATEGORY = "D-Nodes/Switches" to CATEGORY = "image/postprocessing".

[shadow oak]

I'm just using the node in the repo:

[noble shoal]

Never used this one. I think it allows you to tile either in x, y or both directions, right?

[shadow oak]

https://github.com/FlyingFireCo/tiled_ksampler

Testing it now.

[noble shoal]

My node just stiches the tiled images together in a grid to have something nice to look at, after the whole sampling process.

[wicked frigate]

the code there is a little odd 0.o

(A) it has lasting side effects on the source model and (B) why is it implementing the ksampler node instead of just being a middle node?

The SwarmTiling node is more carefully built: https://github.com/Stability-AI/StableSwarmUI/blob/master/src/BuiltinExtensions/ComfyUIBackend/ExtraNodes/SwarmTiling.py feel free to yoink that (or just run your comfy through swarm so you get swarm stuff automagically)

you just hook SwarmModelTiling after the checkpoint load and before your ksampler

and use the tiled vae decode instead of regular vae decode

[noble shoal]

👀 I wasn't aware of this hidden node. Let me check it out.
(A) it has lasting side effects on the source model and (B) why is it implementing the ksampler node instead of just being a middle node? I have no idea, you might ask the genius who wrote the code.

[shadow oak]

Thanks, Can this run standalone or will I need the swarm?

https://github.com/Stability-AI/StableSwarmUI/tree/master/src/BuiltinExtensions/ComfyUIBackend ?

[wicked frigate]

That can indeed run standalone, though I don't see why you'd ever not just use swarm

it's a node like any other and can be just saved into your custom_nodes and it'll work

[shadow oak]

There any issues with running swarm and headless API?

[wicked frigate]

swarm runs fine in a headless env

[shadow oak]

Thanks so much mate, I just wanted to confirm this is the best I'll be able to get, am I expecting to much to think it would be perfect?

It's close, but a little off:

[wicked frigate]

that looks wrong

[shadow oak]

kk, I'll keep playing with it. She's a mess, I'm dropping the swarm tile before my second ksampler

[wicked frigate]

better to have on both if you can - and you need the VAE Decode to clean up the edges fully

[shadow oak]

My ksampler is going into that vae decode on the right, is that what you mean or is there a different node?

[wicked frigate]

TileableVAEDecode

that sample tiled btw

properly seamless

i usually check with https://www.pycheung.com/checker/

though the node aimingfall gave above seems like a good option too for tile preview

[noble shoal]

🥹 Thank you

And thanks for the SwarmModelTiling node. Implementing it in a Custom model loader at the moment. My next abomination.

[shadow oak]

That was it, thank you!

Absolute final question, any way to keep upscales? Is my only option lower the noise?

[wicked frigate]

Er, not sure what you mean?

[noble shoal]

Are you using the "patched" model and the "TileableVAEDecode" also for upscaling? If not, you may do that

[shadow oak]

Tossing the second image in a upscale, you letting me know there's a vae made the first image tile perfect, then I take that image and upscale and you can see the result, it's close but not perfect.

[shadow oak]

testing.

[wicked frigate]

ooh, you have one of those mega-all-in-one-nodes that hides away the insides

so the tiled model should pass in fine but the VAE won't

[shadow oak]

😄

[shadow oak]

So no luck, I'll just have to get close with my initial image?

[wicked frigate]

nah hol up 1 sec

Tweaked the code https://github.com/Stability-AI/StableSwarmUI/blob/master/src/BuiltinExtensions/ComfyUIBackend/ExtraNodes/SwarmTiling.py to be what it probably should've been from the start

you can now just

and then use any regular VAE Decode or presumably even the fancy upscale all-in-one

[shadow oak]

This is dope mate, thank you so much.

Is that your repo?

I'll return the favor by doing a TODO.

[wicked frigate]

(it's possible though that the all-in-one messes with seam handling itself, considering it does its own tiling for the upscale, so might conflict)

I'm the lead dev on Swarm yeah

[noble shoal]

Interesting, very interesting. Now i only have to combine both nodes into one and make a "on/off" int or string.

[wicked frigate]

don't need that - select the node and hit CTRL+B

that Bypasses the node ie disables it but sends the input forward

you can also CTRL+click and drag to select multiple nodes at once and then hit CTRL+B to toggle em all at once

-alternately- if using swarm anyway if not doing a fancy custom workflow you can just enable Advanced parameters on the generate tab and then just checkmark the Seamless option

[noble shoal]

Thank you for your effort trying to safe my time, but its sunday, i have nothing to do and i like the challenge 😜

[wicked frigate]

if you want to practice coding for a challenge, do whatever you want I guess - just, don't publish nodes with on/off switches made for the sake of challenge

[noble shoal]

Hell no. No on/off. Just a passthrough if set to "No"

[shadow oak]

ABSOLUTLY, last question here, This is my upscale running off the tiler on my base checkpoint do I need to make multiple tilers for each upscale model too?

[noble shoal]

I think nobody is judging you for your questions. Something still seems to be off with your workflow. You have multiple models in your workflow?

[shadow oak]

swarm is just connected to the base:

[noble shoal]

I think then the answer is yes, because the "upscale" model is also altering your image

[shadow oak]

testing

[shadow oak]

No, didn't work for the upscale (The base image is perfect), any ideas?

Off the checkpoitn refiner

[wicked frigate]

that's an on/off. You're baking into a node a manual copy of something that's built into comfy itself already

there's already a dedicated keybind (CTRL+B) to do exactly that

[wicked frigate]

you'll have to use an upscaling setup that is built to work with tiling then

-or- if you have the VRAM for it just do a direct upscale instead of the tiling thing

[shadow oak]

Doing a upscale off the image with the update you did to the swarm tile is producing that image:

base ,(is perfect tiled)

[wicked frigate]

for clarity btw a "direct upscale" is something like this:

or in the swarm generate tab is done with the Refiner option

[shadow oak]

correct, my output image from my ksample is saved and also put directly into the upscaler:

[noble shoal]

But i can't convert a key bind into an input

[wicked frigate]

No, the example i gave is what to do instead of using that upscaler node

[wicked frigate]

ah, that.. some what makes sense

what are you outputting to put as an input on that tho? 0.o

[shadow oak]

perfect tile,

upscaled off that tile:

[noble shoal]

My main control note will have a "Yes/No" output.

[wicked frigate]

aaand what if you do a direct upscale

[shadow oak]

If by direct scale you mean feed the base image into the upscale that's what I'm doing.

[wicked frigate]

no

[shadow oak]

Going to try using just a regular upscale node.

[wicked frigate]

^ I mean do that

[shadow oak]

I'll try that

[shadow oak]

My upscaler was dropping it's own styles, thanks mate. Appreciate everything

[stone fossil]

Bad kitteh. 😄

[dim steeple]

Hello

What are the best ways to upscale in SDXL img2img?

using SD upscale extension?

[dim steeple]

settings?

nice work

[dim steeple]

how did you make this?

[zinc cargo]

dynavision xl model

[peak dove]

IMHO best models are DyanvisionXL, NightvisionXL, UnstableDiffusersXL and CounterfeitXL ... but that's just me 😄

[dim steeple]

whats best model for cartoony stuff?

[oblique cargo]

/style

/help

[cosmic fern]

@stone fossil eggsplat 2

[stone fossil]

Hamster for measurements.

[steady grove]

when i caught these vibes, i was thinking "so cool" so now a few days later i'm training a lora with these mad balls. we'll see how it comes out i guess. here's a couple training samples of a zombie dog ball. source images mostly suck for resolution so i'm just training at 768

[steady grove]

american history z

[solar bone]

bluepencil XL model is taking 7 minutes to generate a single picture on rtx 3070 ti https://civitai.com/models/119012/bluepencil-xl?modelVersionId=212090

is something wrong?

[native knot]

Yes, obviously. You have some other setting that is causing it, but there are far too many variables to tell you what without knowing a lot more about the settings you have setup.

[solar bone]

i'm pretty sure im just on default a1111 settings

i barely changed anything

and these are my generationi parameters

[native knot]

Those settings look fine. Are you running from an SSD?

[solar bone]

yes

nvme ssd

[steady grove]

i've never used an exponential sampler. i don't want the ai to gain awareness or nothing. scary

[native knot]

Have you tried closing and restarting A1111?

[solar bone]

yes

ive tried restarting too

[native knot]

I use exponential all the time..shouldn't cause a 7m gen time.

[steady grove]

yeah 7m sounds like cpu tbh

[solar bone]

even on another sampler

it still takes 7 mins

my gpu spikes to 100%

vram full

so its not cpu

switching to sd1.5 model makes images take less than 2 seconds each

[steady grove]

might be a vae problem, like it's hanging hard on the final step? do you know bout ollins?

[solar bone]

ollins?

no idk what that is

[steady grove]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix vae that can work without --no-half-vae enabled

[solar bone]

if it matters here are my vae settings

its default pretty sure

[solar bone]

oh whats that

[solar bone]

my progress bar was moving relatively slowly though

[steady grove]

yeah it'll usually fail on the half vae attempt and then do a full vae render by default. that's checked

[steady grove]

ah so it wasn't that final step.

[solar bone]

yeah

im pretty sure its vram

[steady grove]

ollins is still a cool one to use though

[solar bone]

my gpu fans were barely spinning

[steady grove]

still a good tip for other reasons

[native knot]

Feels like there's still some other underlying setting that's wrong in A1111

[solar bone]

but vram was loaded to full

[solar bone]

what does it do lol

im confused

[steady grove]

allows the half precision to be used in the vae step

[native knot]

Have you ever run A1111 with SDXL successfully?

[steady grove]

faster

[solar bone]

by deafult does a1111 use fp16? ik it should be a lot faster than rtx series

[solar bone]

never

any sd1.5 and below models are super fast

under 2 seconds

[steady grove]

for nvidia cards that supprot it it defaults to half precision, fp16

[solar bone]

is there a way to check?

although i dont think

thats the problem

i think its vram

[native knot]

Are you running --med-vram as a startup parameter?

[steady grove]

--no-half would have to be used to make it full precision

oh med-vram is a bucket of goodies

[solar bone]

this is what happens when i generate

and i dont even think my gpu fans are spinning

[solar bone]

i havent

what does it do / how do i do that

[solar bone]

oh nvm it is

[steady grove]

aha. 8gb. it'll be using too much and instead start using system memory. didn't think it was an 8gb card mb.

they seem so old

[solar bone]

if this matters heres msi afterburner while sdxl runs

[steady grove]

you'll want --med-vram potentially.

[solar bone]

ohh

[native knot]

Yeah

[solar bone]

yeah blame nvidia

[solar bone]

how do i do that?

[native knot]

open your bat file

[steady grove]

with notepad ^

[native knot]

add --med-vram to the commandline_args

[solar bone]

this?

[native knot]

yup

[solar bone]

or without the user

okok

[solar bone]

alr just did

[steady grove]

it'll slow down your 1.5 gens a bit too warning

[solar bone]

lemme see

[solar bone]

oihhh okay

[steady grove]

but it'll make sdxl all fit into your 8gb too

[solar bone]

lemme try it

hm

is that a typo or smth

i just searched it up theres no dash

its --medvram

lemme try that

[native knot]

Ah...I'm not looking at it, so just doing it from the top of my head

[steady grove]

see, if nvidia's old drivers where they didn't fill the system memory when over allocated were in play, it would just throw a cuda out of memory error. That's actually a control panel option now

i totally remember there being a dash

[solar bone]

uhhhh im like

stuck at 97%

with huge cpu spikes

making my ui lag

and 26gb of ram

well image is done now

[steady grove]

yeahhhh budy look it go

[solar bone]

but for 40 seconds my cpu just spiked

[steady grove]

flex that 13900

[native knot]

A lot better than 7 minutes.

[solar bone]

yeah fr

[steady grove]

that'll speed up the medvram mode a lot!!! part of why it slows down is because it's doing a lot of memory bank swaps

[solar bone]

it was stuck at 97%
is it the vae thing?

[steady grove]

i got a 12700k measely i7 so i feel it when its enabled

ugh sucks being obsolete

[solar bone]

oh

like swap between cpu ram and vram?

is there a way around that

its freezing my entire computer for abnout 40 seconds toward the end at 97%

[steady grove]

oh yeah that's the vae thing. it fails that first time then swaps to fp32 mode which is a bit slower

[solar bone]

ohh

so should i install the fp16 thing

[steady grove]

it'll not fail that next gen so it'll be faster, but not as fast as half

[solar bone]

will it have downsides

[solar bone]

lemme try

[steady grove]

not that i've seen. works good

don't know why it's not fixed by default yet tbh. ollins is still my golden boy

[solar bone]

hm yeah it hung at 97% again

[steady grove]

model makers aren't even baking it into their models yet. i dont get it. oh well. yeah it's good cause it doesn't fail to full prcision 👍 !

[solar bone]

lemme install that

what even is a vae? im still confused

and what is a sampler

[steady grove]

you can configure auto too, so that it has quick select settings. thats whta the model select is. you can put theh VAE selection up there too. i like doing it that way instead of going deep into settings menu

[solar bone]

ohh

and which one do i download

[steady grove]

VAE is a variable auto encoder, i think?

sdxl_vae

[solar bone]

yeah but what does it do

[steady grove]

the diffusion process creates a "latent" image. out of some sort of lower dimensional abstract math space. i geuss. the VAE turns that into pixels

[solar bone]

what an improvement

from 7 minutes

[steady grove]

more like it

[solar bone]

it still took ~15 seconds at 97% though

[solar bone]

hm i see