✨｜sdxl

[high skiff]

it warmed up more

10.3 tokens/s while using pooled RAM

this is insane

[boreal bough]

cause that actually needs clip training to avoid damaging the model - except I've not nearly tested clip enough to pull that off yet

[high skiff]

I could load massive models this way

I wonder if I could do 30b on a 3080 like this, cause I have 42GB total GPU RAM

[boreal bough]

flat color go brrrrrrr ❤️

[high skiff]

when I get my 3090, I till have 56GB RAM to play with

[nimble heart]

normal weight

[uncut steeple]

No 4090?

[high skiff]

no, 3090 can do everything the 4090 can do, with less power, and WAY less money lol

I don't care about speed, just functionality

[uncut steeple]

Fair

[high skiff]

much rather a $700 top of the line 3090 over a $1900 top of the line 4090

[alpine pine]

technically, a restricted 4090 does what a 3090 does with less power

[high skiff]

[uncut steeple]

Got my 4090 at home for gaming, the one in the office is for SD

[boreal bough]

this man diffusions

[uncut steeple]

Only had to buy one of those myself so its fine

[boreal bough]

sketch, owl sitting on a branch

[nimble heart]

lol sometimes it actually understands it

[boreal bough]

I mean I have a lora with my usual 80% hitrate 😏

[uncut steeple]

My workflow takes about 100s for a 2048² image on a 4090, not very optimal but I like the outputs

[nimble heart]

thats a long ass time how many samples is that?

[uncut steeple]

You can check it yourself, its embedded in all my images

I just switched it to use wildcards right now

[west breach]

lol at some of the prompts from sandbits prompt compilation file: A male fairy with dragonfly wings runs along a beach holding up a fishing net. He is chasing a flock of flying shrimps creatures. The fairy wears a wide-brimmed hat, thai fisherman pants and has a beard. Graphic novel illustration by Alphonse Mucha

[uncut steeple]

Im running through that one too

My fav from his file so far

[boreal bough]

sketch, drawing of boto waiting on his computer to finish the high quality drawing by the man stuck in the unet
(lora is gender friendly! #biasFree)

[west breach]

the mushroom is downright adorable

[uncut steeple]

[nimble heart]

2048^2 is 70s @ 26/4 + 15/5 samples for me. For 100 on a 4090 that must be like 50+ samples low and high

[vale eagle]

Do you use leco to extract the concept in the model?

[true wing]

Well I managed to crash comfyui somehow. That's a first

[boreal bough]

leco?

[vale eagle]

This one, https://github.com/p1atdev/LECO

[nimble heart]

1 or 2?

[uncut steeple]

I like 2 more

[alpine pine]

2

[vale eagle]

2

[alpine pine]

eyes feel more not dead

[charred beacon]

2

[chrome flicker]

Hi, anyone know how to use an input reference image in sdxl with the same way like midjourney(midjourney can get style, object, color from reference and generate creativate related image) ? If this can be true, sdxl could replace midjourney!

[nimble heart]

damn ok

2 took over double the runtime :(

[upbeat summit]

SDXL loves that concept 😉

[vocal rapids]

hi-res fix work so good using @high skiff workflow

[uncut steeple]

I remember you posting that one

[upbeat summit]

yeah - sorry for reposting. emerging patterns are always interesting

[uncut steeple]

No worries I do the same

[upbeat summit]

really like the artwork you've been posting. beautiful stuff

[west breach]

that's about how much sauce they give you at a pasta restaurant

[true wing]

It's called an Italian restraunt, show some respect

[alpine pine]

there is a large difference between pasta restraunt and italian restraunt

[thin nova]

I hand masked 1800 images in an hour while listening to other stuff. Here's hoping the masked lora training code actually works.

[west breach]

https://tenor.com/view/the-sopranos-tony-soprano-nothing-chin-flick-gif-22794125

[boreal bough]

no sdxl support yet. but cool method of doing it!
I did it by essentially bombarding sdxl with 41mb of extremely biased input.

[vale eagle]

OhIsee

[ionic dragon]

got cha

[nimble heart]

o no i forgot about that

[true wing]

So I've been messing with anime styles since the complaints about it not working and I think they work great... with the right prompt

[ionic dragon]

@nimble heart normal(left), boobie(right)

boobie huge boobies

[uncut steeple]

[vale eagle]

I think we really need to learn how far could the base model reach.

[urban breach]

who's really sitting on the throne 👀

[boreal bough]

same. if you use vit-h on any anime image, then put that exact prompt back into sdxl, you get amazing anime images. but I get that that isn't really userfriendly

[true wing]

I'd say further than most of the finetuned models already. Maybe not better than the best 1.5 finetunes, but pretty up there. Maybe TOO far.

[true wing]

What's vit-h?

[uncut steeple]

I uhm... 🍄

[west breach]

A man with one butterfly wing for an eyelid, so accurate!

[uncut steeple]

Yes

But it was portrait of a mushroom warrior sadly

[lilac wren]

he looks very... strong 😳

[nimble heart]

maybe this one won't be as easy. A or B

[uncut steeple]

light does look better in B id say

[alpine pine]

B looks better in general, but both have odd eyes

[nimble heart]

yea im working on pure latent upscaling and I'm tryna figure out if an iterative approach is worthwhile or not

[nimble heart]

due to the latent nature it's gonna be a lil jank

[boreal bough]

ummmm that is not a bird eating a worm!

[uncut steeple]

I gave up on pure latent upscale

[nimble heart]

In both tests one image goes straight 1024 -> 2048 while the other increases over multiple steps by a factor of (2048/1024) ** (1/3)

[vale eagle]

I feel like the pixel upscale has issue on background. It is doing pretty well on the subject and ruin the background a bit.

[nimble heart]

latent does OK for me @ 1536 but anything higher is going to need inpainting I fear

[lilac wren]

Cat Vador

[nimble heart]

stupid idea lemme test it real quick

hope to god it doesn't fucking work cause it'd be stupid

[uncut steeple]

[lilac wren]

https://tenor.com/view/rat-mouse-mice-exercise-workout-gif-18172436

[thorny frost]

I have like a zillion images of anatomy for artists (not they are that you think :p) can I use them to train SDxl with a 3090ti?

[true wing]

https://tenor.com/view/pumping-iron-arnold-schwarzenegger-workout-fitness-gym-gif-15067436

[alpine pine]

yes, idk about legally

[nimble heart]

o no he got the armpit pics

[alpine pine]

but functionally, you can

[lilac wren]

does (word strength:1.4) still work? with SDXL?

[ionic dragon]

hey wait, i generated something similar few hours ago

[nimble heart]

there's a chance it actually helps a little ☠️

[uncut steeple]

Yours do look nicer, mine is a little on the blurry side

[boreal bough]

can you? yes
finetune? ehhh... technically yes with gradient accumulation. but it's gonna be race for if the heat death of the universe occurs first, or the finetune finishes
LoRA? yes - but bigger datasets make training harder, as you have to get your captions real good. which comes from lots of practice on small datasets

[uncut steeple]

Face in A but details of B

[ionic dragon]

i used sytan's workflow, workflow in the image too

[uncut steeple]

I had a look at his workflow earlier, interesting approach to upscaling, but not too sure if its really that good

[west breach]

It's hard to tell how well an upscale process is working on busy images. Like I tried for ages to get a latent upscaler working, and got it to a decent looking image, but as soon as I changed the style I could see all the faults the other image was hiding

[thorny frost]

So it is not like throwing the images into the model (haven't done it before obviously lol) but a bit more complicated.

[boreal bough]

essentially, anatomy is currently the hardest thing to train by far, with the most pitfalls. I'd highly recommend starting with something easier, like the style or feeling of an artist

much easier to train, and you need a dataset of like 50~150 images

[true wing]

Yeah I've been thinking the next revolution in AI image generation is probably going to involve a better understanding of 3D space and how objects fit in it. That seems to be a big limitation in how the AI currently function

[true wing]

Dr Stranger Man?

[thorny frost]

Would be nice to make the basic scene in 3d and then render it using an AI engine instead of traditional render engines

[vocal rapids]

Fantasy Artwork style is awesome

[true wing]

I think some of these high quality render assets like Unreal's textures and whatever their human generator are called mixed with real photos would be a good start. Start with comparing images of the 3D model to the 3D model, then mix in pictures and try to get a 3D model out of that. I know there's been some work in that space but the results have been mediocre

[uncut steeple]

Balls

[lilac wren]

https://tenor.com/view/call-me-kevin-kick-the-baby-kick-kicking-punting-gif-18237653

[boreal bough]

Dr Stranger Vampire!

[true wing]

For some reason these remind me of all the women attached to some villain in James Bond

[upbeat summit]

definitely. it's like the glamour photoshoot version of a james bond film intro 😉

[lilac wren]

Eva Green ... Bruh ! 🌡️ 💥🍆

[true wing]

You remember Gold Finger? Well get ready for Gold...

[boreal bough]

sdxl has very big opinions about "woman from james bond movie"

lol. this causes more boob focus than when I tried with boobies earlier 🤣

[thorny frost]

I think a different ai generation technology would be needed. One really understanding 3d figures in 3 dimensional space, not working on denoising images only. Such thing exist I think as there are AI generating 3d models already but it is still in very alpha stages. I think diffusion models are only the beginning.

[true wing]

I've actually been wondering about a system of merging specialized AI with one 'overseer' to distribute tasks. It's something I've thought about alot and apparently there was a recent-ish paper called 'Mixture of Experts' or something on the topic. There's also speculation GPT-4 might be exactly that.

[boreal bough]

sketch, landscape from shaun of the dead

[true wing]

That almost looks like a panel from The Walking Dead comic

[uncut steeple]

Yes we need more ultra wide stuff, I need a new wallpaper for my 5120x1440 screen

[vocal rapids]

what a artwork ....

[upbeat summit]

if I do a google image reverse search on it, will I find it?

[thorny frost]

That would be basically getting closer to a general AI, which is unavoidable probably. Next year's will be awesome, if chatgpt 8 doesn't go skynet and decides to exterminate humanity in a nanosecond after getting self awareness.

[true wing]

I keep forgetting not everyone thinks AGI is the goal of AI. In my mind they're one in the same but that's probably because that's where I want things to go.

[uncut steeple]

Probably not

[nimble heart]

has SD ever put out a text2img result that could be reverse searched?

[nimble heart]

besides like the mona lisa i guess

[upbeat summit]

"I have these shoes in my mind - let me quickly generate them and look if somebody sells them"

[boreal bough]

yes. absolutely yes XD overcooking can even happen in the base model

[west breach]

$1337 dollars lol

[thorny frost]

I think AGI is the only logical way for AI to develop, unless people gets scared or something and try to forbid it , which would only make AI development secret.

[nimble heart]

ah. All those examples I saw of supposedly re-creating some deviant artist's piece was clearly img2img so I didn't think it was actually possible

guess if you just fry a single image of your waifu into a lora that'll do it

[true wing]

I agree but we can already see people getting scared. At the same time I don't know how you'd really stop it, I think the best you could do was slow it down.

[true wing]

I think in one of the ongoing lawsuits one of the image generators was able to spit out almost exact copies of some movie posters. I don't remember if it was stable diffusion though. They were trying to use a statistical artifact to 'prove' AI is just copying and remixing existing work.

[nimble heart]

"copy and remixing" make it sound like automatic photoshop lol

[upbeat summit]

oh yeah that has been around for a while. if you read the study they did nothing is really clear about it

[nimble heart]

If a human spent a lot of time in a museum with 5,000 Iron Man 2 posters they'd probably be able to draw it pretty well from memory

[true wing]

It's dumb but they have to prove it's a 'derivative work', which means being substantially based on an existing work. But it's impossible to store the amount of information used to train AI Image generators in the model sizes used so I guess it's going to come down to how dumb the courts are about it.

[thorny frost]

Yep, first lawsuit came from artist. Don't know whenor what will the court rule but will be interesting.

[boreal bough]

if greg rutkowski sues, I wouldn't even be mad if he wins though

[true wing]

I read the one on the 'secret keeper' or whatever it is by google about extracting numbers. I just skimmed the lawsuit because it struck me as kind of stupid and copyright law is subjective.

[nimble heart]

0 steps kinda looks wrong

[true wing]

Only if every artist (or their estate) with a similar style to his that came before him is allowed to sue him in turn.

[peak dove]

Looks like one of my prompts - excellent work 🙂

[true wing]

Also don't you dare ask artists about using 'reference' images, they get real mad

[grizzled root]

genial sdxl and refiner with comfyui , but I can’t use the (Loras sdxl) and (style model with load image). if anyone has any info?

[thorny frost]

Which is dumb itself since any human artist dies exactly the same as AI. Copy other artists work and mix them up to build his own style. Being a real medium artist myself I only see these lawsuits extremely dumb and arbitrary

[boreal bough]

https://civitai.com/models/117635/greg-rutkowski-inspired-style-lora-sdxl
what a convenient name change that page had

[uncut steeple]

Not a huge fan of this, he even links his patreon in the first sentence of the description

[true wing]

Eh it's the same thing. You can't copyright a style, somebody probably just got a scary letter about it.

[boreal bough]

less about the style, more of a 'specifically targeted and impacted' thing

[true wing]

Right but it specifically says 'style' in the earlier version so their bases were pretty well covered.

[upbeat summit]

I don't know when AI will take over or surpass us in everything but it will definitely play a big part in entertainment and visualization

SDXL image

runway gen-2

[true wing]

adding 'inspired' doesn't really change much

[hard fractal]

Bet ya this guy's doing ddim-u

[true wing]

Yeah weren't a bunch of people upset because Secret Invasion used some ML/AI Art? I haven't seen it but I've heard there was some controversy

[upbeat summit]

yes the intro had lots of ai generated imagery. most people who are not familiar with it just thought it was a cool art style, but if you know how AI art works and looks there's no question that they used it

[uncut steeple]

Im doing 2S ancestral for base and upscale and 2m karras for refiner

[civic sigil]

I thought it was pretty sub par tbh

I heard a couple other people disliked it but they might have been biased cause its AI art

[true wing]

I've been using UniPC and my animals have come out great. I don't know what the complaints are about.

[uncut steeple]

Have a non cursed cat

[true wing]

What's cursed? They're just chilling

[nimble heart]

its got chinchilla hair

[west breach]

I think 0.9 had a higher success rate for coherent elephants, from what I tried a few days ago

[hollow hare]

hi guys, it's kinda off the track but there was a trend for this Korean profile picture generated with AI. I was wondering if it's possible to train using SDXL LORA to achieve this kind of result? i also noticed when i generate using SDXL with asian woman prompt, it has the tendencies of korean beauty style (which is from the datasets i presume).

user upload their face and got this template images

[nimble heart]

maybe their dark skin worked better with its stronger offset noise

[uncut steeple]

[boreal bough]

can confirm. cats look great with UniPC. 10/10

[true wing]

seriously though I did have to take a break from generating cat spiders because it was starting to upset me and that doesn't happen often 🤣

[thorny frost]

Can somebody do a furry octopus? That seems to be the ultimate challenge for SDXL. The most I have achieved is a furry cuddle toy in the shape of an octopus but not a real one. Dall-E otoh makes it wonderfully at the firstborn attempt.

[boreal bough]

instruction unclear. send halp.

[true wing]

If you grab some of my images I think you can just drag that into comfy UI and see what I did to get the spider cats if that helps.

I'm not using any custom nodes so the workflow should just work

[boreal bough]

best my lora can do with its few lines XD

[upbeat summit]

this makes me so uncomfortable. I love it

[true wing]

Thanks! Uncomfortable is sort of my specialty.

[boreal bough]

sketch, drawing of a car in the city at midnight under the lights of streetlamps

[peak dove]

Octopus covered in layer of fur

[peak dove]

Looks like the ink preset in Wombo Dream

[thorny frost]

That looks like a mix of elephant and octopus. Still closer than I got

[peak dove]

An octophant

[thorny frost]

Pretty amazing, but not much hair there

[west breach]

Since sandbit's file is so massive, I added word search to my wildcard pattern matching

[thorny frost]

That is Dall-E first attempt.

[peak dove]

I will try "obvious fur" "long fur" "heavy fur" ... 🙂

[keen phoenix]

Does anyone know if the controlnet dev team is working on SDXL ? Or if there is a place where we can follow their progress ?

[upbeat summit]

a furry muppet octopus, from a 1980 tv series

[noble shoal]

Now add "under water"

[thorny frost]

Yes, muppet furry octopus work. But put in under water and hair will disappear. (It is only evolutionary logical anyway lol)

[peak dove]

Octopus covered in heavy fur/long fur

[true wing]

nice

[upbeat summit]

my specialty with SD 2.1 was putting fish above the water

[boreal bough]

there. did it properly now

[thorny frost]

I think that is a winner

[upbeat summit]

a extra furry muppet octopus underwater, from a 1980 tv series, national geographic

[boreal bough]

A head on portrait photograph of a furry octopus covered in thick fur at sunset in a reef
Supporting: hair, pelt, thich hair, fur, thick fur, head on, fujifilm, bokeh, f1.8

[true wing]

I've done some of that too!

[peak dove]

What made SDXL sit up and do FUR properly? 🙂

[urban breach]

Did he finish ravaging the diner? 😎

[thorny frost]

Lol love the curly hair

[true wing]

Nah getting home from work.

[charred beacon]

the meatball rebellion

[urban breach]

An entire series about an angry shark who think he's a person

[true wing]

done.

[boreal bough]

enough messing with sketch, time to do ink lora next

[noble shoal]

This was my attempt of generating a furry octopus from a few days prior. Call it a success or not.

[urban breach]

We need some pics of him doing chores around the house and sleeping

[upbeat summit]

hybrids are always a success

[urban breach]

perhaps getting into a dispute with the neighbor

[true wing]

Mine keep coming out like cartoon characters

[boreal bough]

sketch, cutest friendly pet octopus I have ever seen!

[thorny frost]

First one is perfect, second one is more a mushroom octopus

[ionic gulch]

why doesn't primitive nodes size get fixed with higher priority? makes it impossible to create good looking "frontends" and should be an easy fix, if other nodes already save their size:
https://i.imgur.com/uMlrcGA.png
also some kind of on/off switch to turn on and off sections like lora, upscaler or switch between txt2img and img2img would help a lot. other great features would be an option to hide values you never want to change in a node (set value once, then hide) and input nodes that allow to define a list for a dropdown selection without editing the code externally.

[true wing]

I think it's supposed to be a bowl cut, in hindsight I probably shouldn't have used the word hair 😄

[west breach]

yeah that would be good if they didn't automatically shrink

[upbeat summit]

omg mikey.. this is great!

[true wing]

Here's a decent one.

[trim totem]

are there any sdxl loras yet? I'm looking for something to create better hands. Thanks!

[thorny frost]

This one is 4000×4000 done with SDXL and ultimatescale. Gives an idea of how much richer and variated xl is compared to previous models. Only issue is anatomy imo.

[west breach]

lots

[trim totem]

can you tell me where for example? TY

[upbeat summit]

give her/him a parachute!

[west breach]

go to civitai and add a filter for sdxl 1.0

[thorny frost]

I mean ultimate-upscale

[west breach]

what did you have the denoise value set to and tile size? There's a lot going on!

[thorny frost]

Tile 1024 x 1024. Denoise 0.2 or so. Did a little inpainting too to eliminate some nonsense .

[uncut steeple]

my settings

[noble shoal]

The most fluff i can get atm

[thorny frost]

I did it in a1111. An important thing I found is to put 1.1 or more in the img2img noise multiplier to create extra things which did not exist on the low res image. Don't know hom to do that in comfy.

[west breach]

I created a tile size calculator node for dividing up the image evenly. didn't like how sometimes you get a border of weirdness

[uncut steeple]

usually doesnt happen for me

[thorny frost]

That is a proper furry octopus I would say

[noble shoal]

Thank you.

Now i want to pet him

[true wing]

wait... I just had an idea

[west breach]

this is with denoise set to 1.0 just to illustrate. first image you can see if you don't evenly divide the picture, it creates tiles with the leftover space. 2nd image has the tile size calculated to get evenly distributed tiles

[true wing]

I think this one is my favorite

[upbeat summit]

sophisticated furry octopus muppet

[west breach]

so cute with the goggles!

[noble shoal]

Interesting idea. Let's give aquatic creatures unnecessary diving equipment.

[thorny frost]

Yes, it is necessary to divide it evenly. In my case the high res image was the low res one multiplied x4 both width and height

[modest mulch]

i am loving comfyui , never going back to automatic1111. i have reduced image generation to under 1 minute even with 40 steps and 1024 resolution, and gpu v ram never takes ,more than 6 gb with rtx 3070

[true wing]

Man I love mixing two things together

[west breach]

feels weird going back to A1111 now

[modest mulch]

increase ur pc ram its cheaper, because i saw comfyui takes some huge amount of pc ram compare to automatic1111

[true wing]

Perfect.

[thorny frost]

Those spiders are like a 20% of octopus. I wonder if SD understand percentages

[upbeat summit]

plushy!

[west breach]

I just remembered I had a felting style

[true wing]

It might be because I was trying to swap the tentacles with spider limbs, rather than doing a full hybridization. I got either octopi standing like spiders or spiders with tentacles. I'm sure I could get it exact if I messed with it enough but I'm happy with the results.

[west breach]

what the hell did they catch

[thorny frost]

octopus with plasticine brain

[upbeat summit]

storytelling in SDXL is so good now 👌

[thorny frost]

That is looping the loop

[true wing]

I think it's because it understand the prompts so much better so it's easier to get consistent results

[upbeat summit]

that's exactly it

[thorny frost]

Yep, you can ask chatgpt, bard or whatever to write a history and SDXL will summarize it wonderfully in a single image

[upbeat summit]

that's really cool 🙂

[thorny frost]

Even better than Dall-E in that aspect

[true wing]

I can't wait for someone to get control nets trained on this thing. It's already so good at just understanding the prompt, with controlnets it should be easy to get the perfect image pretty consistently. It will probably take a few months though.

Which reminds me I think my next step needs to be practicing img2img and inpainting, I haven't even touched that yet because I wanted to get my workflow right.

[west breach]

Anyone try the reference only node that is pinnedd?

[civic sigil]

There are depth controlnets

[boreal bough]

inked, drawing of a girl sitting on a rock, grand canyon in background
inked lora coming along nicely c:
(40/200)

[true wing]

Interesting I'll have to take a look. I've seen some people asking for them and hadn't heard any were released. Thanks for the heads up.

[thorny frost]

Img2img is where real control is. I have got my own real world drawings perfectly finished with minimum denoising and continuously retouching the results in Photoshop. Not controller at all. It is laborious but I get exactly that I want

[true wing]

That's really cool. My art is decent but my proportions and perspective tend to be off so it sets the whole thing off.

Although in fairness that was 1.5, still haven't tried on XL

Alright I've been at this too long, I should take a break. Here are some of my 'failed' cat spiders for anyone I've traumatized.

[west breach]

not a single chainsaw man prompt in sandbit's file, and only 1 with saitama 😦

[indigo vine]

a meatball playing electric guitar at a spaghetti concert

[paper phoenix]

which genre do you think left and right meatballs are playing?

left gives me wedding band vibes, but right looks like he could be playing nu-metal

righty could deffo play static-x riffs

lefty would play something self indulgent, like an hour long santana riff nobody's ever heard of

oh god, why do these things have headcanon now?

[noble shoal]

May i ask what the sandbit file is? I think i missed something.

[west breach]

#🌠｜show-and-tell message

I had to strip out the double new lines before I could use it as a wildcard file

[noble shoal]

Thank you very much. Now that's some file.

[zinc cargo]

where can i see your stories man?

[meager canopy]

https://tenor.com/view/beaker-meh-looking-good-cool-gif-9589492

[noble shoal]

I just did, and it feels like an img2img with 1.0 denoising ,which behaves controlnet like and keeps the subject in place? I actually have no idea what it is suppose to do but it could be helpful.

[jade forum]

do weights in ComfyUI work the same way as A1111? example (((red hat)))

[zinc cargo]

can you suggest a good workflow for img2img?

[west breach]

I think it's supposed to help guide the generation. Found this example on reddit

[west breach]

use (red hat:1.3)

[ionic dragon]

which one do you prefer?

[peak dove]

Left side

[west breach]

right

[ionic dragon]

now i am confused

[thorny frost]

Depends on what you want to achieve. If it is about retouching your own art I recommend using Photoshop and some Photoshop SD plugin like stable.art which uses a1111 API mode.

[zinc cargo]

actually i dont have a set goal, i just want to expriment with img2img as i've never did

[lilac wren]

Totally agree, Stable.art is really great

[peak dove]

Some rare beauty Sytan's w/flow

[ionic dragon]

whats the prompt?

[west breach]

love the orange and teal colour pallette

[peak dove]

PROMPT = a vivid watercolor depiction of tinga tinga art, diverse Rococopunk Afrofuturist women with beautiful and bold head wraps walking toward a huge moon in the background, holding hands, walking away from the camera, collaborating in a creative and productive environment, women empowerment poster, photography, inspired by the styles of tinga tinga art, victo ngai and vladimir kush tinga tinga art

[lilac wren]

@thorny frost did you try Stable.Art with SDXL? Its working?

[ionic dragon]

how did you divide it, like this?a vivid watercolor depiction of tinga tinga art, diverse Rococopunk Afrofuturist women with beautiful and bold head wraps walking toward a huge moon in the background`````` holding hands, walking away from the camera, collaborating in a creative and productive environment, women empowerment poster, photography, inspired by the styles of tinga tinga art, victo ngai and vladimir kush tinga tinga art

[peak dove]

All in primary positive prompt - in secondary positive prompt = warm colors, translucent, glow

[thorny frost]

Then I don't know, feel free to experiment like the possibilities are infinite. Try different models over the same image, if you use a1111 set noise multiplier >1 to add details, use Lora and inpainting for changing faces. Don't know so much things...

[ionic dragon]

oh ok, tq

[peak dove]

But divide it up as you like ... see what happens - a creative accident might arise - serendipity!

Some facial features benefited after upscale; but a lot of detail gets lost after upscale ...

[thorny frost]

Not yet. But since stable.srt is nothing but a1111 rendering inside photoshop I suppose it works exactly the same. The base at least. Don't know if the a1111 refiner extension will work too. I suppose it does

[high skiff]

does anybody know where to put your args in oobabooga?

seems they changed that as well

[west breach]

session tab?

[lilac wren]

Yeah thats what i was thinking too

[high skiff]

MAD MAN

thank you

I am not sure why they make the standalone and manual installs so massively different

[west breach]

Love the prompt @peak dove

[peak dove]

Yes, the watercolor is so soft ...

[west breach]

I'm still really new to the oobabooga

[indigo carbon]

no refiner

yes refiner

[high skiff]

I used it monts ago, and they changed it so much

[west breach]

would be nice if the refiner didn't turn the grass into little strings

[indigo carbon]

hm, could this possibly happen if I used too high denoising?

[west breach]

I tried over a hundred different step settings with this image and could not get away from the grass looking like green fur or strings

[peak dove]

Pomegranate Girl

[indigo carbon]

so, I inferred the refiner correctly?

this is with the new A1111 branch

[west breach]

I find it's hard to get blades of grass like this

[indigo carbon]

same gen, lower denoising strength.

[ember cargo]

some pics generated with the improved Blame lora I finished training a while ago

the girl is supposed to be Emma Watson lol

[west breach]

love the inky style

[thorny frost]

SDXL interpretation of Bing prompt-story: "After a deadly virus wiped out most of humanity, only a few survivors remained in isolated bunkers. One of them was Mia, a young woman who loved plants and flowers. She had a secret project: to create the last garden on Earth, using seeds and soil she had collected before the outbreak. She hoped that one day, she could share her garden with someone else. But she was not alone. There was someone watching her from the shadows, someone who wanted to destroy her garden and her life."

It is amazing how well SDXL understand normal text

Would exchange that for correct hands any day though

[ember cargo]

i passed that prompt to my lora lol

[thorny frost]

First one is spot on. Second one where is mia?

[ember cargo]

somewhere in the latent space most likely

[west breach]

try this workflow I've been working on today. it upscales to 2k and fixes faces and hands pretty well

[peak dove]

Pomegranate Girl

[thorny frost]

some nodes missing. Working on it...

[west breach]

you also need the detection models. if you have the manager extension installed, you can find them in the install models menu

[ember cargo]

"a robot sitting in a Greek marble throne room by nihei tsutomu" upscaled x2, see original to see the details

[thorny frost]

i already did. manager cant find UltralyticsDetectorProvider though

[west breach]

that's in the impact one

[west breach]

https://github.com/ltdrdata/ComfyUI-Impact-Pack

[ember cargo]

amazing

[thorny frost]

working now

[thorny frost]

that is ok? i lack the sparkling lora

[west breach]

Oh don't need that for your prompts

[west breach]

looks great!

[west breach]

https://civitai.com/models/119711/sparkling

[upbeat summit]

"he doesn't care about the backdraft"

[thorny frost]

ty. looks pretty nice workflow. Will be my default for now. Will add some intermediate preview image to know what is rendering before finished because it takes a while. 🙂

[west breach]

yeah, it's pretty thorough 😄 good idea

[upbeat summit]

no more tears!

[west breach]

I think my sparkling lora helps a bit with hand poses since it was trained on outputs from a 1.5 model that had predictable hand poses

[ember cargo]

i tried combining my lora with pixelart xl lora 🤔

[upbeat summit]

nice

[peak dove]

When I have a queue of prompts in ComfyUI - is there any way to see the names of the prompts as they are passing thru at all?

[ionic dragon]

you can click load

you can get the worflow of that specific execution

and you can view the prompt in prompt nodes

[thorny frost]

that is interesting. preview of the default prompt with sparks lora. Will try anatomy now.

[upbeat summit]

you could output them to the console with the Text to Console node or I use the show text node from https://github.com/pythongosssss/ComfyUI-Custom-Scripts. you connect it to your prompt source node / or processor (can even contain wildcards) and it will always show you the current processed prompt

[thorny frost]

um truck driver hands, but those are the sdxl best feet i have got till now

have to go now, will try more later

[stray warren]

How do you setup wildcards?

[upbeat summit]

you could use Mikey Nodes by @west breach. The Prompt With Style node does have support for wildcards, wildcard files and more options to do dynamic prompting: https://github.com/bash-j/mikey_nodes

[stray warren]

Oh I guess I missed that, thanks!

[peak dove]

...

[peak dove]

tq

[indigo carbon]

@west breach I think I inferred the refiner correctly here?

[west breach]

best I've seen 👍

[indigo carbon]

thanks! I made it so the refiner stays in CPU RAM and base stays in VRAM and when it's time for the refiner to stop/start it swamps the UNET in CPU RAM with the UNET in VRAM and back.

[noble shoal]

Does anybody knows a tool, workflow or tutorial to get pixelart outputs pixel perfect? 🤔

[indigo carbon]

@west breach is that smart or stupidly complex?

[west breach]

sounds pretty smart to me!

[brazen patrol]

The real question to me is: regarrdless of not having the refiner hurt the result, did it actually help it over the non-refined version?

Can you say: oh this is so much better with it

[indigo carbon]

yes, what ever it did it's better than what I did before. see: scroll up in this chat

[brazen patrol]

THat's the point: I did. I compared this last result with the original one labeled 'no-refiner'. I canoot even see one iota of improvement. Maybe I am missing it?

[brazen patrol]

refined and not refined

your posts

[brazen patrol]

and if in doubt:

no refine

[indigo carbon]

they are both the last one I uploaded. this is the one I was talking about:

[brazen patrol]

refined

[indigo carbon]

this is without refiner, labeled wrong

[brazen patrol]

I am talking the result aftr you claimed to have solved it

[brazen patrol]

here

[wet nacelle]

refining works by having the base render at X steps then the refiner rendering the rest of the prompt.

You should only have one output image.

[indigo carbon]

yes. this is the refined result

this is base

[brazen patrol]

are the workflows in tose images?

[indigo carbon]

the refiner did improve it; and in the way I implemented it doesn't effect performance that much.

[stray warren]

Looks like it's not useable without the refiner model 😦

[wet nacelle]

If you're talking about the examples on the git then no.

I have two comfyui workflows that I can share that use it properly.

[thorny frost]

pretty reasonable hands and feet from @west breach workflow!

so the secrete is in the sparks?

[wet nacelle]

Each of them are using the SDXL-VAE and an upscaler setup as well.

[west breach]

They look awesome! Dunno if my lora is doing all the work, but I think it's nudging it in the right direction

[brazen patrol]

I'm confused. I thought he was the author of the pink cubes, not you

[thorny frost]

the statue looks perfect

[wet nacelle]

I'm just confused here I think.

If you are using the refiner as a way of getting an output image after having the base already rendered then you are using it wrong.

I don't know what you guys are doing I think.

[brazen patrol]

I am not using the refiner at all. I already explained I get consistently worse results or sideways, and demonstrated this

[wet nacelle]

That's wacky man. Sad to know that you're having issues when I and most other people aren't.

[thorny frost]

i have lots of reference images for artists. Will try to train some lora on anatomy. Never have done it though before, so probably will get crap.

[brazen patrol]

I haven't had a single example where the refiner was a clear win over not using it.

[wet nacelle]

I mean are you rendering the base as a blob and letting the refiner render the blob out into a full image?

[meager canopy]

@west breach It takes some coaxing sometimes, but nice when it sparkles 😄

[wet nacelle]

comfy ui?

[brazen patrol]

it either gives a different flavor, but not better, or it hurts the result

[cinder aspen]

100 people give 100 different opinions on which workflow works and which doesnt

[meager canopy]

My first images didn't sparkle at all.

[brazen patrol]

Blob? I just use the standard work flow with 80% Base + 20% refined

[cinder aspen]

I bet the most basic and default setup works well

No custom ones

[brazen patrol]

and then compare it to Base 100%

[west breach]

yeah dunno what happens sometimes when the don't show up, but loving the sparkle in that image 🤩

[brazen patrol]

so 60+15 vs 75

[wet nacelle]

This is why I'm getting confused still.

Are you rendering the base image fully and then letting the refiner literally refine the smaller details of that output?

@brazen patrolAre you using ComfyUI

[west breach]

It wasn't trained on daytime images, so less effect usually with those. I can see it added glitter texture to the jacket on the second picture. It can add glitter, sparkles and stars

[thorny frost]

can i post nude marble stautes i will get banned by the spanish inquisition?

got some amazing anatomy

[brazen patrol]

Yes, standard ComfyUI setup

[meager canopy]

I was getting some of the glittery tops without the lora though. I think it just takes the trigger words from the prompt and creates it.

[meager canopy]

If you do, it'll be unexpected 😉

[urban breach]

nice workflow, what is __sandbit?

[thorny frost]

[west breach]

a massive text file made by user sandbit who scraped the showdown channel for prompts

[meager canopy]

https://tenor.com/view/hiding-reverse-gif-19206163

[wet nacelle]

sorry for not reading what you sent while I was typing. My bad.

I'll tell you this. I have a workflow that uses a full steps of 40. I render the base to 20and then refine the rest at at again 20 (to equal that 40)

I don't let the base image fully render out.

[thorny frost]

i think the classical pose contributes to the perfection of the srtatue

[urban breach]

Interesting, what's the significance? Are you somehow choosing all keywords with this wildcard? Sorry this is a new concept

[meager canopy]

The steps depends on your sampler, but I use 30 steps and base ends at 21, where the refiner starts.

[brazen patrol]

I am well aware of how it works. The ideal balance is not so much refiner though (20+20), so in your case it may completely change the end result since the refiner also interptts prompts differently. Hence refiner as a term

[wet nacelle]

That too makes sense.

[west breach]

The node detects the __sandbit__ and opens the txt file and inserts a line from the sandbit.txt file into the prompt

[urban breach]

Awesome thanks

[brazen patrol]

so if I shut off the refiner, I also disable the end with noise setting

[stray warren]

@west breach is it possible to use your v3 without the refiner?

[wet nacelle]

I'm done with this. I clearly don't know what I'm talking about.

[west breach]

yeah, it just needs any clip input into the refiner clip, so just feed the base into it

[stray warren]

oh great, thanks!

[brazen patrol]

The devs used some system to poll people on Base vs Base+Refiner. I only wonder about how they did this. I used their bots, enjoyably, but often would get two very different results, and not always remotely what I wanted. A or B? wel.... A? but hard to really know. I never was shown an A or B where the diff was subtle or small, indicating a genuine Base vs B+R comparison. So if randomly throwing in with or withouts and mixed results, they would have a huge amuont of noise in the results

[thorny frost]

correct way to render in sdxl is to render the first 70 or 80% with the base and then finish it with the refiner, not to render the whole imagen with the base and then do img2img with the refiner. I think stabilityai said it. In case someone dont knoes it yet.

[wet nacelle]

Oh you're fucking right.

Oh shit. It can be much more difficult to get what you want when the model itself is built for something else while also being talked about as an everything model.

[wet nacelle]

That is hard to read

[late marsh]

https://tenor.com/view/mandalorian-this-is-the-way-gif-24826742

[brazen patrol]

It's ok, I parsed it well enough. 🙂

[brazen patrol]

I doubt this is a problem, since most people will be using pre-baked workflows such as Comfy's, which come with it already set up correctly.

[meager canopy]

@west breach Any chance of adding a dropdown menu, to your v3 node, for the styles? Like "SDXLPromptStyler" uses?

Makes life so much easier when selecting

[thorny frost]

only satying, just in case somebody didnt have it clear.

[west breach]

I had this in the original and the v2. But my list of styles was getting so long, plus when you type it in you can add multiple styles

[brazen patrol]

I hear you. I mean, I see these reports about wacky combinations such as Scott Det's R+B+R, and it is not that I doubt the end-result was enjoyable. I just question whether this was not just random noise as a result of multiple unpredictable sequences. In other words: could I not easily get as good results without such an elaborate setup?

[meager canopy]

With such a long list, it's impossible to know them all to make use of them, without having to refere to a seperate list all of the time. It may as well then be an embedded list you can choose from, with also the option to type it in.

[meager canopy]

Maybe I'll just use v2 🙂

[thorny frost]

i think it is ok to try weird not official things anyway. Maybe there are workflows not even stabilityai knows that gives amazing results. That is that makes comfy great. We can experiment without really knowing what we are doing .

[little gazelle]

Anyone used the diffusers library to load in SDXL on Google Colab? Getting 3.67s/it

[deft coral]

what LR is this

[west breach]

I'll see if I can add it to the bottom of the node without it breaking the workflow. V2 won't have the lora:lora_name syntax support

[meager canopy]

I'm not worried about that, I have lora options all over the place already 😄

tinyterra nodes

[brazen patrol]

Oh absolutely. I love it! Please don't misunderstand me. I know the caveats of this node setup, mostly in appearance and figuring out what is doing what, but love the power it gives to experiment. No, what I meant is this: suppose we see an image. Any. and I tell you this was the result of R+B+R+ 123 steps at Karras+ 15 with... whatever and so on. And then conclude: this is thanks to my absurd sequence. Now, this is not false mind you as it DID come from it. But it leaves the impression that you could not get this without it. When maybe just Euler at 45 steps and the right seed would get the same thing.

[azure oxide]

Sure but mabey the sequence leads to it more consistently

i could be wrong but that should be how the refiner came about -- SAI didn't need to include it but they did it anyways to try it out

in the paper, they mentioned that some results werent clear and from here probly had a several options to proceed forward in solving this problem, refiner was probly just 1 of them

[thorny frost]

absoplutely. BTW talking about steps, yesterday i found that doing 125 or more gives a totally different image, like a new seed. Bet stabilityai didnt know that.

[boreal bough]

let it be known, I deeply respect this.
Going straight for the endboss of LoRA training XD
protip, overfitting is the easiest solution - just get 50 images of different people in the same pose. done & anatomy learned (only results matter, right?) XD

But if you wanna do it properly, then you'll be doing a lot of manual tagging + getting equally distributed datasets of all the anatomy parts you wanna train/improve.
(unknown anatomy parts take longer to learn, so you'd have to eventually train those more than everything else)

[wise edge]

Do you guys know how to curb this kind of concept bleeding? My starting prompt was just "photo of curious little (mechanical robot humanoid:1.2), standing next to a (large fluffy dog:1.25)" Or another example, if I try to give someone red hair, they'll end up having red everything.

[thorny frost]

gasp, sounds like a lot of work!

[brazen patrol]

Or not. How much detailed testing was done? I am a super methodical person as a rule for such. So for example, in the last two days, not having the benefit of the time powerusers here already enjoyed, I put a single seed through its paces. Meaning every sampler, then tested numbers of steps, then permuted a single word, the order, in parentheses, and so on, and ran it all again and carefully compiled the results. HUndreds of images were done this way to really see what was happening. Oh, and yes, with and withut refiner, and varying it with few steps, more steps and slight variations of %. I.e. Refiner for 40% or 30, 20, or 10 (and of course zero)

[brazen patrol]

it was super instructive

[azure oxide]

this isnt an sdxl exclusive thing, https://github.com/hako-mikan/sd-webui-regional-prompter is one method of solving it

idk if comfy has an equivalent tho

did a quick search and came up with this for comfy
https://github.com/BlenderNeko/ComfyUI_Cutoff

[wise edge]

Yeh, any kind of composition or "art direction" gets difficult... I think I saw something like that in the custom nodes in the manager too

Thought to ask if I'm missing something obvious to help it even a bit. "Just do this dummy" -type of a thing 😄

[thorny frost]

kohya contine beiung the best way to train things, isnt?

[boreal bough]

your best option is to find a single or 2 prompt words that essentially define the act of both animals being present in one frame

"cat chasing a dog"
"cuddling"
"friends with"

[glass dove]

Hey friends, has anyone seen a good consistent technique for legible text?

[boreal bough]

other than training a lora? no
there are some words that clearly exist often enough in the dataset that it works, but you can't actually "write" text.
with the word "dog" would be the way to prompt for words. ('text' doesn't work at all, 'word' makes the magic happen)

[glass dove]

ahhhhh, will try word, was using text :)

and roger that on Lora

thank you :)

[cosmic matrix]

why we are keep voting A and B in bot rooms?
its supposed to continue improve for a better model training or something?

[vocal stream]

yes

[vale eagle]

I am thinkg about training a text lora with specific words. Do you think it is posible in SDXL?

[vocal stream]

this seems clearly untrue #1100170514670039070 message #1100170514670039070 message

[meager canopy]

I've seen this with 20-40 steps on some samplers. You don't need to run so many steps to see it 🙂

[urban breach]

Here we combine this exquisite mansion with a dorm room style "sleep on the couch" theme

[wet nacelle]

ha

[boreal bough]

and it also works without 'text' or 'word' if you use sign.
basically 'text' feels a bit like a garbage collection tag, where everything texty looking was put it. Yes it can work, with the right context.
'word' however usually can be made to work more often than not in out of context situations.
like making the word "sdxl" neon glow and float in the air in a dark forest

[boreal bough]

sketch, While cleaning her room, teenage girl Yuu discovers an old book containing a stunning map with glittering stars indicating an unknown island. The map is very detailed and hand-drawn, with intricate compass roses and sea monsters, as well as a red cross reminiscent of the maps of ancient explorers. The artwork is classic and vintage in style, with warm colors and soft light evoking curiosity and adventure. The image should be ultra-realistic
+sketch lora XD

[vocal stream]

in the first example it's only picture with text, not sign with text

[thorny frost]

DreamShaper XL1.0 gives me better results than the default base consistently. And it is in alpha... There is a lot of room for improvement.

i think SD will got better than midjourney.

[barren bluff]

This is taking forever, it got stuck here for like 5 minutes

I am using a 3060 12GB with 16GB RAM

[broken hare]

I tried it last night and got some good results too. Have you tried XL6 -Hephasitos? I saw a video showing results so i gave it a try.....pretty good at people

[thorny frost]

downloading......

[indigo vine]

I have the same card and while it's not as fast as Comfy, it's far from 5 min freeze. In options, stable diffusion, check that you have "Checkpoints to cache in RAM" and "VAE Checkpoints to cache in RAM" set to 0

[broken hare]

i keep seeying people with that card have issues....some have said that you need to revert drivers to previous verson

[pearl shoal]

How to force comfy ui to use xformers? Mine is not using it and there is no flag that turn it on. There is a disable xformer flag no enable xformers flag tho.

[broken hare]

has sort of a strange face fix mixed in, but people's bodies and clothing looks good......and you almost dont need to run refiner

[azure oxide]

if its still processing, i'd check the console to see if it hung. it could also be that its trying to load both the base model + the refiner into your RAM and its stuck waiting for more space

[barren bluff]

Hmm

[azure oxide]

this is how you manually install it btw, shouldnt need to do that

[thorny frost]

lets see how it combines with mike workflow and sparky lora...

[barren bluff]

It hung on the "creating model" part

[pearl shoal]

I do have xformer installed, that is not the issue

Comfy is just not using it

[wet nacelle]

ah okay sorry then

[azure oxide]

should be able to just input --xformers into run_nvidia_gpu.bat but i have yet to use comfy with any args so i could be wrong on this

[thorny frost]

now is a face with personality

[native moon]

@visual glade hey i love comfyui and it is nice to play with it around. I woder if u also had the idea of adding node conection to all the numbers and settings. And math nodes. Simular to blender. Thus you coud use math nodes to calculate something and controle for example the power of a lora or how many steps a sampler does. This would be soooo cool . And another cool thing would be loop node setups. With these you could loop throgh a generation and each time changing a setting or maby even more complex stuff like shifting the noice imput image.🤔

[azure oxide]

check your task manager (ctrl shift esc) and see if your RAM or vram are full

idk too much about a1 and sdxl but i do know that it is awful in loading the models in

[thorny frost]

curiously always paints the same woman, but feet and hands are nearly perfect

[broken hare]

not bad right?

[wet nacelle]

Does the terminal on startup say this?

Do you see the exformers here?

[broken hare]

i did get repeated likenesses too. I hope this gets merged and used for more looks because there is something good in there

[thorny frost]

yep

[urban breach]

Now we're talking lol

[pearl shoal]

No, it used to but then I started experimenting with flags and it does not show it any longer

[wet nacelle]

Ah okay I see. I don't know what that is and hope that you get the help you need.

[barren bluff]

RAM usage at 96% and the VRAM is barely being touched

[pearl shoal]

Thanks anyway

[azure oxide]

yeah looks like its stuck hahah

your best bet is to kill the whole thing, mabey restart it and try again, or mabey try another ui

a1 is kinda awful with sdxl

[urban breach]

Simba's kingdom is looking a little... urban lol

[barren bluff]

I guess I will try comfyUI instead

[thorny frost]

male hands look really good too in a consistent way. Only some minor errors.

[still dove]

yeah!! 😄

[thorny frost]

correct number of fingers but too many laser sabers. Cant get everything ok! 🤣

[dense chasm]

finally, sdxl 1.0 lora training works on 16GB GPU RAM,params optimization is so important with convolution network

[thorny frost]

moar loras!

[broken hare]

yep....the newer models are still not quite there, but pushing the needle a little bit at a time

i just saw a flood of new ones on civitai since yesterday....have you tried any of the brand new ones

[thorny frost]

things are going pretty fast indeed, much faster than with previous models. It is only a week since xl1.0 released and we have a number of custom checkpoints already and loads odf loras! I havbe tried most of them and i think the better most equilibrated adn "proffesional" one is dreamshaper, which was launched first btw.

[broken hare]

most look pretty weak by the preview pics, and comments.....but i think those two models do give better results for humans.....but for other things, i dont see much of an improvement

i dont care much about celebrity loras, but there is one that The Last Ben did for a photography style that is pretty good

[thorny frost]

IMO XL is great in almost everything except humans.

[uncut steeple]

Trying LORAs in my workflow with the voxel LORA

And the DFunk Lora

And the cyborg Lora

[prisma wyvern]

Can someone generate examples created with Tiled Diffusion vs Tiled VAE, and what is the difference between them? Which one of these methods generate faster images with SDXL? I tried running Tiled diffusion on my system it took way tooooooooo long time which produce 2 tiled split screen result and turned 1 dog into 2 dogs in each tile split horizontaly 2 + 2 = 4 dogs 🤔

[smoky patrol]

oh, love it

[smoky patrol]

CAT!!!

[dense chasm]

pure coding no UI lora training,loll.

[dense chasm]

estimate 1.5 hours finish 10 epochs for training lora,16G RAM runs so well

[solar tusk]

why is there still voting going on? how will the data be used?

[ionic gulch]

Yes, Euler A does it a lot - often multiple times if you increase steps - that's one of the reasons why I prefer different samplers even if results with Euler A are good in general.

[ionic gulch]

There are addons for math, the real problems are ways to switch different behavior by addings switch nodes and allowing nodes to handle empty input.

[hollow halo]

Comfyroll addon has switch nodes

[ionic gulch]

Thanks, I will check it later when I get home.

[ionic gulch]

are they general switch nodes or only for a single purpose?

[barren bluff]

ComfyUI is really fast lol

[dense chasm]

use Euler A for my lora sdxl sampler training, loss 0.13-0.15,good result

[barren bluff]

The hands are kinda wierd though, maybe it is me who forgot to put the negative prompts lol

[peak dove]

Leopard Lady etc

[zealous horizon]

reroutes are just sooo big, therefore I made this (can be toggled and allow moving and connecting while in dot mode)

[peak dove]

Pareidolia

[bold osprey]

how can you toggle the connection routes to a boxy style like this

[next ruin]

Hi, I would like to learn more about mashing various styles for generating images. For example Barbie movie style applied to Harry Potter scenes. Do you know good resources to read/watch?

[zealous horizon]

also part of my extension: https://github.com/failfa-st/failfast-comfyui-extensions

[bold osprey]

perfect thank you

[half ivy]

i am here

what have i missed o0

[wet nacelle]

not gore

[wet nacelle]

oh fucking wicked

[bold osprey]

are you using the voxel lora so far?

[unreal plover]

yes

an alien skull face medieval knight action figure riding a horse, castle in the background, extremely detailed,natural lighting,film grain,((voxel syle)),lora:VoxelXL_v1:1

[dense chasm]

sdxl lora training of hitokomoru successful, prompt:1girl, aqua eyes, baseball cap, blonde hair, closed mouth, earrings, green background, hat, hoop earrings, jewelry, looking at viewer, shirt, short hair, simple background, solo, upper body, yellow shirt

[bold osprey]

I'm afraid of going back to A1111 at any point in time after using Comfy

[next ruin]

you can't comfortably use refiner with a1111

[wet nacelle]

https://tenor.com/view/fishtank-josie-fishtank-live-josiecel-gif-15644911615554340556

[bold osprey]

heard you can use it as a highresfix now

[next ruin]

but you can't use hiresfix then 😄

[bold osprey]

[next ruin]

would be cool if people would post their prompts and seeds alongside the images you know.. for educational purposes

[bold osprey]

yeah putting it into comfy isn't worth the hassle most of the time

[fresh path]

prompt?

[zealous horizon]

Copy image directly from the context menu "Copy Image" 🙂

[copper rose]

I just finished training my 400 image style lora, it can do almost anything and is only 40mb!

[trim orbit]

loras are really cool like that ❤️

[bold osprey]

prompt: (LEGO tiger:1.1) sneaking through the jungle, orthographic view, color graded, wide shot

[copper rose]

my previous one was 800mb and was essentially taking over the training pool for SDXL. It was having a ton of difficulty generating variety -.-

[cinder aspen]

first time setting up

god knows what will come out

[trim orbit]

they're easy to over train with sdxl

[copper rose]

I mean, I think thats just because of the way people are training atm, I used DIM 8/1

[inner ruin]

my inpainting isn't transitioning very smoothly right now in SDXL 1.0. Any tips to make it blend better?

[copper rose]

Prev I used 128/128

[cinder aspen]

holy crap is it slow with 4 batch

for me

takes a few minutes

[trim orbit]

yup. i've found no need for more than 8/1 too.

[cinder aspen]

use generative fill in Photoshop or Firefly

[trim orbit]

thats sort of what i mean. because sdxl is very easy to overtrain, people are stumbling all over those pitfalls

[inner ruin]

I want to do it in SDXL purely lol

[cinder aspen]

oh, then i dunno xD

[copper rose]

Unless you want to only produce your dataset yea LMAO

lora vs original

[vocal rapids]

I'm so hungry ....

[wet nacelle]

Should I cancel my Mid Journey subscription.

I don't use it that much anymore and I've completely moved over to stable.

[trim orbit]

answered your own question. if you're not using it why pay for it

[wet nacelle]

I've had it for over a eyar now

[trim orbit]

saas' biggest caveat

[wet nacelle]

I know that

ahhh

[trim orbit]

bite the bullet

do it

cancel the plan

[wet nacelle]

It's just difficult

[trim orbit]

do it!

[wet nacelle]

I'll archive all of my prompts. You aren't allowed to view your prompts a while after you leave. @trim orbit

[trim orbit]

will you ever look at them ever again?

[wet nacelle]

Yes. I use them for when I have nothing else I can think of.

I've made 40+ thousands images.

[trim orbit]

cancel it all. the whole plan. all the prompts.

[ionic gulch]

https://i.imgur.com/fXX4Kpb.jpg

[wet nacelle]

@trim orbitnaw. I'm going to save all of the prompts.

https://tenor.com/view/jerma-gif-20630865

[trim orbit]

https://github.com/topics/midjourney-download these might help

[wet nacelle]

Thanks.

[peak dove]

MidJourney still has a finish totally unavailable in SDXL, especially Niji

[wet nacelle]

What does that mean?

[trim orbit]

midjourney fans might not like to hear this, but sdxl is a better model than what MJ uses. all objective tests reveal this.

Subjectivity YMMV as we all have preferences of course. Where ever objectivity can come into play, SDXL wins. Simple as that.

[echo pine]

What did you train it with and any chence u could share your config?

[copper rose]

Kohya and I used the config @boreal bough posted!

Original vs Lora

This is lora after hi-res fix :3

[echo pine]

where was it posted? can't find it

[copper rose]

https://discord.com/channels/1002292111942635562/1026382406279770152

[hardy cipher]

anyone know what would cause a "pooled_output" figured out it was the efficient loader causing the error. but can't figure out why.

[copper rose]

Has anyone figured out an effective way to reduce twinning in resolutions over 1024x1024?

[midnight shuttle]

Use a supported SDXL resolution.

1152 x 896
896 x 1152
1216 x 832
832 x 1216
1344 x 768
768 x 1344
1536 x 640
640 x 1536

[echo pine]

thanks

[hardy cipher]

^5

[copper rose]

TYVM ❤️

[late marsh]

Peekaboo!

[autumn forum]

If you want bigger the best thing to do would be to upscale the resolution from the trained sdxl resolutions, then run a base pass over it again, or just use the upscaled as your final image.

[unreal plover]

An action figure of a cyberpunk cyborg cop controling a punk group of muscular bad guys, cyberpunk street in the background extremely detailed, natural lighting, film grain, ((voxel syle)), lora:VoxelXL_v1:1

[cinder aspen]

So the first few results of SDXL were really good although it took some time to render

And m laptop even lagged for a few seconds

[cinder aspen]

I also didnt have to install Pytorch etc

[trim orbit]

when the memory usage overflows the vram into the system ram, the entire main bus will be saturated and you'll get obvious system latency

pretty cutting edge stuff

[cinder aspen]

I took the settings from a youtuber maybe thats why

[trim orbit]

well you've got a laptop so i'm going to assume 8gb vram

[cinder aspen]

4096x4096 width and lenght to 1024x1024 target and then i changed also to 4 batch which results in 4x Base images, 4x Refined

8 images alltogether

I dont have any clue what those other changed settings where tho tbh

with noises etc

I just followed along

[keen phoenix]

I tried the DreamShaperXL checkpoint, it's actually really nice !

does anyone has good upscaling method ?

[fresh path]

cool

[meager canopy]

What is the purpose of the lora? You say it can do "everything", which seems a bit general 😄

[copper rose]

to make psychedelic imagery, I wanted something to encapsulate my personal taste. I also wanted it to have depth in its scope.

[meager canopy]

"Ultimate SD Upscaler" is good. It's in my workflow, here :
https://civitai.com/models/119257?modelVersionId=129558

[keen phoenix]

thank you !

[fleet harness]

Okay, here again tryna ask for some help.

was-node-suite-comfyui always bugs out and won't load certain nodes. I've tried everything from fresh installs to manual installs to asking for the nodes themselves. I'm looking for TEXT MULTILINE and a few others. If anyone knows what to do please help. It's so annoying because half the workflows use was-node-suite-comfyui and I can't get all of the nodes to install!

i tried tech help. no go there

[unreal plover]

someone tryed 1.5 + xl refiner ?

[shy kelp]

hmm, have you?

[unreal plover]

i will try on auto1111

[shy kelp]

ok i might as well try too

but first .. im installing the refiner extension so i can use refiner in txt2img w/o having to go into img2img

[cinder aspen]

Is img2img available on XL? Didnt look uo

[shy kelp]

and with that ext you can also load up base and refiner inside txt2img with one click output

[cinder aspen]

didnt even look out for Controlnet

I guess CN is included with ComfyUI?

[shy kelp]

well not sure about CN with XL

dont think they are made to work with XL yet

[cinder aspen]

Thats okay, i use SDXL only for specific purpose

[sharp flame]

Experimenting with upscaling with the SDXL refiner model, starting to get a better understanding of it. The number of total steps effects how much gets denoised each step, and you can change which step you start denoising

[fresh path]

loving the speeds today

[sharp flame]

In this picture I did 50 upscale steps total, with the start_step at 20. I set the base model to stop and the refiner model to start at step 40.
Its sort of like a denoise at .6 with 35% of the steps being refiner

[shy kelp]

but does going above 30 steps show any visible difference?

[fleet harness]

YO

[shy kelp]

30 steps with base and 5 steps with refiner

[late marsh]

https://tenor.com/view/mando-way-this-is-the-way-mandalorian-star-wars-gif-25407126

[vital wolf]

does anyone have a workflow with only one checkpoint loader+lora+vae+upscale+hires with multiple prompting section?

[late marsh]

Searge is tinkering with something, can be found in Civit

[vital wolf]

oh, I will wait for it then.

[vital wolf]

but it has refiner checkpoint

[late marsh]

True

[alpine pine]

you can remove the refiner by setting the start at to the equal to the total steps

[vital wolf]

oh

also, i can't change the lora in searge's workflow

it's fixed to offset lora by SAI

when i clock on it, only offset lora is being shown

*click

[soft bone]

insert lora loader into sytan's

[high patio]

Any ideas what I can do to get the same image results as DreamStudio? I get much better results on DreamStudio than in Automatic1111 using the sdxl 1.0 model. Width, height, steps, and seed are the same accross both.

Does DreamStudio add additional details to the prompt or negative prompt?

[unreal plover]

I don't know if they use the refiner, surely, so try to install the extension

[soft bone]

#✨｜sdxl message

they're using their style prompts in the back

[high patio]

Thank you. In DreamStudio I did not select a style. Does it select one by default if I don't select one?

[soft bone]

probably not. but i think they might have a permanent negative. they also might be using loras, i know they've developed in-house style loras that the bots have used in the past

[trim orbit]

dreambooth has the style selector, which adds styles passed to clip L iirc. you can get this from an a1 extension

[soft bone]

i heard that on the bots the style prompts were passed to both l and g

[trim orbit]

i'd like to look into that more. maybe you're right

[soft bone]

i really dont like having to divide my prompts. i hope a guide comes out explaining the advantage of separating them vs passing same text to both

or comparisons at least

[trim orbit]

They're like having two separate actions on the latent space rather than just one for invoking images. You want them to act in concert, but you can also divide their attention across a wider scope of latent focus. Sytan's workflow has a nice bit of text about it. Well, i haven't seen the 1.0 yet but his first one did.

[shy kelp]

What are the GPU requirements for XL model?

[cyan crown]

I really need Controlnet working with sdxl

[late marsh]

https://tenor.com/view/star-wars-yoda-patience-sw3-gif-21070142

[hot rover]

Is there something better than WD14 tagger for interrogation? (It's good enough not great) Something more like clip interrogator that works inside of comfyui?

[ionic gulch]

is there a bug in step calculation in KSampler(Advanced) node? steps and start_at_step don't match the actual executed steps. works for lower numbers, but if you go higher with your steps, the actual steps (14) are a lot higher than what should be the right amount of steps (11):
https://i.imgur.com/ng78t4r.png

[vale eagle]

Outside of comfy, just tried a cool project. blip2+wd14+flamingo and summarize with llm.

[minor glacier]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/commit/91704abbae38a0e1f60d433fb08d7f7d99081d21
Look what got commited an hour ago

[ionic gulch]

same thing for 75/80 - should be 5-6 steps, but it does 9 steps:
https://i.imgur.com/sKJUJ1E.png

[midnight path]

What was the reasoning for making 1.0 use 0.9 vae now?

[midnight shuttle]

Looks like they finally admitted there was an issue. On the first day of release they were very dismissive about it and even claimed it was supposed to be that way. So I'm glad to see they changed their stance.

[midnight shuttle]

Artifacts in generated images with 1.0 VAE, especially when using refiner in Comfy. Artifacts not seen with A1111 base model only.

[ionic gulch]

for lower values it works and 25/30 gives you the correct 6 steps:
https://i.imgur.com/jcPsAgw.png

[shy kelp]

o.0 how can the vae be this big?

[midnight path]

Oh thank you 🙂

[midnight shuttle]

That is not a VAE. That is a base model. The name indicates that it has the 0.9 VAE baked in.

[shy kelp]

ahh ok

[hot rover]

That's pretty cool, I'm trying to refine detected faces in a conditional stage of an SDXL workflow, I wanted to generate descriptions of the face on-the-fly with interrogation. WD14 does well, but I would really like some better quality automated text to influence the generation more. I'm sure more options will be available in the coming weeks.

[soft bone]

https://github.com/theovercomer8/captionr

[thin nova]

also i put this together if you want to use instructblip https://github.com/briansemrau/kohya-ss-sd-scripts/blob/sdxl/finetune/make_captions_instructblip.py

I get well written captions like "The image features a young girl standing on a hillside, looking down at a house. She is wearing a yellow shirt and has a backpack on her back. The house is situated on a hill, surrounded by trees and a stone wall. The girl is standing near the edge of the hill, with a view of the house and the surrounding landscape."

[neon fable]

I also have a question for anyone who may help. I'm trying to find a way how to prompt SDXL to replace part of an object with something different. Is there a way to use something like "instead" in prompt? I know I could go for inpainting or regional prompter (once it's working with SDXL), but I'm searching for a way to do this within the prompt itself. An example would be "a photo of hamburger, cigarettes instead of meat". No matter what I do (give more strength to keywords, use negative "meat") it renders normal burger with meat in it.

[vale eagle]

I think wd14 is not that good at training a SDXL lora or model. I used 15 images with hand made prompts. Only 1 epoch and 10 repeat, it could get pretty decent result in 150 steps. Therefore, I tried that mixing tagging project.

[grizzled viper]

A dozen cigarettes, sandwiched between two brioche buns (?)

Could also try prompt editing… to get the composition of the burger the switch it to cigarettes between buns

[neon fable]

Trying it now. That would mean that "instead" wouldnt work, I'd have to find a way around i guess 🤔

[grizzled viper]

Might have some effect if used afterwards… dozen ciggs between buns, cigarette burger with cigarettes instead of meat

[midnight shuttle]

SD doesn't understand sentences the way ChatGPT does. It just sees words, actully tokens. That's why if you say a prompt like "cat with dog's head, dog's feet, looks like a dog only in the head and feet" you will probably get a bunch of dogs and a bunch of feet in addition to your cat.

[trim orbit]

https://tenor.com/view/phil-trailerpark-gif-5118951

[urban breach]

the workflows are just getting so good

[neon fable]

Yes, now I'm getting bunch of buns. Mostly I'm going quite simple. I'll try to find a way around. I mean, impainting would be the key then i guess

[midnight shuttle]

Yes, inpainting is key. SD does not understand complex relationships between objects. That is written in the documentation.

[hoary saddle]

anyone know what method CLIPDROP uses for their background removal? i'm sure it's ffmpeg related

[cinder aspen]

Might come eventually tho since its possible to do so and there is something that does that already

[midnight shuttle]

I assume you mean Midjourney?

[twin bluff]

can anyone help, how to make a prompt to put subject a little further from the "camera", it always crops the bottom and the top of the subject... negatives like "out of frame" don't help really

with sd1.5 don't have these issues. it's sdxl issue, can't get it no way. and no controlnet yet. maybe some "field of view" trick or something?

[trim orbit]

there's some level of relationships between tokens with CLIP but yeah, for the most part, its nothing like LLM

[inner ruin]

what sampler do you use?

[upbeat summit]

The Voxel Simpsons!

[azure oxide]

ymmv but ive had good luck with rule-of-thirds photography as a positive prompt. SDXL should actually have this happen less than 1.5 because of how it was trained, autocropping was the issue and sdxl had sort of a workaround to help prevent this issue

[upbeat summit]

it's probably your prompt or the resolution you're using why it frames it like this. try mid-shot portrait of or just portrait of

[urban breach]

i usually put full body as the second or third term in prompt. then anything that gives you detail on parts you want to see like shoes etc

[ionic dragon]

how can we add hires fix into Sytan's workflow?

[static prawn]

its allready in there

i just recommend to set the steps in the highres fix higher than recommended

[twin bluff]

tried... it works 1 in 10 times

[cinder aspen]

I mean Firefly/Photoshop generative fill. Midjourney doesnt even have inpainting ^^

[ionic dragon]

is there hires fix?
i dont think so

[indigo carbon]

MJ is out of the equation right now. it's not even close to SDXL.

[cinder aspen]

In terms of image quality or all in all/general?

[urban breach]

it also helps to change the resolution like more of a portrait style to help it give you full body shots

[indigo carbon]

both

[static prawn]

i think so

[cinder aspen]

For image quality MJ is very well toe to toe, for customization...well needless to say

[twin bluff]

thank you, but maybe you're right, the resolution im trying is 456x1472 upscaled by 1.7 to 775x2502 ... but already the first pass cuts the legs

[static prawn]

imo thats a highresfix

[unreal plover]

DPM++ 2M SDE Karras

[upbeat summit]

yeah, try a different resolution. SDXL works most reliable if your resolution fits into 1024x1024

[indigo carbon]

nah, midjorney doesn't even get close to this level of image quality.

if for you it's not, you're not using SDXL correctly

[cinder aspen]

From what i have witnessed and also used myself definitelly disagree there

But also depends always

Where MJ definitelly will remain behind is customization and advanced tools

[trim orbit]

MJ doesn't even have 1024 base resolution. Your quality comparisons are likely thumbnail glances

[cinder aspen]

Until they bring out their web UI SD and Firefly will be much further eventually

[twin bluff]

yeah, which is a little problematic for people, who are tall and narrow... if you need just the people, no need to make it rectangular... that's how I see it... so I use short width and long height, and I keep adding height but it always cuts the legs... at some point it starts making double legs or something, rather than making one person whole 🤣

[indigo carbon]

fuck no, mj isn't even a match. I NEVER saw MJ get to this level of quality and/or detail. if you disagree with this you're most definitely biased

[trim orbit]

subjective biases can be very strong and hard to seperate from

[midnight shuttle]

MJ and SD are different tools for different use cases. No need to defend one or the other.

[twin bluff]

for example, a simple prompt like centered full body photography of a woman in boots at 456x1224 upscaled to 775x2080 results in this, which would be acceptable if the hat was 20 pixels higher and if the feet were visible.

[trim orbit]

i woudln't call MJ a tool as much as it is a service

a very limited service

[midnight shuttle]

MJ and SD are different things for different use cases. No need to defend one or the other.

[cinder aspen]

I bet more like you are biased man lol

[cinder aspen]

That is true

You will find those three have different target audience

[indigo carbon]

oh, please. do you really think MJ is capable of producing an image with this quality?? and don't even get me started with text

[trim orbit]

Across all technical and objective categories, SDXL succeeds. This is just something MJ fans need to contend with. The MJ SAAS business model is only going to succeed for a blip. They've suceeded well but if they don't pivot, they're done.

Wouldn't be surprised if something like deviant art ended up acquiring whats left of MJ in a year

[keen phoenix]

What is the best upscale model ?

[trim orbit]

no. I think MJ uses old technology for their current version. A very refined and capable form of those older papers, but nothing like SDXL and can't compare to SDXL at all

[indigo carbon]

yeah, it's not even close

[hot rover]

I think he'll sell to Adobe in a year lol

[static prawn]

could someone explain me how i could add a save node to comfy?

[trim orbit]

Adobe doesn't need MJ at all

MJ's business model is hyper exploitive, and their real competitor is Adobe. Adobe will break them.

Stability will just keep pumping out open models. They're competing in a different way .

In a matter of week's after launching adobe's generative userbase doubled MJ's

[twin bluff]

tried adding rule of thirds as well, still cuts the legs at ankles height...

[tribal jackal]

Tried another lora training last night. File size and training speed were significantly improved, but it couldn't generate my dog like before- only random dogs. A couple settings got mixed up, caught in them review.

Made one wrong setting during overnight lora generation, lead to another model at 1.7GB. Currently retraining, not any faster sadly but this file size is lookin' real good and the sample images confirm it's able to recognize my dog when prompted with the unique text identifier.

final product quality remains to be seen

[trim orbit]

describing the photo rather than the position of the camera. "magazine cover photo", "rule of thirds", "perfect composition", "professional portrait" "studio photograph", "patterened backdrop" all give a context to where a photo isn't cropped. Without context then anything goes. Selfies are always cropped tight because most people suck at composure when doing selfies

[twin bluff]

thanks @trim orbit i'll give that a shot

[trim orbit]

"magazine centerfold" gives good full body compositions

[twin bluff]

it works, not every time, but every 2nd or 3rd it makes full composition

[autumn forum]

I’ve found the dimension rank is what leads to large file size higher the dim higher the files size. Dim 8 leads to 40mb file haven’t tested if higher ones were better yet tho

[trim orbit]

it's a helper is all. i'm trying more to describe the frame of thought for building prompts.

[twin bluff]

well thank you, it's a good starting point

[hardy cipher]

would a prompt be able to remove artifacts? I actually don't know

[true wing]

I'm not sure what's caused this but for some reason it looks like SD pasted a new image of a shark over a completely different image

It's following the prompt, it's just a weird render I should say

[trim orbit]

naw thats just one of those there bioluminscent air sharks

[hardy cipher]

I got one yesterday where it literally pasted different images in? not sure how to describe it. let me see if I can find it

[true wing]

I should mention part of the prompt was trying to turn the shark into glass but... that's not what that would look like 😂

[hardy cipher]

no, just a weird workflow. unclip models and whatnot

[true wing]

I think I was just using one of the two ComfyUI stock workflows with minimal changes. Although I haven't run into it with Sytan's so maybe there's something to that

[hardy cipher]

I left it to run quite a few while I was gone. only one that did that

[true wing]

.... what?

[trim orbit]

those are beautiful artifacts

[cinder aspen]

No way

[hardy cipher]

unclip models are very interesting. I still don't know how they work. made some interesting stuff though

[cinder aspen]

@trim orbit i agree, i mean SD, MJ and Adobe with FF have different audience target

[hardy cipher]

like what sort of fever dream did this come from?

[cinder aspen]

Gotta test more with SDXL on friday

[hardy cipher]

might be the same woman there

[stray wave]

What is the main advantage of sdxl?

[hardy cipher]

gets better results with less finetuning?

[cinder aspen]

#📣｜announcements message

[trim orbit]

resolution

[hardy cipher]

with all the loras and tools with 1.5 I can make things that can compare with what a lot of sdxl does. but with just the base model I've been exceeding the quality of about anything I did previously. basically a much higher percentage of results are wins

[ionic gulch]

https://i.imgur.com/Ik8OqeC.jpg (SDXL, no refiner, just finetuned upscaling)

[trim orbit]

resolution improvements always kick ass in any it field

[hardy cipher]

indeed. also takes soooo much longer to render. but it's all a tradeoff, lol