✨｜sdxl

[rustic garnet]

vintage anime by andreasc75

it's such a great lora 🤩

block weight lora means that you specify weights for each layer independently instead of using only one weight for the complete model

the middle layer is for composition the later output layer determine style. So if you want the lora only for style but not for composition you set the weight for the middle layer to 0

[vale vapor]

pls where can i find this lora?

[vale vapor]

wow i didnt know this, how exactly do you this? is this in a111 or comfy

[wet nacelle]

Can you explain how to actually use this? I understand what it can do butnot so much how I myself can get it to do anything.

[rustic garnet]

in comfyui you can use inspire-pack for that. In auto111 I don't know but it should have something like that, too

[rustic garnet]

he removed his loras from civitai 😦 maybe ask him here on discord

[buoyant crow]

Please translate my note using GPT or something similar.
This is everything I have at the moment.
I also have the workflow for ComfyUI available.

[vale vapor]

yo plz check your dm

[kindred shoal]

Is it possible to create just the closeup of an eye using sdxl and/or pony?

[dense sequoia]

Anyone tries to reproduce MJ showcase gens using SDXL?

just for fun to see what comes out

like this one for example https://www.midjourney.com/jobs/fa8de2a0-0106-4c3b-9e0b-4b175b4c3db0?index=0

that's mine

[copper kraken]

try typing closeup photo of an eye lol

[dense sequoia]

uhd macro photography

[kindred shoal]

i figured, but i wasn't sure if it'd be that easy or not.

[copper kraken]

your eye has an eye lol

[dense sequoia]

do a goat eye

[worthy orbit]

I used Surreal Lora

[kindred shoal]

man, pony really likes adding extra fingers.

it makes the hands look good, except for the amount of fingers

ahhh there we go. just had to fiddle with the seed.

[worthy orbit]

Shamouna

Ahaham hummm oh yeah

bang

Mr T (!?)

[terse tundra]

love the style! What is the lora used for those?

I would also be interested in that lora, where can it be found?

[tawdry current]

why do all these have the face of a child?

shits kinda creepy man

[terse tundra]

yeah, it was the lora I used... removed it again from the workflow

[cyan crown]

I lora I made.

[tawdry current]

@rustic garnet since we were talking about style loras yesterday and coincidentally the IPadapter creator also talked about lora blocks for things like style transfer in their latest YouTube video, have you seen or tried out https://b-lora.github.io/B-LoRA/

Apparently you can pretty much get full style even out of a single image with a very tiny amount of training. You'd have to use the block weight node in comfy to use it right, but I wondered if you had tried this out yet

If I get some time tonight, I'm going to test it out

[rustic garnet]

it's the same

b-lora was a paper about exactly that: train a lora, then separate middle layers and output layers

[tawdry current]

From what they say in the page, it actually does it even better with even less chance for concept bleeding

And also, you don't waste time training everything, only the two blocks I guess

[rustic garnet]

I think middle layer + next two output layers determine composition, the layer afterwards determine style. But would have to check

[rustic garnet]

you should train both, style and composition layer, even if you remove on of them afterwards

[tawdry current]

Would be hella sweet to make style loras in a minutes vs hours though

[tawdry current]

I'll read the rest of the paper on it to see if there are any differences vs the way you're describing

[rustic garnet]

I think most loras are unnecessarily large anyways

[tawdry current]

Yeah I finally got dora training down pat in koyha. With even an 8dim network, it works great and the files are tiny

Like megabytes

correction, 100mb, but i cant remember if this one was 8 or 16 dim

[rustic garnet]

dora should be same size as lora

[tawdry current]

it is, but you can use faaaaar lower dim with doras to get equal or better quality vs a regular lora. like an 8 dim dora can be better than even 64 dim lora, so the files can be much smaller

[stray warren]

Mobius ❤️

No upscale, raw output

[terse tundra]

Is it available for download somewhere?

[native moon]

it feels like older bing images wehre more realistic (bing)

[marble swift]

a girl

[worthy orbit]

I really liked mobius, but also I've trying this model, also uses clip skip 3 and according to description it is specially trained for using DPM++ 3M SDE. https://civitai.com/models/448483/mr-xotox-nasse-waende-sdxl

[misty hull]

This is a one pass generation?

[worthy orbit]

yes!

40 steps, CFG 6

[misty hull]

Wow, there is a great amount of details and very coherent as well

[copper kraken]

i don't think there's really any way to train for a certain sampler or scheduler

[kindred shoal]

he got that army guy hand

where you slot the toy gun into it

[worthy orbit]

Yes I have no idea about, I took it with with a grain of salt. Sometimes it is good to be biased and believe in a model

[copper kraken]

so you got it lol

nice

[kindred shoal]

only took an ungodly amount of inpainting attemps, hah.

[copper kraken]

i bet

[gloomy lark]

prompt: a tornado of library books, weird supernatural ghosts flying around sliming the library

[peak dove]

How do you run Mobius in ComfyUI? 🙂

[worthy orbit]

I used in Automatic but I also used Mobius in Comfy, I just reproduced the recommend settings there. It generated different but good IRC

Used a node to use clip skip

[peak dove]

I need a PNG with metadata or a json?

[peak dove]

I don't have a clue how to reproduce the recommended settings otherwise?

[nocturne dove]

I was doing some ipadapter style transfer with some comic book/ink drawing images and I started getting this style image instead for some reason. It doesn't look much like the input to the ipadapter but it's an interesting style anyway

[peak dove]

PiXart-Sigma+SD15+SDXL+PAG+Face Detailer

[mellow tendon]

That is an outstandingly good model, I thought he had taken it down as the link I had was dead, thanks for sharing.
I

[worthy orbit]

embeded is the workflow as I understand it to reproduce the settings

you only have to change the model

[meager canopy]

Isn't it loaded just like any orther model?

[stray warren]

Thank you, I'll give it a try

[meager canopy]

Mobius

[misty hull]

Insane man. 🔥👏

[misty hull]

Can you get sharp results with mobius or you use another model for second pass? @meager canopy

[meager canopy]

The ones above used the same Mobius model for Ultimate SD Upscale then face and hand detailer. This is RAW Mobius, using the recommended settings

This is with the other refinements

50 steps are too many. I find 30 are plenty.

Same prompt with Mobius...

...and Xotox

[misty hull]

Ah, when i tested it, I noticed that the upscale and img2img results were producing some noisy and unusual textures, particularly on the hair compared to the usual model I use. I'll give it another attempt

[misty hull]

That's really good 😍

[misty hull]

I gave it a shot, and it is a really really good model. Very sharp images for realistic

[meager canopy]

This is how it looks using my own model. The raw output from mine is much cleaner

https://civitai.com/models/240590/gtm-ultimateblendxl

This is raw output

Upscaled

[misty hull]

wow, it looks realllly nice 😍 . Sharp and detailed. Fingers, skin texture and hairs 👌. I didn't know you had models on civit. I'll definitly give it a try

[keen bobcat]

the creature

generating weird stuff like this is fun

[slim sandal]

guys, what do you think about Align Your Steps (AYS) at 10 steps vs hyper lora, or even both? I did try to combine hyper with AYS, but don't know if it worth it as I can't really see noticeable improvements while loosing ability to go lower 10 steps

[gloomy lark]

I use hyper 8 step Lora with 12 steps, so ays works well with that

[dense token]

create a girl with red hair and blue eyes

[meager canopy]

Here is the image you requested

[slim sandal]

maybe will stick with AYS too...
Managed to heavily speed up SDXL from 20s on rtx3060.
Hyper + 10step + TensorRT + AYS + cfg 1.5 = 6 seconds
Hyper + 8step + TensorRT + cfg 1 = 3.3 seconds

TensorRT is really gamechanger if you don't need loras

Comfyui does not support tensorrt loras but you can bake into model while conversion, as I did with hyper lora

[gloomy lark]

Interesting. Didn't realize it would work with both. I tried installing tensorrt and it always fails to import. I'll have to try it on a more clean machine.

[tawdry current]

What's weird is I had it working fine the other day, but then today I had to completely reinstall comfyui(broke it working with something else) and now tensorrt won't work for me

I guess I probably did some package stuff with cuda for a different plugin in the past and that it let it work, but I don't remember what it was lol

It's probably some pytorch+cuda thing I need to rerun where you manually pick the version off the pytorch site

Like pytorch2.3.1+cuda 11.8 or 12.1. I'll test it out in a bit to see

[paper parrot]

Pix-Art Sigma with sdxl refiner. ComfyUI Workflow in CivitAI

[peak dove]

Not Mobius ... 🙂

[peak dove]

I double-click ComfyUI desktop - but no Mobius features appear ... ?!

Mobius is in my ComfyUI Custom Nodes folder - but does not appear in the Nodes Menu ... ?

[meager canopy]

Why do you need nodes? It's just a model.

[peak dove]

OK, so where do I initiate it?

As I say, I cannot decide which nodes/tensors/loras etc to use?

[meager canopy]

The same as every other SD model 🤷🏻‍♂️

[peak dove]

So I put the Mobius folder into the checkpoints folder?

[meager canopy]

Yes

folder?

It's a model

This? https://civitai.com/models/490622/mobius

[peak dove]

Yes I have a folder see the contents ...

[meager canopy]

Forget that, you only need the model from Civitai

[peak dove]

OK

So that Mobius folder is another implementation?

[meager canopy]

You have them both n that folder, just rename them and delete the rest

[meager canopy]

yes

[peak dove]

Rename them to what? 🙂 Sorry for all the questions

[meager canopy]

"Mobius", or "Mobius-fp16" 🤣

Whatever you want to keep and use, then move it to models/checkpoints folder

[peak dove]

So the Civitai checkpoint is this Mobius safetensor I already have?

[meager canopy]

https://tenor.com/view/jack-nicholson-yes-nod-agree-gif-11928667053720670991

[peak dove]

Hurrah! Boy its confusing as an artist "with some tech savvy" some™

[meager canopy]

I think you were just trying to overcomplicate it 🙂

[peak dove]

Mobius - OK, nice, no LoRAs

Mobius+2xLoRAs = Ethereal Grace and Poetic Style (two of Andreac's LoRAs)

Ouch - it has some acute visual acuity alright!

[safe stump]

Is there any open source image generation models other than sdxl 1.0 and pixart

[peak dove]

Thank you for the prompt!

Very rich colours and clear lines ... me likey Mobius 😄

It's getting on for MJ standard imho ... is Mobius

[paper parrot]

Pixart sigma + sdxl refiner

[peak dove]

Mobius

[peak dove]

Mobius+LoRAs+Refiner

Mobius+LoRAs+Refiner

Mobius+LoRAs+Refiner (ComfyUI)

[safe stump]

Anyone here using Hunyuan dit if so can you please share some examples

[gloomy lark]

@safe stump So hunyuan like the other ones needs an sdxl refinement stage in comfy, so these are all hunyuan with refinement.

[peak dove]

Mobius doesn't really work with Face Detailer - the faces seem to come off worse than when left alone!

[tawdry current]

I'm working on getting lumina reinstalled, but i've been stuck building the flash attention for like an hour lol (it takes a long time even with ninja). i want to try out the new finetune of it they dropped yesterday

[misty hull]

I'm more scared of his brother in fire flying in the background 😄

[glass forge]

So is Mobious the new hot checkpoint?

[gloomy lark]

really depends. it's good as a native sdxl model. Most of my stuff is rendered with pixart/hunyuan/ella etc and i need a good sdxl model to refine that. I'm mainly using zavychroma07 at this point for that. I found mobius is rather soft and too stylized for it whereas zavy gives sharp results.

zavychroma on left, mobius

on right

mobius is kind of a mess for refining/upscaling model.

[misty hull]

I'm using mobius to generate base genation since a couple of days. It's really creative. I have a style lists node and it's reacting better to it than any other model

[misty hull]

Yes... I gave up. It's seem to even make the image worst in certain spot 😦

[gloomy lark]

this is just straight mobius, no sd3 and then upscale. even direct it's kind of a mess. it's a cool picture, but it's "too" stylized... much like midjourney does to cover for it's inabilities.

[misty hull]

The composition is still good. when passed to another model, it will be washed out anyway, only keeping the composition just like you do with ella

But most of your images has that MJ "styling" 🤔

[gloomy lark]

totally, but i'm not refining sdxl with sdxl. 🙂

[misty hull]

Why not?

[gloomy lark]

the composition from the other new T5 based models are better than mobius alone. and way more prompt following.

for me, an sdxl model is useful as a refiner or not at all at this point.

mobius/realvis/zavychroma

i'll try upscaling mobius more. maybe it only shines after a few stages of it.

[misty hull]

Mmmm, I have to say mobius is crazy too. You can give him long crazy prompt, and he'll come up with extremely wild and surprising results.

[gloomy lark]

it's really more about the limitations of sdxl. I have high hops for him with sd3.

hunyuan

hunyuan and pixart have really cool composition

[misty hull]

Do you like it so far?

[gloomy lark]

yeah hunyuan is awesome

that said, sd3 is awesome too though. if sd3 weren't coming out in 2 days, I'd say you should try it out.

sd3 makes a really good version of this prompt as well.

[misty hull]

what's the prompt? I'll give it a try with mobius/mhxlAventisHorizon

[gloomy lark]

A post-apocalyptic video game art style standoff where survivors and mutated beasts are taking place against the backdrop of a ruined cityscape with crumbling skyscrapers. In front there's a gigantic mutant with multiple arms roaring in rage. High resolution, sharp focus, studio photo, intricate details, high quality, game cover

[misty hull]

very true. If sd3 is the last model they release, these types of models could remains highly valuable in a close futur

[gloomy lark]

this is sd3 refined with zavy. it's awesome

I may not use anything other than sd3 after it releases... it's got really good composition and I can always just refine it like I'm doing with these other ones if need be.

[gloomy lark]

what is mhxlAventisHorizon ?

[misty hull]

you are used too with the api version. You think the smaller version will still kick other model's ass?

[misty hull]

a model i really enjoy on civit. Give it a try man

[gloomy lark]

hah it really whips the llama's ass... that old winamp phrase, but now it's "it really whips the llama3's ass"

I searched for that phrase but didn't find it. what's the name I should search for on civit?

[misty hull]

https://civitai.com/models/331672/mhxl-aventis-horizon-sdxl

[misty hull]

holy shit, i can't wait to push the button to download the weight

https://tenor.com/view/happy-gif-27563888

[gloomy lark]

https://civitai.com/models/502097?modelVersionId=558077

whoa

holy crap

a real serious pixart sigma checkpoint.

[misty hull]

😱

[gloomy lark]

yeah this could be serious...

pixart is seriously good... better than the others, just needs training. this may actually make it a real thing now.

[misty hull]

That's so true

In addition with his prompt adherence and composition, it art capabilities are totally insane (drawing, painting, etc...)

[gloomy lark]

pixart / new pixelwave checkpoint.

he may need to put it back in the oven for longer.

😦

[misty hull]

there is a tiny difference with the demo images 😄

We only see the quality of the image when we try it ourself haha

[gloomy lark]

he has a workflow that i'm trying now. it's rather different than what I'm using for pixart stage

[misty hull]

I'm curious. I'll take a look

[gloomy lark]

so this is with what?

[misty hull]

there is not 30 ways to run it in comfy 😅

mobius and mhxlAventisHorizon

[gloomy lark]

so i tweaked his settings. he has a cfg of 4, but it looks like garbage with his settings at 4. 7.5 is optimal.

that's still refined, but obviously looks really good. I'll compare against native pixart.

[misty hull]

hard to see when it's refined.

[gloomy lark]

it still looks like butt unrefined. hardly something you could use without refinement.

not sure what the point is if i still have to refine it.

that's raw

[misty hull]

Yeah, it veed some more work obviously.

refined with mrXONOTOXNaSSEWAENDE_v20. it's better

[misty hull]

hahah it's totally out of it's comfort zone 😄

[gloomy lark]

zavy chroma / aventhorizon

avent might be better with this one... it's more choice of content vs. quality. they're both good.

[misty hull]

aventhorizon is extremely good for 3d render, and that kind of stuff, so we can see it tend more on the side

[misty hull]

That's the key 🙂

[gloomy lark]

zavychroma / aventhorizon.

avent on this one too... looks more like will smith plus it put those evil little creatures which is a big plus

and it's ultra sharp too..

[misty hull]

Same for me. I saw those creature dancing and 😍 🤣

that's a great image by the way. Sick compo 👌

[gloomy lark]

In grotesque, hyper-realistic style with intense texture detail and unsettling close-ups, Will Smith stands amidst a swirling vortex of viscous, gelatinous spaghetti tendrils that writhe around his body like snakes, hundreds of grasping, slimy-fleshed children clinging to his arms, legs, and torso as if he were some sort of twisted, otherworldly host, their eyes glowing with an unearthly green light, while god rays pierce the dark, smoggy sky above, casting an eerie glow on the scene, smoke and steam wafting from Will's mouth and nostrils like a noxious cloud, his own eyes blazing with an unsettling, pulsing yellow light that seems to sear itself into the viewer's retina.

it's a great prompt. ellla makes some seriously insanely evil versions of it.

this is probably the tamest version. the rest i can't post

[misty hull]

sick ❤️ we are far from the famous first spaghetti version of will smith

[gloomy lark]

from pixart 512 res refined.

so far i'm very unimpressed with pixelwave for pixat. native pixart is better at composition

same seed on pixelwave version.

the native pixart version is so much more interesting.

see this is the problem.

these new models are capable of so much more. if people sit there and train sd3 on boring portraits again, it's going to actually make it worse than base model.

[misty hull]

Totally

I don't enjoy portrait so much, so I'm freaking tired of seing the same type of models and image over and over .... Hopefully, some will train these model to keep their originality

I don't mind getting these sharp images, but try to keep these great brush stroke so unique that only pixart can generate (for now :-P)

[subtle swift]

#1237459938901491852 a fat man 40 years drinking beer funny picture

[gloomy lark]

ok fine mobius refined with aventhorizon is pretty neat. 🙂

[misty hull]

😻 That's really sharp

[gloomy lark]

yeah this is some really good output for an sdxl only mixture.

[kindred shoal]

that is fucking gorgeous.

[gloomy lark]

the workflow should be in that picture.

[misty hull]

That' different then ella. I think Sdxl can sneak more content as the base image. So the second pass model will define which style direction to take. I usually prompt a lot for the second pass to enhance details, add stuff

[kindred shoal]

it saves as a .webp, so i dont think the workflow will be there

[gloomy lark]

click it and then click open in a browser, that should be the native png

with workflow

this is the girl with the skull: Gothic cyberpunk woman robot, holding the skull with both hands,face the skull, profile , realistic illustration, sad and dark mood, red japan circle in background

[misty hull]

That's the prompt from the other day?

[gloomy lark]

here's my workflow

[kindred shoal]

oh, i had my discord window too small, hah

[gloomy lark]

the samurai butterfly one is : close-up of a a Chinese samurai sitting meditating eyes closed on his kimono. Amidst a dreamlike landscape of Organic Symmetry, clad in armor made of iridescent, butterfly-wing tiles and branches tangled with glowing, ethereal silk threads that weave a symbolic tapestry of balance and unity shimmering mist of Cultural Fusion infuses the air, blending the essence of ancient traditions with the serenity of the natural world.

yeah it's not a new prompt.

[kindred shoal]

oh dear, i see how this is going to be a complex venture...

[misty hull]

It was so similar to the image the other day 😄

[kindred shoal]

i had to install a few custom nodes i was missing.

[gloomy lark]

yep that's the one.

so that was ella I think

[misty hull]

I did it with my workflow and you tried it with ella

[kindred shoal]

I think I might have to stick with my simpler workflows 😅

[gloomy lark]

man i'm really loving this. it's doing incredible work with all my saved prompts.

[restive ice]

yes try it you will see improvement

[gloomy lark]

man, I've gotta say, I've just put it through tons of prompts. it's right there with pixart and ella on most of it.

kind of in disbelief right now

[misty hull]

Also, I've been using the ella you provided a while ago, and it's a fresh air to use a sdxl as base model. It outputs different type of image

[gloomy lark]

yeah definitely... although this is crazy good, with the 2. where mobius outputs kind of a mess, avent cleans it up based on the good composition from mobius.

[misty hull]

yesss. It's basicly the same principle then ella

It might be a bit more gpu costly.

[gloomy lark]

i can't seem to stump this thing

35mm photo of a disney princess setting a car on fire with a torch. found footage.

[misty hull]

what do you mean?

[gloomy lark]

it's insanely prompt following.

ok so i can stump it on multi-subject. happy boy next to a sad girl... but aside from that, it's getting everything

[misty hull]

i know man. I realized that when I was adding styles using this node with mobius. Even with very long prompt, it get the style like a chief. With other models, even when small prompt it sometime don't get the style

here is multiple image with the same prompt, i only switched styles

[gloomy lark]

definitely. 99% of the time you say "oil painting of..." and it's not actually doing that.

[gloomy lark]

wow

[misty hull]

exacly

that's freaking crazy how good it is

the style is appended at the end of the prompt, and it still get it right

[gloomy lark]

In a darkroom-lit, high-contrast, with a mix of warm and cool tones, a bespectacled man sits amidst a sprawling mess of glowing computer screens, wires, and keyboards, his face illuminated only by the soft blue glow of a few screens, as he frantically attempts to corral his army of tiny cat minions, each donning miniature spectacles and typing away on their own miniature laptops, surrounded by scattered papers, coffee cups, and discarded pizza boxes, with a haze of cigarette smoke wafting through the air, casting an eerie mist around the chaos, while in the background, rows of humming servers hum their gentle tune, and the distant glow of city lights casts a faint orange hue through the grimy windows.

so it has its limits.. but this is more of a mult-subject thing.

there's no army of cats, but if you do that in pixart or ella, it gets the cats.

so his stuff but trained on sd3, would be something really special.

[misty hull]

Does it gets it if you reprompt or it's just too much?

[gloomy lark]

that said it's nailing these every time. awesome concepts it gets all the time.

as long as it's not 2 different subjects.

i'd say this is peak sdxl.

[misty hull]

Hahha it's pushing the limit.

[gloomy lark]

mobius / hunyuan / pixart

so it's really great, but multi-subject beats all a lot of the time.

Trump wearing greek god armor with his hands in the air while donuts are flooding from a chute in the ceiling. Adoring crowds are so happy about the donuts. god rays, faces in shock,

hah although no armor in the so called multi-subject ones. oh well.

[misty hull]

you probably have to play a bit more with prompt weight with sdxl and you'll get there

[dense sequoia]

Did you add "googly eyes" to the prompt?

[misty hull]

changed the prompt as well: Trump wearing greek god armor with his hands in the air while (donuts are falling from the ceiling:1.15). Adoring crowds are so happy about the donuts. god rays, faces in shock,

[gloomy lark]

Surrealism; surreal black-and-white image featuring a person with their back to the camera, their dark skin dotted with numerous googly eyes of various sizes, contrasted by white patches, and holding a white cup also adorned with googly eyes and a straw, set against a stark white background under soft, diffused lighting, and a table covered in scattered googly eyes.

I just know you're gonna have fun with that one. 🙂

[tawdry current]

Finally managed to get hunyuan up and running. It passed my "red ball on top of a blue cube on top of a green table" test on the first try

[dense sequoia]

i tried my other prompts but with "googly eyes" for fun

[gloomy lark]

with your prompt... seriously adding perturbed node does insane things with regular sdxl. i stopped using it because it doesn't work with sdxl... but dropping hyper, it does so much for image quality

but no crowds.

hah ok fine next seed, crowds.

[gloomy lark]

man that looks incredible.

[tawdry current]

God that was a conda nightmare, but I got it running with diffusers

[misty hull]

What do you mean for hyper?

[gloomy lark]

hyper-sd

about a 2x speed boost.

[gloomy lark]

it works flawlessly with zavychroma7, but doesn't work at all with mobius or this adventhorizon. the output is just blurry. for whatever reason, zavychroma keeps its sharpness with it.

but then you went ahead and showed me better checkpoints, so i'm back to half speed for the increase in quality.

[misty hull]

ohhh gotcha. You don't lose quality? I haven't experiment with it

[gloomy lark]

with just that one checkpoint.. so it's finicky.

i assume it uses a reduced set of blocks in the unet to do its thing, so if your model isn't heavy on that block, it's going to come out bad.

[misty hull]

ok. Ill try it with chroma. THanks

The same apply with deepshrink. With some model, it's an absolute killer and can keep the composition together. But with some, it's an absolute mess

[gloomy lark]

definitely.

[misty hull]

Do you use hyper with upscaling?

[gloomy lark]

yep

[misty hull]

damn, it must be brutal if it works with zavychroma

[gloomy lark]

upressed image in 7 seconds

so 1700x1200 res image in 7 seconds.

and the output is sharp

[dense sequoia]

how about using "Trypophobia"

[gloomy lark]

they also have it for sd 1.5. so my ella workflow goes through tons of upscaling stages super quick.

[misty hull]

Here if you want to add the styles. The default had an hundreds, but someone in here shared a google link with a list of couple hundreds of styles. I added 100 more. Drop that in this folder: custom_nodes\ComfyUI-Easy-Use\styles\

[misty hull]

oh sweeeeet thanks man 🙌

[gloomy lark]

i put that in and got his.

[mystic prairie]

Does anyone know the best way to make D&D Battlemaps using SDXL?

[gloomy lark]

A group of adventurers huddled around a large wooden table lit by flickering candles surrounded by dusty tomes and scattered dice, their faces illuminated by the soft glow as they gazed down upon a sprawling, intricately drawn battlemap sprawled across its surface.

[misty hull]

and drop that in custom_nodes\ComfyUI-Easy-Use\styles\samples
They were generated using pixart 1024 basemodel

[gloomy lark]

very cool, thanks. that demo you had before was really neat.

what style was this one?

jpeg says fluid

[mystic prairie]

Nice image! But i am looking for like an top-down map to use in an encounter for my campaign. Is there any model that can do that?

[misty hull]

that's the 3D___Abstract_And_Fluid_Dynamics

[gloomy lark]

that i don't know.

[gloomy lark]

have you looked on civitai.com ?

[misty hull]

You added a guider 😮 What is it doing?

[mystic prairie]

yes, but all of them are on sd1.5 and look kind of weird

[gloomy lark]

the benefit of the guider is that it perfectly sets up the CFG which is rather crucial for hyper to work best.

[gloomy lark]

mobius keeps making great stuff.... based on a describe of yours

[dense sequoia]

what is mobius?

[gloomy lark]

https://twitter.com/Corcel_X/status/1796657183850905903

Corcel.io (@Corcel_X) on X

Big News! We're so excited to unveil Corcel Research Labs' two new image models: Open Vision & Mobius! ✨

Both fully open sourced under the apache-2.0 licence!

Check them out here:
https://t.co/xbcmWLCacy
https://t.co/wo9x4ocCiV

💖Open Vision: Finetuned from S19's MJ

Twitter · May 31, 2024, 9:37 PM

[misty hull]

damn it's really fast. a lot faster than my setup

[gloomy lark]

mobius is a new sdxl model that has really amazing prompt adherence and style/concept knowledge. but it can be a bit messy looking sometimes, so KevenG just got me going with a aventishorizon sdxl model refiner which puts out incredible quality in the end with the composition of mobius.

[gloomy lark]

yeah, it breaks my heart that it doesn't work with mobius or aventis. it literally does, but the quality is garbage.

[misty hull]

😦 Did you try with the clip skip -3?

[gloomy lark]

I don't think so.

no

i don't have that for my current workflow either...

do i need it?

i'm getting killer outputs without it.

[misty hull]

it's recommended for it yes. You'll get better result with genaration

i'm not sure it will helps haha 😄

[gloomy lark]

so the above one was without it, this is with -3

i assume you mean like this

@misty hull ok that explains why you were getting the crowds and I wasn't i guess. 🙂

jeez... crowds every time now whereas almost never before.

[misty hull]

that's crazy 😱

[gloomy lark]

i love that his armor has his own face on there... that's so "him"

[misty hull]

probably. I'm not sure what's it's doing under the hood, but you'll get some bad composition, like too zoomed close-up without it.

[gloomy lark]

A grotesque Trump mask made from human skulls and twisted metal hangs upside down from a rusty chain suspended above a trash-strewn alleyway illuminated by flickering fluorescent lights casting eerie shadows on crumbling brick walls.

never let anyone say that llama v3 is censored.

it just came up with that.

[misty hull]

😅 . That's creapy

caricature

And the funny thing, i went into the hirefix from the caricature, but forgot that I had an image loaded instead of the base image. So it turned out...

[gloomy lark]

haha that's awesome

[misty hull]

that was the image haha

[gloomy lark]

mobius vs sd3

[dense sequoia]

mid level boss

[dense sequoia]

it's different but is one better than the other?

[gloomy lark]

A cartoonish orc in a bright yellow hazmat suit holding a gigantic rubber ducky while standing amidst a backdrop of mushroom clouds and nuclear fallout signs amidst a desert landscape under warm golden light.

the sd3 is way more prompt following

and mobius made the mushroom clouds into literal mushrooms

that said, mobius makes awesome orcs

[dense sequoia]

they both did

[gloomy lark]

FINE. 🙂

less so.

better duck.

although I do like the NUCLAR.. that's fitting for what an orc would have on the name

[misty hull]

Both are crazy. These giant mush on the right image look so nice. Nuclear mushroom 🍄 😋

[gloomy lark]

definitely

he's doing the no look walk away from the explosion

[edgy wave]

Devil leprechaun

[peak dove]

Here is my Mobius+SDXLRefiner+FaceDetailer json

[peak dove]

Mobius+SDXLRefiner+FaceDetailer

[peak dove]

Mobius+SDXLRefiner+FaceDetailer

Mobius+SDXLRefiner+FaceDetailer

[peak dove]

ComfyUI+Mobius - Kardashian and cookies!

[twilit loom]

Hello everyone, can you suggest me a good model for open pose ?

[kindred shoal]

thibaud_xl has been good to me

it's for sdxl

[midnight valley]

A young Chinese male bodybuilder with ultra short hair, wearing only a tight pair of shorts, well-developed chest and thigh muscles, tight calf muscles, bare feet, solid soles, walking upright on stage with steady steps. Real photos show full body photos of the feet.

[meager canopy]

Here is the image you requested

[stray warren]

@gloomy lark Was it you who thinks you can't use Mobius with hyper?
These are with hyper, 8 steps for Euler A Trailing, and TCD, 4 asteps for DPM++ SDE Trailing (in Draw Things, I think it's called sgn or something in comfyui)

First two TCD, second row DPM++ SDE Trailing (hyper 8 at 75% strength), last one os Euler A Trailing

Borrowed the prompt from @meager canopy

[gloomy lark]

Very cool, I can't look at it now, but mind dropping your json on here so I can try it later?

[stray warren]

I'm not using comfyui, it's draw things @gloomy lark

[misty hull]

What is drawthings?

[stray warren]

mac app, you can use it on mac book, iphone, ipad, imac

local app

[misty hull]

I'm not getting sharp image like these ones. Is it str8 out of the model or its refined ?

[stray warren]

Just raw output

https://drawthings.ai/

[misty hull]

👍. Really impressive. Does it generally output images as sharp as that?

[stray warren]

Depending on the image you need to tweak the settings

There are alot of 'buttons' you can tweak

[misty hull]

Oki. I'm wondering if it's just because these image are in the training data

Can you give a shot at this prompt when you'll have a sec?
A post-apocalyptic video game art style standoff where survivors and mutated beasts are taking place against the backdrop of a ruined cityscape with crumbling skyscrapers. In front there's a gigantic mutant with multiple arms roaring in rage. High resolution, sharp focus, studio photo, intricate details, high quality, game cover

[stray warren]

Same settings as before, so no special tweaking for this one

[misty hull]

Interesting. That's what @gloomy lark got from comfy with the same prompt
#✨｜sdxl message

[stray warren]

Looks like he's not using hyper

[misty hull]

Yours are better

He's not. But he should get similar results, or close

[stray warren]

Not really, results differ a lot depending on samplers

And methods like the fast loras

Eeven within one and the same sampler like TCD you can get various results just by changing the strategic stocastic sampling percentage

And then there is the sharpness and clip skip

Also the prompt

[misty hull]

Gotcha. 👍 need more testing 😫

[stray warren]

Yeah needs a lot of testing, also cfg

[misty hull]

Do Yyou know if we can tweak that in comfy?

[stray warren]

I haven't used it for months now, so I think you should ask that the regular users

[tawdry current]

Anything is tweakable in comfy if you're brave enough to do some python scripting lol

[misty hull]

Defjnitly. Mobius don't react like other models with cfg

[rustic garnet]

just dreamshaper turbo without anything special

[misty hull]

😁🤣

[tawdry current]

sometimes i luck out and actually do something that works, other times, not so lucky lol...

[stray warren]

I have several models which can do that prompt with ease 🙂

[peak dove]

This Mobius is doing a lot of left-arm/right hand - right leg/left foot!

And it is not particularly good linking the Mobius Model to Face Detailer ... do it via an SDXL intermediary setup.

[tawdry current]

i tried mobius out for a couple days, and while it can make some decent stuff, it's kinda meh. realvis4 stomps it most of the time

and i don't even make people 99% of the time

[stray warren]

My point was that it can do hyper

[peak dove]

Must try realVis4 ...

[tawdry current]

https://civitai.com/models/139562?modelVersionId=344487 for the non-lightning model

[tawdry current]

the thing is that hyper is a lora and maps to the base model. these fine tunes have stuff added in them usually, so the lora probably won't work well with a lot of them

i remember trying it with some models and it was hit and miss

usually works well enough though

[stray warren]

I'm using hyper for all models, it works well with TCD

[tawdry current]

yeah tcd works great for it, but so does AYS

[stray warren]

agree

[tawdry current]

like if you're using the 12 step hyper lora

[rustic garnet]

honestly, if a model doesn't work well with other loras this is always a good indication that the model is rather bad

(yes, this includes pony xD)

[tawdry current]

why? it just means the weights are different and that the lora isn't mapping to all the weights it expects. like you can't use regular loras for pg2.5, yet the model is great for a lot of things

you can train pg2.5 loras though

but then they cant be used for sdxl models

[rustic garnet]

the sdxl base model was trained on billions of images

some custom models overfit so heavily that they destroy the base weights and then yes, their shiny new custom model excels at everything close to the few training images they provided but forgot everything else

[tawdry current]

yeah but almost all custom models are just merges with the base model or other fine-tunes. only a few are actually trained from the ground up

personally, i hate overtrained models since they are everywhere and typically only good for portraits of waifus, but you'd be surprised how well some of them can still do completely unrelated stuff

[rustic garnet]

I refer to the models that are fine-tuned. If a model is trained from scratch this is of course something different.

but not many have the resources for doing so

and quite often these models from scratch are trained on much less data with higher quality (usually midjourney or dalle images...)

they look great, but are also less versatile as consequence

[tawdry current]

well pony is one of them from what i hear and well you see how commonly used and praised it is

[rustic garnet]

are you sure pony is trained from scratch? I think it's a finetune

[stray warren]

you don't need to send me those, i have realvis for ages and a lot of other models

not pony models btw

[rustic garnet]

and yes it is praised, although its a horrible model in my opinion. But either people don't care how bad a model is as long as you can make porn with it, or maybe I just have a very different subjective opinion in this regard than other people here

[tawdry current]

yes, it is from scratch from what i've heard. the guys has a small a100 farm like 4-6 of them and spends months training them

[stray warren]

hear hear

[tawdry current]

just showing how poorly hyper does with some models

[rustic garnet]

there is even a pony-turbo merge - that wouldn't be possible if it is trained from scratch

[tawdry current]

oh im sure he has some merges, i just know he has his own training set

[stray warren]

hyper works well with my realvisxl

[tawdry current]

idk, never used it, but he was on here a week or so ago complaining about SD3's license

[rustic garnet]

"Pony Diffusion V6 is a versatile SDXL finetune " <-- this is even written in the description

yes, he might have a big training set, but it's still a finetune. he did not train the model from scratch

[tawdry current]

yeah i don't know the exact specifics, i know he has millions of images used and blah blah blah

all i know is that it's its own thing and normal stuff doesn't usually work with it. no clue what the merge ratio is vs the base model and all the technical stuff

[rustic garnet]

if you finetune your model in a similar way as the base model, things don't break that fast

many people do weird stuff, though. Like juggernaut was trained with pyramid noise making it highly incompatible with other loras

pony uses a totally different captioning scheme and who knows what else

[tawdry current]

yeah that's the big difference: his captioning. and like i said, who knows what else he does to the base model. maybe he squishes it down in some kind of 90:10 merge or something

i remember reading somewhere professional that technically, you only need around a 1000 "perfectly" curated and captioned images of humans, to have almost perfect anatomy and whatnot. so who knows what you can do with millions of images of even mediocre captioning and quality

[copper kraken]

i didn't know that... interesting

https://wandb.ai/johnowhitaker/multires_noise/reports/Multi-Resolution-Noise-for-Diffusion-Model-Training--VmlldzozNjYyOTU2 haven't read much on the effects of different types of noise in training, this is about it... but it certainly piqued my interest

[rustic garnet]

I'm not a big fan of all these noise models :/

[copper kraken]

why's that

or any particular ones

[rustic garnet]

fine-tuning the model on a new noise model is always harmful and, in my opinion, unnecessary.

like offset noise was introduced to solve the zero-timestep problem and allow models to generate dark images

[copper kraken]

gotcha, your concern is more that the base and the finetune were with different noise models?

[rustic garnet]

but instead of messing up training you could have just add a little slider in auto111/comfy that determines the brightness of the image, allowing the user to make dark images without using offset noise models

yes

[copper kraken]

yeah, that's what i've been doing

to get dark images, instead of a straight offset to the initial latent, i've been offsetting on a decaying schedule after each step within the sampling loop

better/more reliable results

just been sticking with the default offset for dora training

it seems like doing anything significantly differently than base was trained is just not a great idea

[rustic garnet]

yes, unfortunately sdxl was already trained with offset, so you have to use it, too

[copper kraken]

yeah

[grand cargo]

🌟 🎸 Well it's been a bit of time, but we have a new Top 10 SDXL model - Art Universe 3 - Excelled in photorealism but strong adherence as well. Worth checking out - https://docs.google.com/spreadsheets/d/1IYJw4Iv9M_vX507MPbdX4thhVYxOr6-IThbaRjdpVgM/edit?usp=sharing 🤩 🥂

[sharp scarab]

New system available ♥ - https://www.youtube.com/watch?v=-mTT49TDxIE

[minor pecan]

not to be crass but do they have down syndrome

[dense sequoia]

https://civitai.com/models/471609/chromosome

[minor pecan]

LOL

[peak dove]

It's nice to see a community that is inclusive of the more challenged in our society!

[peak dove]

Mobius+SDXL

[ebon remnant]

How is sdxl different from using sd 1.5 and upscaling it? Doesnt that acomplish nearly the same thing.

[lusty wolf]

SD 1.5: Includes basic capabilities for text-to-image generation, but might lack some of the advanced features found in later models.
SDXL: May include additional features such as better handling of complex prompts, improved coherence in generated images, and advanced control over image attributes (e.g., style, lighting, and composition). SDXL represents an evolution in Stable Diffusion technology, offering enhanced capabilities, better performance, and higher-quality outputs compared to SD 1.5.

[worthy orbit]

Hi, I'm looking for a SUPIR workflow that might work with my 12 GB GPU. I'm trying some but still no luck}

[lusty wolf]

I have tried several. It will kill your GPU. I have 12GB too... There are other ways. 'Ultimate Upscale'.' Mara Scott'. 'Pixel KSample Upscaler Provider'

[worthy orbit]

oh, bummers

[ebon remnant]

@lusty wolfthat does sound a bit more interesting if its not just having higher resolution by default and includes other things. which sort of makes it a new model.

[lusty wolf]

This what I'm doing here, and just because I hate wating I set it low. I think the quality is good on the eye..

[undone crest]

Made guide to my X profile how to create the Viral soft anime style
https://x.com/transparen96143/status/1800235665331798061?s=61

Transparent Art Process (@Transparen96143) on X

🧵 How to create cute soft anime style that got viral by @petravoice. Full guide bellow:

Twitter · Jun 10, 2024, 6:36 PM

[native knot]

It literally is a different model. SDXL is not just "SD 1.5 but bigger".

[ebon remnant]

ill get to test it in 1 to 2 days

@undone crestOhhh. twitter is no more? Just x now. BTW that is too soft for muh eyes.

[undone crest]

Twitter is no more? Thank you for reply! Maybe bit too soft, did my best as got requested from many how it was done :))))

[rustic garnet]

the quality difference between SDXL and SD 1.5 is huge. It's not like SD 1.5 and SD 2.1,

[tawdry current]

dig my crappy little ui for hunyuan-diffusers-distilled... (doesn't seem to be supported in comfy's extra models node yet, since they expect pth files and not safetensors? idk, every time i tried with this version, it seemed to spit a bunch of errors). i still need to add a couple things to it, but at least it's getting there lol

[native moon]

Ned a ds slider lora xD

[copper kraken]

https://github.com/ClownsharkBatwing/RES4LYF

will add workflows later

got some more code cleanup to do but it's in much better shape now

https://huggingface.co/ClownsharkBatwing/CSBW_Style/blob/main/csbw_v1-7_225tkn_r64a32_12epochs_base1-0-fp16_save-fp32_mixed-no.safetensors this is the dora i've been using

[vital ermine]

anyone here use supir?

[copper kraken]

sometimes

[vital ermine]

Does this look right because it is only a 2x upscale and takes 5-10m on my 4090.

[copper kraken]

yeah that's not right, is that the stage it's frozen at?

[vital ermine]

yeah, this is a wf I downloaded and man is this node so slow

about 1 step every 10-20 seconds

sometimes half a min

[copper kraken]

that's really weird

should be more like a min or so for the entire thing

[vital ermine]

yeah

[copper kraken]

and that's with 40 steps

[vital ermine]

reason I am wondering what setting is doing this in that node?

the wf came from here - https://comfyworkflows.com/workflows/b563ba19-8639-4c83-8f74-ac7c9cf96978

everything else, but it, in the WF is fast

[copper kraken]

what resolution are you going up to

2048x2048 or equiv?

[vital ermine]

I wish, as I just looked and it is a weird rez of 2250x2250.

[copper kraken]

i think that increases the number of tiles by something like 2.25x total

4 to 9 i think?

still doesn't account for THAT much time

[vital ermine]

1500x1500 then it seems mul 1.5

yep, 2250

[copper kraken]

weird resoluton

[vital ermine]

very

[wooden field]

XD

It was WAY EASIER 3 years ago, AND IT WAS FREE!!

[vital ermine]

changed to 1024 and much better

Much much faster

[wooden field]

I think it was much easier to generate beautiful images when you could do the promise for FREE, the options were much better before the video option, now I cannot make anything I want to

[fervent dock]

beautiful

[wooden field]

remember the CONTROLNET?

ÇA c'était la belle époque

[wooden field]

tu as payé stablediffusion pour ca?

[copper kraken]

nope

[wooden field]

je vais ou pour ne pas me faire arnaquer?

[copper kraken]

i don'nt speak french lol

[wooden field]

I mean, where you wrigt the prompt?

[copper kraken]

it's my own model on my own 4090

with nodes i coded

on comfyui

[wooden field]

I'm looking for image a woody autochtone Knife, simply

wood knife

[vital ermine]

man, supir is nasty

w/o and with

so much ugly noise in supir gen

[copper kraken]

Your s noise is really high

Try cutting that to 1.003

[vital ermine]

I tried even 1.0

[copper kraken]

Hm

I'll get you my settings in a couple hours, got a training run going

[vital ermine]

alright, but I have never managed to get supir right

[copper kraken]

Best I've gotten is subtle sharpening on edges with subtle softening between them

[vital ermine]

not sure why all the videos show SUPER results then I use their wf (for the ones that offered them) and rubbish results?

[vestal plume]

https://tenor.com/fKwAeF62Cs.gif

[peak dove]

Mobius+SDXL

[vital ermine]

restore cfg 10?

[meager canopy]

Anything up to 20

[vital ermine]

what does that even do?

mine was -1

[meager canopy]

You do need to play around with the settings depending on what image you're using

[meager canopy]

The model you use also makes a difference

[meager canopy]

5376x6769 (no SUPIR)

[misty hull]

That's sick 🔥

[wary ridge]

its about time someone painted something decent

[wary ridge]

pics or it didnt happen

[meager canopy]

There're literally pics all over the discord 😄

[shell flame]

1ère fois que je vois un fr ici

[gloomy lark]

Hah sd3 whoops

[misty hull]

Two things models don't like. Cards and poker chips 😰

[gloomy lark]

For non sd3 that's really good

[misty hull]

Yeah, I think I got lucky with the seeds. It took only 3 or 4 tines to get something decent 🤪

Did you saw zavychromaV8 released yesterday?

[gloomy lark]

I did, just didn't get a chance to try it yet. You spoiled me with that aventiswhatever from the other day. Now I'm using that. 🙂

[misty hull]

Thats kind of funny, I switched to zavy since I tried your upscale wf 🤣🤣

[gloomy lark]

Hah ask me next week and I'll probably say something else. Having too many models is a good problem to have. 🙂

[meager canopy]

Is that what you used for the poker image? That is great!

[meager canopy]

Was she playing against the "Cheating twins"?

[upbeat summit]

SD Ultra

[peak dove]

How do you setup/download Ultra for ComfyUI?

[upbeat summit]

You sadly can't. I used the paid API for this

but you can use the Stability AI API with the official ComfyUI nodes: https://github.com/Stability-AI/ComfyUI-SAI_API

[misty hull]

I used mobius to create the composition then mrXOTOXnasseWAENDE_v20 for double deepshrink to get details. Here is the base image genarated
Prompt:A beautiful menacing girl materializes within a bar, her presence distorting the air, as she summons a swirling storm of razor-sharp playing cards that orbit her like a deadly asteroid belt.

[misty hull]

If I can join, Yes!!! 😄

This batch was particulary good starting point

[upbeat summit]

1h left to midnight (CEST)

[cosmic quiver]

I'm not sure if they are going to release it perfectly at 0:00

[upbeat summit]

me neither - it was just sarcasm 😉

[cosmic quiver]

I hope it will be something around 10:00 tomorrow. I took personal leave at work to try it 🤡

[gloomy lark]

wow really cool stuff. i guess I'll have to go see about that xotox checkpoint now too.

[misty hull]

I get you haha. I have to stop myself checking new model on civit ai since i also want to generate images and not only testing model 😄.

[upbeat summit]

1 minute left until midnight (GMT+1)

[native knot]

Bro, how you gettin' pics of my woman? 😄

It's ok...we'll always have

[wet raven]

(images of SDXL Turbo custom model)

scale model toy scene

(using LORA + downscaling then upscaling)

making weird images on purpose

[copper kraken]

i make weird images by accident

[wet raven]

Many people believe SDXL Turbo is always inferior to SDXL models. I disagree. All these images are made with SDXL Turbo (UltraSpice) model.
Of course I am sharing specific images, not every image I made. It can't do everything this well.

[sonic void]

Who knows how to train an xl lora model that is very reminiscent of film and television characters, and what parameters are needed? I found that 1.5 version training has better restoration degree

[dawn tundra]

/Sea, sea level, sky, freighter,

[peak dove]

I tried your safetensor - got this error

[glass forge]

This is pretty neat actually. if there was trains in mad max.

i mena the basic concept is... the deatils ugh

[meager canopy]

Not good on a windy day though 😄

[copper kraken]

it's a dora, so you use it in the lora node

recomend using with base sdxl 1.0

[copper kraken]

last three were by scheduling the offset of the fractal noise

using the tan scheduler node

[glass forge]

So

these ar ethe last few hours of SDXL glory....

history in the making

[copper kraken]

lol

last few hours, before the return of sdxl glory via ppl angry as f there's no contronets or finetunes

[glass forge]

haha!

2 more weeks for those

[peak dove]

As a rule, do DoRAs go slowly?

[copper kraken]

inference is the same speed as lora, takes more vram

i usually save any lora/dora i use a lot as a checkpoint cuz it makes all of them faster

all you have to do is connect a save checkpoint node to the lora node

[peak dove]

float_000004 is slow 🙂

Good, but slow

[copper kraken]

oh that's the text encoder only fyi

if they say "TE" in the file, it's just the clip/text encoder

the others are unet/model only

can't train both at full precision with 24gb vram, have to do em separately

[peak dove]

So how do I set it all up to work?

[copper kraken]

load one of these WFs

and it'll all be there

https://github.com/ClownsharkBatwing/RES4LYF nodes are here

[meager canopy]

Oh, you tidied up! 🤣

Still so many unconnected nodes though 🤷🏻‍♂️

[copper kraken]

they're there for convenient experimenting lol

[half gyro]

Is there a guide in this server

On how to train a personal

Lora?

[peak dove]

Lovin' the "shiny immediacy" of the csbw DoRA!!!

[peak dove]

#✨｜sdxl message

[copper kraken]

the text boxes in the default workflow in these images is full of triggers

it's actually the entirety of the captions i trained on in the boxes that start by talking about oil paintings

[peak dove]

Mobius+SDXL+csbw_DoRA

[copper kraken]

i love the wind effects

[peak dove]

I borrowed KevenG's prompt = A beautiful menacing girl materializes within a bar, her presence distorting the air, as she summons a swirling storm of razor-sharp playing cards that orbit her like a deadly asteroid belt

[copper kraken]

add stuff about the color, style of the image (oil painting, surreal photo, graphic illustration, pop art, etc), mood (mysterious, creepy, electric, etc)

and you'll get crazy style

[peak dove]

Agreed, I have one or two LoRAs stacked in there

Poetic Style and Ethereal Grace (by Andreac)

[copper kraken]

ahh yeah good ones

oil painting, thick colorful paint, a surreal low-light photograph of a beautiful menacing girl materializes within a bar, her presence distorting the air, as she summons a swirling storm of razor-sharp playing cards that orbit her like a deadly asteroid belt

[peak dove]

I'm trembling 😄

Don't see no 'deadly asteroid belt' though?!

[copper kraken]

hah

it's trained on 691 of my outputs

each with at least 2 separate captions, most with 4 plus, rotating through them with each epoch

[peak dove]

This one has a "third hand"; but I think it's the Casino Owner "making sure the house wins!!!"

[peak dove]

So each of those many files will give a separate and different look?

[copper kraken]

that's the training set

[peak dove]

... I'm learning

[copper kraken]

the goal was to get them to blend and it's actually kinda worked out

[peak dove]

So which of the files is a fully-fledged DoRA?

[peak dove]

(I come to this as an artist with a little tech savvy, so I'm glad of clarification)

[copper kraken]

https://huggingface.co/ClownsharkBatwing/CSBW_Style/blob/main/csbw_225tkn_r96a48_12epochs_base1-0-fp16_save-fp32_mixed-no_prodigy_decoef08.safetensors

i recommend using just this one

and not bothering with the text encoders

[meager canopy]

Shit

Inspector and croupier are not to be messed with.

[timid garnet]

👋 .

[meager canopy]

Is that SD1.5 walking off into the sunset?

[glass forge]

He is not going anywhere. You can get some interesting results still if you refine a generation with SD 1.5 the skin and face details are still better than SDXL

[timid garnet]

Let me doubt about this, lol

[glass forge]

Try it

[timid garnet]

I did, 1 year ago

[peak dove]

Using ClownShark's DoRA inside SWARM

[silver fiber]

An Apple

[meager canopy]

100% Pixart_Sigma

[copper kraken]

Here is the image you requested.

[stray warren]

Crazy people 😂

[thick nexus]

anyone know of a way to export a sequence of images where vram gets cleared after each image export? im trying to animatediff at a higher resolution but batch rendering a bunch of large images crushes my 4090 card's max vram. I only get images to about 512 x 760 but no bigger.

[cerulean summit]

y'all does pony diffusion have a discord server?