✨｜sdxl

[coral canopy]

hmm, got the link to it? any mixed model you might suggest?

[ruby kayak]

I use SDXL models and mix SD 1.5 with them

I use the Fooocus UI

[coral canopy]

you mix?

or download a mixed version

[coral canopy]

ah im using forge

[ruby kayak]

I use SDXL as the primary and SD 1.5 as a Refiner

I use Forge too, but Fooocus has its own merits

[coral canopy]

what use is a refiner?

sounds good

[ruby kayak]

early SDXL models required Refiners to render correctly. Many newer SDXL models have a refiner built in

If you like Photorealiostic images. Some SD 1.5 models render better skin details

[coral canopy]

guess ill install 1.5 then

can you link me the newest 1.5 version?

[ruby kayak]

you can use 1.5 or SDXL models in Forge

[coral canopy]

yea ik

where can i find the newest 1.5?

[coral canopy]

also do you know how i can activate the refiner?

[ruby kayak]

This is where you activate Refiner in Forge

Its called SD VAE

[chrome flicker]

anyone can help me ? I want to generate a high-resolution detailed map, with the same layout and structure as the reference image I give. The prompt is : Top view of a game map, deserted island, fuchsia earth, forest, spider nest, white spider silk, lake, island.

I only have the rough map and don't know how can generate a beautiful map with same structure

[copper kraken]

Here are the images you requested.

[chrome flicker]

Too imaginative

[copper kraken]

Here is the image you requested.

[chrome flicker]

oh, it looks good

[meager canopy]

You can download most models from Civit. Example: https://civitai.com/user/galaxytimemachine

[topaz jacinth]

here's what I got

[meager canopy]

@gloomy lark Thanks for the prompt 🙂

Has anyone used "Boltning Hyper"?

I prompted for a "kingfisher bird", and with CFG 1.0 I get images like this...

However, if I change the CFG to 1.2, it gives a load of images like these 🤣

Apart from the slight CFG change, nothing else changed.

Similar with CFG 0.8...

[chrome flicker]

structure changed

[chrome flicker]

this is good 👍

[meager canopy]

Nothing strange here, all perfectly normal.

[copper kraken]

got my eye on you.

[meager canopy]

https://tenor.com/view/sauron-lotr-lord-of-the-rings-mordor-the-eye-gif-5619682

[peak dove]

Searge's SDXL w/flow

[stray warren]

Why so dark, go light and cheery 😊 @copper kraken

[copper kraken]

🤣🦈

[stray warren]

Hahaha you found it!

otoh

[copper kraken]

So much greatest in that one

Clownshark heads, one flying like a missile looking shocked

And one with batwings

I love seeing a random shark soaring past the moon like Peter Pan in images

You have a lot of great stuff on there btw, nice work

[stray warren]

And Homer and Cartman 😁

Thanks 😊

[gloomy lark]

It's one of the few times where svd actually puts out a good video

[stray warren]

Did you download the seperate vae?

[meager canopy]

Yes, it's just the SDXL VAE, isn't it?

The other images I created, above and below those examples, are also from the same flow and model. I guess it must have the correct VAE or they wouldn't have been as good.

[gloomy lark]

I'm usually interested in giant epic scenes of something huge, I should try this closeup stuff more often

[stray warren]

I dunno which vae it uses, but i guess you used the one they link to
It's just because of the washed out colors of the kingfisher bird that I thought of it
Very strange result with the cfg settings, does it mean it's overtrained?

[meager canopy]

No idea, but I was trying it with other prompts and changing the CFG didn't do that. I like the model, but there's something a little odd about it 🤣

[stray warren]

"a little" 🤣

[unkempt wagon]

https://tenor.com/bmi9t.gif

https://tenor.com/bJNqw.gif

[peak dove]

^..^<

[peak dove]

Use the word "fauvism" in your prompt - brings the colour saturation up a treat!

[stray warren]

I really like my new cover image for civitai

[gloomy lark]

Me after switching to Mac.

[uncut gull]

"I am now more powerful than my computer."

[acoustic spruce]

Is anyone here good at img2img upscaling? I made this (first) image, which I want to upscale. I'm playing around with it a bit, and found that these settings (second img) produces insanely good texture and sharpness (third img) but it also changes the motif of the image too much. I'm using basically the same prompt as when I generated the original image, just added sharp focus and smoke. Any way to work around this?

[ruby kayak]

Although I've not used img2img I do know that Denoise levels do impact the final results when upscaling in txt2img.

[acoustic spruce]

That's correct, it does change the output image, for better or worse. Thing is that you need enough denoise to change the image a tiny bit with the upscaling, but not too much so it changes the whole picture. I feel like im losing the sharpness and texture when I lower it too much, and if I put it higher it changes everything to pure chaos

[peak dove]

Surreal Fun with i2i in SDXL

[gloomy lark]

Img2img is great fun

[meager canopy]

Is 4k big enough?

[acoustic spruce]

Actually it's too big 😄 I'm usually settling for 2k!

Whats the workflow?

[meager canopy]

2k version

[meager canopy]

Nothing simple! 🤣

[acoustic spruce]

MY ARCH NEMESIS COMFY

shiiiet

[acoustic spruce]

Looks sick my friend

Too bad for me I will not be able to replicate it, it looks like

[meager canopy]

Before/after comparison

[meager canopy]

A less smooth version

[acoustic spruce]

WHAT IN THE ACTUAL FUCK MAN now you're just teasing You really wanna lure me into spending +500 hours into learning comfy? :((((

[glass forge]

Comfy is big boy/gal

the rest is amateur hour

[spiral bridge]

Comfy isn't very comfy 💀🙏

[native knot]

ComfyUI isn't difficult at all. The difficult piece us understanding exactly what each node does that you are using and how they are supposed to connect to provide the results that you seek. If you start very small/simple and gradually add to your knowledge, you will find great success. Not because starting small/simple is required, but because the act of slowly maturing your workflows and understanding is teaching you the skills you truly need in order master it:

1. Knowing how Stable Diffusion and other related tools work, and...
2. Knowing how, where, and when to search for information in this space.

This is where the act of learning ComfyUI allows one to exceed tools that are less capable overall. Once you understand it to the depth you need, you can begin to break past the boundaries of most of the other tools in the space due to the freedom ComfyUI allows.

[glass forge]

i used to say the same thing now i make noddle kits with 100 nodes

[spiral bridge]

Doesn't make it comfy fr still tragic with what u have to do (does give most flexibility)

[glass forge]

its like LEGO

gets addictive hooking stuff up

[misty hull]

@meager canopy That's really insane how you keep the details while upscaling man. In the best upscaling I've seen 👏

[acoustic spruce]

I mean you're not wrong. I discovered this text2image local stuff around 2 months ago, having zero experience with programming or coding. Checks out

[kindred shoal]

does controlnet work with ponyxl?

im struggling just a little bit, but maybe because i have no uh. aiding in the prompt

it just refuses to listen

[glass forge]

This represents our reality in so many ways.

[copper kraken]

loras don't work, guessing cnet might not either

[kindred shoal]

mnh that sucks, aight.

like i'v egotten it to work before, i just dont remember how

and whenever i try certain sdxl cnet models, it pops out an error.

i'll try another sdxl model.

[worthy orbit]

Are you using SUPIR?

[kindred shoal]

alright, does anyone have any tips on how to word things in sdxl for someone who's been using 1.5 for a while

[high sand]

draw: AI VS press

[kindred shoal]

well i cant seem to get controlnet working at all, wtf.

im assuming the wrong model i'm using is the only one working, as the rest just TypeError: 'NoneType' object is not iterable

[worthy orbit]

It is much simpler, don't put anything in negatives, unless you don't want something in the picture.

I usually do

Medium / Subject / Location / Style / secondary attributes / secondary details about style

[kindred shoal]

this might just be a forge issue.... i hope this isn't a forge issue lmao

okay so does anyone know how i can access the a1111 ui from my profile on the computer?

if i need to go to tech support, just tell me lol

[worthy orbit]

We are generating your image, please hold.

[surreal ferry]

/txt2img A white man , dressed in an OpenAI logo black T-shirt , writing on the blackboard with white chalk ，the blackboard have two pieces and can be moved up and down. The content he wrote is "Transfer between Modalities. Suppose we directly model P (text, pixels, sound) with one big autoregressive transfoer. What are the pros and cons?" Shot from the back.

[worthy orbit]

Here is the image you requested

[kindred shoal]

i can't even get controlnet to work with sdxl, wtf

just a normal sdxl checkpoint.

im surprised it's not yelling at me for the negative prompt i forgot to clear

just...what am I doing wrong? I'm so confused.

trying different sdxl cnet openpose models and it's...just not listening to me.

cranking the weight did jack fucken nothin.

[valid yacht]

draw : A logo for a learning management website named 'Aibou.' The logo should reflect a user-friendly and engaging platform that emphasizes active learning and interactive flashcards. Incorporate elements that symbolize learning, growth, and collaboration between students and teachers. The design should be modern, dynamic, and inspiring, with a focus on educational tools like flashcards. Consider using vibrant colors that evoke a sense of motivation and enthusiasm for learning. The logo should be simple yet distinctive, suitable for both web and mobile platforms.

[worthy orbit]

Openpose is pretty frustrating, but you could try other Openpose models, and non anime first https://huggingface.co/lllyasviel/sd_control_collection/tree/main

thibaud_xl_openpose.safetensors

[kindred shoal]

will give a shot, thanks

[kindred shoal]

that one didn't work

i've tried like 5 diff open pose models at this point

time to try 1.5

yep it ignores it in 1.5 as well.

that's annoying af.

[meager canopy]

Not in those examples, but the last one I posted included it as part of the workflow.

[copper kraken]

https://huggingface.co/xinsir some new controlnets

[kindred shoal]

does anyone know why controlnet would just randomly stop working?

i can't get it to work on any model i use.

sdxl or not, it just downright refuses.

[tiny storm]

fish made with origami

[meager canopy]

It doesn't look like origami to me

Maybe this one?

[copper kraken]

Here is the image you requested.

[tiny storm]

nice

[copper kraken]

Here is the image you requested.

[tiny storm]

shark man!

fish

[meager canopy]

Desert clothing 🤣

[dark tangle]

Go to the desert

[meager canopy]

Is he carrying his packed lunch?

[meager canopy]

Darth in the Hood!

[glass forge]

damn it renders vader well

[meager canopy]

Other side says "Food"

[sweet mason]

Darth in the Hood!

Darth in the Hood!

[meager canopy]

Cascade, but SDXL refined

[gloomy lark]

pixart/refined

[kindred shoal]

god these are so cool. one day i'll get that good

[glass forge]

i have no clue how some if these are done

i do realism only

u just type in vader in pointillist impressionism bauhaus or something?

[gloomy lark]

Mine was: gangsta Darth Vader head floating in the air with gold chains pointing a gun, in the style of gritty graphic novel, background is sticker sheet of chibi cute stormtroopers

[vital ermine]

https://civitai.com/models/483223?modelVersionId=537435

[vital ermine]

https://civitai.com/models/150871?modelVersionId=538463

[mortal marsh]

Screenshot from #1230212453288378418 message

[worthy orbit]

Simcity vibes

from far away, zooming a Q-Bert level

The site is crazy

[dawn cedar]

https://tenor.com/YwL7.gif

[gloomy lark]

Anime-style dramatic scene featuring a determined white-haired, red-horned woman in dark, glossy armor with red lines, brandishing a fiery red blade, intense battle unfolding with robotic soldiers in matching armor, background ablaze with orange flames and flying rubble, towering gray cliffs frame the action, sunlight piercing through clouds, crisp details, and contrasting fiery and cool color palette.

[misty hull]

@gloomy lark 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.

[gloomy lark]

that's a neat prompt

[misty hull]

Wow, looks really nice

[gloomy lark]

are you using ella?

[misty hull]

Yesss sir.

Best way to go. Ella or pixart. I alternate 🤪

[gloomy lark]

I'm using juggernaut reborn for the sd 1.5 model, and zavichromaxl_v70 as the refiner.

Just now, I'm trying aniverse 4.0 which is supposedly highly regarded, I wanted to see if it was more prompt following. obviously did a good job with this one.

aniverse 4.0 is sd 1.5

what's the prompt for that top one you posted with the woman and knights?

I just zoomed up the samurai one you posted. really good sharpness on that.

[misty hull]

I think I used realisticvisionv5(computer is closed) for sd1.5 and Cheyenne for refiner

[misty hull]

I haven't tried these two, but I use chroma a lot. One of the best models

[misty hull]

I used your prompt

[gloomy lark]

so far i'm liking the output after refinement with this aniverse.

will smith with god powers with hundreds of children riding on top of will smith. God rays, laser beams, smoke, glowing yellow eyes. Perched on a mountain of spaghetti.

[misty hull]

That's is really good. It did an amazing jobs on the hand 😱😱

[gloomy lark]

i have clownshark to thank for that. he's doing amazing stuff with the res_momentumized sampler

squeezing every last bit of detail out with it

[misty hull]

Indeed, it's really amazing.

I'll have to get into it. I did a couple of generations, but definitly not enough

[earnest crow]

apple

[copper kraken]

Here is the image you requested.

[peak dove]

i2i-SDXL makeover using images I made at the Venice Carnival in 2017 - original then SDXL versions

[peak dove]

i2i Searge's w/flow

[lusty wolf]

What you mean there is too few fingers...

[peak dove]

A Trifingerant!!!

[lusty wolf]

Usually there are way too many...

[midnight valley]

A lotus comes out of the water, with huge flowers blooming and several lotus leaves covered in water droplets on the surface.

A lotus comes out of the water, with huge flowers blooming and several lotus leaves covered in water droplets on the surface.

[meager canopy]

Here is your image

Cascade

[meager canopy]

6912x4032

[fiery cradle]

A group of children happily celebrate Children's Day

[cyan crown]

https://civitai.com/models/486306

[peak dove]

i2i Searge SDXL workflow, checkpoint is NightvisionXL. Prompt made using Jan.ai (Llama 2 Chat 7B Q4) = Imagine a vibrant and colorful mashup of Peruvian arpillera and Hawaiian jazz, set against the lush backdrop of Cocoanut Grove. The year is 1950, and the island is alive with the sounds of music, dance, and song. In this imagined scene, a talented Hawaiian singer, is taking center stage, accompanied by a group of skilled musicians and dancers.

As you observe the performance, you notice that the artists are drawing inspiration from various artistic movements, including Cubism, which is evident in the abstract shapes and vibrant colors used in the arpillera. Meanwhile, the jazz musicians are channeling the spirit of Henri Rousseau and Vladimir Kush, with their improvisational skills and lively melodies.

Suddenly, a burst of energy fills the room, as Victo Ngai, a renowned artist known for her bold and colorful works, joins the performance. She brings her unique blend of Calypso and jazz, infusing the music with a sense of movement and energy.
As the performance continues, you find yourself lost in the magic of the music, dance, and art. The fusion of Peruvian arpillera with Hawaiian jazz creates a truly unique and captivating experience, one that blurs the boundaries between art and music, and celebrates the diversity and creativity of the artists involved.

[crisp owl]

Did Searge ever show up again or is he still MIA?

[crisp owl]

Mmm seems still MIA, last updated workflow on civit still last year

[cyan crown]

https://civitai.com/models/485548

[kindred shoal]

is anyone struggling with civitai download speeds?

ookla speedtest says my speed is fine, but civitai is only downloading at 1-2MBps

[gloomy lark]

Yeah it's often not great. Sometimes they include a hugging face link on their info page as well.

[kindred shoal]

whatever was wrong with it, it's good now i think

[gloomy lark]

https://huggingface.co/Corcelio/mobius/tree/main

looks like it's out now

[kindred shoal]

in my journey of messing with area composite, i've made this thing

[tawdry current]

Make sure if you're going to try it that you set your clip skip to -3, otherwise, the outputs aren't all that great. (It states the clip skip on the hf page)

Didn't think I'd see an sdxl model in 2024 needing clipskip, but I assume that's part of how they substantiated their lower training cost claims

[tawdry current]

figured i'd torture others with some abominations

I've been on an underwater/fish kick lately

[peak dove]

i2i Searge SDXL w/flow

[worthy orbit]

Interesting, I want to try it but

[glass forge]

cant wait to make these people act and come to life

[vale vapor]

what checkpoint nd workflow is this??

[glass forge]

Phenomenal hands.

[kindred shoal]

I didn't even need controlnet for it! SDXL just popped out really good hands.

[peak dove]

i2i using Searge's SDXL w/flow, NightVisionXL checkpoint

I will try and locate Searge's w/flow at Civitai ...

[vale vapor]

whats the prompt those look absolutely incredible. i lowkey thought it was sd3. i didnt know nightvision could do that. i thought this searge stuff was just the normal workflows but put together in a node that makes things everything easier to use. is it adding extra hot sauce while the nodes are cooking?

[peak dove]

https://tinyurl.com/CivvittAI this is a collection of Searge's w/flows

[peak dove]

I made the prompt via Jan.ai - using Llama 2 Chat 7B Q4 - and it is enormous ...

[worthy orbit]

interesting nice model!

[peak dove]

I added an automated i2i input; and Face Detailer as well.

Change the denoise in the Searge w/flow for interesting results

[worthy orbit]

I looked at that but it doesn't want to work, it says it needs portable Comfy on Windows, I'm on Linux

[peak dove]

I'm using System ComfyUI - I get conflicts using Windows Portable

The same results are available using a basic Base/Refiner/Face Detailer setup ...

I don't think that most of Searge's w/flow - although very clever and meticulous - is much better than that basic setup for results.

Here are some pictures just output ...

[worthy orbit]

Looks really good

[peak dove]

The w/flow should be in each png

[worthy orbit]

At 100 % percent it has some kind of TV lines (?!), like some upscaler artifacts, but at less zoom it works really good for making it realistic

[kindred shoal]

i think discord deletes that metadata? i'm not sure though

unless i'm just dumb.

[worthy orbit]

maybe it is the idea

[worthy orbit]

No, it doesn't

[kindred shoal]

yep i'm just dumb

[worthy orbit]

I think you have to open the image in browser and save it there

[peak dove]

I think the success of these images is A) high quality i2i images and B) having the original prompt improved using an LLM (Jan.ai) - plus my go-to checkpoint NightVisionXL

[worthy orbit]

how much denoising?

and steps?

[peak dove]

If you use 0.2 to 0.4 - much of the original input is preserved - from 0.5 to 0.8 - you start to get totally different (from the original input) but still good.

45 Steps, CFG 5, Karras, DPMPP 2M SDE GPU

[worthy orbit]

ty, yeah, the thing I'm never quite pleased with the result, it is hard to find the sweet spot (!?)

[peak dove]

Try and marry your input photos to the prompt. Don't try and make grunge using a Hawaiian Sunset palette - if you know what I mean?

And if you lack good quality input - get some Public Domain/CC0 photos from Unsplash, or Wikimedia, or Pexels ...

Sweet spot - it's all trial-and-error! Good luck

There are two ways to get great photos: extend and expand the artistic possibilities available to you; OR extend and expand the technology (sigmas, loras, tensors, etc etc)

I am definitely for the artistic route. And grateful for the technology!

[worthy orbit]

what's sigmas?

[peak dove]

https://stable-diffusion-art.com/samplers/

Sigmas elude me, but are something to do with the Noise Schedule ... I'll need to read-up on this!

[worthy orbit]

yes I knew Karras etc was the scheduler (!?) of the sampler, but not much than that

[peak dove]

I ran it with Diffusers - it made an image - no clue as to where the image is stored? 🙂

[trim bane]

Anyone else keep getting 'NoneType' object has no attribute 'rsplit' randomly?
It happens with img to img, usually (but not always) after changing the cfg scale.
Closing and re-entering the image sometimes helps, relaunching the program always helps, it's pretty annoying, and it only started happening with xl models.

[wanton river]

nice can you share workflow/what model used?

[gloomy lark]

Dark fantasy digital illustration with eerie green lighting focusing on a will smith king adorned with a cracked, jewel-encrusted crown with glowing green eyes and tattered clothing draped in a shadowy misty background with oozing, melting flesh and ominous black wisp-like elements floating around, creating a haunting and decayed atmosphere. His throne room is filled with the undead praying to him.

[dire pulsar]

A large and menacing gargoyle stands guard over an aged electric guitar in a dark and eerie alley. He is surrounded by fantastical disturbing creatures causing a scene of surrealism. Dead roses and rotted thorns adorn the guitar.

#🏞｜general-with-images A large and menacing gargoyle stands guard over an aged electric guitar in a dark and eerie alley. He is surrounded by fantastical disturbing creatures causing a scene of surrealism. Dead roses and rotted thorns adorn the guitar.

[vital ermine]

https://civitai.com/models/489470?modelVersionId=544285

[gloomy lark]

In a vibrant, futuristic Neo-Expressionist style with exaggerated forms and bold colors, a sea of iridescent microbes clad in tiny activist accessories like megaphones and picket signs, each unique with expressive facial features, assemble in a crowded, pulsing metropolitan petri dish, surrounded by eerie bioluminescent lighting and dripping with tendrils of glowing, electric-blue microbial goo.

this has to be one of the stranger ella outputs based on the prompt.

[vital ermine]

https://civitai.com/models/489470?modelVersionId=544662

[rain granite]

生成一个亚洲女性，带着一个棒球棒，侧脸45度，衣服和帽子都是白色，美女，脸部稍微丰满，可爱类型

[peak dove]

i2i Searge's SDXL w/flow

[vital ermine]

https://civitai.com/models/274262?modelVersionId=545270

[solar rune]

I implemented my own "prompt injection",

Now I can... prompt SPECIFIC BLOCKS of the UNet in different ways, And I implemented three different kinds of nodes for different workflow usages Each of these on the screenshot does the same thing, inject a specific prompt in OUTPUT0 and OUTPUT1, but the inputs/nodes is different

If every block has the prompt with "blue dog", BUT only OUT0 and OUT1 is fed the prompt with "white cat"... This is the result

Now, this is the image without any injection, all blocks being fed the blue dog prompt

this is the white prompt cat when it is fed to allll the blocks

So the thing is, this allows for more control
cause like, if u like the composition of a prompt, but want to replace the subject/content with something else, if u just do the usual way of replacing, composition changes entirely, u see how the entire cat prompt contains 2 cats?
But if I use the dog prompt to all blocks but then feed white cat prompt to only the OUT0 and OUT1 blocks, then I get the same composition as the dog prompt... BUT WITH A CAT
usually the prompt is fed to all Blocks, this basically allows you to replace or define which prompts goes into which block, for any block you dont specify, like IN4 and etc, it will still use the original prompt

in SDXL, based on a paper, and based on tests I also did myself, The content/subject understanding of the model is almost entirely contained within the MID0 and MID1 blocks

Like, in that image there, revolver ocelot, or "Kazuhira Miller" only appeared when I apply lora to blocks OUT0 and OUT1
meaning that... if someone wants to train only subject/content plus a bit of their style, the only thing they need is a Lora that affects blocks OUT0 and OUT1

https://github.com/DataCTE/prompt_injection

[gray crescent]

画图

画图

[gloomy lark]

Hi there, you should post that here, you might get more uptake. https://discord.com/channels/1148340499866980402/1148864356906442793

[solar rune]

i don't have access to that channel

[gloomy lark]

I have to send it privately. One sec.

[rough moss]

create a sign with the word: "Schäf"

[meager canopy]

Will has been eating too much spaghetti!

I think the prompt was from @gloomy lark

[gloomy lark]

Hah loving the style on this one

[slender coral]

A recent push to comphy is breaking all model.xxx does anyone know what the new properties are or fixes?

                                                                 ^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'device'```

[gloomy lark]

Refresh your browser and make a new ksampler node maybe. It's trying to push device when it's not there in the new version maybe

[slender coral]

Yes, thats why I'm asking what changes were made to the model module. .

[misty hull]

The Event

The After-Event

[misty hull]

The perfect mac & cheese chain. The color contrast looks really nice

[glass forge]

Just looking at it gives me diarreah.

[vale vapor]

please can you also post in comfy discord and reddit..plus include detailed examples in the github cause there are no images that explain what these do on there

[glass forge]

Watch out SDXL! SD3 is coming!

[peak dove]

SD3 Weights drop on June 12 #📣｜announcements message

[kindred shoal]

omg i just realized she has the right amount of toes XD

[glad grove]

she has the right amount of everything 👌

[meager canopy]

...apart from years, perhaps?

[glad grove]

you'd have to ask her

[meager canopy]

ID PLEASE

[kindred shoal]

LMAO

I did not put 'child' in the tags if that's any help

or 'teen'

... i dont think i put adult either so it just kinda did whatever

yeah no i didn't put adult in there, lmao

seeing how much different it is with the 'adult' tag

[glad grove]

should have put homo sapiens female,1 torso,2 arms,2 legs rookie mistake

[lime fractal]

Design a T-shirt

The theme is the 25th Anniversary.
The target audience is company employees.
The color should be elegant and simple, suitable for everyday wear.
It must include the company logo, which is the logo in the picture.
The design style should be simple, modern, and elegant.

[copper kraken]

Here is the image you requested.

[solar rune]

There is a comfy discord? I won't be posting it on reddit

[meager canopy]

No Discord, just uses element.io

Cascade + SDXL

[marsh roost]

gun

[meager canopy]

https://tenor.com/view/going-down-duck-get-down-police-squad-lieutenant-olivia-benson-gif-14246478

[glass forge]

tax bot

[meager canopy]

eh?

[glass forge]

Dead or alive your money comes to me. Not the queen.

[meager canopy]

ah! 🤣

[glass forge]

Pay me sedated peasants!

Mars is mine.

[glass forge]

Dubai calls. Oh wait now, that's worse.

My queen I am bending over take a good look.

All yours.

Sweden? Stalin?

do you love me?

Lubricated.

[meager canopy]

Have you been spending too much time around clownshark? 🤣

[glass forge]

Nah

😄

Reality and my frien;d reality

[meager canopy]

Contagious madness

[glass forge]

We have to laugh

or we go insane.

[meager canopy]

...but should we be laughing with the insane?

[glass forge]

Do you want to be succesful as a self employed ineterpreneur ?

of cours enot. why because you dont want to support the government

All goverments are corrupt to the core

When you are down they ignore you

when u try to succeed theyre interested

who are these people?

i never m,et them

did you

oribably not

youll see

just wait

[glass forge]

yes thats how i look

being tired of life

sigh

the whole planet is communist

[meager canopy]

The crown must weigh you down

[glass forge]

lol

We're all communists.

The incentive is gone.

Time for vodka and cigs.

I am so sorry kids.

You will never know what the American dream was.

I grew up under totalitarion communist regime. I see what it does to people. And now I see the same thing again but now it is happening globally.

Many chines ehere yes?

If you let it rot I udnerstand.

Even in america or europe we feel the same

why do anything.

Let it rot.

[meager canopy]

I don't think any of us are here for your political views 🤷🏻‍♂️

[glass forge]

its not politics its how it is

😄

anyways

often talks in AI rooms go into this sort of things

😄

it shows people or tense

and ready to blow up!

world over

guess what

not ponly 9 out of 10 businesses fail even thsoe who make it taxes and inflation will take eveyrhtign you amke

[meager canopy]

You're rambling nonsense?

[glass forge]

so

maybe AI( will be the last nail in the coffin of the economy

lets hope

[glass forge]

I wish...

How old are you?

[meager canopy]

Why?

[glass forge]

Takes some time to see through the BS.

We tell lies to ourseklve sbecause we want to believe certain things so we refuse to see the world as it is

[meager canopy]

What's that got to do with my age?

[glass forge]

But i see many very young people especially in asia give up today ebcaus ethey are smartt enough to see the system is rigged against them

good for them

well if you are young you may be naive

[meager canopy]

Have you considered you're telling lies to yourself because you refuse to see the world for what it is?

I'm 58

[glass forge]

well then you are way ahead of me

so you must knpow what i sya i strue

please dont tell me ypou buy the party line at 58

pleaZe

unless you inherited a fortune or are sheltered af theres noway

[vale vapor]

yes there is, you want an invite?

[solar rune]

i would love a invite!

[glass forge]

isnt that just here?

like asub channel

like it's just there

points comfy-ui

[soft merlin]

hey guys, can SDXL do outpainting? I want to expand some images

[copper kraken]

just a simple 1024x1024 generation with some sampling tricks

[balmy mirage]

The three-dimensional intelligent fish farming plant on land, plus photovoltaic and breeze power generation

just a simple 1024x1024 generation with The three-dimensional intelligent fish farming plant on land, plus photovoltaic and breeze power generation

@balmy mirage

[copper kraken]

Here is the image you requested.

[balmy mirage]

thx

[copper kraken]

Here is the image you requested.

[candid heart]

is there an SDXL discord bot or only SD3?

[vital ermine]

https://civitai.com/models/173888?modelVersionId=549100

[glass forge]

workflow plZ

[copper kraken]

it's embedded

but you'll need some custom nodes i made

and a crazyass sampler i made lol

changes the frequency of ancestral noise throughout the generation

[glass forge]

😏

[misty hull]

Mobius creativity is blowing my mind 🤯

[glass forge]

Nvidia Town

[lyric roost]

A captivating anime illustration of a young girl with a unique sam yang style, wearing a bright red jacket with intricate patterns. She has long, flowing red-orange hair and possesses a playful, yet determined expression. The background showcases a bustling cityscape, with towering skyscrapers and a vibrant neon cityscape. The overall ambiance of the image is energetic and lively, with a futuristic touch., illustration, anime

[lusty wolf]

Image you requested:

[stone fossil]

He asked for samyang style....

[winter bridge]

a apple in outer space

[signal mirage]

this is not the image you requested:

[craggy basin]

!diffuse --prompt "3D rendering of an industrial structure with enhanced three-dimensional effects"

[blazing kernel]

that's sdxl?

how did you manage to get clear text

[misty hull]

I used Mobius to generate the base image, which is doing realllly well with text. Then I kept the denoise fairly low for the rest of the wf to keep the text right

[blazing kernel]

is that a finetune or a whole site/service on its own

I couldnt find it

I found a model named moebius

[misty hull]

This is the model sir! https://huggingface.co/Corcelio/mobius

[blazing kernel]

oh thanks!

[misty hull]

My pleasure 🙂

[blazing kernel]

oh I didnt know that

I'll check as soon as I get a new ssd

my current "checkpoints folder" is about 500gb lmao

not entirely diffusion models, have few LLM's but mostly diff models

[misty hull]

This is also another thing 🤣

I won't forget it haha

What model is that? Look really nice

Thanks 🙏 I'll try it later

[stray warren]

😉

[misty hull]

Great job man, I enjoy it a lot :-).

[misty hull]

Colors are crazy on these last ones 👏

[charred perch]

#🆕｜sd3 girl

girl

[misty hull]

?

[charred perch]

how to use

[copper kraken]

new sampler tweak there

using fourier transforms to impose the spectral shit from a source image onto the noise generated by the noiuse sampler

source, a woman with wings, a gunship with wings

[misty hull]

I don't get the spectral shit part 😄 but it surely works.

It's similar process to unsampling?

[copper kraken]

only in that noise is involved

[gloomy lark]

An imposing cookie monster in a cream suit with bold orange polka dots stands among towering poppies, dark skin contrasting against bright floral backdrop, golden sunlight casts sharp shadows, sky a pale azure, flowers vibrant with delicate petals, combining natural elements with stark geometry, creamy white hair and solid brown boots grounding the composition, captured at a low angle.

[dim steeple]

What is the best way to draw text in SDXL?

with quotes?

[signal mirage]

Depends on the scenario. If you just want to add text “on top” of the image use a simple image editor. If you want to add text to objects in the image I would use the image as depth map and add a near white text at the position where I need it.
3rd way would be using lora or models which are at least slightly better with text. (Mobius model is able to generate text but not at a 100% score)
Or you wait till next week and use sd3….

[shy kelp]

🙏 ✨

[glass forge]

is this better than DPO lora?

=0

hmm its a text aid lora but those example results below are not really all text related...

😏

[glass forge]

This looks like literarily you just typed in your discord name.

[tawdry current]

Why not just use Photoshop(or krita) rather than wasting an exponential amount more time trying to get lucky? I mean I understand if you're trying to have it make some complicated neon light design for text, but for simple crap like black text on a whiteboard, just Photoshop it lol...

[tawdry current]

made another waifu for you thirsty guys, drink up! 💦

[copper kraken]

snakeloop

[meager canopy]

https://tenor.com/view/mobius-gif-11033415

[glass forge]

someone out there will be batin to this

[misty hull]

I think mobius dont work very well with lora 😰

I havent, but i read it on another channel.

[glass forge]

thre is a spanish movie called moebius

about subways and stuff

No drugs were consuemd in the making of that movie.

[worthy orbit]

yes it is very nice movie, language is spanish but it is not spanish, is an argentine movie 🇦🇷 , filmed in Buenos Aires subways.

I'm from Argentina BTW

[peak dove]

We shd have some matté

[worthy orbit]

lol we need a SD XL mate Lora https://civitai.com/models/49455/mate

[cyan crown]

I'm sorry to tell you that I canceled all my loras from Civitai.

In last 2 moonths I began to receive requests to do specific loras for Pony (which I didn't do) and then continuous notifications of porn images made with my standard lora and uncensored models. I trained my loras for other types of images. And filters on the gallery are not enough. Since we are not talking about a pair of tits but about extremely vulgar images, I preferred to delete them today. If the situation returns to normal sooner or later I might go back to publishing them in the future. But in the last two or three months these contents and images have really multiplied uncontrollably.

[gloomy lark]

well that's just extremely sad. I use your stuff in most of my generations on here.

You're now the 3rd lora/checkpoint creator in the last 4-5 months that has removed all their stuff for one reason or another.

[cyan crown]

I assure you that I am not a bigot, but this is really exaggerated.

[cerulean summit]

I don't think civit users are likely to change from that any time soon. Huggingface is somewhat less like that for what it's worth

[gloomy lark]

I totally get it. As long as it's not underage or unconsenual, I'm a live and let live kind of person, but I totally understand what you mean, the stuff on civit at times is kinda crazy.

I would agree with MrCheeze, your stuff would still be used for whatever anyone wants, but there wouldn't be a gallery of that on your page.

[cyan crown]

for gallery there's a filter, btw everyday I receive notifications of post made with them so...since I have two children it's not for me. And really I'm not a bigot. and also I'm a live and let live kind of person

[void skiff]

The massive numbers of the few. When you look you often see that the most ditry ones upload so much mote than normal users do, I have never uploaded images there for I do not feel a urge to do so, but folk with a kink are often exhibitionist too. Welcome to the Internet.

[gloomy lark]

I'm in a similar situation. There's been a couple of people on here who's posted underage suggestive stuff and it freaks me out possible more than someone without kids.

[cyan crown]

yes usually are the same users

[peak dove]

I absolutely love Andreac LoRA's - Ethereal Grace, Poetic Style ... (I think I got the names right?)

[cyan crown]

yes

[peak dove]

Fabuloso

[cyan crown]

thanks

[peak dove]

⭐

And the Shaun2 😄

[gloomy lark]

yeah i use cute 3d render on pretty much everything

it often takes more serious stuff and adds a light hearted nature to it

[ember gale]

Hi

[peak dove]

Is Cute 3D Render a LoRA or just a prompt phrase?

[gloomy lark]

it's a lora.

that andreac75 made

[peak dove]

OK, will check it out ...

[gloomy lark]

well they just removed it all so i don't know that you can anymore.

[peak dove]

Oh, one of Andreac's?

[gloomy lark]

yeah

[peak dove]

Sad, he does some excellent LoRA - he will be missed!

(If Andreac agrees, send me a Google Drive link to share?!)

[gloomy lark]

@cyan crown I really hope you'll do a version of cute3d render for sd3 when it's out and find another venue for hosting it so we don't have to miss out.

or patreon or something

[cyan crown]

thanks. Maybe SD3 doesn't need

[gloomy lark]

sd3 is amazing but it needs training big time, by people like you.

[cyan crown]

did you try the candidate or api?

[gloomy lark]

I have the api as part of automation I've got going. The sd3 api is awesome at some stuff and worse than the other new llm t5 based Chinese models a lot of the time. Sd3 is capable of more, but it needs a ton more training which the community will need to provide

[cyan crown]

i tried api too. But I hope 2B model will be better in terms of quality

[gloomy lark]

I'm sure it will be. The few samples we've seen are already better than the 8b beta that's on the api.

[stray warren]

Irt's sad that you had to do it, but you're absolutely right. ❤️

[cyan crown]

thanks

[warped sedge]

nasty civit strikes again....

[rustic garnet]

I really loved your loras and frequently visited civitai just to look if you published something new 😥

However, I can totally understand your feelings towards pony and this strange porn community on civitai

[gloomy lark]

(made with andrea75c's cute 3d render lora)

[kindred shoal]

is adetailer easier to use in a1111? I feel like my workflow isn't the greatest for the comfyui detailer...

not doing adetailer yet, but i know how to make a basic workflow lmao

[kindred shoal]

is there a controlnet that works decently with ponyxl?

like openpose for example

i was able to get thibaud to work, but it doesn't work with the openpose i'm trying to copy

[kindred shoal]

i can't get dwpose working on comfyui...

is this image just too wide?

i think it's too wide.

[glass forge]

Well that's sad. by removing your loras you deprive legit users who try to use them for nice things and the vulgar porn will just keep comign anyway so i guess they win

even ic civitai banned it the porn will keep coming

AI pornhub next

shrug

shame to see trainers remove their content due to somethign theyc ant stop anywya and removing the loras an checkpoints wotn solve the issue anyay

[kindred shoal]

can i uh...get some help here with comfyui?

i can't figure out why my workflow is being a mess.

like it's...not printing out the image i want.

unless the dw is getting so confused it's not doing the fingers right.

regular openpose doesn't work here

just prints out black

[glass forge]

theres a comfui room

they might now better

[copper kraken]

can you post the lil green image here?

the output?

[kindred shoal]

Well sure, if y'all don't mind the abomination

that should have the workflow though...not sure...

if anything i can send the json.

like it looks like it's trying, but it just...can't.

it's the only sdxl model i have right now, but i can try another one if necessary.

[kindred shoal]

i don't know what the fuck im doing wrong and it's making me upset.

[copper kraken]

well what's the main subject of your source image

that takes up most of the screen

maybe try prompting arm instead of woman

might be worth trying a full openpose cnet too

[kindred shoal]

i tried that on the image and it doesn't work.

it just prints out black.

[kindred shoal]

up here

...i think it's the model.

this model is fucking up everything i throw at it.

trying a different model after it downloads...if it will.

...of course it's 2mb/s now on civit ai. ffs.

[copper kraken]

yeah word of warning, openpose is very hit or miss with sdxl, often mostly miss

and very model dependent

it's often a good idea to start with a partial denoise

and to use something like mistoline controlnet at the same time, as well as depth

[kindred shoal]

it works fine with uh..pony, but it's only thibaud.

i'll figure it out after this model downloads. which could be a while.

thanks civit for having horrendous speeds sometimes.

[gloomy lark]

Pixart

Cinematic slow-motion photography with high contrast and sharpness, focused on a Cookie Monster running with intense determination next to a black car amidst a massive explosion in an urban setting, vibrant flames and shattered debris flying, rugged textures of cracked asphalt and buildings, glowing tones of orange and yellow lighting, lifelike details, smoke and chaos in the background.

[runic geode]

https://tenor.com/LaGsRT8YDd.gif

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

[glass forge]

So there is an SDXL Lora for dirty toilets.

[peak dove]

SD3@ClipDrop

SD3@ClipDrop

It can do "bad" eyes!

[peak dove]

PiXartSigma+SD15+SDXL+PAG

[robust blaze]

A set of 6 chairs, 1 sofa, and a table made of dark-colored wicker rattan stands on a beautiful lawn with artificial grass near a cozy house.

[gloomy lark]

wow, which one is that, and is it still available? 🙂

[cyan crown]

sorry no. Yet never published

[gloomy lark]

I have an image like that, that i've been using with ip adapter for style transfer, but it would neat to have a lora.

[cyan crown]

produce like 10 images and train a lora

it's so simple

[gloomy lark]

yeah, I'll see if I can do this one with that style

[gloomy lark]

for that kind of thing, would I put it through onetrainer without captions? I've done a couple, fully captioned, but I haven't tried style ones yet.

[cyan crown]

I caption also for style

[gloomy lark]

so just have the image but put a tag or 2 for the style type?

[cyan crown]

I just describe images

so I don't have trigger words in lora

[gloomy lark]

ok interesting.. so just by having those pictures regularly captioned, that's enough to transfer style to all images generated after that?

[cyan crown]

yes

[cyan crown]

I usually use 40 images but done also with 5

[gloomy lark]

ok neat stuff

[supple spire]

Is this the default single mode? It’s freakin insane!!

[gloomy lark]

That's ip adapter style transfer based on one of andreac75's images

[supple spire]

I wish I knew what any of that meant… I’m just like very new to all this 🙂

[kindred shoal]

god, she looks so pretty....

i love the dress details

i love usdu but god the tile count going up to like...30 really hurt lmao

are there any tips with doing full body in sdxl? I keep getting warped faces and hands

i'm gonna have to use face detailer aren't I 😆 😭

it's ignoring controlnet again, but I think on a deeper level this time. it's not showing that it's reading the controlnet model in the uh... terminal.

i love when controlnet completely ignores me

pony needs to get cnet models for it...this would make life so much easier...

okay so it listens in a1111, but not comfyui...why though.

[floral quiver]

what do you use to train? can this be done right from a1111? I’ve done some searching but can’t find anything that delves into this specifically

[hexed totem]

一个现代化的室内台球厅。台球厅内有多张台球桌，桌子上摆放着台球和球杆。环境设计风格偏向工业风，天花板上裸露的管道和灯具显得非常时尚。墙壁和地板采用了灰色调，营造出一种冷静、专业的氛围。台球桌旁边有一些舒适的沙发椅，供人们休息。整个空间宽敞明亮，灯光设计合理，适合进行台球运动和社交活动。

[glass forge]

well if the bees are the current thing an image with nicholas cage covered in bees is mandatory

[peak dove]

My w/flow plus a new Andreac LoRA? 🙂

[peak dove]

I'd love to make my own LoRA - but get confused by the parameters in Kohya.ss

[tawdry current]

You'll find like a million guides all with conflicting information and that all swear up and down they have the "best" settings. There are a ton of settings that can affect it all, like batch size, network alpha, dropout and learning rate are all kinda tied together in terms of modifying how quickly or slowly it learns. If you're following some guide that sets them all, but you can only use batch size 1 instead of 4, the rest of the settings would need to be adjusted to compensate

And then depending on what you're trying to do, it means different settings as well. Like copying likeness of a person vs some art style. Oh and tagging is a bitch too

[buoyant crow]

Recently, a revolutionary ControlNet model has been released, so I'm sharing a summarized article about it.
I think it might simplify the workflow.
Since the information is scattered in Japanese, I believe it is difficult for overseas people to access, so I have summarized it.

https://note.com/kokekokkosann/n/n46b3967fe895

[copper kraken]

death star

subtractive gauss noise sampler

[oak shoal]

a girl

[rustic garnet]

interesting... I always wondered how large your training data is. For my largest style lora I used ~150 images and I still find it overfitting for certain stuff (in particular humans always look the same). Your Loras also often overfit (like dark side of the world always wants to add something circular in the ceiling xD), but some of the loras are extremely versatile (like the vintage anime lora)

it's surprising though that your Loras are, from the file size, larger than your training data. Did you ever tried using smaller Loras? Like I usually work with Loras of a few megabytes in size. Did you found that you get better quality with larger loras?

What I often do to train a style lora is a iterative approach where I first use few images or prompts to get a overfitted lora that is good in style but tends to produce always the same elements/compositions. Next, I use a large bunch of training images in any style and feed them into a depth control net to force the overfitted model to produce images that follow their composition. This way I produce new training data for the style lora

what also works quite well is using slider loras on these pairs of (original image, restyled image). They tend to do less overfitting than just training on restyled image.

[peak dove]

PiXart-Sigma+SD15+SDXL+PAG

[tawdry current]

When you're testing them out in apps like comfy, are you just rawdawging at 1.0 strength? Even with badly overfit loras, you can usually still use them if you just turn the strength down. Like one I made a while back, the sweet spot was around 0.65-0.7. I didn't bother wasting retraining it

[rustic garnet]

yeah, of course you have to check strength. But with overfitting I rather refer to things like "the same compositon is used all the time" or "all people tend to have the same face".

what also works really well is using block weight lora

train a lora first on all blocks and then remove all but the output blocks. This also has the effect that the Lora is only applying on style and not on composition

e.g. Andreasc75 Vintage Lora on "white tiger in the jungle":

without Lora:

with block weight lora:

this is full strength and of course, 0.7 would be enough. just wanted to show how well that works

[tawdry current]

Are you also training the prompt encoders? Because in comfy, you can vary the strength of the lora on the model and the prompt separately

[rustic garnet]

yeah sometimes. I guess for style models its better to not train text encoder, though

[tawdry current]

yeah, i mostly only make style loras and very rarely train the encoders. they just end up screwing things up 95% of the time

[tawdry current]

i've been playing around with only using the output blocks and it really is surprisingly decent at keeping composition (inspire pack's load lora block weight node set to sdxl out all). i'll have to experiment with this a bit more. it's breathing some life into some style loras that didn't quite meet the mark before

[peak dove]

PiXart-Sigma+SD15+SDXL+PAG

[boreal sparrow]

#1237459938901491852 a close up of a person in a costume with a sword, scorpion from mortal kombat, character from mortal kombat, mortal kombat 11, in mortal kombat, chaos nightmare ❄️ amour venom, black armor with yellow accents, style of mortal kombat, 8 k highly detailed ❤🔥 🔥 💀 🤖 🚀, mortal kombat, cyborg satan, menacing. unreal 5

[smoky shard]

Testing advanced style-aware inpainting on hand-drawn art

[royal geyser]

/drea,

[sharp scarab]

New system available (: - https://www.youtube.com/watch?v=IQn3CJ9cTNA

[cyan crown]

smaller loras does not achieve the result. Usually 40 to 50 images are enough. I think overfit is for steps and not for size. Since I do around 2K step per lora

[vale vapor]

what checkpoint is this and wats block weight lora?