✨｜sdxl

[worthy orbit]

have you tried Isometric Cutaway Lora? https://civitai.com/models/69211/isometric-cutaway

looks pretty nice

[dim ember]

thanks 🙂 please make some creations with it on civitai and post them to the page I wanna see what you all can make

its my first lora I've made (that is not of myself)

I've made some 1.5 loras on my old pc that just didn't work at all, now I'm starting to understand things a little better

SDXL training seems to be pretty good without captions

https://civitai.com/models/299285

forgot to link it on civit

[uncut gull]

Cool 😊 But not really groundbreaking. I do like schematic generation, but I would love to be able to auto-detect blocks of messy text and plug the image into Llava to have it describe what the text should say, then replace the image text. No idea how to start, and too many other projects that are easier.

[worthy orbit]

No only for block test or schematics, but it adds good isometric 3D view to some generations. About the text I think it would be easiear to inpaint the text area until it has some nice desing and then try to write in Photoshop, the text you can generate in any LLM, usually I just don't care about the text if it is non-intelligible.

[uncut gull]

I mean like adetailer but for text. I don't want to fix the text myself, I want a tool that auto-fixes all text.
I wonder if segment anything can detect text blocks. Hmm.

[worthy orbit]

At some point I guess there will be something like that

This is also for sketches https://civitai.com/models/294229/futscribe?modelVersionId=330658

[cyan crown]

https://civitai.com/models/298188

[icy brook]

Nice subsurface scattering!

[cyan crown]

sorry what do you mean?

[icy brook]

How the light absorbs and reflects back through the skin. On the first image it’s most obvious. Look up the concept and you’ll see what I mean.

[cyan crown]

ok thanks

[hot bane]

what's the prompt for this? also what model

[hoary saddle]

is from the mspaint model made by @noble shoal ?

[shy kelp]

mspaint + apu apustaja on dreamshaper xl

along with downsampling to 256x256 in comfyui

[uncut gull]

The model is https://huggingface.co/dataautogpt3/ProteusV0.2. Its specialty is following prompts closely.

poster of body of a spaceship anatomy with labels, layers: hull, wiring, and skeletal support structure engineering technical blueprint Steps: 75, Sampler: DPM++ 3M SDE Exponential, CFG scale: 7, Seed: 3677720763, Size: 1024x1024, Model hash: 60e96ba05c, Model: SDXL-ProteusV0.2, Denoising strength: 0.61, Hires upscale: 1.25, Hires upscaler: 4x-UltraSharp, CADS Active: True, CADS Tau 1: 0.6, CADS Tau 2: 0.9, CADS Noise Scale: 0.25, CADS Mixing Factor: 1, CADS Rescale: True, CADS Apply To Hires. Fix: False, Version: v1.7.0

Wow, Llava really knows what's up. 😮

[hot bane]

thanks

[frail pawn]

The Kodama:

[idle linden]

Cat #✨｜sdxl

[gusty arch]

https://tenor.com/bnAAQ.gif

[feral minnow]

draw a girl

[crisp owl]

https://tenor.com/view/satisfying-gifs-oddly-satisfying-drawing-how-to-draw-drawbook-gif-22206075

#1047610792226340935

Usually bot channels for that

[merry tapir]

https://tenor.com/bSqKH.gif

[naive ginkgo]

My first proper AI image 🙂 Attaching the first original prompt so by proper I mean there are quite a lot editting - outpainting it into 16:9, inpainting some details, upscaling etc. I know there are some flauts even now and I know I made a lot of mistakes during the process which took me quite a lot of a time to repair. At least I have some new experiences for next generations 🙂

[dapper dragon]

https://tenor.com/view/john-c-reilly-motherfucker-angry-pissed-off-mad-gif-11627105

[foggy quartz]

superhero wheelchair model for dog

[molten gull]

hi guys

i haven't been active for some weeks

what did i miss?

[viscid crest]

generate a picture of sunset

[tiny frost]

picture of a greenhouse with large boulders and obstacles with different kinds of robots (legged robots, flying drones, crawling insect bots)

[twilit tundra]

Portrait of cute girl with red hair

[native knot]

Portrait of users who don't read #1047610792226340935 on the server.

[dim ember]

How did ya make the model house one those are cool

[worthy orbit]

"isometric cutaway"

[dim ember]

thanks 🙂

theres so many prompts you can possibly think of lol

[worthy orbit]

hey thanks, good Lora!

[dim ember]

Appreciate it

[dim ember]

Funny thing I am pretty sure I trained about 12000 total steps by accident

[worthy orbit]

All have incipient baldness xD

[dim ember]

I plan on making a Seinfeld 2.0 soon

[worthy orbit]

great

[worthy orbit]

I would love to train Lora but became lazy for all the work and the training

Also on SD XL

I trained some Lora on 1.5

[dim ember]

Doesn't take terribly long also if I wanna train something long I can turn it on before I go to bed

Gathering the data set is what can take a while

[worthy orbit]

yes, maybe I over think it or over do it, I try to give good datasheet and images

Then the Lora is deformed

[dim ember]

Yeah the first several times I made loras they didn't work

Sometimes when I make one it still doesn't work right

I got a Rick and morty style lora I made the other day I might release

[worthy orbit]

I'll should give it another take any time soon

[dim ember]

Look up sdxl lora training tips on reddit

There's a post or two on there somewhere that are real good

[valid delta]

oh

how to use this

[radiant harness]

girl

[meager canopy]

No, it's a tree

[molten gull]

morning guys

i've been away some weeks, what big changes did i miss?

[somber roost]

dream promted

[molten gull]

what does that mean?

[shy kelp]

wizard on edge of a cloud with round glasses and an elf hat

[blissful scaffold]

A dog run on

[meager canopy]

I don't think there has been anything major in the past few weeks, just the bots being down while being migrated.

[sweet summit]

a young French adult with Chanel hair, light brown eyes, French style clothes, serious expression, digital art, in the city, modern urban clothes style 2024, black cargo style pants, small size running jacket, black shirt under the jacket, small jacket, race small jacket style, jacket cropet style

[sweet fable]

Lel

[fast gyro]

but where is the exhaust?

[clever verge]

Other side obviously 😂

[fast gyro]

https://tenor.com/view/oh-you-you-pointing-gif-9478893

[carmine gale]

anyone know what this error is?

[uncut gull]

Why would a tron bike have exhaust though? Where is the lightstream?

[hearty ginkgo]

lol

[hexed charm]

I just put a default_negative prompt, but turns out it was in the "prompt" instead of "negative prompt"

so it created this abomination

when i tried writing "dog"

[rose aurora]

héroes, Vinieron a la tierra para seducir a los hombres de la tierra., ((La ropa interior también se puede ver en la falda.)))) La hermosa mujer en cuclillas frente expuesta., pecho grande y enorme, (((La ropa interior también se puede ver en la falda.)))

[south frigate]

Why sdxl turbo must have low cfg scale? Cant find in it google.

[rustic garnet]

CFG is a hack in general. It's surprising that it even works so well ^^ The reason why it works is probably because diffusion models somehow repair themself during many timesteps. But if you only have one or a few timesteps then this is not possible

it's quite obvious for one timestep:

the way cfg works is that it predicts two images (or more precisely, it predicts the separation of noise and image two times, but to keep things easier I will just pretend it would predict images). One time it predicts an image with the positive prompt and then it predicts an image with the negative prompt. Then it subtracts both images from each other and adds the difference on top of the positive prompt image

I think it's quite obvious that this cannot work if you only have a single step

[south frigate]

And how do you get exactly what you want from a prompt with such a low CFG value? It is very hard to get what you want if the prompt contains a little more description. @rustic garnet also, thanks for info about cfg

[remote saffron]

Hi, does anyone know how to get around training ema models for sdxl? Also which weights work the best?

[round lily]

sun day

[meager canopy]

It's MON DAY here

[fast gyro]

Yes it is. damn it lol

[clever verge]

I have no clear answer to give you but I know that bghira has built a trainer that does just that so there's probably things to learn from both the code and it's settings regarding EMA.
https://github.com/bghira/SimpleTuner

[uncut gull]

Mon day 'mon.

[rustic garnet]

it works quite well in most of the cases. For the 5% of the cases were low cfg doesn't work you just use a non-turbo model with high cfg and make a second img2img pass with a turbo model

[worn lark]

school

[restive anchor]

COW WEEK #23- 1st Place & Staff Pick- "Stable Pets"

[junior field]

Where did u create this images...this are too realll to be true 💀

[vernal mica]

aegis72's workflow using comfyui

[boreal bough]

finally trained a sdxl base -> dnd/fantasy character portrait lora

got almost every race working. lore accurate. all in the same style. fixed base sdxl eyes as well

even prompts not trained at all, still work c:

sry for spam XD happy to finally get this working

[cyan crown]

https://civitai.com/models/142352?modelVersionId=340839

[dire raft]

/pro casa

[vital ermine]

https://civitai.com/models/143415?modelVersionId=341092

[crisp jolt]

Cow

[peak dove]

Github is down??? At least here in London UK ...

[minor glacier]

Its up here, now, Midlands UK

[meager canopy]

Is this what is coming next in the bots (whenever they eventually return)? https://stability.ai/news/introducing-stable-cascade

Training scripts are already provided for training, fine tuning, Loras and controlnet

[polar ravine]

https://huggingface.co/stabilityai/stable-cascade/tree/main

[shy kelp]

generated some monstrosity

[smoky badge]

vram 20gb 💔

[meager canopy]

There are 'lite' versions of the models that may reduce this.

[shy kelp]

is dpmpp sde and dpm++ sde the same?

[raven bison]

https://github.com/Stability-AI/StableCascade
https://huggingface.co/stabilityai/stable-cascade/tree/main
Which model to choose? And how much vram does the 14 GB model consume? It is stable cascade.

[meager canopy]

Yes, the pp stands for plus plus

[meager canopy]

Have you got a UI that supports it? I guess it will take up 14GB of VRAM.

The larger sizes will likely only be needed for training.

[pseudo flint]

https://tenor.com/view/0001-gif-17282391190974969363

Hey guys

I'm new to generating images, how can i use SDXL?

[junior field]

Can u help me with it please...i didn't get 😅

[junior field]

Same question, please help

[limpid radish]

happy birhdays lisa

[minor glacier]

If your card supports bfloat16 it'll work at just under 15GB, and less if your throw the first model out of VRAM.

and before anyone asks it is broken wth float16 hence the bfloat16 mention.

[fierce hollow]

can also encode the prompts and unload te first, I got it working in under 10gb locally offloading the prior stage before decoding (with diffusers)

generations are actually pretty fast, and yeah, fp16 seems to produce nans

[minor glacier]

Yep, that what I meant with the throw the first model out of VRAM ( which I guess its really the second model 🙂 as its stage b ).
Shame about float16.
I really need to try out the lite version of the stage b model

[fierce hollow]

oh, I completely forgot there was a lite version

[minor glacier]

Yep, might be useful for peeps like me stuck with the float32 version for stage b

[manic plank]

Stable Cascade works on Google Colab (A100)!
https://twitter.com/AICUai/status/1757464834189443447
GitHub
https://j.aicu.ai/StableCascade

AICU Inc. (@AICUai) on X

Google Colab で試す Stable Cascade での新時代のテキスト画像生成

AICU media @AICUai #note https://t.co/f1Qts5K79F #GoogleColab #StableCascade

Twitter · Feb 13, 2024, 6:00 PM

[shy kelp]

Do you guys have any idea how to fix the teeth on dreamshaper xl? i always get bad teeth or mouth. 😦

[velvet shore]

I have just Made new stable cascade model demo anyone interested in testing?
https://huggingface.co/spaces/FumesAI/Fast-Stable-Cascade

[meager canopy]

Try adding "perfect white teeth" to the prompt. That helps if I add it to facedetailer prompt.

[shy kelp]

Will try this one. Thanks!!

[last abyss]

想象一个宏大的蒸汽朋克风格场景，五位财神各自占据着画面的不同部分，共同构成了一个多元且和谐的财富象征。中路的数字财神站在中央，手持发光的数字化账本，身边环绕着数据流的光纤和数字代码。东路的双碳财神被绿色植物、旋转风车和太阳能板包围，身穿融合了环保材料的蒸汽朋克服饰。西路的人工智能财神被复杂的机器人手臂、神经网络图案和AI芯片环绕，展示了科技与智慧的融合。南路的文旅财神身旁摆放着古老地图和旅行箱，身着带有古地图图案的斗篷。北路的元宇宙财神站在虚拟现实头盔和数字化风景中，体现了对虚拟世界的掌控。每位财神头顶的红色横幅上，以精致的简体中文书写着他们的名字：“数字财神”、“双碳财神”、“人工智能财神”、“文旅财神”和“元宇宙财神”。整个场景应体现为一幅杰作，具有详细聚焦、动态角度、32k超高清分辨率、最佳质量、专业摄影水平的高度细节和深度场景。

[vapid wave]

我想要一个教室的图片，里面有中国女学生

[wary raft]

#🌠｜show-and-tell Draw a skwtch of King Shahryar from the book of one thousand and one nights (Arabic Nights). He is a persian king, tall, handsome, regal and his age is maybe in a middle age.

Draw a skwtch of King Shahryar from the book of one thousand and one nights (Arabic Nights). He is a persian king, tall, handsome, regal and his age is maybe in a middle age.

[halcyon plank]

/settings

[elfin echo]

Draw a skwtch of King Shahryar from the book of one thousand and one nights (Arabic Nights). He is a persian king, tall, handsome, regal and his age is maybe in a middle age.

/how?

[uncut gull]

There is no bot AFAIK guys.

[soft wigeon]

/style

[shell stirrup]

/dreams

[cyan crown]

https://civitai.com/models/304822

[flint warren]

Why 30% of my generated images are completely blurred ? It's very annoying

[muted yarrow]

Is AUTOMATIC1111 otherwise called as stable diffusion or is that totally different? Sorry for the dumb question. I just started using stable diffusion and I see many here have also spoken about AUTOMATIC1111, so thought of double checking

[vivid lake]

#🌠｜show-and-tell

[hoary saddle]

anyone have a good sdxl workflow for combining 2 images into 1 using ipadapter and/or clip vision?

[wary willow]

i have from YT i think

check u like this or not

[hoary saddle]

ty kindly, will give it a whirl

[keen mantle]

night river

[amber lodge]

我想要一个教室的图片，里面有中国女学生

[vital ermine]

https://civitai.com/models/144527

https://civitai.com/models/150046?modelVersionId=343791

[flint warren]

Why 30% of my generated images are completely blurred ? It's very annoying

[hearty ginkgo]

send me a ss of your workflow

[shy kelp]

I’ve made a sdxl negative embedding https://civitai.com/models/305043/beyond-sdxl-negative

[frail pawn]

The HeartBreaker:

[oblique tartan]

Stable diffusion is the general term for the ai image generation. There are more nuance and detail. but Generally when people said Stable diffusion. they refer to the AI image tech. Automatik1111 or A1111 for short is the UI people use for generating image.

@muted yarrow Feel free to dm me if you have other question and i'll answer if i know it.

[shy kelp]

Cool

[uncut steeple]

Heyo, this channel is for sharing content regarding sdxl, please dont post random stuff

[slender wigeon]

A camel is coming out of a cave on a mountain which is surrounded by a desert area, this illustration should be in watercolor. This illustration is to be placed in the story of the nation of Thamud.

[meager canopy]

#1047610792226340935

[carmine gale]

anybody know what the best alternative to uDDetailer is for auto1111? Need segs (yolo8) and inpaint. Thanks!

[peak dove]

ADetailer?

[peak dove]

ADetailer?

[peak dove]

Won't work!

[velvet shore]

Yeah it was an error it's fixed and now it's working

[spiral bridge]

Idk if this is the channel to ask but any of u guys know any weapon loras or any for of way to get close to realistic looking weapons like handguns, ar, dmr etc

[dapper dragon]

check civitai

there some guns / poses with guns

[dim ember]

you guys try proteus .3 yet?

[vital ermine]

https://civitai.com/models/307795?modelVersionId=345479

[prime swift]

how do i generate imagesD:

[lucid rain]

#1047610792226340935

[worthy orbit]

Just noticed all the seam lines of the upscale, I used 16 mask blur and 64 padding, 1024 tiles, as I upscale by 4 (original 1344x768) with Ultrasharp, what settings to prevent the seams? I duplicate the base values as I was going x4, but maybe it doesn't help.

[naive ginkgo]

hi, could you give me some help with this error? I can see it's about memory, but can't understand what to do with that. I'm running the ComfyUI on RunPod with RTX 3070

[peak dove]

Keeps making the same image even though I change Seed/Art Style ... ?!

[velvet shore]

Yeah I know it happens with stable cascade you can increase or decrease the steps value by " 1 " or you can add or remove some words from the prompt then it will create a new image
Or you can try this one it has 6 models and it don't have bugs like stable cascade
https://huggingface.co/spaces/FumesAI/Best-Image-Models-Demo

[minor glacier]

Yes.

And one that's Proteus V0.3 with the LCM lora to speed things up.
Both images are straight T2I no extras (except the LCM LoRA on the second one)

[meager canopy]

Ultimateblend_xl v2.0

[peak dove]

Just been using IPAdapter+FaceID - installed via PINOKIO. Fun, for about 5 minutes. Then yawn (but that's just me!!!) 😄

Now installing AUTOMATIC1111 - which I have used for eons; but recently the extensions would not update; and a YAML file took 10 minutes to create before I could change the checkpoint!!! Hopefully, PINOKIO'll put me back in the fast-lane?!

[coarse lotus]

Im new here. Trying to find the right channel to post in. Im using SDXL and trying to use openpose to pose a nonhuman pixar style apple. Is that doable with any controlnets>?

[meager canopy]

I found SD.Next to be better than A1111. If you haven't tried it, it's normally more advanced than A1111.

[peak dove]

There was a time when checkpoints would load instantly; but now there is a 7 minute/10 minute delay to build a YAML file!!! Makes me give up 😄

[quasi iron]

how does using Ipadapter for the same image with the same prompt and same ksampler, turn it into this multiple faces

[cyan crown]

https://civitai.com/models/303619/misty-vintage

[worthy orbit]

Mm I tried SD Next for a bit (right after SDXL came), and it feelt really fast and with good options, but then I returned to Automatic after the SD XL update, and it felt good, then I tried SD next and I feel it was pretty buggy and broken, compared to Automatic

[delicate grotto]

Oi
Anybody heard about sora?

Just heard bout its announcement

[hasty smelt]

Witch is the model? And Upscaler?

[spiral bridge]

Guys I'm struggling to find a model but if I was gonna create very realistic exterior amd interior architecture what model would u guys recommend idm SD 1.5 not SDXL models

[rustic shadow]

happy snails 🐌

[dim ember]

Trying to train a bong hit lora since there's no model that can accurately do a bong hit correctly. Well see if it works

[slim raven]

a coastal city skyline against the backdrop of the ocean, with advanced desalination technology highlighted, symbolizing the utilization of ocean water to fulfill the city's water needs.

[dim ember]

heres some cherry picked results of my bong hit lora. a lot of smoking from the stem instead of the top. wonder how i can train the model with captions to always use this position instead of the stem

at least it seems to have a firm grasp of what a bong is no other models i've seen or loras seem to know what a bong is

im off to bed 🙂

[swift star]

#🤝｜tech-support

[meager canopy]

It uses a hires-fix script with lanczos upscale https://civitai.com/models/240590/ultimateblendxl

[mighty frigate]

Using SDXL turbo for quick first pass and feeding it into a higher sample model. Pretty pleased with the results

[icy brook]

Released Aetherverse XL yesterday.

https://civitai.com/models/308337/aetherverse-xl

It’s got great dynamic range, from deepest blacks to brightest, nuanced whites. Follows prompts well and is clean and sharp.

[quasi iron]

is it possible to use sd 1.5 lora with sdxl, I would so like to use some great ones out there, or is there some other work around

[next urchin]

what are u trying to do?

have u tried FaceGen AI ?

check this

https://www.youtube.com/watch?v=jjRFa1JugA0&t=130s

[quasi iron]

its not for faces, its for buildings and urban environments

also I work on comfy

[next urchin]

https://www.youtube.com/watch?v=VFl1VFCrB5M&t=2s

https://www.youtube.com/watch?v=oDCkxRV8IRs

[next urchin]

https://www.youtube.com/watch?v=4OHzY_dayeU

[quasi iron]

This one is with IP adapter, I think?

[next urchin]

i thought this one makes more sense, because the training on enviroments is much easier

[quasi iron]

ok

will try this out

[next urchin]

have fun mate

[quasi iron]

I am trying to stick with xl checkpoints cuz, I think overall there are much less errors without a lot of fine tuning

[next urchin]

in comparison to what?

[quasi iron]

1.5 models

I could be wrong but so far that what I saw in the comparisons

personal comparisons

[next urchin]

yeah iam still sticking to xl too

but trying out cascade later on

[quasi iron]

need bigger gpu for that thing

🥹

12gb vram wont cut it

hopefully they optimize it

[next urchin]

hehe

[next urchin]

@quasi iron u can downlaod some good urban env models here

https://civitai.com/search/models?sortBy=models_v5&query=urban

[quasi iron]

yes

[next urchin]

sdxl as base#

u can use them as refiner

[quasi iron]

ok, so xl models as refiners then?

[next urchin]

no, u use sdxl as base and them as refiners

[quasi iron]

ah, the 1.5 models as refiners?

[next urchin]

yyep

[quasi iron]

so the lora based on 1.5 wouldn't pose any issue then

[next urchin]

possibly not

ye

[quasi iron]

I will try this out

[wary willow]

My test with lora dissolve and linquviera, shark hybrid with human

[fossil turtle]

If you have comfyui you can already use a ~10GB lowvram version of cascade

[quasi iron]

ok

will try it then

[dim ember]

bong hit Lora is out https://civitai.com/models/310085?modelVersionId=348025

had the best results with zavychoma - trained it on sdxl base model

[ionic dragon]

@west breach can i dm?

[stray warren]

In front of a galaxy of hearts Emad The white haired Mad Hatter confidently with a flirting smile holding a heart close to his chest, his eyes filled with love and devotion. His hat is askew and his clothes are mismatched, but his expression is full of joy. (Bright, cheerful colors)

neg: painting, drawing, illustration, glitch, deformed, mutated, cross-eyed, ugly, disfigured

cyberpunk version

[west breach]

Of course you can 🙂

[dim ember]

dang apparently a lora that emultes bong smoking is not allowed on civitai mine got removed

[glad grove]

drugs bad alcohol good

[crisp owl]

And all the other stuff they allow on there 😂😂

Heavens forbid the evil bong

[vital ermine]

https://civitai.com/models/310682?modelVersionId=348716

[peak dove]

Using IPAdaper-FaceID (PINOKIO) I have swiftly tailored some pictures of a young woman eating cotton candy. Barbie-ish color scheme!!!

[bright valley]

Whoops forgot to tag you @peak dove 😇

[peak dove]

Cool!!! I see that Cascade is (almost) natively producing text?

[bright valley]

yeah it looks decent i wonder how it will do finetuned

unfortunately seems like the main upgrade lol

Seems a little odd to me to release an entire new base model that seems otherwise in the same league as the SDXL base model

but hey, not my chair, not my problem, that's what i say

[quick lake]

hey, why does my code not work? How can i load my checkpoint (https://civitai.com/models/271592/big-head-3dxl?modelVersionId=306137)
Code:

from diffusers import StableDiffusionPipeline
import torch

def generate_images(prompt, num_images=5, local_model_dir="/home/yago/bigheadsdxl/bigHead3DXL_v10.safetensors", output_dir="./generated_images"):
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    torch.cuda.set_device(device)

    try:
        pipe = StableDiffusionPipeline.from_pretrained(local_model_dir, local_files_only=True)
        pipe = pipe.to(device)
    except Exception as e:
        print(f"Failed to load model from local path: {e}")
        return

    for i in range(num_images):
        generated_image = pipe(prompt, guidance_scale=7.5)["sample"][0]
        image_path = f"{output_dir}/image_{i+1}.png"
        generated_image.save(image_path)
        print(f"Image {i+1} saved to {image_path}")

if __name__ == "__main__":
    generate_images("a sweet cat")```


Error:

Failed to load model from local path: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/home/yago/bigheadsdxl/bigHead3DXL_v10.safetensors'. Use repo_type argument if needed.

[meager canopy]

Not true. Try to reliably create this in native SDXL base at 2560x1536, no refining or upscaling.

[rustic garnet]

very different architecture from SDXL. Internally they probably test several different architectures. Of course they could just release the single one that works best, but why complaining about them releasing also the intermediates?
Furthermore, the way this architecture works might be interesting to play around with. I could imagine that it's much easier to add new features/plugins to the stage C sampler

[bright valley]

I wouldn't even if I could, I've not had a use case yet that would require it

[bright valley]

That's absolutely true, it's using the Wuerstchen architecture, which was the result of the methodology of this paper https://arxiv.org/abs/2306.00637

I may be incorrect but I believe there was an official implementation of this architecture in a model, and the 3rd version of that model is what became "Stable Cascade"

[rustic garnet]

yes, Wuerstchen is there since a while. They just trained it with more parameters and more data

it's a shame, though, that they keep the stage_c with such an extremely high compression factor

[cyan crown]

https://civitai.com/models/309044

[solar cipher]

Hello

[peak dove]

Very Route 66/Patricia M Highsmith. Also quite some Ed Hopper too. Nice series.

[dapper dragon]

Thank you @peak dove 👍

[steady grove]

thats cool research. more efficient training is obvioulsy the future. why spend energy you don't have to

[broken hare]

Anyone else trying stable cascade in comfyui, then switching over to the newly releases v9 of Juggernaut XL, using same prompt?

Listing for v9 says he will release a Juggernaut Cascade version before v10

[cyan crown]

https://civitai.com/models/311756

[dim ember]

tried making a couple renderings of a house drawing using controlnet

[low garden]

create the main page of a portfolio website for a web designer, in a minimalist style. ux. ui. ux/ui

[high skiff]

Does anybody know how many layers there are in the Down mid and Up blocks of SDXL? I am trying to train with custom block LR's, but I have no idea how many layers SDXL has

I can't find any info on it anywhere

[high skiff]

This is the only info I can find, but it doesn't detail the actual block weights for LoRA training

[dense mica]

1

[high skiff]

bruh lol

[copper kraken]

Got this anywhere else now that civitai nuked it? Ffs, it's legal for most ppl now, it's legal in my state

[dim ember]

if theres another place i can upload lmk

[copper kraken]

File.io allows up to 2gb

@median mulch since you were making creepy dolls and creepy faces...

[vale eagle]

You could print state_dict to see what does the unet have. I remember sdxl base has 0-9 in input, some middle and 0-9 in output.

[shy kelp]

i keep getting red error somethig happend on our end why is my dream buggin so hard

i have 590 credits i have restarted 4 times

[rustic garnet]

this image is accurat

the up- and downblocks in SDXL are not all the same

you have two up and downsamplers

for Lora you usually only care about the Transformer Layers, though

[high skiff]

I know that, I was just hoping to be able to see the layers of them like we can for SD 1.5. I wanted to do individual block LR training for my LoRA

[high skiff]

I was trying to figure out which ones were transformers, and how many layers in total there are in every block. You can't do block weight without knowing how many exactly there are. One value too man or too little and it crashes

[rustic garnet]

the down blocks have [0, 4, 4] transformers. The up blocks have [30, 6, 0] transformers

the mid block has 10 transformers

instead of changing the LR for blocks you can also just mess around with the alpha value

like assigning a individual alpha value per block or layer or whatever

maybe that's easier than messing around with whatever commandline tool you use for training

[tiny zodiac]

a house

@high skiff create a house

[high skiff]

@tiny zodiacWhy are you mentioning me, and why did you call me?

[tiny zodiac]

Oh I think you are a model

Sorry

[high skiff]

😅

I am very much not lmao

[tiny zodiac]

Hahah

How can I use SDXL

can you give me some suggestions

[peak dove]

Using Shaun2, Melting candles, and MoDernArt LoRAs - plus Reactor Restore face, and TurboXL_TurboVisionSuperfast checkpoint

[oak nacelle]

whats a good negative prompt?

[stable egret]

Thanks

[peak dove]

bad dream, text, closeup, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, pens, pencils

Cool idea!!!

[stable egret]

@smoky patrol

[stable egret]

thx

[slender coral]

Holup, there's a new sdxl?

[copper kraken]

that's cascade i think

[slender coral]

https://github.com/Stability-AI/StableCascade ?

[copper kraken]

yep, that

[cyan crown]

https://civitai.com/models/313318

[copper kraken]

that's great stuff

[peak dove]

Thanks to @hushed bobcat

[nocturne kayak]

hi

[stark sedge]

Having no negative prompt

[meager canopy]

These are Cascade! Sorry, wrong room 😄

[shy kelp]

Y'all are so good a prompting, i suck.

[rustic garnet]

woah, that's beautiful!

[cyan crown]

thanks

[native knot]

Nightmare at Freddy Mercury's

[native knot]

Yoda the Hutt

[plucky crow]

Is there any reason to finetune on SDXL instead of Cascade?

[broken hare]

more accessible to more people at the momment.....fooocus and folks with modest graphics cards

but i do want to see cascade fine tunes

[warped sedge]

https://huggingface.co/ByteDance/SDXL-Lightning the 2 step model is giving sub second generation on a 2080ti

[rustic garnet]

is Cascade really better than SDXL? I'm not sure yet. For me the images from Cascade look to "midjourney-like", probably overfitted on MJ images. Compared to base sdxl it might look a bit better, but compared to custom models cascade looks much worse.

for me cascade might become interesting when we play around more with the stage c layer. I could imagine that it's easier to train fancy controlnets or other kind of conditionings on it.

[west breach]

I have trained PixelWave from scratch with better captions. It should now be more flexible and adhere better to artists and styles from the base SDXL model. Less bias towards photorealism. Please check it out 🙂 https://civitai.com/models/141592/pixelwave?modelVersionId=353516

[wanton meteor]

a girle

[peak dove]

Shaun2 and Ethereal Grace LoRAs and again my thanks to @hushed bobcat for the prompt!

[scenic dove]

I have trained PixelWave from scratch with better captions. It should now be more flexible and adhere better to artists and styles from the base SDXL model. Less bias towards photorealism. Please check it out 🙂 https://civitai.com/models/141592/pixelwave?modelVersionId=353516

[peak dove]

How do I d/load this repo WITHOUT all the large files? I see something about "prepend GIT_LFS_SKIP_SMUDGE=1" but my senses are too dull to make much sense of this? 🙂

[unkempt fulcrum]

i got this error when i get first sample of train
UnicodeDecodeError: 'utf-8' codec can't decode byte 0xe9 in position 252: invalid continuation byte

[unborn warren]

An saphire and amethyst skeletion watch with black chrome and rose gold accents and tastfully inset black diamonds and a black dial, compass, crystal digital display, and mc esher infinity inspired watch band. The watch is on a red silk surface. 8k splashscreen hyper realistic, realistic, detailed, sharp, hour hands, luminesence. v5 3:5

[woeful patio]

hey guys, I've been gone for a while. Is kohya still the best way for making loras or is there some new program that makes it better?

[peak dove]

I have 43 LoRAs inside my A1111 LoRA folder - ComfyUI shares this folder - and can see ALL of them; but A1111 only sees 37 of them?! Why, so, because?

[warped sedge]

i downloaded a full 2 step checkpoint, not the lora and used their workflow

[peak dove]

A1111 NightVisionXL Checkpoint, Futuristic Cubism and Liquid Flow LoRAs

[glad fulcrum]

anyone knows how can I use a controlnet video in deforum without being affected by my camera zoom?
It generates zoom because I want it but it applies the zoom to my controlnet video as well which I don't want.
I want my controlnet to stay fixed

[peak dove]

SDXL Lightning!!!!! I cannot believe the speed!!!!! Or the kwallittie.

Just 4 iterations, NightVisionXL, and the 4-step LoRA

[civic gyro]

https://twitter.com/HikariUchu/status/1760364229377630403

Hikari Uchu (@HikariUchu) on X

Explore the intricate 'Symphony of Life' where Flora, Insects, and Fungi weave a complex web of existence. 🌸🐝🍄 A world where every organism contributes to the planet's harmony and balance. #NatureSymphony #EcosystemInterconnectedness #EarthJourney #nature #ai #aiart…

Twitter · Feb 21, 2024, 6:01 PM

https://twitter.com/HikariUchu/status/1760358793811210323

Hikari Uchu (@HikariUchu) on X

Embark on a journey beyond the stars with the dawn of the Space Yachts era 🚀✨. Luxurious voyages across the cosmos, where elegance meets exploration. Ready to sail the universe in unparalleled style? #CosmicVoyagers #SpaceYachts #EarthJourney #ai #aiart @stabilityai SD Model:…

Twitter · Feb 21, 2024, 5:40 PM

is this the right channel to share some XL art? 😄

im new at all this

[shy kelp]

One message removed from a suspended account.

[onyx gate]

Hi Guys, question how do you generate a proper image with a text? Im using the api, what are the configurations should I check?

[native moon]

is it possible to train textual inversion for sdxl models in automatic 1111?

[dense mica]

1

[weak elk]

现在这个在哪里生成呀

咋找不到频道了

[fossil shale]

H

[night goblet]

best face i've gotten with turbo

[ionic gulch]

InstantID + IP Adapter:
https://i.imgur.com/4d2CBBC.png
https://i.imgur.com/iatX7WF.png
https://i.imgur.com/OAOrWze.png
https://i.imgur.com/G32afhl.png
https://i.imgur.com/s1DBq0w.png
https://i.imgur.com/FNw2vFa.png
https://i.imgur.com/gm0KZ7N.png
https://i.imgur.com/fPhZwZV.png
https://i.imgur.com/azOzrta.png
https://i.imgur.com/ItX1uCJ.png
https://i.imgur.com/yDszuQK.png
https://i.imgur.com/L25G8nU.png
https://i.imgur.com/wVHUgik.png
https://i.imgur.com/lRM8JoU.png
https://i.imgur.com/czTTW7Q.png
https://i.imgur.com/P0WASL3.png
https://i.imgur.com/yPwHxbp.png
https://i.imgur.com/JMOInZY.png

[chrome flicker]

Did anyone can keep the same anime character face using instantID ?

[broken dove]

Do you have a workflow to share?
I tried InstantID but feel that it's less flexible compared to just IP Adapter. All generated pictures will have the same face position and pose unless you run it through a controlnet with a different face position. I miss the flexibility of having the model generate the person and face and then IP Adapter controlling the face features to make it look similar to the reference.

[ionic gulch]

@broken dove the workflow includes unreleased nodes, so it wouldn't be very helpful. hope to finish the nodes + workflow next weekend, so i can upload both to github.

there is a separate input image in instant-id for the angle of the face (separate from the face you want to transfer). so you can control that part easily. the position and size in the generated image depends on the face position and size of that secondary input image, so you can control it by padding/cropping the input image. overall it's quite flexible, but needs a good amount of preparation, especially for the support image, that controls face size, position and angle.

my images all use the same support image, that's why angle, size and position are almost identical in all the images i've posted. the setting and other details come from ip adapter (different for each image) and text prompt (same for all images).

[broken dove]

I do agree that it's certainly possible. I just enjoy the flexibility of IP Adapter that just generates faces in different angels and positions freely. I guess one could run a 2 step approach.

1. Generate image that has the composition you want with a random person
2. Use that image as the second input to control the angel and position of the face with instant-id.

Doable, just a few extra steps that would be nice to skip 🙂

[peak dove]

SDXL Lightning - TurboVisionXL_Superfast, 8_Step_LoRA, MoDernArt_LoRA, Shaun2_LoRA, euler/sgm_uniform, 8 Steps, CFG 1

[ivory aurora]

/prompt

/prompt The huge stone statue of Du Fu in the mist-shrouded Taishan Mountains

[fossil shale]

Create a photo: Long shot photo of a Bangladeshi beautiful girlr ests beneath a coconut tree by the seashore. The gentle waves lap against the shore as the evening sun casts a golden hue over the scene. In this serene moment, she finds solace in the tranquil embrace of nature, allowing the sea breeze to caress her soul and the rustling leaves to serenade her senses.

She is adorned in a stunning Salwar Kameez,

draped flawlessly, with intricate designs and vibrant colors that catch the light beautifully,

She is an Aesthetic, she has an inverted triangle body shape, she has a tear drop shape 38 size of chest and 28 waist, she has long, black, wavy hair,

Looking as Humanly Possible, Fashion Photography, Magazine Cover Photography.The photo should be a masterpiece, in UHD (Ultra High Definition) quality, 32k, high regulation, altra realistic, captured in natural light, and must look realistic and natural. There should be no worst quality, low quality, blur, missing tooth, missing ear, double hand, double finger, watermark, illustration, 3D, 2D, picture, cartoon, or sketch.

[sweet geyser]

/prompt The huge stone statue of Taylor Swift in the mist-shrouded Taishan Mountains

[hot river]

Anyone here run sdxl on a 1080ti?

[frail pawn]

God Emperor of Dune Concepts (based of Blender Render (Img2Img)):

[peak dove]

New SD3 anybody?

[brave gazelle]

+1

[stone cove]

/prompt

/prompt /prompt The huge stone statue of Taylor Swift in the mist-shrouded Taishan Mountains

[timid gate]

/prompt a young girl run at the seaside , while there is a monster behind her/

[random trench]

1

/1boy

[autumn patrol]

/prompt a young man

[lament meadow]

有人吗

[uncut gull]

Will SD3 be able to do this? (SDXL obviously cannot.)
A colorful pie chart with the labeled ratios: 5%, 20%, 75%

Will SD3 be able to do this? (Another SDXL fail.)
From left to right: A shiba inu, a husky, and a gray wolf. The shiba inu is taller than the husky. The gray wolf is taller than the shiba inu.

Will SD3 be able to do this? (Another SDXL fail.)
Six dice on a table, showing each face from one to six pips.

(Hint. Hint. SAI please let me test SD3. 🙏 🙏 🙏 )

Will SD3 be able to do this?
An antique clock showing the time three-thirty.
(SDXL attempt.)

Will SD3 be able to do this?
A hand, and its shadow on a wall. The hand is doing a thumbs-up, but the shadow is doing a thumbs-down.
(SDXL fail.)

Will SD3 be able to do this?
A grid of boxes with three columns and four rows.
(SDXL fail.)

[uncut gull]

Will SD3 be able to do this?
A hand fanning out five playing cards. Each card is different.
(SDXL fail.)

(Obviously all these prompts require advanced reasoning from the model. I am not trying to make any model "look bad", I am very interested in searching for any emergent behaviors the researchers may not have noticed. SDXL learned to count "things with faces" up to 3, and Cascade learned to count things in general up to 4 or 5.)

[surreal sundial]

I think the best way to do it would be generate an image with correct proportions using Excel/graph stuff. then use that as a "canny" controlnet extension, that's what I'd do atm, but I think SD3 might stand a chance by itself if the hype is anything to believe

[uncut gull]

Yeah, of course. I'm not intersted in making graphs, I'm interested in SD3's emergent behaviors.

Will SD3 be able to do this?
A sign beside its reflection in a mirror. The sign says: OKAY
(SDXL fail.)

[surreal sundial]

It think you're more interested in limitations than 'emergent behaviours', but I don't think SD2 was ever meant to be that great at really understanding a prompt, it just had the customisable everywhere else. I think from what we've seen so far "SD3's" focus appears to be on better semantic understanding than anything else

But I think the answer is yeah, it'll be iterative, I'm sure it wont be perfect, but im 100% it's going to be mile better than 2 right 😄

[uncut gull]

Emergent behavior is about generalization from specific examples to abstract reasoning.

SDXL's training data included "three people" and "three puppies" etc., but from that pattern emerged the generalized capacity to count things with faces. SDXL can understand the prompt "three Einsteins", which was not in its training.

I'm looking for emergent behaviors including positionality (inside/above/left/right/behind etc.), physics (light color blending, shadows, reflection, fluid conformation, etc), comparison (shorter, wider, darker, different/same etc.), variable binding (e.g. "the woman is taller. The taller person has a hat."), and more. That's what I mean with these posts. 😊

[meager canopy]

According to the examples it can. There's no point asking until somebody has access to it to verify that it's not cherry-picked images.

[uncut gull]

I know I'm just so impatient. And to be clear: It won't have every behavior associated with abstract reasoning. It might understand "behind" but not "inside", which would be both amazing and so sad. I just really want to know what it can do!

[meager canopy]

Apparently, from what I've seen, it does

[uncut gull]

Will SD3 be able to do this?
Two hats that are the same. vs. Two hats that are different. Is SDXL beginning to show some emergent understanding of same/different?

No. It is not.
Prompt two pokemon that are the same vs. prompt two pokemon that are different.

[native moon]

is it possible to do texual inversion text embeddings for sdxl in a1111

[astral path]

does anyone know a good way to automatically search for generation failures in images? e.g. malformed hands etc

[rigid laurel]

I don't think so, because it's very hard to automate deciding whether or not something is "good enough"

if you look hard enough, there will always be problems

[uncut gull]

Adetailer. (I don't use it, so I don't know much, but it works by just fixing everything, not fixing broken stuff.)

[astral path]

sure, just would like to filter out at least the ones with obvious mistakes, like hands that are misformed etc... cogvlm/agent requirements are a bit high

[civic gyro]

Hope you all having a kickass Friyay - https://twitter.com/HikariUchu/status/1761111069244994046

Hikari Uchu (@HikariUchu) on X

Step into 'The Hidden World Within Trees' revealing nature's resilience and beauty.🌳🔍#TreeMagic #NatureUpClose #trees #自然接続 #natureconnection ＃自然 #nature #ジェネラティブアート #generativeart #ビジュアルアート #visualart Animation: @stabilityai #stabilityai Model:…

Twitter · Feb 23, 2024, 7:29 PM

[rigid laurel]

I get the intent, I still think you're underestimating how hard it is to parameterize an "obvious mistake". You could look for specific problems, like malformed hands, but how do you tell the difference between an image with no hands in it and an image with hands warped so far they can't be recognized?

You could do somethink like rigging the image output up to a controlnet pose preprocessor and filter outputs so you only get images where it could recognize a face and/or hands

You can also use aesthetic scoring to filter out images that just -look bad- in very general ways that aesthetic scoring models have been trained to pick up on

but matching general 'mistakes' is a problem that is hard to formulate

[supple knot]

What is Pinokio, is it like A111? Why do you need it to run Cascade

Do you get same results running Cascade in ComfyUI and are the outputs worth it ?

Is Cascade a new process?

[rigid laurel]

cascade is a whole new type of model

[uncut gull]

SDXL Lightning 9 very high quality images in 16 seconds.

And the prompt following is still excellent. 4 steps with Euler sampler. a cheetah kitten wearing a racing helmet, sitting at a computer desk with his mouth open in astonisment, shock, and awe; with lightning exploding in the background <lora:sdxl_lightning_4step_lora:1>
This 100% replaces LCM and the LCM lora.

No text though.

[silver oxide]

你好

[chilly dirge]

a car parked in a field with a sunset in the background, abandoned vehicles, nostalgic photography, magic hour photography, golden hour photography, orange sunset, wlop |, looking out at a sunset, rusty vehicles, during a sunset, by Marcin Zaleski, car, car, transparent black windshield, colorful sunset, car commercial photograph, old car, vibrant atmospheric colors

https://t.me/ai_generate_image/935

[jolly brook]

@smoky patrol

[daring smelt]

boop

[uncut gull]

It's a cozy Saturday morning.

[dense tendon]

fashion photography vogue style

[vital ermine]

https://civitai.com/models/319667

[atomic bloom]

anyone have knowledge about forge ui? I am unable to load controlnet fast load extension?.... what to do

[vital ermine]

https://civitai.com/models/319826?modelVersionId=358631

[smoky patrol]

So cute!

[uncut gull]

I tried tiling it, and this is actually seamless. 😮
seamless tiling texture, soft gray noise, paper texture 4k

[copper kraken]

Whoa

[uncut gull]

It's an app background. 🙂

[steady grove]

noise is easy to make look seamless

recent vid showcases what i mean. even if the noise is structured around a line it fades away immediately. https://youtu.be/TdTMeNXCnTs

[timber terrace]

/create wide close view from top shot of an african american women face with wide open mouth ,taking a drag of air, very wide open eyes looking to a side, oil painting, loose brushes, expresive ambient -

[sly gust]

/create wide close view from top shot of an african american women face with wide open mouth ,taking a drag of air, very wide open eyes looking to a side, oil painting, loose brushes, expresive ambient -

[copper kraken]

Here is the image you requested.

[peak dove]

Those spiders just erupt outta nowhere?!?!?

[copper kraken]

lycoris-ia3 blows lora out of the water for likeness, wow

[spare mirage]

/create a mother

[strange niche]

expressive ambient, indeed.

[stark basin]

/create a beach

[final notch]

here's ya pic, mate

[stark basin]

/without clown Please

[copper kraken]

Here is the image you requested.

[final notch]

aight, I put clown in negative prompt for you

[stark basin]

thanks

/make this photo beautiful

/make this photo beautiful

[red tulip]

/create a tiger roaring in wildlife

[copper kraken]

Here is the image you requested.

[cerulean dew]

/create a human "i'm sorry man"

[stark basin]

/This is not a human, please create a human

[clear hound]

aight, I put clown in negative prompt for you

[shy kelp]

One message removed from a suspended account.

[cerulean dew]

Prompt: Anything

[queen oar]

Is this a bot?

[final notch]

no

[queen oar]

Ha! Sorry.. green noob here

[eager tide]

sick

[lapis parrot]

There was a lot of talk about there not being a controlnet tile for sdxl earlier, but now there is and no one is talking about it? Is it fake, or what am I missing

[copper kraken]

what it/s are people getting for 1024x1024 with sdxl with a 4090? i'm maxing out at about 1.3 it/s with onetrainer, slightly less (1.1-1.2) with kohya

[copper kraken]

Here is the image you requested.

[copper kraken]

a1111 really is something else... 22.3 GB running batch size 3 with 1024x1024 lmao

[copper kraken]

many of you are probably aware but wow, combining a couple lycoris-ia3 and a couple loras trained on the same data in different ways, then playing with the weights, gives a stunning level of resemblence and consistency

[icy solstice]

/create

[final notch]

Here is the image you requested.

[icy solstice]

Hahahaha

Was a test, dont know how it works

[final notch]

there's no bot

[icy solstice]

Ibsee

So how would i be able to do things?

[final notch]

run sdxl locally, use any of the hosted services, or wait until they've set up the new bots for sd3

[copper kraken]

Maxing out around 1.3 it/s with a 4090 and 5950x CPU... Seems the GPU isn't running at full load

One of the cores for the CPU does however appear to be hitting 100% pretty often... Not using official monitors here just hwinfo, but wondering if this really is a CPU limit or not

I'm willing to build anotder PC to house the 4090 if it'll double my training speed by going from <50% load to near 100 on the 4090

I'm on windows 11, 64gb ddr4 ram @ 3600mhz

This is with OneTrainer which I've found to be a bit faster than kohya which has me at 1.1 it/s max

[tawny yacht]

/create a Stephenson Rocket Steam Locomotive

[civic gyro]

Good morning 😄 https://twitter.com/HikariUchu/status/1762179443093869035

Hikari Uchu (@HikariUchu) on X

A visual odyssey from the simplest of organisms to the complexity of the human brain. 🧠 #EvolutionJourney #nature #ai #aiart @stabilityai @Lykon4072 #dreamshaperxl #stabilityai ＃アニメーション #animation ＃デジタルアート #digitalart #美人 #beauty #hikariuchu #visualart #vj…

Twitter · Feb 26, 2024, 6:14 PM

[steady grove]

With the sd3 hype coming up, i see a lot of people saying sdxl's training set had no censor filter on it. I don't believe this is true. Can anyone confirm?

[night goblet]

4 steps

[jolly brook]

@daring smelt 👆

[daring smelt]

tyyy

[pallid path]

⚠️ Light Gore ⚠️

open original and zoom into it

hopefully with SD3 (and its finetunes) in the future, highresfix will make them near perfection like this

I'm so excited for the future

⚠️ Light Gore ⚠️ (Michael Jackson as a zombie)

[keen nimbus]

Etc ethereum classic is rocketing

[copper kraken]

Here is the image you requested.

[minor pecan]

https://tenor.com/view/cucumber-whirl-gif-24703408

[sleek pumice]

GM!

[woeful patio]

is there a way to put 2 lora characters in the same image yet or still difficult?

[meager canopy]

I think it's possible, but still difficult.

[woeful patio]

damn hoping on some improvement on that in sd 3 but I doubt it

[meager canopy]

Prompt following in SD3 looks much more controllable, so you may be in luck. Have to wait and see 🙂

[sleek pumice]

or use regional prompt and loras, is not problem

[woeful patio]

I'll look into that, thanks

[sleek pumice]

in automatic1111 regional prompt is much easy than in comfyui, but in comfyui u have more control 😉 segmentation, face/body/clothes detection and inpaint with lora etc

[woeful patio]

is it supported for sdxl? my output looks like this:

[sleek pumice]

i used it 4 mont ago in auto111, now im use only comfyui and it works

or... get my WF to comfyui (without regionasl) is in last image

[woeful patio]

could you share your comfy workflow? a1111 is really unkind to people with low vram

[lyric finch]

which laptop to buy for "stable diffusion"?

[woeful patio]

nvm I figured it out, had to change the mode from attention to latent no it works fine, thanks for the help

[sleek pumice]

not laptop 🙂 or only to use i connector via www to PC 🙂 you need GPU with 12 gb vram or more

[shy kelp]

I use a GPU with 4gb and it takes forever to create anything lmao

[umbral lichen]

is this normal

it looks bad

[minor pecan]

[woeful patio]

what is the state of video outpainting I'm looking to outpaint a fairly easy video. Is there any way to do that?

[ancient badger]

How to fix?

[mellow tendon]

I know Adobe showed off Video Inpainting a few months ago, I don't know if it is released yet? I haven't used it, but it did look good: https://www.linkedin.com/pulse/aiyou28-adobes-project-fast-fill-revolutionizes-video-greggory-elias-fwfze/
I haven't used much open source for video apart from AnimateDiff.

[ancient badger]

help me pls

[mellow tendon]

Wrong VAE most likely.

Try setting VAE to "None" or "Automatic".

[uncut gull]

Juggernaut SDXL Lightning works with control net. 😮

[shy kelp]

One message removed from a suspended account.

[pallid path]

wow these are great, are the first two images done with some noise offset lora or something?

[bright valley]

The best laptop for Stable Diff is a PC

[sudden osprey]

chromebook ^

i heard they have rtx 5090 ti in them

[misty oar]

change the color

[lyric finch]

merci

[loud cedar]

cute bee robot

[slate wagon]

cat

[upbeat hill]

sheep dance with wolf

[finite nimbus]

Juggernaut

[low creek]

Make a logo for the human centered ai lab called HCAI

[tardy moss]

https://huggingface.co/stabilityai/stable-cascade/tree/main/comfyui_checkpoints

anyone tried using new comfyui files? I get this error:
KeyError: 'conv_in.weight'

[meager canopy]

They are Cascade models, you're in SDXL here

[eternal fog]

Has anyone had a play with those Koala models?

On the commit it says use the UNET Loader to load it in, but it doesn't say how to use the Text Encoder, I've tried to load it as clip but it doesn't work. I can use a normal SDXL Text Encoder, but the output looks borked.

[fringe prairie]

does forge ui handle sdxl checkpoints well?
i am running on 4090 but getting a lot of bugged images or no output

[eternal fog]

Ok you have to load both clip models, and it looks pretty much exactly like the base SDXL Clip model anyway.

[grizzled viper]

So, in both A1111 and Forge I am using the newly released Tile model controlnet, and doing an img2img with output size 3x and denoise of about 0.5 I'm getting good results.
Just wanted to see if anyone else is going it like this... in the past I always used Multidiffusion in conjunction with tile mode, but it doesn't even seem necessary now.

[mellow tendon]

Is there a new Tile ControlNet for SDXL? I have been waiting for one for 7 months now, please say it is so....

Or are you using X-Adapter to use SD 1.5 ControlNets with SDXL? still seems like a solution.

[grizzled viper]

There’s now an SDXL tile model

[uncut gull]

It's the best thing that's ever happened to SDXL! Came out 5 days ago. Right in time for SD3. https://huggingface.co/bdsqlsz/qinglong_controlnet-lllite/tree/main
(Most recent one in the list.)

Now we can spend the next 7 months waiting for an SD3 tile controlnet.

[copper kraken]

WHAT

how did i not know there's SDXL TILEEEE

great. huggingface is down. now i'm going to die waiting

[balmy lance]

What's tile controlnet?

[mellow tendon]

It allows for more consistent tilted upscales, so you can upscale images to 4k or 8k with a higher denoising strength and more detail added, than you are able to without ControlNet Tile.

[copper kraken]

so has anyone tried it yet?

[glad grove]

yes i tried it with ultimate upscale and it sucks

[copper kraken]

oh. wonderful lol

[glad grove]

try it yourself maybe u can make it work

[copper kraken]

huggingface has been down for a bit so yeah, whenever it's back i'll look

[native knot]

Yeah... maintain for a bit.

[slow halo]

Hey @loud briar could u share a workflow? Looks great! Thanks

[earnest trellis]

Where are those bots to create sd1.0 images?

[grizzled viper]

There’s a message on here somewhere that they are all offline

[earnest trellis]

Oh so now those aren't available

[grizzled viper]

#🗣｜artisan-support-feedback message

[toxic sinew]

[sage lion]

Anyone have a workflow to using the new SDXL tilenet with SDXL Lighting ?

[limpid sphinx]

I am quite new here, is there a way to see what prompts have been used for the above images?

[rough meadow]

gril

[rustic garnet]

that happens all the times. Every few months somebody releases SDXL controlnets and they never work -_-

the only controlnets for SDXL I found useful so far were the ones from SAI

[mellow tendon]

If people have left the prompt information on the image you can download it by clicking "open full image" (or something like that) and download it and drag the image into ComfyUI or the prompt box in Automatic1111.

[granite vigil]

A horse balancing on top of a colorful ball in a field with green grass and a mountain in the background.

[spice fox]

ss

[oblique gust]

Japanese fresh style. A 16-year-old Japanese girl wearing a navy uniform is dancing

[vital ermine]

https://civitai.com/models/328075?modelVersionId=367672

[vital ermine]

https://civitai.com/models/328415?modelVersionId=368029

[minor pecan]

https://tenor.com/view/neuron-activation-neuron-gif-1455356329209493514

[prime bison]

/create a photorealistic closeup of an unsettling cockroach with hairy spider legs and long sharp teeth leading a flood of the same that fill up the entire canvas

😊

[cunning cradle]

how does /create work? is it just a request for someone to fulfill?
Just curious cuz I have a 4090 and not really using it right now, so if anyone wants to borrow it(processing power) I can do that

[eternal fog]

I think he's just memeing, /create doesn't do anything at all

[cunning cradle]

well my offer still stands, feel free to dm me ill do any prompts

[cunning cradle]

bro out here posting some furry bait and im all for it

[wet nacelle]

Yes

[glad grove]

https://tenor.com/view/freddy-freddy-fazbear-fazbear-fazbear-pizza-place-freddy-hitting-gif-5328662341160349848

[eternal fog]

And when it goes wrong lmao

[cunning cradle]

does anyone here get ducks when they type in jibberish?

[eternal fog]

Depends what you mean by jibberish

[cunning cradle]

prompt: FPKOEWNFOKEWNFOWJNFOIEJSFNE OWfnn OFNWONFO NWONFO NWF , MFOEWNFOEWF, eOFM W<FOPW<F WOPF< WOEPF<SDP<F<f,,,,e FIWOMFOWF< WOFMOWPNMFOPFNf871þç²i1+3δ1a651↓65[í

[eternal fog]

I just got a woman, but I'm using juggernaught, which goes that way anyway

[paper phoenix]

Thats not gibberish, its how ducks type

[cunning cradle]

well technically half of that was my cat, so ducks type similar to cats, apparently. teh more you know

[paper phoenix]

But yeah if theres nothing for the model to find among training data related to the gibberish itll fall back on whatever its finetuned towards

[cunning cradle]

this is sdxl base with default settings, no refiner either, still its funny I kept getting ducks even after 5 tries

[errant iris]

/create a girl

[meager canopy]

https://tenor.com/view/weird-science-kelly-le-brock-80s-eighties-the80s-gif-24111662

[rose cosmos]

https://tenor.com/view/prairie-dog-stand-gif-13202825

extremely stupid question, but how do i get sdxl to generate me pictures of a prairie dog and not a dog in a prairie? thanks

like damn

[azure oxide]

you could try prairiedog or prairie_dog or prairie-dog

could also try negatives of dog or golden retriever or any other dog breeds

[rose cosmos]

thank you

[rose cosmos]

woaw!

amazing!

happy

[hushed bobcat]

Elements

[gray ember]

perfume for spring

[uncut gull]

A word fitly spoken.

[proven isle]

is this SDXL 1.0?

[uncut gull]

Everyone uses fine-tunes these days.

[proven isle]

what does that mean

[uncut gull]

No one has used SDXL 1.0 since the first week it was released. Everyone uses fine-tuned variants from Civitai or Huggingface.

[proven isle]

ah right

yeah same lol

[uncut gull]

Oh no a shower pic. Might get modded.

[uncut gull]

The sampler you use makes a difference, you know? LCM for straight lines.

[rose cosmos]

woaw! amazing

[civic gyro]

https://twitter.com/HikariUchu/status/1763697057511207295

Hikari Uchu (@HikariUchu) on X

Dive into 'Cosmic Evolution' and discover how our aquatic heritage might illuminate our path to the stars 🌌🌊 #AquaticRoots #SpaceFuture
#nature #光宇宙 #HikariUchu #人工知能が生成 #artificialintelligencegenerated #ai #aigen #aiart #インスパイヤ #inspire ＃絵画 #painting…

Twitter · Mar 1, 2024, 10:45 PM

[wet nacelle]

fizzy fizzy

[shy kelp]

How did you do this?

[hushed bobcat]

Image to image. The original was close to the pink cherry blossom. I very roughly added green, blue and red to the leaves for the initial seed images and used different prompts on each.

[chrome silo]

https://tenor.com/bSl6n.gif

[short roost]

a cute cat in the big city

[violet zinc]

/create a car

[copper kraken]

Here is the image you requested.

[copper kraken]

Here is the image you requested.

[eternal fog]

lol

[violet zinc]

Create a image of a car being washed that is transformed from rusty and old to brand new

[copper kraken]

Here is the image you requested.

[nimble palm]

make a Photo of a soilder

[copper kraken]

Here is the image you requested.

[nimble palm]

Make a evil soilder pointing a gun toward an innocent kid and her mother waiting for Aid, the overll inviroment should be war and chaotic

[copper kraken]

Here is the image you requested.

[prime bison]

You forgot the soil

[humble juniper]

any1 had a problem before were you can't load loras using comfyui?

name displays as undefined

[gloomy lark]

hah, i asked for a giant mechanized beetle crawling up the size of a building.

[balmy coyote]

https://creator.nightcafe.studio/?ru=mick

[prime bison]

What Arron said. I was making fun of people. Also, you should check out AI Horde if you wanna lend out your GPU. See #1080946152318443610

The Ikea Lora is great

satanic Donald Trump demon and Satan Claus

[gloomy lark]

hah

good stuff on the deepbluexl for mecha. now I have to try and think of all the stuff to try.

[gloomy lark]

the list of parts is the best part

[prime bison]

I tried asking for an Allen wrench because it keeps putting them in, and got this 😦

[gloomy lark]

can never have enough tools

[wintry raft]

OMG dunno why, you made me think of Saroumane at the end of the novel in the third tome, when he wanders across the Shire in a kinda frightening "wain". I could totally see Saroumane and Grima on the roof !

[violet lantern]

a guy that worked as a plumper back in the 80’s. He preferred to wear a brown moustache. He also had a bit big belly. He lives in his blue carpenter’s pants and his red flamboyant pirate shirt that shows a bit to much breast hair. He always wears that grandpas hat that matches the blue on his pants. He is struggling in those big green pipes that just goes on forever. Some evil turtles sitting on the clouds and throwing stuff at me all the time. But now I am happy because I just found a MASSIVE magic mushroom which most probably is psychedelic. My HOT girlfriend who is the Princess and absolutely beautiful with great shapes and wears a pink robe that you almost can see through it. Her hair is platina blonde and her tiara is shining as gold and the diamonds in the tiara really lights up in different colors. It’s an hour before sunset and they both hang around in Marios mushroom-looking house. Mario is the main character on this frame, don’t ignore anything I have asked for.

[formal vigil]

where

[lofty cargo]

#🌠｜show-and-tell Afternoon, bright sunshine, in the living room, a boy (reading), (rocking chair)

Afternoon, bright sunshine, in the living room, a boy (reading), (rocking chair)

#✨｜sdxl Afternoon, bright sunshine, in the living room, a boy (reading), (rocking chair)

[hollow ether]

#🌠｜show-and-tell paiting high 4k resolution image of Radha Krishna kind of detailed best to suitable for wall paper

[prime bison]

Prompting nonsense I found online.

As with Darwin, biological observations precede and spur the development of good scientific tools producing measurements, valid interdisciplinary interpretations, and their applications```

Various nonsense

[copper kraken]

Here is the image you requested.

[swift valley]

paint a scene of cultural revolution that happened in china 1967

[glacial phoenix]

1 girl

[gloomy lark]

hah putting it through my llm, it also gives this, but no idea how historically accurate it is

[lapis wave]

logo for MENX telegram Channel

[native knot]

No.

[shy kelp]

[keen nimbus]

etc ethereum classic rocketing again.

[shy kelp]

[full wadi]

doglike employees

[proud scarab]

A young woman with fair skin

[copper kraken]

Here is the image you requested.

[copper kraken]

Here is the image you requested.

[jolly crest]

a stunning stylized lacquer painting of a sika deer with mountains and rivers, gorgeous, luxury, beautiful, golden, black background

[copper kraken]

Here is the image you requested.

[copper kraken]

Here is the image you requested.

[rose yarrow]

malevelon creek has fallen

[mossy lodge]

Is there any way to use IP adapter models with a1111? I have tried to use it many times with a1111 but it always gives error, while it runs perfectly in comfyui

[mild gust]

Chinese President Xi Jinping eats hot dry noodles at a food stall

[frigid flint]

shinkai

[worthy orbit]

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

[cobalt mason]

solo,1boy,child,kawaii,smiling happily,a little boy is eating an apple,holding an apple in hand,kneeling on the ground,masterpiece,Full HD,dynamic lighting,ultra high detail,ultra detailed,Best quality,within four walls

[cyan crown]

https://civitai.com/models/331085

[livid dust]

Should I write something here?

[glad grove]

nein

[karmic girder]

soccer ball,lottery ticket,logo,simplified chinese text,

[fiery sable]

food

[copper kraken]

Here is the image you requested.

[karmic girder]

soccer ball

[gloomy lark]

What kind of low effort horror gore is that? Pikachu with no electricity? 🙂

[south horizon]

pretty sure I've been to this Shell station

[cyan crown]

https://civitai.com/models/332246

[stone fossil]

Hmm yes.

[copper kraken]

https://civitai.com/models/330313/tplanetsdxlcontrolnettilerealisticv1?modelVersionId=370104

anyone try this?

[minor kettle]

does XL support negative prompts?

[copper kraken]

yes

[minor kettle]

hmm, mine don't work for some reason

[copper kraken]

we have a controlnet tile that actually works for upscaling!!!!!!

[glossy ravine]

https://tenor.com/bqXP1.gif

[shy kelp]

Show us it works for me

[cyan crown]

https://civitai.com/models/332552

[stray scroll]

musk

[crisp owl]

I challenge you to a dual vile creature of the dark!!

[stone fossil]

Guys... Im having some trouble with the lawnmower in the kitchen again. 😦

[toxic belfry]

this basket with 5 wooden essgs

[crisp owl]

I had trouble with my kitchenmower earlier this year, did you change the wiring to the headlight? That seemed to do the trick for me

[stone fossil]

Ah yes this could be it, the headlight could be leak, Ive not pumped it up in quite some months.

[crisp owl]

I get it, it's hard to keep up the maintenance on everything

[copper kraken]

I've found sdxl really really wants joints to have a cig filter

[stone fossil]

😄

Im testing me cursed lora hehe.

It is quite a mess.

One could say.

[meager canopy]

Pikatumor

[shy kelp]

indian king with sword

[open kraken]

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

[rustic garnet]

out of curiosity: how many images do you usually use for training your lora?

[junior niche]

MAKE A NEW RABBIT