✨｜sdxl

[silent kite]

So we all just moving back to xl?

[gray lagoon]

@half gyro traffic being hell, still omw back

[limber kindle]

All in 3.5K, full view recommended

[frail pawn]

Amabie Yokai:

[trim orbit]

futura concept nice

[twilit loom]

Hello everyone, can anyone suggest me why my generation of mask is not applying on the left side of the image. I am keeping the dimensions - 1016 x 504 but there is this left black patch that is coming again and again

[split lynx]

SDXL BETA BOT

[potent wagon]

Hey guys, I have a questions regarding the licensing of AI images. Can I license the images generated by stabilityai/stable-diffusion-xl-base-1.0 under a Creative Commons license (CC-0-1.0, CC-BY-4.0 or CC-BY-SA-4.0)?

The stable-diffusion-xl-base-1.0 model is licensed under CreativeML Open RAIL++-M License. Thank you!

Sounds good, Thank you! I will contact them, do you know how long usually I can get a reply from them?

No worries, thanks for the help!!

[meager canopy]

I'm pretty sure that the images created by SDXL are public domain.

There is no ownership of them.

It's the same license on the model.

Anyway, best to read and interpret yourself: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md

Whatever that means: "The Output You Generate. Except as set forth herein, Licensor claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License."

😄

[potent wagon]

Interesting, so I guess it means we can use the generated image whatever we want?

[meager canopy]

That used to be the case, but I'm not sure now that they've introduced the creators license.

[storm root]

Trypophobia intensifies...

[potent wagon]

I am only going to use the model for non-commercial use, and no money change, so I guess I should be able to license them?

but then they also give an example:

Imagine a company wants to use your model in order to develop a version for a commercial chatbot. The company accesses the model, modifies it, and finetunes it to be the technical backbone of the chatbot app. Firstly, these actions will be governed by the OpenRAIL license. Secondly and worth to note, according to the terms defined in the OpenRAIL License this is considered a derivative of the Model. Thus, the use of the chatbot will be governed by the use-based restrictions defined in the OpenRAIL license, and accordingly, when commercializing the new version of the Model by means of a commercial license (or any other type of legal agreement), the latter will have to integrate these use-based restrictions as part of the subsequent license.

I guess it is pretty complex, haha. Thank you!

[meager canopy]

I don't think you're allowed to license it, because you don't own it.

[meager canopy]

He uses the faces of his enemies as armour.

[green lantern]

Yo how do i generate stuff?

[meager canopy]

...or #artisan-faq

"How do we drive this thing?"

[potent wagon]

Hey, does anymore know the rate limits for inference API for Stable diffusion XL model on hugging face?

[limber kindle]

All very cool stuff ❤️ Do you apply some kind of DOF in post?

[meager canopy]

No

Spot the hidden spycam

[meager canopy]

https://tenor.com/view/fizzer1k-5597-3689-move-along-nothing-to-seehere-gif-12353147

[meager canopy]

Medieval Iron Man

Zoom in on the deets

What's the problem with hands anyway? 🤷🏻‍♂️

[elfin girder]

how are you doing text with sdxl? these look pretty good

mind sharing it?

yeah just realized lol
thank you

[river scroll]

Zoom in on the deets

[elfin girder]

the consistency and clarity is sd3 level, if not better (sd3 struggles with other details of the image sometimes)

[meager canopy]

How all that jewellery above was made

[elfin girder]

imma try it once i get home
im at work rn
i haven't used any text loras before

oh definitely
thank you
i usually use dreamshaper XL, hope it'd work fine with that

[elfin girder]

lord
whats your gpu?

[meager canopy]

4090

[elfin girder]

24G?

[meager canopy]

Yes

[elfin girder]

Damn

[limber kindle]

@meager canopy Very first test with hidiffusion. 4kx4k

I mean aside from the hands, kinda cool

[meager canopy]

https://tenor.com/view/pillo-pillo-pillo-gif-25873550

[limber kindle]

: DD

[timid garnet]

👋 .

[timid garnet]

I think you got a world record here

So funny

But very interesting image

[limber kindle]

Lol yeah. 4x the resolution, 4x the fingers!

At least it doesn't mess up faces. Tweaked it a bit, will see whedre it goes 😛

[native knot]

Needs more fingers.

https://tenor.com/bGgFl.gif

[limber kindle]

What guidance scale do you guys usually use in sdxl?

[red minnow]

🍭 lollipop

[uncut gull]

Did the mods leave or something?

[timid garnet]

For A1111 and Forge users :
https://www.reddit.com/r/StableDiffusion/comments/1dkh5cv/a1111_forge_improve_your_images_with_a_single/

I finally wrote a post about this

[stray warren]

Off to dreamland

[keen bobcat]

is there an inpaint controlnet model for SDXL?

[copper kraken]

several

[old mulch]

Can anyone that knows how to finetune loras in particular about captioning images take a look at #🔧｜finetune for me please.

[idle cargo]

modern hotel lobby

[tacit garden]

good morning

[onyx meteor]

/gile

[hearty light]

A small white cat

[meager canopy]

Like this?

[meager canopy]

More like this?

[gloomy lark]

hunyuan

[glass forge]

That shoe tho its in an interesting place. If you quint...

[timid garnet]

👋 .

[meager canopy]

https://tenor.com/view/fruit-dancing-food-gif-4808470

[gloomy lark]

They're having the best time. 🙂

Futuristic steampunk warrior in dynamic pose, bright neon blue energy illuminating surroundings, fighting oversized, animated evil fruit in a chaotic grocery store, intricate mechanical armor glowing with blue circuits, swinging a plasma blade, anthropomorphic apples with fierce expressions, bananas with menacing grins, flying strawberries with sharp teeth, aisles filled with overturned products, flickering fluorescent lights casting eerie shadows, shelves toppled over in the background.

[meager canopy]

Which model are you using?

[gloomy lark]

https://civitai.com/models/89804/ip-design-or-3d

[gloomy lark]

It's ella with that sd 1.5 checkpoint, then cutejack 0.3 on civit as a refiner

[gloomy lark]

I was playing around with various sd 1.5 checkpoints the other day and it's by far the most prompt following with Ella that I've come across across so far. Was using juggernaut reborn before this.

[meager canopy]

I just don't get good images from Ella, at all! Must be my workflow 😄

These have been improved with controlnet tile too!

Look at these full screen! 😎

[tacit garden]

.

[tame ore]

\dream

\dream 该项目位于迪拜酋长国谢赫扎伊路AL SAFOUH地区，是业主投资建设的超高层五星级酒店公寓项目，该项目设计为2B+G+M+7P+GYM+14 HOTEL+SERVICE+27 HOTEL APARTMENT+SERVICE+ROOF层,占地面积为5072.7平方米，总建筑面积为96298.3平方米，建筑高度226米。设计一个高大威猛的中铁建塔吊来施工这个建筑

该项目位于迪拜酋长国谢赫扎伊路AL SAFOUH地区，是业主投资建设的超高层五星级酒店公寓项目，该项目设计为2B+G+M+7P+GYM+14 HOTEL+SERVICE+27 HOTEL APARTMENT+SERVICE+ROOF层,占地面积为5072.7平方米，总建筑面积为96298.3平方米，建筑高度226米。设计一个高大威猛的中铁建塔吊来施工这个建筑

[meager canopy]

Coming to Netflix soon!

[gloomy lark]

Sure, here's my workflow. 2 stages of ella, then a stage or 2 of sdxl refinement. you'll need the extra samplers custom nodes, and i'm using the resolution selector from pixart.

I find that the res_momentumized samplers do especially well with fingers and hands.

[meager canopy]

Nice, thanks! 🙂

[meager canopy]

I like these 😊

[limber kindle]

Anyone in here using allor by any chance?

[dim steeple]

New project I am working on using SDXL for characters - 24/7 always on live stream of historical figures reacting to twitter live: https://www.twitch.tv/timelesstakes

[stable dune]

/create

[limber kindle]

She looks like an alien ❤️

[glass forge]

Being very wealthy would make contemplating mortality too unbearable.

[tame ore]

https://tenor.com/bFWdy.gif

https://tenor.com/Qyi0.gif

[warped trellis]

Not very good as-is, but there are some nice merges.

[deft ivy]

\dream generate image of hours

[gloomy lark]

Cascade/pixart

[tame ore]

\dream generate image of hours

Cascade/pixart

这个怎么玩的呢

能告诉我是什么程序吗

ok

thanks

[rich marten]

👍

NICE

[tame ore]

💯

[oak sandal]

I would go sd3 for environments. It really shines in backgrounds and landscapes without characters.

[shy kelp]

the new Juggernaut XI is great for landscapes I spent a while with it today

[gloomy lark]

Wow really cool stuff

[limber kindle]

Midsummer \o/

[gloomy lark]

lumina

[timid garnet]

👋 .

[limber kindle]

@timid garnet Yo! Can you link me that article you wrote on your a1111 workflow. Can't find it back

[timid garnet]

This

Was curious, but no one seems to care ahah

[whole osprey]

sdxl flash from sd community, 3 word prompt, 8 steps in 10 secs on A100, fast asf and not bad

[tall garden]

Imagine a powerful Viking, adorned in traditional fur-lined armor, engaged in an epic struggle. On one side, he grapples with a fierce bear, its claws extended, while on the other side, a majestic lion roars, teeth bared in defiance. The Viking's stance is resolute, muscles tense with effort, encapsulating the raw intensity of the battle between man and beast. The lines capture the dynamic motion and strength of each participant, evoking the ancient clash of bravery and ferocity

[dim steeple]

Using SDXL to generate avatars for a 24/7 livestream of Michael Jackson, Marilyn Monroe, Einstein, and Steve Jobs reacting to a live twitter feed. Fully AI generated conversations and interacting with chat: https://www.twitch.tv/timelesstakes

[copper kraken]

Here is the image you requested.

[peak dove]

Automated SD to Photoshop node/plugin - SDXL Models = Mobius, NightVision XL and Albedo XL

[timid garnet]

My upscale is always, more or less, washing the colors (contrast)... I'm struggling to understand what's going on here...
Visual and obvious example :

Is it only cause of the upscaler ? (foolhardy)

[limber kindle]

Yeah it comes from upscaling. In comfy I use a color match node, that takes the constrasts and saturation from base image and applies it to the final output. No idea how to do it in A1111 though

[timid garnet]

There is a color match setting too, but it's just enable/disable, and pretty meh

foolhardy / 4xNomos8kDAT

[limber kindle]

Best way to upscale imo is using a tiny bit of hallucination ontop, for additional resampling

Not much, just to trigger it

[timid garnet]

hallucination ? 🤔

[limber kindle]

Well running it through a sampler again changing the noise just a tiny bit. Basically like magnific, just with low settings

Just googled real quick, in A1111 what I mean is this

https://www.youtube.com/watch?v=wZiPYppLu10

[timid garnet]

Well i'm using ultimate sd upscale after the txt2img part with extra noise. The pics are split in 9 tiles and get 25+ additional steps

[limber kindle]

Yeah but those are upsacle steps, what I mean is re-rendering with a sampler. Just watch the vid, you'll see what I mean

[timid garnet]

Totally what i'm doing. But i'm also adding extra noise, for details to emerge further.
I'm using a 0.46 denoize.

ah, he's using CN

I can't

vram limited

As i didn't test, i'm not sure what the tile CN add to the process

This is just the USDU part :

foolhardy / 4xNomos8kDAT

This is not bad

[shy kelp]

I think this one is a bit better than the dat

https://openmodeldb.info/models/4x-Nomos8k-atd-jpg

[timid garnet]

DAT models are soooo long

[molten gull]

flatworld 🙂

[shy kelp]

🤨

[frail pawn]

elesh norn

[nova temple]

https://civitai.com/models/471439/0002-pony

[limber kindle]

prompt blend with gradients

[meager canopy]

Did I break Luma?

[limber kindle]

Haha nice animation

[humble cape]

Imagine getting 10 hours of queue for static animation

[timid garnet]

Scrolling the gallery. I love this one

[peak dove]

"Get Gud Skilz!!!" ™

[timid garnet]

👋 .

[peak dove]

Wow!!! This CFG++ is really making a difference - REALLY!!! Incredible!!! Just drop the CFG from 8 to about 2.5 or 3 before you start!!!

[meager canopy]

I can't see any way to set steps from/to with the custom sampler, so I'll wait until it's part of the other samplers.

[peak dove]

Ollama Up Baby!

[peak dove]

Get friendly with an LLM! "Get Gud Skilz!" ™ Using CFG++

You may be doing overkill if you implement CFG++ with a ReFINER

[meager canopy]

It's not good for anything realistic.

Not from what I've tried so far anyway.

[peak dove]

Drop the CFG right down ... and use realistic models like NightVisionXL or RealVis

[meager canopy]

I use low CFG all the time, even at 2.0 it's not so good.

The premiere of "The Lion King"

[molten gull]

some new pictures from the last couple days 🙂

the first picture you mean ? it was kinda an experiment with comfyui and an unexpected result

it's an s-kite

or are pure-xxl, not 100% sure

it's been a while 🙂

yeah, it's a pure xxl

and it does pull like a truck 🙂

it's an ancient thing 🙂 i doubt there's many people left flying those

🙂 it's like not from this millenium

it's a fun-flight, you cannot really kiteboard/buggy with it, but in low winds it flies astonishingly stable

i still have a lot of old kites down in my (very dry) cellar, i dont really get to fly much in the last years

got some fun stuff 🙂 a c-quad 8.5, a 20er stacker of speedwings, two big s-kites

a 12qm power lifter

should check my cellar again these days

those were fun, too 🙂

[peak dove]

#🆕｜sd3 message

[meager canopy]

You don't seriously think that looks realistic, do you?! 😄

[peak dove]

Sorry, I am mainly an artist - everything looks 'realistic' in strictly artistic terms!

But it ain't photographic realism, no!

CFG++ is a fine arrow to string to your bow - at least at "Torcello Towers" it is!! LOL

[peak dove]

Try 0.4!

[haughty meteor]

"I currently have a drawing task, which involves creating a picture for a child's name, specifically for the name '田羽靖'."

[south horizon]

neat details

[timid garnet]

I'm dying, i can't use adetailer anymore with Ultimate SD Upscale

[brazen umbra]

why?

[shy kelp]

50$ gift
steamcommunity.com/gift/2918J0193717

[timid garnet]

It's bugged since i updated something. Not sure if it's an extension or the webui/
If you try to enable :
-Soft inpainting
-ADetailer
-Ultimate SD Upscale
in img2img, the first tile will be processed, then return in a black image.
Then you just process the black image.

This is quite devastating for my "workflow"

I'm using a1111*

[silent quest]

Are there any compltley free Stable Diffusion models?

[timid garnet]

All SD models are on Civitai or Huggingface and are free.

[silent quest]

whats the link

[shy kelp]

https://civitai.com/

https://huggingface.co/models?sort=trending

[silent quest]

If downloading a model on Civitai do I have to connect to the official SD API to run them?

[dapper dragon]

SDXL racer 😀

[shy kelp]

nice retro lighting

[shy kelp]

yeah the fog and rain are great

wait is that a AE86 lmao

[dapper dragon]

you can try "green and red neon glow"

[shy kelp]

ah okay thanks

[dapper dragon]

low cfg also helps to increase the foggy stuff

[shy kelp]

yeah I use low cfg almost always now
because low cfg also helps get softer skin

[dapper dragon]

SDXL keeps rocking 🙂

[shy kelp]

yeah I'm staying SDXL mostly for a while
I think I only like models after a lot of fine tuning

it took me a long time to leave SD1.5 even

[dapper dragon]

Totally agree 🙂

[dapper dragon]

I switched to SD2.1 it was not the best in the world but could do proper 1024x1024

[shy kelp]

I actually skipped SD2 and SD2.1

I might go back some time

and see how they are

cos maybe for certain subjects they have a unique look

[dapper dragon]

you didnt miss too much. Overall it was not that good. But there was nothing better. Try Illuminati 1.1 it can make very unique stuff

This shit knows proper details.

[shy kelp]

which checkpoint is this

ah i found it https://civitai.com/models/11193/illuminati-diffusion-v11

[dapper dragon]

Realistic Vision 4 SD XL ...

[shy kelp]

ah ok

[dapper dragon]

Yeah this one is a good one 2.1

[shy kelp]

realvis is already my main model lol

[dapper dragon]

Very good choice 🙂

For illuminati you also need the 3 embeddings

they fix the model problems

provide on that page 🙂

then just check the examples

[shy kelp]

Dalle also does better in the rain and fog

I think rain and fog are good ways to make models more real

[dapper dragon]

I don't know DallE that much and prefer local stuff 🙂

[quartz yew]

Hm how do I get a casual look in my images? like someone shot it on its crappy phone instead super high quality studio photos ? "shot on iphone" and similar doesn't really work. "Polaroid" gives some believable results. But is there something even better ?

[wet nacelle]

cascade

cascade

[wet nacelle]

cascade

[iron rover]

Cascade

[iron rover]

I use them in my iPad on the app draw things it’s free runs sd great

[copper kraken]

also cascad

[iron rover]

Beautiful 😻

[wet nacelle]

Cascade

[copper kraken]

idk if you're able to run comfyui but i got some nodes that are getting crazy shit outta it

[wet nacelle]

Do you mind giving me the flow to try?

[copper kraken]

sure, no prob, you'll need the nodes https://github.com/ClownsharkBatwing/RES4LYF

git clone https://github.com/ClownsharkBatwing/RES4LYF

^run that in your custom_nodes folder

then cd into it, and type "pip install -r requirements.txt"

[wet nacelle]

okay. I'll figure this out.

How many gigs do you think this will take up after full setup?

@copper kraken

[copper kraken]

of vram or disk space?

[wet nacelle]

disk

[copper kraken]

for disk space, nothing beyond what you need for cascade itself, unless you want to use my lora, that's about 700mb i think

[wet nacelle]

Alright kool shit

[copper kraken]

i see half the discord is using cascade again 😄

giving me february flashbacks

[iron rover]

And ppl asking for bots 😂 images I still have some of ur funny images saved

[copper kraken]

oh god i miss the prompt zombies

[iron rover]

😂

[copper kraken]

we made a sport out of it

[iron rover]

We sure did

Mack and I would cascade battle out of funn too

[copper kraken]

i wish we still had that cascade channel

there's what, like 7 or 8 of us that are using it daily now at least?

[iron rover]

Me too found one

[copper kraken]

hahaha

[iron rover]

Yes u me and Mack I don’t really recognize anyone else

Least I give u credit thou had to save them good times

[copper kraken]

holy shit

cascade is so creative

[iron rover]

It does horror so good

[copper kraken]

it does... this was something generated from my paint lora

[iron rover]

Best free model

[copper kraken]

so that's a 3D paint sculpture

yeah, for real

it got overlooked cuz the SD3 anouncement came a week after cascade was announced

[iron rover]

Yes I’ll have to go digging but I did save some of ur work

[iron rover]

Yes and XL too

They all released with in a week and we lost the bots

Now there’s way more drama

[copper kraken]

what happened with XL?

yeah, the cascade launch was waaaay more fun

[iron rover]

Good question kinda got hidden with Cascade and sd3

[copper kraken]

this was 100% me when i saw how nasty things were getting in the sd3 chan

seems to be calming down now though

[iron rover]

Yes same Mack and I were getting annoyed because the time kept going up and ppl acted rage 😡 with out reading and knowing anything about the model … insane to see that much drama

Worse part it was over one prompt

[copper kraken]

i mean... a lot of it was funny at first... but then it turned really mean

[iron rover]

😂 like seriously ppl all this over grass 😂

[copper kraken]

does sd3 have some big issues? yup, but jeez, keep your humor ppl and don't get so serious like you're actually going to die because your free AI art model isn't what you hoped it'd be

[iron rover]

Yeah very mean

Cascade is better

My fave SD model and I agree

Not that I hate on sd3 just knew hands were gonna be a issue

Called it from day one

[copper kraken]

yeah

i think the big challenge is going to be learning how to wrangle a three text encoder model

from a training perspective in particular

i'm trying training experiments every night... getting SOME Results, but nothing terribly exciting yet

[iron rover]

I’ll share my horror with u if that helps because right now it’s trial and error

[copper kraken]

it's ridiculously good at illustrations and paintings

better than any model i've tried

[rotund surge]

🤗 nice images

[vital ermine]

https://civitai.com/models/536487?modelVersionId=596395

[surreal flax]

I have an issue with SDXL models

I'm on a new PC (3060 laptop), installed Auto1111, and my generations times are inconsistent

These are my .bat arguments:

set COMMANDLINE_ARGS=--medvram --medvram-sdxl --xformers --no-half-vae

I've gotten 1it/s on some generations and 2s/it on others

[copper kraken]

see if you have the same problem with comfyui or stableswarm

[surreal flax]

Exact same generation settings for both

Even the seed was exactly the same

512x512, 20 step, 7 CFG, I know 512x512 isn't recommended for XL but I'm just testing this GPU

Essentially same image generations vary from 20 seconds to 1 minute+, it's odd

[surreal flax]

I'll try SwarmUI now

Around 25 seconds with SwarmUI with the same generation settings

[copper kraken]

is it consistent now

[surreal flax]

Wait, SwarmUI was using 1024x1024 instead of 512x512

I entered 512x512 and now it's 10 seconds

Significantly faster

[copper kraken]

yep

a1111 is slow as f

stick with swarmUI if you can, it's gonna be getting lots of upgrades

[surreal flax]

I've done it 5 times with various prompts and each generation time is consistent

[copper kraken]

yep as it should be

[frosty rapids]

pls share more on this prompting?

[surreal flax]

Removing --medvram-sdxl from Auto1111 did help generation speed for some reason, but it's still 2 times slower than Swarm

[frosty rapids]

we're looking to put guides together so results from any experiments are helpful and we'll be grateful

[copper kraken]

cool! yeah, i can get some stuff together for ya tomorrow.

[surreal flax]

Oh god, now I'm getting 35s/it on Auto1111, I'll enable medvram-sdxl again

[copper kraken]

i'm getting a significant shift in style now - it took a ton of steps, but it looks pretty awesome now and tehn with the right captions... haven't figured out how to make it reliable

[copper kraken]

the prompt there is some shuffled components of captions for the training set for the lora i was using for that (cascade)

the biggie is the sampling strategy

i took the RES sampler from @lyric grove 's awesome extra samplers node pack and threw on tons of bells and whistles that allow scheduling every parameter throughout every step of the diffusion process

there's very real benefits from that, i've found

[surreal flax]

Even with this argument, I'm getting 33s/it, I think 6GBs of VRAM is just not feasible with SDXL on Auto1111

[copper kraken]

yeah, honestly, best thing you can do is just switch to comfyui or swarmui

they're going to be much better supported in the future, and they're better right now in almost every way

Cinematic photograph of a menacing plant with spiked petals and blood-red sap dripping down, set in a dimly lit, dense forest, the sharp contrast between the vibrant red sap and the dark green foliage, detailed textures of the plant’s spiky petals and leaves, creating a chilling, eerie mood.

[copper kraken]

cascade? good stuff

[wet nacelle]

yee

[iron rover]

Sd runs better on cuda sounds like u do not have that https://stealthoptional.com/tech/stable-diffusion-runtime-error-how-to-fix-cuda-out-of-memory-error-in-stable-diffusion/

[gloomy lark]

A young adventurer wearing a headband and boots holding glowing blue spiked fruits illuminated by the warm light of a setting sun vibrant, high above jagged rocky cliffs fierce dragon-like creatures with sharp teeth and colorful scales snarling, broken shards of rock and fruit suspended in the air, dramatic clouds swirling in the sky a sense of tension and exhilaration.

[nova chasm]

https://tenor.com/bmc8q.gif

[shy kelp]

The fact that the poor kids prosthetic leg is breaking down is just another insight into a dysfunctional and dystopian health care system 😆

[wet nacelle]

Cascade

[sharp scarab]

https://www.youtube.com/watch?v=zo6w-Md7DUo

[timid garnet]

Finally tweaked my upscale. Need your impression peeps.
https://imgsli.com/Mjc0NTQx/1/2

Raw vs Old upscale vs New upscale

Click on the link for the comparison with a slider and the images choices. The raw output is there, for reference.

I need to manage again the contrast. Need to think about it. Probably a prompt issue.

[grand cargo]

💖 🤖 Strong results have put Boltning Hyper D into 2nd place on the SDXL Top 10 model list. It did pretty well with photorealism too 💗 🥳 https://docs.google.com/spreadsheets/d/1IYJw4Iv9M_vX507MPbdX4thhVYxOr6-IThbaRjdpVgM/edit?usp=sharing

[shy kelp]

face is far better on the new one

you lost some contrast though

[iron rover]

Don’t quote me but I believe it’s 2

[meager canopy]

Neither is very sharp and clear, but I like this little addition 😄

[meager canopy]

Mind sharing the prompt for this please?

A lot going on in this one! 😄

[glass forge]

i like the extra saber blades just for good measure.

AI: "star wars means lightsaber blades, a lot of them everywhere"

[tiny nebula]

Images of people performing different types of cold therapy (cold showers, ice baths, cryotherapy).

[iron rover]

Here’s the image you asked for

[copper kraken]

that's cool as F

[sonic shard]

Letter

[timid garnet]

Hey, here it is. May be random with another model lol
prompt:Photo of an amazingly beautiful creature with flowing white hair towering over a forest floor, its fur barely covering the delicate scales. Behind her lies stone staircases and ancient trees filled for miles above shorelines as rain fills each path to catch fish before it reaches, 2000s vintage RAW photo, photorealistic, film grain, candid camera, color graded cinematic, eye catchlights, atmospheric lighting, natural, shallow dof, High level of detail to create a photographic-like image, focusing on lighting, realistic textures, hyperdetailed. steps:40 width:832 height:1216 guidance_scale:5.0 sampler:DPM++ 3M SDE scheduler:exponential seed:1589594602 data_model:ZavyChromaXL negative_prompt:NSFW, nude, naked

[meager canopy]

Thanks! I think it's interesting seeing how it varies between models 🙂

[timid garnet]

I need something to automatically tweak the contrast/colors in a1111

Boring, my pics are so washed out when i try to improve my upscale...

[copper kraken]

prompt: "creature with multiple large eyes and sharp teeth, the scene is whimsical yet detailed, surrounded by lush green leaves and smaller golden blooms, a creepy"

[meager canopy]

The prompt is a little confusing for AI 😄

I changed it a bit

[timid garnet]

Got one like this, nice

[meager canopy]

Looks good in wide format too 🙂

[atomic minnow]

https://tenor.com/W4ck.gif

[shy kelp]

new king of the jungle

[desert echo]

what model is this??

[wet nacelle]

Cascade

[dapper dragon]

Yeah 70s muscle car rocks. I agree it could be a little better, but it's so amazing, pure prompt no lora. I cant complaint bro.

haha dont worry bro 🙂

Did this one too

[dapper dragon]

not bad 🙂

Sexy

🙂

[peak dove]

CSBW wf

[dapper dragon]

😂

it already knows a lot I still amazed what it can do with no lora

[peak dove]

CSBW w/f

[clever cave]

realvis xl lightning with ncp dpo lora is just insane! i didnt even add anything crazy to my prompt
prompt is

masterpiece, high quality, 8k, best quality, an horse in the space

[silk smelt]

Hello chat, im new here. I guess this channel is for all kinds of sdxl models?

[clever cave]

i believe so

[lusty wolf]

Still love what you can do with SDXL.

[silk smelt]

I still love sd15, but mainly use sdxl. and try sd3 a little

[copper kraken]

try cascade

cascade > sdxl, just needs the community to dive in

[clever cave]

No good finetunes of cascade, that’s the problem.

Cascade base is def a lot better then sdxl base tho. Really good text too

[copper kraken]

and the reason for that... is cuz everyone thought sd3 was gonna be the next big thing

it trains more easily than sdxl, and now i just showed that the talk about training B being useless is false - it really is all you need to add that bit of polish

i'm getting better generations than i've gotten with sdxl now

[clever cave]

You gotta try sdxl with this new dpo Lora tho. The detail is insane.

https://huggingface.co/alexgambashidze/SDXL_NCP-DPO_v0.1

[copper kraken]

yeah i've used some of those

it looks cool but tends to blow out images imo

[clever cave]

New technique tho, it just came out today I think

Bit different then the original sdxl dpo Lora’s but this improves quality massively. The original sdxl dpos were ok-ok

[copper kraken]

depends what you're aiming for, i s'pose

to me a lot of those examples look kinda cooked

but i go for a different style myself 🙂

[clever cave]

Yeah true, imo I like these Lora’s too. They kinda make sdxl do real art

https://huggingface.co/alvdansen

[copper kraken]

nice!

[vital ermine]

https://civitai.com/models/540293?modelVersionId=600680

[carmine aurora]

If you want a little more vitality, a little younger flat wind illustration vertical

[meager canopy]

Cascade 🙂

Yes, but it's only using a captioner

https://tenor.com/view/minion-any-questions-question-question-mark-confused-asking-gif-15404852

Why?

You mean, training for that stuff you may want to use in the future?

Once I post on Discord I delete the image locally, so what's here is the only version of what I made 😄

If it helps build the models of the future, then I'm glad it goes to good use.

It's all clean here 🤷🏻‍♂️

[lucid surge]

蔷薇花

I am in nearly 90 discords, i like control over any traces i might have left anywhere.

[wet nacelle]

Casacde

[wet nacelle]

Cascade.

[spark abyss]

Characters:
Boy:
Attire: Wearing a light-colored short-sleeved shirt, sturdy khaki pants, hiking shoes with rubber details for grip, and an olive green travel backpack with side pockets.
Appearance: He has short, dark hair slightly tousled by the jungle breeze. His face is lightly tanned, with a neatly trimmed short beard. He is tall and athletic.
Pose: Holding the girl's han, he looks forward with a slight smile, his gaze full of curiosity for the adventure ahead.
Girl:
Attire: She wears a light-colored tank top, comfortable yet durable trekking pants, robust hiking shoes, and a colorful backpack with floral details.
Appearance: Her long dark hair is gathered in a ponytail for convenience, with some strands delicately falling on her face. She is slender and has an energetic and determined expression.
Pose: Holding the boy's hand affectionately, she has a slight expression of wonder. Her gaze is directed sideways, curious to explore the surrounding jungle.
Setting:
Amazon Jungle:
Vegetation: The environment is dominated by majestic trees with massive trunks and exposed roots, adorned with hanging vines. The foliage is dense and varied, featuring tropical plants with large, vibrant leaves and colorful flowers scattered throughout.
Extra Details: Colorful butterflies flutter among the trees, while rays of sunlight filter through the canopy, creating patterns of light and shadow on the ground. A narrow trail disappears into the vegetation, inviting discovery of hidden wonders within the jungle.
Composition:
Character Placement: The boy and girl are at the center of the illustration, walking away from the viewer into the jungle. Their left hands are clasped firmly, conveying a sense of unity and shared adventure.
Background: The jungle fills the entire background of the image, imparting a sense of depth and mystery. Tall trees and dense foliage envelop the characters, while the atmosphere is imbued with excitement for the impending exploration.

[copper kraken]

well, that makes it suck extra badly that #stable-cascade is gone

[gleaming magnet]

China's Loong Boat Festival, the background is three giant mountain Zongzi, surrounded by clouds, Loong Boat racing, Loong Boat moving on the water, the river winding from the huge mountain Zongzi to the distance, the background is misty mountains, rich in details, daytime, fairy fog, clouds, movie lighting, super high quality, super high resolution, ultra high definition, ultra clear details, 8K, HDR, Chinese style

[copper kraken]

Here is the image you requested.

[copper kraken]

the problem is it's all already gone and fruitt refuses to make them available again for whatever reason

thanks

[uncut gull]

Now that's a painting. Really good simplification.

Can't really get people with a painterly feel though. This is too good.

I know there are infinite models. I'm using CreaPrompt Hyper/Lightning v4 at 4 steps, with no hires fix. Maybe could get better results with a painting or art model, or a lora.

Scrolling through Civitai, all the paint models and loras have the same problem: The face colors are too subtle / realistic. No painter could blend this many shades this perfectly over such tiny surface changes.
Doesn't really matter anyway. I was just curious. Maybe someone could figure it out.

I mean the changes from pink to yellow to green on the skin of her face. Those subtle hue shifts are impossible with just blending pigments.

[dapper estuary]

full body shot, miumiu fashion show, irregular shape, morden chinese digital style

[uncut gull]

Yeah. Reference from Gurney. You can try to paint those colors, but they'll never look as subtle and smooth as real life. SDXL is just too good at those gradients.

You don't seem to get what I'm talking about. Maybe you can't see it.

Pigments are exactly 1 hue. When you blend between two pigments, one hue fades out, and the other hue fades in. You can't add any other hues in there. It's physically impossible.

That has blue, pink, yellow, and green, but they're all blended over a huge, smooth area independent of the brush-stroke impressions.

[weary juniper]

dog, excited, happy expression, spacesuit, helmet, oxygen mask, floating, space, planets, cinematic, full-body, 4K, 9:16 aspect ratio, photorealistic, detailed, vibrant colors, majestic, otherworldly, visually stunning, classic composition, masterpiece, exquisite, color correction, amazing visual effects, intricate details, sharp focus, super high effect, HD, 16k --ar 9:16 --v 6.0

[copper kraken]

Here is the image you requested.

[peak dove]

Mobius checkpoint

[gloomy lark]

Stumbled upon a cool prompt prefix: unreal engine,c4d renderer,high resolution,Octane render, iconic bird view top shot of a

[gaunt dock]

Hi everyone. I want to create female face shapes that are different from the same shape that sdxl always uses. What kind of words do you suggest using to modify face shape? I tried 'diamond shaped face" or "heart shaped face" type prompts but they seem to either not work or to put pictures of hearts on her cheeks. My overall plan is to make images then turn them into a lora for an original character. Any suggestions appreciated

That is a good idea. I have used controlnet but never tried doing that

Im trying to build an australian girl but the model doesnt understand that face shape.

I initially tried making a lora of instagram girls with that face but their camera quality is too bad

Those are the kind of words I was after

There are 100 breast slider loras but no face slider loras like youd find in skyrim

[copper kraken]

you have been posting some cool as fuck shit

[vernal cloak]

Sorry guys. But i have a question regarding SDXL inpainting. Particularly in ForgeUI v #🤝｜tech-support message

[iron rover]

Here is the image with Lora you requested

[copper kraken]

made a decent amount of progress on cascade stuff btw

[copper kraken]

@gloomy lark @iron rover i combined your two images with cascade clipvision

[urban badge]

/git

little girle

[copper kraken]

Here is the image you requested.

[glass forge]

Turbodong, is that you?

[copper kraken]

the teeth are my fav part

[unreal bough]

/git
little Rabat

robot

[rigid laurel]

"Y'all need jesus, and I'm fresh out."

[iron rover]

That’s so awesome 😎

[iron rover]

That’s so awesome 😎 … I love ur stuff … I merged 3 horror artist together , using Cascade came out so good… artist Hp love craft , Wes craven and H.R. Giger

[iron rover]

Misty does not play very well 😂

[meager canopy]

@vestal breach Can you ban the spamming bot please? ☝🏻

[gloomy lark]

cascade

[stray warren]

Good old base SDXL

[copper kraken]

Those are some great ones

Loooove creepy and weird

[gloomy lark]

cascade

[gusty zealot]

Erstelle mir einen Charakter aus dem Film Alles steht Kopf der für die Emotion zwang steht

[gloomy lark]

cascade

[gloomy lark]

lumina

[iron rover]

Yesss add me on here we can collab I mainly do dark horror fantasy art

[copper kraken]

done 🙂

[gloomy lark]

lumina

[gloomy lark]

cascade

[vast galleon]

I converted one of my models to CosXL. You can get some interesting results when you mix the Original model with the CosXL clip. From left to right, Original model, Converted CosXL model, Original model + CosXL Clip.

[gloomy lark]

I usually set output blocks 0-4 at 0.01 and don't use the cosxl clip at all because everything in the more recent advanced models is better.

[vast galleon]

yeah, i thought that was interesting that it didn't really make my converted model any darker, but the outputs are COMPLETELY different, which I kinda like.

[gloomy lark]

yeah, they're compeltely different because the clip and training is very different than standard sdxl base.

models like mobius and aventishorizon etc that have been trained on really good GPT created captions and others are all going to be more prompt following than cosxl

[vast galleon]

I also noticed merging loras into it doesn't really work out

[gloomy lark]

I found in a/b comparisons, using the cosxl generally made the model dumber to put a blunt end on it.

[vast galleon]

yeah, things that should be photographic are coming out as paintings which i thought was odd.

[gloomy lark]

until I realized I didn't need to bring any of that into the merge, and could leave hte clip behind, and only have 0.01 for the output blocks, preserving the most of the original

yeah, loras are based on sdxl base. unfortunately cosxl is effectively a different base model.

they really should have just released a version of sdxl base that ignores the brightness/darkness limitations.

[vast galleon]

yeah. what're you're thoughts on mobius? i've seen it get a lot of hate and doubt

[vast galleon]

i'm trying out their clip now.

Interesting. Mobius on the left, SDXL on the right.

[thin smelt]

Both good, though SDXL seems just a tad better.

[vast galleon]

It didn't get black and white either. things seem further away in it too, which can be good. trying some merges out on it.

[gloomy lark]

my favorite sdxl at this point is mobius with an aventishorizon refiner stage.

@vast galleon some mobius from today.

[vast galleon]

nice

[clever cave]

mobius is nice, but whats going to be really useful is the training strategy they use. mobius as a model itself is decent but isnt some model that beats every other model.

[clever cave]

why is there harley quinn in the bottom lol, and what prompt?

[iron rover]

My watermark says my name too

I do that not to only share here but also on deviant art and fb etc

[vast galleon]

I merged it's clip slightly with the model I'm working on, in addition to the clip from the converted CosXL model I made. The differences are small, but it improves some things slightly.

[clever cave]

nice, i cant wait for the mobius training strategy. If what they say is really true, then it can really improve sdxl generation.

[iron rover]

@clever cave the prompt >

Easiest way for me to keep track of my prompts wordpad

[clever cave]

pony i guess? I never really got good results with pony. Any settings you recommend?

[iron rover]

Yes but I don’t use it for nsfw just for the poses

I recommend these

[clever cave]

yeah it definitely seems a lot better then other sdxl model with poses, il try it with those loras

[iron rover]

Pony Diffusion v6 xL if u are looking for similar results as me as for the horror side like mine that’s actually me drawing by hand adding in the sfx

But u can cut the prompt down use the trigger words and say no to nsfw

I’m like u I’m not into that type of art either just love the poses

Great for horror when I draw

Don’t forget to change the dress color and hair that’s how I got Cinderella and Belle is obviously the Lora

[clever cave]

yeah example images show how good it is, is it also good for prompt following compared to models like proteus or just other sdxl finetunes?
for realistic things, i guess this is better?
https://civitai.com/models/372465/pony-realism?modelVersionId=534642

[iron rover]

Thank u I will def do that…

Meaning check it out yesss it’s perfect with bodies and I’m not knocking sd3 but it’s great for what we want none nsfw

But I do love Cascade too!

[vast galleon]

Gonna be dropping Mangled Merge XL V3 in the next few days. 1800 loras merged. The last 600 were using the DARE/TIES method. So far it seems capable of both photorealism and pretty good 2D.

[marsh sail]

create a image of fantacy race inspired from snakemen

[iron rover]

Here’s the image u requested

[iron rover]

Spider girl Jasmine

[meager canopy]

"Are you a betting man?"

[shy kelp]

Those war scenes are gritty for real

[shy kelp]

that's beautiful

[sick fulcrum]

anyone know how i can use two shot/latent coupling for sdxl?

[clever cave]

whats the best overall sdxl model for human anatomy and overall except pony?

[gloomy lark]

Cascade

[wintry marlin]

Might be useful to many here: https://www.magicflow.ai/insights/read/sdxl-realistic-keywords

[grand cargo]

🚀 🍪 Congrats to @mellow tendon who's JibMix has entered the realm of the SDXL Top 10! Nice rich model! 🤘 💖 https://docs.google.com/spreadsheets/d/1IYJw4Iv9M_vX507MPbdX4thhVYxOr6-IThbaRjdpVgM/edit?usp=sharing

[gloomy lark]

Hunyuan 1.1

[shy kelp]

its ok

Hunyuan is behind the others I think

[gloomy lark]

Really only beaten by sd3 8b on the api

They recently released 1.1 of hunyuan which improved things, and 1.2 is coming shortly.

[shy kelp]

you've gotta compare models on photo realism

its a lot weaker in that area

[gloomy lark]

Photorealism is boring. None of the stuff I do makes sense in a photorealistic setting.

although I guess I could rephrase the trumptopus as photorealistic, let's see.

[gloomy lark]

so that's sd3 8b with multiple photo realistic keywords. it refuses to do it.

[shy kelp]

SD3 goes way more photo realistic than this
you've gotta get the famous guy out of the test LOL

its triggering the censorship probably

here is a good example of sd3
https://old.reddit.com/r/StableDiffusion/comments/1dlvaej/sd3_can_make_real_instagram_photo_somtime/

[gloomy lark]

another from hunyuan/refined

[gloomy lark]

more hunyuan/realvis

[shy kelp]

there's basically no detail at all on their face

I don't understand how that compares to this at all https://old.reddit.com/r/StableDiffusion/comments/1dlvaej/sd3_can_make_real_instagram_photo_somtime/

[vital ermine]

https://civitai.com/models/551577

[meager canopy]

Is it the new Hunyuan v1.2?

[nova oxide]

AI bg changer:

[verbal inlet]

generate image of a art model front facing with hands on waist half clothed posing for artist

generate image of a art model front facing with hands on waist half clothed posing for artist ✨ | sdxl

[peak dove]

I made 8 different images with the SD A1111 Photoshop Plugin - and cut+pasted into this composite

[misty hull]

Man, your images (not only these ones but since weeks) has been so freaking cool 👏

Oh yeah??? Your lately image has been using your new checkpoint ?

[gloomy lark]

Hunyuan 1.1. Apparently they released 1.2 right after I went to bed for the night. 🙂

[meager canopy]

It was uploaded 3 days ago, but I've not managed to get it running in Comfy yet

[half gyro]

Anyone here tried IP adapter

On realistic faces?

[gloomy lark]

https://huggingface.co/Tencent-Hunyuan/HunyuanDiT-v1.2-Diffusers-Distilled/tree/main

This implies last night

There was a captioner, but this is the image model.

[meager canopy]

Ah! https://huggingface.co/Tencent-Hunyuan/HunyuanDiT-v1.2/tree/main

[gloomy lark]

They keep releasing stuff constantly

[meager canopy]

Maybe I've downloaded the wrong stuff, and that's why it won't work.

[gloomy lark]

Oh ok you're right. Well, ultimately we want the distilled anyway. 🙂 it's better quality for the same steps.

The model file is double the size, but it's "faster", but higher quality if you keep the same higher number of steps

So I'm doing 25 steps of ddim/ddim-uniform, but with distilled you could do 13 if you wanted to. I use that extra speed to just keep it at 25

[meager canopy]

I deleted my original install, and can' remeber what goes where now 🤷🏻‍♂️

[gloomy lark]

this is the workflow I'm using currently. So the hunyuan 1.2 distilled is in safetensors format. that's good compared to the old pickle, but the city 96 extra models nodes don't work with that, so I'll have to ask them about updating the nodes.

[meager canopy]

Which directories the models go into 🙂

[gloomy lark]

https://github.com/city96/ComfyUI_ExtraModels

if you go here, and scroll down to hunyuandit, it'll get you going with the right files and directories. i'd say that you want to search for the hunyuan 1.1 distilled model though and put that in there instead of 1.0

[meager canopy]

Ah yes! Thanks 🙂