✨｜sdxl

[heady vale]

2160p UHD single shot gens

[peak dove]

After about 75 attempts - finally MAJESTY appears - Harrlogos2 LoRA

[crisp owl]

Not what I was going for, but.....enjoy

[peak dove]

Good Morning!

[crisp owl]

😅

[peak dove]

Hurray - one I didn't have to throw back!!! 🙂

[woven ibex]

#1010934719455707218

[cyan crown]

new lora is baking

and it seems very good

[peak dove]

ALoRAbaster?

Alabaster?

[cyan crown]

Shadowy Linens

[peak dove]

Waxy Linen

[cyan crown]

https://civitai.com/models/248866

[peak dove]

Rainbow and Underwater LoRAs

Rainbow and Shadowy Linens LoRAs (almost Biblical scenes)

[cyan crown]

wonderful

[peak dove]

Boooootfull!

[south frigate]

Is the A1111 suitable for SDXL? My impression is that it's hard to get good results on it.

[peak dove]

SDXL works wonderfully well on both Comfy and A1111

[cyan crown]

I use A1111

[strong copper]

Spring Is Coming!

[bright valley]

Love it!

Maybe I can help with consistency

I just did a batch of 41 and got 2 misspellings

[peak dove]

The more common the word the easier for Harrlogos. I would like some help on consistency 🙂

[lapis gale]

congrats 🙂

[lapis gale]

intermediate stage 🙂

[stone fossil]

😄

[left dock]

Hello lads! I just wanted to ask something, do you happen to know if it is possible to create posters for an event on sdxl? I could use a hand to make a quality one, or even some tips would be fine. Thank you!

[lusty wolf]

Maybe this? https://comfyworkflows.com/workflows/894a9b88-c129-4eec-9e9f-135bb74a5682

[left dock]

I'm using webUI unfortunately 😦

[bitter juniper]

https://tenor.com/view/ra-gif-20536639

[lapis gale]

depends on the style and event type... e.g. this is simply "((perfect art nouveau poster)) (((saying "event name"))) a cool retro background" with SDXL

[left dock]

Thanks for the reply! Actually I needed the poster for an event in the city, in a nightclub, could you give me some tips? Thanks again

[lapis gale]

or "((perfect art deco poster)) car (((saying "car event"))) a cool dynamic background"
and repeat until the text is ok 😉

upscale to 4k and your poster is ready

[left dock]

and for a nightclub event? What would you write?

And checkpoint?

[bitter juniper]

https://lexica.art/?q=poster

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

But I would just use base version and a strong prompt

problem with custom model is they train them on a database and they may not have thought about poster art

like dreamshaper is really good. But don't ask it weird art styles

[left dock]

do these work on stable diffusion?

[left dock]

Thank youuu

[lusty wolf]

But which webui? of what provider: A1111, SDnext, comfy? There is lots of WebUI's?

[left dock]

a111 sorry, I didn't know😂

[lusty wolf]

Comfy is more versatile. I started with A1111. Did you manage though? Otherwise I can look-up my archive. Anything is possible...

[left dock]

yeah, I'm trying to do it, lemme know if you like it! (ps: it's for a nightclub) Obviously text will be added soon.

[prisma peak]

Watch out, I'm taking SDXL and making a AI recreation of Steamboat Willie since its now in the Public Domain*

In the style of an Anime

This will take me several hours of non stop GPU usage.

Days even

[prisma peak]

115 hours

[prisma peak]

This is what the starting opening transition looks like. I guess the boat is an anime girl now

[hollow kindle]

Can I ask a really dumb question… are the SDXL checkpoints and Loras compatible with SD at all?

I heard they work with SD 1.6… is this true?

[rustic garnet]

no. SD 1.5, SD 2.1 and SDXL are all incompatible to each other

[uncut steeple]

I think there was word about Loras from 1.5 working for SDXL or something, but not all of them

[hollow kindle]

Thanks for the answer. So I can’t run SDXL on my SD

Been putting off installing SDXL cuz I dunt wanna re-make all my Loras… lol

[gaunt dock]

Can anyone reccomend an equivalent of latent layer camera hypernetwork for sdxl?

[earnest rapids]

installing? it's just the model you use isn't it?

I've got 9 SDXL models and 15 SD models on my drive so far, and looking for more 😄

[left dock]

Hi guys, I just wanted to ask you something. Why does it give me an error on xl every time I try to add 1.5 embeddings? PS: I'm using kaggle

[earnest rapids]

as in SD1.5? that wont work on SDXL

[vale eagle]

x-adapter. No code yet

[lapis gale]

depends on the event... i'm only using SDXL since it's so flexible (realisticfreedom).
e.g. this is simply "((poster invitation)) ((saying "nightclub")), people dancing, ecstasy, colors, happiness, haze smoke, stage lighting, action" with SDXL advert style... maybe better not to try for text and add it in photoshop later 🙂

[lapis gale]

same prompt, cubist style (i consider this SFW since they are highly stylized and clothed)

[left dock]

omg so coool

thank youu

[lapis gale]

or analog film style

last one, psychedelic style, also a nice touch. have fun!

[cedar tusk]

Anybody use Comfy via API in python? If so, is there a way to get the resulting image in a python notebook other than navigating to the directory and grabbing it?

[bitter juniper]

hand and sdxl....

[lapis gale]

well, you know, AI ... "a perfect hand" and you get THIS 🙂

even more cursed

well at least it's good for really getting a grip on things 🙂

[vale eagle]

one more finger

with a few attempt

[left dock]

guys do you have any suggestions on how to remove these texts?

[vale eagle]

inpainting

[bitter juniper]

I wonder if there is something I can do in prompt. Or describe the hand ?

[vale eagle]

yes, describe the hand gesture would help

[lapis gale]

i was intentionally posting horrible results for fun - SDXL does hands quite well most of the time, and if not I just hit "generate" again, no big deal with a beefy GPU. I don't believe in prompt magic and 300 token negative prompts and anything, I just go with the flow 🙂

[bitter juniper]

ok thanks

[vale eagle]

cat gundam

[orchid acorn]

Is it possible to edit images instead of generating them? I'd like to edit some selfies of mine

[left dock]

omg! Thank you! Can i ask u how did you do the impainting?

[vale eagle]

denoise the masked area. use watermark, text, in neg prompt

[bitter juniper]

much better than before

The trick was when I trained the lora, I increased the number of repetition at 8

But now it is harder to mix it with other lora...

And I want my Jaffa with a karakesh !!!!

Let me check your temperature...

Hummm... Im close to something...

but impossible to get my jaffa !

https://tenor.com/view/glee-will-schuester-this-is-completely-unacceptable-this-is-unacceptable-unacceptable-gif-24413284

time to pressurize my ai...

[desert bobcat]

I also tried but it replaces it with a blur or puts images, do you use controlnet?

[vale eagle]

no

[desert bobcat]

i tried all type of denoising

[vale eagle]

I adjust the strength and do multiple times

[desert bobcat]

fill, original, latent noise but nothing worked

[vale eagle]

denoise the masked area

add random noise

[desert bobcat]

ahh okok, i will try, thank you ❤️

[vale eagle]

it should be fill? I didn't use webui along time.

[desert bobcat]

i need to leave all text

[rustic garnet]

best is to just overpaint the area in your favourite graphics program and then do normal inpaint

[desert bobcat]

okk, thank youu

[lapis gale]

google for "inpainting", lots of resources 🙂

[left dock]

guys it's like 2 hours that i started fixing his right hand with inpaint. No results at all. Any suggestions?

[light sapphire]

anyone know why my images kep coming out like this? im using sd_xl_base_1.0

[west breach]

change the vae to the sdxl vae

or set to none to use the vae which should be included with the sdxl model

[frail pawn]

Was inspired by this short film: https://www.youtube.com/watch?v=dt4Nz-69_Yk

[worthy orbit]

maybe this could help https://civitai.com/models/200255/hands-xl-sd-15 , sometimes it gives a good base. But hands needs a lot of tricks. Check references and make a morph (or photobash) so it is anatomically correct, then inpaint over until it blend.

The more you can progress in an image editing program, the better. I think it saves time and effort at the end, and get you in better shape for next battles.

[bitter juniper]

I can make a dhd now

but I cannot make a DHD in front of a stargate....

...

I wonder if can inpaint that thing ?

[bitter juniper]

don't work. I will try to modify my database... I wonder why it is harder to train object than character ?

[vale eagle]

might be the object data is less than human form character in the model

[bitter juniper]

in sdxl database or mine ?

[vale eagle]

the base

[bitter juniper]

hummm maybe

But I think nobody has never thought to create a model only focused on objects

[vale eagle]

there is

[bitter juniper]

really ?

[vale eagle]

some product design model

[bitter juniper]

oh yeah I found some on civitai...

https://civitai.com/models/117230/ornordic-modern-style-interior-design

create your Ikea kitchen 😄

hummm, I will try that

my lora should understand what I want

and fill the background with anything it wants

people thinking ai will take our jobs. Clearly have never used ai....

https://tenor.com/view/ncis-angry-pissed-technology-ziva-david-gif-7468619

these things need a lot of effort in order to work correctly

[steady grove]

AI just has to be smarter than the dumbest person and the great replacement will begin.

not that tall of an order tbh

[lapis gale]

start from scratch? 🙂
TBH i find it easier to generate something fresh from the start than optimize an existing result

have a festive cup of cocoa and relax

[ionic dragon]

@lapis gale which is the best model currently?

[tawdry vine]

Hi, I am trying to create sdxl lora for product images
I have 6 product images with white background
But it struggles in color
After I finish training it sometimes generate different color products
does anyone can help me?

[meager canopy]

Easiest is just to remove it 😄

[molten breach]

Prob shouldn't drink that though....holly berries are poisonous 🙃

[agile haven]

Hi, I have a problem I would like to install SDXL, but I have an RX 7800 XT, and I don't know how to do it once I download the SDXL folder to put DirectML in it.

Can people help me please

I don't know what to do with this

DirectML (AMD Cards on Windows)
pip install torch-directml Then you can launch ComfyUI with: python main.py --directml

[left dock]

😂 😂 😂

[meager canopy]

Any better?

Not perfect, but something to work with.

[left dock]

thank youu

it's perfect

[lapis gale]

depends on your style and purpose. i would always go for SDXl for better variety and prompt understanding, but some concepts are unknown to SDXL and better served with SD 1.5 models (mostly due to the NAI leak). personally i'm nearly always using realistic freedom SDXL, the september version since the more recent release is too much photo oriented in my opinion and drowns out the other possible styles (love the SDXL style selector)

[lapis gale]

may be too small a training set

[lapis gale]

with embeddings, it was also about using tokens during training that should NOT be part of the "yes this is XYZ" learning - e.g. if you want a generally usable embedding but train it on people wearing t-shirts, add the "t-shirt" to the learning prompt so the training realizes that the shirt is not an integral part of the concept to be learned

[peak dove]

GPT Mixlab Node 5 - ComfyUI

[worthy orbit]

There is a new ControlNet preprocesor that automatically makes a depth map of the hand, check it out

I haven't try it but the idea is super good. Previously the best method for fixing hands was to use a depth map hand image, but it took a lot of work to generate that model or to use one from a library (as you have to choose the correct pose and angle, and manually place it), so lots of operations.

This pre-processor generates said hand depth map based of the hand, place it in the right place, so it is like an automatic depth hand library.

[bitter juniper]

finally work

not perfect

but it works...

by itself it is magic

I have a lot of garbage but when it works, it works !

[ancient cairn]

Is there any normal map estimation controlnet for sdxl? Want to use normal controlnet in sdxl workflow but can't find anything

[peak dove]

Pixart Alpha

[ionic dragon]

dayum, looks great
are you working on a new model?

[formal panther]

This is going to sound silly and dumb, but..
how do I set Auto1111 to use sdxl?
I have everything downloaded and installed, but I forgot what to set so I can actually generate with it without giving me an error

[peak dove]

No, iits here ... https://pixart-alpha.github.io/

https://github.com/PixArt-alpha/PixArt-alpha

Pixart Alpha

[stone fossil]

Tinfoilisa.

[lucid rain]

You should be able to just load the model and go.

[fair yacht]

Sorry if this is a dumb question. I tried searching, but I don't know how to phrase it for the search bar.
In short: How can I tell if a Lora, Lycoris, or whatever like that, is going to work with my SDXL models?

I just recently switched to SDXL, from SD... 1.5, I think? Or something like that. Allll my loras used to just work fine. Now, almost none of them even show up for my SDXL models, and the ones that do show up, don't seem to do anything at all. I'm trying to figure this all out, but it's difficult, as a noob that just types words and clicks generate
p.s., sorry if this is the wrong place to ask.

Oh, if it matters, I'm on Automatic1111 and an RTX 3070.

[native knot]

SD 1.5 LoRAs don't work with XL. You need different ones. End of story.

[fair yacht]

Yes, I figured that out. Sorry, I meant, when I'm looking for new Loras to get.. How do I know if they'll work with SDXL?

[native knot]

Wherever you are getting them from will tell you.

For instance, Civit shows it here on the page:

Tensor.art shows it here:

[fair yacht]

Ah, gotchya. Thank you!

[peak dove]

Pixart-alpha, a better SDXL

[peak dove]

ComfyUI Mixlab Node 1 - makes two photos, masks them, then recombines them - so that there are always two models in each combined photo

New combined photo, masked photos (black areas), individual originals

[peak dove]

Try jan.ai - it is a freebie - and can do great things for your prompts, and is almost as good as ChatGPT - (but jan.ai is FREE!)

[peak dove]

A1111 NightVisionXL model - prompt (via jan.ai) = Craft a complex and dynamic portrait in the style of tenebrism, as characteristic of Catherine Abel and Ewa Juszkiewicz, representing Charles and Camilla as a blissfully married couple. Use a dark cyan and amber palette, play with chiaroscuro and light to perfection, striving for the juxtaposition between cartoon violence and brutalist architecture. Aim for a joyful and optimistic mood, despite the intense dramatic contrasts, embodying the Romantic academia aesthetic. Incorporate characterized animals, and create captivating, award-winning figures. Even in this unusual coupling, their love and mutual respect should be palpably sympathetic

Change "Charles and Camilla" for Batman/Robin, Spongebob/Venom, Joe/Kamala, Peppa Pig/Chthulu ... etc etc

[bitter juniper]

finally get my Jaffa with Karakesh !

https://tenor.com/view/robot-devil-torture-satan-boulder-gif-4788082

stupid ai...

it takes too long to get something so easy

[deft kestrel]

word "if" 3d

[bitter juniper]

There can only be one Ra!

[ionic dragon]

did SD just die?

[bitter juniper]

Footages of the first Stargate movie made in 1994

[jolly brook]

Lookin more like 1894

[bitter juniper]

anyway, not our century...

[bitter juniper]

training a lora for face seems completely different.

[bright valley]

@wet nacelle

[wet nacelle]

Wicked

[lapis gale]

right 🙂

[peak basalt]

What dea go on

[bitter juniper]

I dont know why but it is harder for me to create a Lora about movie stars...

less flexibility, more garbage

I cannot blame ai this time, I think I just haven't figuret out yet what are the perfect parameters

[peak dove]

Sytan's Workflow

[bright valley]

@peak dove

[frail pawn]

New Desktop Wallpaper:

[frail pawn]

Kit-bashed using a few assets in Zbrush:

[peak dove]

Very nice! Thank you!

[bright valley]

Np man

[peak dove]

Using Harrlogos2 on my upcoming Valentine's Stock

[lilac wren]

wonderful, I'd be curious to know your settings for having Gandalf look so much like him, and for his seat on the tiger to be so good.

[stone fossil]

The meta data is in it and quite simple, however it's done with a unreleased lora for now.

lora:3U4rt_001-step00025500:1, 3U4rt, gandalf riding a tiger in the jungle, using heavy brush strokes
Negative prompt: double tails, lose tails, floating tails
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 2668400334, Size: 1536x1024, Model hash: 80f46bd4bf, Model: Paradox_2.0_180000_Steps, Seed resize from: 1024x1024, Lora hashes: "3U4rt_001-step00025500: 74a2602206d6

[peak dove]

Gandalf riding a tiger in the forest, style of Henri Rousseau ... Rousseau is good with forests and tigers!

[stone fossil]

Tigers can be mounted in various ways.

[signal pagoda]

can I use SDLX as base model and a LORA trained on V1.5?

[stone fossil]

No.

[signal pagoda]

sheesh

guess I'll have to download 1.5 base them

[stone fossil]

Or retrain the lora.

[signal pagoda]

yeah I guess I could do that for the ones I trained, but most lora available online are for 1.5 so...

using SDXL feels kinda limited for now regarding community assets and usage

[stone fossil]

I love XL there is no way I can go back to 1.x or 2.x.

[signal pagoda]

or maybe I'm not savy enough to make the best use of it since i'm a nwebie

[stone fossil]

But it depends on what you want out of it I guess. 🙂

[signal pagoda]

Right! For some weeks now I learned the basics before jumping into pre-packaged workflows and stuff. But now I kinda want to see whats out there available by the community

and feels like 1.5 is the most adopted so far. lots of content, guides, workflows and stuff for it

[stone fossil]

1.5 is ofc the most adopted mostly because of hardware limits of XL.

But that will shift in the future. 🙂

People will get better hardware.

[signal pagoda]

Yeah! Today i trained my first lora on XL. 38 images with 40 steps(?)... it took almost 4 hours on a 4070 and the lora filesize is HUGE compared to ones i've seen on civtai... 600meg

[signal pagoda]

true!

[rustic garnet]

that's rather a problem of your parameters

[signal pagoda]

This is dope! Made with XL?

[rustic garnet]

my sdxl loras are always just a few megabytes in size

[stone fossil]

Yes ofc.

[rustic garnet]

if you use a too large rank then filesize and training time will increase. That has nothing to do with sdxl or 1.5

[signal pagoda]

Oh, the Network Something option?

[rustic garnet]

dim

should be between 4-48 depending on what you train.

[signal pagoda]

hmm

i used 128 / 64 ☠️

[rustic garnet]

try 8

[signal pagoda]

i was training a face

I'll give it a try!

[signal pagoda]

mind to share what sorcery you used to make the painting style?

lora or checkpoint i mean

[stone fossil]

The checkpoint is Paradox 2.

The lora is yet to be released.

🙂

https://civitai.com/user/Thaevilone

[signal pagoda]

Damn it is getting amazing tho!

[signal pagoda]

I'll follow for sure :­D

[peak dove]

"Dangerous Site"

[peak dove]

"Dangerous Site"

[hasty tapir]

do not visit that link you will get hacked!

[stone fossil]

I do not see the problem I visited the link any my SD output seems just fine.

Let me check my phone.

Yep, nothing unusual.

[hoary saddle]

who has a good ipadapter for putting 2 images into a new image?

[peak dove]

Scott Dettweiler

[hoary saddle]

found his, ty Torcello

[lilac wren]

I remember the release of 1.4, where I tried so hard to put Trump on an Alpaca. Today it's so easy...

[lean forge]

it made the earth flat

[vale eagle]

working on a hand gesture finetune.

[dark crypt]

Why are some SDXL models so much slower than others?

[dark crypt]

I'm gonna try --medvram and drop --no-half... maybe that will help until I need those things.

--medvram isn't cutting it even with 12gb vram. Yay Radeon.

[stiff moat]

I'd really appreciate some help with generating a cobra logo.... SDXL keeps making snakes.

Most go like this, where the hood probably keeps getting absorbed into the 'snake'

Although this one is pretty funny

a black and white drawing of a cobra on a white background, with a black outline of a cobra, Aya Goda, an illustration of, cobra, gradient, gradient background, grey background, greyscale, monochrome, no humans, open mouth, teeth, hooded cobra
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 2, Seed: 2300522753, Size: 768x768, Model hash: 673887ace8, Model: opendalle_v11, Clip skip: 2, Version: 1.6.0

[south flume]

Cobra is snake:) funny how it gets non-english...

[stiff moat]

Cobras have a hood though, whereas snakes do not 😩

[south flume]

Think is a coincidence Shelby logo is a snake? Think again....

[dark crypt]

Try using ControlNet with a source image of a cobra? Depth or one of the others perhaps. That should give the right shape

[snow wasp]

Adetailer is turning everyone into stoic anime boys

[dark crypt]

What do you use to get Warhammer images? If that's what it is.

[snow wasp]

I trained a Lora

I will release it soon, but its only trained on cadian guardsmen since its my first lora

[cyan crown]

The most useless Lora in history is ready: Cappuccino Style. https://civitai.com/models/259619

[lusty wolf]

If you want to play with 'Madness', here we go, at your own risk...

[keen turtle]

hwy guys! I'm trying to use the sdxl typescript rest api, and im getting this error:for stable-diffusion-xl-1024-v0-9 and stable-diffusion-xl-1024-v1-0 the allowed dimensions are 1024x1024, 1152x896, 1216x832, 1344x768, 1536x640, 640x1536, 768x1344, 832x1216, 896x1152, but we received 400x289',

this doesn't happen when i use the sandbox, which uses the same typescript api. Any thoughts?

[zealous pine]

@keen turtle do you use stablestudio?

[cyan crown]

a new Lora is baking. Tomorrow it will be ready

[soft bone]

s k i n

poke her face

[heady vale]

the model or lora you are using is placing the ears too low on the head. the top of the ear should be closer to the temple on the face

[soft bone]

i thought the same, checked her originals, and no thats just her. same with slight strabismus in eyes. she also constantly looks up a bit

here's an original

[lilac wren]

have lot of work... But you will enjoy it guys

[snow wasp]

looks dope, what is it?

[keen turtle]

no ive been using the api

[pure crystal]

(logo illustration, line art:1.4), hooded black cobra outline on a white background, Aya Goda, an illustration of, cobra, gradient, gradient background, grey background, greyscale, monochrome,open mouth, snake fangs try this instead

[carmine gale]

Does anybody know a version of dynamic prompts for comfyui that works? https://github.com/adieyal/comfyui-dynamicprompts

[dark crypt]

How are you getting good text?

[lusty wolf]

SDXL is good with text, but you can help it with a Lora:
https://civitai.com/models/176555/harrlogos-xl-finally-custom-text-generation-in-sd

[gaunt dock]

how do you guys decide which checkpoint to use?

[lucid rain]

You just find the one that matches the style or quality you want.

[high skiff]

I have been trying to figure out how to use controlnet tile in Comfy, and I am just not understanding how its supposed to work with just conditioning

All you can do with controlnet in comfy is conditionings, and I have no idea what I am doing

don't really know what all to do with this TBH

and I can't find anything online for how to use controlnet tile in comfy

[high skiff]

got it working

my dumb ass was sending the conditionings to the wrong sampler

🤡

in all fairness, I am stupid sick AND on cold meds lmao

[mellow tendon]

Is there an SDXL ControlNet Tile for photorealistic images out yet?

[high skiff]

I believe SDXL tile has been out for several months by now

[cyan crown]

One of me best Lora : https://civitai.com/models/259682

[lusty wolf]

https://docs.google.com/spreadsheets/d/1IYJw4Iv9M_vX507MPbdX4thhVYxOr6-IThbaRjdpVgM/edit#gid=0

[cyan crown]

https://civitai.com/models/261219

[open aurora]

I started with SDXL for the first time today, please note that I create stuff that some people might find disturbing.

What if cosmic horror met cyberpunk?

Really liking how intence loras are with XL models. 🙂

[shy kelp]

Hi guys,
I am new to this community. I want to train Stable Diffusion on pictures of my face and then want to create images with my face in it. Is this possible? If yes could anyone guide me how to? I trier different GUIs, watching different videos and reading multiple articles on Medium and Github but didn't understand the code much. There is so much to study and understand, especially all the hypermeters.
It would be great if anyone could guide.

Thank you.

[rustic garnet]

is it? I haven't found it

[fierce hollow]

there's no sdxl tile afaik, closest is kohya's blur controlnet but it's not that good

[gaunt dock]

Anyone got tips on making an image with character directly facing the viewer, so half the face is on each side of the image?

[rustic garnet]

just take any image that is like that and make img2img with 90-100% denoise

[peak nimbus]

I trained a lora for SDXL but it only appears for 1.5 stuff 😐

why? like where would be the problem

[west breach]

there is a sdxl tile controlnet trained for anime: https://huggingface.co/bdsqlsz/qinglong_controlnet-lllite

[polar epoch]

I really need a new gpu. Took my 3090 55 minutes for this near 8k res image lol. 768 resampled by 1.5x 5 times.

Sadly this result was too low denoise at 25, so testing 30 now lol.

[peak dove]

ComfyUI and some really cool LoRAs!! 🙂

[icy brook]

Is there a site where I can use the rembg (remove background) extension or similar that's in automatic1111, but for free?

and preferably without any sort of login

[peak dove]

remove.bg

[icy brook]

found it also and tried. but adobe express does it for free however many images wanted. that is with login though 🙂
https://new.express.adobe.com/tools/remove-background

[icy brook]

hey! long since I spoke to you 🙂

but maybe you hang out in here all the time, bc I rarely do these days

[stone fossil]

Hah im always around. 🙂

[icy brook]

I just go to the show channel on here these days

but not much happens there even 😄

[stone fossil]

I like this one a lot because I do not think anything I post is "show worthy". 🙂

You made any new cool loras?

[icy brook]

depends on when we stopped talking.

8 newest to oldest

[peak dove]

Since Christmas all the SD channels have been relatively quiet

[stone fossil]

I think ive seen you post em all. 🙂

[icy brook]

Aether Snow went very well I think

concerning realistic detailing of texture

[peak dove]

I'm using Aether Aqua, but find that Liquid Water is more defined

[icy brook]

defined? you mean in terms of subject composition?

[stone fossil]

Im here these days having a lot of fun with img2img.

Just gradient and maybe a silhouette.

Noise it up a bit and go. 😄

[icy brook]

the thing is I set out to try and make stuff as photoreal as possible both in terms of texture and composition. since there are no clouds, no water or snow etc that takes the shape of these things naturally most of the time, too defined features does not make it look realistic. liquid water does define shapes better, but I don't think that's realistic. it needs an element of abstraction to make it come off as real. I deliberately balance my dataset that way.

some like it the other way, and that's fine. but it looks more painted than real for sure. also: water isn't always bluish, lighting changes color nuance with water a lot depending on context.

[icy brook]

fun is fun 🙂

[stone fossil]

Its quite nice to use and direct a bit, almost like controlnet wise.

But more leaving creative part. 🙂

Prompt: Gandalf.

Nothing less, nothing more.

[icy brook]

Are you good with inpainting?

[stone fossil]

I used it but not much.

So no not good.

[icy brook]

I just try and remove a person from an image. I can’t do it 😅 And I know a lot about ctrlnet but can’t even fix that 🤣

[stone fossil]

Oh I can do that lol.

I loved to take like a screenshot of a doom game.

And then strip it down till there is nothing left. 😄

Fun little games. 🙂

[icy brook]

I will dm you now instead

[naive ginkgo]

I'm newbie in Stable diffusion and this is my latest creation after few days to this area. What do you think? What to do better, what to focus on? 🙂

[peak dove]

Cool!

[stone fossil]

Focus on what ever makes you happy. 🙂

[grand mural]

Why it gives pics as files
I can't download it at all

[polar epoch]

Shit. Hate that i can't do 0.295 denoise. Top is 0.29, bottom is 0.3 lol Takes 55 min per attempt, so gonna try lowering latent upscale and see if that can help lol
https://image.duckers-web.site/hEja1/cIserOyu56.png
https://image.duckers-web.site/hEja1/zELASiJe58.png

[clever verge]

Discord is broken at the moment.

[little umbra]

Yo guys i open sources my miniSDXL model.
Built upon SDXL, of course ❤️

Enjoy :)

https://huggingface.co/kopyl/miniSDXL/

[naive ginkgo]

🙂 I meant more like what to focus on to make it looks even better, what to do differently etc 😉

btw. is there a way how to modify the picture a little? For example change color or style of hair but with the same face or slightly modify the eyes or lips? I use ComfyUI..

[dark crypt]

Use the same seed or change the number slightly and modify the prompt slightly to see what changes and how much. Work with weights like (red dress:1.3) and adjust the weights and see what changes. You'll want to experiment with different models and different settings, and even changing the pixel count can drastically change an image.

If you want the same face, save the image and use ReActor or another face swapping tool. Then you can change everything else but keep that consistent

[naive ginkgo]

thanks for tips! changing seeds - here change the Control after generate to Fixed and change the noise seed by 1 for example?

[dark crypt]

Yeah, give it a try and see if it makes a slight change. Change nothing else to test it.

I don't use comfu but I think you want increment to have it do this automatically

[stone fossil]

Random ofc.

[dark crypt]

Sundance is learing about making slight changes to impact the image.

[stone fossil]

Then still random till you get something nice, lock seed and then bump up the steps.

[dark crypt]

Yes, that's what I was saying.

[stone fossil]

It will give slight variations.

Thats 30 vs 120.

[naive ginkgo]

for example what to do if I just want to repair the green freckles but keep everything else the same?

[stone fossil]

Haha SD is not a restoration tool

You can inpaint some such.

[dark crypt]

Use ReActor, save the face and then in the new model, reduce the number of freckles.

Or you can use the same seed and put negative prompts of (freckles:1.3)

[naive ginkgo]

thanks! will try to prompt around and will look what the ReActor is and how it works 🙂

[clever verge]

I guess it's because you prompted green eyes?

[dark crypt]

Inpainting won't work without the same negative prompts on the inpainting, it will just do the same thing again unless you do so much that it just redoes it all from scratch pretty much.

[naive ginkgo]

yes

[clever verge]

If so you could try to inpaint green eyes after

[dark crypt]

I would first and foremost try using the negative prompt of freckles.

It's the quickest and easiest way to get there.

[naive ginkgo]

I have freckles in positive prompt to have them there..

[dark crypt]

Then lower the weight of the freckles in the prompt with (freckles:.5)

[stone fossil]

I still think the most fun is txt2img, alll mine are. 🙂

0 touch edit AI spit outs.

[clever verge]

I think he wants both freckles and green eyes

[dark crypt]

Ah, I thought he was saying there were too many freckles.

[naive ginkgo]

thanks for all the tips! will try them all. But it tooks me around 10 minutes to generate 1 image so it will take some time..

[naive ginkgo]

just the green freckles bothering me..

otherwise I wanted the green eyes and some freckles (not green ones 😄 )

[clever verge]

Yeah, it's normal for the model to bleed between concepts unfortunately.

[dark crypt]

Are you using SDXL? Try using a turbo model if it's that slow. Otherwise I'd use 1.5 and lower the resolution. That's really harsh.

[stone fossil]

Freckles fetish hmmmm. 😛

[naive ginkgo]

wouldn't say that.. just found out today it adds some kind of reality to the pictures 🙂 less of the extremely smooth skin 😉

[stone fossil]

I do not judge. 😛

Must say they can look kinda sexy.

🙂

[dark crypt]

What's your actual prompt?

[naive ginkgo]

yep XL.. this one has juggernautXL as base and SDXL refiner as a refiner..

I tried some LMC yesterday but it didn't work at all 😦

[naive ginkgo]

positive: 1girl, portrait, red hair, (beautiful:1,2), perfect face, freckles, green eyes, masterpiece, high quality, 4k, 8k, photoquality, nikon, high quality
negative: text, watermark, bad quality, ugly, (bad hand:1,1), disgusting,

[dark crypt]

Consider moving to 512x512 on an SD 1.5 and then when you find what you like, upscale it.

[naive ginkgo]

isn't the 1.5 worst than the SDXL?

[dark crypt]

You have commas instead of periods... it should be (beautiful:1.2)

[dark crypt]

It's lower quality but much faster and can be upscaled.

[naive ginkgo]

oh, thanks, didn't know.. will redo

[dark crypt]

Ask CS1o on #🤝｜tech-support for details on how to upscale well

[naive ginkgo]

haven't got to the upscale worklflows yet.. for now trying to learn how to prompt and playing with that.. 🙂

[dark crypt]

Your tolerance for 10 minutes an image is much greater than mine! lol

[clever verge]

something like turbovision could save you some time but still get good quality

[naive ginkgo]

well I would be glad for 2 seconds like on RTX 4090 😄 but just started and using my M1 Macbook.. start generating and in the meanwhile reading some articles or watching some tutorial videos on youtube to learn how to do it properly 😄

[stone fossil]

I would always upscale using Chainner after hand if wished for, up scaling images you have no idea what the outcome is, is quite wasted time.

[dark crypt]

A few modifications should really help, though. Switching to a turbo model cuts down the number of needed steps from like 30 to around 4. And it's good quality.

[naive ginkgo]

so this one https://civitai.com/models/215418/turbovisionxl-super-fast-xl-based-on-new-sdxl-turbo-3-5-step-quality-output-at-high-resolutions ?

how is it used in the "base and refiner" workflow?

[clever verge]

There are several ways to use that as a refiner for another model. Either you do a latent upscale or and img-to-img with a denoise after

[stone fossil]

Dreamshaper is a awesome refiner.

[dark crypt]

I can't say, I don't use ComfyUI, I just turn on the DreamshaperXL Turbo and ignore the refiner and go for it.

[dark crypt]

I haven't used that specific model.

But you're on the right track

[naive ginkgo]

so the Dreamshaper is better https://civitai.com/models/112902?modelVersionId=251662 ?

[dark crypt]

I couldn't say better, I just know I've gotten good results with it.

Try both if you have good download speeds

[naive ginkgo]

ok

[stone fossil]

And if ur brave enough try Paradox 2. 😄

[dark crypt]

Important, CFG scale must be between 1 and 2 (or 3 to 4 for stylized stuff), and sampling steps should be around 4-7.

[dark crypt]

Then, when you have that, move the convo to #1178805966335119472 if that's screwing up. lol

[naive ginkgo]

btw. what should I learn and focus on if I want to create an instagram model like this one https://www.instagram.com/fit_aitana/?next=%2Ffourhandsfurniture%2Ffeed%2F&ref=Dirty%2BInstagram&hl=cs ?

[dark crypt]

Sorry, not my area. No idea.

[naive ginkgo]

ok

[peak dove]

Use ComfyUI and Portrait Master node

I was using Sytan's w/flow, and three LoRAs

[naive ginkgo]

I was thinking about what techniques to learn, how to make the one character and generate her in different situations etc.

[naive ginkgo]

btw. this is quite nice style!

[clever verge]

Torcello is the master of styles

Regarding the instagram model my guess is that they have trained a model with images generated by a model

or some kind of ip-adapter or a combination of everything

but when you have your model it's easier to put it in different environments

[naive ginkgo]

how difficult is it to train a model like that?

[clever verge]

It's doable if you put the effort in.

[naive ginkgo]

great, I don't mind putting the effort in

[wraith ice]

/drame big house

[naive ginkgo]

green freckles are almost gone, but still generating portrait even when i put there the "full body:1.1" prompt 😕

[clever verge]

if you prompt for clothes it will change perspective too

[naive ginkgo]

added some body, but still not full body.. and significantly worse quality of face..

[clever verge]

seems like it does what's it supposed too. to recover detail you have to do like a latent upscale but that will be hard on your hardware

prompt feet or legs or something else, look at civit what people have prompted for and what perspective that resulted in

[naive ginkgo]

will check the civit

[polar epoch]

As AITemplate hasn't been updated for months and doesn't work on newest comfy, is there an alternative that can speed up gens? As AIT sped up gens by 46% for my 3090

[ocean ermine]

This is quite possibly the most beautiful unintended generation I've ever had in my year and a half using SD. 😅

[naive ginkgo]

what should it be when you mentioned it was unintended? 🙂

[ocean ermine]

I had put the prompt in the wrong spot of the node, so in essence it was prompt free, heh. You know you have a good workflow though when you hit generate without a prompt and it gives you a stunning result.

[naive ginkgo]

so quite good workflow 😄

[dark crypt]

Have you tried ADetailer? Also, change the ratios so it's tall and narrow or you can do something like this as a fun trick.... Prompt: A path through a forest, [1girl(with the rest of your prompt):2] .... The brackets and :2 means that it won't generate on the first step, so it will put full focus in the background with a path, then your girl will show up in the existing background, and that's more likely to show full body.

[dark crypt]

Here's a good time to use the ReActor I mentioned. Take the high quality face image, then use that image in ReActor to replace the generated face on the new pic.

[dark crypt]

But you should possibly weight the remaining prompts, sometimes it just ignores the step 2 prompts and just gives you the path in the forest.

Yet another option is to just build a background and then add a figure with inpainting and then also replace the face with ReActor. That allows a small part of the picture to benefit from high resolution. There are so many ways of accomplishing things that it's hard to say what way is best. heh

[cyan crown]

https://civitai.com/models/263504

[naive ginkgo]

will try that later, now I'm trying to do a workflow with upscaling. Created this simple one but works quite fine. 1024x1024 took 333 seconds 🙂 I tried using different model for upscale part (SDXL refiner) but it lefted a lot of noise in the final image so now playing with this one 🙂

[clever verge]

There are many ways, this is what I mainly use.

I mess around with the settings depending on what I do but basically this.

width and height are if I go manual, so I don't use 512x512 for XL. 😄

[naive ginkgo]

wow that's far beyond what I can understand now 😄 too many nodes I've never seen 😄

[clever verge]

it's bottom up, so in reverse

pretty similar what you do in your flow, I just have a image to image upscale at the top. this is normally inactivated unless I want to save an image.

[naive ginkgo]

is it better to upscale decoded image than the latent image as I did?

[clever verge]

I'd say no, but each time you do a latent upscale there's a possibility that you change the image quite dramatically. That could be something you may want or may not depending on the use case.

[naive ginkgo]

really? I thought it's always the same immage when I keep the same seed. And in quality point of view is there any difference?

[clever verge]

Latent upscale is like you would just resize an image to make it larger. You end up with a latent image (/8:th of the pixel image) thats been enlarged but no new information.

So you pass that through the ksampler with some noise added back to have the neural network to remove it and hopefully have something similar again.

[naive ginkgo]

so if I don't add any noise to the second ksampler it would just enlarged it but the image quality would be poor as I just duplicated the pixels (like digital zoom on photography)? so I need to add some noise to make it generate new pixels to the blind spots?

[clever verge]

Yes

[naive ginkgo]

great now I understand, thanks 🙂

[clever verge]

It’s a bit more complicated but in essence what you do.

[naive ginkgo]

and what difference would it make if I use different seed for the first and second ksampler? would the second one start all again generating whole different image?

[clever verge]

It would just add other random noise, you could hook it up to see what difference it makes. If you don’t make the second pass 100 noise (defeats the purpose) you still have something that will remain and guide the result of the denoised image

So the random noise in the second pass will affect the result less than changing noise at creation of the empty latent image

[naive ginkgo]

ok thanks 🙂

and thanks you all guys for todays guidance! going to sleep now and will return tomorrow

[cursive tree]

Is there any sort of simple tutorial on how to get ControlNet working with SDXL on A1111?

[gaunt dock]

If I am using auto1111 with sdxl do I need to download a special inpaint model or should I just use the regular txt2img checkpoint?

[peak dove]

https://youtu.be/bsAwtEbVStM?si=jaJJ6RG1QowozyNw

[gaunt dock]

I will say openpose with sdxl/controlnet is hit and miss, you do it the same way twice and one will work and the next time it wont, needs constant playing with, depth and canny work great though

[peak dove]

Unsampler?

[peak dove]

Portrait Master ComfyUI SDXL

[lusty wolf]

Thought it looked familiar ... 💁🏻‍♂️

[ionic gulch]

@lusty wolf don't know where the source image (the female) is from. it was in my input images for a while now - maybe from something (image or prompts) you posted in the past.

the male version was created by using unsampler from the source.

[ionic gulch]

yes, i'm getting source images from here and other sites and use them for my ip adapter, img2img and now unsampler experiments (or for general tests of my workflow)

the other source image (the girl with the flower suit) is also from here or another source afair. don't remember that i created it myself - at least not from a pure text prompt.

[naive hare]

mmm needs more feathers 🙂

[naive hare]

trying to animate diff it at the moment, so fiddly getting the settings right

[stone fossil]

Here not into video yet because the limitations are quite harsh.

[native knot]

R2 needs a fill-up sometimes.

[stone fossil]

Perfect idea I love it. 🙂

[naive hare]

heh, should be behind the counter serving other droids or something 😄

[naive ginkgo]

wow! how are those made?

I let it go for an hour or so generating the same vague prompt with randomized seed and I'm quite surprised the outcome is really simmilar one to each other. Almost all are close ups, not even portrait, none of them is full body photo. The most surprising fact is that many of them have heterochromia in eyes. I did some heterochromia attempts earlier today, but there wasn't anything about heterochromia in this prompt not even anything about eyes in general..

[stone fossil]

Well its quite easy.

You take a silhouette and a gradient.

Then you image blend these 2.

[naive ginkgo]

any ideas why is there so little variability? every one has brown eyes, etc.

[stone fossil]

And then we run first and 2nd pass.

[naive hare]

i love gradients 😄

[stone fossil]

Me 2 lol.

[naive ginkgo]

thank you! that's quite far from where I'm now with my ComfyUI learning 😄

[stone fossil]

So.

You can also use a pure black backdrop.

And use img2img as text2img.

But the backdrop will guide it more towards the backdrop.

This is why it will listen good to gradients also.

[naive hare]

@Sundance, maybe some of my tutorials might help you out? https://www.youtube.com/channel/UC6e4Bgtt63X05SBApnOBtvg - no discord you cant stop me

[naive ginkgo]

thanks, will take a look 🙂

[stone fossil]

😉

[naive hare]

fundamental ones are good if you are a bit newer to things, the noise one is really cool 😄

[stone fossil]

Just prompting on top of gradients is badass.

[naive hare]

have you tried out comfyrolls radial gradients?

[stone fossil]

Yes we can as noise also. 🙂

No.

But just black backdrop is also very cool or white.

[naive hare]

yah, just got to avoid full solid colors otherwise you have to go really high denoise

[native knot]

Disney/Paramount/Embracer crossover

[naive hare]

I combine vextra's gaussian noise with my gradients, lets it add a bunch more detail 😄

[stone fossil]

Yes.

SD is the best tool since mspaint.

😄

[naive hare]

ive been messing with photoshop integration, its soooooo good

[stone fossil]

I rarely use PS hehe.

Have it installed but its not needed in my workflows.

Yet its a sick tool non the less, cant deny.

[naive hare]

im lucky to be trained in all of that and have a digital tablet XD

[stone fossil]

Looks like something Piscabo would make. 🙂

[native knot]

Disney going all out this year with their crossovers.

All prompted, btw...no tricks.

[naive hare]

ive always thought the space guy looks like earthworm jim.

[native knot]

Buzz Lightworm Jim

[naive hare]

😄

[stone fossil]

Can I get a buzz buzz.

[naive hare]

probably showing my age there, the kids shouldnt know about earthworm jim lol

[ionic gulch]

they only missed bad controls :)

[naive hare]

hahah!

i think I broke my thumbs off trying to finish EWJ

[native knot]

That was an excellent idea, btw...nice crossover character.

[naive hare]

😄

because why not share some actual art using img2img 😄

[stone fossil]

PErfect!

[naive hare]

original lol

[stone fossil]

Slap this in and use the same prompt. 😛

[naive hare]

i would but unfortunately im rendering a loooong animation atm

[stone fossil]

1536x1024.

😄

[naive hare]

workflows embedded if you wanna grab teh prompt tho

[stone fossil]

Nah was just throwing crazy ideas. 🙂

[naive hare]

i use a slight perlin noise plus gaussian noise over my sketch the promt itself is like... just for the features I want so it actually jsut makes lots of teeth and stuff on a plain image without any sketch underlying it

[stone fossil]

Ofc I use it.

On the 2nd pass.

[naive hare]

heres somethign to try @stone fossil blurr an image until it looks like that gradient with a few random colors and stuff in it, see what you can make with it 😄

i occaisionally just grab random stuff off google and blur it completely to generate off 😄

[stone fossil]

Its better to do so in your workflow.

[naive hare]

i mean true

[stone fossil]

Well better, more lazy. 😄

Because we can just gen on top of this.

😄

But its even more lazy if your workflow can gen this. 😄

So u dont have to download it lol.

And trust me, I am very lazy.

[native knot]

Iron Lightyear

[naive hare]

preaching to the choir

[stone fossil]

Noice. 🙂

[naive hare]

I usually use 3 color gradients, and most of the time I have a full random gradient, XD but lately ive been using clusters of random color radial gradients with random sizes, i like blending in some perlin and adding some pixelated noise on top 🙂

[south horizon]

pretty neat

[stone fossil]

https://github.com/XmYx/ainodes-engine

[sturdy abyss]

official support on Deforum discord https://deforum.art/

[naive hare]

this is what im messing with atm 😄

[stone fossil]

Many respect to the deforum crew. 🙂

[native knot]

If Woody looked a bit more like Tom Hanks.

[naive hare]

the only celebrity i bother generating

hes just hillarious when you put batman in every day situations while looking sad

[stone fossil]

Perfect.

[naive hare]

some gradient batmans at different denoise strengths 😛

[lusty wolf]

No you right, you did not create it, I did...(with my original text prompt) 💁🏻‍♂️ You should at least give recognition I think...

[ionic gulch]

@charred bluff i've removed them. sorry. didn't expect anyone to bother about using images as source for img2img experiments. and as i don't know the sources (in most other cases) it's better to remove them all.

[naive hare]

:<

you can borrow my stuff if you want @ionic gulch 😄

here, free dragon 😄

[ionic gulch]

thanks, but it's to complicated to handle "author" per image. guess i will just do some img2img / ip adapter techniques with enough differences from the source in the future for all images i store in my input folder. so they are closer to the used model/style than to the "original" image.

[naive hare]

my advice if you are going to remix other peoples art, it is a great idea to credit original. me, i dont care about the credit lol even though I did draw the base image for hte above dragon, its like 40% not my work anyway so eh.

[ionic gulch]

in many cases "images" are just pulled from open source checkpoints, loras, etc. guess very few people really draw templates, masks, etc. themselves to create them. that's why i didn't bother to record the sources in the past. but if some people don't like it i will change my behaviour in the future. it's not too hard to create something similar with ip adapter or unsampler, that is different enough to avoid situations like that.

[naive hare]

Ive had to get in touch with photographers a few times to source images, contact people about various artworks to get permissions and stuff in my line of work and legit, most people are like oh you want to use it? Awesome! just stick my name somewhere 😛

meet some cool people by reaching out to artists too 🙂

[ionic gulch]

i understand that for commercial usage or if someone really made some drawing as template like your dragon, but most pictures on here aren't created like that. it's just some prompts that pull everything from open source checkpoints.

[naive hare]

I mean it depends how your using a thing, even then though you can literally get sued lol, dude I knew madea free to play game, wasnt earning anything and used some guys freely provided models but didnt follow the licence which said put his name in credits. dude got sued was so silly.

[ionic gulch]

afaik you can't copyright ai art at the moment

[naive hare]

nah, you can

[ionic gulch]

(in most countries)

[naive hare]

t2i, you wouldn't have much standing

[native knot]

Oh shit, Mega Man grew up...

[naive hare]

but if you messed with the workflow enough and had like img2img. yeh, you can copyright it

lke your input has to be 'transformative' enough to basically be a significant change that the AI couldn't come up with on its own

[naive ginkgo]

could anyone explain me the basics or point me to some good article about the basic parameters and what do what? I'm just watching this video about upscaling and don't know, on what basis he setting the denoise, why it's 1 at one place and 0.3 at another place.. or what exactly the KSampler do in the first time.. You know, just some basics, so I will understand the generating more in the technical way so will be able to use it properly 🙂 https://www.youtube.com/watch?v=t3lT_04A9j0&list=PLIF38owJLhR1EGDY4kOnsEnMyolZgza1x&index=5

[naive hare]

pretty sure ive got an upscale tutorial in my list 😉

i do cover the denoising stuff

scott tends to miss certain details, hes the reason I started making tutorials in the first place 😄 dont get me wrong, he does great work tho

[naive ginkgo]

I didn't mean the upscale in particular.. more so I would use some basic stuff.. what params do what, what to focus on etc.

[naive hare]

more inspired me when i saw lots of people not going in depth ya know

so you know what denoising does yah? (no offence, gauging skill level)

its basically saying, how much of the original do I keep? 1 being none and 0.0 being all

so when you do for example, upscaling, like in his video you are working with an image which is larger than it was but usually either distorted or blurry in some way, so you have to judge what degree of this distortion you want to keep and how high your denoise has to be to remove any bad stuff/add details. do it wrong and you end up with extra limbs and stuff, do it right and it should be close to original with more sharpness and minor details

typically in iterative processes you want to slowly drop the denoise as you get closer to the end image

[naive ginkgo]

thanks for this explanation!

I'm new to this AI or SD so I would be glad to find some article about all those basic stuff, params, how it actually works and all..

[naive hare]

https://youtu.be/hdWQhb98M2s?feature=shared >.> not sure about a good written article on it tho

you could ask the comfyui reddit, they are super helpful most of the time

[naive ginkgo]

I’m not familiar with reddit, but will learn 🙂

[ionic gulch]

Neon Genesis Unicorn

[ionic gulch]

unsampler - pig to horse

[ionic gulch]

blur as negative:

[naive ginkgo]

that didn't come out as intended.. Do I understand this images combining wrong or did I proceed it wrongly? I thought there would be the girl and the car from the loaded images..

[pseudo shard]

Is SDXL still slower and less resource efficient than SD 1.5?

Or things have changed and its now as much efficient and quick as 1.5?

[crisp owl]

SDXL is producing a much larger image, so, it will of course require more resources.
But there are turbo models now

[pseudo shard]

Yeah, heard about them, I wonder how those turbo models compare to SD 1.5 efficiency and speed

[crisp owl]

I haven't used 1.5 since SDXL came out, so, I'm biased

[pseudo shard]

Well I'm about to find out

Oh damn thats quick

Like insanely quick, whats that black magic

I can't believe its not running on an external massive server

[crisp owl]

I read up on it when it came out, but tbh forgot much of how they did it. But yeah, it's quick.
It does sacrifice some quality for speed, but not really much depending on your desired output images.

[glad grove]

1.5 still faster than turbo

[rustic garnet]

depends on your hardware and vram. If you have the vram, SDXL turbo is much faster than 1.5

but also normal SDXL is at same 1 megapixel resolution not slower than 1.5. It's just on lower resolutions where 1.5 is faster

[glad grove]

no,sd 1.5 is still faster,on a 4090 u get 35% difference in performance compared with SDXL

[pseudo shard]

I gave SDXL turbo a long try, and even though it is relatively quick, it is so damn messy and not compatible with SDXL regular loras, it seems quite unstable, gonna stick to SD 1.5

[peak dove]

A1111 Pop Art img2img

[uncut steeple]

Jumping the egg hype

[half ivy]

@exotic ruin this is what im inpainting now

going on lyk 2 months it feerls lyk

SD then PS, SD then PS, doesnt end

[exotic ruin]

gawdamn

that is art right there

[half ivy]

bro

not a shill will delete after an acknowledgement or summin but damn, the creativity

[exotic ruin]

damn from a logo?

[half ivy]

brrrroooooo

[exotic ruin]

yes?

[half ivy]

yes

line art control net

do you know control net

??????

[exotic ruin]

yea why

[half ivy]

BEST
THING
EVER!!!!!!

[exotic ruin]

yea haven't been using it for a while

missed it

[naive hare]

ahh you found the way 😛 better yet live photoshop integration + controlnet

[naive ginkgo]

why is there so little variability in the face? All of them looks like one girl.. I used this prompt:
beautiful woman, portrait, age 29, {with|without} glasses, (skin imperfections:1.4), {blonde|pink|light blue|green} hair, blue|green|brown} eyes, masterpiece, high quality, photoquality, realistic, nikon, 4k, 8k
this settings: TurboVisionXLSuperFast checkpoint, 1024x1024, 5 steps, 1,5 CFG, dpm++ sde, karras

[peak dove]

Use Portrait Master node in ComfyUI for really different faces each time

[naive ginkgo]

oh, have this one open here to download, so tomorrow will try that one 🙂

[dark crypt]

TurboDiffusionXL is doing great for my Warhammer 40k sisters of battle.

[rustic shadow]

Be more creative... is she Irish? African? Round face? etc

[cyan crown]

https://civitai.com/models/266125

[vestal peak]

I put your prompt into Fooocus and got these: https://i.imgur.com/wj903Q8.jpg

[cyan crown]

https://civitai.com/models/265129

[dark crypt]

These are fantastic. I'm loving these.

[naive ginkgo]

I know I can be more specific in the prompt.. I just thought it will give many different faces when I intentionally give it quite open prompt..

[naive ginkgo]

nice! quite detailed! what settings did you use?

[meager canopy]

If you don't specify differences in the face, SD will choose one and keep using it.

[naive ginkgo]

oh I see.. probably my misunderstanding on how it works.. thought the differencies will be there from randomized seed..

[meager canopy]

Using:
Lora https://civitai.com/models/258106/obsidianfirered
Model: https://civitai.com/models/240590/ultimateblendxl

[peak dove]

ComfyUI Portrait Master 2.3

Icelandic Woman

Cuban Woman

My Upload to Printler today

[naive ginkgo]

Magnificent Millie is great 🙂

[dark crypt]

I did a ReActor face swap on this: black hair girl in Warhammer 40k, carrying a sword with electricity dancing along the edge, black robes, psychic blank, gold skull on her chest, 8K, highly detailed, intricate, ultra-realistic, award-winning, professional, highly detailed, (Raw Photo:1.3) of (Realistic:1.3) (using the Golden Mean:1.3), highly detailed, elegant, smooth, sharp focus, beautiful face
Negative prompt: cartoon, drawing, painting, bikini, swimsuit, anime, ugly, deformed, noisy, blurry, distorted, grainy, nipples, nipple light, headset, headpiece, helmet, long neck, asymmetrical eyes,
signature, watermark, baddream, unrealisticdream, closeup,
Steps: 8, Sampler: DPM++ SDE Karras, CFG scale: 3, Seed: 15, Size: 680x1024, Model hash: 1524a9699f, Model: turboDiffusionXL_v112, Version: 1.7.0

[naive ginkgo]

thanks for the details

[uncut steeple]

Really cool work

[cyan crown]

thanks

[icy brook]

I'm pretty sure I saw a while back a LoRA or something that helped remove the always bokeh/blurred backgrounds of portraits. Anyone know what I am talking about?

[ionic gulch]

@icy brook using blur as negative helps

[icy brook]

yes, but very slightly 🙂

especially on closeups

sdxl is overfitted on bokeh

[kind pollen]

has anyone had/seen ongoing issues with SDXL returning blurred out images? Not like a little blurry, but as if a strong gaussian blur has been applied to the entire image.

[wet nacelle]

Post this here as well. https://discord.com/channels/1002292111942635562/1002602742667280404

[wet nacelle]

kitty cats

[naive ginkgo]

I'm still having troubles with full body gerenation.. I'm using the Portrait master node, sed Shot type to full body, set shot weight to max (1.95) and still gotting this result.. tried different checkpoints (turbovisionxl, sdxl base, juggernaut..)

[proper cypress]

anyone else getting long necks while using juggernaught v8? or am I doing something wrong?

[rustic garnet]

change aspect ratio

you use 1024x1024. A full body shot in this aspect ratio (without describing anything in the background) would mean most of the image is empty

[proper cypress]

by juggernaught I say what thing they wear on their legs (skirt, pants etc.) that works for me, is I use 768*1355 as aspect I stil get only the face

[rustic garnet]

change it to portrait format (like 768x1344) and you should get full body shots

[rustic garnet]

oh yes, that also helps a lot! If you just mention their shoes they will often automatically get full body shots

[proper cypress]

while still using (full body)

[rustic garnet]

on the other hand, if your prompt mentions many facial features (natural skin, skin imperfections, freckles and so on) then you will more likely get portrait photos (as you wouldn't see these details on a full body shot anyways)

you can add these details afterwards using Face Detailer

[proper cypress]

btw does anyone know if a 3080 10gb can handle sdxl training

this is what I mean with long neck, I did put long neck in negative and short neck in positiefe

[meager canopy]

This is a problem with using this node. It adds a lot of facial details to the prompt, so the emphasis in the image is towards the head. I've managed to get some images with more of the body showing, but that node makes it VERY difficult. I gave up using it for this reason.

[rustic garnet]

yeah, to be honest: You don't need the node. Just look into the prompts it generates to get the right trigger words

e.g., I never used "natural skin, skin imperfections" and was surprised how well it works

but the node is only concatenating prompts. You can just do that yourself

[peak dove]

I'm getting full body images, different hair color and styles, photographic styles etc by modifying the LCM-LIST files. Add to pose list jumping, walking, running, lotus, sideways etc etc

[naive ginkgo]

@rustic garnet @proper cypress @peak dove thanks guys for your insights 👍🏻 so basically is it better to create a face in one generation then a body in a second one and then switch the face on the second one with the first one? or what is a best way to do that?

[rustic garnet]

no, generate the complete body first

then you do "upscaled->inpainting->downscaling" on the face

this is sometimes called "detailing". There are also special workflows or plugins for that which do this automatically on faces

what tool are you using? ComfyUI? Auto111? InvokeAI?

[meager canopy]

Most of it can be done in the prompt, but the face detailer just adds polish to the finish. All of these are just from a prompt, before face detailing...

Using https://civitai.com/models/240590/ultimateblendxl

[naive ginkgo]

I use ComfyUI.

[naive ginkgo]

that looks great! it's really one step with one prompt?

[naive ginkgo]

why upscale and then downscale? can you hint me some advice on the workflow or plugin?

[rustic garnet]

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

[rustic garnet]

SD in general has problems to get small details on low resolution right. This is mostly due to the VAE (internally, SD compresses the images to 1/8 of their original resolution), but SDXL is particularly bad in this regard (much worse than SD 1.5)

So if you have e.g. a resolution of 768x1344 and a full body shot, then the face in the image is maybe only 160x160 pixel in size. Internally, due to the VAE compression, the face would be then only 20x20 latent pixel in size. Now the diffusion has problems to get facial details into such a small image

the idea is that you cut out the face, upscale it to higher resolution (e.g. 1024x1024) and then do img2img on the face. Afterwards, you shrink down the face to the previous resolution. Of course due to the downscaling you will loose some of the details again, but overall it will be still much more detailed than before

of course you could instead do upscaling on the whole image (e.g. with SD ultimate upscaling). That is also possible and would have the same effect (more details into all the smaller parts of the image).

So if you want to do higher resolution anyways you don't need Detailing. The Detailing is only necessary if you don't want to go to high resolution but still want detailed faces

[naive ginkgo]

wow thank you for this explanation, now I understand the technique behind it, much appreciated! 🙂

[naive ginkgo]

this is the plugin to do all the steps for the full body - face - upscale - face swap - downscale? 🙂

[rustic garnet]

unfortunately that is a plugin that does one million things 😩

it has, next to many other features, this workflow for face detailing

I don't know if there is another plugin that is more focused and easier to use

[naive ginkgo]

ok thanks, will check it 🙂 is there some "standardized" name of the whole process? or what should I look for when searching it? 🙂

[rustic garnet]

I can make you a minimal example

[meager canopy]

Yes, 1 prompt and the "ksampler cycle" node, which upscales x2.

[meager canopy]

Some ComfyUI workflows here https://civitai.com/images/4654927

[naive ginkgo]

would be great!

[naive ginkgo]

this is it? https://civitai.com/models/119257/gtm-comfyui-workflows-including-sdxl-and-sd15 ? so I just clicl the version I want above the pictures and download it?

[meager canopy]

Yes

[naive ginkgo]

thank you! 🙂 about the versions - sdxl v5.1, xlite or cleanUI v1.2?

[meager canopy]

The latest ones are to the left, older ones to the right. Give the cleanUI v1.2 a try, but it may require some additional custom nodes. If you have the manager installed, that should auto install missing ones.

[naive ginkgo]

ok will start with the cleanUI v.1.2. I have the manager so if needed will download missing nodes

[rustic garnet]

this workflow takes an image as input and refines it

[naive ginkgo]

great, will try it! thank you very much

[naive ginkgo]

could you tell me what are the models in SAMLoader and UltraLyticsDetector?

[rustic garnet]

I hope/thought they would download automatically whe installing ImpactPack

SAM is "segment anything", a model that basically classifies each pixel in your image. it allows you to select e.g. all pixels in the image that belong to a face

UltraLyticsDetector detects the faces in the image, however, it does not detect pixels but just the bounding box where the face lies. Both detectors are somewhat combined (although I'm not entirely sure why you need both)

[south horizon]

Fun--will you share the prompt for this?

I'll scope out that ultimateblend xl model.

[honest marsh]

recently I moved to A1111 and i dont know why is it so slow

in comfyUI i get 2-3 it/s

i have a rtx2060

[ocean ermine]

Among other reasons, this is why I had to move to Comfy for SDXL. If you're getting that slow of speeds, it's likely that the VRAM it needs it more than your GPU has and thus it's using system ram instead--which will obviously be much slower.

[honest marsh]

i previously had comfyui so it doesnt bother me

[ocean ermine]

It bothers me and I've been using ComfyUI for months.

[severe grail]

https://tenor.com/view/this-is-fine-gif-24177057

[meager canopy]

photograph of a female ice hockey player with ponytail hair and (muscular thighs), wearing a short cropped sheer hockey jersey, thigh high socks, and a pleated skirt, smiling cheerily while looking at viewer, posing for a picture, in an ice hockey rink, (Instagram:1.3)

[naive ginkgo]

nope, at least not for me 😦 I tried to reinstall it now and still no models 😦 I found the SAM model online and download it, but can't find the models to Ultralytics.. even the folders in models - ultralytics - segm and bbox are empty 😦

[naive ginkgo]

finally I found some models here https://huggingface.co/jags/yolov8_model_segmentation-set

[cyan crown]

A little news from the lunch break at work: A new distinctive style.: https://civitai.com/models/268599

[naive ginkgo]

first inpainting attempt 🙂

another try.. is there a way to fix the details around the jeans shorts?

[lucid rain]

try StableSwarmUI. Nice front end but back end is comfy.

[hoary saddle]

does anyone know the ramifcations of sdxl turbo as far it running in auto queue? wondering if it's harmful to the gpu more than AI is alreaady, does it continue to heat up when auto is on but no generation is being made?

[lucid rain]

I have not specifically tested that but I have not seen any extra use of the GPU unless images are being generated.

I'm sure just having comfy on and the model loaded takes some sort of power.

[hoary saddle]

was my thought, just curious if anyone was seeing something different