✨｜sdxl

[spring fulcrum]

Should i put the t5 first?

[shy kelp]

its a very bad design choice

in Comfy the yellow noodles called Clip can carry multiple models

its confusing

[spring fulcrum]

yes it is. and I am lol

Ok I think I've added the right things and integrated this right.

now to try it out

[spring fulcrum]

would it be a bad idea to use the same clip twice?

[shy kelp]

will likely throw an error

for best model use you want both, generally

[spring fulcrum]

Does it matter which one is first?

[spring fulcrum]

What do you recommend for shift? these look so bad since i added the model sampling flux node. 😛

[shy kelp]

I don't use shift I use custom sigmas

its not really possible to recommend a number cos it depends on so many factors

[spring fulcrum]

What in the flux is custom sigmas. I swear every time I get on this discord you teach my like 100 new things.

[shy kelp]

so in Comfy the schedule is a list of floats

float is a number like this 2.3, 6.5 etc

each individual number is a sigma

for every step of the model, the model looks up the sigma for that step

and it takes the sigma into account when it does the denoising process for that step

[spring fulcrum]

are you willing to put an image up with the embedded workflow so I can see how that would work then I can Frankenstein that into my workflow 😛

[shy kelp]

sure this is my base workflow these days

[spring fulcrum]

They say we have only explored 5% of the ocean.... It may take me a decade or so to explore all of the nodes in your workflow... 😛 Currenly at 5%

[shy kelp]

haha

[spring fulcrum]

explores more.... thinks (i understand some of these words)

so you connect certain nodes for certain use cases?

[shy kelp]

no its just that my workflows aren't really made for sharing they are personal
so they are always a mess and half made
I don't mind sharing them but they are not organised

there's like 20 sections that have been remade and deleted this is just the nodes that survived until the end LOL

[spring fulcrum]

lol

Is there a particular set of nodes that you are currently using more than others. if so can you take a screenshot of that section?

dont worry i will use a map and goe locate where it is on the map you provided lol

[shy kelp]

I do completely different stuff every time really

there's not that many good nodepacks you will have tried them all by 6 months time

[spring fulcrum]

what i find amazing is that you know how this works so well. In my mind it is like when i take an engine apart and lay out all the parts on the garage floor then rebuilt the engine.

[shy kelp]

I mean there's loads of tiny ones but I mean the big, structured packs
its not a massive number to go through

its kinda similar to engines I guess yeah

[spring fulcrum]

Is there a node that can take an uploaded image and tell me the size? Width, Height in pixels. I just want it to be in the middle of the process, and then I can put another of the same node and have it display the size on the new image?

Something like this. Tells me size of original image... another to tell me size of created image?

[shy kelp]

KJ nodes has a get image size node

[spring fulcrum]

That worked. I needed these too it seems.

[spark onyx]

hi guys do anybody know how to create several arts with same scene(Background)?
img2img isnt very effective
photoshop is the only way?

[meager canopy]

1. Create a background image.
2. Create a subject image and remove background.
3. Blend the subject and background images.

You can create a ComfyUI workflow to do this.

[spring fulcrum]

Do you have a link to any good resources on how to do that? I was wondering that myself. I already have a great workflow for removing background. But getting comfyui to create just a backgroud with either low details or no people can sometimes be difficult. plus it has a hard time getting scale consistent with subject not in the image.

[rocky sierra]

Hello! I am using "Structure." The generated images have darker lighting compared to the uploaded images. How can I adjust the brightness to make them lighter? I would appreciate your advice.

[raw idol]

Hello!

[turbid orbit]

What are some current good controlnet models to use with pony and illustrious based models?

[humble cape]

pony and illustrious have controlnets ?

[turbid orbit]

I have been having some luck with cn-anytest, but can behave a bit strangely for certain preprocessors. Was wondering if there was anything newer

[spring fulcrum]

Does anyone have any suggestions for a great model to run with open-webui through a local ollama install. It would be running on an RTX 4090?

any yes I am aware this is not the ollama or open-webui discord server. I also run flux with comfyui. A lot of folks in ai art run in the same circles.

Beuller?... Beuller?

(hears crickets)

Does anyone have any suggestions for any good diffusion models? I currently run flux dev gguf q_4 k m. along side an ollama model for vision that describes an image and creates the prompt to generate the image from.

[normal cliff]

Testing sdxl depth ControlNET model.

[potent dove]

Does anyone have any suggestion of how those SDXL-lighting models from community are trained, since SDXL-lighting is not open source yet?

[steady wren]

/3D desk#

[rough gazelle]

I'm trying to make a lora from SDXL Lightning, but i'm just getting mutants

[rough gazelle]

from what i read, the easiest way is to train a lora and merge that with the base lightning checkpoint

[lofty helm]

i've never been impressed with lightning models. they're heavily distilled and have less prompt comprehension on top of sdxl's lackluster prompt comprehension

if you're just a single person working with images, how fast do you need it to be? 30 seconds vs 10 seconds? not saving you too much in the creative process. If you're a bulk generation service that's serving up 1000s of images a minute, it's better for those purposes

[shy kelp]

they are merges

[warped sedge]

video upscaling, animatediff, things like this lightning was more useful to me

i looked i have 2 lightning checkpoints that i haven't used since February 😆

[shy kelp]

would recommend TCD/TDD/PCM/DMD2 loras
they are newer

[spring falcon]

@pseudo mortar AI大脑，体现科技感，用于系统宣传图

[steady wren]

@primal lichen There is a little boy in the study, a yellow person studying

@pseudo mortar There is a little boy in the study, a yellow person studying

[potent dove]

Have you tried (lighting lora + community base model) -> finetune? Compare to training a lora on the base lighting model, which one could be better?

[potent dove]

i am currently training DMD, but my current result is not as good as community lighting model

[shy kelp]

the original DMD or the newer DMD2?

there are different pros and cons to all these methods

[potent dove]

DMD2

[rough gazelle]

I have no idea, i'm noob at this, sorry.

[toxic wasp]

Oh! I LIKE these!

[twilit lagoon]

if you are using juggernaut XL 11 can you tell me if it's better to use it with clip skip 2 because on civitai uses that by default

[chrome atlas]

Very very beautiful pictures and dress , very nice subject.💖💖💖💖💖💖💖💖💖

[fading pike]

does anyone have technical experience using the sdxl pipelines available from huggingface?

ive been trying to compute gradients withing a callback on step end function but the problem is that the base pipeline has a no grad wrapper that seemingly makes the computation of any gradients whatsoever impossible

ive tried working around this with a custom pipeline that removes the no grad wrapper but at 1024 by 1024 images the memory requirements are too much for an a100 gpu

(ive also tried generating images at 512 by 512 and also 256 by 256 but these images seem to be malformed, even with the unmodified sdxl pipeline)

I have a feeling this may not be the best place to ask but if anyone knows where I could find some guidance with this I would really appreciate it!

[lofty helm]

clipskip is only something used by anime models like pony xl. it was more important during the sd15 days, but astralite decided to use it on pony. i dont know why.

Juggernaut doesn't need it. The civit generator uses it by default because 90% of image gen tehre are pony xl for "reasons". And the civitai team doesn't want to spend time coding a dynamic setting that changes based on the model

[sullen escarp]

An enchanting night sky illuminated by a massive drone show featuring vibrant, detailed holiday-themed displays. Start with a glowing Thanksgiving turkey surrounded by swirling autumn leaves, transitioning into a magical winter wonderland scene with snowflakes, pine trees, and shimmering ice sculptures. Highlight a bustling gingerbread village with intricate details—houses covered in icing, candy canes, gumdrops, and cheerful gingerbread characters. Make the drone formations vibrant, dynamic, and perfectly aligned in a clear, starry sky. Add festive lights and a warm holiday atmosphere throughout."

[trim spruce]

i have a question
what prompt to use to avoid having a character who looks too young, I'm looking for a prompt so that he seems to be in his thirties

[faint shore]

Instead of describing someone as male or female, try describing them as father, mother, husband, wife. This can make them older. Also, professional titles can help as well. If you use mother or father you may need to add child/children to the negative to prevent kids from appearing.
Sometimes, I just specify it.
a 30 years old professional woman teaching a class.
On another note, some models can't do aged people well. As you have discovered, they can only make 20 somethings. Try another model, a well rated one with good prompt adherence.

[glacial star]

One message removed from a suspended account.

[next quest]

Guys, quick question: which schedule type should I select when using Euler A?

[faint shore]

@next quest Try ddim_uniform.

[proven mason]

What Setup should i use with the instantid Id in sd Forge webui? I have instantidid and ipadapter plus 2 and all of them but i dont know what combination i should use to get only the face to be identical to my generated caracther

[glacial star]

One message removed from a suspended account.

[deep iris]

hello frens! do the newer models require SDXL? is Pony XL different, or simply the model they use? How to update my Auto1111 install so it works with latest models?

[clever cave]

Sdxl is a model itself by stability ai. PonyXL is a finetune of sdxl to make it know more knowledge, styles, and characters.

Not sure what new models mean, but you might mean flux/sd3.5.
Flux is a model trained by Black Forest labs and it is much better in prompt following, humans, and can do text as well.
Sd3.5 large is a bit worse in the above but knows more styles and knowledge.

Auto1111 is kinda outdated, I would recommend something like forge or comfyui now which does support these new models.

[glacial star]

One message removed from a suspended account.

[forest pasture]

how can i use sdxl to creat pictures?

[neon temple]

for real?
you install a checkpoint of sdxl and use it with a webui like a1111, forge, Foocus, comfyUI or so

CivitAI is vvvveerrry eassy to use for the beginning or Leonardo AI

[tribal dome]

does anybody know of good ultra realistic models

[gray lagoon]

Depends on your specs or are you gonna run it thru the cloud?

[robust pawn]

https://civitai.com/models/133005?modelVersionId=782002

[keen phoenix]

is it possible to combine 3 faces into one synthetic face?

has anyone heard of this being done

[normal monolith]

https://tenor.com/view/waving-the-monster-frankensteins-monster-bride-of-frankenstein-hello-gif-18796531

[gray lagoon]

what UI do you use? Swarm has a feature called alternate

[keen phoenix]

i just automatic111

[gray lagoon]

Ab then im afraid i dont know, i switched to swarm a while back and its been a great upgrade for me

I do miss some extentions though but nothing too important

[keen phoenix]

with alternate what does it require to create the new face
l10 images of each face and interpolated together?

[gray lagoon]

Nope, alternate is just tool in a prompt so it switches every step so in the example image

It alternates between cat and dog while generating

So if your model knows ben affleck and ryan gosling™️ it would mix them

[keen phoenix]

oh i see

can you adjust the weights of their face?

[gray lagoon]

Yes, normally you use alternate:cat,dog,horse but you can either repeat the cat or use <alternate🐱0.4||dog:0.2||horse0.4||>

Discord formatting being a dingus

But its very well documented in the github and in the UI

But if your used a1111 its some getting used to

[keen phoenix]

ok interesting

it’s to create a lora for me personally

[gray lagoon]

Hmm i use civitai for that but once i run my pc less actively im training them locally

[heavy coral]

Hey is anyone good with control net?

Im still learning it, i thought i followed the tutorial correctly but no matter the settings i do the control net pose for open pose doesnt have the same pose when i generate a render

[gray lagoon]

it wont end up exacty the same but what model are you using?

it has problems with illustrious atm, needs a different controlnet

[faint shore]

@heavy coral Also check your seeds. Make sure they are all fixed while you debug the controlnet.

[heavy coral]

How would i know? Im still new, ive been sticking with sd 1.5 and sdxl checkpoints

[heavy coral]

Maybe that its the cause, i dont think i touched the seeds. But the art style im using is an anime art style so not sure if that is the cause

[gray lagoon]

well are you using Pony, illustrious or some other, the checkpoint name could be important

and iirc dont quote me on this SDXL used a different controlnet model then 1.5

[gray lagoon]

talking about the SDXL model name*

[heavy coral]

I just checked the checkpoint it shows SDXL 1.0

Let me see if the seed would do anything, i even tried a depth map with a image i made but even that was off

[gray lagoon]

ah your using SDXL base, hmm could be a sdxl controlnet issue

[heavy coral]

Ideally my plan with SD is to make my own web comic but i want to be able to control the backgrounds and also control the characters in the shot with control net.

Im not sure if thats possible since im still learning SD currently

[gray lagoon]

hmm i would say its ambitious but not impossible. i know a few that make their own games with it

rpgmaker but still

you are using automatic1111?

[heavy coral]

Any idea on how to set up SD for this workflow?

Yeah, i tried comfy ui but its a bit too complicated for me currently

I was told to use Automatic1111 first until im comfortable with it

[gray lagoon]

yeah thats solid advice, once you know A1111 and want to move on to something slightly more complicated theres SwarmUI

but since you use A1111 ill Dm you a link to stable diffusion art tutorial for controlnet and SDXL

the last time i linked something i got a repremanded lol

[heavy coral]

No worries, ill send you a friend request since i dont think i can get messages from non friends lol

I made some great art in Leonardo but trying to get the same art quality in SD is hard for me atm

[gray lagoon]

hmm i generally do themes not suitable for this discord but i can show you some prompting examples and some sources that i use

[craggy finch]

Hi everyone! I'm looking for models and/or loras to do realistic inpainting with SDXL and Flux in Forge. Any help?

[gray lagoon]

hii, have you checked civitAI for inpainting models?

[queen elbow]

soon

[elfin burrow]

stoke

[craggy finch]

yeah, but I only found epicrealism and is for SD 😦

[gray lagoon]

Try looking for juggernautXL or "impainting" in the searchbar in civitai and use the filter to SDXL kn the left

Though i dont do realstic impaintig that much

[craggy finch]

I think juggernautXL does not do inpainting as far I know

[gray lagoon]

Theres a impainting variant

[craggy finch]

oh, I've just found it

[gray lagoon]

And a epicrealism impainting var

[craggy finch]

thanks!!

[gray lagoon]

Np!

[undone tide]

Anyone know how to not get portraits everytime? Doesnt seem to matter what i enter as prompt :/

[gray lagoon]

whats the prompt?

or do you use a specific lora?

[undone tide]

This is one

artiangel, (male, green eyes, brown short hair, angelic, gothic, glowing, translucent, bioluminescent, imposing, ethereal, white historical armor:1.5, gold details:0.5), poster art, bold lines, angel wings, hyper detailed, expressive, award winning, (professional, finest details, masterpiece, best quality:1.5), looking at viewer, dynamic pose, full body view,wide angle view <lora:artiangel:0.5> <lora:RPGAngelsXL:0.2>

Also tried similar to "male standing on a hill"

Trying to create a male angel. with some Loras

[gray lagoon]

hmmm do you have the same problem without the loras?

[undone tide]

i will try with out it, and get back to you :p

[gray lagoon]

tried running the same prompt (different model) though no portraits for me

[undone tide]

Hmm could the GPU matter in what the prompt gives you? and not only the time it takes?

[gray lagoon]

shouldent really\

the SEED determines output. unless theres some wacky vram things going on

[undone tide]

What should the seed be set to? it was set to -1 automaticly when i set up webforge

[gray lagoon]

seed being -1 is the intented

because thats random

but lets say you got a image your happy with but want to change one thing lets say eye pupil color or a shirt colour etc

you can reuse the seed and possibly not change the whole image too much

i can DM you some examples

[undone tide]

Aa alright! 🙂

[glacial star]

One message removed from a suspended account.

One message removed from a suspended account.

[gray lagoon]

theres a guard uniform lora for squidgames but not a contestant

[glacial star]

One message removed from a suspended account.

[gray lagoon]

yeah?

the tracksuits with numbers?

[glacial star]

One message removed from a suspended account.

[glacial star]

One message removed from a suspended account.

[gray lagoon]

i mean it looks pretty solid

[glacial star]

One message removed from a suspended account.

One message removed from a suspended account.

[inner quarry]

Anyone know if it's possible to run OneTrainer SDXL Lora with 8gb 3070ti? I searched Google but none of the solutions worked for me. Always out of memory error

[glacial star]

One message removed from a suspended account.

[drifting girder]

nou

[gray lagoon]

This isn't like midjourney

[tacit sparrow]

Anyone have success posing someone doing choreoraphed dance moves in Pony? I've even tried a Lora, which didn't seem to really work, and I'm running out of prompts to try. Something like this: https://cdn.canvasrebel.com/wp-content/uploads/2023/07/c-PersonalRhapsodyTaylor__IMG32002_1688167242343.jpg

[gray lagoon]

Open pose

[drifting girder]

use controlnet openpose

[tacit sparrow]

I haven't gotten to that level yet, haha. Plan to soon. I'm working between CivitAI, and InvokeAI.

[drifting girder]

look into learning comfyui

it wil be intimidating at first but once u get the hang of it

easily become my main ui

[tacit sparrow]

I have it installed, tried it. I keep flopping aropund

[drifting girder]

best way to learn is by using other ppls workflows and seeing how things are connected

[vagrant flower]

How do I get better looking outputs?

This is what I have btw

[gray lagoon]

I have to say thats an INSANE negative prompt

Since you have easy negative already you can trim it a lot

[south horizon]

that also appears to be a pony model, and your prompt structure is not correct for pony models

[drifting girder]

way way way way way to much negative lol

thats sd 1.5 over prompting days

[gray lagoon]

Lolled @ the facebook in negative

Id remove anything after disfigured and replace the embeddings with, blurry, low quality bad quality, monochrome

[idle solstice]

i dont think im using the pixel art xl lora correctly

[torpid pecan]

Hey guys am a bit new to Stable Diffusion so I need some help. I am looking for a model to generate realistic human poses while also being somewhat fast at image generation. Can someone recommend me a model ? I have RTX 3050 and 16 GB RAM

also i want to controlnet with openpose so a model that can fit in all these categories 🙂

[neon temple]

https://civitai.com/models/277058/epicrealism-xl

https://civitai.com/models/25694/epicrealism ( but i dont like to use open pose with that)

[solar trellis]

hi question. If you are training a character Lora in SDXL, is it advisable to train it on the base model or a custom checkpoint that you want to generate the Lora on?

[torpid pecan]

thanks man

[rocky mesa]

Any suggestions of how to train a SDXL lora for jersey patches (think soccer uniform logos/patches) that are stylized? I was able to train a lora with a full uniform and that worked OK but the final image was a bit too photoreal (presumably because the training images were photos and not stylized). I guess I'm asking how do you pre-stylize images for lora training? I tried using the stylized model (DynavisionXL) but even at low denoise it mangles the logos/patches. Feels like a chicken/egg problem.

[gray lagoon]

Do you only want patches or a image where theres patches on clothing

Because if you just want patches youd train on images of patches only with nothing else

[rocky mesa]

Ultimately the goal would be logos/patches on clothing. I think I can make that happen but how do I get the logos/patches to be stylized before training?

[gray lagoon]

How do you label the patches

Do you just do shirt with patch or do you do "Barcelona patch" "brazil patch" "uk patch" etc

Still the smaller the patch on the shirt the more mangled its gonna be

[lofty helm]

Playing with SDXL today and i noticed i still have the SDXL refiner saved in the folder. Has there been any use of that since SDXL came out? lol i don't think i've ever found a workflow that implements it.

[rocky mesa]

I haven't tried a lora with logos/patches yet. I've tried full jerseys. It worked ok but it skewed the final results to be too photorealistic.

[shy kelp]

some mad lad on Civit used SDXL refiner on Flux and it worked really well LOL

[lofty helm]

it wouldn't use the same latent space though right? he would've used the vae on it then passed it to flux that way

[shy kelp]

yeah they used vae decode/encode
it was flux first and then SDXL refiner after

[gray lagoon]

Some people also gen with flux and impaint with sdxl

[lofty helm]

makes sense, but then it ends up with the 8 channel vae. big downgrade imo

[shy kelp]

yeah I agree that is a downside

its tricky

I do SD 1.5 or SDXL tiled upscale on Flux images a lot and having to use the lower quality VAE is rough

[timber tree]

@lean kelp

[worthy orbit]

yes, after working with ai images, at the end of the day and if you use each tool properly, it can work well in the final image.

I mean, some models have more "quality", but if you have a good eye for it, you can inpaint with a "lesser" model, just watching it keeps the quality... But in reality, sometimes it is not about the quality in some area but the quality of the whole img, what separated a higher model... so inpainting in SD XL works fine

I would love to do it with SD35 or Flux, but it takes so much time that I guess it must be a nightmare. But at the same time, maybe for some great changes, it may be better.

for example you can generate in one model, img2img with other model, then inpaint, or do some photoshop in the middle, the upscale

I don't like automatic workflows that do all that, I like doing one process myself. Also those workflows take a lot of time, and tiresome set up, and sometimes I don't want to work with an image that much.

I don't understand how people use those workflows that inpaint the face, details, upscale, for an image that maybe isn't good...yes it is automatic and you can leave it working and will have some finished good pics... but well

[gray lagoon]

you first generate a bunch without. if only the face is off then you do the rest

[glacial star]

One message removed from a suspended account.

[analog slate]

https://huggingface.co/cagliostrolab/animagine-xl-4.0 cagliostrolab came back with a new finetune for SDXL, exciting

[echo crane]

Have any of you managed to get SDXL controlnet inpainting work properly on non-comfy UIs?
It seems that no matter what I do (multiple controlnets, multiple checkpoints, different areas and scales), it either doesn't work at all (forge / reforge), the result is total garbage (A1111) or very bad quality (Fooocus).

[gray lagoon]

Are you using the proper control net model? Theres one for every main model you know

Theres a seperate one for illustrious & SDXL & Pony etc

[echo crane]

I tried all three SDXL inpainting controlnets I could find (contolnet union, Ecom, destitech). None worked at all. SD 1.5 controlnet inpainting works fine (in A1111, it appears to be broken in reforge).

[gray lagoon]

?? What model are you using to generate?

[echo crane]

I tried thinkdiffusion, juggernautxl and realvisxl). All resulted in sinilar garbage results (ie. Something was clearly broken, not just less than stellar quality),

[gray lagoon]

Okay lets first settle on a single model lets say juggernautXL

juggernautXL has its own impainting model so thats something

Since its a XL model it needs a controlnet made for SDXL

Where are you getting these?

What are your settings, do you get any errors in the logs

[echo crane]

JuggernautXL is from civitai, controlnet union from official xinsir huggingface repo.

Controlnet union (promax version) works just fine for other types of controlnet (depth, scribbles, lineart etc). The page claims it has inpainting & outpainting support (with example images).

[gray lagoon]

Stupid question but why arent you using the image to image impainting feature?

[echo crane]

I am (that's afterall necessary to do any inpainting in A1111 in the first place without dealing with multiple manual masks in A1111 img2img tab). If I do the exact same things in SD 1.5 (just swapping checkpoint, controlnet and resolution), things work perfectly.

[gray lagoon]

I could've sworn inpainting in a1111 didnt require controlnet unless you want automatic masking

[echo crane]

Controlnet is needed when there is no good quality inpainting model (which I haven't found for anything other than SD 1.5 checkpoints and even there it can be iffy depending on checkpoint)

[gray lagoon]

Hmm feels like it's more of a UI problem tbh, ever considered switching to swarm? Where this is absolutely not a problem

But jf you have error logs etc i recommend hopping over to #🤝｜tech-support

[echo crane]

I might try SD.Next. It seems that controlnet union support just landed there a few weeks ago.
I tried an older version of swarm (that was on thinkdiffusion back then) and got fed up with the UI within 15 minutes.

[shy kelp]

its hard if you are on A1111 cos very few people still use it so its hard to get help

the two big communities are Comfyui/Swarm and Diffusers

Swarm contains Comfyui, its a bit confusing

essentially Comfyui and Swarm are the same

[steel rain]

I use forge

is that stable diffusion?

[echo crane]

Forge is broken currently with a lot of functionality. Otherwise I'd go with it.

[steel rain]

what's broken with it? Im new and installed it yesterday. Had no issues creating, even tho it took a while for each image

like 15-20 seconds

[echo crane]

Controlnet fixes have been pushed back " a few weeks" since last July

And it of course broke a lot of stuff when he transitioned to Gradio 4

[shy kelp]

Forge is based on A1111

[echo crane]

For certain values of "based" now that he's revamped it so much

[shy kelp]

yeah I don't want to offend people by describing it badly
IDK exactly what the current Forge backend is like

[gray lagoon]

Web-ui forged is still pretty good for beginners who dont do a whole lot with it yet and isnt as intimidating as comfyUI

[shy kelp]

I started in A1111 too, I don't think it was a bad place to start

[echo crane]

Alas, I need that extra functionality (working controlnets) and I'm completely allergic to Comfy's idea of user interface

[gray lagoon]

Use swarmUI

It has both

Optional comfy but a functional UI for controlnets etc

And lots of updates etc

[echo crane]

Can SwarmUI be significantly revamped from this layout?
https://raw.githubusercontent.com/mcmonkeyprojects/SwarmUI/refs/heads/master/.github/images/swarmui.jpg
Because if it can't, it's an immediate no go for me

[gray lagoon]

Why? I mean the theme can be changed yes

And it scales with your window size

Im on a ultrawide and the side bars go nicely to the sides

But if you want to move tabs or hide them probably not. (Advanced options give you 3x the amount of configuration)

[shy kelp]

I didn't like the Swarm GUI either

GUI is very personal thing

you might like Invoke

otherwise SDNext is made by a good dev

[echo crane]

Because it wastes 75% of the screen on things that are irrelevsnt to me 90% of the time

[gray lagoon]

You can resize iirc.

[gray lagoon]

I prefer swarm made by a ex stability dev. ||And a pretty cool guy in general||

[steel rain]

Im a beginner

[shy kelp]

I did use Swarm for a bit

[echo crane]

I'd kill (read: pay actual money) for a good and actually working Photoshop plugin that supported local generation. Alas, as far as I can tell no such plugin exists.

[gray lagoon]

You could hire some developers to make that project a reality if they think its feasable

Ive seen ads for like 1200 for 6 weeks of dev time but they were college students iirc

[shy kelp]

https://github.com/NimaNzrii/comfyui-photoshop?

maybe

[fast anchor]

Can i create an image now

[gray lagoon]

You always could? Or do you mean in this discord?

[fast anchor]

In this discord using SDXL ai

[gray lagoon]

If you had an artisan subscription but otherwise its a community discord

[fast anchor]

There should be a chatbot here creating images

[gray lagoon]

We mostly generate locally in here. Otherwise i recommend using a free service like civit ai (5 images a day ish iirc)

[gray lagoon]

There is, #artisan-faq

[fast anchor]

So can some of yall create images for me?

[gray lagoon]

If someone wants to you can ask but i cant rn

[fast anchor]

Allright thank u

[gray lagoon]

Also nsfw requests probably wont be appreciated in generalchat lmao

[fast anchor]

Nah i justed wanted to try out SDXL 1.0

Appreantly im in thr wrong place

[gray lagoon]

I mean you can use civit ai. If you make a account on there you get some free "buzz" where you can try different stable diffusion models on their website

This is probably the best place to ask for advice but people aren't keen to make stuff for others all the time

[fast anchor]

Allright .. advice is enough

Lets see what civit Ai is

[gray lagoon]

Civit ai green is the SFW only website

But if you got a strong pc you could also run it on your computer

[fast anchor]

Im try to get janus 7b to to run on my phone

😆

[gray lagoon]

Goodluck with tha5

[fast anchor]

Mirror the pc to phone its no problem

Anyway thanks for the advice

[echo crane]

In case anyone scrolls up back to the SDXL inpainting discussion, turns out that Controlnet Union inpaint requires different mask values than SD 1.5 inpainting controlnets and A1111 Controlnet extension doesn't do that (and hasn't had a single commit after merging the initial Controlnet Union Promax support).

[neon temple]

yes, and somtimes the modle you use plays a big role too

[echo crane]

Well, now I have SwarmUI running and I'm trying an inpaint with xinsir Controlnet Union Promax. The result has no change unless I lower inpainting controlnet influence to super low strength (eg. 0.25 or something) in which case the prompted result appears but there's no consistency at all. The comfyui workflow doesn't have any obvious problems.

[echo crane]

In case anyone needs it, I found a way to get xinsir Controlnet Union Promax inpainting work in A1111 / Reforge. You need to use Inpaint upload tab and upload the image and inpaint mask separately. Then upload an inverted copy of the same mask to the controlnet tab as independent control image. Make sure to use soft inpainting.

[shy kelp]

you might like brushnet in comfy

https://github.com/nullquant/ComfyUI-BrushNet

brushnet is like controlnet but a bit better designed for inpainting

[wise viper]

can anyone tell how do I even use this server to generate images?

[gray lagoon]

Hi, check #artisan-faq

But most people do it locally on their own pc

[autumn frigate]

"/dream"

Парень в зелёной рубашкн

"/dream" парень в зелёной рубашке

[tough cliff]

Looks like a black alligator

https://tenor.com/view/alligator-explosion-gif-19265301

[shy kelp]

Wow that alligator toy got violated

[spring fulcrum]

I forgot where to put bbox files in comfyui. Can anyone help me with that?

[shy kelp]

its not a native comfy type

so it would depend on the node pack

I would recommend instead storing the co-ordinates of the four corners as 4 seperate ints or floats

if this is impact pack its probably in their docs

I don't use impact pack because I have the opposite opinion about how it should be done, but impact pack does work very well

impact pack focuses on a new data structure they made called SEG, whereas I like to have everything seperate

[wheat rose]

it's in ultralytics

[light bone]

DRAW

[tough cliff]

A medieval game art scene

[faint shore]

Playing around with the IPAdapter style and composition example workflow.

[atomic rain]

Architectural illustration: a street with a coffee shop, where a couple with cups is reflected in a heart-shaped window. There is an open Valentine on the table

[tulip briar]

its very annoying i am using sdxl and i am trying to get a person wearing red jeans, so if i say, man in red jeans looking at the sunset, his shoes or shirt can become red, i am using sdxl atm what would you recommend i use, is it better with sd 1.5?

[rough thorn]

bruh

[unborn anvil]

Can anyone recommend me a good model for making creative art? I am using the standard SDXL with SwarmUI but it's not giving me what I want, prompting is hard man lol

[wheat rose]

sd3.5 medium is excellent, and artsy

[unborn anvil]

Thanks buddy!

[proud palm]

High-resolution, photorealistic image of a single person standing outdoors in a serene natural setting. The person, a [choose: young woman / young man / person of indeterminate age], is extending their arms forward, palms open, as if reaching towards the sky or embracing the sunlight. The setting is a breathtaking landscape – perhaps a vast open field, a mountain vista, or a peaceful forest clearing. The lighting is soft and natural, golden hour light, creating a warm and inviting atmosphere. Focus on capturing realistic details in the person's features, clothing, and the surrounding environment. The composition should emphasize the person's solitary presence and the feeling of peaceful connection with nature. Emphasize high detail, natural colors, and a sense of serene solitude and expansive freedom.

[stable verge]

Does anyone know an SDXL model which can easily pull style of know artists without lora ? Like old Deliberate able to reproduce quite a few artists from this page : https://supagruen.github.io/StableDiffusion-CheatSheet/index.html

[gray lagoon]

#artisan-faq

[frozen cradle]

is this the channel to ask questions about how to accomplish something

[gray lagoon]

Depends if its a prompting question or if you have need for #🤝｜tech-support

[frozen cradle]

Well I guess it’s a workflow question for comfyui

[gray lagoon]

You can try to ask but there might not be a answer if it involves custom nodes

[frozen cradle]

Alright, well for starters my main curiosity is how I should approach using a reference image that has the style of background/setting I desire and the character is also wearing clothes I want to base my prompts off of so I can aim for a similar result. Where should I start? I assume this is an img2img process.

[gray lagoon]

Hmm yeah mostly, i assume but for the best results id recommend a lora if your planning on making a TON of them

Hmm ip adaptor can help a lot for face consistency and style but since your in comfy idk

[frozen cradle]

Thanks, I’ll see if I can find a Lora

[torpid pecan]

you can try using Florence model to get a detailed prompt from an input image.

[shy kelp]

Male Tennis player on tennis court

[wet lily]

／The back of a woman on a cliff.

#The back of a woman on a cliff.

[frozen cradle]

are there any benefits or downsides of using multiple embeddings at once? not one in each but for example multiple positive embeddings. also, what is the syntax for embeddings in comfyui, especially if i ihave more than one like i mentioned?

[tidal garden]

id:guide A massive 3D Bengal tiger running through a jungle, covered with millions of honeybees. The tiger looks very distressed, shaking its body to get rid of the bees, but they keep sticking. Ultra-realistic lighting, cinematic quality, 4K render.

[glass forge]

IS this legit?

[spiral sail]

generate a picture of a portion in a magical forest

[torpid pecan]

Go to artisan channel for that bro

[glacial ruin]

Arthur morgan :>

[gray lagoon]

No

[glass forge]

trading view is web based - u cant sell "cracked" versions,

mods get this fool

[rare oyster]

tf is wrong with my stable diffusion i was generatin pictures and all of a suden it start generating 4 fingers on each hands 24h???

[smoky shard]

Here you go, buddy

[brisk hollow]

So, just wondering, how does one go about generating videos with SDXL/Pony?

Does animatediff even work with XL? The motion modules surely won't work

[gray lagoon]

animate diff is some ancient stuff by todays standards

[brisk hollow]

Yeah but I wanna generate vidoes locally

And I don't have a 32G card

[gray lagoon]

technically it should work

[brisk hollow]

I see txt2vid workflows for Hunyuan but not XL

[gray lagoon]

well yeah becasue hunyuan is a video model

and XL is a image model

Hunyuan is also capable of images if you set its frames to 1

[brisk hollow]

Any idea how much VRAM it takes to run Hunyuan?

[gray lagoon]

if you want a 8 second video IIRC it takes between 15-30 minutes on a 4090

since its like 240 images + consistency

but if you dont mind lower quality:
https://www.reddit.com/r/comfyui/comments/1huxvsd/hunyuan_video_running_on_4gb_vram_rtx_3050/

[wheat inlet]

please generate a black and white guernsey cow pattern

[gray lagoon]

🐄

[wheat inlet]

can you create it without the cow. just the pattern. and bigger. to be used for a website background

[faint shore]

@brisk hollow I'm running Hunyuan on an 8GB card. I can get 153 frames @736x416. Try generating a single frame as suggested first. Keep bumping up your batch length until you error out. Keep non-Hunyaun nodes to a minimum, or don't use them at all.

[pale sluice]

If you lower it to 640x360 (still a 16:9 ratio), how many frames do you think you could get?

[faint shore]

An important thing about Hunyaun is to make sure you are using their resolutions values. Drop down an empty latent node and click the width height. Find the closest values. So 640x360 may work, but a better setting is going to be 640x368.
Also, a big discovery for me is that the shorter your prompt, the more frames you'll get out of the system. Don't overload Hunyaun with a massive FLUX-Style, chapter-sized prompt. Start small, like "a red dog".
Once you start getting some output, try increasing the batch/frame length, and/or increasing the resolution. Then refine the prompt a bit more.

[pale sluice]

My reasoning for trying 360p is that it is exactly half of what the model is trained to output. Mathematically / Theoretically it should be better than using 368

[brisk hollow]

So are you generating in Comfy? Can you share your workflow with me, I have never used hunyuan, and for that matter I haven't been using Comfy all that much, and don't know the nodes well enough to build my own

[pale sluice]

He posted it here. #▶｜stable-video-diffusion message

[pale sluice]

Another one to get rid of when you get a chance @tidal venture, spammed a bunch of channels.
Account name is yam6666 in case they try to run.

[slim wren]

they can run but they can't hide.

[south horizon]

doing pixely stuff?

[tepid geyser]

Yeah! I found a node that converts things to pixel art and then coded in some new colors :)))

[torpid pecan]

Sounds interesting can you share the workflow ?

[tepid geyser]

I'd be happy too! But the metdata also contains the wf, and honestly there's really nothing special to mine. It's a pretty standard WF but the node itself is what made such a huge change. I will post that node's github in the server. That being said, the node I used was a modified version of that which allowed me to get the different colors

[torpid pecan]

Sure bro, that helps

[tepid geyser]

https://www.reddit.com/r/comfyui/s/WomfAeibFq

[tepid geyser]

There is the original node I got. 🙂
And then, if you want to do like me and modify it with different colors, I can show you the websites I used.

[torpid pecan]

hey, thanks a lot for this

[pale sluice]

Have you tried the DB16 or DB32 color palette?

[pale sluice]

I found an older but really interesting thread on Github that shows off what a mad scientist has done with SDXL.
https://github.com/cubiq/prompt_injection/discussions/1

[shy kelp]

yeah its so cool

[olive temple]

how do i use sdxl here

[slim wren]

You don't.
The only bot you can use here is sd3, cf #artisan-faq
For the rest you'll have to use online services (eg: https://stability.ai/stable-assistant) or install it locally.

[half hollow]

#✨｜sdxl How can i use different models using automatic1111 text-to-img enpooint. current when i set any model on webui the text-to-img endpoint use that model only, how can i will pass the model into text-to-img endpoint request

[sharp scarab]

Liminal Found Footage - [Flux Experiment]

More experiments, and project files, through: https://linktr.ee/uisato

https://www.youtube.com/shorts/dP3yCiqkHzA

[worldly glacier]

I'd like to transform a person's face photo into a cartoon-like character while keeping their recognizable features (just like loverse.ai does).

Questions I have:

1. SDXL vs Flux for this specific task - is one clearly superior, or are people just following the hype?
2. IP-Adapter configurations - is there a "golden setup" that actually works consistently, or is everyone just guessing?
3. Has anyone ACTUALLY created a workflow that matches commercial quality?
4. What workflow end-to-end to get same or better results?

I've seen countless tutorials claiming to solve this, but the results never match services like loverse.ai. Who's actually figured this out?

If you've got real insights (not just theories), I'd love to hear them.

[signal mirage]

First of all, not one picture of the loverse.ai site has cartoon characters. They got simply a ton of prompts for humans in different setting and use face swap.
If you would for example make a anime person out of a real picture you will get larger eyes, smaller lips and more, so that you would be still able to match the eye or haircolor/style but you would lose all kind of specific head elements.
So basicly i would use a simple face swap workflow (e.g. Flux + Pulid, Reactor + SDXL, Instant ID,...) and feed them good creative prompts always including a more realistic photo / image style.

[worldly glacier]

yes but in this type of examples as you can see they even match the color of the person and also the newly formed image is not exactly the same as the style, does that mean that they are using some sort of IP adapters? also based on what you said what would be the output the face and a prompt only or, prompt, style image and face image?

[signal mirage]

Well you would start with a llm or neural net to determine different characteristics of the person in the image. Skin, glasses, gender, age.
This will be included in the prompt. To make sure some elements appear (like the ice flakes) you could use a general composition and style ip-adapter.

[lament jungle]

1girl,black long hair,walking in the forest,sunshine,

[gray lagoon]

👧 🌲 🌲 🌲 🌥️

[gray lagoon]

Jk #artisan-faq

[marsh sundial]

Anyone could share a working WF using InstantID with SDXL please ?🙏

Would you mind to share your wf using instantid?

[glacial star]

One message removed from a suspended account.

[fallen magnet]

imagine an online store dased on darktheme ui,ux design selling shampoo, conditiooner, texture powder

[humble peak]

A boy dance

[dire basin]

Style helps a lot

[lost shoal]

Recommendations on how to get better text with SDXL?

[copper kraken]

bongsampling with RES4LYF

[dire basin]

Prompt takes 50%, while checkpoint and style attributes take the rest…
Need keep trying and tuning, no shortcut at the moment🥲

[faint shore]

Dropship stylized concept.

[hot juniper]

Modern living room includes: TV wall, table and sofa, wall hanging, door frame, decorative lights

[tranquil jay]

A whimsical scene of a ship sailing through an ocean wave,with its sails made entirely out of fluffy blue and white fabric.,

[gray lagoon]

🌊 ⛵ 🌊

#artisan-faq

[dire basin]

Same prompt, different device, different outcome

[stark sedge]

Try the dmd2 Lora.

[marsh sundial]

Anyone could suggest a good site (I m ok to pay) for a lora training for a character ?

[white iris]

Hi everyone, does anyone here have experience fine-tuning an SDXL model checkpoint on a large number of images? I'm hoping to improve its base performance. Some examples are like JuggernautXL , RealVis XL etc. Were they trained using dreambooth as well? Is dreambooth the best method to train a general model? As a lot of the examples of dreambooth seem to be only about training on a specific subject.

[gray lagoon]

Dreambooth is outdated and breaks the old web-ui

idk what the recomended way to train is for checkpoints but its generally the concensus here

[pale sluice]

Here's what you need. @gray lagoon You too if interested. https://github.com/Nerogar/OneTrainer

[plain hare]

/imagen

[spring pine]

anyone knows how to transfer pose of an image (using another pose image input) while keeping character face, clothing consistent?

[charred roost]

Here where I fixed bias and artifacts in an SDXL image:
https://vm.tiktok.com/ZMBMsNyMp/

@wheat rose

[wheat rose]

most people call that stuff hallucinations. i just call it poor quality and use photoshop to fix it

[faint shore]

Regional prompting using SDXL.

[prime pebble]

((cartoonish style), (Q版 fantasy)),
main elements:
smiling sun character with straw hat (拟人化太阳),
wheat fairy holding scythe (木属性精灵),
dynamic composition with wind-blown wheat waves (火性动感),
color palette:
orange sun (丙火),
emerald wheat (乙木),
light gray clouds (金属性弱化),
avoid deep blue or silver (忌水金)),
text overlay: "庚午匠心" in bold calligraphy (火属性印章)

[pale sluice]

https://tenor.com/view/luffy-wano-gif-22439049

[atomic ember]

"A retro sailboat docked at a small beach pier at sunrise, with coconut trees swaying in the breeze, gentle waves lapping at the shore, and a large, glowing sun rising over the horizon, in a 1950s-inspired illustrative style with a serene, pastel color scheme and sharp silhouettes.

[pale sluice]

You always have insane images, what model do you use?

[meager canopy]

Thanks! These ones were made using one of my own models: https://civitai.com/models/240590?modelVersionId=474681

[pale sluice]

Awesome, thanks. ^_^

[languid sedge]

A picture showing Two main types of bone include spongy (trabecular or cancellous) and compact (cortical) bone. osteons in compact bone and trabeculae in spongy bone. Figures showing osteons in compact bone and trabeculae in spongy bone, including osteogenic cells, osteoblasts, osteocytes, and osteoclasts and blood vessels.

[empty furnace]

There is a field of flowers underneath the clear sky, and the model lies on her chest facing the grass. Fashion film wearing colorful tank top dress and pink stockings and purple shoes, fairytale

[fringe herald]

Is there a complete list, potentially with example images, for someone not educated in art/design, covering all the different styles and whatnot that SDXL understands? I've only used SDXL for anime and semi-real generations, so I'm well-versed enough in the tag-based prompting of Danbooru, but stuff like "silhouette art, glitch art, virtual, colored shadow, chromatic aberration, polarized, etc." (taken from a prompt of a picture on Civitai) mean very little to me, plus I didn't even know these are understood by SDXL.

[fleet quartz]

‏“A man wearing old, tattered clothes is swimming in the deep ocean while being swallowed by a massive whale. The whale’s mouth is wide open, and the man is halfway inside, with his arms stretched out in panic. His face shows extreme fear and shock. The dark blue ocean is turbulent, with waves crashing around. Sunlight penetrates the water’s surface, casting a dramatic glow on the scene.”

‏“A man wearing old, tattered clothes is swimming in the deep ocean while being swallowed by a massive whale. The whale’s mouth is wide open, and the man is halfway inside, with his arms stretched out in panic. His face shows extreme fear and shock. The dark blue ocean is turbulent, with waves crashing around. Sunlight penetrates the water’s surface, casting a dramatic glow on the scene.”

How i can generate image???

[pale sluice]

I don't know if all of it still applies to SDXL, but there is this resource. https://github.com/SupaGruen/StableDiffusion-CheatSheet

[stray warren]

Here you go

[gray lagoon]

Hi there. theres gonna be a a really really long list

since illustrious and other fine tunes know different styles better

[fringe herald]

I wasn't specifically asking for styles, but rather artistic elements in general. That aside, is there a list like that?

[pale sluice]

define "artistic elements". Do you mean shadows, lighting, specific angles, type of shots, etc?

or do you mean genres like "film noir", "cyberpunk", "animation", etc?

There's no real cheat sheet for any of this. It all depends on what the model was trained on, and how verbose they were with captioning the image dataset.

[fringe herald]

Well, what are these "silhouette art, glitch art, virtual, colored shadow, chromatic aberration, polarized"? I guess the first three would be some kind of specific art direction/style/category? What would you call stuff like "colored shadow, chromatic aberration, polarized"? Like I wrote initially, my understanding of art, styles, direction, composition and everything else art related may as well be non-existent. I'm not even sure what exactly I'm asking for, really.

[pale sluice]

"chromatic aberration" is exactly what it says, you can look this one up for examples. And people really just know keywords that produce the effects they like. "chromatic aberration" for example, I throw in my negatives as it's complete crap in my eyes.

[atomic laurel]

How many images and epochs/settings would be best for an SDXL likeness model and how could I train that lora on a specific checkpoint

[paper parrot]

new illustrious merge

[haughty bloom]

[dire basin]

https://x.com/w4t_3r/status/1903761929816162563?s=46

🐱🚗🚀 (@w4t_3r) on X

Open for commission isn’t a deal unless no violation of law lol
#sdxl #aiart #aigirl #aianime #aiイラスト #stablediffusion #stablediffusionart #stablediffusiongirls #stablediffusionwaifu

X · Mar 23, 2025, 10:53 AM

[sharp scarab]

https://www.instagram.com/p/DGlVKAJxjOo

uisato_

Found it.

Technique consisting in a new synthetically trained AI model [FLUX.D LORA], and some ComfyUI wizardry, with the objective of accurately reproducing a 'found footage/liminal' aesthetic.

Both music and visuals by myself.

You can access these [new FLUX.dev LORA] + ComfyUI workflow + 2350 images and prompts in metadata + 26 img-to-vid e...

Likes

1265

Instagram

[dire basin]

https://x.com/w4t_3r/status/1905122623010529487?s=46

🐱🚗🚀 (@w4t_3r) on X

#sdxl #aiart #aigirl #aiイラスト #stablediffusion #stablediffusionart #stablediffusiongirls #aigirlart #aigenerated #aiartworks #aiartcommunity #runninghubcommunity #runninghub

X · Mar 27, 2025, 5:00 AM

[inland kayak]

Sleek product banner showcasing 3 gold real wax flameless LED candles (4", 5", 6" heights) with their realistic warm flickering glow. A modern remote control is placed neatly beside the candles. Clean, slightly blurred background suggesting a stylish home or event setting. Emphasis on convenience and modern technology. Professional studio lighting, sharp focus on products, photorealistic. --ar 3:1

[safe epoch]

Resume Photo

[kind nebula]

I am trying to do image captioning in Kohya using wd14 captioning. I uploaded the file with images I want to do captioning on but when I check Logs it says "found 0 images". Why is that, can someone help please?

[grave gorge]

Can somebody help me, please?
Why am I getting this result? I tried using a pony base model with the VAE it includes, but it doesn't work.
However, when I try using novanimeXL without the VAE, it works fine

[violet zephyr]

I know that whenever I use a model that requires a vae, I have to change the setting for vae from "Automatic" to the specific vae. The vae's aren't always labeled properly, so auto1111 doesn't pick them up sometimes.

[dire basin]

Work for model.pt?

[mint zealot]

hi, i'm a newbie in image generation and I can't get to generate what I want, can someone help me please ?
I'm trying to generate an image of a wounded wolf with signs of "corruption" (purple tendrils mushroom like), but I just can't make it work, it does a wolf indeed but there are no scars, wounds or anything (as you can see on the image)
i'm using juggernautXL

[paper parrot]

new illustrious model : https://civitai.com/models/398636/reymixxl. this one : #✨｜sdxl message

[rotund galleon]

anyone know why can't get a transparent image there is always a background apear on final output ?
[i'm using layer diffusion]

[true beacon]

https://civitai.com/models/1426048/saitomstylesdxl i trained a lora about the style of saitom(character design of xenoblade 2 and 3). Welcome to download and use if you are interested !

[fallen stream]

Is there a website version I can use of Stable Diffusion XL ?

with no restrictions

[dire basin]

civitai?

[torn spade]

how do you train it i wanna do that but i have no idea how

[gray lagoon]

Im gonna be honest chief, your completely new so id focus on getting great images first, and loras second

[torn spade]

i managed to get some good images tho is there like a tutorial for it

[fallen stream]

Might be a stupid question but I am new to Stable Diffusion, It's a noticeably better quality AI image generator than for example GrokAI, Chat-GBT or Meta but those chat AI understand exactly what I want in one go and generate way better images for what I want. Is there a way to make Stable Diffusion this way if you know what I'm saying

[normal monolith]

I'm no expert, but stable diffusion was trained on more "quality" images, hence the controversies about copyrights etc, and it's mostly just that: basic AI with tons of image data. It isn't nearly as "smart" as chat bots. If you have access to all the tools, why not generate what you want with bots and then "dress" the image in desired style with SD?

[fallen stream]

I'll try to do that even though I have no idea how to lol

but if SDXL had the understanding of the prompts as well as Chat-GBT or Grok it would be perfect !

[true beacon]

i trained with sketch to imitate the style

[normal monolith]

what style are you after? realistic/photo, anime, cartoon, 3d?

[true beacon]

the dataset is made of 45 imgs

i tried 145 imgs but the result is worse

[fallen stream]

Oh It works it's easier than I thought! Thanks mate

[echo crane]

That level of prompt understanding and good SD 1.5 finetune fine detail aesthetics would be quite something

[fair mural]

create an image of A dramatic historical painting-style scene of the French cavalry crossing a cracked frozen lake in winter. Soldiers in blue uniforms with gold accents struggle to control their panicked horses as jagged fissures spread across the ice. One horse has fallen through, its rider desperately grasping at the broken edges. The atmosphere is tense, with a cold, overcast sky and distant snowy landscapes. Highly detailed, cinematic lighting, realistic textures of ice and fabric, evoking a sense of impending disaster. Art style reminiscent of 19th-century military paintings with dynamic composition."

[sleek bramble]

/settings

[ornate bough]

create 3 images:A glamorous woman at a high - end soirée, with porcelain - like skin and almond - shaped emerald eyes. Her long, wavy chestnut hair is styled in loose curls, cascading over one shoulder. She dons a floor - length, form - fitting scarlet gown with a plunging neckline, adorned with intricate diamond embroidery. A pair of diamond - studded chandelier earrings graces her ears, and she holds a small, bejeweled clutch in her hand, standing in a luxurious ballroom filled with crystal chandeliers.

[tired viper]

Create an image of solar system

[glacial star]

One message removed from a suspended account.

[brisk jetty]

Create an image of end of times. An ul=nlimited barren land and all of the humanity standing there waiting for the judgement

[jaunty crow]

nizsys

nizsys

gopwu

Dior

[river agate]

hammed burger

[faint shore]

Some stylized hearts.

[grim nexus]

I use deforum with SDXL, I noticed that the images degrade heavily over time. E.g. prompt

Prompt:
High detail, illustration, dark purple colors, night sky, gentle glow over the whole scene, texture lora:SDXL_Space_Cowboys:0.2 space_desert, film grain, lora:Ethereal_Illustration:0.1, lora:acid_illustration:0.1, space desert landscape, acidillust, depth of field, paper texture, fine lines

Stellar Renegade with glowing mono glass sunglasses, Nebula, Riding a Space Bike, trippy desert bar, with cacti, illustration style, motion blur, high-speed, dark night scenery

Neg:
low detail, low quality, photo, nsfw, 3D, cgi, child, anime, manga, waifu, text, watermark, comic, human

If I run it through txt2img, or img2img I get a way better quality as compared to the quality after a bunch of frames in deforum.

It gets that more comic, thick lines, less detailed background look over time. I use strength between 0.3-0.7 and 0.065-0.15 with the beats of music. So my assumption would be that the image should reset quite good.

[glad cedar]

Hi! I'm building an app with SDXL-Turbo and realtime prompting, like the Stability AI video example. Anyone knows which python library can replicate that? But without ComfyUI, i want to integrate it to the app without starting the comfy server. I attempted with Gradio but can't do the realtime prompting, but waits until the next image is generated. Thanks!

[cyan cosmos]

#🆕｜sd3 Close-up view of a corner connection for stacked shelves (thin galvanized steel rectangular tube). The top surface of the lower shelf corner has small metal blocks welded to form a square locating pocket or fence. The upper shelf has a 10cm tall square spacer foot welded underneath its corner. This foot fits neatly inside the locating pocket/fence on the lower shelf. Show the 10cm separation created by the spacer foot. Detailed, metallic, industrial design, 3D render.

[rigid laurel]

why'd you link the SD3 chat?

[finite basalt]

https://tenor.com/lesdVyBsEB9.gif

[gray lagoon]

They think they can generate an image here

[rigid laurel]

yeah but why make a request in the SDXL chat and add a discord link to a different model chat in the request 😂

I'm just so curious about their reasoning.

[rich kindle]

Good day. I have a question regarding this live AI panel method. What is this called? is there any documentations or videos about this?
https://www.youtube.com/watch?v=KhgABU6mnPM

https://www.youtube.com/watch?v=MWWvc4yBMHY

i guess he is using this plugin?

https://youtu.be/i__ciRbs3VA

found it

[marsh sundial]

anyone has a pulID workflow working with SDXL ? i tried the example i got error : RuntimeError: Error(s) in loading state_dict for IDEncoder:
Missing key(s) in state_dict: "body.0.weight", "body.0.bias", "body.1.weight", "body.1.bias", "body.3.weight"

[buoyant lava]

/image_dream prompt: A cozy indoor scene featuring a modern fabric-textured bedside lamp with a white lampshade, sitting on a white surface next to a bottle of Cetaphil moisturizer and a baby bottle. Sheer lace curtains allow natural light to softly illuminate the neutral-colored wall and ceiling, creating a calm and serene atmosphere image:

[gray lagoon]

you could rent a GPU

vast AI or similar services

[gloomy mulch]

hello there im looking for a method/workflow for ComfyUi to make images with multiople characert, in wich each character has its own LORA if someone knows a way please let me knoe or at least guide me the right way...thanks

[past roost]

What would be tag be for a straight portrait shot of a face? Like the ones on your ID

[faint shore]

"Facing camera" or "Facing Forward" or "Facing viewer"?

[gray lagoon]

dont forget the tag

portrait

[past roost]

any upto date control net tutorials? I remember doing the uploading image, it turns all black with white lines to mimick the poses etc, idk how to do it now

[past roost]

right, i alreadfy have 1 of them installed

but idk where this is pulling the tags from

(regardless of ui)

[gray lagoon]

Danbooru?

[past roost]

its not, half of these tags dont exist

nothing shows up in danbooru when u start typing 'presenting'

hence, why it cant tagauto complete from the extension

[gray lagoon]

Hmm maybe derpibooru hmm yeah that's why i don't like random extensions that take over instead of trying my own csv

[past roost]

actually they do show up now, idk

maybe i misstyped or sum shi

but not all of em

Is it possible to make my own default settings? im tired of changing sampling, scheduler etc every time i launch

[gray lagoon]

Depends on the UI ngl

[past roost]

found it in the config.json

or settings tab>default ui

Does anyone know which scheduling type does euler a use when set to automatic?

[gray lagoon]

normal or simple useually

[brittle garnet]

Anyone with a rtx 3060 12gb know how long it take to train an sdxl lora? I've tried searching but can't find an answer.

[faint shore]

You could skip training all together and use the IPAdapter SDXL style transfer to enforce a likeness.

[past roost]

takes me hour and a half with 36 pictures on a 3070. U have more vram, so expect similar/better results.

[faint shore]

That doesn't seem bad at all. So you could train a lora overnight on a home system?

[brittle garnet]

That's actually pretty good, thanks

[desert spade]

How to create image

[old cargo]

Now I give you a sentence, can you generate a sentence describing the related picture for me

[raven idol]

Is anyone else having issues with SDXL image-to-image? I have uploaded a product, with white background, I have put in multiple different prompts for it to change the background of the product in the image and not the product itself, but it just doesn’t listen. I am using the api, image strength 0.8, if someone has had similar issues please let me know

[signal mirage]

Well i would say nearly everyone who would use the image 2 image without masks or controlnets will have the similar issue. As you allow the AI to modify the image (everywhere) not only the background. Controlnets, IP-Adapter etc. will help to guide the AI to change only the parts of the image you would like to change. UIs like Krita (with AI Plugin and for example Comfyui as Backend) or Invoke let you easier "edit" images .

[raven idol]

Okay, thank you!!!

[ebon breach]

is it possible to inpaint a clothing from a character lora ?

i cant seems to get it working

[tall knoll]

Okay I have some doubt so see in my training my sdxl model the images which i have used to train the model aree all of same size 1216x832 and when I used trained model to generate images previously I was using output image size of 1024 X 1024 but images were not that much good but when I used exact output size of 1216X 832 then the generated images were amazing why is it like this if anyone can guide me in this why is it like this

[vernal hinge]

Just started learning SDXL weeks ago. Here's what I got so far

[gray lagoon]

scam

[warm hare]

I'm curious what people are using for img-img upscaling scripts? I used to use the ultimate sd upscaler using the UltraSharp model, but that doesn't seem to work since I've migrated to SDXL/Illustrious.

[shut geode]

Can anyone give me suggestions on how to remove the (what i think is called) artifact in her eyes? That white glossy spot? It tends to appear on the pupil, nipples, sometimes nose, any pointed part of anatomy really.

Using an illustrious model on Forge WebUI. Happens regardless of model I choose. Current settings:

[brittle horizon]

Hi, is it real? To use text to image Model, English is the only language supported for this service?

I tried to use api with latin Words, it works too! 🙂

This is so bad, to use it for Content of international Websites!

[proud bridge]

hi, i am new to to stable diffusion , can anyone explain why it generating non realistic images, do i need to modify something
current gpu:- rtx 4060
using:- automatic1111

[gray lagoon]

looks " realistic" to me?

tough your using the base sdxl i see

maybe on civitai you can use a realism finetune?

[proud bridge]

should i change the model

[gray lagoon]

you could try

yeah

[glacial star]

One message removed from a suspended account.

[neon sentinel]

Try a higher rated model from SDXL on e.g. Civit.ai
Then use the resolutions preferred, like 1344x768, and rerender with img2img when happy, e.g. with resize 1.5x.
SDXL seems to prefer high resolutions to generate details. E.g. inpaint with 320x320 will generate worse results on an area than with SD 1.5, so get your final pic up in the resolutions, as well as inpaints (e.g. 1024x1024 for inpaint).

My 2 cents

[small pine]

Hi! Do anyone know how to use model L3-Stheno-Maid-Blackroot-Grand-HORROR-16B-D_AU-Q8_0.gguf in Forge? I tried a lot of settings and I didn't get the right one?

[gray lagoon]

That's not an sdxl model is it?

[small pine]

It seems it is a llama 3 model!

[warm geyser]

Anyone know where I can find some tricks to alter the features of different characters in the same image? They all just tend to get jumbled up no matter what I try

[dapper wasp]

as in, you already have an image? or do you want to generate an image with multiple characters that have distinct features

[gray lagoon]

#artisan-faq

[gray lagoon]

People cant read😭

[light karma]

This is actually one of my favorite things with this server. Seeing people randomly throw prompts in the wrong channels. Especially the really lewd prompts.

[gray lagoon]

Especially if they keep trying in multiple channels lmao

@formal vector prompt: read the #artisan-faq channel to learn how to generate images in this discord

Or he doesn't speak english or is a bot idk

[near prawn]

hello

[round plume]

Fooocus, it generated me an image based on certain prompt, how do I get more variations of this single model from the photo. This faceswap shi aint working or i just dont know how to use it. Anyone willing to help

[ancient haven]

Hey guys - is there any really good guide showing use of SDXL + a refiner (if need be) + upscaler (if need be)? I'm having a lot of trouble figuring out how to get face's not to look distorted, even though I think i've completely copied what people are doing online. I'm using bigaspv2 which i know is not a normal model but it has TONS of activity on civitai so i figure it must be possible to make some good stuff

[vernal hinge]

https://civitai.com/posts/16736334

learned gimp and in-painting with this one

[humble cape]

Illustrious ?

[vernal hinge]

Yea. Most of what i do is illustrious https://civitai.com/user/TK503

[pearl sable]

Girls shopping in a mall, lively, three-dimensional architectural scene, warm colors, flat style

[glad hound]

Hi all! I'm very new to this space and just experimenting a bit – but I’d like to start my own long-term workflow for an AI character project.

Before diving in fully, I’d love to have one or two LoRA models trained by someone with more experience – just to see how far I can push the quality and learn what’s possible. I already have 30–50 images (generated via Astria AI) – same face, consistent style.

Happy to pay for the work (PayPal/Fiverr/etc.). If anyone’s open for it or knows someone, feel free to DM me. Would mean a lot. Thanks! I am from Germany and my „Girl“ will be a realistic Charakter“

[humble cape]

which fine tune model and lora you use ?

[vernal hinge]

all the info is in the prompts

[dapper wasp]

roughly how much extra vram does it take to load 1 controlnet?

sigh
surely there's a mod role or something I can ping at times like this

dude I am not clicking on your shady discord invite from a person who's never talked on this server
especially when you could've just answered with a number

[gray lagoon]

Yeah its spam, who pinged you but you shouldn't ping a mod for this regardless

Hmmm

Id feel like it heavyly depends on the controlnet

2-3gb? People with 8gb vram should be able to use controlnet with no problem on nvidia cards

Will take longer though

Unless the mod comment was for the spam link then yes i agree i wish

[dapper wasp]

yeah I was talking about pinging mods for spam links, oh well
ty for your answer btw; tho follow up, can I run controlnet at like fp8 or smthn to make it take less?

[gray lagoon]

Hmm well if your on nvidia a 6gb card "could" run it too maybe

Will increase times massively tho

Fp8, im not sure

[dapper wasp]

I am on 8gb, though I think my card's vram modules have some problems cause my screen sometimes starts artifacting if I try to go beyond like 7.2gb

[queen vector]

Whats ideal epoch, repetition, network dim/alpha, unet lr and txt encoder lr for 200 photos?
I sought help from chat gpt twice and both the times it ruined it, one lora produces great results but result does not resemble the person in data set, second time it burned, all i can see is orange burn on image.
Chatgpt and grok both suggest something vastly different, i am running on cloud gpu, so cannot learn by trial and error a lot.

Any help is highly appreciated

Using OneTrainer cloud

[winter gust]

Can anyone tell me what the best furry models currently are?

[copper kraken]

just added style transfer (not ipadapter) to sd1.5, sdxl, sd3.5 (and flux, and hidream, wan, lightricks ltxv etc etc)

style guide left

[covert saddle]

Did you ever get it working? I am having the same problems.

[sharp scarab]

Kinestasis Stop Motion / Hyperlapse - [new WAN 2.1 LORAs]

https://www.youtube.com/watch?v=3rYPtYxq_3c&ab_channel=uisato

[cedar haven]

I used prodigy and consine. Set LR:1, TE:1 and UNet:1. But somehow LR turned into 0.00001 or 0.00002 after training. Why? SOMEONE HELP

[reef zephyr]

Looking to generate realistic human images in Stable Diffusion based on height and weight (e.g., 172 cm, 57 kg). Since exact body proportions can’t be controlled by prompts alone, what’s the best approach or tool—prompt engineering, control models like ControlNet, or post-processing techniques—to get accurate visual matches? Any recommended pipelines or methods?

[signal mirage]

Sorry for being confused but there are tons of elements you would need to consider to get an accurate image generated. For example the distance from the viewport, the angle of the “lens” or the hight of the camera.
Additionally a 57kg person might wear clothes which let her look like 58,5kg,……
There are some character studio tools outside in which you can use sliders for size of the characters and generate an image. You could use these (line, canny, depth) as input for controlnet.

[reef zephyr]

I've been using MakeHuman to create 3D human models and generate depth and pose maps from 2D renders. However, I'm having trouble getting consistent height in the generated characters. Are there any ControlNet models or alternative methods you'd recommend for better height control?

[signal mirage]

As mentioned with depth maps or canny you could influence the generated image. You might need to render a scene (living room, couch, lamp,…) then insert (gimp, krita, photoshop,…) the human at the right distance, convert to depthmap or use a canny preprocessing to feed into the controlnet.

[runic sun]

https://tenor.com/dFQZxcy9UP8.gif

[opaque marten]

this is so freaking cool

[obtuse tartan]

https://youtu.be/VHaVdqp7RRQ

[tulip briar]

I am trying to find a way to run stable diffusion xl in python but where it gives me good result, for example if i runt comfyui or fooocus i get better result bevause the have refiners etc but how could i run an "app" like that in python? I want to be able to run LoRa combined with image prompt and inpaint (mask.png). Does anyone know a good way?

[eternal sorrel]

sorry, noob at this
can i use controlnet with sdxl from diffusers library to control generation????\

[smoky shard]

SD 2 is still a thing? I thought it was displaced by SDXL

[vernal hinge]

not sure why i said SD 2. Im using SDXL

Also, who da girl? 😂

[smoky shard]

Marcille ||Donato||

[vernal hinge]

she looks like a cartoon version of my character

[burnt osprey]

A cozy indoor scene of a mother and child playing on a whimsical carpet, Pixar-style animation, 3D rendering, expressive exaggerated facial features, soft dynamic textures, warm smiles, child holding red and yellow toys, mother in light sweater with red trim, soft cinematic lighting, natural daylight through white curtains, muted rainbow and car patterns on carpet, glossy toys, plush fabric books, geometric shapes, bookshelf with plants, floral tablecloth, white cylindrical appliance, harmonious vibrant colors with earthy neutrals, subtle motion blur, dynamic angled perspective, nostalgic warmth, hyper-detailed textures, --ar 16:9 --style 4b

Key Parameters Explained:

1. Style Anchors: Pixar-style animation, 3D rendering ensures the CGI aesthetic.
2. Visual Details: expressive exaggerated facial features, muted rainbow/car patterns, geometric shapes aligns with Pixar's stylized realism.
3. Light/Color: soft cinematic lighting, vibrant colors with earthy neutrals replicates the studio's signature balance.
4. Composition: dynamic angled perspective, subtle motion blur adds cinematic depth.
5. Technical: --ar 16:9 for widescreen framing, --style 4b to enhance 3D/animated qualities.

[burnt osprey]

how to create image by image

[tribal laurel]

Hello!
I have the following scenario: i want to animate bats flying from a static cave image using transparent (alpha) bat video. (generated). Is there a workflow i can use as an example? thank you!
I have made a depth map and i use SDXL

[copper kraken]

regional prompting with SDXL

also supported:

SD1.5
SD3.5 (medium and large)
Flux
HiDream
AuraFlow
WAN

unlimited zones may be used, demo is 2 for simplicity

[drowsy cove]

Cat

[supple knot]

https://youtube.com/shorts/-yVzpvhKmVk?si=TH8iF5uK7DGQeL56

[faint shore]

Two different composition images, using the same style frame.

[surreal stump]

crazy how diff they feel with the same frame both hit hard in their own way🔥

[scarlet cosmos]

we're looking for a comfyUI designer to join our agency full-time, long term position, fully remote🙌🏻 anyone interested?

[final kelp]

image

https://tenor.com/8x8Y.gif

Just started learning SDXL weeks ago. Here's what I got so far

[queen vector]

Some help.

I found initial few success in lora training while using default.
But i am struggling since last night.
I made the best data set till now, manually curated high res photo (used topaz ai to enhance) and manually wrote proper tags individually.
264 photos of a person.
Augmentation - true (except contrast and hue)
Used batch size 6/8/10 with accumulation factor 2.

Optimiser : adamw
Tried 1. Cosine with decay
2. Cosine with 3 cycle restart
3. Constant
Ran for 30-40-50 epoch but somehow the best i got was 50-55% facial likeliness.

Learning rate : i tried 5e-5 initially then 7e-5 and then 1e-4 but all got similarly non conclusive result.
Txt encoder learning rate i chose 5e-6, 7e-6, 1.2e-5
As per chat gpt few times my tensorboard graphs did look promising but result never came as expected.
I tried toggling tag drop out on and off in different training , dint make a difference.

I tried using prodigy but somehow the unet learning rate graph moved ahead while being at 0.00

I don’t know how do i find the balance to make the lora i want. Its the best set i gathered, earlier on not so good dataset jt worked well with default settings.

Any help is highly appreciated

[old anvil]

hey guys, im trying to outpaint using the sdxl inpainting, but it keeps generating these borders. sometimes they are very subtle, but you can see them. hwo could i make it not to happen?

[surreal void]

try taking the "grow mask by" number down to 0. that may do it.

[old anvil]

it seems it was a model issue. I switched to 1.5 inpaintin and dreamsharper and it worked

[quiet loom]

Hi, I'm trying to inpaint with comfyUI and illustrious sdxl but am completely green, is there a some kind of guide how to or workflow with explanation for beginners ? I know that I need inpaint checkpoint but couldn't find a dedicated one for illustrious or am just dumb as hell or blind.

[maiden marlin]

does anyone have a recommendation which sdxl checkpoint i should use to train on this style?

[trim totem]

why can't i see "create masked area"?

[gray lagoon]

Scam

@trim totem dont join, its scam

[trim totem]

i know hahaha, thanks tho

[gray lagoon]

Recommend making a post in #🤝｜tech-support for this one tho

[faint shore]

@maiden marlin Do you really need to train something? You could supply those exact images to the IPAdapter style transfer node for SDXL. It's like having an on-the-fly lora.

[maiden marlin]

i already tried ipadapter but it didnt copy it well enough. and i tried training lora as well and its also less than satsifactory. tried both flux and sdxl. im trying dreambooth now, it probably should be better from what ive gathered

[silk current]

PA realistic standing image of Lord Kalabhairava, the fierce form of Lord Shiva. He is depicted with a terrifying yet divine expression, with three eyes glowing like fire. His complexion is dark as a stormy night, adorned with garlands of skulls and serpents. He stands powerfully in a cremation ground, surrounded by blazing fires and spirits. He holds a trident, a drum (damaru), a noose, and a skull bowl in his four hands. His hair is matted and flies wildly, crowned with a crescent moon. His feet are adorned with golden anklets, and he wears tiger skin. A dog stands loyally beside him. The atmosphere is mystical, with storm clouds and divine light behind him, capturing the essence of time and death. Style: Hyper-realistic, high detail, divine and intimidating aura, traditional Hindu iconography.

[stable ridge]

insouciant Incredible Hulk, King of the World, Fire And Ice, clasp portrait, aquiline features, crystal, haughty, aristocratic, Artwork By Boris Vallejo, Artwork By Louis Rojo, Artwork By Arthur Rackham, High Quality, Detailed , blizzard, masterpiece, darkness, goth, biome

[odd dagger]

Hello peeps

[velvet matrix]

/bot

[surreal stump]

Yo this looks insane, like some next-level fantasy world straight outta a dream! mind to show more of this?

[spare summit]

I try stable-diffusion-xl-1024-v1-0, how to deal with the edge

[smoky shard]

Here's full prompt:

1girl, scenery, masterpiece, (ancient ruins:0.5)
Negative prompt: worst quality, lowres, from behind, blurry, worst detail
Steps: 32, Sampler: Euler a, Schedule type: Karras, CFG scale: 5, Seed: 3207308689, Size: 1216x832, Model hash: c364bbdae9, Model: waiNSFWIllustrious_v110, Denoising strength: 0.67, ControlNet 0: "Module: reference_only, Model: None, Weight: 0.8, Resize Mode: Crop and Resize, Processor Res: 0.5, Threshold A: 0.8, Threshold B: 0.5, Guidance Start: 0.0, Guidance End: 1.0, Pixel Perfect: False, Control Mode: Balanced, Hr Option: Low res only", Hires Module 1: Use same choices, Hires CFG Scale: 5, Hires upscale: 2, Hires upscaler: DAT x4, Version: f2.0.1v1.10.1-previous-636-gb835f24a, Source Identifier: Stable Diffusion web UI

And this was the image for reference controlnet:

[glad hound]

hey is here a lora trainier for realistic ?:D

[rigid laurel]

oh man that's a lot of people who got hacked at once

[surreal stump]

the image looks vibrant and dynamic, but it could use a bit more realistic lighting.

[viral isle]

A cozy indoor scene of a mother and child playing on a whimsical carpet, Pixar-style animation, 3D rendering, expressive exaggerated facial features, soft dynamic textures, warm smiles, child holding red and yellow toys, mother in light sweater with red trim, soft cinematic lighting, natural daylight through white curtains, muted rainbow and car patterns on carpet, glossy toys, plush fabric books, geometric shapes, bookshelf with plants, floral tablecloth, white cylindrical appliance, harmonious vibrant colors with earthy neutrals, subtle motion blur, dynamic angled perspective, nostalgic warmth, hyper-detailed textures, --ar 16:9 --style 4b

#✨｜sdxl A cozy indoor scene of a mother and child playing on a whimsical carpet, Pixar-style animation, 3D rendering, expressive exaggerated facial features, soft dynamic textures, warm smiles, child holding red and yellow toys, mother in light sweater with red trim, soft cinematic lighting, natural daylight through white curtains, muted rainbow and car patterns on carpet, glossy toys, plush fabric books, geometric shapes, bookshelf with plants, floral tablecloth, white cylindrical appliance, harmonious vibrant colors with earthy neutrals, subtle motion blur, dynamic angled perspective, nostalgic warmth, hyper-detailed textures, --ar 16:9 --style 4b

[viral isle]

dancer of a boy on photo in sketch style

[slim wren]

a well written channel explaining how to use discord robot #artisan-faq

[surreal stump]

bro the detail in this is wild!...like, every pixel’s pulling its weight. like it

[earnest crow]

anyone got celebrity loras?

[rigid pelican]

https://youtube.com/shorts/qJWgfHrSt8A

[surreal stump]

love the concept...waves and lighthouse looks bold and striking

[fluid pasture]

Urban Guardian – Post-apocalyptic patrol deep within the ruins. Focused, armed, and ready. Inspired by cinematic survival thrillers and urban decay photography.
#AIart #Cinematic #SDXL #PostApocalyptic #StrongFemaleLead #UrbanDecay #SurvivorScene #🏞｜general-with-images

[trim whale]

is it true that sdxl is limited to 75 tokens?

how do you train a model with more than 75 tokens in tags?

[slim wren]

no

[trim whale]

but arent they separated by 75-75-75

I dont know how does it work behind the scenes

[slim wren]

it does not matter that much actually because tokens are """computed""" before passing the infos to the models.

iirc the limitation comes from CLIP

but basically regarding that topic, the pipeline goes like this :
"prompt" (made of words) -> vector of tokens -> numerical value -> feeding those to the model -> [model magic]

So yes the resuts will be "different" if you overcome the limitation of clip. But you won't loose any information because of it.

[trim whale]

could you help me with a problem then?

do you use lora easy scripts?

[slim wren]

the 75-75-75-75-etc will all be concatenated before feeding it to sd.

[slim wren]

not at all :/ sorry (also it's 2:00am here)

[trim whale]

I just want someone who knows the abcs

its difficult to find people that know about trainings :/

[slim wren]

I never did anything serious training wise, my gpu is only 8gb

[neat fern]

https://civitai.com/user/espiwnl142

[surreal stump]

like the first photo...just looks so natural and smooth

[neat fern]

Thanks:)

[azure cave]

this looks real, @neat fern! good workkk

[surreal stump]

other than stable diffusion, savro, and deepAI, what ai model generators do u all suggest? i’m looking to try more!

[neat fern]

Are you looking to make realistic generations?

[neat fern]

Thanks :)

[neat fern]

https://civitai.com/user/espiwnl142

[queen vector]

Greetings, could anyone please help me with sdxl base tagging?
Been 1 months, i am trying to master lora training, sometime I succeeded but my lora has prompt rigidity issues. In some prompt it works great, in others it breaks.

I am trying to train a celeb lora.
High def upscaled 250-300 photos, used wd eva large tagger, cleaned , curated, dint work. Used chatgpt, tagged accurately but poetic, dint work. Whats the ideal tagging strategy for prompt flexibility and better lora training with diverse data set. Which style of tagging, underscore or no underscore, compound tags or no, tagging cloth, pattern, color or no, hand movement, poses, gaze, anything specific?

[surreal stump]

yeah byeah

[torn sleet]

Create a photorealistic and realistic image with a resolution of 3840x2160 in a cyberpunk style. In the foreground, depict a very beautiful, slender woman with a short haircut, who is half Asian and half Caucasian. She wears thin, tight-fitting clothing with the inscription "Xaero," through which the outlines of her nipples are visible. In the background, show a megacity with dark tones accented by blue, pink, and purple colors, and a cyberpunk-style sports car parked nearby. Please generate 3 different variations of this image. The image should have photographic realism, with detailed lighting, textures, and atmosphere typical of high-end cyberpunk visuals

[covert coral]

message me if you want to create ai model easily

[neat fern]

I think I’m good mate ;)

#mogged

[pale sandal]

Bro you used a Flux lora for this? How'd you come up with that lol

Looking good

[neat fern]

Trial and error :) thanks

[pale sandal]

I didn’t even know that's possible at all. Have you had any luck with others? And how did that polyhedragon etc improve it for ya?

[neat fern]

Yh I am always experimenting with different LoRA’s and combinations. I found the Poly LoRA improves close up details when given the right prompt key words.

[pale sandal]

Very interesting. Thanks for the tip

[glass forge]

so based on everythign right now whats the most best sdxl model on civitai?

[neat fern]

Depends on your goals

[glass forge]

photorealism

[neat fern]

I have had good experience with EpicRealism. PM me if you want to know some more.

[brazen hatch]

Is andrew here from stable diffusion art.com

[neat fern]

https://civitai.com/user/espiwnl142

[final otter]

Create a 1990s realistic portrait featuring Mexican American singer Selena Quintanilla with long dark hair and bangs, she's wearing red lipstick she's smiling

[graceful solstice]

magical warrior

[last tangle]

Close-up professional corporate man headshot, modern business portrait. The subject's head and shoulders are tightly framed, filling most of the image. Focus is sharply on the face, particularly the eyes, with a shallow depth of field blurring the background.
Lighting: Three-point studio lighting setup optimized for a close-up. A soft, diffused key light directly or slightly to the side of the face to minimize harsh shadows. A fill light to subtly illuminate the shadow areas under the chin and nose. A hair light or rim light from behind to add a subtle highlight along the hair and separate the subject from the background.
Background: Smooth, solid, neutral dark gray or deep blue background, completely out of focus to ensure maximum attention on the subject.
Camera & Style: Simulated DSLR photography with a high-quality portrait lens (e.g., 85mm equivalent). The image should have ultra-detailed facial features, realistic skin texture (without excessive smoothing), and professional, neutral color grading suitable for business use. The overall feel should be confident, approachable, and trustworthy.

[surreal stump]

this one smooth bruh..how long u've been using civitai? been testing other ai tool like pykaso, ssavro,, maybe i give it a shott..

[neat fern]

I don’t use CivitAI for generation, I use it as my main hub to share my generations, check out other peoples work, and explore LoRA’s and Checkpoints. Very good site.

[faint shore]

I've added face blending to my SDXL workflow. Trying to make unique faces.

[red vale]

I really need help with my lora training is anyone available?

[azure cave]

cooool. keep it uppp @neat fern

[livid ferry]

how to use it?

[lusty hill]

how to use it?

[surreal stump]

ohhh i see thanksss man

[surreal stump]

this is cool man,,how much time u render to generate this output??

[faint shore]

Those are 94 second gens on Windows 11 using a 3060RTX 12GB. Pretty fast considering the workflow does hand and face fixup on top of a mild up/down scale.

[open crescent]

👍

[surreal stump]

woah that's cool bruh

[fiery jolt]

sakura, white, pink --ar 9:16 --sref 2121577414

[devout lynx]

Mujer con la mente fragmentada

[placid condor]

Hello

[robust arrow]

extremely realistic DSLR photo of a beautiful young woman, light skin, pale complexion, natural beauty, long slightly wavy black hair with middle part, symmetrical and well-proportioned face, full lips, small nose, large brown eyes, thick dark eyebrows, smooth skin texture, expression: tongue out playfully with relaxed eyes, body visible down to thighs or knees, wearing thin see-through camisole, no bra, clearly posed as: standing by window, body turned sideways, one knee lifted slightly, background setting: clean white-walled room with realistic ambient daylight, sunlight casting directional light across wall and bed, full mirror against wall reflecting room elements, rug and slippers partially cut off by frame, strong outfit visibility, confident and natural posture, realistic soft lighting on body and clothing, ambient shadows, high detail skin and facial features, shallow depth of field, 85mm lens bokeh, ultra sharp focus, high resolution, same facial identity, natural skin texture, subtle facial asymmetry, realistic eye reflections

is there any lora i can use to help the facial structure stay intact in tongue out prompts?

[neat fern]

https://civitai.com/user/espiwnl142

[graceful solstice]

Princess Peach racing in dune buggy

[azure cave]

this is nearly perfecttt @graceful solstice <33

[rapid tinsel]

a divine digital painting of Lord Krishna as Radha Ramana sitting beneath a blooming kadamba tree on a carved stone bench, Radha resting gently against his shoulder, Krishna wearing a saffron‑yellow silk dhoti and peacock‑feather crown, softly playing the flute, Radha in a pastel pink and turquoise lehenga with jasmine garlands around her braid, lotus‑filled pond glimmering behind them, morning mist and golden rays filtering through leaves, peacocks and deer in the background, tranquil Vrindavan atmosphere, ultra‑detailed devotional art, cinematic soft lighting, peaceful romantic mood, high‑resolution

#✨｜sdxl #a divine digital painting of Lord Krishna as Radha Ramana sitting beneath a blooming kadamba tree on a carved stone bench, Radha resting gently against his shoulder, Krishna wearing a saffron‑yellow silk dhoti and peacock‑feather crown, softly playing the flute, Radha in a pastel pink and turquoise lehenga with jasmine garlands around her braid, lotus‑filled pond glimmering behind them, morning mist and golden rays filtering through leaves, peacocks and deer in the background, tranquil Vrindavan atmosphere, ultra‑detailed devotional art, cinematic soft lighting, peaceful romantic mood, high‑resolution

#✍🏼｜rules-and-tos #✨｜sdxl
a divine digital painting of Lord Krishna as Radha Ramana sitting beneath a blooming kadamba tree on a carved stone bench, Radha resting gently against his shoulder, Krishna wearing a saffron‑yellow silk dhoti and peacock‑feather crown, softly playing the flute, Radha in a pastel pink and turquoise lehenga with jasmine garlands around her braid, lotus‑filled pond glimmering behind them, morning mist and golden rays filtering through leaves, peacocks and deer in the background, tranquil Vrindavan atmosphere, ultra‑detailed devotional art, cinematic soft lighting, peaceful romantic mood, high‑resolution

#✨｜sdxl

[rapid tinsel]

#✨｜sdxl a divine digital painting of Lord Krishna as Radha Ramana sitting beneath a blooming kadamba tree on a carved stone bench, Radha resting gently against his shoulder, Krishna wearing a saffron‑yellow silk dhoti and peacock‑feather crown, softly playing the flute, Radha in a pastel pink and turquoise lehenga with jasmine garlands around her braid, lotus‑filled pond glimmering behind them, morning mist and golden rays filtering through leaves, peacocks and deer in the background, tranquil Vrindavan atmosphere, ultra‑detailed devotional art, cinematic soft lighting, peaceful romantic mood, high‑resolution

[graceful solstice]

Enchanted Iris

[wraith elbow]

[river forum]

ks

[astral basin]

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, 1 girl, solo, barefoot, black hair, hood, necklace, water, pants, hood, belt, shell, standing, chain, black pants, black t-shirt, t-shirt is covered in blood, blood flows down the T-shirt, against the background of dark entities with fiery red eyes, cape, jacket, red eyes, hooded cape, looking at the viewer, flower, chest, bracelet, walking, red lips, ripples, makeup, long hair, walking on fluids, collarbones, black nails, leather, jacket on shoulders, black coat, empty eyes, ((ominous atmosphere)), everything in red and black colors, ((everything is covered in blood)), there is blood everywhere, a terrifying situation

[native knot]

No.

[graceful solstice]

A cat waking up

[analog ice]

Ultra-realistic photogrammetry 3-D globe named Gloxus, 16-k resolution earth texture, micro-topographic detail on every mountain ridge and river delta, continents carved from obsidian-black basalt with razor-sharp displacement maps, iridescent neon-cyan ocean currents swirling under a thin glass layer, holographic magenta circuit-veins mapping city lights across landmasses, subtle cyan grid lat/long lines hovering 2 mm above surface, cinematic rim-light from a cool white sun at 45°, micro-scratches and fingerprint smudges on glossy protective dome, shallow depth of field f/1.4, 32-bit HDR, octane render quality, ray-traced reflections, photoreal shadows, ultra-sharp 200 mm lens, clean black studio background, --ar 16:9 --cfg 12 --steps 40 --sampler DPM++ 2M Karras --vae kl-f8-anime2 --no text, watermark, logo, frame

[azure vigil]

anime artwork anime artwork thick outlines,outline,masterpiece, bestquality, score_9, official art, detailed painting, remodernism, 1girl, long hair, silver hair, bangs, blue eyes, blush, collarbone, n_pples, p_ssy, navel, medium breasts, looking at viewer, open mouth, smile, teeth, solo, eyebrows visible through hair,Du Qiong,firefly (honkai: star rail), official art, a detailed painting,lineart, score_10,anime,Hi-res,cgi . anime style, key visual, vibrant, studio anime, highly detailed, anime style, key visual, vibrant, studio anime, highly detailed
Negative prompt: dirty, EasyNegative, FastNegativeV2, BadNegAnatomyV1-neg, badhandv4, bad eyes, plastic, deformed, watermark, shirt, dress, bra, clothes, bad anatomy, blurry, (worst quality:1.8), low quality, hands bad, face bad, (normal quality:1.3), bad hands, mutated hands and fingers, extra legs, extra arms, duplicate, cropped, text, jpeg, artifacts, signature, username, artist name, trademark, title, multiple view, reference sheet, long body, multiple breasts, mutated, disfigured, bad proportions, bad feet, ugly, text font ui, missing limb, monochrome, censoring, score_1, photo, deformed, black and white, realism, disfigured,, photo, deformed, black and white, realism, disfigured, low contrast, photo, deformed, black and white, realism, disfigured, low contrast
Steps: 40, Sampler: DPM++ 2M SDE, Schedule type: Karras, CFG scale: 5, Seed: 40841707, Size: 2048x2048, Model hash: da09ed5708, Model: inpaintSDXLPony_inpaintPony, Denoising strength: 0.4, Conditional mask weight: 1.0, ControlNet 0: "Module: canny, Model: diffusers_xl_canny_full [2b69fca4], Weight: 0.75, Resize Mode: Crop and Resize, Processor Res: 512, Threshold A: 100, Threshold B: 200, Guidance Start: 0.0, Guidance End: 1.0, Pixel Perfect: False, Control Mode: Balanced, Hr Option: Both", Mask blur: 4, Version: f2.0.1v1.10.1-previous-635-gf53307881, Module 1: sdxl_vae

#✨｜sdxl anime artwork anime artwork thick outlines,outline,masterpiece, bestquality, score_9, official art, detailed painting, remodernism, 1girl, long hair, silver hair, bangs, blue eyes, blush, collarbone, n_pples, p_ssy, navel, medium breasts, looking at viewer, open mouth, smile, teeth, solo, eyebrows visible through hair,Du Qiong,firefly (honkai: star rail), official art, a detailed painting,lineart, score_10,anime,Hi-res,cgi . anime style, key visual, vibrant, studio anime, highly detailed, anime style, key visual, vibrant, studio anime, highly detailed
Negative prompt: dirty, EasyNegative, FastNegativeV2, BadNegAnatomyV1-neg, badhandv4, bad eyes, plastic, deformed, watermark, shirt, dress, bra, clothes, bad anatomy, blurry, (worst quality:1.8), low quality, hands bad, face bad, (normal quality:1.3), bad hands, mutated hands and fingers, extra legs, extra arms, duplicate, cropped, text, jpeg, artifacts, signature, username, artist name, trademark, title, multiple view, reference sheet, long body, multiple breasts, mutated, disfigured, bad proportions, bad feet, ugly, text font ui, missing limb, monochrome, censoring, score_1, photo, deformed, black and white, realism, disfigured,, photo, deformed, black and white, realism, disfigured, low contrast, photo, deformed, black and white, realism, disfigured, low contrast
Steps: 40, Sampler: DPM++ 2M SDE, Schedule type: Karras, CFG scale: 5, Seed: 40841707, Size: 2048x2048, Model hash: da09ed5708, Model: inpaintSDXLPony_inpaintPony, Denoising strength: 0.4, Conditional mask weight: 1.0, ControlNet 0: "Module: canny, Model: diffusers_xl_canny_full [2b69fca4], Weight: 0.75, Resize Mode: Crop and Resize, Processor Res: 512, Threshold A: 100, Threshold B: 200, Guidance Start: 0.0, Guidance End: 1.0, Pixel Perfect: False, Control Mode: Balanced, Hr Option: Both", Mask blur: 4, Version: f2.0.1v1.10.1-previous-635-gf53307881, Module 1: sdxl_vae

[shy kelp]

#✨｜sdxl Create a stylized image of Queen Grimhilde from ( Snow White 1937 Disney style). Make her face look angry and sharp, just like in the original movie. She should have a small waist, large chest, and wide hips, wearing her iconic purple gown with black cape and golden crown. The background should be her castle’s throne room. Keep the art in classic Disney animation style — vintage, 2D, cel-shaded look. She should look powerful and seductive at the same time.

[graceful solstice]

A black cat in a vintage turtleneck sweater

[surreal stump]

so cuteeeee! what is he reading all about?

[graceful solstice]

Maybe “Why Humans Think Typing Makes Them Fancy – A Feline Investigation”

You wanna peek at what I’m typing at the moment?

[graceful solstice]

Different furniture

[dapper cobalt]

question, if i have the stable diffusion webui, how can i set up sdxl?

[rocky gale]

I m thinking of rentimg my gpu machine fully secure, a6000 ada 48 gb ram, i 9 32 core process static ip ,full power backup, online it costs about 800 dollars - 900 dollars a month , i m willing to give it in 500 dollars a month , if anyone interested ping me

120 gb ram

[wide urchin]

/saveid deeksha

/saveid

[alpine night]

i don't know how to generate images here. help

[faint shore]

Read the Artisian FAQ channel. You can generate in those channels.

[graceful jewel]

can you help me create an image of before and after a home renovation ?

[last haven]

Hey can anyone give me a ballpark number on how much faster a 5090 w 32gb vram is compared to a 4070 super 12gb? Specifically at generating/detailing/upscaling images. Thanks

[warped dirge]

lookup ai benchmarks

[boreal oriole]

Please use your screen sharing tool.

So that I can see your status easily.

[past roost]

i dont have either, but i can say significantly

[graceful solstice]

Wailord and the divers

[celest warren]

a man doing bbq in the backyard. he is holding black meat claws made with nylon. trying to shred chicken, smoked, grilled chicken. night environment. bokeh light in the background

[prisma spade]

Good morning everyone!

I'm trying to train a LoRa on Runpod and I've had several tests with poor results. I was wondering if anyone could help me...

It's a LoRa to replicate the phenotype of an Argentinian woman. I've prepared a dataset of 202 high-quality images. Generally, for image generation, I use Juggernaut XL as base model, but I read that the base model should be SDXL Base 1.0. In my last test, I noticed that if I use the RealVisXL checkpoint, it gets a bit closer to the images in the dataset. Could it be that if I want to generate images with Juggernaut, I should train the LoRa with that base model?

I'm sharing my training configuration to see if anyone can help me.

Dataset Configuration
Number of Images: 202
Repeats: 7
Epochs: 6

Base Model Configuration
Base Model: stabilityai/stable-diffusion-xl-base-1.0

Dim: 64
Alpha: 32
Base Resolution: (1024, 1024)
Enable Buckets: ✅ Yes
Min Bucket Resolution: 256
Max Bucket Resolution: 2048
No Upscale: ✅ Yes
Bucket Steps: 64
Optimizer: 8-bit AdamW
Learning Rate: 1e-4 for everything (U-Net and Text Encoder)
Gradient Accumulation Steps: 1
Batch Size: 4

I realized that by doubling the weight lora:argenta:2, it gets a bit closer to the desired result, but it always generates a similar-looking woman

Thank you very much in advance!

[past roost]

looks to me like you are overbaking the fuck out of ur lora

[graceful solstice]

Fantasi verden

[fluid pasture]

Testing a custom SDXL setup. Post-apocalyptic vibe, raw look, no upscaling. Let me know if it’s worth pushing further.

[untold edge]

Gluttony, is a mythical creature, in the classic ancient Chinese mountain, theyare described as green skin four-legged alien creatures, with the shark teeth,eyes on the shoulder, eyes shining, and seas in the classical Chinese culture isdescribed as fengshen, Chinese god beast, wings of a huge cloud, windfuryhuge explosion, release the power, lightning, movies, the epic scenes, vividlight volume, Fantasy, surreal, highly detailed, octane rendering, light andshade contrast,Unreal Engine,symmetry, 4k,hd

#🆕｜sd3 Gluttony, is a mythical creature, in the classic ancient Chinese mountain, theyare described as green skin four-legged alien creatures, with the shark teeth,eyes on the shoulder, eyes shining, and seas in the classical Chinese culture isdescribed as fengshen, Chinese god beast, wings of a huge cloud, windfuryhuge explosion, release the power, lightning, movies, the epic scenes, vividlight volume, Fantasy, surreal, highly detailed, octane rendering, light andshade contrast,Unreal Engine,symmetry, 4k,hd

[silver lintel]

Hey All, just looking for some help to try to fix my logo with SD, is this where I ask my question? If not, where should I ask? I am using SDXL in Gradio, BTW.

[median sluice]

a panda surfing

[brisk field]

a man with red hat

[echo ginkgo]

I need help. Can someone tell me why, when I use Illustrious XL with a cartoon LoRA, the images I get have deformities and look low quality? When the LoRA is supposed to be capable of generating images with the quality of the second one.

pls

same prompt and config and same issue

[glad snow]

A dynamic vertical split-scene poster featuring a professional wrestler. At the top of the image, he stands dramatically on top of a ladder inside a wrestling ring, arms outstretched, facing a roaring crowd. He has long black hair, turquoise tribal face paint, and wears red elbow pads. Below, in the same image, he is flying mid-air in a Swanton Bomb move, leaping from the ladder toward an unseen opponent. The crowd and arena are in grayscale, while the wrestler is in vivid, high-contrast color. Comic-book style, intense motion blur, dramatic lighting, full body view –v 5 –ar 2:3 –style raw

[vocal cliff]

/generate a retro-scifi close-up illustration of a beautiful gothic girl in a library, glossy red lips, gothic black dress, rows of books in the background, detailed, spooky, Tim Burton style

a retro-scifi close-up illustration of a beautiful gothic girl in a library, glossy red lips, gothic black dress, rows of books in the background, detailed, spooky, Tim Burton style

[violet arch]

#📝｜prompting-help

/generate jdkd

jeidis

[native knot]

Tried your prompt:

[south horizon]

Did you get this figured out?

[echo ginkgo]

Yes, thanks anyway!

@south horizon do you know where I can find a reference controlnet?

[south horizon]

i think regular controlnet does that, or something like ipadapter. Or that new Kontext thing. Not something I really do.

[silk frost]

Hi, does anyone know where i can find a high quality lora trainer for sdxl. Insta influencer / OF quality. I have the dataset.

[graceful solstice]

The wind blows the grass, the clouds float, the stream flows, and the camera moves from the ground to the sky

[nova aspen]

Which settings did you used?
Try lower the lora strength and if you share the lora link I can test,
oh you already got it solved

[graceful solstice]

Gongbi dancai of a white-haired girl in a blue robe, cradling a colorful bird, nestled with a giant leopard in a dreamy birch forest.

[shy harbor]

Are you using any LoRAs for the style?

[short grotto]

naw, but i do a fairly involved process

i use invoke, novaFurryV8b first, then then animeArtflow_v14 for the upscale

+3 creativity -3structure for the top layer

-3 creativity -3structure for the background

then scrub the bad parts away

and inpaint the edges

[graceful solstice]

The samoyed

[echo ginkgo]

If I have my configuration like this, is it normal for the ip adapter to ignore that?

[patent viper]

SDXL BETA BOT

[woven breach]

a big watermelon

[graceful solstice]

Seasonal circulation

[topaz spade]

i use invoke, novaFurryV8b first, then then animeArtflow_v14 for the upscale