✨｜sdxl

[west breach]

Hoard everything!

[uncut steeple]

[fallow kestrel]

what's going on with the ticket tool?!

[lean granite]

Oh yeah id rather keep it too overall

and no idea, ticket is going crazy

[fallow kestrel]

hahaha wow

[west breach]

It's entertaining 😄

[lean granite]

also what is that save image node? I dont seem to have all those options in mine

[fallow kestrel]

I think we have to take it down...

[uncut steeple]

@meager sundial

[fallow kestrel]

IT'S GETTING WORSE

[upbeat summit]

technical support!

[fallow kestrel]

oh

[fallow kestrel]

it's gone

[sour obsidian]

it will be back...

that was temporary

[fallow kestrel]

Aha!

[upbeat summit]

very temporary

[lean granite]

tis back

my save image node looks like this btw

got nothing in it but that x)

[fallow kestrel]

same for me, should be fine :)

[lean granite]

was hoping to get this tho

[fallow kestrel]

there are custom nodes though that could have more options, I don't know where to find them tho

[fallow kestrel]

yeah that looks custom

[hollow dirge]

why does it not see my checkpoints?

[woeful patio]

anyone else have the problem that bodies are getting streched out when you make the height more then the width?

[uncut steeple]

I think its from an extension uhm

[lean granite]

wonder if its any of these

[fallow kestrel]

did you put them in your models/Checkpoints folder?

[lean granite]

image gallery one?

[livid cradle]

anyone got a comfyui workflow for image-to-image with masking generation?

[fallow kestrel]

the model should be a .safetensors file inside your models/checkpoints folder

[trim totem]

can I use the refiner in comfy ui?

[fierce hollow]

what's this? somehow it even has my nodes

[eternal fog]

I was using "textured skin" as a stylistic prompt. But I think it also depends on what style of photograph you are going for.

[uncut steeple]

its one of these, couldnt figure it out yet

[vocal rapids]

lol just did a "high res fix" in comfyui

[eternal fog]

Happens sometimes. I've had some luck mitigating it by using the SDXL text encoder node and setting the height, width, target height and target width to match the aspect ratio of the latent you are using.

[lean granite]

ill just get em all xD

[hollow dirge]

oh thanks... I made a silly mistake in the "extra model path" document :D

[lean granite]

its a node manager, helps with installing

[fierce hollow]

ah that's why I suddenly have more issues on github

[lean granite]

probss

[fallow kestrel]

incredible how there is already a node manager for a model that came out yesterday. Where did you find it?

[west breach]

mine aren't in there... thank god

[fierce hollow]

it's a node manager for comfy as a whole, the ui's been around for a while

[lean granite]

^

[uncut steeple]

I need to figure out how to upscale without loosing so much detail

[fallow kestrel]

Which upscaler are you using?

[jolly creek]

Does comfyui has an api like A1111 has txt2img?

[west breach]

I was thinking we need something that detects how busy an image is and automatically adjusts the number of steps on the sampler running on the 2x image

[delicate grotto]

You can upscale by small margins, 1.1x each time wont lose much

[uncut steeple]

The long run

[fierce hollow]

apparently, yes, never tried it personally but there are some examples in script_examples dir

[delicate grotto]

Differentiql changes are almost always superior to to large ones

[woeful patio]

sdxl is gonna be a game changer

[fallow kestrel]

Oh damn okay sorry, that's above me 😅

[fierce hollow]

the more fields the better the quality

[tropic turret]

Can I use both the Vae and Refiner with the base SDXL model in ComfyUI? Sorry for asking, I am new to this.

[jolly creek]

Hmm thanks. Then I might as well give comfyui a try, at least until A1111 implements proper refiner usage

[west breach]

are you using refiner for the ultimate upscaler?

[delicate grotto]

Optical polarizers have the same thing
Use 1 to change an angle and you lose between 0 and 100% opacity
Use thousands and you can continue spinning it without losing any opacity almost at all

[uncut steeple]

yep

[west breach]

have you tried lower resolution tiles?

[uncut steeple]

I thought 1024 would be fine since its XL
Will do

[fallow kestrel]

My god here I was thinking I knew about 80% there was to know about 1.5, and now everything is so exciting again - I could watch tutorials and try different things for hours, it's so much fun

[uncut steeple]

Watching tutorials for hours = Fun?

[fierce hollow]

sounds like absolute pain with how every channel tries to maximize time, peddles sponsors, ads, etc

[lean granite]

this is the manager node for anyone who wanted to kno earlier:
https://github.com/ltdrdata/ComfyUI-Manager

[delicate grotto]

The dumbest ones are those that take literally 3 seconds to do with 10 minute vids

[fallow kestrel]

yessir

[delicate grotto]

I abhore it that the dislike button is gone

[fallow kestrel]

the longer you watch tutorials, the more things you have on mind like "Oh I gotta try that one!" And you just watch until you can't take it anymore and gotta try it. It's a lot like edging

[fallow kestrel]

well yeah you gotta be picky about the chanels you watch, true

[fierce hollow]

clickbaits be like: REVOLUTIONARY FREE AI MODEL

[fallow kestrel]

"This changes everything!!!"

yeah...

[west breach]

Is SDXL the 1.5 killer?

[delicate grotto]

10 things you should know about ai/sd scary
Went wrong

[upbeat summit]

thanks for sharing your workflow and how it is set up! I really like some of the ideas

[vocal stream]

you must be really suspectible to clickbait if you are consinstetly playing those ones

[soft zealot]

my pleasure. It is heavily laid out for my personal usage

[fierce hollow]

I think it's more of an issue with every google search yielding those,but probably getting too offtopic

[delicate grotto]

Short vids wont generate lots of cash, so yt doesnt promote them, they are usually at the page 50 or below

[vocal stream]

page 50? if you look for a specific extension or method or something you'd see one in top 5

[delicate grotto]

Top 5 if you are looking for something that is known

[vocal stream]

are you just searching ' ' and scrolling until you get a vid you need, i dont get it page 50

[uncut steeple]

I mean, maybe?

[delicate grotto]

Thank god GPT is here

[fallow kestrel]

if you watch a lot of channels you trust, the algorithm will recommend you more of those. Just don't be tempted by clickbaity bullcrap

[fallow kestrel]

Hey, that's me!

[vocal stream]

skin still seems a bit more overprocessed then everything else but pretty good

[fierce hollow]

As an AI language model I can't recommend any YouTube tutorials since they might be unsafe.

[fallow kestrel]

What does the ascore inside the Clip Text Encoder?

[delicate grotto]

Btw, GPT has lost its accuracy due to PC bs... it cant even do math now

[delicate grotto]

I meant just asking gpt directly instead of yt/google/reddit surfing

[fallow kestrel]

Just use Falcon Locally with OogaBooga

[delicate grotto]

Cant
Takes too much space with the stuff it downloads

[woeful patio]

how can you outpaint with comfy?

[vocal stream]

no chance is asking GPT about stuff like sdxl that have came out 2 years after what it's trained on is better than just google

[fallow kestrel]

Oogabooga with Falcon 13B is like 20GB? 🤔

[fallow kestrel]

You can collect all the documentation on SDXL and train a LoRA on your LLM lol

[fierce hollow]

btw if you thought GPT was too PC try Llama-2 with the default system prompt, it thinks even mayonnaise recipes are harmful...

[delicate grotto]

I meant what it downloads into disk C
had like 70gb free, after oobabooga had like 35, and i still cant find where in the FUCK DID IT PUT ALL OF ITS SHIT THERE

[vocal stream]

eh local gpt, especially one small enough that you can train even loras for also doesnt sound like it'd be that insane compared to just ctrl-f-ing and skimming the documentation youd train on

[fallow kestrel]

lol that sounds odd. Oogabooga has a one-click installer that's usually quite slim. I'd recommend having ai stuff on a seperate SSD from your OS-Drive anyways tho

[clever verge]

Here https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main

[vocal stream]

but i guess gpt4 + large context + docs yeh i can see that

[fallow kestrel]

true

[fallow kestrel]

AHHH!

[delicate grotto]

I have a fully uncensored model, ask it how to make an explosive shuriken and it will lend a hand

[uncut steeple]

Yeah im still looking for the right settings

[fallow kestrel]

same ^

[fierce hollow]

yeah just saying the base models are kinda out there with the safety stuff

[proven crane]

it seems like there are 2 comfy workflows now, one is using the refiner on top of a complete image from the base model and the other one is refining an image with leftover noise

[fierce hollow]

nous-hermes trained on llama2 works great for that matter

[vocal stream]

that does look slightly more realistic at first glance (tho whether it is better aesthetically, not sure)

[proven crane]

i'm not sure which one is supposedly corect

[fallow kestrel]

Just browse the huggingsface leaderboard for LLMs and look for the score on HellaSwag, as this seems to be the best benchmark rn

[delicate grotto]

The safety is whats killing their intelligence sadly
They became delusional at a certain point where it lies to you due to censorship

[vocal stream]

if you use gpt4 through the API (if you have access) and give it a different system prompt it's way less censored

[fallow kestrel]

Although I don't disagree, it's kind of hard to pinpoint what exactly made chatGPT worse. I doubt that the only problem is HRL and alignment

[uncut steeple]

I wouldnt say there is a correct way, just different approaches and we will see which yields the best results sooner or later

[proven crane]

yeah we can be pendantic about what's correct or not, the community can find new ways to use it

but i guess official way maybe a better word for it

[delicate grotto]

Well, i know when it happened
Sadly talking about it will get me in trouble here

[proven crane]

i guess i could look at the original repo to see roughly how it's "supposed" to work

[fallow kestrel]

I love nous-hermes

[tropic turret]

I tried loading in your workflow and got this issue

[orchid ledge]

fundamentally they are the same workflow just with different levels of noise

[vocal stream]

there's a new model from july (you can see it in the api), but you can still access the model that was there from march and is 100% unchanged if you need

[fierce hollow]

I'd assume the example comfy workflow for sdxl is what would pass as official, since sai and comfy seem to be an item now

[vocal stream]

you just need to use the api access instead of the site

[uncut steeple]

You can replace it with a normal image save node

[trim totem]

hey I'm working with a developer on image creation for fun, what do you guys use for this case? google colab? With colab the session interruption is super annoying any solve for that?

[tropic turret]

Okay

[clever verge]

Agree with this, there's no right or wrong within limits of course. 🙂

[delicate grotto]

Im using gpt4 for coding, but ill keep that in mind for math and phys related stuff
Early 2023 was the last time the model didnt screw up math badly

[fallow kestrel]

if you need an LLM for coding, check out "Starcoder"

it was trained on a lot more code afaik

[delicate grotto]

I think i know that one
Visual studio addon right?

[vocal stream]

for coding there's the alpha testing for the new gpt4 version of codex or whatever if you request access from github

but I never tried it after getting access

[fierce hollow]

you can also try phind though it's getting dumber lately somehow

[fallow kestrel]

I haven't tested it personally but thought it was another .safetensor model for the OogaBooga Webui

[unreal plover]

OVNI files unclassified

[fallow kestrel]

wow this is actually great 😂

[delicate grotto]

Mmmm yeah oobabooga is a no go
Ill try to load it in a collab ty

[west breach]

I tried the new chat copilot for vscode. I didn't find it very useful

[fallow kestrel]

the best copilot is still your brain as of now

[delicate grotto]

Gpt4 works great, but you need to be precise with what you need

[fierce hollow]

exllama has pretty minimal requirements and webui if you just want to run some llm locally btw, ooba feels kinda bloated

[wet merlin]

Question... is there going to be an official version of this? Is this a good way to fix it or is there a better way to use FP16? https://civitai.com/models/117188/sdxlfixedvaefp16

[fallow kestrel]

true, I think ooba is more for building upon and creating a front/backend for products

[west breach]

I mostly just use copilot to autocomplete my code. When I tried the new chat thing, I found a lot of the suggestions were no good

[delicate grotto]

3.5 screws up a lot

And i mean a LOT

Use 4

[fierce hollow]

if this is the same as the huggingface vae then it works well, I trained a lora with it and wasn't getting any nans

[vocal stream]

I'm not sure what the legality (both for my contract and stuff like gdpr) are with sharing everything open in my vscode or repos to an US company so I've never tried it

[wet merlin]

yes it's a merge based on this https://huggingface.co/madebyollin/sdxl-vae-fp16-fix

[delicate grotto]

Im doing a project in dataset management and training, gpt4 is priceless

Sorting it all, center cropping, finding corrupted pics that are most likely useless, binary corrupted img search

Dataset purity sorting
Everything

[delicate grotto]

In all honesty, without gpt4 it would be impossible

[vocal stream]

the few times I tried gpt4 for more complex stuff it helped but I still had to change it at the end + prompt it a bunch of times and at that point it didnt seem faster than me just reading online, trying and adjusting

[delicate grotto]

You need to make sure that your logic doesn't have a backdoor
Or it might givr you a statement that it will just autofill that backhole to close the function without warning

And make sure to not ask it to make anything longer than 2k tokens

[vocal stream]

I have access to 16k I think or maybe 8k

[delicate grotto]

So ~50 lines or less

How well does it remember at 8k 16k tokens?

I remember it losing context with more than 2k with its acvuracy

[vocal stream]

(just checked it's 16k) not sure haven't experimented recently

[delicate grotto]

Mmm i remember a graph comparing some models to gpt4 at longer context
It got worse with more tokens, but its better than anything else

There was a llm that could handle 10m or 2m tokens, but it had ~ 30 40% context accuracy

Relative to what they tested

[vocal stream]

I'd still rather get a finetuning API so I can effectively handle really long 'contexts' like we could do before

[delicate grotto]

Mmm im going to try to make one in august, one with memory instead of token knowledge

[vocal stream]

like with gpt2 I just trained it (for a month on free colab) on all my conversations and I'd want to recreate that for gpt-4 https://svilentodorov.xyz/blog/gpt-15b-chat-finetune/

[unreal plover]

Titans are woking up

[delicate grotto]

Mmm you're not starting to train it with hard material right?
You should train it from easy to hard

Much faster learning

[vocal stream]

eh, you do a bunch of epochs and at least then it's random sampling from the training data

given that you start from what it was trained on already, nto from scratch

from scratch it'll take forevr without a big budget I'd imagine

[delicate grotto]

Mmmm try to sort it a bit

Btw, mind sharing the notebook?

[vocal stream]

I've last done it with gpt2. You're much better using the llama 2 stuff if you are doing it today

[delicate grotto]

Got it

[shy kelp]

so far SDXL 1.0 looks about the same as SD v2.1 just with better PR and marketing
rememmer this is by the same guys that wanted to cancel Automatic1111 and now doing the same by designing the XL model in such a way that it works better on ComfyUI

[vocal stream]

ive never seen sd v2.1 handle text

[delicate grotto]

I used 2.1 since release up to few months ago, it is obsolete to 1.0

[vocal stream]

not to speak of 2.1 being way more lackluster at understanding prompts in general

[delicate grotto]

2.1 was bad at everything basically beside having an exteemely fast embedding training time

It needed sometimes ~600 prompts in the neg to get same img quality as none in sdxl

[vivid silo]

show me base 2.1 creating this masterpiece:

[unreal plover]

Someone have a good tutorial material to make a Lora with XL 1.0 ?

[delicate grotto]

Eyes, hands, anime, faces etc... obsolete relatively to 1.0

[delicate grotto]

I had some with my embeddings

Crystal embeddings, bird embedding and 3d can do that in 2.1, but with messed up eyes and feet

[vivid silo]

Yeah it's definitly possible but not just with the base model

[fierce hollow]

I think auto just wasn't very interested in the whole nda and stuff so he's lagging behind

[versed locust]

Can you use Roop on Auto with SDXL?

[delicate grotto]

There is a technique to do that with 2.1 base

Someone made a prompt that made gpt 3.5 do prompts based on your input

It was extremely precise

[lilac wren]

I'm really surprised how much effort people put into not sharing their Workflows. You take advantage of an open source work that took years to develop, you steal the prompts, the settings from everyone, to end up being selfish. It's profoundly lame as reasoning.

[delicate grotto]

Insane quality

[shy kelp]

I missed something during yesterdays live. they said something about sdxl now being more drag and drop ready. what was that about?

[unreal plover]

Did you know how are made Spaghetti ?

[vocal stream]

I'd imagine they meant it's more usable out of the box rather than with extra tools and finetunings

[shy kelp]

i think there was more to it than that but im not sure

would be really great if they could reupload

[fierce hollow]

the audio kept cutting out so not sure

they talked about some new ui thing, might have been that

[shy kelp]

wondered if they somehow merged the refiner into the base model

[fierce hollow]

pog

[lilac wren]

my GTA V workflow?

[woeful patio]

for the people that dont know about this website, it's awesome:
https://supagruen.github.io/StableDiffusion-CheatSheet/

seems to work better than with 1.5 too

[meager canopy]

Thanks for sharing 🙂

[lilac wren]

Sure, sharing is not really the watchword these days

[civic forge]

Im a n00b with comfyui, do you know how to use the sdxl1.0 vae on It?

[west breach]

Also this one is good: https://proximacentaurib.notion.site/e28a4f8d97724f14a784a538b8589e7d?v=ab624266c6a44413b42a6c57a41d828c

[civic forge]

I mean, I have the workspace for using SDXL 1.0 and the refiner but I don't see a way to add the VAE

[fierce hollow]

there's a vae loader node, you can plug it to encoder and decoder instead of the checkpoint vae

vaes should be placed in models/vae by default

[civic forge]

I'll try

[civic forge]

Yes, I did that, so the problem was I wasn't using the right module.

So, It would be like this:

[fierce hollow]

yep

[civic forge]

Thanks, It was pretty obvious once you told me, but I wasn't able to pull It by myself.

[clever sky]

can i install sdxl on i7 intel iris xe laptop 8gb ram?

[uncut steeple]

What GPU?

[clever sky]

I think it's integrated gpu, intel iris xe

[uncut steeple]

Then its a no

You need a beefy GPU

[clever sky]

Even if the generation of images is slow, it is not possible?

[lilac wren]

A good GPU, and 32 GB of ram is suggested

[tribal moat]

is someone facing problem of similar generation with small variations only? whenever I try to generate images by adding tag "Art by Yoji shinkawa", I always get similar images unless I change rest of the prompt but still, it mostly follows same vibes. Unlike in other models, every other Different seed always give different image.

[delicate grotto]

Cpu mode might do it, but you need a pruned model with 8gb ram

[fierce hollow]

you should be able to run it on cpu with enough ram but the speed will be abysmal, also fp16 will be upcasted on cpu iirc

so it will require more ram than gpu

[delicate grotto]

768 img should take you ~2 to 5 minutes with cpu

[civic forge]

comfyui runs on CPU but good luck with the speed, It's faster to draw it yourself in paint.

[uncut bison]

yeah, CPUs can't do FP16

[clever sky]

I have worked on it in google colab but it seems that I have to pay

[civic forge]

Yep, If you use too much colab they make you pay

[delicate grotto]

Collab gives you free use for a certain amount of CU, and it wont give you a gpu if they are all used

Buy cu, its not that expensive

But buy it with free use, not the one that disappears each month

[tribal moat]

just use different google account. lol

nowadays everyone got multiple account

[fierce hollow]

there's only so many google accounts you'll be able to make before they start asking for verifications and stuff

[tribal moat]

2 google account basically allows me to use google colab for whole day.

[clever sky]

It is the first time I have used colab I have not seen any option to use what you say for free

[tribal moat]

idk

I do alot of training and automation tasks using google colab

longest runtime was 10 hours.

[fierce hollow]

it depends on your overall usage and luck I guess

[tribal moat]

after that, I wasn't able to use it anymore, then I switched to another google account to continue train my dreambooth.

[tribal moat]

🤔

I'm saying switch to another google account...

[clever sky]

I don't know what you mean, how can I run sdxl in colab for free?

xddd

[tribal moat]

You can run Auto1111 with SDXL model

tho it's slow, better option is use ComfyUI but i don't know if it has colab version.

[clever sky]

what account?

[tribal moat]

https://cdn.discordapp.com/emojis/768544542633754635.webp?size=64

[clever sky]

I don't understand sorry

[trim kiln]

hello - when is v1 available on Beckrock please?

[tribal moat]

You said that google colab is no more giving you free access, right?

[unreal plover]

name this creature

[clever sky]

It is the first time I use it, I have tried to start this and I get an error. https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb#scrollTo=p4wj_txjP3TC

[tribal moat]

I almost thought of Night King...

[tribal moat]

what kind of error?

[clever sky]

in the last step it gave me an error

[tribal moat]

in last step?

[clever sky]

I'm going to run it again I don't remember anymore, I'll give you a screenshot wait

[clever sky]

yesss

[tribal moat]

Ngrok token...

register and get a token from here. https://ngrok.com/

If you don't provide token, there is no way to access hosted site by colab, Ngrok acts as tunnel between you and colab. because you cannot directly access colab's local web server.

[humble wigeon]

QUESTION: on my video card 1080 ti 11 gb, 1 image is generated for 1 minute sdxl 1024-1024. is it possible to speed it up somehow? what speed do you have on which video cards?

[tribal moat]

30 seconds for me on RTX 3060ti. 8gigs vram

[candid pond]

Hey fellas. I've been used SDXL on Auto1111 and I've seen that switching btw refiener and base takes hella long.

[tribal moat]

That is if I use comfyUI*

almost 8 minutes in auto1111.

[candid pond]

I have 3060 12GB and 32GB system ram.

[tribal moat]

[clever sky]

I already registered but I have no idea what that is

[tribal moat]

Are models stored in HDD?

[fierce hollow]

refiner by itself needs like 6gb additional vram so the unets are switching between cpu and gpu, hence the slowdown

[lilac wren]

Here, now I do like almost everyone here, I show the images of my good Workflow, hiding my EXIF ​​data so as not to share. It's fine not to share when working on open source.

[candid pond]

Yep. Should I move to SSD?

[tribal moat]

https://cdn.discordapp.com/emojis/663787183613673506.webp?size=64

[tribal moat]

ofcourse dude.

[humble wigeon]

I also use comfyUI. I have 50 seconds generated . it's a very long time. ((

[lilac wren]

?

[candid pond]

Oh my. I always installed Auto1111 on my D HDD.

Let me swap to SSD and test it that way.

[tribal moat]

It's 130 seconds if I also use refiner.

[fierce hollow]

hdd mostly affects the initial model loading speed to cpu ram tbh

like on cold load, but if you're keeping them cached it shouldn't be an issue

[tribal moat]

no, nothing, it was great image you generated there buddy. Legend of Zelda!

[shy kelp]

'A City build by pigeons for pigeons' lol

[candid pond]

In Auto1111 with how much denoise should I use the refiner?

[tribal moat]

yeah but It needs alot of ram to store both refiner and base. and it isn't the case for Auto1111, it keeps only one which you have currently selected.

[lilac wren]

Is that supposed to be sarcastic?

[fierce hollow]

you can change how many models are cached in ram in settings

[west breach]

halp! colonel sanders is haunting my house

[candid pond]

Oh, how do I change that?

[tribal moat]

Nah, It was an instant reaction for being so closed with workflow, but it's fine, no offense. Sorry if that offended you.

[candid pond]

I would like to have both refine and base cached if I can with 32Gigs ram

[fierce hollow]

just ctrl + f cache in the setttings tab, I haven't used auto for a while but it should be there

[shy kelp]

is it just me or is sdxl not actually that good

[soft zealot]

@visual glade (or indeed anyone else) is there a way of embedding URL links inside workflow ?

[candid pond]

Oh nice. Yeah that worked.

[unreal plover]

interesting fail

[candid pond]

Oh you wouldnt say that if you have seen the 1.5 base.

[upbeat summit]

very fluffy 🙂

[shy kelp]

I have seen 1.5 :D and yeah that's kinda a bit bad too, 2.1 is okayish

[lilac wren]

This is exactly where I want to go with my message, people here come to take everything, and do not share when they find something interesting. I remember a time when everyone worked to pull SD up, not to keep its secrets like a jealous bird incubating its eggs.
All the images in the chat have been removed from the EXIF ​​data, so as not to share... Well done guys, great mentality

[tribal moat]

dam, It looks super cool tho

[shy kelp]

hello everybody

[candid pond]

Hmm. The 1.0 Vae seems somewhat baked.

[shy kelp]

how is sdxl treating u guys?

Not well lol

[tribal moat]

chill buddy. many people copy image directly instead of uploading, Copied images are stripped of metadata.

[unreal plover]

what a time te be alive

[shy kelp]

I am playing around with how I prompt it though

[candid pond]

This is after refiner.

Either I am doing something wrong or there is something wrong with the vae.

[fierce hollow]

there is something wrong with the vae, try the fp16 fixed one from hf

[upbeat summit]

hmm

[candid pond]

Now I have a big question.

How much Vram do we need to fine-tune a LORA for this?

[fierce hollow]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors
this one doesn't have the weird horizontal artifacts and works with fp16 training

lora - around 10gb vram + 10gb ram

[upbeat summit]

yours looks way better - different style. but still similar in composition

[fierce hollow]

finetune - a fuckton

[candid pond]

10GB's is enough?

For LORA?

[tribal moat]

yup

[fierce hollow]

I'm using 9.5gb right now with 1024x1024 buckets, so yes, should be

kohya seems to have fixed some memory issues

[candid pond]

Damn thats really well optimized.

Can you send the parametres you're using?

I need to go and grab a pack of cigs and then I'll be right back.

[tribal moat]

[fierce hollow]

I'm not really doing it the script way but hold on

[uncut steeple]

Feel free to try my workflow

[tall lion]

It's free upto a point. I use Google Colab with SDXL 0.9 for my gens

[tribal moat]

hey buddy, you seem to have alot of knowledge about comfy, can you please tell me how can I interage things like BLIP or deepbooru in workflow?

[tall lion]

Use any ComfUI notebooks and you should use CPU and GPU for free to a certain extent. Use two accounts to use it for 10 hrs

[candid pond]

This looks pretty well: https://civitai.com/models/116197/jenna-ortega-sdxl

1.8GB's for LORA though

What network dimension did they use?

[fierce hollow]

vae = "madebyollin/sdxl-vae-fp16-fix"
network_module = "networks.lora"
network_dim = 64
network_alpha = 64
network_train_unet_only = True
mixed_precision = "fp16"
save_precision = "fp16"
caption_extension = ".txt"
full_fp16 = True
optimizer_type = "Adafactor"
lr_scheduler = "constant"
lr_warmup_steps = 0
learning_rate = 1e-4
min_bucket_reso = 512
resolution = 1024
max_bucket_reso = 1536
bucket_reso_steps = 64
max_train_epochs = 1
save_every_n_epochs = 1
train_batch_size = 1
gradient_accumulation_steps = 10
vae_batch_size = 1
enable_bucket = True
flip_aug = True
gradient_checkpointing = True
sdpa = True
cache_text_encoder_outputs = True
cache_latents = True
cache_latents_to_disk = True
noise_offset = 0.0357

these should be the ones that matter somewhat

[clever sky]

I have the token, what do I do with it?

[tall lion]

It seems that ComfyUI is the way to go for SDXL based on few posts I have seen in comparison to A111. ComfyUI on Google Colab takes about 125 seconds (after loading the models) to generate

[fierce hollow]

I'm not sure if there are any nodes for deepdanbooru or blip, you'd probably have to code it yourself

[tribal moat]

paste it in input box where it asks for ngrok token.

[tribal moat]

[tall lion]

I get around the same time indeed with COMFYUI ON colab.

[candid pond]

Thank you!

[fierce hollow]

comfy isn't really the best tool for tagging datasets imo

[tribal moat]

What about Lora training? are you using comfy ui for lora?

[fallow kestrel]

What's your preferred Sampler for SDXL?

[tribal moat]

nvm, you are using koyha.

[fierce hollow]

I'm using kohya's scripts for training but you can use the loras with comfy, yes

[tribal moat]

okhy, thanks for info.

[fallow kestrel]

Do I have permission to download this? It's the best thing I've seen in months

my depression is gone

[tribal moat]

nah dude, I don't use colab.

It's local for me.

[candid pond]

The first one is without the refiner and the seccond is with refiner.

The ees always have artifacts, what might be the cause of that?

[fallow kestrel]

colab is fine until you boot up a UI with it

[fallow kestrel]

what's your ratio of steps from Base to steps from Refiner?

[tribal moat]

It's same for me too

for now, I use base model only

plus it's faster to use single model.

[candid pond]

50 base, 10-25 refiner. Using 0.3 denoise.

[fallow kestrel]

Are you using the VAE "rasterized" image for the refiner or the latent image? I had much better results when handing over the latent

[candid pond]

I have no idea how I can send the image without VAE.

[fallow kestrel]

"Where do you see yourself in 30 years"

[candid pond]

How can I send the latent image?

[tribal moat]

If Auto1111 then you are using Img2Img for refiner meaning rasterization. simply put you have decoded latent into image. or else it's not possible to use Img2Img.

[fallow kestrel]

The red line 😅

[candid pond]

Oi, I'm using Auto1111.

[fallow kestrel]

Ah, sorry

[clever sky]

what should i do now?

[candid pond]

Guess I'll have to wait until they incorparate a decent workflow.

[tribal moat]

you got the link buddy

delete it before anyone else abuse it.

[fallow kestrel]

Probably. I was 100% for auto1111 aswell but switched yesterday to Comfy, I don't regret it

[candid pond]

Cause this loading base and then refiner is not really optimal.

[tribal moat]

Delete the image*

[fallow kestrel]

risky frisky

[jolly creek]

Living on the edge

[tribal moat]

blue text is link. click on one which says ngrok as domain.

[clever sky]

it is bad? what's happening?

[tribal moat]

nah

it's success

Open the provided link and get started with image generations.

[lilac raven]

Is there a way to zoom in/out ComfyUI on Android? It always opens the menu to add a node

[zinc cargo]

so glad that comfy is taking off right now

[candid pond]

Does Comfy have inpainting support?

[clever sky]

I don't see any link

[fallow kestrel]

Node Based Workflows on android? I'd recommend just not doing that, sorry if that's not very helpful 😅

[fallow kestrel]

I think so, haven't tried it though. Would interest me too - has anyone tried it?

[tribal moat]

can you dm me the image? I'll point it out for you.

[clever sky]

yes thank you

[fierce hollow]

you can do inpainting in comfy but there's no way to use a brush inside the ui like with auto, you'd have to make the mask yourself in photoshop or something and then load it (can be a separate mask or alpha channel)

[fallow kestrel]

Must be rough for poor automatic and what he went through with all the leaks and drama involved, but as a consumer, Comfy is fantastic. I'm used to Nuke and Davinci Resolve (VFX editor by day) so node-based workflows are much preferred imo

[fierce hollow]

there is a clipseg node that kind of works for inpainting, you just specify what in the image you want to edit and it makes a mask for you

[fallow kestrel]

Thanks for the info - this out-of-the-box painting tool in automatic always blew my mind, shame it's not in comfy. But oh well, could probably code a bridge between Gimp and Comfy or smth

[fallow kestrel]

Oh! Like the SEG-controlnet?

[fierce hollow]

yeah

[west breach]

You can use the load image node for inpainting. right click on the picture > open in mask editor

[fallow kestrel]

that's cool!

[fierce hollow]

https://github.com/biegert/ComfyUI-CLIPSeg

[fallow kestrel]

Oh! Wait that's brilliant

[fierce hollow]

also preprocessors https://github.com/Fannovel16/comfy_controlnet_preprocessors can do segmentation as well

oh there's a mask editor now?

[fallow kestrel]

No way! That already works with SDXL?

[tired yew]

SDXL1.0 Base+SDXL1.0 Refiner. I think i fell in love with SDXL1.0!

[west breach]

always been there

[fierce hollow]

mind blown

[fierce hollow]

controlnets don't work with sdxl but preprocessors only rely on pixel space images so there's no problem using them for various stuff

[shy kelp]

Did they filter out nudity in sdxl training lol

[fallow kestrel]

well yeah it's not a problem to generate a depth mask, I'd rather have a way to let the depth influence my result in the way controlnets worked with a1111 afaik

[fierce hollow]

that's probably going to take a while, I reckon illyasviel would have to train new models

depth maps are still pretty useful on their own

[sour obsidian]

we already have some cooking 🍳 just not quite ready

[fierce hollow]

oh that's nice to hear

[west breach]

This is correct.

[lean granite]

love the colors on this!

also my gpu seems to go up to 106c whenever I generate a picture x)

[west breach]

Icey, Icey

[shy kelp]

Well anyway even if SDXL is harder to prompt right now, it does look promising for the future. I can totally see how this is a step towards getting really good high quality images out of image gen AI

It does what it does really well

[tired yew]

starting to get messy on full body SDXL1.0 Base+SDXL1.0 Refiner.

[west breach]

new dekstop background

[delicate grotto]

when you put 1 too much 0's into cfg

[shy kelp]

why the size difference?

[pallid sail]

I've noticed that generating images with no prompts in SDXL results in more coherent outputs than SD 1.5's nightmare fuel

[zinc cargo]

it was cold outside...

(sorry)

probably pruned, right?

[magic dawn]

No, it's reduced precision, fp16 vs fp32

[zinc cargo]

ya'all running comfy?

[heady vale]

yep

[magic dawn]

comfy is comfy

[heady vale]

A1111 is just too slow

[prisma adder]

anyone on AMD GPU ?

[strong field]

been trying comfy with 6950 xt and i have no luck over 512 generations

[pallid sail]

About how long does a generation with 50 iterations split across base and refiner take in comfy?

[prisma adder]

Someone told me about SD.NEXT which is supposed to make AMD GPU with windows work

[tribal moat]

I heard AMD is best to work on Linux instead of windows....idk

[prisma adder]

haven't had the time to look at it yet

[heady vale]

Im using 60 steps total on a 3080. takes just under 13 seconds

[strong field]

yeah i read that PyTorch with ROCM might help but havent bothered testing

[jolly creek]

A1111, haven't tried comfyui yet but thinking about it

[trim totem]

when will A1111 be ready for sdxl 1.0?

looking for a colab version

[jolly creek]

Base model is supported but refiner usage is not implemented yet

[jolly creek]

You have 8gb or 10gb version?

[little gazelle]

Has anyone had any luck finetuning your own style with the new SDXL model?

[heady vale]

10

[west breach]

lora, yes

[jolly creek]

I see, I have 8gb

[little gazelle]

Is that using Kohya ss?

[west breach]

yeah

[clever verge]

It doesn't realy follow your prompt. 😄

[pallid sail]

Is that at 768x768? I'm only getting 1.2 it/s and it took me 2m 18s to make this on my 3080 😢. I'm running my own implementation in Unreal so I have a lot of optimization to do

[clever verge]

SDXL will do nude if you prompt tribes it seems

[heady vale]

full 1024 sq

2.68it/s

[copper linden]

Why is tiling not working in SDXL?

[tired yew]

good idea, ill try without tribes promt 🙂

[tall lion]

pruned versions I would surmise

[fierce hollow]

keep in mind that's probably only the generation part, with 10gb vram you'd have to offload to cpu between base and refiner and also use tiled vae decoding, all of that adds a lot of time

[cunning geyser]

Training Precision = bf16 <- fp16 can and will cause NAN errors in many situations
Clip Skip = 1 <- no novelai in sdxl
resolution = 1024x1024
Gradient = true by default
Cache Latents + to disk on by default
Net Dim = 8
Alpha = 1
Train on = Unet only <- those who dont know the difference, should REALLY use unet only for now
Learning Rate = 1e-3
Optimizer = AdamW8bit
LR Scheduler = constant with warmup
Warmup Ratio = 0,05
Save Precisioun = bf16
bucket res = 512 x 2048


offset noise = 0 (cause then it auto sets it to the correct value)

Notes:
• good captioning is still relevant
• Dataset size of 30~60 to avoid typical issues. 10 or more will work fine if you know what you're doing. under 10 works if you really know what you're doing.
• set repeat to 1, do 40 epochs. 20 epochs will prob be your 'ideal' lora. so save every few epochs. 200 for super complex, multiple concept/object loras.
•image count x repeats x epochs = steps``` This is still good for lora training right

[fierce hollow]

also I'm getting around 1.3it/s at 1024x1024 with comfy and a 3060 12gb for the record

[prime pulsar]

Hi, I need some guidance on how to preserve a white and transparent background persistently. Occasionally, white backgrounds are successfully generated, but mostly the generated color appears to be a shade of grey instead or a totally different background. I've noticed an inconsistency in the results, although I've been using the words 'white' and 'transparent' in my prompts quite often.

[pallid sail]

Yeah, I swap models as needed which adds a few seconds delay in and I have attention slicing, cpu model offload and vae tiling enabled - I use Diffusers as my backend

[fierce hollow]

you can use fp16 with the fixed vae, in fact you should use full fp16 training, it will reduce the vram usage a lot, clip skip is unnecesary, other than that it's all as normal, kohya claims adafactor works a lot better for sdxl, offset is 0.0357 by default for sdxl iirc

[cunning geyser]

...Ahuh

[fierce hollow]

this vae solves nans in training with fp16

[uncut steeple]

[fierce hollow]

that sounds about normal then, diffusers implementation is probably a bit slower than comfy

[pallid sail]

Ah cool, good to know. Unreal is taking up a few gb VRam as it idles but I disable realtime rendering in the viewport when generating an image which doubles the generation speed

[static prawn]

im kinda sad sdxl is so not really working in auto for me, hope they improve the performance. on Comfy its just fine

[light stag]

When will sd beat midjourney

[strong field]

already does

[sullen glade]

is there a workflow update? or which workflow are you using in comfyui? mind sharing?

[light stag]

Is this better than mj v5

[light stag]

How

[sullen glade]

the answer is: depends. But in some areas it surely outperforms MJ whereas in others it falls behind

[static prawn]

nono im just using a normal workflow, just comparing comfy with A1111

i cant use sdxl on Auto bec. of the performance

[sullen glade]

if you want creative output that does not do exactly what the prompt says MJ does a better job at suprising you, but if you talk about doing exactly what your prompt says SD outperforms MJ by quite some ammount

question though: Does someone know how the "style" works. Is it just a text addition to the prompt, a lora or a specialized model of the base?

[static prawn]

a picture in comfy takes max. 2 minutes, when i generate in Auto 1111 it takes 20 minutes or more

[light stag]

Nono im saying the quality Which is better

[soft zealot]

WAS NOdes Suite ahs a Image Remove Backgtround Node.
You could further refine by then adding/merging theis to a plain coloured background

[sullen glade]

depends.

[static prawn]

Auto1111's performance never was really good, especially compared to diffusers, but with SDXL it got unusable for me 😄

[jolly creek]

You mean in A1111?

[cinder aspen]

SD, MJ and Firefly are still having different target audience pretty much

[sullen glade]

no, on the website and the bot

[jolly creek]

#✨｜sdxl message
As far as I know they use these, could be wrong tho

[unreal plover]

I

[dusk mica]

how do people get it follow the prompt so percisely

[jolly creek]

I've been using them and style is implemented quite accurate

[little gazelle]

Any tips for using the Kohya ss for SDXL on Google Colab?

[unreal plover]

I just have "drawing and 3d" on negative prompt

[little gazelle]

Keep running out of memory when I start training

[cedar bison]

Anyone know how I can run SDXL models in automatic 1111 I've updated it but I was told SDXL works by loading two models or something ?

[jolly creek]

In A1111 right now simply use base model

[shy kelp]

did they merge them or does it simply not support refiner yet?

[cedar bison]

do you have qauility loss doing that ?

[jolly creek]

Doesn't support refiner yet. Technically you can use refiner for img2img but that is not really the same

[tall lion]

not noticeable

[jolly creek]

Depends on picture you prompt

[shy kelp]

yeah i know 😅

[jolly creek]

Some get more impact than others

[tall lion]

if you want to use Base+ Refiner + Upscaling workflows with SDXL please use comfyUI. Much better control

[uncut steeple]

Yeah comfyUI is the go to right now

[cedar bison]

any links to how to set that up ?

[indigo vine]

and if you have an working auto1111, you can reuse models+venv in ComfyUI so the switch is quick and reversible

[uncut steeple]

What a mess

[jolly creek]

Cool, easier will be for me to try out

[uncut steeple]

You can also drag and drop the images made with comfy to get the workflow used

[cedar bison]

I'm seeing on Civitai a few of the people putting down SDXL models now don't seem to have refiners

[dense chasm]

4x-ultrasharp is better

[tall lion]

Check out this video. This has good directions for Comfy on Google Colab and also locally on your computer (Assuming you havea a powerful GPU). - https://youtu.be/FnMHbhvWUhE

[sullen glade]

brilliant! thx a lot

[soft zealot]

oh the old "Powerful GPU" sketch again

What exactly is a "powerful GPU" ?

I quite happily use a 6 year old 1080Ti

You do not need a latest generation card.

Sure its nice to have but you don't "need" one especially if you are just tinkering for fun

[cunning geyser]

Ok does anyone have a config for sdxl loras

Cause this is getting real questionable how people are even making one that's 49 mb

Cause I can get 90mb but that's as close as I get

And most of the answers I get are fuzzy or do not line up well.

[trim orbit]

i would definately still call a 1080ti powerful. it was flagship in it's day

[soft zealot]

its just a phrase that grabs my Goat

When I started I was quite happi8lyusing a 980Ti, I only upgraded for the extra 5Gb VRAM

[trim orbit]

oof. wasn't the 980ti the one that split the vram so it had a fake 8gb?

[soft zealot]

no , it had 6Gb VRAM OOTB

[trim orbit]

i was an amd boy back then because i was only playing video games and was happy to compromise on the price

[trim orbit]

i was thinking of the 970

[fierce hollow]

8 dim 8 alpha should be around 50mb I think

the other params don't really matter much size-wise

conv-dim too I guess

[soft zealot]

which was a 4Gb card but yes I do seem to recall something about some mfrs doing something dogy

[cunning geyser]

Question

So if I use an anime character for a sdxl lora

Given sdxl's limited dataset due to photorealism, I shouldn't expect it to pull it off until we get an anime model by someone right?

[fierce hollow]

pretty much, yeah, your mileage may vary

[cunning geyser]

Yeah cause its a 4 arm character

[trim orbit]

what do you mean limited dataset?

[fierce hollow]

you can totally train a decent Ranni I think, she'd just be less anime-ish

[cunning geyser]

Because you can't do anime that well with SDXL

Also meant a char like this

Purple arms, 4 arms, etc

It can't do it, and the lora doesn't DO IT

[trim orbit]

sdxl isn't exclusively photorealism

[cunning geyser]

It's HEAVILY though

I say that becuase 90% of posts are people doing realistic characters/creations vs anime type stuff here.

And the anime ones are extremely mediocre in comparison to sd 1.5

[fierce hollow]

I don't think it was trained on 6 b m booru images like nai so it's not that great at anime tbh

[trim orbit]

i don't know what that means? people are mostly generating photos with it yeah. anime only gets so good. 1.5 exists for that. but here, simple prompt Anime robot in the style of studio ghibli, ponyo

[cunning geyser]

Yeah

[cunning geyser]

That's why I say should I wait til someone makes an anime model

[trim orbit]

sdxl is an anime model. it has crazy anime capabilities. comics too.

[cunning geyser]

Aight let me try that rq

[fast vector]

I mean all 1.5 anime checkpoints have their parentage in NAI and not default 1.5. SDXL has better anime outputs than default 1.5 tho

[radiant grail]

https://civitai.com/models/117259/anime-art-diffusion-xl

[vocal citrus]

ooo, thx for the link

[quartz sequoia]

@visual glade i saw you mentioned yesterday that SDXL TensorRT works right?

does it have any limitations compared to a normal SDXL?

[radiant grail]

No problem

[rustic surge]

What should I do if I use SDXL in auto1111 and it works badly?

[tall fog]

How do you set Train on = Unet only in kohya? Just set TE LR = 0? Or a flag?

[cunning geyser]

Idfk i was just told the settings the other day

I haven't tested them entirely

[fast vector]

Comfy

[cunning geyser]

Im letting everyone else crash and burn to learn then asking to get the easy way out

ok so yeah they were right

[rustic surge]

Need a base model + a refining model?

[cunning geyser]

i need to reference an anime artist lmao

[echo pine]

I'm adjusting your workflow a bit with sytan's workflow I'll send an image after I test a bit

[uncut steeple]

Let me know what you come up with

[fast vector]

Only base required but have both. Depends on the workflow

[vocal citrus]

wtb memory upgrade for my 4080 so i can train stuff lol

[trim orbit]

not trying to be right. just trying to dispell disinformation

[uncut steeple]

Its called 4090

[cunning geyser]

Its prolly my lora based on these circumstances

[vocal citrus]

@uncut steeple

[cunning geyser]

Or maybe its just the anime dataset is too honed to avoid customization of this calibre

[trim orbit]

there are a lot of rumors being pushed that suggest sdxl needs a massive amount of training to be capable of anythhing. it's not a photo model. it's a foundational model.

[cunning geyser]

4 arms do be weird

[uncut steeple]

I know you dont wanna hear it, but I have 2

[cunning geyser]

Yeah

I expected that, so I'll wait til custom models drop more

[wet nacelle]

https://tenor.com/view/jerma-jerma-sus-jerma985-sus-jerma-jeremy-gif-25376340

[fast vector]

"Oh it did. I think it took about 3.5 hours to train this for 2 epochs on my 3090. I've trained my last few 1.5 LoRAs at 1024 this took substantially longer."

[wet nacelle]

https://tenor.com/view/jerma985-jerma-burger-meat-grinder-gif-25978363

I'll leave

[trim orbit]

i've been making loras with my 4080

batches of 2

[delicate grotto]

not perfect, but it is still quite good, some fine tuning and it would be a lot better

[kind pewter]

the texture of the skirt 😍

[delicate grotto]

im sure civati ppz are on it

[cunning geyser]

I'm aware

I have the entire collection of civitai on a document.
Models and loras

[uncut steeple]

She looks like a stickbug

[cunning geyser]

I'm the guy. The one who decided to be a masochist listing this all

[delicate grotto]

ah

umm

[cunning geyser]

So im constantly updating to hope some loras come out too

[fast vector]

https://huggingface.co/thehive/petrichor-SDXL-Finetuned-Fp16

Doesn't seem to be released. But here sdxl anime finetuned model

[cunning geyser]

I aint donwloading 1gb+ of a lora tho

[orchid oar]

Why does the bot make good looking anime pictures but locally it looks bad with the same prompt. How do i add anime as style like the bot has?

[trim orbit]

its released just epnding review of application

[vocal citrus]

@trim orbit ah, ill look into that

[winged tree]

try adding "Detailed, hq, high resolution, " etc... any time i add anime to a prompt it just goes full 90's anime

[delicate grotto]

the L prompt, like in here should do it - wasn't made by me, but a remake of my ver

or what i have here(my workflow, made by me)

[orchid oar]

Already tried that, still doesnt look like the bot, but thanks

[orchid oar]

These look amazing, i will try that

Thanks

[trim orbit]

there's a prompt node that has a "style" option, much like dreamstudio and the other bot services offer. i haven't found an actual release of that node though. Joe has teased it in some images. i think you'd want something like that to invoke the internal "style" or however they're calling those presets

#✨｜sdxl message

[jolly creek]

Try this:

Style: Anime
Positive: anime artwork {prompt} . anime style, key visual, vibrant, studio anime, highly detailed
Negative: photo, deformed, black and white, realism, disfigured, low contrast

[orchid oar]

Thanks, i will try using comfy ui and look for something like that node.

[orchid oar]

Where do i put the "Style:Anime"

[jolly creek]

Nah ignore that

I just copied from file

[orchid oar]

Okay thanks, will try that

[jolly creek]

Of various styles

[orchid oar]

Thx

[tall lion]

Anime using SDXL base+refiner

[orchid oar]

Damn, that looks nice

Do i need to use Comfy for using the Refiner?

[vocal stream]

the people (especially legs) are really messed up

[wet merlin]

the metadata seems to be lost on discord, is there a way to get it or see your prompt? Or would you be able to share it?

[orchid oar]

Put it in Png Info

[little gazelle]

How would I use a LoRa.safetensor with SDXL? Using diffusers

[orchid oar]

I can see the Prompt

[soft zealot]

I should retry this in SDXL1 (this was 0.9)

[soft zealot]

or if it was generated in COmfy open in Browser, savem dragndrop into COmfy

[old mulch]

Comfy is so ugly and confusing. Do we know when a1111 comes out

With full support

[cunning geyser]

I am getting comfy rn for this

I cant do this with automatic1111

[orchid oar]

Tried that locally and as with the bot. The bot (right) still looks way better somehow. Installing Comfy now

[cunning geyser]

yeah ^

Thats what im saying

[old mulch]

Way better lol

[candid pond]

I'll atttempt to fine-tune a LORA with Samara Weaving now.

[soft zealot]

its i s alearning curve but I've settled on a Daily layoutthat works for me and is a tadge more pleaseing onthe eye (plus its quicker & more memory efficient than A1111 for many people)

[fast vector]

Bot>>>SDXL

[trim orbit]

that's wicked good. 0.9 in some ways makes better pics since they structured the release of 1.0 for better training. my loras i made in 0.9 even affect 1.0 way better

[cunning geyser]

If I didn't fight god every day, that layout would be terrifying.

[dusk mica]

The bot sdxl looks so much better idk why

[wet nacelle]

https://tenor.com/view/autism-gif-9632833

[cunning geyser]

That's offensive to autistic people.

[soft zealot]

really ?

Quite structured to my Eye

You have your main parameters on the LH side, the main image output in the centre and a HiResFix thing (with options) on the RH side

[orchid oar]

Would be cool if the showed us how do use "style" locally

[candid pond]

Here is the weaving dataset if anyone wants to try aswell.

[wet nacelle]

oops. I am one buck.

[cunning geyser]

Yeah don't group them in with that idiot

[cunning geyser]

[soft zealot]

I like how an autism meter gets called out but pets on a plate don't

[kind pewter]

that's an angry cloud!

[cunning geyser]

Nah he was insulting autistic people.

They don't want him

[wet nacelle]

https://tenor.com/view/jerma-discord-jermacord-jerma985-twitchtv-gif-26246559

[cunning geyser]

[old mulch]

Funny thing is i am autistic, so he was correct

[orchid oar]

Damn

[wet nacelle]

zamn

[orchid oar]

Did you use Comfy?

[soft zealot]

yup

[wet nacelle]

Are you using multi prompts?

[soft zealot]

yup

[tall lion]

Studio Ghibli, Anime Key Visual, by Makoto Shinkai, Deep Color, Intricate, 8k resolution concept art, Natural Lighting, Beautiful Composition
, artistic, creative, contrasting, detailed, 8k, expressive

[orchid oar]

Thx, will try that now too

[trim orbit]

the bot is sdxl but with a lot of the settings randomly changed for you to a lot of best case kind of settings.

[tall lion]

Empty class room anime

[tired yew]

nice

[glad fulcrum]

does the supporting terms make any difference than putting it on the linguistic positive? or is it just for workflow?

[wet nacelle]

Is that an extension?

How the fuck do you have that in stable?

[soft zealot]

is what an extension?

[trim orbit]

comfyui is engineered diferently than gradio UIs

[wet nacelle]

multiprompting was introduced in MidJOurney I thought.

Stable has never had that.

[fast vector]

has anyone tried something like xyz grid in ComfyUi to check samplers?

[trim orbit]

has it now. SDXL has two text encoders

[wet nacelle]

I still can't understand how that can be.

It just has multi prompting?

oh shjiot

[soft zealot]

My Prompt Input Area

[wet nacelle]

shiot

shit

interesting

[fast vector]

random?

[uncut steeple]

[trim orbit]

one clip layer takes a linguistic style prompt. so word salads work less good there, and the other is classic vit clip and works just like the old prompt styles

[soft zealot]

yup, its a WAS Suite Node, pulls a random prompt via API from Lexica

[trim orbit]

i've been manually using one button prompt extension in webui, to create a bunch of randomized prompts then feed them through comfy ui

[cunning geyser]

Question

How do i get someones thing

[soft zealot]

I tehn run everything through some text manipulation steps to feed to the correct place and also to output the prompt used to a txt file (still have some work to do on that)

[soft zealot]

take them round the back of the bike shed and whisper sweet nothings to them

[cunning geyser]

No i mean

how do i get the nodes of a picture on civitai

it has 29 nodes but idfk how to copy it to comfyui

[delicate grotto]

How did you make them do a sharp turn?

[wet nacelle]

metadata should be in the imaghe

[cunning geyser]

nvm got it..

[orchid oar]

same prompts but higher res (1024x1024) on 3070. Much better

[cunning geyser]

i copied the nodes

[wet nacelle]

just drag drop to see if that works

[cunning geyser]

pasted it in comfyui

gto this

[wet nacelle]

drag drop the image

[cunning geyser]

nah i copy paste the node thing

[soft zealot]

Link Tidier. Save in \ComfyUI\web\extensions (0=straight 1= slight curves 3 = Standard) https://drive.google.com/file/d/11pAM5HW3S72DQv4Qw6cXQXTwCrLFlN2o/view?usp=sharing

[cunning geyser]

LMAO

[wet nacelle]

.json?

https://tenor.com/view/heavy-rain-shaun-gaming-gif-21533780

[kind pewter]

I can't make a planet shaped like a spoon :(
no matter if I phrase it or make a word soup

[worthy bear]

Hi everyone, I downloaded the new SDXL base, refiner and lora. It is an available option in the checkpoint drop down menu, however I cannot seem to select it. It just loads for a while then reverts back to the originally selected model. Any idea why this might be happening?

[glad fulcrum]

terminal messages?

[wet nacelle]

I feel you.

[delicate grotto]

Ooh, ty

[trim orbit]

it tries then fails and reverts back. could be a few things. wrong yaml file. need to update the ui. or you need some kind of launch option enabled.

[dusk mica]

how do you guys get bot level images? whats hte secret?

[trim orbit]

good prompts

[wet nacelle]

Can you give examples of what you mean?

I'm using the SDXL 1.0 model which by default outputs great.

[uncut steeple]

The right setup

[delicate grotto]

Seems like it fails due to size changes, can you show us your workflow?

Ah wait, that is au1111?

[fast vector]

Honestly, comfy is messing my brain. By the way has anyone succefully implemented this type of workflow:

1.5 Fine Tuned Model -> Highres fix -> SDXL Base+ Refiner and vice versa

[placid geode]

How do you get sdxl image portraits, which are NOT out of focus ?
I tried everything i could think of :
Neg Prompt: blur, blurred, out of focus, bokeh etc.
Pos Prompt : different cameras, tele lenses, sharp, crisp, f-stop, etc.

I cranked up the weights, which somehow sometimes had the opposite effect.
Different Samplers, cfgs scales and steps.

But to no avail, only some parts of the face are in focus the rest is not.
I generated 400 already and none were in focus.

[visual glade]

no loras and every checkpoint needs to be baked in

[glad fulcrum]

have you updated auto1111 today?

[cunning geyser]

question how do u use that entire UI with sanity

[worthy bear]

sorry im quite new to this. how can I show you my workflow?

[worthy bear]

no actually, havent updated for a while

[trail obsidian]

does anyone use comfyUI? i dont find DPM++ 2M Karras sampler? Or what is it called in ComfyUI?

[glad fulcrum]

that's your problem right there.

[delicate grotto]

Seems like you are using au1111, not comfy so yeah you cant

Go to tech-support channel

And also do a git pull

[unreal plover]

guys what is the use case for refiner ?
Creating an image at low res then refiner with higher scale with 0.5 denoising strenght ?

[delicate grotto]

Native au1111 cannot use sdxl as before of ~2 weeks ago

From 2 weeks ago to date i have no clue

[worthy bear]

Seems to be fixed now! Thanks

[visual glade]

pick dpmpp_2m as sampler karras as schedules

[wet nacelle]

I gotta say that 1.0 looks great but can also look awful what there is a lot of visual clutter.

It well just mesh smaller details together even when refined.

It's a great model and I can see it working much better when the community makes checkpoints from it.

[cunning geyser]

i tried copying your layout, most of its red?

[flat ibex]

Does anyone know when dreamboothing will be supported for SDXL? I tried both the A1111 extension and the Joe Penna dreambooth repo to no success. I know it's brand new and devs are still working on things, just curious if anyone has had any luck 🙂

[trail obsidian]

ahhh i see thats explains it ty mate!

[trail obsidian]

thanks!

[soft zealot]

check all of the below

---

Credit & Notes:

modified using original from https://github.com/SytanSD/Sytan-SDXL-ComfyUI

HRF modified from https://civitai.com/models/107144/sdxl-09-semi-official-workflow

Seed with Text from https://gist.github.com/alkemann/7361b8eb966f29c8238fd323409efb68

Multiple Nodes from WAS Suite, Efficiency nodes & Deerfu ::
https://github.com/WASasquatch/was-node-suite-comfyui

https://github.com/LucianoCirino/efficiency-nodes-comfyui

https://github.com/Derfuu/Derfuu_ComfyUI_ModdedNodes

styles sheet with SDXL Styles + others https://drive.google.com/file/d/1IZq_0CGTbfxlAdIMsjz3VwRzHjl3JL-n/view?usp=sharing (config in WAS Node Suite fpr location)

you need to add the Path to the styles2.csv (or whatever you decide to call it) in "\ComfyUI_windows_portable\ComfyUI\custom_nodes\was-node-suite-comfyui\was_suite_config.json"

So in my case its

"webui_styles": "Z:\AI\ComfyUI_windows_portable\ComfyUI\styles2.csv"

SO you can save it wherever you like

Link Tidier. Save in \ComfyUI\web\extensions (0=straight 1= slight curves 3 = Standard) https://drive.google.com/file/d/11pAM5HW3S72DQv4Qw6cXQXTwCrLFlN2o/view?usp=sharing

Hide

[trail obsidian]

Anyone have a good workflow for comfyUI for the new alpha dreamshaper model?

[haughty iris]

Thanks!

[glad fulcrum]

will custom checkpoints be lighter in terms of size? Like pruned 1.5 ones? Using refiner in auto1111 is too slow workflow. it has to load the model every time you change from base to refiner.

[scarlet pine]

My A1111 gens take around 8 minutes on my 3080. Using ComfyUI, they are taking roughly 15 seconds. Do I just have some setting wrong with A1111 or is it just not working that well currently.

[haughty iris]

A1111 isn't well setup for it right now

[clever verge]

try this prompt: positive: iphone photo {prompt} . large depth of field, deep depth of field, highly detailed negative: drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly, shallow depth of field, bokeh

[dusk mica]

A1111 needs tons of optimisations

[scarlet pine]

Just making sure I wasn't missing some crucial setting

[dusk mica]

what does . do in a prompt?

[trim orbit]

a1 was cobbled together quickly at the beginning but i think sdxl proves that it has outgrown it's uses. The community will learn better workflows as we grow.

[trail obsidian]

now we only need controlnet and animatediffusion for ComfyUI

btw does the dreamshaper alpha model need a refiner?

[reef sky]

it doesn't, in my experience

[haughty iris]

If you ask me, now is a good time to try experimenting with getting used to ComfyUI, if you haven't already. There's more reason than usual with A1111 being behind and it'll benefit you in the long run!

[heady vale]

can comfyUI do outpainting?

[trail obsidian]

ok so just upscale after basemodel and thats it so cool

[clever verge]

You can also try the prompts in the sticky. #✨｜sdxl message

[trail obsidian]

i really like nodes if i can do controlnet animatediffusion and roop eventually im happy and stay its sooo much more performant

[placid geode]

Thank you for the reply, i´ll give it a whirl 🙂

[lyric cedar]

invokeAI's latest build supports SDXL and is looking real sleek, I might consider it an alternative to both auto and comfy

[quartz sequoia]

Well that’s not nice

[haughty iris]

Personally like the control Comfy gives me, but I do want to ask, how does it compare performance wise?

[clever verge]

Not sure if it does anything

[placid geode]

Thats even better, awesome, thx 🙂

[haughty iris]

doubt it has too much effect, if any but if it does, you'd probably have to dive into the actual weights to find out, It's probably really good for artificially making other tokens less connected by putting a few to a ton between them

CLIP is wild on on the inside, lol

[dusk mica]

i heard people saying CLIP is like magic

[trim orbit]

https://github.com/twri/sdxl_prompt_styler/blob/main/sdxl_styles.json this looks good for prompting

[vocal stream]

i wish stuff like that would be baked into the metadata of the model and that UIs like comfy and a111 would just read it from the metadatata and make dropdowns based on them

[scarlet pine]

What json setup would you recommend

[vocal cosmos]

wheres the download for sdxl 1.0?

[trim orbit]

i dont have a recommendation yet. i'm still figuring out what works good for me

[haughty iris]

If you mean SDXL, here: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 and here: https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0
It's in the pinned

[clever verge]

Those are just a few examples that work. You can come up with your own.

[vocal cosmos]

ah thanks thats the last thing i was thinking whew

shouldve known its pinned

[vocal stream]

they aren't just examples that work though, some of them are based on how the model was trained and not random

[trim orbit]

oh hey that's twri's script! this guy knows what he's talking about

[haughty iris]

Just glad to save you from tearing your hair out looking for it!

[echo pine]

@uncut steeple try this

[clever verge]

I made it, and added a few prompts other than the sai ones.

[trim orbit]

its very nice

[vale eagle]

Offset lora: 0->0.5->1

[vocal cosmos]

oh do i have t make an account in hugging face to dload?

[clever verge]

The most incredible thing is SDXL, it responds very well to basic styling.

[uncut steeple]

I will once I got the missing nodes

[clever verge]

Thanks, happy you like it.

[haughty iris]

I think if you just git clone it should let you sidestep that

[wet nacelle]

hank

[echo pine]

ComfyUI_experiments/sampler_tonemap
ComfyUI_experiments/sampler_rescalecfg

[woeful patio]

Is it me or why is text i put in the prompt coming back in the pictures it generates?

[echo pine]

i think thats it

[azure oxide]

ngl im a human and i wouldnt know what to even attempt to draw if you asked me to draw a planet shaped like a spoon lol

[trim orbit]

haha yeah it happens now. SDXL sometimes tries to make text

[wet nacelle]

a decal is how I thought I could trick it.

[azure oxide]

@wet nacelle wait im guessing you got it to work??

[wet nacelle]

fuck no

hell no

[azure oxide]

unfortunate

[wet nacelle]

I thought

[azure oxide]

someone else also had issues last night, i dont think they resolved it either lol

[rich coral]

Yah all those guts hanging out, shudder

[trim orbit]

closest i've gotten

[uncut steeple]

Looks really similar to mine, any real improvements?

[kind pewter]

wow not bad

spoon-planet merger

[vocal citrus]

oh cool, i can train a sdxl lora with 1.2gb of vram to spare with 4080

[echo pine]

yeah, it's almost the same, just some stuff so you can use a higher cfg scale without burning the image and I added A score stuff from sytan
basically you'll see less burn artifacts and the A score makes the images better i think

[floral island]

any way to know which selection was taken from dynamic/wildcard prompt in comfy, or is it impossible to see?

[azure oxide]

i was wondering this myself. seems to be no but i would be completely happy to be proven wrong

[trim orbit]

kind of a neat result too

oh oh dialing it in closer now

[kind pewter]

damn this one looks cool as fuck

[azure oxide]

cosmic god cereal

[polar tiger]

The styles used in the bot channel - are these natively embedded in the new SDXL model, or are they just adding "extra" prompts to my prompt when I choose them?

[trim orbit]

more sortof but not quite. here's the prompt i've gotten to. 3d render of a spoon orbiting a star, in cosmic glory, colossal sized spoon with earth like features and textures across it's surface

[deft coral]

[delicate grotto]

You can do a 1st 8 steps with
A ball on a spoon
Or a ball in a spoon

On the 2nd pass lets say steps 8-12 a planet in a spoon

[glad fulcrum]

ghoul

[deft coral]

pregnant woman in sunset

[trim orbit]

some prompt editing might help a lot. there's hope yet @kind pewter

[wet nacelle]

How are you using multi prompting?

I'm trying to use it and am not receiving any changes.

[azure oxide]

now thats an interesting trope never seen before, zambies eating like civilized people

[uncut steeple]

Please reread #✍🏼｜rules-and-tos , epsecially Rule 4

[kind pewter]

thank you! & @delicate grotto

[deft coral]

who broke rule 4

[soft zealot]

DM me and I'll look later. On a conference call for work

[trim orbit]

Sytan's workflow is set up for multiple text inputs

[wet nacelle]

Can you send it?

comfy right?

[trim orbit]

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[dense chasm]

walking dead lol.

[glad fulcrum]

a winning image #🏅｜pantheon message

[deft coral]

bro this is clean

[molten gull]

it is ?

[royal python]

? si ti

[echo pine]

I also added random seed and the seed will be the same on all passes, to me the details are nicer but I haven't tried with other prompts yet

[uncut steeple]

Yeah that happens

[trim orbit]

ngl. they look 15

[safe spear]

As someone who hasn't tried LoRA before, what notebook (or tutorial) should I use to train a model for SDXL 1.0? I used Dreambooth when it came for SD 1.5 but I never tried it for SDXL 0.9 and I also never used LoRA

[deft coral]

who

[trim orbit]

kohya-ss with the gui is what i use. its nice and easy to work out

[safe spear]

Thank you, I'll try to use that one then

[azure oxide]

kohya's has always been touted as the best since dreambooth apparently breaks with updates

[woeful patio]

anyone know why my faces look like they've had lawnmower accident?

[trim orbit]

yes. your images. lets not expand on it. You do you. i'm just noticing the same trend with woman portraits as 1.5 had

[little gazelle]

Does anyone get this when trying to use a LoRa they have made on Kohya ss? modules.devices.NansException: A tensor with all NaNs was produced in Unet

[high escarp]

(8k UHD, soft lighting, high quality, film grain: 1.1), a cow
Negative prompt: NSFW, cartoon, painting, illustration, (worst quality, low quality, normal quality:2), (deformed iris, deformed pupils, semi-realistic, CGI, 3d, render, sketch, cartoon, drawing, anime:1.4), long neck, Photo of a zombie, Photo of a television screen, Photo of a camera lens, photo of unspeakable Horrors, Photo of a pile of body parts and blood, Close up photo of filleted skin, Photo of a cat, Close up photo of a tooth removal
Steps: 10, Sampler: LMS Karras, CFG scale: 3, Seed: 1409750279, Size: 768x768, Model hash: 31e35c80fc, Model: sd_xl_base_1.0, Version: 1.5.1(8k UHD, soft lighting, high quality, film grain: 1.1), a cow
Negative prompt: NSFW, cartoon, painting, illustration, (worst quality, low quality, normal quality:2), (deformed iris, deformed pupils, semi-realistic, CGI, 3d, render, sketch, cartoon, drawing, anime:1.4), long neck, Photo of a zombie, Photo of a television screen, Photo of a camera lens, photo of unspeakable Horrors, Photo of a pile of body parts and blood, Close up photo of filleted skin, Photo of a cat, Close up photo of a tooth removal
Steps: 50, Sampler: DPM2 Karras, CFG scale: 7, Seed: 1409750279, Size: 1024x1024, Model hash: 7440042bbd, Model: sd_xl_refiner_1.0, Denoising strength: 0.4, Version: 1.5.1

[vapid roost]

how do I use refiner in auto1111

[azure oxide]

1.5 had this issue, it was fixed with "restore faces" i believe. idk if that works with sdxl tho, mabey it does

[cold mica]

I've removed it, please refrain from posting content that is generally uncomfortable to view (mark as spoiler if needed) - esp for death/gore stuff

[trim orbit]

a lot of these services are trying to build brands by buying exclusivity deals. i hate it.

[royal python]

nah

[floral island]

skeletons (non-gore type) are ok, right?

[cold mica]

Yep that's fine!

[tender timber]

Does the ComfyUI not work? It lists Intel Arc as a card that can be used.

[trim orbit]

codeformer still works super well when it needs to. adetailer extension does a good job if you use it with a lora too

or embedding even, or just a good prompt

[azure oxide]

ya i was also going to note how 1.5 has that issue. could be a training thing. makes sense that images tagged as "girl" would be younger folks

[woeful patio]

stupid question but what's codeformer?

[trim orbit]

boy and girl token tends towards children aesthetics. woman and man tend towards adults. sdxl uses the same clip for one of the text encoders.

[little gazelle]

@trim orbit have you tried training a LoRa for SDXL yet? I'm trying to run a LoRa I made in the Kohya webui but getting thrown a NaN issue

[floral island]

ping, this is the answer

to get women, either prompt woman, female, or feminine

[tender timber]

What is the render time for 32 steps on the AMD?

[trim orbit]

it's another AI that's great at restoring faces

[floral island]

girl will get you ~8-20 year olds most of the time, depending on other context in the prompt

[low plover]

great.it's a very interesting thing

[deft coral]

does this look 3d

[floral island]

a bit, seems a bit blurry for 3d

[trim orbit]

i got a few NAN's when is tarted. in the GUI there's a tab called guide, offers good tips. you need to --network_train_unet_only, use bf16, and cache the text encoder outputs. ever since going that route, i've not had problems with nans

[cunning geyser]

anyone know a good upscaler for sdxl

[deft coral]

iv got 3d in the negative

[floral island]

remacri

[deft coral]

but idk what to do

[ember cargo]

do you guys know any proper workflow with lora and refiner combo? I currently don't use the refiner though

[little gazelle]

Great stuff, cheers

[trim orbit]

i find even when youre using "girl" in more adult contexts, the face still looks like they're working through puberty yet

[high escarp]

Super stable results at 768 one to one aspect ratio

[glad fulcrum]

yeah... I guess vegan art is banned

[royal python]

maybe try analog photo or something like that

[trim orbit]

good luck!

[deft coral]

can someone give me a good prompt

for detailed skin

[trim orbit]

a good prompt for detailed skin

[floral island]

can't disagree with you there though

[high escarp]

(8k UHD, soft lighting, high quality, film grain: 1.1), Raw photo of Hugh Jackman as a steampunk clockwork
Negative prompt: NSFW, cartoon, painting, illustration, (worst quality, low quality, normal quality:2), (deformed iris, deformed pupils, semi-realistic, CGI, 3d, render, sketch, cartoon, drawing, anime:1.4), long neck, Photo of a zombie, Photo of a television screen, Photo of a camera lens, photo of unspeakable Horrors, Photo of a pile of body parts and blood, Close up photo of filleted skin, Photo of a cat, Close up photo of a tooth removal
Steps: 10, Sampler: LMS Karras, CFG scale: 3, Seed: 1409750279, Size: 768x768, Model hash: 31e35c80fc, Model: sd_xl_base_1.0, Version: 1.5.1