✨｜sdxl

[ivory blaze]

looks like a screenshot from that cat game I like

i forget the name

[indigo carbon]

who here wants volunteer for this conversation? I don't have the mental capacity.

[vital ermine]

Yep. He was a joke in the general with images channel for a long time and no one watched him.

[mossy canopy]

stray?

[soft zealot]

@native knot same prompt as earlier but I've used different models and style settings(theyre in the image filenames if you open in browser)

[ivory blaze]

yes

[vital ermine]

Some days I didn't sleep and it christmas I was lucky they let me have 24h on colab.

[soft bone]

Olivio (SEcourses) is active in this channel. Probably my fave youtuber as well, not that I use youtube for this stuff. But I got the best reg image dataset I've ever seen from his patreon

[ivory blaze]

for god's sake I'm joking. Except about the comedy part. He cracks me up how excited he gets over simple things

[native knot]

Interesting differences. I like how the 2nd image has the flower in the hair due to where it put the color.

[ivory blaze]

yeah I love his stuff. I just listen to him while I drive

He's very, I don't know, he is so happy doing it, it makes me smile lol.

[vital ermine]

Olvio is a good guy and so cheerful all the time.

[ivory blaze]

yeah, exactly

[indigo carbon]

good. I probably couldn't tell because of my condition

[ivory blaze]

It's okay, text and sarcasm blindness is common among 100% of internet denizens

also the mr t stuff was intentionally bad, this my friend, is true art.

[soft bone]

he does not like vegetables

[vital ermine]

Eons ago I got a ban because they mistook what I said. When I said I even used 🙂 to show I was joking they said that people do that even when they aren't. I gave up on the internet denziens after that.

[ivory blaze]

yeah, very much. humorless limps are no fun, but a ban from such a place is no ban at all, but a gift imo

[ivory blaze]

no, there is a child out of frame I had it generate, he's angry the child wants his vegetables

[vital ermine]

It was, but I got my just deserves when they went bankrupt a year or two later and shut down. Was cool, but when nasty people get power, if even in a limited capacity, nasty things happen.

[soft bone]

whose child wants vegetables wtf

[glad grove]

do u have a moment to talk about our savior Jesus Seagal

[ivory blaze]

it is AI art, it need not be real

[crisp owl]

mine

[ivory blaze]

meh, мені слід сказати, я дійсно не люблю Стівена Сігала. Він симпатизує росіянам і не є доброю людиною.

[soft bone]

https://tenor.com/view/cookie-monster-vegetables-veggies-diet-pissed-gif-4651500

[steady grove]

think diffusion got confused and thought jason segal

[ivory blaze]

lol

[crisp owl]

100+ s/it
I currently have --disable-smart-memory flag on.
Would that effect your changes you just made?

[visual glade]

how much are you upscaling to?

[soft zealot]

this was:
ProtoVision for 3 steps in a Preconditioning type stage (refiner prompt from attached)
DynaVision for 13 Steps in a Base type stage (Base G & Base L from Attached)
ProtoVision for 9 Steps in a Refiner type stage (refiner prompt from attached)

[crisp owl]

just x2
img2img process goes

upscale model 4x > upscale image by 0.5 > vae encode > ksampler for 10 steps > vae decode

[native knot]

Ok...so you're starting to approach Cruella DeVille territory, which is interesting to me. That's definitely a hot look, though.

Didn't even notice the glove until just now...haha.

[ivory blaze]

I've looked but is there any no merger nodes, where I can set complete circuits within a node and combine them into a single reusable set

[vital ermine]

I don't know of any mixer nodes which would rock

[ivory blaze]

Wait I just thought of a really easy way to do this actually, a three node set with an ID value, the input and output will correlate to a second nodes input and output, and the input and output nodes will be placed at the edge of the circuits outside of the primary area

[vital ermine]

would require parallel nodes though

[glad grove]

come here my son

[ivory blaze]

That's true, see I don't know how it works, I'll have to look. Maybe if within the code it replicated the structure by the number of nodes that are used for a single input and output. Creating non-visual node sets that are copies of the original

[vital ermine]

Got me as I have only just begun to really use comfyui

[ivory blaze]

For example let's say you have no input a, and no output a, and they take and output a latent signal. For the first instance it would use the design that is outside of the primary structure, passing the signal remotely to the input a in and taking the signal from output a in the third node which is only an input and output.

Now while the input and outputs have an ID of A, the node could be A_1 which would be the first instance that is seen and designed by the user, whereas A_2 if it exists would copy 1 internally, not being displayed on the actual ui, to avoid the requirement of parallel pathways

[steady grove]

heeeres steven

[zinc cargo]

i remmeber someone a workflow with face repair for sdxl, right?

[ivory blaze]

I will have to look to see how it works actually I'm going to do that when I get home because I really really need emergency. Because spaghetti hell...

I need to look at the actual API I haven't ever really looked into comfy yet. I've just been treating it purely as a node base system. But I wonder if I could create perhaps a API note of sorts that would allow for definitions via selected Python extensions. Plant without having to create a specific node for each and that method

[crisp owl]

it works best last, after upscale. At least when I tested it

[steady grove]

segal as vorhees. he kills it imo (heh)

[glad grove]

real

[glad grove]

repent or else

[soft bone]

I'm gonna need to enlist the help of a node wizard soon.

Impact Detailer node is great but there needs to be a way of attributing the denoise value to the whole image rather than the cropped+zoomed image, so that the further away the face is the more it denoises. Right now it's not scalable whatsoever. Auto1111 works correctly where "inpaint at full resolution" gets it's denoise from the zoomed fraction of detected face and not afterwards

If there was a way to intercept Detailer after it has zoomed, and say "if you zoomed by x amount, scale denoise value up by x"

[crisp owl]

So running all three different canny models, nothing else changed:

Control-lora-canny-rank128
1630.46 seconds

adapter-xl-canny
207.32 seconds

diffusion_pytorch_model.f16
170.8 seconds

[thorny frost]

some img2img testing

sdxl is bad for this. had to use 1.5 to turn it into a photography

sdxl tend to lose detail blur and mess it all 😦

[vital ermine]

👀

[west breach]

That's odd. It worked well on my 3060 6gb. What upscaler model are you using there? I was using the 4x ultrasharp

[indigo carbon]

is there an interrogate node for ComfyUI?

[vital ermine]

SDXL requires its own and it was hard to find anything to make my captions for training.

I settled on llama-7b

[indigo carbon]

tf, LLaMa can describe images?

didn't know that

[vital ermine]

pretty damn well, yes

13b is even better but I found nothing that works with it

oob I gave up on

[indigo carbon]

is there a node for that?

[vital ermine]

¯_(ツ)_/¯

be damn nice if so

[indigo carbon]

I'm making a version of the AIT workflow that makes an image, describes it, and repeat

[nova osprey]

from this to this. i still have work to do with the scar coz inpainting always seems to fade it out and im clueless on how to make it more noticeable instead.

[sweet bane]

metadata + workflow included

[uncut gull]

SDXL model CrystalClear. See the PNG info for gen parameters. (If you didn't know, Euler A doesn't work with SDXL. It always produces blurry images. You probably already knew that, but here's a working prompt for clear results you can edit for other stuff.)

[vital ermine]

Wait a min I use euler A all the time in Automatic1111 as it was the best of them with less noise.

[indigo carbon]

figured out the interrogate thing. is there a node that takes the text from conditioning?

[vital ermine]

unclip

though I am not sure that is exactly what you need for this

[indigo carbon]

eh, it wants clipvision, definately not it

[vital ermine]

revision uses that node

close though and is doing the same thing. sees a pic to interrogate it.

might need to make our own node. I do know they have a chatgpt node

[indigo carbon]

not what I needed. I needed something more similar to how A1111 used to describe stuff

[vital ermine]

that is broken af though for XL

[eternal fog]

There's this that does Booru tags, maybe you could pull it apart and use something else for SDXL with it? - https://github.com/pythongosssss/ComfyUI-WD14-Tagger

[vital ermine]

wd tagging went poof with xl too

seems for training wd tagging and xl is a hit or miss now

[indigo carbon]

will check out

[eternal fog]

This also has some BLIP Analyse nodes - https://github.com/WASasquatch/was-node-suite-comfyui

[vital ermine]

that would be closer

[soft bone]

1000%. wd tagging fixed all problems with clothing artifacts but fucked the perspectives and proportions beyond repair. I removed them and instantly got the model i was hoping for

[vital ermine]

Yep. I always used them now forget it

Losing them is why I ran to try and find anything that would caption to only find one thing that would work locally for me

[soft bone]

I was super careful too. I pruned all tags of human characteristics and turned threshhold way down, so the only tags were the background, clothing, and facial expression. Still completely broken

I tried blip too though. For training faces on XL it seems captioning never really helps.

[vital ermine]

Yes, faces are bad many are saying and for me the first thing that hit me was the faces being smooshed in, missing, etc... that I never had in 1.5/2.0/2.1

I finally got it mostly sorted out but captioning was not an answer

[soft bone]

same. spent a long time wondering why the hell SAI told us XL is "easier to train" when i was finding the opposite. I'm finally happy with my formula but it takes an hour on 18 images

[vital ermine]

@upbeat summit you home yet?

[vital ermine]

Yeah, I am lol'ing inside because a friend of mine I said EXACTLY the same thing to two weeks back.

[steady grove]

i've been training loras in 30min with 16gb. i train at 896 buckets with batches of 2

[vital ermine]

I think one thing hurting us may be the two TE stuff, but not sure. All I know is XL has issues with training that I never saw in any version before.

[steady grove]

20-30 images usually

[vital ermine]

I wonder how many actually touch TE now or is everything --train unet only?

[soft bone]

Caith has been recommending TE training so most people active here are using it now. Even SEcourses and aitrepreneur are training TE

I'm training it too but only with token

[steady grove]

--network_train_unet_only

[vital ermine]

You know I took my friend's training and used windows and OOM instantly. I have to train on Linux or else. On linux 18-19GB on WIndows 24+ GB

[soft bone]

each time you raise dim, memory goes up by 2gb+. I recommend 64 maximum to train successfully on 24gb

[vital ermine]

Just saying that on Linux it trains fine while on Windows, changing only folder paths, no way no how. WHY?

[soft bone]

Caith and Olivio seem to think higher dims damage the model on XL. It might, but it still helps with details

[upbeat summit]

I never left?

[vital ermine]

I use prodigy so it sets all that

[autumn forum]

i see dims as like the more info you are trying to fit in the higher the dim. but it donesnt take much. i could be wayyyy off but it makes it easy for my pea brain to understand

[vital ermine]

Oh, lol. You tried the locon? btw, I posted it on reddit and the bots hit me. Each time I refreshed the number was going down and up and steadily down.

[steady grove]

i dont think the TE needs training tbh. it knows a ton of tokens very well already and training only the unet is faster and works.

i'd really like to see some comparisons. followfox might have something soon. 🤞

[upbeat summit]

not yet - but I will now. is the post still up? let me upvote it!

[azure oxide]

Smaller modules

It's possible we don't need to use dim=128 and adapt all attn layers. I suspect that we can reduce the size by quite a bit if we are careful about which layers to adapt.

https://old.reddit.com/r/StableDiffusion/comments/1223y27/im_the_creator_of_lora_how_can_i_make_it_better/
LORA creator thinks >128 isnt necessary either as long as you're careful of the other stuff

[soft bone]

yeah, i see it as another way to control overfitting and detail amount. I want the actual freckles and skin pores of the real person to be present in outputs. That aint happening with "dim 24" as recommended

[vital ermine]

I didn't dare post it in the old sub reddit as they hit me hard last year so I can only imagine now

[soft bone]

256 is definitely overkill for any case tho

[vital ermine]

I wish I could tell what my dim/alpha is at but it is automatic

[autumn forum]

idk ive never went over dim 12 and it got freckles and skin detail for sure. just dont caption the skin or anyuthing about the skin

[soft bone]

accelerate command should say

[soft bone]

I'll give it a shot. I remember my outputs being smooth af and i hate smooth

[vital ermine]

goes by so fast I never see that stuff. I will look next time but it is automatic past accelerate it isn't static set

I used to set all that by hand but it was a nightmare so I let prodigy, etc... set it with some long ass commands and optimizer settings

[soft bone]

i didnt know schedulers could change anything beyond LR

[ivory blaze]

my new potato LoRA

[autumn storm]

Quick question. I have used the model in automatic1111 based on the github model provided. As well as I used the sample on clipart. It seem like clipart produces much better quality results? What could be the explenation for the difference in quality? Is the model I am running locally not working the same as the model on clipart?

[vital ermine]

Well, I used to use iA3 and forced to switch to locon. Love the iA3 as it trained way faster and did better for sub 1mb.

905k is now 43.5m

[ivory blaze]

I said nothing of wating a panini

why is he eating a panini

[indigo carbon]

nah, I concluded the caption output is too close to the original- I just wanted to see the wacky results.

[ivory blaze]

?

[glad grove]

repent

[upbeat summit]

https://tenor.com/view/this-is-a-knife-crocodile-dundee-gif-10637727

[indigo carbon]

I was making a version of my AIT workflow that makes an image like the original, describes it, send to prompt, repeat

[ivory blaze]

oh i missed that

comment was just kind of - there

hmm I have yet to play with AIT

i should probably

[glad grove]

sensei 🙇‍♂️

[indigo carbon]

the AIT workflow I made is now excepted as the official one for SDXL with the AIT node, it's included in the repo

[ivory blaze]

oi

uh

i guess it did as I asked

))))

@glad grove the masterpiece is complete. I can leave generative AI forever, and collab with AIpretenaeur in videos about uncensored language models speaking french

[steady grove]

sd literally thinks wilfred brimley is mr potato head. thats all this prompt makes. a potato in the shape of wilfred brimley, russet potatos in a basket

[ivory blaze]

lol

lol

so this is because you are using the As/Is dialogue, try "a realistic carving of a potato, wilfred brimley, etc. "

[ivory blaze]

that is if he WERE a potato, not displaying a potato AS him

[ivory blaze]

mine was "photograph of an unfinished carving in a potato, fat disgusting steven segal "

[ivory blaze]

dear god that is horrifying

Wilfred is all carbs

[steady grove]

diabeetus

[soft bone]

tried lora-FA yet? heard it converges twice as fast as standard

[vital ermine]

???

[steady grove]

dia beeeet us

[ivory blaze]

oh lord

[soft bone]

brand new lora type available only on kohya dev2 branch

[vital ermine]

I was using -b dev then found out dev doesn't work with locon and he said to go back to master

[ivory blaze]

so this is what got me banned from DALL-E2 forever ago,

[soft bone]

git checkout dev2. locon works. 2 seconds to switch back if it doesnt

[ivory blaze]

while we'pre on beets

apparently kim jong il as a bowl of borshch is not okay :\

cos openai right banned me

[vital ermine]

does it work in gui for training and more impotantly does it work in comfyui?

[ivory blaze]

(during the beta long while back)

[soft bone]

yes and yes

[steady grove]

its probably the intent behind the image that they don't want their brand associated with, more than anything else

[vital ermine]

damn. my gui is updated and not even shown

[ivory blaze]

The guy is A) dead, and B) a public figure and C) it's borshch, everyone loves borshch

[ivory blaze]

wilfred brimely, ultra detailed carving, in potato

[soft bone]

#🔧｜finetune message

it wont show unless youre on dev2

[ivory blaze]

not sure what it happening, but it is.

[indigo carbon]

I did finish building the AI telephone workflow, but SDXL's clip is too accurate it was useless, so yeah

basically, this would be entertaining on old models like 1.5, but because SDXL is so good, it's useless

[vital ermine]

so -b dev2?

[soft bone]

idk what -b is i just do "git checkout dev2". is that a linux thing?

[soft bone]

The FA stands for full attention ig

[vital ermine]

well, I will have log into linux and do the lycoris shuffle git clone -b dev2

if locon did 15 epochs at 1800 steps I wonder what FA lora would do?

[soft bone]

#✨｜sdxl message

i havent tried it myself yet

[vital ermine]

Screw 300 epochs.

I will see what prodigy does with it

[upbeat summit]

pretty cool LoRA! got inspired by your samurais

[vital ermine]

VERY cool

[vital ermine]

That lora will also do realism

[upbeat summit]

yeah I just saw that and checked the settings if I was doing something wrong 🙂

[vital ermine]

I was doing it by accident

you can add realism, and ambient occlusion (that one is a great one) and it will switch away from cartoon

it is a great effect

It is a nice lora but so many people are back on 1.5 now

[upbeat summit]

but SDXL can be fine-tuned so well

it's amazing what you can do

[vital ermine]

anime people are back on 1.5. They say 1.5 for anime and realism XL.

I could not change their minds so we have two groups now splintering the community

same as it was with 2.x vs 1.5

[glad grove]

yea cuz theres not a good anime model yet

[vital ermine]

precisely

[glad grove]

maybe later there will be a good one

[vital ermine]

they said thanks to NAI being jacked anime/1.5 lived

seems everything is a descendant of NAI if Anime related

they said if we can get an anime model training on millions of images like NAI did but for XL then they will come. Made me a bit sad and perplexed.

[glad grove]

😔

[vital ermine]

Reason I stopped releasing for 2.1 is everyone run back to 1.5. Now I am getting the same vibes for XL.

[glad grove]

maybe someone will figure how to transform training data from 1.5 models into XL ? or not possible?

[vital ermine]

not possible

SAI already said that with 1.5 to 2.1

now, what they could do is retrain and that would work

[glad grove]

😔

[vital ermine]

a lot of those models came from people with 8-12gb cards and on XL it laughs at puny cards for training. 😦

[ivory blaze]

use XL img2img and upscale to 1024 and then use those images as the dataset to train a new model 😄 or do you mean the already refined models to XL format?

[vital ermine]

YES, or straight to 1024 as the gen source as most of those models do it

I am laughing

[ivory blaze]

well...

Currently allocated : 6.59 GiB
Requested : 5.00 MiB
Device limit : 8.00 GiB
Free (according to CUDA): 0 bytes
PyTorch limit (set by user-supplied memory fraction)
: 17179869184.00 GiB

I need more GiB

[vital ermine]

see

sort of drawn but not

[upbeat summit]

yup 🙂

[vital ermine]

it really is a good lora that will fizzle and die on the vine.

[steady grove]

very very cool

[fresh path]

Upscaling a dataset is a terrible idea

[vital ermine]

yep

I tried that once in a training and oh, my

it saw shit I never saw

[fresh path]

that shit is like cocaine to a diffusion model

[ivory blaze]

hence the emoji... I was joking

[vital ermine]

Seems XL likes drinks

[ivory blaze]

and brussels

[ivory blaze]

and potatos of course

[vital ermine]

hehehe

[steady grove]

i upscale datasets all the time to great result. only a double though. a ton of cheap photos around 600x to 800x sized photos of varying aspects. throw away everything with significant artifacts. i crop slightly only to minimise bucketing. so if i can shave 10 pixels to make it match another better i will. then i pretty much ALWAYS double from there. any image that is smaller than 1024 in any dimension.

after that, kohya sizes down all the images so they fit buckets built around the resolution i define. cropping is more about ratios than actual pixels because of this end step. i've never had any problems with gan artifacts showing up where they shouldn't and i actually pull great detail into the loras i train.

don't upscale shit though. don't even use it.

[ivory blaze]

nah, I mean when I have, I just use CLIP out as prompt and img2img at 2x @ .6-.85 seems to have good results. upscale itself is garbage

wtf

it just put out a nude potato lady

:\

[shy kelp]

idk i feel want to go back to sd 1.5

[steady grove]

OH you're still pumping it though, you're just not using the upscale models specific. cool!

[ivory blaze]

yeah

it's just using the CLIP output of the original

[steady grove]

hot potato means you gotta give it to the next guy

[ivory blaze]

and then regenning something LIKE it

but usually kind of close, but it's good for a dataset

since the tags will be corrolated since it CLIPs

works not so bad

cheap quick and dirty, but works need be

not at all what i asked for, but k

[shy kelp]

what is the different sd2.1 and sd1.5? why many user using 1.5 insted of 2.1

[ivory blaze]

@shy kelp it's... friendlier

[vital ermine]

Careful, I almost got banned for that and had a HUGE adressing to by a mod for doing the same thing.

[shy kelp]

wdym?

[vital ermine]

wise move.

[ivory blaze]

oof, I mean it isn't like, dirty, it's clearly artistic style, but yeah

[vital ermine]

censored boxes is a nono it seems.

[ivory blaze]

and really, they aren't even per se, what you see. it's smooth, no nips or any detail

just didn't wanna risk it

just weird because I asked for a woman carved into a potato with potato skin filagre

[vital ermine]

yeah, not worth it. Hell, I no longer show any children because I got a 3 day vacation when I did absolutely nothing wrong but a resdident pearl clutchers took offence at the showing of children and down I went. smh

[ivory blaze]

oof

yeah uhhh

adam sandler and tiny tim's love child? maybe? wtf

[vital ermine]

@soft bonefatal: Remote branch dev2 not found in upstream origin

tip toe through the tulips

for a 1.25 Paul.

[lethal prairie]

For SDXL, what's the current opinion on A1111 vs ComfyUI vs ??. Just looking for the best frontend to run it locally.

[vital ermine]

you a grease monkey or a Sunday driver type?

[lethal prairie]

Somewhere in the middle. I can tinker but I do appreciate things that work easily.

[vital ermine]

Stick with the dev branch of Auto1111

until a proper frontend that is friendlier to novices is made. As Comfy said ComyUI was made as a backend and the ui it has was thrown on.

[lethal prairie]

I was already building a brand new instance of that, so I guess I'll get it going and see how it is. I just saw someone else complaining that it didn't integrate the refiner very well.

[vital ermine]

I am sure someone is going to come out with a really good frontend soon

[heady vale]

interesting result using a 2.1 embed with XL

[vital ermine]

most are not really using refiner anymore as Sytan has stated it is not good.

[upbeat summit]

nice!

[lethal prairie]

Oh, I didn't know this. I haven't used SDXL at all yet so I'm unfamiliar.

[vital ermine]

well, on auto it is implemented really bad and on master branch it will crash you to a BSOD if it runs out of page file. I have 48GB of ram and my page file was 128GB and growing. I limited it and BOOM. In the dev it loads base, then unloads it, then loads the refiner, then unloads it to reload the base. Really bad.

[lethal prairie]

I think that was the basis of the complaints I was reading.

[vital ermine]

there is one other issue

[glad grove]

they still havent fixed it? its been more than two weeks already

[vital ermine]

sure haven't

the other issue is that the images are foggy/lot of noise and the same prompt in Comfy is crystal clear

made me think my lora was jacked

[lethal prairie]

Well that seems like a big problem.

[vital ermine]

it is and isn't. For me I finally gave up due to that but what was keeping me is the iA3 I trained in. In Comfy that will never be implemented so I stuck with Auto as long as I could.

If I am going to release loras/locons they need to be made for what people will use them on, if anyone is still around with XL

[vital ermine]

was that dev 2 of lycoris or of the gui?

[soft bone]

no idea what you're doing. i use regular kohya on windows and dev2 has been there for months

dev2 of bmaltis kohya

[vital ermine]

see for iA3 you had to go into the gui, delete the LyCORIS folder. do a git clone https://github.com/KohakuBlueleaf/LyCORIS -b dev then venv and pip install .

[wet raven]

Frankchieng talked about https://huggingface.co/stabilityai/stable-diffusion-xl-1.0-tensorrt
I am wondering if this is for 1080 Ti video cards or not

[vital ermine]

no TC in pascal cards @wet raven

[soft bone]

https://github.com/bmaltais/kohya_ss/tree/dev2

[vital ermine]

they didn't come until Turing for consumers

[soft bone]

literally all i do to switch is open command in kohya folder, type "git checkout dev2" and close

[wet raven]

ok thanks

[vital ermine]

welcome

[vital ermine]

do you use windows or linux to train in?

[soft bone]

windows

i set up linux once but too many hoops to jump through each time, considering i train all day every day

[vital ermine]

Linux is superior at this than windows but all my tools are windows only :/

even triton is Linux only the windows one does nothing but shut up programs complaining

[soft bone]

accessibility vs. function. a tale as old as time. Comfy vs. auto1111, linux vs. windows, android vs. apple, pc vs. console

the only one i havent opted for function in is windows

[vital ermine]

should I install cudann? I know his versions are out of date

[soft bone]

i believe its much faster

[vital ermine]

was no difference in Linux with them or without

[ivory blaze]

second pass,

[vital ermine]

A fatter adam sandler again

[ivory blaze]

oprah getting in on the assasination of rasputin

[vital ermine]

I am just waiting for a new training scheme to hit to compensate for XL

nvidia is supposedly going drop something for 4090s but who know might have just been all hype talk from them

[ivory blaze]

there we go

[tepid surge]

your pfp so cuute >u<

[wet raven]

I wish i remembered who made it for me. Its a Toad + Snail hybrid.
My invention, but not my artwork.

[vital ermine]

Oh, and I tried deepspeed and that was a nightmare so I opted out

[tepid surge]

uwu so cute, glad you have it

[soft bone]

There's a 4 inch long vinegaroon in my house somehwere. should i kill it

[vital ermine]

something I never knew is the last accelerate config option of no, fp16, bf16, or fp8 does it matter?

[soft bone]

for mixed precision or save precision?

fp16 for mixed precision is causing problems

[vital ermine]

I mean we set that in the gui

[soft bone]

oh i did bf16 for that

[vital ermine]

my 4090 will do fp8 but kohya does not

[wet raven]

RandomPower#3956 Made my Discord Avatar on Game Dev Underground server

i saved that as a note so i'll remember

[soft bone]

welp im gonna kill it

bitch is bigger than the one in harry potter

[wet raven]

oops dal_mac, didn't mean to DM you

[ivory blaze]

Benito Mussolini, liked children's cereals.

[soft bone]

its right there m8

[vital ermine]

exactly, I was showing you there it is

I have my previous iA3 model I trained and will redo it now in FA

[soft bone]

ok lol. my first time switching i actually didnt see it for a minute

[vital ermine]

on Windows to see

yeah

I hate how it autostarts without the dark mode so I have to type the url with it. Annoying, and blinding

Just changing my paths and move that to FA

[soft bone]

"Freeze A and only train full connection layers. In simple, better memories efficient and training faster" purpose of FA i guess

[vital ermine]

we do not need any special flags?

[vital ermine]

Reminds me of a side scroller type game

[heady vale]

prompted for Terraria landscape

[soft bone]

knew it

was about to say that

[vital ermine]

hehehehehe

I tried that game and never could do it.

[soft bone]

same

[vital ermine]

I went back to minecraft

[soft bone]

but i cant even get into minecraft

[heady vale]

it can get quite difficult

[vital ermine]

I can't now but back then it was alright

[soft bone]

i had minecraft on pc alpha in 2011 or something and got bored

[heady vale]

Ive been playing MC modpacks for years. I like the expert packs that take a few months to complete

[soft bone]

never looked into modding it. I extensively mod tho, in skyrimVR and now Kerbal

[vital ermine]

RUDY

I just lack the room, or bank account now, to VR

this damn 4090 hit me hard

[soft bone]

I have been using vr for 6 years now, and to this day i keep my feet planted. the only room i need is for my arms to swing

[vital ermine]

when I train my case gets so hot while the card goes around 69-70c. Fans that came with the case suck so 140mm is OOS but is next.

[soft bone]

my 3090 stays at 85 while i train so

i live in the desert

[vital ermine]

well, nice thing is when it gets cold this is one heater for the room

[vital ermine]

the case fans that came with it are 120 and just can't push a feather at max rpm. New ones do almost 70cfm a piece at over 2mb/h20. Arctic makes good fans

now that we are headed back into covid masking and lock downs I need to get while I can get

[soft bone]

lian li uni for aesthetic, noctua for power. I got an industrial 140mm that rivals my $140 room tower fan

[vital ermine]

my side glas right at the back where the actual gpu is blow so hot the glass gets almost too hot to touch

well, I need 4 but only come in a pack of 3 and I don't wish to spend 40-100 per fan

[soft bone]

2400 pwm or some shit

[vital ermine]

damn

these do 0-1900 and 0 is for real

never gonna do 0 though

[vital ermine]

I bet these fans aren't even for pressure so even worse

I finally went with a aio and I am in love. When it dies I want to go to a full on open loop but the block for this 4090 is almost 200 USD, sheesh

btw,

I am about to train but I am unsure what to do and how to test if it is as good? 17 epochs with 1 repeat 130 images

that was the locon/ia3 but for this?

[soft bone]

i've never used a type other than standard before so no idea

i just know FA takes less epochs than standard

[vital ermine]

standard is 10-17 epochs for me with 120-130 images 1 repeat

[soft bone]

i'd do 10 and save each one

[vital ermine]

damn, almost half? Yeah.

save each 1 sounds like a good idea

tensorboard is pretty useless I found

[soft bone]

@vale eagle Said it was much less than half for them

[soft bone]

i just ranted about that in finetune channel last night. changing the learning rate drastically didnt change the slope of my loss whatsoever yet people are using that to determine "overfitting"

[vale eagle]

No, I do another test which slower than normal lora but it have another effect

[soft bone]

with FA?

[vale eagle]

Haven't post it yet

[vital ermine]

Precisely. Friend, and I, train a lot and can't determine a damn thing based off the nonsense of tensorboard. Change stuff and the graph stays the same,

[vital ermine]

good or bad?

[vale eagle]

yes, it is weird. The test conflicted with previous result. But it should be more accurate

[vital ermine]

well, we don't need a slower way to train

[soft bone]

i wouldnt mind, if it fixes clothing artifacts when not using captions

thats my last remaining problem

[vital ermine]

what sort of artifacting?

@vale eagledid you use any special flags for it, or arguments?

[vale eagle]

#🔧｜finetune message

[soft bone]

if i prompt the subject into a context where someone would be wearing a certain thing, but dont specify what that is, i.e. "ohwx man at a fancy gala" it puts them into a tux, but their clothing from the input images messes with details of the tux, such that its not believable anymore.
It's overfitting slightly, sure, but to guarantee likeness I cant afford to overfit any less. captions solved this completely but as we know they have to be perfect or its fucked.

its very small details. things like too many buttons too close together, or a necklace not following the laws of gravity

perhaps lowering dim can help but I imagine anything that helps this would also hurt likeness

[vale eagle]

Lora-fa has more room for prompt even it's overfitted.

[worthy orbit]

beautiful, always loved those kind of level map in game magazines

[vital ermine]

weird

[vital ermine]

Well, not sure I like the FA results

base, ia3 17 epochs, f4 10 epochs

base, ia3, fa

I wonder what dim/alpha I should use and is this using conv too?

[twilit nexus]

that is cool

looks very usable in games

[heady vale]

yeah it could be for sure

[twilit nexus]

btw is there a way to make the textures seamless

i made some textures but they ended up not totally seamless

[heady vale]

Ive not tried seamless gens yet

[dense chasm]

https://www.youtube.com/watch?v=fEtHfsYvU0U

[glad grove]

speaking of wukong

[vital ermine]

I upped the dim and alpha to 32/1 like a normal lora so it is over 4 times larger than previously and not a lot of difference.

[strong field]

Testing out new workflows and well im actually impressed

[vital ermine]

it must use conv since lora hides them but FA they appear

was trying to decipher the papers

[twilit nexus]

you just have to put tiling option with texture style and check y n x tiling and it tiles in sd

now i don't need to manually fix my images

is using sdxl worth it?

m still using 2.1

[heady vale]

for sure. you can gen a lot more things

[twilit nexus]

i just make game assets

like fixing textures n such

mid journey was amazing for making concept art for games

[glad grove]

depends if u want anime style or realism

[heady vale]

I used 2.1 only for months and at the end felt quite restricted

[twilit nexus]

no to realism never

seen enough of reality lol

i want something like hand painted like league of legends

sdxl is more realism?

[glad grove]

yea or comic style,anime style 1.5 still better

[shy kelp]

how others got realistic face on sdxl

especially eye

[half ivy]

BBRRRRROOOOOO

[dense chasm]

Hugging Face just raised a 235M / 4.5B series D!

[half ivy]

I am here

is it still SDXL if the init was made using SDXL but cleaned up in Auto 🤔

[vital ermine]

8 epochs at dim rank 16 else I lose the ring of light in the middle on the bottom. It did not matter if I did 20 epochs it would not bring it back until I went from 8 to 16. I have not tried between those.

[soft bone]

is the ring of light in the training data or something?

[vital ermine]

no, it is in the base image though

[soft bone]

until finetuners have time to train. consider where 1.5 was at 1 month in (absolute trash)

[vital ermine]

this is without my lora

[soft bone]

what exactly are you training

[vital ermine]

right off, regardless how low I went see the top of the saucer? it isn't smooth it is actually bulk plating.

it refined it.

I only train styles not objects

so the bulkhead plating is awesome

[soft bone]

thats with FA? thoughts vs. standard so far?

[vital ermine]

against locon, or ia3? Not as good

base, fa, ia3

fa barely changed it even when I did 20 epochs

16 dim

or 32 dim

[soft bone]

hm. different for faces im sure

[vital ermine]

base, fa, ia3

ia3 I get the same with locon only locon is larger

[soft bone]

FA is improving on base without changing architecture. looks best to me so far

[vital ermine]

not for styling as that is not what I want.

[soft bone]

yeah impossible to judge your results without knowing your dataset

[vital ermine]

dataset is irrelevant as I am looking for the greatest changes when I use a lora/locon.

this, I lay odds, is more for objects/subjects.

[soft bone]

thats dumb. for the greatest change you can just do LR of 1e-2 and overfit the shit out of it. plenty of change. shit quality tho

[vital ermine]

did you know the reason we have locon is for styles as a lora isn't really made for that

1 layer deeper which is where we get the styles from

so, for you and clothes on your subjects this would be good I think

[soft bone]

i havent done styles since DB. if i wanted to do styles again id probably start with finetuning

[soft bone]

could be. will try

[vital ermine]

if you notice it is refining the base image

[dense chasm]

still remembered the floppy disk drive and internet explorer DOS and office https://twitter.com/Rainmaker1973/status/1694597311723180089

Massimo (@Rainmaker1973)

#Today in 1995, Microsoft released Windows 95.

[📹 PC USER 486]

Likes

17119

Retweets

3223

▶ Play video

Twitter · Aug 24, 2023, 6:27 AM

[glad grove]

i remember when games used to come in those big 8 inch floppy disks

[uncut fiber]

i still remember games on those compact cassette 😄

[glad grove]

yea i heard some of them are pretty expensive now ( only if they are sealed in mint condition) could be rich now 😔 💸

[vital ermine]

Hey, I still remember tic tac toe on a stone tablet. fun, but a lot of work when it was time for the next game.

[glad grove]

fighting with a stone axe was my fav sport

[vital ermine]

head splitting. Damn, those were the days.

[autumn storm]

Quick question. I have SDXL model in automatic1111 based on the github model provided. As well as I used the sample on clipart. It seem like clipart produces much better quality results? What could be the explenation for the difference in quality? Is the model I am running locally not working the same as the model on clipa or their playground in terms of quality.

[uncut fiber]

The first is cats can play as well and second is cat's bosses are playing too.

@pure crystal cfg, VAE if quality is realy poor. Feel free post image to compare

[pure crystal]

Could you explain? Not sure what you mean

[uncut fiber]

it looks good to me. @pure crystal
CFG can be too high to affect image quality. And VAE is think that in some time can hurt quality as well.

[pure crystal]

what settings would you adjust in the prompt/settings?

[uncut fiber]

i am not in A1111 and load sdxl models takes ages, can you make screen of your A1111 where prompt is?

i were at dentinst 😄

[pure crystal]

@autumn storm post some example images

[uncut fiber]

@pure crystal sorry i mistaken you for @autumn storm

[pure crystal]

SDXL is so much fun!!!!!!!!

[uncut fiber]

you have @pure crystal base model?

[hardy cipher]

https://tenor.com/view/dinie-gif-20207710

[pure crystal]

dynavisionbeta.safetensors

[uncut fiber]

great i am using copax or sorach, dont know 🙂 @pure crystal
And comfy or A1111?

[pure crystal]

Comfy for SDXL for now. A1111 for everything else

[uncut fiber]

yes seems right

[pure crystal]

Once I get comfortable with controlnet and inpainting workflows in Comfy I think I'll use it most of the time

[uncut fiber]

he reminds me that ugly con from Galaxy quest 🙂

[uncut fiber]

GAPA!

[ionic gulch]

IP-Adapter + Canny:

[vale eagle]

third is canny source, second for ip-adapter and the first is the result?

[ionic gulch]

yes

[vale eagle]

did ip-adapter need text prompt?

[ionic gulch]

only some info for the style: 3d imax shot, 8k detailed photograph, centered, uncropped, dramatic lighting, cinematic, vivid,

[vale eagle]

If you use the same image of ip-adapter for clip vision g, would it be better?

[ionic gulch]

you mean in addition to ip adapter or instead?

[vale eagle]

in addition

[ionic gulch]

no with my first try - will try again later after work:
3d imax shot, 8k detailed photograph, centered, uncropped, dramatic lighting, cinematic, vivid,

[bronze arch]

How is control lora different from t2i adapters?

[sour obsidian]

T2i adapters only run inference one time and have a lack of understanding of the prompt or timesteps so they are pretty limited in how far they can be used with the model, the benefit is they are very fast. ControlLora adapts the models own encoder (insteading of needing a whole 2nd one like controlnet) which then has all the generative knowledge to generate its own controlling context including the prompt and timesteps so it not only is better for simple tasks but possible to have it handle much more complex ones.

[molten gull]

can some of you guys explain your current workflow of how to upscale stuff ? (or how to get more/better details into an image) ?

[ionic dragon]

any plans to launch t2i adapters too?

[vale eagle]

what would be coming up in next sdxl release?

[wraith apex]

How is 3060 12gb for sdxl? Is it enough? With controlnet? And how much better a 4060 16gb would be?

[sour obsidian]

Not really atm. We had some internally at one point but they really don't have the same strength or usability controlnet does and now with control-lora we arent trading off vram requirements anymore

[bronze arch]

i see

thanks

[sour obsidian]

Hopefully stuff people like haha, nothing guaranteed atm. Lots of testing of new ideas going on (or old ideas in new ways)

[west breach]

@sour obsidian are you working on a tile controlnet?

[sour obsidian]

Would love some better upscaling systems. Ill say its on the list to dive into real soon, as above though, no guarantees

[west breach]

was my favourite controlnet for SD1.5

[shy kelp]

is ipadapter and clipvision a different thing?

[vale eagle]

Revision is pretty impressive. Hoping for other great work.

[sour obsidian]

Yea its pretty impressive, really like what it does conceptually

[sour obsidian]

Im glad you like it 🙂

[peak dove]

SDXL Good Morning Comfy-ists!

[icy brook]

The King of the Dead, LOTR

[ionic dragon]

so controlloras dont consume extra vram as they are being used?

[sour obsidian]

From a glance it looks like a modified transformer block that splits the cross attention and projects a clip image embedding in as a new input instead of text and then mixes the outputs cross attn for text and img. Hmm interesting. Wonder how a fully tuned revision would compare. Cool approach, I like the idea of the split. Is it slower for inference time running both inputs?

[sour obsidian]

Much much less than reg controlnets since there is no secondary model like with controlnet

Its just turning the models own input blocks into its own controlnet

[uncut gull]

See SiliconThaumaturgy's sampler testing. Euler A doesn't work for SDXL. You'll get smooth, fake-looking skin / hair etc. and blurry backgrounds. https://youtu.be/JAMkYVV-n18

[ionic dragon]

so control loras use a technique where the seed starts with controlnet?

[sour obsidian]

Kinda. It essentially has the model run its own encoder blocks 2 times. Once with the lora applied where its acting as a controlnet, and another time fully through like normal without the lora

[ionic dragon]

oh ok

@sour obsidian any info about multi gpu training?
as its a very anticipated thing, as SDXL training consumes a lot of vram, it helps people to use multiple gpus so they can accumulate the vram

any plans on that

[sour obsidian]

Yea, what you're looking for is sharded training. Id check out deepspeed or fsdp. They should work on any pytorch based trainer but of course takes a little know how to set up. Its not really XL specific using them as they are basically required for very large LLM tuning now haha. Check out this video for some info on deepspeed but should give a great gist of the concept (https://www.youtube.com/watch?v=pDGI668pNg0). Id imagine these could prob be integrated into something like kohya np with some initial configuration for anyone who has access to mutli-gpu

[ionic dragon]

will check it out

[sour obsidian]

We use sharded training though for basically everything internally, typically with our internal variant of this for SD stuff: https://github.com/Stability-AI/generative-models

[upbeat summit]

hey @sour obsidian!

[sour obsidian]

Heya 🙂

[tepid surge]

thank youuuu

[ionic dragon]

@sour obsidian i dont see many community models coming out as much as 1.5 did, is it bcuz XL is OP?

[vale eagle]

Does the generative-models is a training tool? We could use that to fine-tune the base model?

[sour obsidian]

Haha maybe? We certainly put a lot more effort into making sure the base model was really usable vs the older ones, im sure the model also being a bit larger and needing better hardware to train contribute as well

[sour obsidian]

You totally could yea 🙂

[ionic dragon]

yes and yes
its very gpu intensive and comparatively takes more time to train

[icy brook]

maybe I should train a new Double Exposure SDXL version? 🙂

[vale eagle]

Does 24g vram enough to use the generative-model?

[peak dove]

Great Idea!

I am back using A1111 in Base Mode only ... it works a lot better than at the outset of SDXL!

ComfyUI is great - but I miss my favourite A1111 Extensions (and .pt files!!!) (Latent Mirroring, Image Browser)

[sour obsidian]

With the right config it should be possible. Ill be honest though, I only use it with larger cards so I am not quite sure how its low end is 😅
Likely kohya is better unless youre trying to train on chunkier datasets

[peak dove]

A1111 in Base Mode only

[icy brook]

this is from my 2.1 model

[peak dove]

Cool!

[vale eagle]

OK. Used kohya already. Just want to know if it would be another option or not.

[sour obsidian]

It is a large and very capable trainer. Lots of knobs and easy to dev on. Its built though with larger scale training in mind but not limited to it

[sweet bane]

This is brilliant!!

[icy brook]

Thanks! Made with this I made yesterday:
https://civitai.com/models/133412/aether-ghost-lora-for-sdxl

[vale eagle]

would take a look. Hopefully, I would rent A100s to do something training with it. 🤣

[thorny frost]

wow, i dont know what was wrong with my prompt, but tried yours and the result is amazing, at the first attempt.

my prompt was: cinematic photography of a twisted dead tree in a deep forest background and sun rays coming through the branches, bluish fog , metaphysical, very detailed, intrincated . 35mm photograph, film, bokeh, professional, 4k, highly detailed
Negative prompt: drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly, low contrast, blurry, desaturated, drawing, painting, lines. I used copaxTimelessxlSDXL1_colorfulV2

[ionic dragon]

oh wait, so this tool can be used to finetune?
and supports sharded training?

[sour obsidian]

It doesn't natively support deepspeed or fsdp though it might be updated to have it natively in the future. It should support ddp_sharded though atm

[ionic dragon]

the statement is too technical for me

[thorny frost]

that is pretty nice, specially the god rays, but the point is to make it similar to the original drawing

[shy kelp]

i didn't notice it image2image haha

[thorny frost]

trying to get similar godrays but it turn them on branches

[shy kelp]

noicee

[thorny frost]

sometimes it does a photo, sometimes a drawing, depending on the seed. It is curious.

[shy kelp]

how to use this node

[thorny frost]

no idea. it makes a new model from text and another model?

[thorny frost]

final result, only a slight sharpen in photoshop

[shy kelp]

looks good

openpose is fun

[tepid maple]

RIP

[zinc cargo]

anyone encoutered this before?

[tepid maple]

prompt very long ?

[rustic garnet]

looks more like something on your timesteps is causing the error

[uncut fiber]

it was solved yesterday, that this is just warning, but should be solved.

#✨｜sdxl message

but message he was answering was deleted. It was same screenshot as you are posting

[zinc cargo]

actually, it was me, so sorry for the dupe 🙂
totaly forgot i asked it, and yes, i was answered as well. thanks!

got to love those realiastic models 😉

[thorny frost]

a bit more balanced and detailed. (last one, i promise)

[thorny frost]

what is doing to those tomatoes? nothing good probably

[zinc cargo]

tomato core

[uncut fiber]

What model is it?

[zinc cargo]

i think proton... with some loras.
the info is wihtin the PNG
just drag it to your comfy

[uncut fiber]

o.k. some are deleting it. And got opened my now, so probably later. Thank you!

[zinc cargo]

i have no secrets 🙂
i've learnt from @upbeat summit's 2.1 images exactly like that, opened at my local SD and try stuff out with his prompts and settings.

[uncut fiber]

i am learning now from magic prompts and so. It is part of dynamic prompt. Nice long prompts, but problem is i dont speak english well, so some basic words i have to find in translator.... 😦

[shy kelp]

recreate

[zinc cargo]

what's your first lang.?

what's magic prompts?

[uncut fiber]

cze. And magic prompt is part of dynamic prompt. You can find in manager and how to IT is on github

[shy kelp]

it enhanced your prompt

[zinc cargo]

you feel it's required with sdxl?

[uncut fiber]

what do you mean? It is working with comfui, so why not using it

[thorny frost]

Nice. Fog is gone and tree is having branch loss issues.

[shy kelp]

became bonsai

[zinc cargo]

i mean that i thinkg that the prompting with sdxl can be simple and still give great results

[thorny frost]

That is pretty 😍

[uncut fiber]

yes of course but i like for example randomize input to be surprised. Or it can your idea -> whisky and soda change to different prompt like vodka with chicken legs or similar. It is not limited to make prompts longer @zinc cargo

[shy kelp]

yea on sdxl, doesn't have to put a long prompt, and it'll create an amazing images

[uncut fiber]

@thorny frost godrays seems to have more sources.

[shy kelp]

wrong person

[thorny frost]

Yes, too much sources

Many

[uncut gull]

Now I'm very curious. I feel like your prompt should have worked. The most powerful parts of my prompt are positive "Unreal" and negative "depth of field". I would have to experiment with your prompt though to see exactly what effect those negative art terms have on the result.

[thorny frost]

@uncut gull Are you using comfy? I am using a1111. May be that. I got very different results with different seeds. Some drawings, other more photographics.

Seems unstable

Otoh loaded the 1.5 model, and boom, 100% photo at first attempt

[tepid maple]

anyone good upscaler table site ?

like comparison for all

using LDSR at the moment good but slow

[uncut fiber]

https://upscale.wiki/wiki/Model_Database
not sure it is what you want @tepid maple

[tepid maple]

yep perfect

lots of example I can look at

[uncut fiber]

then

[tepid maple]

time to get a 1tb ssd

ssd is full

cant d oshit

rip

here goes 110 bucks to ssd 😄

spend 200 bucks on 3060 12gig too

[thorny frost]

That is going to favorites asap

[tepid maple]

yep

[zinc cargo]

deleted all pics!!!

[tepid maple]

its just 240 gig ssd man it was never goign to survive

[zinc cargo]

lol

time for a purge

[tepid maple]

I have 4tb hdd but it is slow

time to go new ssd

give me some advice for my website

what do you guys want ?

every feedback is appreciated....

my aim is to make really easy prompt finding and sharing hd images only

all iamges will be HD or more

[thorny frost]

I use SSD for the checkpoints I usually work with the most, and a big ass mechanical hd for everything else, including less used checkpoints LORAs and such

[tepid maple]

my models and loras alone 300 gig atm

[uncut fiber]

how about m.2?

[thorny frost]

Mine a couple of teras, included some LLM models. But you must be selective and leave the things you use daily in the ssd

[thorny frost]

Have my motherboard covered with them.

No place for one more

[uncut fiber]

😄 i have only one, but it is truly high speed.

[thorny frost]

Yep, much faster than SATA SSDs

[shy kelp]

crazy how ai images can create a fake news, from peaceful city to apocalyptic city

[zinc cargo]

controlnet?

we should start a challange creating utopia from places not disasters...

[shy kelp]

yup using canny

[uncut fiber]

bit of peace 🙂

[ionic gulch]

is there are reason why things like lora, ip-adapter and revision don't work well or even give broken result or console errors if used with refiner instead or in addition to the base model (in a usual 80 base / 20 refiner split setup)? this reduces the usefulness of refiner a lot, if the last 20% of the process work without and against the influences of lora, ip-adapter and revision.

[shy kelp]

yeaa, any suggestion?

[rustic garnet]

because the refiner is a different network with different weights and slightly different architecture

[ionic gulch]

do you know what would be needed to change for ip adapter and revision to make them work in the refiner part of the workflow?

[rustic garnet]

you would have to train them again

you can't use them for the refiner for same reason as you cannot use an 1.5 controlnet on sdxl

[ionic gulch]

so ip adapter and revision also are also based on some training? not only models and lora?

[rustic garnet]

okay... maybe I don't understand what you mean with ip adapter and revision

[zinc cargo]

alos none vanilla models...

[ionic gulch]

https://huggingface.co/comfyanonymous/clip_vision_g/blob/main/clip_vision_g.safetensors
https://github.com/laksjdjf/IPAdapter-ComfyUI

[rustic garnet]

cool, didn't knew that one

but yes, the IPAdapter is trained on the unet of the model

[shy kelp]

do you know how to install the ip adapter

[ionic gulch]

for comfyui you need the nodes and https://huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/ip-adapter_sdxl.bin + https://huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/image_encoder/pytorch_model.bin
i already have it working in my workflow since yesterday. so the next update of my workflow (released in 10-20 hours) will include an ip adapter section.
the current version of my workflow already has a revision section: https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[uncut gull]

I'm using A1111. 1.5 and SDXL behave differently, for sure. But SDXL is better overall, for all things like chair legs that don't fade into the ground, animals with four legs, buildings with windows that don't turn into billboards halfway through, and of course, hands with four fingers and a thumb more than 50% of the time. Learning to prompt for it is worthwhile IMO, and hopefully we'll have the full suite of controlnets in A1111 soon.

[ionic dragon]

how did you create that, canny?

[shy kelp]

using canny node and controlnet

[ionic dragon]

ok

[pure hazel]

Can IP adapter be used in Auto1111?

[west breach]

superhero parade brought to you by wish.com

[uncut fiber]

dont know how to make hamburgers smaller. When not mentioned size it was even larger, now tried Palm sized hamburger, still same. But love the mood.

[upbeat summit]

I don't know either

SDXL just likes to make huge burgers

[uncut fiber]

and Hotdogs as well 🙂

[upbeat summit]

haha - nice

[shy kelp]

how to use this nodes

[ionic gulch]

connect to an image you want as image prompt, connect clip-vision to a loader, insert between model loader and ksampler

[pure hazel]

This would be awesome in comfy https://github.com/naver-ai/DenseDiffusion

[pure hazel]

Great question. Should post on the adapter github.

[west breach]

detail boost

[zinc cargo]

can you give whole workflow?

do you upscale size whilte you're taking this step?

[west breach]

no upscale

[native knot]

Like a rhinestone cowboy.

[uncut fiber]

Photographicographicographicographicographicographicographic
what a tag 😄
Thinking about Christmas.

[west breach]

christmas already??

[native knot]

Like a rhinestone alien.

[west breach]

I made a customer sampler that has the boost trick built in

[tepid maple]

🙂

[uncut fiber]

@west breach yes thinking because atmosphere, peaceful atmosphere 🙂 I come to that mood doing snowy images.

[dense chasm]

LSTM can produce midi music after training, long time memeory and compose music

one of my short videos vistors traffic just over 400k

according to my analysis,the vested interest and the social class solidification is a big issue,the young generation is struggle for surviving
the comments and emotional classification is another key point of AI research,it can help predict the tendency of stock market or social life ,ppl opinions on e-commerce site products or social network topic bias or whatever

[tepid maple]

rememebred the gem hunters show

I wish that show continued

[west breach]

also made an image caption node today

[vital ermine]

Interesting.

[native knot]

DPM++ 3M SDE GPU for the win.
Edit: karras > normal

[vital ermine]

Be aware the video started saying unipc, ddim, and PLMS can't be used with XL when that is true for the master branch on automatic1111 but 100% wrong for the dev branch of automatic1111 as well as ComfyUI.

[vital ermine]

Not even listed as an option on civit as I found out yesterday when I was filling in the info for my upload.

[native knot]

It's pretty new...but they really need to update to add that.

[uncut fiber]

i tested ddim on a1111 and it told me it cant work with sdxl, so in some time it was true

[vital ermine]

Yeah, I had to lie when asked about which one I used for the images

[vital ermine]

smh, get off that master branch

[uncut fiber]

i updated and it said something about RC, i got 1.5.2 but it say something about 1.6.0 so not sure. Will check it

[vital ermine]

it will say dev branch. You are 100% on the master branch which is bad

[uncut fiber]

yy

[vital ermine]

that branch caused me my first BSOD in years

[uncut fiber]

and how about devs and performance?

[vital ermine]

much better. not perfect but I was happier in it

they switched stuff around but I got used to it fast

[thorny frost]

i dont see the dpmpp 3 gpu in a1111

remote: Enumerating objects: 3, done.
remote: Counting objects: 100% (3/3), done.
remote: Total 3 (delta 2), reused 3 (delta 2), pack-reused 0
Unpacking objects: 100% (3/3), 353 bytes | 27.00 KiB/s, done.
From https://github.com/AUTOMATIC1111/stable-diffusion-webui
a6cedafb..4c678864 dev -> origin/dev
Updating a6cedafb..4c678864
Fast-forward
modules/ui_loadsave.py | 17 +++++++++++++----
1 file changed, 13 insertions(+), 4 deletions(-)
venv "K:\a1111xl\webui\venv\Scripts\Python.exe"
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: v1.6.0-RC-7-g4c678864
Commit hash: 4c6788644a367704d5dcb684a6c74a4ad1d1b078

Is that the developer branch? I think so

[dull atlas]

¯_(ツ)_/¯

[ionic dragon]

custom lora?

[zinc cargo]

say, can anyone tell me how to turn off previews from the ksampler efficent advanced?

[cyan crown]

hello, do any of you have a good json for kohja to train a style for sdxl with a 4070?

[thorny frost]

dpmpp_3m_sde_gpu. normal, karras, exponential, sgm, simple, ddim

[bronze arch]

https://github.com/Stability-AI/stability-ComfyUI-nodes This is the only node available that i can find

Where do i get this?

[thorny frost]

have you tried with the manager?

[bronze arch]

what manager

[thorny frost]

there is a special node calledmanager which will dowload the missing nodes automatically, among other things. look for it.

https://github.com/ltdrdata/ComfyUI-Manager

[zinc cargo]

well that didnt work right out of the box 😛

[bronze arch]

that worked thanks

[crisp owl]

change to None in Manager

[thorny frost]

it has many other functions. It is a must for comfyui.

btw Google Bard is pretty dumb, but can create some nice prompts

[crisp owl]

yeah bard is bad
But I imagine it's decent at prompts because it has internet access and can pull from other prompts formats now?

[uncut fiber]

like very much first picture, may(j)or tom i believe, and David Bowie isnt so great 🙂

[thorny frost]

possibly. At least he knows what a prompt is, apparently

For example: "A young woman sits on a bench in a park, reading a book. The sun is shining, and the birds are singing. The woman is wearing a long, flowing dress, and she has long, flowing hair. She is completely absorbed in her book, and she looks happy and peaceful."

A sunset over the ocean. The sky is ablaze with color, and the waves crash against the shore. The sun is setting below the horizon, and the sky is filled with shades of red, orange, and yellow. The scene is breathtaking, and it is a reminder of the beauty of nature.

A cityscape at night, with all the lights on. The skyscrapers are lit up, and the streets are full of activity. The city is alive and vibrant, and it is a symbol of progress and modernity.

A forest in the fall, with the leaves changing color. The trees are ablaze with color, and the air is crisp and cool. The forest is a place of peace and tranquility, and it is a reminder of the changing seasons.

[tepid maple]

ok found a great upscaler for myself test almsot all of them

SwinIR is the best if you have a decent cpu

like a good Ryzen

it doesnt KILL your ram or vram uses cpu more

perfect

[tepid maple]

model?

[soft zealot]

in the file name

[tepid maple]

i wish this model forks wrote their trained data list

every fork is too blind

[soft zealot]

Buts its 3 steps dynavision
13 steps NightVision
9 Steps Dynavision

[tepid maple]

9 step

hmm

[soft zealot]

do the maths, thats 25 total split across 3 passes

[tepid maple]

how much time one image takes ?

[soft zealot]

plus a further 5 wuth a 0.12% denoise onwith NightVision on the upscale

[soft zealot]

thsi is a batch of 10 random seeds

[thorny frost]

do you use some special node for that?

[soft zealot]

and another batch of 10 random seeds with a mosdified prompt

[soft zealot]

uses the method in @high skiff s workflow https://github.com/SytanSD/Sytan-SDXL-ComfyUI

although I have added an additional "preconditioning" stage which is the 3 initial steps

obvioulsy there was prior work on it, these are variations on a peompts I started posting yesterday

[steady grove]

what benefit are you trying to achieve with 3 different models on a single image?

[soft zealot]

you mean 2 different models on 3 passes?

its a variation on the OG Base+Refiner pass for SDXL

only difference is I'm using 2 different models

[thorny frost]

@soft zealot your workflow is extremely complex! i am trying it but lack the seed with text node. Manager cant find it.

[soft zealot]

If you scroll out and go to the highlighted black box it has a readme

[thorny frost]

ok, ty. looking for my glasses...

[soft zealot]

its wonderful what you can do with copy & paste into notepad and then increase the font size

FYI if you refer back to my overview

Green Boxes are the only ones you need to use ona daily basis
Purple you may occaisonally need to use
Blue you are unlikely to need to do anything with
Yellow is old Shit I've left ibn for reference/

As Isay though its only there because I dont believe in stripping workflow out and f you want to follow along you really need to start with something way simpler first

[steady grove]

ooooooor windows 11 ninja new awesome shortcuts. start + the actual + key. then start + esc to get out

[icy brook]

What types of models/LoRAs are sorely missing on the SDXL landscape?

Styles, people etc.

I am interesting in training to fill the void 🙂

[noble shoal]

Maybe something practical like user interface stuff. SD1.5 had an awesome Button lora. 🤷

[uncut fiber]

🙂

[crisp owl]

nice

[rugged crow]

hello stable people

[crisp owl]

^||unstable||

[crisp owl]

ooh those are neat

[noble shoal]

Thanks, but that are just photos from how i live.

[crisp owl]

better connected with nature than 99% of people then

[noble shoal]

Thats ture. Let me check if i can generate my PC setup 😜

[crisp owl]

I'm expecting potatoes

[uncut fiber]

Nice steps imho, and second bearded woman or another issue.

[noble shoal]

That was a good idea!

[crisp owl]

Perfect potato pc 👌🏻

[steady grove]

one step away from being a really good couch potatoe metaphor

[queen adder]

my NMKD superscale doesn't work with SDXL. is there a close equivalent upscale that's compatible? i really liked that one

[crisp owl]

I use
https://huggingface.co/uwg/upscaler/blob/main/ESRGAN/4x_NMKD-Siax_200k.pth

[noble shoal]

A Potato Couch?

[steady grove]

https://www.youtube.com/watch?v=3a9BO_BN0Ro

[queen adder]

hmmm this doesn't work for me either. must be the checkpoint of the SDXL model i'm using maybe.

[crisp owl]

Hmm yeah then something else is wonky inyour flow.
I use it on nearly 100% of my images

[strong field]

same

workflow can be simple as this

[uncut fiber]

and how it would look to not scale exactly 2x or 4x?

[strong field]

Look you say??!

[uncut fiber]

yes my bad english, i mean workflow how it would be

[strong field]

Haha, nah i was cracking a joke

[uncut fiber]

there is such think for say better to downscale, nearest neighb... bilinear and bicubic.

[strong field]

Yes, what are you trying to achieve, upscale the image by a set amount?

Or by a factor?

[uncut fiber]

yes

not sure what to do with 4096x4096 images 🙂

[strong field]

use this node, and downscale by .25

and then resample, with a decent denoise so you dont have artifacting

you making me give away my secret sauce 😛

here's an example

[uncut fiber]

This i know o.k. on other hand it spend too much time to upscale.
Thank you 🙂

[crisp owl]

Yup exactly that above.

I always just upscale by 0.5, to achieve a 2x upscale using a 4x model

[uncut fiber]

Thick paint 🙂

[stone fossil]

Thick cracking paint.

[uncut fiber]

Now comes pseudo artsy images 🙂

[marsh ether]

cat made of liquid light 😛

[icy brook]

Very nice heavy use of palette knife 👌

[marsh ether]

My character from old dnd campaign, tho I couldn't get eyes blue

[uncut fiber]

Does anybody knows what is benefit of using upscale latent space by... ? Tested it and same results, and same performance with it or without it.

[crisp owl]

It's picky based on the sampler methods, so I tend not to even bother with it.

perhaps there's reasons to use it and it's best for, but eh, I currently couldn't be bothered with it.

[ionic gulch]

JPS All-In-One-Workflow 5.0 with IP Adapter, Revision (Clip Vision), TXT2IMG, IMG2IMG, Control Net Canny, Control Net Depth) now on GitHub:
https://user-images.githubusercontent.com/142158778/263359034-b000458e-9635-4ab4-97dd-4d0b0cac57f2.png
https://user-images.githubusercontent.com/142158778/263359221-86572383-f3a9-42a5-8fce-8122de7a6133.png
https://github.com/JPS-GER/JPS-ComfyUI-Workflows/tree/main

[tepid maple]

prumting

prumcant

[tepid maple]

why it is s ohard to prompt this office

[tepid maple]

all of the models halucinating crazy when it comes to office

[uncut fiber]

tar Puma? on white skin. Owners must be happy 🙂

[tepid maple]

so many twins so many weird looks

[uncut fiber]

how about those @tepid maple

[tepid maple]

pandora ?

where avatar people

[vital ermine]

Hehehehe

[uncut fiber]

that second from top @vital ermine too high cfg?