✨｜sdxl

[trim orbit]

/leaves to pour bleach onto my internal vision

[ionic dragon]

dude

[soft zealot]

you asked, I answerd

[indigo carbon]

yeah, bitches eating shit

[boreal bough]

wth XD
found a prompt my lora is very very unhappy with
girl turns into an amanita muscaria mushroom, raw photo, 8k
without/with lora

[trim orbit]

partially jk ||but any mention of such horrors should beckon a permaban|| but jk

[soft zealot]

so theres an issue about two girls sat around a table sharing one cup of tea?

Apoologies but as a good friend of mine uused to say "Its not the mind it comes out of it's the mind it goes into"

If you choose to interpret 4 innocent words completely differently to me then nothig I can do about that 🙂

[trim orbit]

yes my mind is quite damaged by the early years of the internet. i cna't deny this. i'm a product of the world.

but you knew what you intended too so 😛

[soft zealot]

prove it 🙂

[ionic dragon]

i was seriously asking

[trim orbit]

i dont have to. you brought it up. now i'll forever consider you a poop fetishist

[indigo carbon]

I have some bad news for pretty much every time you drank tea

[autumn forum]

Faces lmao

[soft zealot]

that was close to a typo

sand

[autumn forum]

What if it was a typo😉

Lmao jk!

[trim orbit]

you know like that guy who cuts one inappropriate joke about hitler and then everyone considers him that guy forever? is it eveyrone else's fault for him cutting that joke? naw

/godwins law

[trim orbit]

if i can answer the question, it's not as good at anything v3 at creating creative mechas

on 1.5 models , especially the anime ones, i could create delorean mecha

Haven't caught one from xl yet

[autumn forum]

From my experience with sdxl it is not good at faces at a distance, nor is it very good at eyes at all. There’s always a bad one. Wheels on cars are still kind of an issue but it’s definitely not bad.

[ionic dragon]

still?
oh, thats something which need a model, so i dont have much computing power yet i have the dataset sadly

[trim orbit]

bad faces are usually solveable with hires fixes, stronger prompting, and detailer passes

[ionic dragon]

which v3?
what are mechas?

[indigo carbon]

if you have old school type issues with SDXL this means you likely didn't infer it properly

[trim orbit]

anything v3. it's the quintessential anime model imo. anything v5 is good too. anything v4 was just a fake knockoff merge from a stooge who didn't like v3's popularity

[trim orbit]

mecha's are like, giant robot mechs from japanese anime

[autumn forum]

I get good results but it requires more than what the base and refiner has to offer. Sdxl is very very good. Not saying it’s bad but can be better. Heck i don’t even use the refiner for most of my gens

[ionic dragon]

i didnt get what versions you are referring to

[indigo carbon]

OK WTF, that's an apple logo

[trim orbit]

the name of the model is "anything"

whose on first

[ionic dragon]

oh, i think i got few

[ionic dragon]

so basically v3, is like finetunings?

[trim orbit]

yeah it can do them for sure. but not creatively like "delorean mecha" or "1977 honda civic model mecha"

[ionic dragon]

@trim orbit do you have lora settings to train a face, got bad results from 3 trainings

[ionic dragon]

i can train a lora, if you want

[trim orbit]

it's a sd1.4 based model that is extremely specialized to anime

[ashen oracle]

@indigo carbon i ran your workflow but for some reason can't find the output image lol...

[trim orbit]

i dont need them. i just came across that case that sdxl couldn't do. I always generate deloreans on every model. delorean mechas too

[indigo carbon]

you are likely missing required nodes. it's pretty self explanatory

[ashen oracle]

it runs / completes and node manager says i have everything - just not seeing anything in the actual Output folder which is strange

[indigo carbon]

oh, I know what's happening 😄 , you don't have the upscaler I use, it saves the image after upscaling

[ashen oracle]

i used my upscaler in place of it though

maybe breaking something?

[indigo carbon]

likely. it should work perfectly fine.

anyways, who said SDXL can't do mechs?

[trim orbit]

creative mechs. like delorean mecha

[indigo carbon]

the car?

[ionic dragon]

if sdxl is lacking that, i can train, but tryna find params before i kickstart training diff loras

[soft zealot]

@boreal bough Inial impression is good (ignore the hands I was just using my usual test Seed lol)

Left to Right

No Lora, Defaulat SAI Offsest LORA, your LORA

Loras at 0.6 strength

[trim orbit]

it can do both mech AND it can do deloreans, very well

i'm not sure how you'd train a lora for better zero shot capabilities here

[indigo carbon]

isn't this that car? idk what a delorean is, but this is what it makes from "delorean mecha"

[ionic dragon]

so then what is it lacking exactly

[ionic dragon]

i gotta get some exp

[trim orbit]

i donwloaded anything v5 because i didn't have them anymore. here's a result. it often does them like a silver haired man too. i think because its pulling info about mr delorean's portrait hahaha.

[trim orbit]

hard to say. zero shot means it can generate something it's never seen. while it's very great at zero shotting photos like Furkan has shown, riding a dinosaur and a lion, it's also limited when turning cars into robots

it might be because cars are very over fit

[ionic dragon]

i think, it wasnt trained on a great dataset of transformers

[trim orbit]

transformers and mecha are slightly different aesthetics but yeah, maybe that'll be a better approach. prompting is king after all.

i'm firing this motif off on "psy animated" now and its actually getting some better results than i had with the base

the base would often draw a mecha AND a delorean, entirely separate

similar situation but the mech is much more aesthetically a delorean now

the more i use psy animated the more i like it

[boreal bough]

forgive me, for I have prompt sinned

P: abstract expressionism art,(ahegao:1.2),(a 29 yo woman, (Zooey Deschanel:0.2), (redhead:1.1), cute smile, detailed face, perfect eyes),neon,hdr10,they are watching into glowing wires everywhere,transformers,woman as a transformer,megatron prime

N: monochrome error,multiple views,blurry,uncloned face,disconnected limbs,fused limbs,long body,long legs,body,fused limbs,long legs,missing limb,missing arm,no arms,short limb,nonsense,huge face,bad eyebrows,ugly beard,mutated,

[trim orbit]

why zoey would ever be a 0.2 is beyond me. this is the greatest sin

[trim orbit]

oh til. it actually is zooey

[soft zealot]

for appearing in the crap remake of HHGTTG in 2005

[crisp owl]

Tried your flow as is TD and it didn't like me

[ionic dragon]

As for lora we only use few images, so we can caption using minigpt4

[indigo carbon]

CMD log?

[trim orbit]

i loved the movie and feel it's the best version it could've ever been of the first book

[crisp owl]

same error as in the UI

[soft zealot]

I used to have it on vinyl 🙂

[trim orbit]

bombed in boxoffice though so i dn't argue that with people who insist it's bad. i just keep on enjoying it

[trim orbit]

the radio show? oh god. now that's bad

[indigo carbon]

that's probably not true. that type of error usually shows like "imput:0 error" or something like that. if it's still broken for you you can try manually placing the AIT modules.

[boreal bough]

close enough?
a (transformer:1.4) Autobots in the shape of the delorian with a woman sitting inside it, 29 yo old, Zooey Deschanel, Autobot

[trim orbit]

hah . that's the kind of results i get a lot of the time. a delorean next to a robot.

cool delorean!

[indigo carbon]

oh wait, I see what's up. you're on SM75, aren't you?

[boreal bough]

I'll get there! XD I'm already getting closer

[trim orbit]

prompting is king! i believe in you.

[soft zealot]

front end looks more Fox Bodied Mustang than deloreac

[trim orbit]

that one is more of a firebird tho

[trim orbit]

i agree. it's blending models

[boreal bough]

and they believe as well!

closest I'll get with zooey hogging all the latent space XD

[trim orbit]

she really does have commanding screen presence

chonk delorean

so i was wrong

sdxl through psy animated has been getting good hits

[boreal bough]

without her its much more doable ^^'

[indigo carbon]

did you try DSXL or CCXL?

[maiden gale]

These are bomb!

[crisp owl]

Shouldn't be reloading the checkpoint again, but it will have to run through your text encoders again.

[soft bone]

which should be fast. show your workflow

[crisp owl]

Unsure regarding that

[elfin cobalt]

How much memory do you have?

[indigo carbon]

if your GPU gen is under 3000 series, it's not supported yet for AIT. however, in a few days all the modules will be ready.

[elfin cobalt]

No, main memory

[soft bone]

yes that should be fine

[elfin cobalt]

It does not.

How much memory do you have?

[crisp owl]

Ah gotcha, 2060 so yeah. I'll keep an eye on it

[elfin cobalt]

Like I said, main memory.

Not graphics memory.

[glad flint]

I think when you change the prompt it has to redo the the CLIP encoding or something
I know trying to use wildcards in my prompt was KILLING my performance, lol

[elfin cobalt]

That isn’t enough.

You need about 24, minimum, to get full performance

[glad flint]

I have 16 GB of RAM 😄

[soft zealot]

but will it support a 1080Ti ;o)

[elfin cobalt]

ComfyUI sits at 18GB memory use for me

It might simply be dropping unused data in your case, but then it has to reload it to compute new embeddings.

[glad flint]

im about to run a prompt so I won't be able to type for a bit~~ lol

[elfin cobalt]

I think it’s likely. And RAM is cheap these days anyway.

Also, are you looking at its VRAM usage or RAM usage?

Both matter, but obviously they’re not the same.

Depends on your motherboard.

Do you have 2x8 right now?

4x8 should work, but you might lose some performance.

Should try to make it the same model though.

DDR3200.

Match the model if possible, match the speeds and clock rates if not.

[vale eagle]

@indigo carbon Using your workflow. Want to replicate an anime robot but the lion head in the wrong place.

[elfin cobalt]

Well, yes. DDR 4 and 5 are completely different.

On the bright side, DDR4 is extra cheap. :p

At what speed?

Ok. Well. I can’t guarantee it’ll work, but worst case if you reduce the speed it should.

Running four modules at your current speed.

No, because 3200 is an oversimplification and really they need to match impedance and clock speeds in nanosecond regimes, on wires that are so long relative to the frequency that they act as antennas, not wires.

[soft zealot]

but unless you're stats watching it will probably be fine

[elfin cobalt]

4 more than doubles the chance of problems.

No.

All desktop motherboards are dual-channel, so 2 is optimal. They handle the 2 dimms mostly independently.

With 1, you actually lose 50% of your memory bandwidth.

If you have infinite money? Sure.

Otherwise do 2 or 4.

1 and 3 are both bad.

[soft zealot]

I run 4 x 16Gb Crucial Ballistix BL2K16G30C15U4B and have done for the lat 18 months without any issues

[late marsh]

I want infinite money too thank you

Oh nice, is it like downloading more VRAM?

[soft zealot]

system is stable and plenty fast enough

[elfin cobalt]

As I’d expect. Matching all four usually works just fine.

[indigo carbon]

yes, theoretically you can alter the node's code and compile the module yourself, but that's insanely complicated

[elfin cobalt]

Having four modules sometimes causes trouble, but usually not. Really the issue is going from two to four… because you’re likely to not have four identical ones.

Yeah.

But also.

Just about every combination works fine if you leave it in the failsafe configuration, at 1200 MHz.

[indigo carbon]

it already can, but the older gen your cuda cores are, the less boost AIT will cause.

[elfin cobalt]

Which is the default.

Unfortunately a lot of people buy fast memory, then never flip it to high speed. :p

[soft zealot]

just go all in and grab 2x32Gb sticks, its cheap enough. Pop your existing stuff in the drawer just in case

[indigo carbon]

nice! does the AIT node work for you? many people report issues with old GPUs with it

[vale eagle]

No. Has error while using the node

[indigo carbon]

GPU?

[elfin cobalt]

There’s an AIT node? Where?

[indigo carbon]

custom node, experiential and I seem to be the only one on windows it works for =\

[elfin cobalt]

Maybe? But for DDR5 you’d also need to swap your motherboard, and maybe CPU too.

[soft zealot]

my bad those are SODIMMS

[elfin cobalt]

Hit me.

I’m on Linux anyway.

[soft zealot]

https://pcpartpicker.com/product/nht9TW/neo-forza-faye-64-gb-2-x-32-gb-ddr4-3000-cl15-memory-nmud432f82-3000dg20

[elfin cobalt]

Either will work. 3000 is maybe 2% slower in practice.

[soft zealot]

unless you're chasing numbers just take the 2 old sticks out and put these two in

it will be fine

you will notice the benfit and be able to have even more chrome tabs open whilst generating images

[elfin cobalt]

Hahahahaha…. … yes.

[wet glacier]

What’s the minimum ram quantity requirement?

[indigo carbon]

oh nice. you should be fine. just load the workflow from one of my gens and restore nodes it says you don't have. you shouldn't need to worry about compiling the modules yourself as long as your GPU is 3000 series or above. if it is a newer gen it should handle the modules on it's own

[soft zealot]

0

[elfin cobalt]

I mean, we’ve seen it work on 4GB…

[soft zealot]

because the first one was SODIMMS for latops

Generalised list of DDR4

https://pcpartpicker.com/products/memory/#Z=65536002&sort=price&ff=ddr4

[wet glacier]

I have 256gb @3600

[elfin cobalt]

Don’t but something no-name.

Flaky memory is a real nightmare.

[soft zealot]

depends on how much of a snob and a number chaser you are

[elfin cobalt]

I’d get one of the crucial kits.

[indigo carbon]

ddr5 is more efficient if you want speed

[elfin cobalt]

That one should be fine, sure.

[indigo carbon]

yeah, who needs fancy rgb lights on their ram? nobody lol

[hardy cipher]

pip cache purge
Files removed: 2009

lol

those things add up

[elfin cobalt]

Heh. That’s why I use bcachefs.

Tiered storage is great.

[hardy cipher]

hmm. I'll have to check that out

pip really does enjoy just eating up storage space

[elfin cobalt]

TBH, I’ve had data corruption bugs. Don’t use it unless you know very well what you’re doing.

You can just let me debug those first. It is remarkably good otherwise though—I’m genuinely getting SSD speeds with HDD storage capacities.

[hardy cipher]

my issue is I'll sometimes realize a pip install is 100 gb or something, and my primary drive is a bit on the small side

just purged 25 gb

[elfin cobalt]

Ah. Yeah. I just have everything in a single giant pool.

[hardy cipher]

so one day I'll have 160 gb free space then install a couple things and it's 20 somehow

[elfin cobalt]

Two NVMEs and two HDDs. Mirroring, of course.

[hardy cipher]

well I store a lot of things on hdd and backup in cloud storage. but realized loading sdxl models from my hdd takes FOREVER

like 5+ minutes

[elfin cobalt]

Mm. Wir bcachefs, it writes files to the SSDs on access.

So you still need a decently sized SSD, but it’ll automatically delete the oldest cache elements when that fills up.

In practice, it just means that older stuff I don’t use ends up in the HDDs, and I don’t need to think about what is where.

…also writes start on the SSD, so that’s fast as well.

[indigo carbon]

this chat turned into LTT lol

[hardy cipher]

not sure what LTT is but it sounds like it must be pretty cool

[hasty obsidian]

Maybe just install One Button Prompt in ComfyUI? It's available as an extension in the extension manager of ComfyUI.

[hardy cipher]

there's literally a one button node right? or am I crazy

[upbeat summit]

well hardware is a big part of this - at least when you run it locally heh

[indigo carbon]

true, true

[upbeat summit]

nice background (and foreground) 🙂

[hardy cipher]

thanks, it's weird because all the other generations in that sequence were faces.

[hardy cipher]

did you hit tha twith the face detailer?

[upbeat summit]

yeah well, sometimes you hit a different part of the latent space while exploring it 😉

[hardy cipher]

yes, liminal latent space

[indigo carbon]

isn't the backrooms a liminal space?

[hardy cipher]

yes

[hardy cipher]

I have a backrooms lora for 1.5

[upbeat summit]

tonight - join picturesonpictures on a visit to the backrooms of the latent space

[hardy cipher]

ooh, maybe I should combine USSR propaganda lora with the "truck" lora

soviet truck propaganda

[hardy cipher]

that's a top tier lava lamp

[vale eagle]

@indigo carbon I think they are updating something recently on the AITemplate node repo. It broke something. Wait until they are ready for it.

[indigo carbon]

yes, it works with base SDXL models on SM80 gpus, they said refiner support is coming very soon.

@visual glade said he is working AIT support without any custom nodes, but that was a few weeks ago

[keen cobalt]

do yall know the newsest engine name for api usage

[somber hill]

adding chapters to hopefully coming today my new massive tutorial

[visual glade]

yes but I'm a bit busy in LA right now

[indigo carbon]

yeah it's fine. the custom node (kinda) figured it out, well, for the base atleast

[soft bone]

i need a female photo reg dataset. anyone got one?

[trim orbit]

i am very liking the psy animated xl today. more deloreans

[karmic dew]

Hey folks I’m seeing a few issues with image quality in ClipDrop. Is this the best place to talk to someone about it?

[cyan crown]

in 1h I should finish my last video! WIll post here (Made with SDXL and MJ)

[slender coral]

Comfyui does queing up prompts use the prompts AT the time you quede it up or at the time the prompt is running?

[soft bone]

uses the params at the time you queue it. feel free to change params after queuing

[crisp owl]

Yeah everything is AT the time. Including the timestamp of image creation, so if you use that parameter in your image filename, it will be the stamp when you hit generate, not when the file completes

[keen cobalt]

stable-diffusion-v1 is this the latest stable diffusion engine, or is there a new one i can use for api

[soft bone]

does "shuffle caption" shuffle words within the caption or shuffle captions between images? and if there are no caption files then this setting shouldnt do anything?

[grim hedge]

I have trouble even loading the model. I have a 15gb shared gpu memory and 8.5gb individual. But I can't even get it started somehow 90% of it is dedicated somewhere else.

[autumn forum]

Pretty sure shuffles words within the caption

Shuffling caption between images doesn’t make sense lol

[arctic bloom]

why do you think 15gb of shared gpu memory should be used?

[grim hedge]

Well it just doesn't load

That's why

[soft bone]

in this case, if there are no caption files and the "caption" is the folder name, will it shuffle the folder name words?

[grim hedge]

How much Vram do you need to even start the sdxl model?

[autumn forum]

Doubt it lol

[soft bone]

more than 8gb by default im pretty sure

im sure theres optimizations

some ppl are running cpu only

[crisp owl]

Running in ComfyUI just fine, 2060 Super with 8gb

[arctic bloom]

what do you use?

[arctic bloom]

https://tenor.com/view/danny-devito-danny-nope-danny-no-gif-23264539

[glad grove]

runnin sdxl on dual core apu

[lilac raven]

I think my second attempt at LoRA (the first [supposedly] diverse enough to be character LoRA) has failed. It seems to have too much effect and any attempt of photorealistic, 3d render or photo fails either by still generating 2d sprites or the character is near completely lost. Beyond that, I have to add the details of the character in the prompt, which I think I shouldn't have to (the 2d sprites are not that bad, actually)

[high skiff]

Anybody else having an absolutely terrible experience with training LoRA's for SDXL as of late?

No matter what I do, all of my LoRA's are trash now

[cyan crown]

problem is making lora or using them?

using is fine. Making is horrible. I just can't obtain decent result

[high skiff]

making is suddenly near impossible

I made a few just fine before

now, no matter what settings I use, no matter what captions or dataset, it learns nothing

[cyan crown]

for me it learns wrong in hours 😄

[high skiff]

I have had @boreal bough and other extremely smart people when it comes to LORA's help me, and yet I am getting nothing, and I mean NOTHING

my LoRA's are basically identical to base SDXL

full hour of training on a 3090, loss is identical no matter the settings, LoRA is identical to SDXL

[cyan crown]

loss for mine tests is 0.15 that is too high

[crisp owl]

try training something absolutely off the wall, something which would be extremely easily able to tell if it's doing anything at all. That should tell you if it's doing anything

[high skiff]

this is 2 different training tests with different learning rates, different DIM, and different alpha balues

[soft bone]

what are you training

[high skiff]

A portrait photograph LoRA

my first test had some results, but the effect was very undertrained

I got a bigger dataset and spent way more time captioning, and now my LoRA's are identical to just stock SDXL

[cyan crown]

which level of loss do you end with?

0.?

[high skiff]

its different from run to run in the end, but it never matters

SDXL is just refusing to learn anything now

everything is either stagnant/inconsistent/increasing loss no matter what I do

[soft bone]

ah. see i have a suspicion that SAI was very concious about portrait photography and gave it excellent representation. Without a stupidly large finetune, I don't see XL getting better at portrait photography in general. You could weight it towards a certain photography style or camera setting/effect though.

[high skiff]

it has a long way to go

even my super fast 20 image 5 epoch LoRA already suggested some huge improvements

[autumn forum]

Sytan in all of your tests did you use adamw8bit?

[high skiff]

I used adamw8bit and prodigy

both are horrifically unsuccessful

[autumn forum]

Use adamw or adafactor

[cyan crown]

for me it learn, but not in the way I want. Images are too far from the training ones (style train adafactor)

[high skiff]

if prodigy didn't work, likely nothing will. Its made to adapt as best as possible in real time

[autumn forum]

At least try Sytan

[soft bone]

agreed, but i dont think a lora is gonna get it there. The Dreamshaper finetune has thousands of portraits handpicked over 6+ months, yet it still only barely improves portraits

[high skiff]

even with prodigy, it manages to have a consistently raising Loss

[high skiff]

Thats cause the dreamshaper XL finetune is kinda trash

and by kinda I mean like... very

[clever verge]

What scripts are you using for LoRA training?

[high skiff]

Kohya

[cyan crown]

me too

[clever verge]

Anyone tried diffusers?

[cyan crown]

nut it just doesn't work 😦

[soft bone]

We decided in this channel that dreamshaper is def the winner for portrait photos out of all finetunes currently available

[high skiff]

A friend has, I am just not willing to mess with diffusers ATM

[autumn forum]

If it’s doing nothing try to make it do something. Like adamw 1 epoch high learn rate

[high skiff]

any proof of that?

I'd love to see some images

cause I am pretty confident my base SDXL images will beat them

[high skiff]

nothing

I have now ran 16 different LoRA tests

all of them end the same one way or another

[soft bone]

i said out of all finetunes available. base not included

i'll find it one sec

[high skiff]

start strong, after a few steps, rapidly raise to 0.12 loss, and stay there, no matter how long you give

Either that, or they go up to like .14

[clever verge]

I've been tinkering with inference with diffusers. building a lora environment as we speak.

[high skiff]

If a finetune is worse than the base, then I am not sure why the finetune would be suggested, personally

unless I am misunderstanding

All I know is from what I have seen of dream shaper XL, it was a super rushed and poorly executed finetune

They should have given it more time to be good, rather than rushed to be "first"

[clever verge]

https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora_sdxl.py

Not sure if it's good, but I will find out.

[cyan crown]

there are good Lora that works, I would like to know how did they do.....

[high skiff]

I know there are, I have made some

Its just... all of a sudden, LoRA's will not train for me

[soft bone]

@high skiff #✨｜sdxl message

5 is dreamshaper. unanimous vote. yes there are other factors to consider but for photorealism portraits on any seed dreamshaper was doing the most real lighting, and most dynamic environments and outfits

[clever verge]

Here's a workflow using it https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/SDXL_DreamBooth_LoRA_.ipynb

[cyan crown]

this is strange

I never been able with sdxl

[high skiff]

ah yeah, looks exactly how I expected, very plastic

[soft bone]

they all are

[high skiff]

Its a cool style, sure, but its not realistic at all

[soft bone]

xl is more plastic

[high skiff]

base SDXL can do realistic just fine if you know what you are doing

[clever verge]

There's another too: https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/README_sdxl.md

[soft bone]

and so can dreamshaper lol..... base XL is in that comparison on the exact same seed, yet dreamshaper is LESS plastic

[high skiff]

thats because whoever did the comparison doesn't know how to prompt base, IDK what else to say

[soft bone]

its the same prompt on base and dreamshaper. if the prompt caused the plastic in the base then it would cause even more plastic in dreamshaper if dreamshaper was more plastic than base

scientific process

[clever verge]

The documentation is sparse

[high skiff]

that makes no sense, there is nothing "scientific" about that

They weren't finetuned on the saem data, for all you know the way to prompt realism for one is way different than the other

Maybe sometime I will download dreamshaper XL and run some proper realism prompts through it

maybe then I would have a better feel for how good it works

[soft bone]

Before that guy's test, I also thought the same, i remembered dreamshaper and realismengine and all of those looking waxy on 1.5 and 2.1. Yet due to that comparison I tried it and boosted realism across ALL of my generations. very surprised and pleased

mostly the benefit is the boost to coherent details especially in clothing

[high skiff]

I still have a hard time believing it will be any better. Base SDXL can look leagues better than all of those with some proper prompting, but who knows, maybe dreamshaper can look even better-er lol

all I know is that SDXL base with proper prompting blows MJ out of the water for portraits

[cyan crown]

is there any way to do textual inversion train?

[high skiff]

I believe Kohya can do it now, but don't quote me on that 😅

[soft bone]

base vs. dreamshaper.

This is a horrible comparison for many reasons, but shows the gist of added color, dynamics, less artifacts

[high skiff]

those are a little lest fake looking, I'll give you that one

[soft bone]

it compliments the lora, helps make up for what it damages about the model in training

[high skiff]

this is the difference my first poorly traied very small dataset LoRA makes for portraits

no LoRA

LoRA on the same seed with the same prompt

and its very poorly made

my only confusion now is why it can't train anymore

I should know, these are not the portraits I am referring to when I say that SDXL is exceptionally good at portrait realism

This was just a very quick prompt that I threw together without any real focus on realistic image prompting

Just wanted to see quick results

It trained in like 15 minutes, so it's definitely very undertrained, and the data set that I used was very small and limited, and also had pretty bad quality images in it

My new data set is multiple times bigger, and monumentally higher quality

[ashen oracle]

anyone have a suggestion for a good image to image workflow?

[soft bone]

and you've scaled the epochs up with the dataset?

[high skiff]

No, I replaced the data set as it introduced a shit ton of issues because of the low quality images I originally sourced

I ended up finding a really good repository for the images I needed, so I was able to ditch the low quality images I originally trained on

I went from a 24 image data set where almost all of the images were quite low quality, to a 64 image data set where basically all of the images are near or above 4K resolution

Although the LoRA is training at 1024x

[cyan crown]

I share my last work done with SDXL and MJ

https://www.youtube.com/watch?v=vdZ9exqDFTc

[high skiff]

It's honestly got me very curious what would happen if I use the same data set on a 1.5 model

And still training at 1024x

Even the 64 image data set is still a relatively small scale test

If at some point I can manage to unfuck it, I would want to do a full version of around at least 400 images

[soft bone]

i'd just raise learning rate or steps until i see a change. at my learning rate i would need 4+ hours of training for 64 imgs, but thats for faces

[high skiff]

It's not my learning rate that's causing issues, something else is fundamentally messed up

I've tried to range of learning rates from 0.0001 all the way up to 0.1

They all do the exact same thing, start off strong, then after a few steps go up to 0.12 loss, and then either stay there or continuously rise for the rest of the training

I guess I can try installing derrian again

[soft bone]

you're one of the only people i know that looks at loss graph. whether it looks good at the end is subjective and idc if the math says i did it wrong if i like how it looks.

but it sounds like youre not seeing the change either

[clever verge]

https://huggingface.co/blog/dreambooth

[fathom hound]

I need a Seinfeld XL Lora

[soft bone]

that was right after the original dreambooth paper (maybe even along with it), super outdated now

[high skiff]

The lost graph is not inherently indicative of training performance, however, it should generally at least go down in some capacity as it slowly learns the patterns of the images that you train

For example, my first successful LoRA I made for SDXL used 23 images, and it trained for 20 epox at 20 repeats per image, and it started at 0.15 loss, and ended at 0.06 loss

And I retrain that one a couple different times to test out some values, and no matter what values I changed, it's still at least consistently went down

Whereas a lot of the tests that I'm running currently start at like 0.105 loss, and then by the time I'm done it's at 0.135 or higher

So it's quite literally getting worse at recreating the images as it goes along, which I just don't understand

[soft bone]

that paper is what taught us to freeze the text encoder for improved quality

[clever verge]

Thanks. This one? https://arxiv.org/abs/2208.12242

[soft bone]

i believe so

afaik no one is using dreambooth anymore. it was just a method of using regularization for training but everyone stopped using reg so it was essentially just regular training like in the original Stable Diffusion paper. and then loras got good.

dreamboothing (finetuning) SDXL is pretty unrealistic for consumers, we're mostly stuck with loras for now

[high skiff]

I use reg images for LoRA's sometimes, they can be massively helpful

but they are hard to get right

[soft bone]

i just started using them again, specifically ground truth (real) photos. there's no doubt that it helps with training faces

[high skiff]

I would personally say don't use real images, as thats not what reg images are for, but hey, if it works for you, do you

[soft bone]

see now that right there has been a massive debate since the paper posted above. in the end I'm going to believe the creators of Everydream2, who stick to real images

[high skiff]

to each their own

[clever verge]

How do you go about captioning your images?

[soft bone]

if i come across a dataset of 600+ person images made by XL, i'll do a comparison

[soft bone]

i stopped captioning. works just fine for people on XL

for anything other than people I'd recommend captions

i just used blip but that's far from the best

but the captioning tool in kohya is nice

[clever verge]

As if the model is better at understanding what it looks at already? Is that how it works?

[hexed hatch]

Is there a way to only import prompt and lora from a image instead of everything in comfyui?

[soft bone]

i assume so. a person isnt a very complicated concept for it to get. it makes guesses if you dont caption, and if the only thing in the image is a person then it's easy for it to guess what's happening.

completely theoretical tho

[clever verge]

SDXL seems to be well trained so if the assumption is true it would make sense.

[soft bone]

I have yet to do a direct comparison of captions vs. no captions but my results without captions are very good so idc

[autumn forum]

same... should i make one?

[fathom hound]

Haha, if you want to, go for it! I'm honestly, thinking of making a George lora once I am done working on other projects.

[mellow tendon]

Booru Dataset Tag Manager:https://github.com/starik222/BooruDatasetTagManager
speeds up captioning edits after Blip.

[autumn forum]

we need a lora of kramer doing all his weird poses haha.

[fleet harness]

I'll be you TAILOR

[runic lark]

Do you know if it's possible to change only model path to another drive in comfyui? I change checkpoints every half hour and it would be nice to have them on my SSD

[heady vale]

extra_model_paths.yaml.example
rename this file to .yaml and edit it in a txt editor.
add your checkpoint folder on the appropriate line

[hexed hatch]

any controlnet for 1111 sdxl out yet? I know there is for comfyui but I prefer 1111

[glad grove]

not yet

[jovial arch]

Hi. What learning rate was used for pretraining SDXL?

It's really difficult finding a technical writeup for the model online

[void skiff]

When I got SDXL for my Automatic1111 there was a file called lora:offset_0.2, I download it at the same time but then never tested it until today, what is it supposed to do? It keep my image similar at very low values like 0.002, but for ever step up the images look worse and more crazy.

[indigo carbon]

SPEED

(this is with AIT implemented in just the base)

[cedar tusk]

how's that work?

[indigo carbon]

magic =]

[upbeat summit]

so fast!

[heady vale]

looks so damn fast, and with 67 steps

[upbeat summit]

how did you integrate AITemplate, what are the limitations and what do you need to show us how to do it? 😉

[indigo carbon]

this is not simple. AITemplate isn't fully integrated yet (don't pressure me, I didn't make the module), by changing the code of the custom node I managed to get AIT to work on the base, the creators of the node claim to release a more stable and fully implemented version of AIT soon.

[cedar tusk]

what is AIT and is it hardware-specific?

[indigo carbon]

more complex than that. it's a self optimized optimization that currently gives an insane boost to newer cards

[upbeat summit]

Thanks for the info! I'm kidding! Not pressuring you, man. Just amazed and of course super interested. Who wouldn't want to utilize their hardware better to explore this thing.

The first time I read about it a couple of months ago, I was hoping it would come out at some point, since it looks so very promising - like the jump we made with xformers and also 4000 cards would finally get a real speed boost.

So I'm just really excited seeing it in your setup! Awesome!

[indigo carbon]

keep in mind that what I demonstrated isn't the fastest possibility. in the workflow I made (in the video) I only figured out how to implement it at base stage, so just wait until the refiner will also be implemented, the speed will get even faster. I bet with an RTX4090 you might get INSTANT image generation like that

[upbeat summit]

https://tenor.com/view/parraque-infarto-fainted-gif-16861591

[hearty ginkgo]

What graphics card do you have now?

[upbeat summit]

thanks for showing it off, Tdg8uU. if that doesn't get people excited... I'm only using base right now anyway 😄

[indigo carbon]

4070ti

[autumn forum]

would aitemplate work on a 3090? lol

[cedar tusk]

I'm happy my 2070 can do SDXL at all, so i'm not too saddened that this new witchcraft might pass me by for now

[compact obsidian]

does anyone know why SDXL has the scheduler config "steps_offset": 1

[indigo carbon]

yes, boost might be lower than for 4000 series, but because 3000 series is also SM80, you should be good

[autumn forum]

nice. where might i be able to look for an update on that for when it comes out for comfy?

[upbeat summit]

it will be all over this channel (and the internet) once it's easier to implement - I think you can be sure of that 😄

[indigo carbon]

there are currently 2 pending releases people are waiting for. one is to wait for the custom AIT node to support the refiner, the other is wait for @visual glade to officially implement it into ComfyUI

for now the custom node seems to only work on base

[heady vale]

what about tools to convert the checkpoint?

[autumn forum]

okay ill wait for comfy to impletement. im sure ill hear about that. at that speed everyone will know about it lol.

[indigo carbon]

that's a story for another day. both the custom AIT node and Comfy claimed to provide precompiled modules for different architectures. AIT is NOT checkpoint specific, it's architecture specific.

[glad grove]

is it possible to make it work on pascal cards or only works on RTX?

[indigo carbon]

wasn't tested yet, theoretically it is possible. I'm sure when Comfy will officially implement it, it should work on new cards (including AMD cards) however, older gen cards are unlikely to get any boost from this whatsoever.

[indigo carbon]

the compiling tools themselves are incredibly difficult to use, therefor in the node and (assumingly) in the official Comfy version it will do that on it's own.

[autumn forum]

i have another question then, does it use more vram/ram or is it rougly the same or less?

[indigo carbon]

but well, after you got all that set up, the boost is totally worth it IMO, I'm literally generating extremely high quality images in about 10 seconds.. (without it even being implemented in the refiner)

[indigo carbon]

a little less

[autumn forum]

oo nice!

[somber hill]

spent like total 6 days for recording + editing 😄 https://youtu.be/sBFGitIvD2A

[tender timber]

Has anyone got an aesthetic scorer to work in Comfy? I tried both **ComfyUI_Strimmlarns_aesthetic_score **and **ImageReward **and neither installs, it seems I have dependency errors

[slender coral]

any quick way to save uuid file names in comfy?

[hardy cipher]

I had image reward working but then if conflicted with something

[tender timber]

Thanks, I've installed it again without errors... I think. I even ran the pip command on the requirements.txt file and that went through but wen I start ComfyUI I get:

from .AestheticScore import * File "C:\Applications\StableDiffusion\ComfyUI\python_embeded\lib\site-packages\ImageReward\models\AestheticScore.py", line 16, in <module> import clip ModuleNotFoundError: No module named 'clip'

[hardy cipher]

add "clip" to your requirements.txt file and run "pip install clip"

if you want to use that

[soft bone]

You say here that using regularization images further tunes the model to that style/quality, and that's why you use high quality photo datasets. is that still true for small trainings like on 16 images?

[tender timber]

Will give it a try, thanks

[hardy cipher]

you need to give this a try https://civitai.com/models/124110/furry-wolf-sdxl

[hoary saddle]

still working on my blender addon for comfy, gotta add workflow to import via button

will be cool to see it work with post generated images and creating depth and normal maps from a workflow

[heady vale]

that a funny looking sexy babe on the beach😆

[hoary saddle]

lol dont judge me!

[upbeat summit]

that's awesome!

[crisp owl]

anyone submitting to the civitai contest? Curious what you do for the "missing generation data" if you use ComfyUI to generate your images

[hoary saddle]

coming along nicely, few stumbling blocks

[autumn forum]

i usually dont have that issue with missing generation data

[hoary saddle]

what gpu did you run that on?

[autumn forum]

They are using AI Template on a 4070ti i beleive

AIT significantly improves speed. and hopefully soon to be supported in Comfy

[hoary saddle]

reading the github on it, didnt realize it was made by facebook

[gloomy barn]

what is made by facebook?

[autumn forum]

AI Template

[hoary saddle]

facebook incubator engineering team

[gloomy barn]

what is ai template?

[autumn forum]

a thing that does a thing that helps speed up ai gens

[hoary saddle]

this ☝️

[crisp owl]

Using ComfyUI? SDXL image?

[gloomy barn]

speed of what?

[hoary saddle]

https://github.com/facebookincubator/AITemplate

[gloomy barn]

Yes, I read it

It can make something like SD work without so many middle men apps?

[hoary saddle]

cant tell but think it's more compiled with the cuda 11 package rather than being "part" of comfyui

[gloomy barn]

scroll that engineering team off this page

[hoary saddle]

lmao

[gloomy barn]

I wonder what it actually does....sounds like it removes the need for certain middle apps in the process of making things like SD work...comfy or one of the better UI's like a1111 and SDNext

[indigo carbon]

scroll up, I already answered that question

[gloomy barn]

So it only works with marketypush UI, Comfy?

[indigo carbon]

yep

[gloomy barn]

So not very useful

[autumn forum]

very useful. i love comfy

[gloomy barn]

Facebook developed something to improve comfy?

[autumn forum]

comfy is love comfy is life. lmao

[gloomy barn]

odd how the whole community changed in a flash to talking about comfy

[indigo carbon]

if you hate nodes just use comfybox xD

[autumn forum]

tbf i didnt know about comfy until sdxl 0.9 leaked

[midnight shuttle]

Comfy and A1111 are different tools for different jobs. I hope they both continue to be developed.

[gloomy barn]

The facebook github doesn't even mention comfy

[hoary saddle]

i hated comfy when i first started using it, all my scripts were using a1111's api, exclusively use comfy now just took a bit to understand it

[midnight shuttle]

Because the other UIs were late to fully supporting SDXL.

[gloomy barn]

odd that it's made specically for comfy, but it doesn't even mention comfy

Nobody was late. SDXL doesn't support SDXL...we've lost tools like controlnet

[hoary saddle]

i think it's made for AI generation in general, someone else is making something to make it work "with" comfy

[gloomy barn]

right\

[crisp owl]

No tools were lost, they are just for a different product

[midnight shuttle]

Comfy fully supports the SDXL generation process (can't say workflow anymore because people think that means spaghetti nodes). A1111 doesn't support the full process yet.

[gloomy barn]

The electric start on the packard wasn't lose. The 1960 car you now have to crank start is a different product

[midnight shuttle]

I hope for SDXL inpainting model someday. It really does work better for many processes.

[gloomy barn]

SD NExt supports it

[crisp owl]

I've said it multiple times, SDXL has been out for 2 weeks, and people expect it to have literally everythign that 1.5 has which took months to get to

[gloomy barn]

what aspect of the process...the workflow, which now you're not using because marketing...doesn't A1111 support?

[hoary saddle]

i'm using inpainting in comfy now with sdxl, seems to work, controlnet "soon" we think

[midnight shuttle]

Inpainting works if you inpaint based on the whole picture. But if you want to inpaint only the masked area you need an inpainting model for best results. Same as 1.5/2.1. No different.

[gloomy barn]

When the cars came out in 1960, they had everything the Model T's had

they didn't go back to 1900

[fresh path]

Is there anywhere I can pay to train a lora for sdxl

[midnight shuttle]

Controlnet will come. Base model is base model.

[hoary saddle]

ah

[crisp owl]

Lets continue to compare apples to oranges

[gloomy barn]

they didn't trot out horse carriages, insist they were a step forward, and soon we'd have engines again

[midnight shuttle]

Runpod (not affiliated).

[fresh path]

I’ll look into it

[hoary saddle]

new gpu friday, then gotta learn how to train, then maybe lol

[gloomy barn]

what gpu are you getting smutz?

[hoary saddle]

☝️

[indigo carbon]

just use ComfyBox if you despise nodes so much, if you want the speed boost I demonstrated earlier you need Comfy as a backend.

[gloomy barn]

wow

I have comfy

I like SD next better

[hoary saddle]

wtf is comfybox? lol, now i gotta go look

[gloomy barn]

what part of the workflow...er, sorry, the marketing language is now to call it process, does sdnext not support?

[indigo carbon]

frontend for ComfyUI, usable with the AIT workflow I made.

[glad grove]

does it support animation extensions?

[gloomy barn]

which animation extensions?

[glad grove]

mov2mov,sd animation,deforum,animatediff

[gloomy barn]

Well, I don't know. You mean with SDXL, right?

[glad grove]

yea can u just paste those extensions into sd next or someone has to port them?

[gloomy barn]

they worked for 1.5. I don't know if they work for SDXL now

[tribal jackal]

I rebuilt my PC into a new case today, my 4080 is hitting temps that seem pretty steep and I was wondering if someone else here could monitor their temps during a generation to increase my sample size. Pretty please and thank you.

[gloomy barn]

can you point me to them and how to get them into comfy?

[glad grove]

cant get them in comfy yet

[gloomy barn]

ok\

then what does comfy support that SDNext doesn't?

[midnight shuttle]

Yes those are high. Try to improve GPU airflow.

[tribal jackal]

I'll have to wait til tomorrow then. My corsair fans don't fit the new thermaltake mounting bracket so I'm lacking 3 intake fans

Thanks

[midnight shuttle]

That could be why. Need airflow to move hot air away from GPU.

[gloomy barn]

Put a yurblotz in there

[tribal jackal]

I did make two pretty pictures during testing at least...

[gloomy barn]

A Yurblotz will help with your temps

[tribal jackal]

lol I thought you were just making words up at first

I'll look up yurboltz

[upbeat summit]

how's nido? 🙂

[tribal jackal]

yurblotz

[gloomy barn]

A yurblotz is a kind of tiny frost dragon. I have a few of them at my house instead of AC. They eat bugs, and they breathe out cold air

[tribal jackal]

Just the best boy ever, as always

[gloomy barn]

Tell the dog "Burf" for me.

[tribal jackal]

Will do, after his nap of course

[hoary saddle]

Same

Vibes

[gloomy barn]

I dont have a dog. However, there are some coyote packs around here that are friendly with me

[glad grove]

its true what they say,sdxl does insane realistic cat pics

[gloomy barn]

I have a sure fire winner for the generate contest

the 4090 will be mine

I am working on an image of a man gazing into a crystal ball wherein is shown a fight between a dragon and the men of a town

However, that is not the winning image

the contest won't be a best generation contest. It will be a who gets the most people to like their entry contest

aka...solicits people to go choose it

[midnight shuttle]

Welcome to internet.

[gloomy barn]

they claim that won't be allowed. How can they stop it?

The bottom line is that to overcome that, someone would need an image that gets people to choose it despite going there to choose something else. I have such an image\

[midnight shuttle]

Image McImageface?

[gloomy barn]

nope

my image is like that one that's been making the rounds on Reddit

a boy goes to a witch to curse a pendant....it will make whoever wears it die in 7 days

the boy places it around his mother's neck - the doctor says she won't live through the night

If someone could make that - it's a 4 panel comic - on one image generation, it would win

[glad grove]

i won the contest gimme the 4090

[gloomy barn]

That would make a good pet

An image that might win would be the image of my meeting with a vulture. One day 6 vultures were on my roof. They had caught something, and they were eating it. I turned on the sprinkler to scare them away because when I went out there to shoo them, they attacked me from all sides. Their claws and beaks were sharp. I could nae fight them all off.

Alol of them left except one, the runt. I guess it was desperate for food, and normally it didn't get much because the bigger ones got all the food

I had mercy on it - turned off the sprinkler so it could eat.

Today it returned, spit out at me the skull of a crow. I guess that was my reward for scaring off all the other vultures

[glad grove]

hmm tasty vultures

[gloomy barn]

my arms got all cut up from them

I think they are black vultures. They didn't have red heads

[gloomy barn]

That video of those images generating - it sure was fast

[crisp owl]

How many of this style do you have now?

[gloomy barn]

Get it so that inside isn't just colors, but an actual model of something like a town or mountain...something like that

[hoary saddle]

hah put your text in for fun

[hoary saddle]

dood

[gloomy barn]

If you could get all four panels in one generation - the story is a tear jerker. Of course, only a fool would use black magic for any purpose, even a good one

[midnight shuttle]

No one will stop to read the image while voting. They will look for colors, details, waifus.

[gloomy barn]

I think they'd look at that one

The winner I have, however, will instantly grab attention and appeal to people in this community

What is this Waifu? Is it like Kung Fu, like a waif fighiting, a street urchin who fights using martial arts?

[hoary saddle]

WAIFU - What Actually Is F_d Up in this image 😉

[gloomy barn]

what image?

[midnight shuttle]

Everybody was waifu fighting... 🎶

[gloomy barn]

Kung Fu is not as much about fighting as it is about strengthening the body

[trim orbit]

just thought of a good way to remember sd15. its the one that made every girl look 15.

rip sd15. long live sdxl

[indigo carbon]

how to remember SD2.1? faces generated with SD2.1 look like they are missing 2.1 million brain cells

[heady vale]

even 15 is a little generous..many looked a lot younger

[trim orbit]

and mind. the roots are for any skill that requires an mastery and discipline over time if i understand it right

[gloomy barn]

correct

[indigo carbon]

also had 15 extra fingers and facial features

[heady vale]

[indigo carbon]

yeah, classic

also conjoined twins with wide images

hard to believe that was about a month ago

[trim orbit]

you'd get a lot of centaurs too

[midnight shuttle]

I wonder what we'll have in a year or two.

[trim orbit]

what i call the people riding themselves

[glad grove]

more sdxl asian girl loras

[trim orbit]

holodeck confirmed

[midnight shuttle]

Moar waifu is a given. I wonder if there will be another new model by then or if SDXL is it for a while.

[glad grove]

SDXXXL soon

[tribal jackal]

text2movie

[honest flint]

on consumer hardware? maybe in ten years at the rate nvidia is drizzling us with hardware refreshes

[trim orbit]

seems to require a tonnnn of processing ower for video, which makes sense. i think for most people, video won't be for a few years.

[tribal jackal]

Apples AR headset is gonna have people walking around with GPU backpacks turning all the pedestrians around them into waifus

[honest flint]

people already do that at home

[tribal jackal]

With VR sure

[trim orbit]

i don't think it'll be that integratable. it'll do that with macbooks though

[tribal jackal]

But imagine the meeting at HR...

when your coworkers complain

[trim orbit]

and macbooks got some of the best ML silicon on them

[glad grove]

google glass all over again

[tribal jackal]

lol I just think it's a hilarious thought

[tribal jackal]

Now with ai!

[glad grove]

ppl even used a term for them i think it was called glassholes

[trim orbit]

i'm thinking of getting one just to jailbreak it and put googly eyes on the visor

[glad grove]

they refused to interact with you if u were wearin a google glass

[tribal jackal]

I just wish I could record my perspective when playing with my dogs but pervs gotta ruin everything

This dog's day will come.

[trim orbit]

a loud minority was against google glass. the reason they really didn't move is because they were ugoh and the actual text dispaly sucked. it was pointless

[honest flint]

building your own google glass wouldn't be that hard today

[trim orbit]

cameras are EVERYWHERE. no one cares

[tribal jackal]

Yeah the tech wasn't all that useful yet.

[honest flint]

could probably even make it look decent with some custom 3d printed parts

[trim orbit]

glass failed because it was dumb

there's tons of smart glasses on the market still. raybans that take photos for your phone. all sorts

[honest flint]

the thing that is keeping these things from really being awesome is power... battery capacity is still garbage and cpu's still use too much of it

[trim orbit]

https://www.ray-ban.com/canada/en/rayban-stories these things are actually getting some market penetration

[azure oxide]

smart glasses would be great if they werent bulky+had the capability of high res video streaming

[honest flint]

they could but we have that power issue

[trim orbit]

yeah but now super conductors

[azure oxide]

yep, thats why i specifically pointed that out lol

[honest flint]

nope that LK99 was debunked

[trim orbit]

i know 😦

[azure oxide]

i read about an idea of a smart battery vest to power all your wearables

[trim orbit]

but super conductivity at ambient is going to be a thing.

[honest flint]

never

[azure oxide]

if universal, products could be designed/developed going forward with that in mind as a power source

[trim orbit]

hmm. a vest that could ignite at a moment. i like that cool idea

keeping a bomb in my pocket is something i've gotten used to so lets just vest up i guess

[trim orbit]

it's just an engineering problem. the math is real.

[honest flint]

math is just a story, stories can be made up

[trim orbit]

perhaps math is the greatest story ever told

[honest flint]

or the worst, but there are usually better stories every couple decades or so

[crisp owl]

https://tenor.com/view/nacho-libre-i-believe-in-science-gif-9246419

[trim orbit]

only because children are stupid and have no cultural knowledge so missed that story last time it was told

ugh kids are so dumb it gets on my nerves so much

[honest flint]

everyone is dumber yesterday then they are tomorrow, its how our minds work in this linear spacetime we are trapped in unfortunately

[trim orbit]

no i think some people go the other way

some people get dumber

[honest flint]

well yeh, people have strokes, eat like shit, or just stop giving a fuck

[trim orbit]

or that entire generation of people who were exposed to wide lead contamination from leaded gasoline emissions. that can happen too

[honest flint]

pretty sure covid dropped the global iq average a couple points as well

you are forgetting the leaded paint that was quite popular around that time and is still very much present today

[trim orbit]

that long vid has me dealing with concussion recovery again. i def got stupider

[glad grove]

i love microplastics on my food

[trim orbit]

nope. i am not forgetting the leaded paint on all of the toys and bedrooms and cribs and lead in everything.

story goes lead paints taste sweet so there'd be kids who ate the paint chips

[fathom hound]

Yall didn't eat paint chips as children?

Am I the only one??

[honest flint]

or irradiated seafood from nuclear power plants dumping their waste into the pacific...

[heady vale]

looks like XL used MSPaint for one of the eyes

[glad grove]

or fake sugar, fake colors and fake fruit flavor on my icecream

[trim orbit]

i'm less worried about the radioactive water since the pacific is huge

[honest flint]

i haven't checked lately but after that disaster in japan the levels of radiation were quite high in the fish/crab/etc being pulled out of the pacific, even as far as NA coastlines

[trim orbit]

i often finds it gets eyes just a little off taht way. like on one the reflection is trying to be sharp and the other it's soft. detailing pass on the face is my favorite for portraits, but i've noticed the settings i got dialed in makes them sort of plasticy. maybe to much denoising or steps

effort to get back on topic. existential dangers are a rabit hole

[honest flint]

eyes have been an issue all the way back to the SD 1.x days, usually from a VAE that needs more training or a jacked up text encoder... wouldnt be surprised SDXL has one or both of those things going on

[trim orbit]

higher than normal. a banana has more radiation exposure. remember, the pacific is literally huge. one of the hugest things on the planet

take all the water in the pacific and put it in a ball, it'd be huge

[honest flint]

relative size of the pacific doesn't have much literal correlation with the risk posed from an industrial sized nuclear powerplant dumping all of its material into it

[trim orbit]

i think all powerplants are industrial sized. they're kind of industry that way

[honest flint]

wtf thought it would be a good idea to build a nuclear powerplant on the beach in an area that frequently seas tidal waves

[trim orbit]

but yeah. it does. the pacific is huuuuuuge

[honest flint]

not really on a cosmic scale

[trim orbit]

that radioactive water would have abad local effect def but the wider? its not homeopathy

[honest flint]

it did tho

[trim orbit]

cosmic scales are unfathomable. huge still sort of it

[honest flint]

it has measurable impact on the radioactivity of the fish being pulled up of the north america coast

[trim orbit]

measurable yeh. higher than typical yeh.

[honest flint]

if its measurable its not homeopathy 😛

[trim orbit]

i live on the pacific coast directly across the currents from fukushima. the debris washed up on my shores on vancouver island. i'm informed on the matter. i live it

worry more about a solar flare than fukushima. there's been at least a dozen that dump more radiation into the atmosphere since then

luckily the atmosphere is huge

[honest flint]

we also have layers of both physical, magnetic, etc composition that filters out most of that crap

not saying a huge ass solar flare couldn't royally screw us tho

[trim orbit]

not the ones that break through and cause auroras and measurable effects on the elctrical grid

[honest flint]

well keep enjoying your uranium infused fish, ill go eat my lead paint chips

[trim orbit]

also the other problems are free radicals and other ionized particles. but they're in such diluted amounts that literally a banana exposes you to more

[heady vale]

one to beam down, Scotty

[trim orbit]

you'll never beam me lucky charms

[heady vale]

hahaa

[glad grove]

why is she wearing the captain uniform and not the admiral uniform

🤓 🤓

[boreal bough]

sphinx of black quartz, judge my vow

[heady vale]

Cant make Spock do the Vulcan salute. dont think XL knows what it is

[soft zealot]

have you tired the phrase £ making a Jewish Priestly Blessing" of " Kohanim Priestly Blessing" ?

[heady vale]

will try

[trim orbit]

haha yeah it loves doing him with a hand up, but fingers splayed

[heady vale]

nope. He just has his fingers raised

[heady vale]

amazingly im seeing 5 fingers consistently

[trim orbit]

higher cfg

[heady vale]

stick em up

Soooo close

[gloomy barn]

Live well and propagate.

[trim orbit]

spock on the bridge

[high skiff]

Spoculues

[upbeat summit]

amazing

[trim orbit]

i mean if you want good palms, i guess prompt spock

[trim orbit]

these hands are outstanding

[heady vale]

yep

magic hands

[trim orbit]

but they can't do the vulcan salute

[heady vale]

sadly not

[gloomy barn]

its called the jewish blessing of aaron

you can do it with controlnet

[trim orbit]

got a controlnet for sdxl handy?

[crisp owl]

hehe, handy. I get it

[gloomy barn]

still missing a finger

Imagine Spock mind melding with Stable Diffusion

[heady vale]

my latent to your latent

[trim orbit]

my prompt to your prompt

[gloomy barn]

"My mind...to your mind. My thoughts...to your thoughts. It is in the stars, when at war always with the Sith there are two...master and apprentice. AT LAST we shall have our revenge on the Jedi! When the stars trek, if a hand has six fingers, then hours are like days. The repairs will be complete within four days....if hands have six fingers...or four."

[soft zealot]

did anyone see the SNG episode recently where Spocks ears were made human and he had to ear a beanie to meet his mother?

[gloomy barn]

Terrible show

[soft zealot]

really ? I'm loving it

[gloomy barn]

anyway...controlnet could easily do this

[trim orbit]

[:splayed:0.25] in negative is aiding a bit

[soft zealot]

wheres the fun in that lol

[trim orbit]

helps to investigate the latent space's mobility without controlnet still. i like this character study

prompting is king

[gloomy barn]

Situations sometimes work

I have found that if I cant get a clothing style, I can describe a place and/or activity to get it

try holding your mouse to your mouth like a microphone, and then speak the prompt

[trim orbit]

yeah prompting i always put a location. context helps sdxl a lot more than 15

[gloomy barn]

"Computer....give me an image of Spock making the vulcan salute"

[vital ermine]

anyone been able to use comfyui and load iA3 created lora files? DyLORA as well?

[gloomy barn]

https://www.youtube.com/watch?v=gjQVxLxXAC8

[trim orbit]

postcard from vulcan

[gloomy barn]

This output is highly illogical

[trim orbit]

all the spock pics were probably all captioned with "holding hand up" so it got blended with knowledge of every other hand

future dataset building opportunities

[heady vale]

The Borg

[trim orbit]

john spock. you don't want to see what he's like when you kill his tribble and steel his shuttle craft

[gloomy barn]

Spock Motor Sales is #1 because our prices are always logical.

[gloomy barn]

The most deadly type of Borg. The Bjorn Borg.

[high skiff]

The Bjorg

[gloomy barn]

https://youtu.be/EJ5msrmf5Tw

[soft zealot]

why has noone done 7of9 yet ......................

[heady vale]

I tried. got someone completely different

[strange mist]

me too, man 😁
that is the most enjoyable trek in recent years.

[trim orbit]

psy borg

[gloomy barn]

Star Trek Woke Frontiers is not enjoyable

Here's another of Seven Eight Nine

Here is from the awful Picard series

[soft zealot]

Trying to remeber the name of the fan based series I warched aon YT a year or two back that basically redid the TOS episodes anew, It was most excellently done

[gloomy barn]

at least she didn't plastic surgery and botox herself into oblivion the way Dcctor Crusher did

[gloomy barn]

Star Trek New Voyages

[strange mist]

it's better than STD or picard, imo 🤷‍♂️
and you can't watch anything without woke these days

[gloomy barn]

Star Trek Continues

Both have some good episodes

[gloomy barn]

I can

[soft zealot]

nope not that one, this one was better

[gloomy barn]

Star Trek Continues, then

[soft zealot]

that was the one 🙂

[gloomy barn]

New Voyages has some terrific episodes - the one with the doomsday machine, sulu, checkov

[trim orbit]

i loved the new episode of strange new worlds where they entered the pocket of space that caused them all to sing

[trim orbit]

and last season they did a freaky fridya episode. it was amazing

[soft zealot]

that was funny as

[strange mist]

you close your eyes when woke scenes appear or don't watch them at all? 🫣

[gloomy barn]

Starship Exeter has some good episodes

[gloomy barn]

I tried to watch some suiperhero show on netflix

Heroes of Tomorrow or something

[trim orbit]

another psy borg.

[soft zealot]

although the SNG episode 2 before where they did the corsssover with Lower Decks

[gloomy barn]

well, I tried to get the vulcman salute\

couldnt do it

[strange mist]

that was your mistake, watching anything on netflix 😀

[gloomy barn]

I dont watch much of anything anymore

I watch anime when I work out or bible study videos and debates

[soft bone]

bard on how to better overwrite a celebrity token for training a new face. does this make any sense to anyone?

[gloomy barn]

they turned dororu into woke garbage

[glad grove]

who is doruru

[gloomy barn]

it's about this baby that is born without body parts, skin, internal organs...his father sacrificed him to demons to gain power. Some sorcerer keeps the kid alive, builds wooden parts for him, and then the kid goes around slaying the demons to get his body parts back

the main character is a boy, a thief named Dororo (I had it spelled wrong) who helps this guy

In the Manga, it turns out Dororo is a girl whose mother raised as a boy in order to protect her from being raped in war torn Japan

in the anime, they change it to that Dororo identifies as a male trapped in a girl's body

the Manga focuses on the demon slayer's relationship with his father and younger brother...the battles between them

the anime focuses on Dororo's gender identity - becomes unwatchable from about half way on

the main twist of the manga is played down in the anime in favor of Dororo's gender identity

[trim orbit]

can we not do subtle hate and prejudice against people? not even in any channel here/ that'd be nice

[gloomy barn]

yes, please don't

[trim orbit]

you seem really angry about woke , wahtever you think that is.

it's probably just a facade for hate as usual.

don't

[gloomy barn]

Im dropping it because I don't want to get kicked out or banned

I expressed no anger - that's the last thing I'll mention

[trim orbit]

well that confirms the context you meant it in then.

[gloomy barn]

Feel free to PM to discuss it

[trim orbit]

if you think you'll get banned for it, then you know exactly wtf you're doing

naw. ew

[gloomy barn]

This would be easy with controlnet depth or cammy...whatever it's called

I bet you could train a vulcan salute lora, too

or just use a hand pose with controlnet

openpose

[glad grove]

since no controlnet for sdxl your best option is img2img

[gloomy barn]

I doubt it would work

I guess that's not bad, but can't get rid of the stubble

[heady vale]

img2img

[gloomy barn]

that's a good one

[gloomy barn]

that looks nothing like spock

[high skiff]

I am getting beyond frustrated with making LoRA's for SDXL

I can't tell if something is just flat out broken with Kohya right now or not

[glad grove]

its been more than 2 weeks and still no Steven Seagal lora for sdxl 😔

[upbeat summit]

how's that new GPU?

[high skiff]

It's sexy, I love it

It's been

No problems

[uncut fiber]

This is Spock 🙂

[upbeat summit]

that's great 🙂 glad to hear it

[high skiff]

Now, if only I could get any results from SDXL training

[dusk mica]

dumb question, how does the StableDreamer bot handle prompt compared to a local comfy ui workflow?

like i feel like bot follows prompt better

[boreal bough]

I present to you - training with the nuclear option known as "caption dropout"

[dense chasm]

unet diffusion_pytorch_model.safetensors 10.3G,long time downloading..for running inference API on huggingface

[high skiff]

At least you have results

[boreal bough]

probably worth setting up your install anew

also, kohya gui got preset updates last night, for sdxl

so those should be useable now - including prodigy

might be a good time to reinstall it

[dense chasm]

git LFS pull is much slower than LFS push

[upbeat summit]

is it in real-time or something?

[boreal bough]

yeah, using kohya samples XD

I added too many new settings to this training run - so I needed at least some way to track it live

[high skiff]

I did

I have an old install, a recent one, a brand new, and brand new Derrian from today

[boreal bough]

yeah no. if none of them are working then there is an issue O_O

python on 3.10.9?

cuda cores are actually working? 🤣 and not starving for power? (/jk, can't image this being an issue)

[dense chasm]

how u train lora?with kohya GUI or just script running?

[boreal bough]

with the gui - since it adds all the working things from the dev branch of kohya-ss

[dense chasm]

what's prodigy params?cuz i just train lora thr python hardcoding,have to modify the training optimization config file

[clever verge]

I'm running my first LoRA training now but the example dataset (pokemon) is large. Should get a smaller set to test on. 😄

[dense chasm]

adafactor's loss around 1.1,if i wanna to minimize loss value,it seems like i have to change another optimization type

[boreal bough]

mine are a copy of this post:
https://civitai.com/articles/1022/sdxl-trainingbdsqlsz-lora-training-advanced-tutorial2best-optimizerprodigy-is-all-you-need

This is the new updated version live on kohya gui: (I didn't test it - but I'm guessing its working if its recommended for everyone?)
https://github.com/bmaltais/kohya_ss/blob/master/presets/lora/SDXL - LoRA prodigy AI_Now v1.0.json

[dense chasm]

your GPU VRAM volumn?what's the minimum requirement?

[unreal plover]

https://catalog.ngc.nvidia.com/orgs/nvidia/teams/playground/models/sdxl

[boreal bough]

I run it at just below 24gb vram
but I've seen in the comments, that people got it down to 8gb vram - can't help on that, as I never tried it

[west breach]

Just published my new Inflated lora https://civitai.com/models/126131
Thanks to @upbeat summit for helping test it out 🙂

[dense chasm]

lora train another character with bokeh prompt,if you generate thr refine model,just make sure chain it with upscale nodes, otherwise the render will be lost with more details,i will update how to do inference on huggingface later,it will be more interesting and time saving

[vale eagle]

what are the difference between no caption and caption dropout?

[boreal bough]

Dropout caption every n epochs
Usually, images and captions are learned as a pair, but it's possible to train just on "images without captions" every certain number of epochs.

This option allows you to specify "drop out captions every ○ epochs."

For instance, if you set this to 2, you will conduct image training without captions every 2 epochs (2nd epoch, 4th epoch, 6th epoch...).

By training on images without captions, it is expected that your LoRA will learn a more comprehensive feature set from the images. It can also help prevent the image features from being tied too closely to specific words. However, if you use captions too sparingly, your LoRA could become ineffective at prompts, so be cautious.

The default is 0, and in the case of 0, caption dropout is not performed.

Rate of caption dropout
This is similar to the "Dropout caption every n epochs" mentioned above, but during the entire learning process, you can train on "images without captions" for a certain proportion of the time.

Here, you can set the proportion of images without captions. 0 means "always use captions during training," and 1 means "never use captions during training."

Which images will be trained as "images without captions" is determined randomly.

For example, if you train LoRA with 20 images, reading each image 50 times for just 1 epoch, the total number of image learnings is 20 images x 50 times x 1 epoch = 1000 times. If you set the rate of caption dropout to 0.1, 1000 times x 0.1 = 100 times, you will train on "images without captions."

The default is 0, and all images are learned with captions

caption dropout info

[heady vale]

hahah works great. nice work

[west breach]

nice 😄 glad you like it

[vale eagle]

Might be I need to do a compare between no caption, caption dropout, full caption.

[vital ermine]

base and a plain lora

[vital ermine]

10 and 20 epochs in comfy

[dense chasm]

left is base model lora training image,looks fine,right is upscaling refine model generation,breathtaking

[upbeat summit]

it's been a lot of fun. very unique fine-tuning!

[west breach]

thanks again for helping to test it out 🙂

[upbeat summit]

you're welcome! thanks for letting me explore it

[west breach]

wonder why it flagged this as 'sexual acts'

[upbeat summit]

I tried to click the arrow

[dusk mica]

maybe check the prompt

[west breach]

it didn't seem to detect any of the metadata either

[dense chasm]

sdxl refine model render details is so good,left base model,right is refine lora upscaling

[somber hill]

this is totally related to number of repeating. if you do more number of repeating it will have stronger effect. because it will use that many images during training as regularazation images

[upbeat summit]

that's great!

[heady vale]

thanks

[west breach]

sigh wrote an article on my nodes on civitai and it just went to a 404 page when I went to publish it, doesn't even save a draft or anything

[upbeat summit]

damn 😦 yes, their page is currently unable to handle the load

I since write everything in my editor or an email draft first heh

[boreal bough]

face lora guide + settings is up 😄
#🔧｜finetune message

[dense chasm]

lora with sdxl refine is awesome

[hexed hatch]

The more I increase steps, the less it follows my controlnet in comfyui, anyone know how I could increase steps, while having it keep the same amount of controlnet control as it would have with 20 steps? I hope it makes sense what I am saying.

[dense chasm]

okay,i admit the character generation is surpass midjourney with different styles and diversity,dark style is so good as well

[unreal plover]

When Controlnet XL ?

[ionic dragon]

even this lora has same settings as the old one?

[soft bone]

I see you're training clip. if it works then what were these "unexpected results" we kept getting warned about with that?

[dense chasm]

interesting poses

[soft zealot]

the head looks out of scale to the shoulders as well

[boreal bough]

training clip is an all or nothing situation normally - but that's why I'm so specific on how to caption it, exactly how long to run the training
it's a very controlled environment

For my big datasets I'm always training clip now - since I have more than 150 images per caption word I'm training - with a total of 4k images 🤣
in that case it works

but the smaller the dataset gets - the harder it gets to train clip and get better results than without training it

[soft bone]

I see. for my purposes I can't manually edit captions so I probably shouldn't train clip, plus my training datasets are always 12-26 imgs

[boreal bough]

yeah :/ was gonna recommend something with user input... but that fails at the individual image level - where you can't exactly ask them to tag if they're wearing a sweater or not XD

[soft bone]

yeah probably would create more problems.

I'm getting good results with unet only right now but there's still a hint of the celebrity used for the token if XL knows their face too well. increasing LR doesn't help and I really don't want to sacrifice more epochs, it's currently an hour for 16 imgs

"professional model" and "model photoshoot" didn't work as tokens

[boreal bough]

I'm pretty confident that the 4 class tokens listed above can be made to work.
also relies on caption dropout.
those would always yield exact reproducible results - but testing various dropout settings would be a pain until you finally get it right 🥲

[dense chasm]

i just uploaded the actress upscaling model on https://civitai.com/models/126239

[boreal bough]

clip skip 2? O:

I haven't tried clip skips at all yet in sdxl. is there value in training with it?

[dense chasm]

since sd2.X,it seems like clip skips is unavaible anymore,anyway,i just leave it for reference.

[fair stag]

My understanding is that the refiner doesn't work with loras, and can't be trained.... Therefore most of the time the refiner actually regresses those learned traits

[soft zealot]

well there is a node in COmfy for it but cant say Ive tried using it

[elfin flare]

Voxel LORA is great.

[dense chasm]

refine model is just render more details than upscaling

[fair stag]

Fair enough, I guess if you keep the denoise low enough

[dense chasm]

https://www.reddit.com/r/StableDiffusion/comments/15nb3wf/realistic_sdxl10_base_model_trained_lora_for/

currently,i'm focus on the hot inference API of huggingface,it takes some time to deploy

[dense chasm]

base model denoise 1.0, refine model 0.25,it's not a big deal,neural work params optimization and be effective and efficiant of work are more important

[soft zealot]

First image is Clip Set Last Layer = -1 , Second Image is Clip Set Last Layer = -2

All other settings identical

(dont count fingers and ignore the foot lol)

[inland vessel]

how could i get something close to this?

[boreal bough]

it was reelvant for the novelAI model back from SD1.5

since they finetuned at clip skip 2, so it gave better results for inference at clip skip 2

[lusty token]

what does "it" in it/s stand for?

[soft zealot]

iterations

[elfin flare]

You shall not pass.

[lusty token]

That's a cool approach, to use a percentage for the base.

[potent flame]

smh ComfyUI is much slower than A1111?

216it/s

[lusty token]

In ComfyUI, is there a way to navigate the Canvas without having to grab outside a node and drag?

Currently I have to zoom in and out then grab and drag a lot. I'd prefer if I could stay zoomed in more often and just scroll around the canvas.

[crisp owl]

spacebar

[boreal bough]

cries in cpu generation

[potent flame]

i am on GPU

[boreal bough]

but are you... really? 🤔

[potent flame]

i am on MX130.

[boreal bough]

jk. you prob have something causing a vram overflow into standard ram. which is causing to run at cpu speeds

[potent flame]

hm.

Chrome may be eating too much RAM

maybe i can try Edge

2GB VRAM is obviously not enough

[boreal bough]

oh god no XD

[potent flame]

but i wouldnt think it contributed by the RAM

[boreal bough]

6gb is the lowest I've seen work

8gb vram 'works'

12gb vram is recommended for 'full' support

[potent flame]

i still got it work actually

[boreal bough]

oh yeah it always works - but not via gpu

[potent flame]

i am always on GPU

[boreal bough]

that's the cpu doing it in the background

[potent flame]

even in SDXL lol

[boreal bough]

what dis?

[potent flame]

@halcyon tusk ( right ping?)