✨｜sdxl

[fast vector]

so, does anyone have afterdetailer workflow for comfy?

[boreal bough]

yeah, many words arent weighted equally. brown hair also dominates most prompts

but I'm getting there

[slender hinge]

the refiner does eliminate much of the texture issues, i still don't consider it an amazing result though tbh a little let down by the quality so far i think. I hope it will improve in the future with community efforts

also the double step process is annoying but 🤷‍♂️

[slender hinge]

oh better i guess

[static prawn]

idk why but the refiner is deleting details for me

without refiner

with refiner

[azure oxide]

are you using a1111?

[static prawn]

yes

[azure oxide]

ya i dont think it's using the refiner in the same way that it does in comfy

[static prawn]

oh ok , i just lose a lot of detail here

i think u can clearly see in the image

[boreal bough]

k. fixed the blur, now I just need to set the style

[slender hinge]

how did you fix?

[boreal bough]

what kinda image are you going for? anything specific

xD

[boreal bough]

selfie of a cute young woman with pastel hair with freckles
(blur, bokeh:1.4), + whatever other negatives you use

[slender hinge]

i was actually trying to make sci-fi movie scenes, but was having issues so i just wanted to try something simple with XL first to figure it out 😂

didnt expect everyone to start making pastel ladies

[boreal bough]

ah! I got you

[boreal bough]

Positive: cat summoning a gigantic tomato using the power of but a single infinity stone, in the style of marvel movie
Style: (horror:1.2) still, grunge, vignette, chromatic aberration, dark, lovecraftian
Negative: deformed, bad anatomy, ugly

[slender hinge]

this is nice, prompt?

[boreal bough]

🤣

dont question my 'story style' prompting!

it gave us masterpieces like these

[static prawn]

anyone has an idea how to use the refiner in the right way in auto?

[boreal bough]

yeah. it can't. quite literally

[eternal crystal]

using the tiled vae got me down to a min thirty, incredible. Thank you 😄

[boreal bough]

sai even confirmed that a1111 does not have full support yet, and will prob take a while

[boreal bough]

perfect! have fun making images c:

[void skiff]

In Automatic1111 I run this prompt "lora:offset_0.2:0.2 plastic doll Dora the Explorer wearing He-Man outfit" with Seed: 169085828 and Sampling method: DPM++ 2M SDE.
It look okay until the end when it glitch out and give me a grainy image, can some test if this is same for other or if it only happen for me.

[slender hinge]

post the grainy image?

[slender hinge]

maybe you still have an old VAE set in the settings?

for an older version of SD, make sure you are using the correct VAE

or automatic

[void skiff]

I run it in 20 steps and it look goot until last frame, and it still do great in the old Euler.

[slender hinge]

sounds like a VAE issue i had the same thing, but if not that, not sure

[bold osprey]

sounds a little like the preview is just taking an estimate

[void skiff]

I run Automatin on VAE and I get same when loading the sdxl_vae.safenor

[bold osprey]

use the separate vae just in case

what is your workflow

[void skiff]

It only happen with DPM++ 2M SDE, the DPM++ SDE look okay

[candid walrus]

i just read some of the comments about the 0.9 vae, doesn't anyone know the story on why, the benefits?

[nimble heart]

They left it in the oven too long

[azure oxide]

i dont think the devs had said anything about it

[candid walrus]

yeah i got this when i tried different samplers, can't remember exactly which ones at present

yeah interesting to see it just appear in huggingface

[wary bay]

there's no "why" for it from what I understand

[candid walrus]

is it's like night and day improvement or like some like a and some like b?

[nimble heart]

anyone know a way to change the default workflow
or better yet load an arbitrary one on launch?

[boreal bough]

that one suffers from not enough enough steps on base model, too much noise AND not enough steps on refiner to fix it

[vast narwhal]

There's token limit on comfyui or works like a1111?

[void skiff]

The bot here, what do it use to generate images, what Sampling Method?

[nimble heart]

it changes

[wary bay]

I don't think it's night/day, but the examples are hard to refute imo lol

[left coral]

I'm training a lora for SDXL. Getting this error trying to generate with it

17:58:40-904771 ERROR    Diffusers LoRA loading failed: crown_vic-000006 'UNet2DConditionModel' object has no attribute
                         ''

(am using kohya)

[void skiff]

Well, for the most part SDXL give me great images so I am still glad.

[candid walrus]

i know i tried 30 steps on the _2m_sde_gpu and got the artifacts that kennylex posted, and thought, if i have to go more than 30 steps, so far... not the sampler for me

[static prawn]

finally a result i like 😄

[candid walrus]

but it begs the question if i do go higher yet on steps will i like the images enough to warrant the time

[boreal bough]

for science. it has a hardcap of 150 steps - at which point the changes stop existing

[void skiff]

What samplers do folk like to use for SDXL? Self I just try to avoid EULER and Ancestors.

[nimble heart]

Euler.

[static prawn]

just used highresfix with the base, seems to work out pretty well for me

[boreal bough]

base only - as refiner doesn't need to be the same

[azure oxide]

i coulda sworn the cap was removed in comfy
i think i did a 300 step generation once just for funsies

was a few months ago tho so foggy on the details

[boreal bough]

nah, there is no cap, but its about when steps stop becoming relevant

[azure oxide]

ah yeah convergence and all that

[copper pollen]

are there supposed to be watermarks in SDXL base model?

[cold mica]

This belongs in #1087493722645725184 ❤️

[boreal bough]

yes and no. please post a photo XD maybe you managed to set ascore to 0

[copper pollen]

This one has some small scribbles in the bottom.
That's base model, txt2img.

[left coral]

Did you ever get an answer for this question? I'd like to know too.

[copper pollen]

@boreal bough - With refiner in img2img, I get another one in the bottom left, + one diagonal across the background, like stock photo watermark google result

[lost blaze]

Mr Bean Ascending (missing the spotlight though)

[boreal bough]

oh. nah. that one is probably just baked into the artist or style you're going for.
I'm assuming you already have "signature" in your negative and it wasn't enough?

(signature:1.4) in the negative in the refiner process should fix it - though you should only use this strength when the default strength isnt strong enough

[slender hinge]

Slightly better results in comfy vs auto 1111

[copper pollen]

Ah, actually yeah, the txt2img WAS with a style LORA, first test XD
Must've come through there.

Both were without any "signature" or "watermark" in the negative. Only "noise, grain, filmgrain" in negative.

[fleet harness]

https://tenor.com/view/ross-coulthard-dave-grusch-ufo-uap-uaptf-gif-7216585467170420799

[eternal fog]

Good shit

[boreal bough]

I see your mech, and I raise you star destroyer

[zealous viper]

I see your Gundam style centaur mech and lower you a rusty beaten up mech that creates paths threw swamps

[viral narwhal]

Copax Realistic XL https://civitai.com/models/118111/copax-realistic-xl-sdxl10

[zealous viper]

srsly though good gen, i've got some rats in power armor that have a bit of the same vibe

[boreal bough]

I have so many questions 🤣

[fresh path]

could anyone explain what It/s means? It says like 17 it per second but still takes a few minutes to generate

[midnight shuttle]

Look more closely at the unit. That is not it/s. It is s/it.

As in seconds per iteration.

[fresh path]

oh haha

why is it so slow :(

[midnight shuttle]

What GPU? What model? What resolution?

[fresh path]

6900xt sdxl1.0-0.9vae 1024x

on sd1.5 I get 20x faster

is it expected performance?

[midnight shuttle]

So AMD card on Linux? Are you sure it's using the GPU? I don't know much about AMD.

[edgy otter]

I can not recommend an AMD card for AI related tasks. All the software prioritize NVIDIA because of its huge market share and CUDA so you may constantly run it problems while setting up libraries etc.

[fresh path]

it's on windows using --direct-ml argument for comfyui

and yes gpu is 100%

[midnight shuttle]

Everything I've heard is that DirectML is slow and trash. Not many people use it so you will need to ask around.

[edgy otter]

I get the same perforfamnce, I have the same card. It's beacause of the resolution and overall size of the model.

[fresh path]

oh cool

thats a shame

[polar epoch]

Automatic1111 isn't really in my favor atm lol

[fresh path]

thanks too, I dont have other options though :p

[midnight shuttle]

Did you get that from inpainting or prompting?

[edgy otter]

Use the base model, not the refiner.

[polar epoch]

txt to image with base

[midnight shuttle]

What was the prompt and CFG?

[polar epoch]

this was an earlier run with basic prompts testing refiner in img2img

[polar epoch]

7 and "beautiful young woman, short black messy hair, white, beautiful eyes, sitting in the park watching birds with a beautiful garden behind her
Negative prompt: fat, , , bad hands, mutated heads, multiple heads"

[midnight shuttle]

It's almost like your negative prompt got in the positive somehow.

[fresh path]

did you find anyway to speed it up?

[polar epoch]

Indeed, automatic has done that lately lol. The more i imply negative, the more it adds it

@midnight shuttleMe and my non existing short term memory

[midnight shuttle]

Yep. Wrong model.

[fresh path]

bro refining noise

[polar epoch]

Yeah, fuck that noise

[edgy otter]

No. If xformers supported AMD we could have gotten better performance but it doesn't. Someone created an alternative to xformers called "flash-attention" and folks at AMD are trying to modify it to work on AMD. Fingers crossed.

The last update was yesterday.
https://github.com/ROCmSoftwarePlatform/flash-attention

[polar epoch]

Oh boi.. This will kill my gpu lol

Also, noticing the generation of automatic is quite slower than on comfy

[fresh path]

thanks, it looks like there is a beta release available, I wonder if comfy cares about implementing it in ui

[midnight shuttle]

How much VRAM and what GPU?

[polar epoch]

Just about made it

[midnight shuttle]

Filling up 24 GB? Wow.

[polar epoch]

Nibbled at my ram, then it went down again :P

[fresh path]

it burped

[midnight shuttle]

VAE seems to use a lot of VRAM in SDXL/A1111. 9.3 GB used then up to 13 or 14 when VAE runs.

[polar epoch]

a 16k i generated with a 8k upscaler in comfy ate all my video memory, 30GB ram and then some lol (as in pagefile on my nvme)

[edgy otter]

No. It's fork of the original flash-attention repo which is why the documentation belongs to the original project. There's no beta release.

[polar epoch]

I even use --medvram, comfy barely uses much compared. It is at half usage when i generate basic res. CAn even make 2048 generations there

[midnight shuttle]

There is definitely something strange with A1111 and SDXL memory usage. It should not be so inefficient.

[polar epoch]

Nope!

[fossil thorn]

In comfyUI, is there any way to pick a single image from a batch and only regenerate that one? When I drag the image back in, I just get the same batch that I originally generated.

[calm shell]

here is the prompt: With Cat-Like Tread
Upon our prey we steal!
In silence dread,
Our cautious way we feel!
No sound at all
We never speak a word,
A fly's foot-fall
Would be distinctly heard!

[polar epoch]

Now to refine this one that already looks great at 80 steps :P

[fresh path]

Okay, sad :/ amd is in a really bad place right now

[calm shell]

artroomAI with sdxl 1.0

[edgy otter]

things are actually improving now 😄

pytorch and rocm's last update boosted my iterations by 3-4 on SD 1.5 512px

[boreal bough]

I see your swamp mech, and send you a bird that really wants to sit on its shoulder

[fresh path]

whats the number so I can compare?

[edgy otter]

I am gettin around 9.5

[fresh path]

with 6900xt? wow, I only get 1

on 1.5 512px

[edgy otter]

You are doing something wrong

[slender hinge]

It doesnt seem to be very good at faces that are not close up, and even close up ones seem iffy

[zealous viper]

been stuck on mechs atm ... but i've got one under fire

[polar epoch]

Not too bad, teeth is a bit off though

[midnight shuttle]

From the model documentation: Faces and people in general may not be generated properly.

[indigo carbon]

I want to punch this shit eater

I hate this duck cube so much

[slender hinge]

some ok results though

[polar epoch]

Is there a "highres fix" upscale by" feature in comfy?

[polar epoch]

Ok, i love it lol

[zealous viper]

I'd have one of those on my desk

[alpine pine]

I need this

[edgy otter]

these are fire!

[midnight shuttle]

Sometimes it is good. And with training it can be better at people. But it is not designed to generate images of people. For that a GAN would be better.

[edgy otter]

I raise you a mech girl

[indigo carbon]

I hate it, if I would see that in real life I'll punch the little shit

[alpine pine]

but its so cute

[slender hinge]

X to doubt on GANs. midjourney does a good job. I guess ill hope for community models to improve it

[zealous viper]

just gonna ignore the fact the tree/rock/explosion thing in the background is a low poly mesh but

[midnight shuttle]

Midjourney has the advantage of being a closed model so they can train it on more things and then filter the results. Open source models need to take into account possible misuse when choosing what to train. Have to be careful when training images of people.

[boreal bough]

confused robot is confused
"how does one... love?"

[zealous viper]

im not gonna lie im dying with that thing on the bottom right. look of just pure disgust/surprise

[edgy otter]

us frfr

[static prawn]

think im getting more into it now

[fossil thorn]

Going to ask again because I think it was missed. Is there a way in ComfyUI to pull a single image from a batch and just generate that one image again.
When I drag the image back in, it just gives me the workflow for the whole batch. I have tried incrementing and decrementing the seed, but I can't find it.

[static prawn]

just need an upscaler

[boreal bough]

despite half a bottle of rum, the man still couldn't handle the endless ruminations of the giant robot. If he could, he would kick his ass and make him walk to the girl.

[fresh path]

is this not the most basic workflow and I still get 1 per second

[alpine pine]

I do not believe so

[zealous viper]

is that actually the prompt

[fossil thorn]

Is there any way to interrogate the image to get its generation parameters? Or are the later outputs in a batch dependent on the earlier outputs.

[south igloo]

Where is the best place to find well made comfy UI work flows? The ones I find are very basic I want one that can do high-res fix and face restore etc

[edgy otter]

What I meant was that your system env is not set up properly. Did you install torch for rocm?

[boreal bough]

use vit-H, it works really well for sdxl

[eternal fog]

I really like this style

[noble shoal]

My eyes hurt just a little bit

[fresh path]

so, the comfyui github states pip install torch-directml, thats what I did

[zealous viper]

Pixel art at the start, i think i had Pixel art game art of -

[boreal bough]

oh, if you mean like get the exact params, then drag the image into comfy, to get back the workflow, or use automatic png info, if it was made in automatic

interrogator + Vit-H, if its just an image

[edgy otter]

idk man, I use the rocm version, works for me. I am on linux btw, rocm isn't on windows

[fresh path]

oh then thats why. I'm on windows, if I remember correctly rocm didnt work for me

its crazy linux has 10x performance for you lol

[fresh path]

1 it/s on a $1000 card is amazing

[fossil thorn]

Basically. I am wondering if there is an equivalent to PNG info in comfy. I have always used batches when testing basic things, but I like to then settle on a single image to upscale, inpaint, etc.
If I can't regenerate that single image, I am going to need to do things differently.

I don't want to do the batch each time because that is just wasting GPU time.

[nimble heart]

if you're talking AMD windows isn't even supported yet. They only just released the basic HIP SDK today, doesn't even have any ML libraries yet

[fresh path]

what do you mean?

oh is rocm the only official amd platform?

[nimble heart]

yes

[fresh path]

fun

[nimble heart]

idk what card you have but if it's an AMD one ROCm is their version of CUDA and it only works on Linux right now. It's coming to windows Soon™️

rocm started as a server thing. didn't even work on consumer GPUs for a while.

[fresh path]

ah I was just talking to s0md3v because they have the same card as me and is getting 10x the speed

[nimble heart]

yea

[nimble heart]

I have a 7900 XTX and it runs pretty fast on linux.

[edgy otter]

whats your iterations/second?

[nimble heart]

sdxl 1024 is like 3.25 it/s or something

[edgy otter]

huh neat. I get 1.5 on 6900xt

[fresh path]

wait I can just install linux

[nimble heart]

yea

[fresh path]

yeah 7900xtx is pretty good

[nimble heart]

I have both windows and linux

[edgy otter]

same

[fresh path]

idk if windows shitty partitions will allow it

[nimble heart]

windows for my Rift linux for everything else

[nimble heart]

the linux installers can forcibly shrink windows partitions lol

[indigo carbon]

Oculus Rift?

[nimble heart]

yea

[indigo carbon]

which one you got? OG CV1?

[nimble heart]

S

the cheap one

[edgy otter]

that's not a good idea. Just create a new partition for linux in windows before you even do anything else. This will make things easier if they are a beginner

[nimble heart]

worked fine for me

bakc up your stuff first though If you've never messed with that stuff before

[indigo carbon]

damn, I'm a huge fan of the original CV1, I had S, CV1 is soooo good

[boreal bough]

yep, is what I said. just drag the generated image back into the comfyui

then it loads all the params

[shy kelp]

Should I buy a100

[nimble heart]

I only got the S cause it's the cheapest headset until valve finally makes a new Index

[boreal bough]

yes if you can afford without hurting your finances. no in all other situations

[alpine pine]

how is the a100 compared to a 3090/4090?

[nimble heart]

if you're 100% certain you really wanna do literally all the AI stuff sure. If AI is an occaisional hobby I'd snag a 3090 or something instead

[shy kelp]

4090 seems better for generation and cheap

A100 is like 2.3 times faster than 4090

[boreal bough]

my 4090 can essentially do everything except pure finetuning - although even that works with gradient accumulation + batch size 1. (would just take too long)

[fresh path]

(10x the price)

[boreal bough]

3090 is the best bang-for-buck option you have right now

[shy kelp]

So like 10x 4090 can be better than a a100 maybe

Not a professional thoo

[fossil thorn]

Yes, but that gives me the batch parameters. I only want one image out of the batch, not the whole batch.

[nimble heart]

I've seen 7900 XT's new for $750, wonder how they'd do in comparison

long as you don't mind the ROCm pains

[dusty turtle]

lol just get an a6000 if you really really need the vram on 1gpu

[alpine pine]

Isnt there something about AMD and double the vram requirements?

[agile quarry]

when you load comfy with direct ML does it display your VRAM properly?

[nimble heart]

¯_(ツ)_/¯

[torn canyon]

It's running well on my 7900XT, getting ~2.9it/sec for 1024x1024 SDXL

[outer oxide]

InvokeAI 3.0.1 just dropped with that sweet SDXL support. (okay, partial support. I think inpainting is still missing?)

https://github.com/invoke-ai/InvokeAI/releases

[nimble heart]

I don't have a 24 gig nvidia card to compare against

[boreal bough]

in that case no - it only provides the batch.
you have to manually increase the seed number to generate just the one image

[shy kelp]

I use gtx 3060 should I upgrade to 4080

[nimble heart]

I do know that the pytorch 2.0 attention is broken on AMD and it causes these VRAM spikes like 2-3x what you'd normally see

[boreal bough]

no

anything but that card

[nimble heart]

so you have to use sub quad or doggettx attention

[boreal bough]

ah, yeah XD 4080 is ok

[dusty turtle]

12g-16gb memory? not really an upgrade

[shy kelp]

Is it 16Gb vram only

[boreal bough]

not necessarily what I'd recommend for ai, but its good for gaming

[shy kelp]

Which one is yhe 24gb one

[fossil thorn]

Do you know what the default increment is for Comfy? In A1111 it is 1, but I have tried 5 seeds in each direction without finding the 3rd image in my batch.

[shy kelp]

Ohh 4090

I want to upgrade 4090

[boreal bough]

3090 (around 700$ used)
4090 (around 1800$)
have 24gb vram - and are the best for AI

[shy kelp]

Ohhhh thats nice

[boreal bough]

4090 is not worth it for speed. while it is a bit faster, not nearly enough to more than double the price

[shy kelp]

3060 still works but I want to learn training fine tuning

[nimble heart]

one thing I'll say is on KoboldAI, my 7900 XTX can load 13B param models with minimal offloading to the CPU, while my 1070 struggled with 6B. The maths math out so it doesn't seem like the VRAM is crazy or anything. Think the high VRAM is just a Stable Diffusion thing.

[boreal bough]

I only got the 4090 since I do gaming and other things, and had the extra money

[shy kelp]

What you use

[dusty turtle]

you shouldn't need any offloading for 13b 🤔

[nimble heart]

16 bit depth

[dusty turtle]

https://tenor.com/view/dies-of-cringe-cringe-gif-20747133

[nimble heart]

bitsandbytes isn't working on my card yet

only RDNA 2 cards

[south igloo]

Win 11 has faster performance too

[nimble heart]

back when I still had my 1070 4/8 bit didn't even exist so I'd say it's a fair comparison

[shy kelp]

So this xl cant work good with 512x512 is iy

[nimble heart]

yea no

use a 1.5 tune for 512

50% faster half the memory better results

[shy kelp]

Hmm

[dusty turtle]

man that must suck

[shy kelp]

So what is this xl for

[exotic path]

Finally got the right to use A100 gpu on google cloud

[south igloo]

Xl sized images

[alpine pine]

1024x1024

[shy kelp]

1.5 can do 1024 too 😭

[nimble heart]

I mean there's a ROCm version someone made that seems to work well for 6900 XT's and similar. I ported some parts of the code to target my 7900 XTX but it miscompiles with some clang problem and I'm too lazy to fix it

[midnight shuttle]

But not well. If you want 512 images use 1.5 or 2.1. If you want larger images such as 1024 or other supported resolution use SDXL. Very simple.

[hearty ginkgo]

How much are you paying 😭😭😭

[shy kelp]

But 1024 way to slow

[exotic path]

Should be less than 1$/hour on spot vm

[midnight shuttle]

Buy or rent better GPU.

[shy kelp]

Still it will be way more slow on better gpu 😭

[alpine pine]

no

[nimble heart]

:(

[alpine pine]

SDXL is optimized for 1024, its a bit slower. But not by much

[vital wolf]

How to enable embeddings in comfyui? Like it's no half vae for a1111

[alpine pine]

its just a better model

[shy kelp]

Yes its like same speed butt 1024 makes it slow

[alpine pine]

it makes 1.5 slow, SDXL is made to run at 1024

[shy kelp]

Yea 😭😭

[alpine pine]

especially when you account for success rate

[nimble heart]

SDXL is faster on 768+

[calm shell]

Bonfire of the Vanities (that's the prompt) ArtroomAI sdxl 1.0 dpm++ SDE

[vital wolf]

How to use textual inversions in comfyui?

[dusty turtle]

I should probably move to linux whenever my upgrades arrive, dealing with all this shit on windows and amd is just pain

its all hacky kinda works *** bs

[calm shell]

type embedding:tinversionname

[torn canyon]

Yeah linux works so much better for AMD cards, it's well worth it

[shy kelp]

What causes this unrealistic look in stable diffusion

[nimble heart]

Use both. Whichever you install first, there's options to leave some of the drive unpartitioned then you can just run the other installer and boom.
Set the Linux install as your primary boot device and it'll add Windows as a boot option usually

[vital wolf]

No need to add any arguments like no half vae in the batfile?

[dusty turtle]

ik, was planning to do it on sd release but decided to wait since i'm swiping out ram an adding a bunch of extra storage space in a few days anyways

[nimble heart]

comfy uses full precision VAE by default

[eternal fog]

😢

[nimble heart]

major oof

And here I felt kinda bad paying $1050 for my OC'd XTX

[calm shell]

price/performance not worth that I think

[nimble heart]

XTX or 4090

[boreal bough]

3090 will always be the better deal, unless you're determined to play cyberpunk without dlss enabled

[calm shell]

my 1660 S does quite well.

but Ii think I need more vram than 6gb

[nimble heart]

man once you get a card that has ai accelerators it's a whole new world. Instead of very carefully generating images hoping they turn out OK you just generate like 10 of them at once and pick the best one

in 1.5 I had my default batch size to 16

[boreal bough]

is what I had to get to make full use of games along with rtx4090
you end up with weird bottlenecks if you go highest end

definitely wouldn't recommend XD

[dusty turtle]

lol how long before it throttles

[nimble heart]

2 TB gen 4 + 1 TB gen 3 == happy

idk anything that can use gen 5 speeds yet

[boreal bough]

most high end gaming boards from last two years. funnily enough we just didn't have nvme's that pci5 until like this year

[nimble heart]

I meant like software wise

[calm shell]

Whats the downside of a 12gb 4060?

[nimble heart]

It's really really hard to be bottlenecked on a gen 4

[shy kelp]

What model is best for photorealistic

[nimble heart]

unless you literally just copy around 100 gig models

[boreal bough]

ah, yeah. windows boot drive.
games - ratchet and clank: rift apart!

no loading while jumping through areas XD

[nimble heart]

nah I put windows on my gen 3 drive fuck windows

linux gets my gen 4

[boreal bough]

I just want my adobe stuff to load faster tbh
windows is just an additional bonus

[nimble heart]

I swear the adobe stuff mines bitcoin or something when turning on

[dusty turtle]

lol intel optane as a boot drive+ pagefile since they have the most absurd endurance ratings of any drive, 1.2pb for a 128gb drive

I think optane is still better than any gen5 for a boot drive even with 2 gens of pcie bandwith for the gen5

[calm shell]

I dumped adobe when they went SAS

But intel dumped Optane

[nimble heart]

optane on nvme drives is a scam. They literally use the same chips 99% of the time

[dusty turtle]

? its 100% optane not the hybrid drives

[fast vector]

Just try 3090. 4090 is overpriced

[eternal fog]

That's so expensive lol

[nimble heart]

what's even the point then? Just a marketing name for their spendier drives or something

it's not optimizing anything if the whole drive is made of the same chiplets

[dusty turtle]

optimizing? its a completely diffrent tech to nand flash, I use it because it has a hilariously high write endurance/sustained write speed vs most junk nand drives
https://www.intel.com/content/www/us/en/products/sku/211867/intel-optane-ssd-p1600x-series-118gb-m-2-80mm-pcie-3-0-x4-3d-xpoint/specifications.html

[heady vale]

PCWorld did a video on optane drives. For most task regular nvme is better. some task optane flies past everything

[pliant acorn]

ROCM is now out for windows. Any ideas how to use it with comfy ui?

[alpine pine]

was PC world speaking on indurance or just performance?

[nimble heart]

the HIP SDK is out but the ML Libraries arent' and there's no torch builds for it yet

[heady vale]

performance only

[pliant acorn]

oh

😦

[nimble heart]

yea

[alpine pine]

I think Robot is primarily speaking about indurance

[slender raft]

Anyone tried latent upscale? I'm getting artifacts when I try.

[nimble heart]

think endurance is oversold nowadays. had this drive for like 4 years across two computers having done a lot wipes + clean installs and in/uninstalling dozens of fat bloated modern games.
total expenditure, 8% lol.

[dusty turtle]

lol 20tb written, i've done like 2tbw in 1 day just tinkering with comfyui

it loads ~25gb of write to the drive every gen

[nimble heart]

Fucking how. I've been using SD since 1.4 came out

mostly on that drive until recently

What's writing 25 gigs?

[dusty turtle]

Lmao. I really want the ram to arrive so this sort of nonsense can stop occuring (this is from just this session)

[nimble heart]

Ah

I've never used pagefiles

ever

swap on windows and page on linux are both 0

wait

swap on linux and page on windows

[dusty turtle]

I tried to get away with not getting more ram but llms are 100% bandwith starved so upgrade it is lol

[nimble heart]

ram's so cheap

[indigo carbon]

also, update on the whole TensorRT situation, NVIDIA stated that on the next TensorRT release they will release a script that converts almost any safetensors(including SDXL) into TRT

[nimble heart]

last thing you should cut corners on

[dusty turtle]

this was a while ago

[nimble heart]

the differene between 16 and 32 gigs on my 2017 machine was like $30

and until very recently 32 was enough to do everything under the sun without swapping

[native moon]

hey does someone know how to use comfy ui? becasue i have some proplems with loras and embedings. i dont know how to activate them i added them to the node tree but idk what to tipe in the clip text box to controll them

[nimble heart]

they're just active

it's not like auto where you need to type in a specific token to activate them

just go base model node -> lora node -> everything else

set the strength to 0 to turn it off temporarily

[orchid oar]

Does anyone of you have a good Img2Img workflow for SDXL 1?

[fast vector]

Trying myself with 1.5 ControlNet

[native moon]

ok thanks

and with the embedings?

[native moon]

1.5 stuff does not work

[nimble heart]

I don't have an actual img2img workflow .json but if you just add an image input into my principled node it behaves like img2img with all the settings you're used to https://github.com/Beinsezii/bsz-cui-extras

[nimble heart]

at all?

[fast vector]

img2img works

[nimble heart]

if XL works 1.5 should 100% work

[orchid oar]

Thx

[white iris]

controlnet with xl works ??

[nimble heart]

No

[fast vector]

no

[nimble heart]

the tooling does but it needs to wait for new models to be trained iirc

[native moon]

1.5 modesl loras or whatever does not work with xl.

[white iris]

right, saw the github discussion, waiting til its out then ill switch over

[vital wolf]

Im getting error" shape mismatch when trying to apply embedding, embedding will be ignored 768 1280" when I try to use embeddings in comfy

[nimble heart]

Oh no.

I thought you meant 1.5 on its own

[soft bone]

I'm trying to build automatic face-inpainting in comfy. But with CLIPseg it makes a solid mask that cant be denoised from the original image ("original fill" mode in auto1111)

Ideas?

[native moon]

no

[vital wolf]

Im getting error" shape mismatch when trying to apply embedding, embedding will be ignored 768 1280" when I try to use embeddings in comfy. Anyone help?

[dusty turtle]

are you using 1.5 embeddings for xl?

[vital wolf]

Yes

Using for dramshaper

Sdxl has different negative embeddings like BadDream or UnrealisticDream

?

[nimble heart]

fucking resetting GPU motherfucker fix your shit AMD fuck

[dusty turtle]

yeah the setup is not recommended now obv, Its super jank but works much better than it should tbh for being at like 99.999% memory usage for hours at a time

[native moon]

i am testing embedings for a 1.5 model

[nimble heart]

speaking of SSD wear I also set my default firefox download folder to my ramdisk and it's really nice 👌
unzip files and dick around with everything at then copy the final folder structure at lightspeed to your disk if you wanna save it. Don't even bother cleaning up cause it all vanishes when you power off at night.

also have my SD output folder on it

[native moon]

if u use comfy ui 32 is still engoh

[nimble heart]

recommend 64 just for that

[nimble heart]

SDXL + refiner I don't go over like 23 gigs

[native moon]

on automatic 1111 i go over 32 but on comfy ui only like a bit more them 20

[dusty turtle]

sounds pretty nice

[slender timber]

So, I left a SDXL lora on back on a 1.5 generation. It is not terrible..

Actually seems to handle upscaling pretty well

[fast vector]

huh? How you loaded SDXL lora inside 1.5 model?

[nimble heart]

very carefully

[azure oxide]

comfyui can do that

[slender timber]

comfy

[azure oxide]

actually it should be pretty easy to (accidentally) do on a1111 too lol

[nimble heart]

lets just answer anyone's questions they'll ever have from now on with comfy

[fast vector]

so, have you tried the other way around? 1.5 lora into SDXL model?

[dusty turtle]

spooky eyes

[nimble heart]

don't think that'd work. won't be big enough

or somehting

[slender timber]

I've not seen great results with that

re: 1.5 -> SDXL

[nimble heart]

I've found it likes to like making anime eyes kinda dead inside

[fast vector]

XL Loras are like insane size. Some on civit AI are at the size of prunned FP16 models

[dusty turtle]

I would not expect any 1.5 lora to produce anything other than random garbage on xl, they aren't even the same architecture

[slender timber]

Is that just a storage constraint? I've never worried about model size before

[alpine pine]

I think it has to do with using sd1.5 lora settings on sdxl, but I am also clueless at sdxl training

[dusty turtle]

I just run them through a 1.5 model at like ~.4 denoise and it seems to work pretty well

[nimble heart]

nah fuck 1.5 it runs like ass at high res

[orchid oar]

How to i get the denoising option for img2img?

[nimble heart]

2 megapixels it's like half speed or something

[nimble heart]

start at steps > 0

[dusty turtle]

lmao (agony)

[fast vector]

I mean the fact that 1.5 loras are like 15-20X smaller. Also you need to load them. So also performance

[nimble heart]

if total is 20 and it starts at 10 that's 50% denoise

[glad fulcrum]

sometimes I get this,.

I queue a prompt and it does nothing

[orchid oar]

Thxxxx, i needed that <3

[azure oxide]

the official sdxl lora example is only 50mb. the folks on civitai posting 500mb loras are doing something wrong lol

[orchid oar]

but does 0 of 20 mean 100% or 0%?

[fast vector]

I mean clearly. Wednesday one is like 1.7GB 😄

[alpine pine]

I think you need to use lower network settings as your replacing less of the model

and in sd1.5 you used higher ones

[dusty turtle]

base 1.5 model= 2gb lol

[nocturne dove]

There are a bunch of different lora types and the appropriate size for each type varies. Then for each type, there are size settings that change the size. You can't just say "lora is big therefore settings are wrong".

[heady vale]

I would imagine a contrast adjustment being a lot smaller than subject loras

[dusty turtle]

surely you dont need 1.5gb of weights for your like 100 image lora

[nocturne dove]

That's true for sure.

[azure oxide]

the concept is the same, you're just training on a (relatively) few images. the size should still end up the same. i think someone spoke about this above but theres been a lot of responses since then, there's settings that you can tweak when finetuning and the ones with huge filesizes are likely setting them incorrectly

[alpine pine]

I believe the setting that influences the file size is relative to the base model. So if you did a 50mb training on SD1.5, then it would be much larger on SDXL without needing to be

[fast vector]

since resolution is 4x, so I would expect 4x size

[heady vale]

assuming theyre using a few images

[fast vector]

Network Weights, etc are more important than they amount of images when it comes to size

[edgy otter]

did someone say street photography?

[queen ether]

76 images of 2k images down in 30 minutes. Woohoo Lora training

[echo stratus]

Waiting for controlnet.

[fast vector]

A400?

[queen ether]

I have a 3060 12gb :((

[glad fulcrum]

can you train Lora with 12 gb vram?

[fast vector]

For SDXL that's freakishly fast

[civic sigil]

Yeah 8 rank at 768 at least

[queen ether]

Easily with 1.5 extremely slowly will a batch size of one and gradient diffusers enabled on sdxl

[glad fulcrum]

I need to learn how to do it then. I have 12 GB VRAM. Do you have a tutorial for sdxl training lora

[queen ether]

The best one was was se courses guy

[glad fulcrum]

I don't understand what you are saying

do you have the link?

[queen ether]

He's a YouTuber

@fast vector 90 images in 30 minutes for a 3060 12 GB? I can't believe how much power sdxl demands :((

[fast vector]

except that I asked one of SDXL Lora's creators and for them it took much longer.

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

[queen ether]

I feel like I might give up on sdxl Lora training and maybe only train a model when I go to sleep or go to work. This is wild

[fast vector]

Yeah 3.5 hours is too long to be honest

[alpine pine]

im still experimenting, but it looks like you can get a satisfactory result in less samples with SDXL.

My exact SD1.5 lora settings on SDXL took 3.5 hours or so on a 3090. Results were bad. I, then, did a different training with new settings (same image count) and less samples and it took 30 min with much better results.

[fast vector]

wait you're mooncryptowow? Or is it just coincidence?

[heady vale]

Stack Overflow are releasing their own Overflow AI, specific for AI/ML. interesting

[queen ether]

Which settings did you change?

[alpine pine]

pretty much everything tbh, I still havent found good settings... just better-than-before settings

[fast vector]

man, I actually wish if comfy had something like autocompletion. It would be great for embeddings, etc

[raw cedar]

does old textual inversion work on sdxl ?

[nocturne dove]

Old nothing works with sdxl. You have to train new textual inversions and loras afaik.

[raw cedar]

i see

[deft coral]

Can SDXL 1.0 even be trained on

i can't seem to train on it

[glad fulcrum]

I think textual inversion will not be able to work on sdxl. At least that's what I remember reading

[deft coral]

i get a error everytime

[fast vector]

I mean there are already loras and even checkpoints

[deft coral]

when i try to train 1.0 i keep getting mussing unet

missing

[nocturne dove]

I haven't tried to train the actual model. Loras and so on weork fine. I use a branch of the kohya scripts.

[alpine pine]

theres a bunch of new and different settings you need

[alpine pine]

and only one video tutorial that is super bloated

[deft coral]

such as?

[alpine pine]

https://github.com/kohya-ss/sd-scripts/tree/sdxl

this has a summary of some

[nocturne dove]

I just grabbed updated scripts, and trained my loras like normal. I only needed to set the res to 1024 and check the box for an sdxl model.

[deft coral]

ya i dont want lora

[alpine pine]

do you have an example of your settings / output?

[deft coral]

i wanted to train a checkpoint

[alpine pine]

ah, sorry I never tired checkpoints

[nocturne dove]

Yeah I have not done checkpoints either for sdxl.

[deft coral]

i just get unet issues

[nocturne dove]

I don't have optimal settings. I don't want to post samples trained from real people here.

[deft coral]

they added 0.9 vae to 1.0

im downloading it

[fresh path]

does comfyui not install on python 310

[deft coral]

no

you need to install that manually

[fresh path]

No I mean

[deft coral]

i got comfyui on 3.10.9

[fresh path]

I tried installing pip dependies and some packages aren’t available

[deft coral]

odd

it works for me

on 3.10.9

[visual glade]

yeah it should work

[boreal bough]

which is the ideal AI python version

[fresh path]

specifically torchdiffeq

[deft coral]

i use torch 2

[visual glade]

I use python 3.11 and the standalone ships with python 3.10

torchdiffeq is pure python so it should have no trouble installing

[deft coral]

3.11 i dont use

since it doesnt support pytorch

[queen ether]

One hour and a half in and I just hit my first epoch out of 20 for SDXL Lora training

[boreal bough]

5k dataset? xD

[visual glade]

well you made me notice torchdiffeq wasn't actually used for anything important

so I'll remove that dependency

[queen ether]

I have a total step count of 1800 😦

[boreal bough]

even then O_o

[deft coral]

bro

i can get 1800 done in 10 minutes

[boreal bough]

you going over your vram by any chance?

[native moon]

hey how do embeddings in comfy ui work. wehn i am using them do i need to put the file name or the embedding name into the clip text box?

[deft coral]

[queen ether]

@deft coral what graphics card do you have

[deft coral]

RTX 4090

[visual glade]

https://comfyanonymous.github.io/ComfyUI_examples/textual_inversion_embeddings/

[queen ether]

@boreal bough I have 12 and I believe it was asking for 17 I got it to work by enabling the gradient something or other

@deft coral 3060 12 gb poverty gang over here

[boreal bough]

oh damn. you're running on standard ram now - which explains the speed

[native moon]

it insnt realy explained ther

[nocturne dove]

Here's the results of a "spider-gwen" lora for sdxl 0.9. I haven't re-trained it for 1.0 yet because I need to change the dataset some and I haven't gotten around to it yet. In this attempt, the fact that I mixed costume, no costume, mask, no mask, made the lora somewhat ineffective.

[native moon]

in ther case the file has the same name as the embeding

[queen ether]

@boreal bough I don't know what that means 🙂

[visual glade]

embedding:filename

[deft coral]

@queen ether

[boreal bough]

basically, once you run out of vram, it doesn't crash nor stop working, but it gets slow as hell
that's what you're experiencing

[native moon]

but 1 embeding file can have multible different embedding names in it

[visual glade]

can you show me one?

[queen ether]

@deft coral my GPU is held up by a paint brush

[boreal bough]

you're missing some optimization settings, since multiple here run it on 12gb vram

[alpine pine]

Is there a guide your following for your settings? They look great

[queen ether]

@boreal bough is that a command argument ?

[native moon]

ok wait a sec

[boreal bough]

are you running kohya-ss directly, or via the ui, or via derrian easy lora trainer?

[queen ether]

@boreal bough UI :0

[native moon]

this one has 2 and ther are others with even more https://civitai.com/models/89484/epicrealism-embeddings?modelVersionId=95263

[boreal bough]

then xformers, gradient checkpointing, 1e-3 unet & learning speed. dim/alpha = 8/1 ,and the command argument for unet only

[deft coral]

does 1.0 even have unet

#

[boreal bough]

--network_train_unet_only

[fast vector]

except that thee are two different ones. One in positive and other in negative prompt

[visual glade]

I only see one

that one should work fine on comfyui

[sour obsidian]

I hope it does

[deft coral]

atm it doesn't

everytime i tried to train its missing unet

[native moon]

it has 2 : 1. epiCRealism 2. epiCNegative and the file name is epiCNegative.pt

[fast vector]

no, there are two files

[queen ether]

@boreal bough I was reading this on the GitHub but I don't know where to put it. I know where to put it in stable diffusion but I don't know where to put it in koya SS I have something filled in in the optimizer extra arguments that starts with scale parameter is that it?

[nocturne dove]

No but if the next one I train works ok, I'll post some details for you here. I think I copied settings from here in my more recent lycoris attempts. https://youtu.be/dUlki1IAB0w?t=460. But I used the prodigy optimizer with learning rate = 1.0 and the loss doesn't drop much after around 1000 steps or so using around 40 images.

[sour obsidian]

id imagine it would be hard to make any images without it having one haha. Its def in there

[native moon]

no it just donlodes 1

[boreal bough]

under the scheduler in the ui, there should be an empty text box for additional arguments, put it in there

[fast vector]

just click on other one

[native moon]

its just 1 file

[deft coral]

my system just locked up

[wet rock]

hi friends, whats the recommended vram for running SDXL?

[deft coral]

i just transfered a 12GB files in 2 seconds

[visual glade]

{'string_to_token': {'': 265}, 'string_to_param': {'': tensor([[-0.0034, 0.0213, 0.0007, ..., 0.0280, 0.0048, 0.0047],
[ 0.0179, 0.0198, 0.0057, ..., 0.0105, 0.0117, -0.0043],
[ 0.0095, 0.0053, -0.0120, ..., -0.0096, 0.0066, -0.0025],
...,
[ 0.0037, -0.0079, -0.0430, ..., -0.0173, 0.0184, -0.0008],
[-0.0271, -0.0072, 0.0129, ..., -0.0072, -0.0059, 0.0009],
[-0.0165, -0.0159, 0.0144, ..., 0.0192, 0.0016, 0.0220]])}, 'name': '_EmbeddingMerge_temp', 'step': 0, 'sd_checkpoint': None, 'sd_checkpoint_name': None}

there's only one

[native moon]

ohhhhhh

[wet rock]

i have 10gb VRAM but when i select the SDXL model, my GPU is stuck at 9.7/10gb 😭

[native moon]

sorry this is the first time i am doing embedding stuff so i am a bit of a noob sorry

[queen ether]

@boreal bough I could keep asking questions forever 😓

[boreal bough]

dont stop ^^ I'll gladly help when it comes to LoRA questions

[native moon]

i didnt see that i have to downlode 2 files i thoght its both in 1 file sorry

[visual glade]

are you on the latest? try update/update_comfyui.bat

[native moon]

at fist i thogh comfy ui looks harder then a1111 but its the opposite.

[deft coral]

wtf

[visual glade]

which lora and can you give me a screenshot?

[static prawn]

is there any sheet for all the styles for sdxl? so we know how we can get a specific style?

[fast vector]

yeah, there is.

[raw cedar]

https://www.reddit.com/r/StableDiffusion/comments/15afvnb/sdxl_various_styles_keywords/

[deft coral]

so i add this to mine right

--network_train_unet_only

[boreal bough]

assuming you're doing normal training, with 1 trigger word at the start

then these are your advanced settings - where xformers and such is included

only the network train unet needs to be in that box (if you're using adamw8bit), if you're using an dadapt or other cool scheduler like that, then yeah, the remaining special settings also go into that box

[sour obsidian]

#✨｜sdxl message also just pinned right here haha

[visual glade]

that's not the proper lora format though

[boreal bough]

@visual glade where do I voice my complaint about the civitai lora promotion? :/
terrible loras are mass produced by the minute right now

[wicked frigate]

"woops, we forget the whole unet, guess the vae's just that good"

[sour obsidian]

only if its clip guided

[wicked frigate]

on the civitai discord probably

[boreal bough]

T.T

[visual glade]

I'm pretty sure a1111 doesn't support that format either

is this the diffusers format lora format?

[native moon]

the bad ones will get downvoted

[deft coral]

you can't train in A1111

don't even try

[boreal bough]

its just a matter of considering 2 basic things
A.) this is not 1.5 - stop reusing the exact same captions - especially when they don't work in standard generation to begin with
B.) use dim/alpha of 8/1 unless you know what the difference is of not using this preset

[queen ether]

i was told to use that argument in there because SDXL doesnt work with adamw8bit im using afactor. i also have almost all the same settings enables as you other than memory efficient attention so youre saying i should switch to adamw8bit and use the --network train argument there intstead?

[boreal bough]

? fake news XD I've trained 20+ loras on sdxl with adamw8bit

all very successful

[boreal bough]

#✨｜sdxl message

Unet Training: 1e-3 <- ideal setting for 95% of situations (From dataset size of 10~1000)
Dimension/Alpha = 8/1
Unet training only! (--network_train_unet_only)
Resolution = 1024,1024
Bucket size = 512,2048
repeats = 1
epochs = around 20 should be where the model is 'perfect'. Train for 40 to be sure. (scales very slowly with dataset size - but not nearly as much as in 1.5)

[visual glade]

is there an example diffusers format lora somewhere?

it's like 2 lines of code to add support so I can add it as long as I can actually make sure it's an actual format and not an output of buggy software

[hard fractal]

ooo

Fancy

[queen ether]

ive been reading about the --network_train_unet_only thing for the past two days now an i just have no idea just ZERO idea how to enable it

[boreal bough]

"easiest way to train stable diffusion xl"
couldnt be more wrong, dear god
almost everything said in this video is not true T.T

[visual glade]

so there's no diffusers lora files anywhere?

[hard fractal]

🤷‍♂️ looks like kohya it is.

[hard fractal]

Perfect prompt following... as long as your prompt is "random noise".

[vital wolf]

Im getting error" shape mismatch when trying to apply embedding, embedding will be ignored 768 1280" when I try to use embeddings in comfy. I saw drramshapersdxl alpha 2 model using BadDream negative embedding, can you help? @visual glade

[hard fractal]

don't tag comfy

for tech support pls

[vital wolf]

Ok, where do I get the support then? I really don't know, thanks

[hard fractal]

here may be good 🙂

[hard fractal]

just don't ping the devs pls

[vital wolf]

Okay.

[azure oxide]

baddream is a 1.5 TI

[hard fractal]

your embedding is prob-- yeah that

[visual glade]

so it will actually work a bit on SDXL but will only apply to the clip-l text encoder

hence the warning

[vital wolf]

That I know but saw the sdxl model dreamshaper using it, saw on meta data in civit

[deft coral]

hmm

[vital wolf]

Okay, so it works to a point?

[hard fractal]

it'll half-work.

and by half, less-than-half... since L is weaker than G

[deft coral]

Joe

[visual glade]

it will do something but it might not be the thing it's supposed to do

[deft coral]

@visual glade can i ask you something

[wicked frigate]

should use kohya gui

[visual glade]

sure

[vital wolf]

So better not use it?

[west breach]

I made a node to apply a HaldCLUT to an image. Would it be possible to apply it to a latent? Not sure how the numbers represent RGB values.

[deft coral]

does SDXL have unet?

[vital wolf]

Oh ok

[visual glade]

yeah it has a unet

[deft coral]

because i can't train it for some reason

i keep getting missing unet

[hard fractal]

comfy, can you check to see if we included the unet with SDXL

[visual glade]

that's probably a problem with your software then or your settings

I don't think people would be complaining about memory usage if it was missing

[wicked frigate]

lmfao

[spark bear]

just stare at the noise like clouds and draw what you see, no need for unet

[midnight shuttle]

Boring old SDXL still requires a unet???

[azure oxide]

guys, theyre NEW lmao

[deft coral]

yes

SDXL still needs unet to train

[wicked frigate]

we might've replaced the unet with a tiny guy that sits inside your pc and draws the pictures for you, not sure

[midnight shuttle]

😕

[boreal bough]

"beat unet guy into submission only"
poor unet man :c always drawing. never rewarded

[deft coral]

so iv tried koyha ss

and no luck

im running out of ideas

[hard fractal]

Try a LoRA first

[wicked frigate]

are you sure you have the correct SD model file? The official standard one not a diffusers file

[deft coral]

i can't even get the lora to work since it needs the main model

[visual glade]

maybe it's actually a wnet from sticking two of them together

[spark bear]

Freon needs to get Everydream working on 1.0 so we can see how screwed we are trying to do full finetunes on 24 gigs of ram

[hard fractal]

oooo

wnet.

I'm in

[deft coral]

joe

[hard fractal]

ryzen

[boreal bough]

using what checkpoint? xD we only have fp16, or the research licenced full 0.9 checkpoint

[wicked frigate]

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

[hard fractal]

okay your issue is trying to do this in automatic1111

[wicked frigate]

i think that's kohya gui, which is just a gradio shell around kohya-ss scripts

[boreal bough]

"no, ignore, bad"

love it XD

[hard fractal]

Ah, got it.

[hard fractal]

https://youtu.be/AY6DMBCIZ3A

Try that tutorial step by step

[autumn forum]

why are you trying to create a model? lmao

[hard fractal]

Then, you can ask beginner questions here: https://discord.gg/wFYWAykv

[spark bear]

heh I guess he's got it working and testing it so we'll find out.. and what happend to the 1.0 with .9 vae sai released today? did that get sorted out?

[hard fractal]

Just renamed the server

Since it's not only dreambooth

[midnight shuttle]

Any insight on why previous models' VAE worked with fp16 but SDXL VAE requires fp32?

[hard fractal]

Now I'm concerned we didn't ship the right file tho.

@sour obsidian - can you check?

And make sure we included the unet?

[wicked frigate]

jonatas

[hard fractal]

alexander

[sour obsidian]

yea, I sometimes forget that, might of slipped my mind

[deft coral]

i don't like lora training

i train full checkpoints

[visual glade]

the previous models VAE also has issues in fp16

at least one of them did

[deft coral]

[wicked frigate]

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix this should work for fp16?

[hard fractal]

fp8 wen

[midnight shuttle]

I hadn't seen that before. 1.5 and 2.1 both worked without --no-half-vae. But SDXL I had to add it. Strange.

[deft coral]

koyah has suport for fp8

[visual glade]

there's enough issues that I made comfyui always run it in fp32

[midnight shuttle]

I've been using that one. I just prefer to use the baked-in VAE so it's one less setting to change when I change models. (Can't figure out how to have A1111 pick up a .safetensors VAE automatically)

[wicked frigate]

when emad gives us even more h100s to play with

[deft coral]

i use no-half-vae

[wicked frigate]

we need at least 50 dedicated to 24/7 cat pic generation before we can get around to productive usages like fp8 testing

[deft coral]

wait

im stupid

i think thats whats causing it

i used --no-half-vae

[hard fractal]

to... train?

[vital wolf]

Is there any equivalent of no half vae for comfy?

[deft coral]

no

[turbid gulch]

Playgroundai.com just launched the first batch of SDXL 1.0 filters. They're super cool.

[deft coral]

in the command line

[trim orbit]

i think stability should run a contest where i win and get a h100

[midnight shuttle]

I don't understand enough about how that would affect the unet. How is it related?

[deft coral]

because no half wont train fp16

only fp32

and if SDXL used fp16 it wont work

[hard fractal]

well, we can neither confirm nor deny the inclusion of the unet.

[trim orbit]

might not be very beneficial with sdxl. the base is so good that none of the refined models that have come out have really impressed me much. they all seem to ruin the general use capabilities

[deft coral]

i know what your saying

[midnight shuttle]

I'm hoping for SDXL hypernet training someday and also inpainting model.

[deft coral]

but i have my own dataset

[molten breach]

Can't wait for inpainting model

[boreal bough]

joe. I just checked that video and it has a few critical things wrong.
it's missing the unet only, it's dimension is set to 256... ... ... and a lot more

essentially any LoRA trained using this video will
• significantly harm clip model. longer prompts wont work correctly anymore - which is what more and more people that rely on this tutorial keep mentioning in this very chat
• create huge filesize, when 43mb is all that is needed. especially for sub 500image datasets
• not mention how incredibly important captioning is, even more now on sdxl
• using ohwx token, which was a 1.5 thing, as we have new token weights now

and more but I dont wanna spam you.

[turbid gulch]

Playgroundai has inpainting already

[trim orbit]

most refiners do too. lots have come out and none have really improved things yet

[turbid gulch]

Nothing local tho

[midnight shuttle]

Inpainting model (works well with inpaint only masked)? Or just inpaint feature in general?

[potent shard]

I guess sdxl has little to no knowledge of slime rancher...

[turbid gulch]

It's all there, including ControlNet

[midnight shuttle]

So they have access to an unreleased inpainting model? That's good news because it means it will be released eventually.

[molten breach]

Can it do latent nothing inpainting that can add totally new objects? That's mainly what I did in 1.5 inpainting

[deft coral]

@boreal bough was i right then>

[turbid gulch]

It's identical to 1.5

[trim orbit]

loras can train on large datasets too. i think the community is just going to move towards loras. they're so much more efficient and when they're performing better than all the base models that are available, people will start making lora merges and all the scene will happen there

[molten breach]

Sweet

[midnight shuttle]

Same here. That's why I am waiting for it and why I am still on A1111. Can't do that with Comfy as far as I know.

[turbid gulch]

And free. They give you 1k images a day for free. Their sub model is like $15 bucks a month.

[boreal bough]

I'm guessing your info came from that video, since it explains the settings

[trim orbit]

inpainting with comfy is possible it's just a pita

[molten breach]

I did some inpainting in comfy with a 1.5 model but I wasn't very good at it

[deft coral]

@boreal bough i never watched the video

thats why i kept asking

[midnight shuttle]

Inpainting whole image is possible. I have not found how to replicate "inpaint only masked" but I don't need it yet since there is no inpainting model for SDXL yet.

[turbid gulch]

Playground is building what they call "Canvas". It's amazing. It does have a couple glitches. But new model coming soon.

[trim orbit]

you'd set up a crop node and inpaint that whole cropped picture. then paste it back onto the coords it came from. that's what auto is doing behind the scenes for inpainting only masked. they just crop to the mask

[boreal bough]

ah yeah, but the training settings in it basically mirrored the ones you had
guessing it's a game of broken telephone in regards to lora settings.

Um the settings I linked you I can guarantee will work out in a good lora

[molten breach]

I'll def check it out this weekend

[turbid gulch]

With Canvas and Davinci Resolve you can do everything you do at Adobe.

Free

[midnight shuttle]

But it sounds like some site has access to the pre-release inpainting SDXL model.

[midnight shuttle]

Hmm...that makes sense. But sounds like a lot more work in Comfy especially for iterative generations (multiple cycles of inpainting).

[trim orbit]

yeah. like i said. total pita

[boreal bough]

@deft coral
if you use these settings,
then I guarantee you a good lora

#✨｜sdxl message

[turbid gulch]

This is what their Canvas looks like. Inpainting is at the top.

4k

[deft coral]

@boreal bough

[midnight shuttle]

I hope A1111 can get SDXL working fully and with better VRAM usage (needs 16 GB right now - 14.5 used max for 1024). Or else mayne SD.next will be an alternative. I still need to look at that one.

[deft coral]

i can't even get it to work

im just gonna wait for A1111 to fix SDXL

[midnight shuttle]

But is it an inpainting model? There is a big difference for certain workflows.

[limpid rose]

A1111 works with SDXL for me. have you tried nuking your venv folder?

[boreal bough]

kohya?

If you use https://github.com/bmaltais/kohya_ss
it should be an easy and automatic install

[deft coral]

koyah doesn't work for me either

[deft coral]

i get the same unet issue

[boreal bough]

https://github.com/derrian-distro/LoRA_Easy_Training_Scripts/tree/SDXL

use that. its foolproof, and uses kohya in background

[turbid gulch]

Inpainting with prompt + controlnet

[midnight shuttle]

I don't understand. Is that a yes?

[queen ether]

after like 2 hours of training only 3/20 epochs completed kohya finally crashed ❤️ i love god

[turbid gulch]

Yes

[midnight shuttle]

OK that's good news. Thank you. It means the unreleased inpainting model will hopefulyl be available someday.

[turbid gulch]

It doesn't work on 4k - 32k images. You have to shrink them, then it works.

Their max upscale is 4k

Davinci Resolve will upscale to 32k

[deft coral]

@boreal bough how can i nuke my ven

[boreal bough]

@golden quarry would it be possible to get an update for kohya?
gonna try and get the lora training tutorial out on sunday, and wanna link your distro for it.

[boreal bough]

just delete the folder, then run setup again.

[turbid gulch]

I've seen your name before. Are you a game dev?

[boreal bough]

venv folder*

[deft coral]

so fresh install

[golden quarry]

yeah I can

[boreal bough]

thank you ❤️

[golden quarry]

I was pretty busy overall

[dry anchor]

Hello guys, is there any lora training guides for sdxl1.0?

[boreal bough]

#✨｜sdxl message

[west breach]

so how do you train a lora without trigger words?

[boreal bough]

depends? what are you training?

[trim orbit]

both kohya and automatic1111 with the db extension load up with xl for me. i don't think the fault is with their commits

i think its a local issue

[queen ether]

--network_train_unet_only is not an option in my kohya and i have no idea where to put ❤️

[west breach]

I was thinking of training an art style

[queen ether]

kill my self

[trim orbit]

in the gui? extra commands are in advanced

[midnight shuttle]

Is ther a good way to get A1111 to detect a .safetensors VAE that is named the same as the model? It only seems to want to use .pt VAE files that way.

[cunning geyser]

So i want to train an anime character on sdxl, can anyone teach/tell me how to do so?

I kind of dont know what settings to sue

use*

[boreal bough]

I would still recommend a trigger word, as it will make your life easier.
but if you wanna do it without anyway:
-> Use Interrogator with Vit-H to auto caption everything. Accuracy doesn't matter too much. its only important you have enough tags.
shuffle on
have a dataset that is big & varied enough that enough concepts show up (car, house, person, toaster, whatever - just make sure its not all of a person, otherwise ALL captions will learn to create people xD)
I'd recommend 100~500image dataset if you wanna do without trigger word
50~100 to get a good varied style with trigger word

[trim orbit]

the hardest part is building the dataset

[cunning geyser]

Oh thats easy

I just don't know what settings work

[west breach]

thanks for the mini guide 🙂 I will give it a go later

[trim orbit]

with dadaptation trainers, settings are 1 and 1

[cunning geyser]

huh

[trim orbit]

https://rentry.org/LazyDAdaptationGuide

[west breach]

I trained a 'grumpy style' lora which makes any face grumpy looking. But you need to add 'grump style' to the prompt

[boreal bough]

does it still work well for sdxl? haven't gotten around to trying the dynamic ones yet

[trim orbit]

i've been using the newest form, prodigy, and getting great results. might try more with dadapt. but yeah it works. i have a 4080 with 16gb and do batches of 2. some models i've even bucketed at 768 pixels

[boreal bough]

noice. will have to give it a go before I post my guide

[heady vale]

some funky eye lashes going on here

[midnight shuttle]

Looks like neurons.

[queen ether]

you've saved my life

[nocturne dove]

I tried one of the other adaptive ones, (maybe adafactor?), and the loss curve looked identical to prodigy but prodigy was just a little lower across the whole training session. So I just went back to prodigy.

[pallid sail]

They look like l-systems

[heady vale]

fractal lashes

[trim orbit]

i believe in you

[deft coral]

bro

this is so good

imagine if i can make my own checkpoints

@boreal bough

this has so much potential

[boreal bough]

👋

my biggest achievement on sdxl so far, is getting finetune level of improvement, by training a 5k datasets, resulting in a mere 43mb LoRA file 😄

got over 100 learned concepts working

[trim orbit]

hmm. i don't know how to make comfy ui stop fudging my symlinks and rewriting new folders over them

one second, going through all my loras, next second, "undefined" cause it decided to make a folder there

[autumn forum]

you gunna enter the compitition ?lol

[trim orbit]

all the checkpoint refinements i've seen so far have underwhelmed. while loras push the base in new ways.

[boreal bough]

true finetuning gonna be expensive as hell if you're in a rush
also we only got the pruned model, so you're either stuck using the 0.9 research license full model
or essentially just merging a lora with the base and calling it a "finetune"

[trim orbit]

Hm. Civit is throwing a contest. But it's also kind of "A model that is in dire need of some tweaking." ish. Which means the contest is for a porn model. If i enter the judges won't pick mine.

[autumn forum]

non nsfw only

[boreal bough]

sfw only

rules! XD always read them

[trim orbit]

no no what they mean is that the submitted example pics are sfw

[autumn forum]

oh i forgot he has me blocked lmao

[trim orbit]

they're very skirty like that. when they blocked NSFW submissions, all they did was make pornographic merges show up in general results, but with their porn images not showing.

the 3 winners will be boobs, bazongas, and butts

[deft coral]

@autumn forum i would if i could fix the unet issue

so i can't

[trim orbit]

when their call to training says that sdxl is in dire need of training, we know what that means

[vale eagle]

How do you caption with this size dataset?

[boreal bough]

either way, lykon gonna get one of those 4090s for the dreamshaperXL base - since by contest rules its gonna beat everything by virtue of being used as a base by many others
and for the other two :/
character is hard as hell to get right, as you'll get to experience all the downsides of the refiner - meaning that lora will kinda cheat by being an existing concept, that is only reinforced well
style is the only true competition, as its also the only one that can be trained on all machines, and is by far the easiest to train

[trim orbit]

automate it with blip and run some filters on it to remove bad captions. you set up processes and work it until you're happy

[trim orbit]

he's also got eyes to start. i'll just buy a 4090 instead of trying to beat a popularity contest trying to boost civit's traffic

[vale eagle]

I used blip but I am not very satisfied with its result

[boreal bough]

yep. what flowwolf said is what I did.
Interrogator running Vit-H

but I really wouldn't recommend it as a starting thing for sdxl, as big datasets are the hardest to get right - and each failure punishes you with 20~40hours of training time (on an rtx4090)

[midnight shuttle]

Curating and captioning data sets is the hardest part of any training. The actual training process is easy.

[trim orbit]

my last lora with just 200 images cooked for 19 hours over 2 runs

[deft coral]

im still salty

[pallid sail]

I've noticed these weird pixel patterns showing up in my SDXL images. Is it to do with the scheduler? Attached image is via Diffusers with SDXL + refiner

[boreal bough]

please use 1e-3 8/1. 14 min are eeeeeennnoooouuugh XD

[vale eagle]

I tested a training with 15 images with selfmade caption. It already able to produce decent features of original images.

[deft coral]

is this a joke

[trim orbit]

apprantly thats the digital watermark making it obvious it's ai gnerated. we have to keep that in according to the license

[deft coral]

24GB for 1 image

[trim orbit]

i think

[deft coral]

[boreal bough]

XD

you got a setting error somewhere

[trim orbit]

yeah. higher learn rates seem fine i've noticed

[boreal bough]

I run a 4090 with 24gbvram and can do batch 8

[deft coral]

how

[midnight shuttle]

The watermark is invisible. There is some problem with the VAE. Try this model: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0_0.9vae.safetensors

[trim orbit]

is that at 1024x1024? i've only tried batches of 2

i think i need to tell windows to put my page file on the new gen 4 ssd

[boreal bough]

10~12 is the theoretical maximum on 24gb vram with adamw8bit
I use 8 cause I still like using my pc

[trim orbit]

its nice to not have 5fps during training

[boreal bough]

ooohhhhhh. yeah I was talking about training a LoRA

[trim orbit]

thanks for the advice. i'll try it on my next training run on that same dataset

[boreal bough]

batch 3 is maximum on 24gb vram

[pallid sail]

Thanks! How can I pull only that file from the hub and replace my existing one? Is there an argument for the hf_hub utility that will let me pick a particular safetensors file from a repo?

[deft coral]

i used to do batch 10

on 24gb

[boreal bough]

and that's using comfy, as A1111 has a lot of unneccesary overhead

[trim orbit]

no this new pixel pattern is different from the vae issue. it shows on comfy gens in the new version

encoding the watermark as pixels prevents the embedded data from being lost in transit

automatic probably won't have it enabled by default

[pallid sail]

Ah, I was wondering if that was going to be the case. If it was just kept in the headers then I imagine it might get lost when being saved/exported into different formats

[dire rampart]

are you denoising in the refiner stage or adding noise in between the two stages? The refiner usually goes over those pixels but if it is being stopped from working on a noised object it might only gussy up the pixelation

[trim orbit]

yeah like there was a quick minute where discord was stripping metadata

[nocturne dove]

How do we know the pixel pattern is a watermark? Where did this info come from?

[turbid gulch]

Emad on Twitter.

[pallid sail]

I generate an image with the base model first, save the latent data and then pass it onto the refiner. I'm using the denoising_start and denoising_end arguments to only refine in the last 30% of the total steps specified

[nocturne dove]

Thanks

[midnight shuttle]

I have never heard of that utility. I just download the file and put it in the right place.

[pallid sail]

It's the library diffusers uses to pull models from hugggingface

[trim orbit]

because if you remove the invisible watermark code, that pixel pattern which is only there if you look for it, disappears

[midnight shuttle]

I haven't gotten as far as writing my own code to use diffiusers. I just use the pre-built UIs.

[pallid sail]

Gotcha. I'm using diffusers in Unreal Engine

[trim orbit]

the vae issue produces interlacing lines, it might even be intentional to give a chromatic abbrasion effect

it doesn't happen on my hardware though so i don't really know

[nocturne dove]

Welp, time to look through the code I guess. Oh well.

[trim orbit]

i was trying to use stable studio this way, but comfy ui hates symlinks and keeps erasing mine

[trim orbit]

i've got stable studio loaded but don't want to use a seperate install. so i'm trying to point it at my existing comfy and models but it all hates symlinks so much

[dire rampart]

That is strange behaviour with the refiner, I would try adjusting the latter percentage and seeing if additional or less refinement is appropriate, my setting of 1/3rd is approximately free of pixellation but if I do a run with no refiner to mass produce compositions then it will have those issues, but I can just run them through again with noise

[deft coral]

so i was right with the unet missing

ill just wait for it to get added to the main checkpoint

[west breach]

@Caith do you have a favourite tool for tagging by hand?

[boreal bough]

unet is not missing. it's like 95% of what makes the model work.
there are currently issues with the vae (color encoder)
so if you use the 1.0 vae, then you may end up with funny colored lines near mouths - but nothing critical

[boreal bough]

absolutely! Hydrus Network

[deft coral]

if thats the case

[boreal bough]

best for manual tagging. also scales well for bigger projects

[deft coral]

why the hell doesn;t it work for me

[boreal bough]

takes a while to learn it once - but absolutely worth it

[pallid sail]

I'll try with 0.95 base and 0.05 refiner and see what happens

[west breach]

Thanks I will check it out. I was using FastCaption before

[midnight shuttle]

Some reason other than a missing unet. Without unet you cannot generate any images from the model. The model is useless.

[deft coral]

and that's what im trying to pin point

[boreal bough]

for automated tagging, use A1111 -> extension "Interrogator", load the Vit-H model, and have that batch generate your captions. best (quick and easy) automated workflow for now

@west breach ^

[midnight shuttle]

I don't think anyone knows the answer. But the unet is not missing. That is not the real problem. It can't be.

[deft coral]

if thats the case then why is koyah ss telling me diffrent

[midnight shuttle]

The error message is not correct. You need to investigate other options.

[west breach]

I used BLIP originally. But the captions were very short and non descriptive

[deft coral]

what other options can i investigate

iv did a clean install

[native moon]

looks cool. so you can setup all the nodes and then use the other ui to just ganerate

[midnight shuttle]

I don't know. Start by reading all available documentation about how the system actually works. Not about how to use it but how it works. Then use those learnings to identify the problem. No one seems to know the answer. So you must do this.

[deft coral]

[midnight shuttle]

It is new technology. If you wish to use it you must be willing to learn.

[trim orbit]

oooo my terran lora works nicely when i don't prompt for what i was captioning and just do natural prompts like i would on the base model

[dire rampart]

haas anyone tried KosMos2 from Microsoft? It would do a visual scan and return tags

[boreal bough]

is one of the many existing models.
the reason I recommend Vit-H, is because it's close enough to sdxl, that it can be used fully automatically