✨｜sdxl

[nocturne trail]

300 works fine too

I mean a friend of mine bought two full brand new active backplate block sets for his two A40s for AU$1200 total

[vital ermine]

see, damn expensive

[nocturne trail]

they're solid copper and low volume

[vital ermine]

That is the only reason I never grabbed one of the server cards

[nocturne trail]

or you can use two $10 fans and a 3d-printed fan duct

[somber walrus]

AYAYA

[vital ermine]

I ceased 3d printing 4 years ago but the fans are loud

[nocturne trail]

people sell the ducts on ebay for about 10 bucks

and the fans don't have to be loud

[vital ermine]

yep

[nocturne trail]

they don't actually need as much airflow as you might think

[vital ermine]

every one I saw using them you have to leave the room

[alpine pine]

would like some noctua 120mm work?

[nocturne trail]

ehhhh

[alpine pine]

😦

[nocturne trail]

generally you want a blower or two 40x28 server fans, which can be found in "reasonably quiet"

it's just not straightforward

[vital ermine]

it has to blow down a channel it has. These were build for datacenters

[nocturne trail]

(been there done that)

I am quite aware

[alpine pine]

ah fair

[nocturne trail]

they're built for servers like the R740XD that my two are in

[vital ermine]

wow, nice

[nocturne trail]

on a 4090? there's no backside VRAM on 4090 or 3090Ti, and an EK block is ~$390 USD https://www.ekwb.com/shop/ek-quantum-vector2-fe-rtx-4090-d-rgb-abp-set-nickel-plexi

the A40 heatsink is actually identical to the A6000 one

and the shell is almost identical

you can swap the cover and fan from an A6000 on

but the bastards depopulated the fan header and PWM drive components, it's a Weird Tiny Connector, and the VBIOS won't drive a fan anyway

you can duct a big ole blower onto the back of it if you have room in your case but a pair of the quieter AVC 40x28 fans is usually the go

[vital ermine]

Just saying what Jay used and some I have seen are 500ish. I just purchased my first AIO and when it goes I will go open loop.

[nocturne trail]

yeah 3090 is a bit different

as are the 48GB cards

because you do need to cool both sides of those

[vital ermine]

A40 seems to be about 5-6k used being a Tesla A40 ampere.

[nocturne trail]

also, little PSA: if you're running a 4090 and you're using xformers, either don't (and just use torch SDP attention) or compile it yourself; the precompiled wheels for xformers are built with a stupid TORCH_CUDA_ARCH_LIST that does not include binaries for Ada or Hopper

[nocturne trail]

i bought two for US$4k total

[runic lark]

I've got no real tech skills, but is it possible with double GPU's, and if so even useful to add an old NVIDIA GeForce GTX1080 8GB GPU?

[nocturne trail]

GTX 1000 series are a waste of time for AI

[vital ermine]

tell me about it with my 1060

[visual glade]

yeah 1080 is pretty slow

[spice skiff]

Thank you. In my MX Master 3S Logi Options+ app, I had it set to "smooth scrolling", which is great for websites, but no so good for ComfyUI as it turns out. Too bad I have to sacrifice one for the other, but Comfy takes precedence at the moment.

[nocturne trail]

Pascal can't do fp16 worth a damn

[vital ermine]

nope, emulated

Turing gave us real fp16

[nocturne trail]

check out the advanced mouse options control panel thing where "enhance pointer precision" lives

[nocturne trail]

technically it's not even emulated

it just runs as fp32

[vital ermine]

Which is what I meant

hands it off to fp32 but accepts the fp16

1650/1660 will not even do that

[nocturne trail]

yeah it's cast back to fp16 when saving to memory

ah, TU11x my behated

[visual glade]

all GPUs should be able to load fp16 weights as fp32

[vital ermine]

Someone did some bad meds when making 1600 cards

[nocturne trail]

the only exception to "pascal considered harmful" being the P100

which was explicitly built solely for fp16 zoom

and is unfortunately kneecapped by being stuck at 16GB

[vital ermine]

Yep, I almost purchased the P100 in Feb but the fans and blowers from hades I didn't

glad I didn't now with XL

[visual glade]

textures can be stored in fp16 so GPUs can load them efficiently

[nocturne trail]

yeah, you get the memory savings of fp16 but no performance benefit

[visual glade]

it's just a question if pytorch is smart enough

[nocturne trail]

also yeah by default CUDNN on a TU11x will decide that fp16 doesn't work

have to do some shenanigans to make it accept fp16 weights and run them in fp32

[vital ermine]

I had to compile xformers for my 1060 around 5-10 times and it was not an easy task. A1111's wheel finally worked but for a while there I had no choice.

[spice skiff]

Thank you, Neggles. Good suggestion. I've got it working fine now. Just had to deactivate a custom setting in the mouse manufacturer's software.

[nocturne trail]

but realistically what you should do if you gave a GTX 16xx is sigh, throw it out the window, and buy a used 3060

yeah the smooth scrolling stuff is not really worth it most of the time

idk why people keep trying to make mice behave like touchscreens

[vital ermine]

what is so aggravating is that businesses are complaining they can't get the 4080 or 4090 as they ask for 5 and get 0-2 because Jensen has stopped making Ada cards to control supply to keep prices high. At the current supply side there is a year's worth of cards in their warehouses.

I could see the supply dwindling on pcpartspicker

Zotac, no thank you. Gigabyte, I will pass. Not really much left.

MSI ceased to make the air cooled Suprim so damn

[nocturne trail]

shrug another friend of mine has appx. 280 zotac 3090s deployed with waterblocks in machines in a datacenter in sweden

they've been spinning since launch and they've had a total of 3 failures, all from the same batch within the first 3 months, zotac preemptively replaced the entire batch (about 40 cards)

[vital ermine]

EU has far greater consumer protection laws than the USA 100%

[nocturne trail]

shrug YMMV but I mean, ASUS and Gigabyte have always made garbage cards, MSI's 4000 series are... underwhelming

and have mad coil whine issues

[vital ermine]

Yep, to all of that

[nocturne trail]

zotac were slightly cheaper than everyone else and actually put some effort into industrial design so here i am
also didn't fall for the 600W power limit meme which is nice

[vital ermine]

I used to be a Gigabyte fanboi then I had a run in with them and will never touch them again.

[nocturne trail]

after the whole PSU debacle they're completely dead to me

[spice skiff]

Another option is 4090 Founders Edition cards from Best Buy (at least in the US and Canada).

[nocturne trail]

FEs have pretty mid cooling though

[vital ermine]

I was going to do FE but never in stock in either BB or Nvidia so I ceased looking last month

[spark bear]

yep, snagged one a few months ago from bb, even got a 10% coupon in , was like 1420 with tax

[vital ermine]

Zotac GAMING AMP Extreme AIRO vs Zotac GAMING Trinity OC

[nocturne trail]

AIRO is the higher end one but they're pretty much the same

i have the airo, it has a slightly better VRM but not in a particularly meaningful way

price difference was $30 if it had been much more i wouldn't have bothered

[vital ermine]

30 USD difference

[nocturne trail]

yeup

it's very quiet too

no coil whine, fans don't have any harsh tones

runs cooler than my friends' STRIXes

[vital ermine]

I am waiting until after gamercon in a couple of weeks as I may just skip this nonsense and settle back to a good life again with an XTX. Not much caring if it can do SD or not just I need a card and not willing to be a victim to gouging.

[nocturne trail]

why on earth you'd pay a $300 premium for a STRIX card that looks like an emo house brick and performs the same as every other 4090 is beyond me

but people keep doing it

[vital ermine]

yep, LOL

[split magnet]

muh RGB

makes the games faster

[nocturne trail]

it barely even has any

the zotac's RGB is prettier

[split magnet]

people pay for the brand, they like Asus

[vital ermine]

Asus with "issues" scared me off

[nocturne trail]

asus is a cult not a brand change my mind

[vital ermine]

I have their monitor that will soon enough be upgraded after years and years of running. Nice IPS mon

[nocturne trail]

with a panel made by LG and electronics made by Samsung

[vital ermine]

BINGO

you got what I was saying

My previous LCD was a samsung 100%

[nocturne trail]

the only good ASUS products are ones they barely made

[vital ermine]

yeah...no touchy. I SAID no touchy. Now sell it.

[nocturne trail]

and after the absolute insanity they pulled with AM5

[vital ermine]

I will not touch them for that "issue" I mentioned.

[nocturne trail]

"ah yes lets just blindly believe the values encoded in DDR5 SPD and completely fail to enable any overvoltage protection whatsoever in our voltage controllers what could possibly go wrong"

"oh, that"

[vital ermine]

Mushroom cloud

[nocturne trail]

"well here's a beta bios to fix it but if you install it it might not fix the problem and also you won't be covered by warranty if it still explodes"

how about no.

[vital ermine]

yeah, what a load of you know what

[nocturne trail]

it was very vindicating tho

[vital ermine]

I hope someone can come up with an inpaint extension for comfy like auto has then it will be great

[nocturne trail]

i've been calling out ASUS' garbage build quality and approach for over a decade and now there's a Big Shiny Fuck-Up I can point at when people ask why

[boreal coral]

What is wrong with Asus cards? I plan to buy STRIX 40xx so pls tell me

[nocturne trail]

huge waste of money

[boreal coral]

Asus Strix are always top performers in benchmarks so wheres the problem?

[nocturne trail]

performance is identical to every other 4090 on the market

there are no performance differences between Ada cards

they all run right against the redline

[vital ermine]

Oh, I had one of those 2008 ASUS boards that they saved 3c per electrolytic cap cause they went to off brand and years later, with no power they blew up as soon as I plugged them in. All of them vented.

[boreal coral]

So all of these benchmarks online are scams?

[nocturne trail]

the only difference is cooler performance and the STRIX runs hotter than cards which are $300 less

show me a benchmark

all the benchmarks carried out by sites that actually follow a documented, repeatable benchmarking process show performance differences in the 1-2% range which is purely silicon lottery

buying a STRIX card is spending $300 to buy a card that's so huge you can't install anything else while it's in there, you'll struggle to fit it in most cases without dangerously bending the 12VHPWR adapter (so realistically you need a 12VHPWR modular cable or a longer extension adapter cable), and it performs the exact same

[vital ermine]

The amp airo is that a single connector or the new 12V one?

[nocturne trail]

12VHPWR

[vital ermine]

good, my new PSU has that

[nocturne trail]

comes with the good-quality adapter but I've got a modular 12VHPWR cable here just not gotten around to installing it

[vital ermine]

My monitor doesn't have DP only old school DVI-D and HDMI so I need a cheap cable to convert DP to DVI-D

dual link

[boreal coral]

The one with water cooling AIO doesn't look so big 🤔

[nocturne trail]

and is even more money for no performance benefit

[boreal coral]

So theres no difference even in water cooled vs air cooled? 🤔

[nocturne trail]

nope!

[vital ermine]

size and loudness

[nocturne trail]

every single 4090 in existence performs effectively the same as every other one as long as you can keep temps below 85C

they run hard against the redline because they can™️

(and because they're afraid of AMD)

[vital ermine]

Another reason the AIBs are mad at Nvidia

[nocturne trail]

yup

used to be the founders edition was downclocked or power limited or both

[vital ermine]

yep

[nocturne trail]

but jensen wants his (your) money

[vital ermine]

or had bad other issues

FE was the only 4090 that would fit my case. I have a new case/chassis now.

[boreal coral]

What about electronics quality? One of the comparisons mentioned Asus uses higher quality internals as some others

[vital ermine]

some brands, like zotac, do scrimp on how many chokes and vrms are used

[nocturne trail]

uh

no

[vital ermine]

uh yeah, I saw them open them up to show

[nocturne trail]

the AMP Extreme AIRO has almost as many phases and the same voltage controller as the STRIX my dude

[vital ermine]

Gigabyte is bad about it

[nocturne trail]

gigabyte are just as bad as ASUS, if not worse

[vital ermine]

You didn't let me finish as I was about to say if you get one that mentions OC on it get that it will have more

extreme or OC

[nocturne trail]

they drop a few phases on the Trinity cards but the VRM is still capable of delivering over 500W perfectly comfortably

and since these cards lose basically zero performance when capped to 350W, who cares

also the AMP is $30 more and has a 24-phase VRM that could deliver 700W no trouble

[vital ermine]

I had my run in with lack of phases on my old mobo. I prefer to get as many populated as I can

[nocturne trail]

zotac: https://www.techpowerup.com/review/zotac-geforce-rtx-4090-amp-extreme-airo/4.html
strix: https://www.techpowerup.com/review/asus-geforce-rtx-4090-strix-oc/4.html

both 24-phase

[vital ermine]

Love it

see those first picts?

Gigabyte has 6 missing

[nocturne trail]

zotac are using a 55A drMOS driver vs the 70A SPS in the STRIX, but that 70A SPS is over-rated and in both cases it's irrelevant as they can both deliver well north of the max the GPU could ever want

yeah no gigabyte are worse

[vital ermine]

a zotec had 3 or 4 but not AMP

[nocturne trail]

i would buy an asus card before a gigabyte one

they drop to 20 phases on the trinity yeah

[vital ermine]

yep

[nocturne trail]

but the card's power limit is lower as well so that's quite reasonable

[vital ermine]

that is what I am talking about

YET it say O/C which I presume is thanks to having 2 more than Gigabyte

[nocturne trail]

except cutting my 4090 Trinity from its max of 495W down to 350W knocks so little performance off that it's within margin of error

[vital ermine]

how did that affect SD?

[nocturne trail]

immeasurably small difference

[vital ermine]

25% less power 10% less FPS in games

[nocturne trail]

495W is slower than 450W by about 1%

maybe

margin of error

350W is slower than 450W by about 2%

[vital ermine]

I have heard that before and that makes no sense

[nocturne trail]

giving the GPU more power doesn't help when it's limited by memory bandwidth

495W is slower because it raises the temperatures enough to knock about 50mhz off the clocks

[vital ermine]

yeah, I will be bottlenecked as it is so doesn't matter

[boreal coral]

So whats the lowest reasonable wattage we can run them lol

[nocturne trail]

below 300 you start to see significant perf differences

[vital ermine]

Alright I was thinking that the 495 was cooking it to death. Got it right

[nocturne trail]

and the STRIX goes to 600 which is even worse

~350 is the sweet spot

[vital ermine]

you limit that in afterburner or how? In Linux greenwithenvy is gone so I hate that

[nocturne trail]

afterburner or nvidia-smi

nvidia-smi -pl 350

[vital ermine]

yeah, my pascal can't even take a minimum setting as it is not supported smi said

[nocturne trail]

there's a min and max limit

[vital ermine]

not on my 1060

[nocturne trail]

nvidia-smi -q -i 0 and look at the power limit section

my 1060 6GB goes down to 50W

[boreal coral]

600 watts lol. Thats what my whole setup drains on max right now

[vital ermine]

min setting I mean for frequency as 135mhz is stupid and makes my desktop sluggish

[nocturne trail]

eg you can't set an A40 below iirc 100W? because the VRAM pulls 60W all by itself

card idles at 78W the moment you light it up from P8

[vital ermine]

I was hoping to set 2GHZ max freq and 600mhz min but it would take the max but not the min and afterburner doesn't work with min freq on this card either.

[nocturne trail]

there's not a lot of benefit to adjusting frequencies even on pascal

[vital ermine]

yeah, there is

[nocturne trail]

pascal there can be some benefits from overclocking

[vital ermine]

for me and the desktop

[nocturne trail]

but pascal was the last gen where you could really make a meaningful difference

[vital ermine]

I fall down to my entire system taking 65 wattrs at idle

[nocturne trail]

apart from some 2000 series cards

ampere and up it's a wash, half the controls/sliders don't actually do anything

voltage control on Ada is almost entirely gone

[vital ermine]

I agree

yeah, 10% I think it was

[nocturne trail]

not even 10%

it's 100 Arbitrary Jensen Units

[vital ermine]

LOL

[nocturne trail]

and dragging the slider up from 0 to 100 only slows your card down by making it run hotter

[vital ermine]

well, if I set my card to run at max always, and reboot my desktop runs so smoothly

[nocturne trail]

afterburner settings don't persist across reboots

[vital ermine]

card gets too hot though so setting a minimum would help but can't on pascal

[nocturne trail]

you can turn down the PL or thermal limits

PL is most effective

but unlike Ampere/Ada you do actually lose performance for every notch down

[vital ermine]

yeah, but that only lowers my max which is counter productive

[nocturne trail]

yeah

[vital ermine]

I wanted to raise my min

[nocturne trail]

dropping idle power is difficult at best especially given how many programs decide to light up the GPU these days

[vital ermine]

On Turing you can

[nocturne trail]

you can force pascal to sit in the first non-standby power state with a semi-undocumented nvidia-smi command

but it's usually counterproductive

[vital ermine]

would that let it sit at more than 135mhz?

[nocturne trail]

fun fact: loading yahoo.com puts more load on a GPU than running half-life 2 at 1280x1024 did at launch

[vital ermine]

1060 is slow to wakeup

[nocturne trail]

yes but it will bump your idle power up 20W for the privilege

[vital ermine]

fair tradeoff

[nocturne trail]

aite lemme find the command

[vital ermine]

I would presume ADA wakes up faster than Pascal?

[nocturne trail]

yeah

you shouldn't really notice in interactive usage though

this flag is intended for cases where you're spinning up lots of very short-lived bursty gpu processes

[vital ermine]

once it hit 1600mhz in web pages it is snappy but 135mhz it can take 1-2 s

[nocturne trail]

h-uh. absolutely shouldn't take that long

[vital ermine]

as long as I can put it back if it isn't what I need

[nocturne trail]

yeah

if your cpu is half decent turning off hardware accel in browser might be worth doing

[vital ermine]

well, it does and always did

only gripe I had for all these years

[boreal coral]

One more question for the crowd. What do you think is the best value card right now? If you were to upgrade from something that is veery old and should last a few years again?

[nocturne trail]

if you're doing AI stuff? used 3060 12GB or bite bullet and buy 4060Ti 16GB

[vital ermine]

if gaming go AMD

raw rasterization price to performance is theirs

[nocturne trail]

or the ARC A770 which is massively underrated especially for linux
on the other hand alchemist is about to get refreshed

with the memory controller bug fixed

[vital ermine]

I am still holding out hope for Druid and Celestial now that the ceo is canned.

his stink is no where on the cesestial.

*celestial

[nocturne trail]

wha?

ARC is Gelsinger's pet project he's killed entire business divisions at intel to keep it alive

[vital ermine]

The one Intel fired

[nocturne trail]

that was years ago tho

Battlemage looks like it's going to be quite competitive at the mid-upper end and if they can keep it up Celestial will be right up there at the top with the best of 'em

[vital ermine]

the one who helped make the turd Vega series

[nocturne trail]

they didn't fire him lol
nor did he make Vega

Raja is responsible for RDNA and RDNA2

he's done the same thing at Intel as he did at AMD - help them make the first gen, get them properly going on the second gen, then get bored and go elsewhere

[vital ermine]

They fired him and it is well known they did. His stench is gone now and good.

[nocturne trail]

the only source on that is MLID who is questionable at best

[vital ermine]

Raja Koduri, the former head of the Intel's AXG unit is leaving Intel by the end of March. As part of the company's reorganization. Corporate speak for they let his ass go which is fired. No one liked him and none cried when mentioned he was leaving.

[nocturne trail]

and he literally quit because he spent 3 months sick in bed and in and out of hospital + his work was done & they're folding AXG in with the habana group over the next few years anyway

shrug he's literally the reason that AMD is competitive today so

[vital ermine]

No

Anyway enough with Raja I am sure he can find some other venture to stink up.

[nocturne trail]

[atomic ice]

QUESTION!!!
I HAVE A QUESTION!!!!
I have a question*
How much Vram do I need in order to train SDXL?
does it already have dreambooth support?

[nocturne trail]

12GB and yes

[atomic ice]

WAIT!?!?!?!?!?

WGHAT!?!?!?!?

SO I CAN WITH A 3080TI!?!??!?!?!!?!?!?

3080 TI?!?!?!!?!??!!?!?!?!?!?

[nocturne trail]

iirc those came in 10GB and 12GB but either way yep

[atomic ice]

I HAVE 12!!

[nocturne trail]

then yes

[atomic ice]

I was planning to upgrade for thissss

so I don't need?!?!?1

[nocturne trail]

use this https://github.com/derrian-distro/LoRA_Easy_Training_Scripts
use bfloat16 mixed precision and don't train the text encoder, 1024x1024 train res batch 1 will work

[atomic ice]

YYYYYYYYYYYYYYYYYYYYYEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEESSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSS

[heady vale]

why not just try it first?

[vital ermine]

training a DB for XL can a 4090 do it yet? Last I saw was just unet was taking 23GB at BS1.

[nocturne trail]

lmao

[atomic ice]

I had 0 hopes

[nocturne trail]

in what world

[vital ermine]

one trainer

dev was saying that

[nocturne trail]

4090 you need to either not use xformers and use SDPA, or compile your own xformers

[atomic ice]

waaaaaaaaaiiiiiiiiiiiiiiiiiiiiiiiiiiiit

it's LORA!!!!!!!!!

[vital ermine]

dreambooth works then to fully train G/L and the TE as well as Unet?

[nocturne trail]

you do not want to train the TE

[atomic ice]

why not?

[nocturne trail]

the SDXL TEs are already very well trained and attempting to train them almost universally results in Worse Output

they're very, very picky

[atomic ice]

oh

makes me want to train them less

[nocturne trail]

you don't need to train them

[atomic ice]

I'll keep making 1.5 models

is it stupid?

[nocturne trail]

just training the unet works exceptionally well

[atomic ice]

wait

lora

is not the unet

is like an extra thing, no?

[vital ermine]

how much ram is needed to train a dreambooth model for XL then and at what BS?

[nocturne trail]

dreambooth is just a word for doing lora training basically
technically you can apply the same approach to a full finetune but

[nocturne trail]

(the prebuilt xformers wheels are built for 5.0+PTX 6.0 6.1 7.0 7.5 8.0 8.6 which doesn't include Ada (8.9) so you're better off using torch SDP attention)

[vital ermine]

The dreambooth I am talking about came long before loras

[atomic ice]

when I dreamboothed 1.5 I didn't Lora it

[nocturne trail]

yeah but it's a bit silly

full finetune works fine as well resource requirements are about the same

[vital ermine]

making a full new model is never silly

[atomic ice]

so I can't?

[nocturne trail]

you can

[visual glade]

I thought the reason why they didn't build it for 8.9 is because it didn't change anything

[atomic ice]

I can only lora

you said

[nocturne trail]

it sure as hell does on my system

mostly because see that 5.0+PTX

[visual glade]

for the forward pass?

[atomic ice]

I'm the creator of OREN-4 by the way

[visual glade]

or the backwards?

[atomic ice]

the world's best model

[nocturne trail]

if it was 8.6+PTX it'd be minimal

[nocturne trail]

i lose a solid 12% of it/s

on this card xformers prebuilt public wheel is significantly slower than SDPA but if i use a container built on top of nvidia's NGC container and compile xformers for 7.0 7.5 8.0 8.6 8.9 9.0+PTX it's significantly zoomier

the problem is 5.0+PTX which limits the PTX intermediate code to friggin' Maxwell functions

[atomic ice]

PTX 3080 TI

[nocturne trail]

this is a silly question and im 95% sure the answer is "no", but, there's no backward in inferencing yea?

[visual glade]

backwards is training

[nocturne trail]

thought so

I haven't compared training perf on the differently-compiled xformers wheels, i just recklessly abuse github actions to build it for me (it does take almost 3 hours to compile on an actions runner)

works out smaller too, at the expense of maxwell and pascal support which (personally) i consider a waste of time but ymmv

[atomic ice]

If you guys wanted to train something

something good and big and important

something like OREN-4, the world's best model

would you

a. Train SDXL LoRA?

b. Train 1.5 with everything

[azure oxide]

c. both

[atomic ice]

Which one first assuming you only have 1 GPU?

[visual glade]

yeah I built xformers on github actions for comfyui once and it took a very long time

[azure oxide]

order wouldnt matter to me unless you have a deadline

i'd do the one you're more interested in

[atomic ice]

Which one am I more interesed in?

[azure oxide]

if you have 0 preference then flip a coin

[atomic ice]

I think that I hate LoRA and love making SD 1.5 models, but maybe LoRA is good now and 1.5 is ancient

does LoRA training even make sense for a giant dataset???

[alpine pine]

Lora was great on 1.5 ...

[atomic ice]

like 3000 images with captions

[atomic ice]

not really, it couldn't do what I needed

it could put a style or a something in the image, but not make a different thing

like how I made OREN-4

[candid walrus]

i often too would get much better results with a full dreambooth model on 1.5 training, but i didn't dump tons of time into it once i had something that worked. But as far as 1.5 even the diffuser docs and scripts only mention lora's as far as dreambooth goes... https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_sdxl.md

[atomic ice]

I made OREN-4 think differently

and do everything better than any model in the world

[alpine pine]

LORA can absolutely make something that isnt in the base data set.

It isnt a checkpoint, but it does its job excelently

[atomic ice]

yeah, a thing, or a couple of things, but not like a different thing in terms of different things

like OREN-4

It couldn't do something like OREN-4

[split magnet]

you've really been out here repeating oren-4 being the best model for half a year

[vital ermine]

Well, it is technically LoRA using dreambooth method. Dreambooth is furthwer up the food chain than a lora.

[split magnet]

nobody thinks that, just get over it

[atomic ice]

because it is

I asked people objectivly and they said it is

people who hated me said it too

want a proof?

[split magnet]

like the people on your reddit post

[alpine pine]

bet, let me give you a prompt

[atomic ice]

they didn't see the comparison and weren't objective

[tribal jackal]

Maybe I should dial the noise down on the upscaler lol

[atomic ice]

give

[alpine pine]

Positive: An anime drawing of an astonaut riding a dinosaur in space

[atomic ice]

no

It's a photography model

[alpine pine]

Positive: A woman wearing a red shirt, blue pants and green shoes with black hair
NO regional prompting

[candid walrus]

yes i understand this, but i'm just saying they don't give an example in the docs, so unless you really know machine learning or find someone that might, they don't give an example of it

[split magnet]

a photography model can do good anime art as well

[vital ermine]

Yeah, I miss Dreambooth from Shiv.

[atomic ice]

not oren 4

[tribal jackal]

High-altitude photograph of (nido dog) on a mountain climbing adventure, rugged peaks, expansive views, gear detail, dynamic weather, photography inspired by Jimmy Chin and Renan Ozturk

Negative: anime, deformed, glitch, blurry, noisy, off-center, cross-eyed, closed eyes, bad anatomy, ugly, disfigured, mutated

[split magnet]

because it's not the best

it's limited

[tribal jackal]

replace (nido dog) with (springer spaniel)

[atomic ice]

It's the best for photos

[alpine pine]

do the second prompt I gave

[high patio]

I'm using A1111. For sdxl should I be using the base model and generating 20 steps then sending the generated image to img2img and then running the refiner model for 30 more steps? What is the typical workflow? Thanks

[atomic ice]

I'll do

[tribal jackal]

too much bloom

[atomic ice]

"she is the hunter of pets"

[nocturne dove]

just because

[split magnet]

nice image with too much bloom, oversaturated, too many arms, weird hands and feet, too smooth and not much detail

[azure oxide]

during the livestream today, a dev was suggesting 32 base 8 refiner (total of 40 steps) but there's no "best" if thatll ever be a thing

[atomic ice]

but Oren-4 automatically make any IDF soldeir with almost correct uniforms and always as a female

[alpine pine]

I want my woman wearing a red shirt, blue pants and green shoes with black hair already

[split magnet]

that's a very niche application

[high patio]

Thanks I will try it. Did he mention anything about the denoising strength?

[azure oxide]

not sure on that one sorry

[atomic ice]

loading SD now

[alpine pine]

ok, and remember:

• no after detailer
• no regional prompting
• screenshot the A111/comfy ui and prove its 3 randomly generated seeds aswell

[nocturne dove]

SD 1.5 models aren't going to be able to get the colors right and asking it to will probably make it mess up the face too.

[alpine pine]

thats the plan

SDXL can

[atomic ice]

no after detailer is upscaler?

[alpine pine]

yes, raw output

The best model can handle regional color prompts and generate a good image every time with any seed

I expect both

[atomic ice]

My model isn't good without the latent thing, it's the best because it overcokes the image in a cool way

[alpine pine]

then your model is not the best in the world

SDXL is

[atomic ice]

NO!! OREN-4 IS THE BEST MODEL

[alpine pine]

colors are all wrong

[atomic ice]

your prompt

because Clip is a moron

[atomic ice]

Check out that it made all of the women hot

This only my model can do

[nocturne dove]

Even if the colors were right, the images were terrible...

[atomic ice]

No

let me upsace and you'll see

[alpine pine]

not my fault, SDXL has better clip, thus is better model

[nocturne dove]

You're going to use something else to replace the faces aren't you?

[atomic ice]

NOOOOOOOOOOOOOOOOOOO

NO!!!

NO

I'M NOT

just latent upscale

[nocturne dove]

lol

[atomic ice]

no restore faces

[split magnet]

could always check the png data

[atomic ice]

Seee????

[nocturne dove]

I think there is one person in the world that thinks your model is the best.

[split magnet]

where are her pupils

[atomic ice]

IT IS THE BEST

She doesn't have them

notice that the pants look almost like IDF soldier pants

[alpine pine]

A. She is not hot
B. It’s clear your trolling

[split magnet]

and they're green

the pants were supposed to be blue

[atomic ice]

because it's trained on IDF soldiers

Here!!

this one came out better

[split magnet]

since when do IDF soldiers wear bright red shirts

[atomic ice]

Notice the boobs

[alpine pine]

Shoes are supposed to be green

[split magnet]

looks like a mannequin

[atomic ice]

They don't but it listens to the prompt and not just IDF soldiers

it's a huge datasets with captions

the captions talk about images like they are normal

but the images are insane

it's the best dataset

[split magnet]

so what you're saying is that the pants could be blue if the model was better

[atomic ice]

NOOO

[alpine pine]

But, purely a question, is it against the rules to advertise a paid model?

[atomic ice]

It gives you a better output than your prompt

green pants are cooler

Feel free to give me more prompts

[nocturne dove]

Wait... he charges money for this? 😆

[atomic ice]

you'll see

[atomic ice]

I made money on this once

[alpine pine]

According to Reddit, yes

Maybe it’s changed now to be fair

[atomic ice]

https://www.patreon.com/orenong

[nocturne dove]

Someday you will make 1 more money and then you can say you made money on it twice.

Probably not today

[atomic ice]

it's in my patreon

give me prompts

[alpine pine]

Anime drawing of an old man

[atomic ice]

It's a photographicccc modelllllll

[split magnet]

do Pippi Longstocking

[alpine pine]

A photo of a woman drawing an anime drawing of an old man

[split magnet]

epiCRealism can do great anime outputs

[alpine pine]

Epic realism is just the best on 1.5 tho, that model is so good

[split magnet]

a girl in the style of studio ghibli

[split magnet]

true, my favourite model

hopefully they'll train it on SDXL soon

[alpine pine]

I hope so aswell. All my mediocre Lora’s were 10x better by just using their model

[split magnet]

Pippi Longstocking in the style of studio ghibli, generated in epiCRealism, a ''photographic model''

[atomic ice]

somehow it works

better than yours

notice the IDF pants color

[split magnet]

see, now you have no excuse for calling it a photographic model 🙂

[alpine pine]

Is this regional or a single prompt?

[atomic ice]

IDK how it did it

WOW

[atomic ice]

MINE IS BETTER!!!!!!!!

[visual glade]

well first pass SDXL + second pass WD1.5, each have their own prompts

[split magnet]

that's not studio ghibli style and it lacks some detail in the brackground

[nocturne dove]

WWhile You Were Partying, I Studied the Braid.

[atomic ice]

Is that a prompt?

[nocturne dove]

No, she just has two braids in her hands like swords.

[atomic ice]

There was a lot of hair in the dataset

[atomic ice]

Used it as a prompt

[nocturne dove]

Oh no

[atomic ice]

You guys must admit that if not the best

it's the most fun model

but it's the best

[alpine pine]

I’ll give you that I’ve had some fun in the last hour

Quite enjoyable

[split magnet]

oh so you are promoting your model

[atomic ice]

no

[split magnet]

which is against the rules

[atomic ice]

oops

[alpine pine]

They have been the whole time

[atomic ice]

deleted

sorry

[split magnet]

I know, I just wanted him to say it

[alpine pine]

Ahhh

[nocturne dove]

Also it's against all logic, nevermind the rules.

[atomic ice]

No I'm just saying that it's the best

nothing non-objective

[alpine pine]

To be fair, you did link your patreon

[atomic ice]

I got to go to make OREN-4

OREN-5***

BYE

[split magnet]

non-objective

[heady vale]

messing about with upscaling. 9.5k square image in 150 seconds on a 3080. now I need to work out how to get more details

[alpine pine]

More details? That image is pure fire

[heady vale]

at 1:1 the faces arent corect

[split magnet]

are they drinking orange juice with their burgers

[heady vale]

XL seems to like juice a lot when you mention just "drinks"

[alpine pine]

Are alcoholic beverages not trained on?

[heady vale]

ive seen shot glasses, so I guess so

[split magnet]

nudity and juice, an AI's guide to the perfect breakfast

[vital ermine]

I have yet to find a workflow I like. Something simple where G/L goes to the sampler and I can see an image then another image is made post refiner. Seems that first image is a horrible mess with everthing I have tried. I thought I could just tap in to the link from sampler base to sampler refiner but nope.

[deft coral]

iu have a dataset of 40 images

im sturggling with overtraining

idk what epoch to use

[split magnet]

well that's mildly disturbing

[nocturne dove]

I thought that was we were doing now. People heads on robots.

[gritty breach]

made this while tipsy, damn idk if ill ever get the ai to recreate this actors likeness so well again

[split magnet]

idk I've been doing this one

[nocturne dove]

Human ears on animals. That's a good one!

[split magnet]

not my pic to be fair, I just ''remastered'' it

it was quite low res

[nocturne dove]

I think human teeth on animals is also very weird. Like smiling and stuff.

[split magnet]

https://tenor.com/view/meow-cat-big-mouth-filter-gif-14392492

[nocturne dove]

Yes like that!

[split magnet]

ah yes, 8192x8192 because why not

[boreal bough]

anyone know the specifics on how caption dropout works?
haven't used it till now

so if i do caption dropout every n epochs = 1
caption dropout = 0.5
A.) so that means 50% of comma separated tags are dropped?
B.) on epoch 2, are 50% of all captions going to be dropped, or 50% of all remaining captions? (essentially until not single token remains)
C.) is keep n tokens immune from this setting?

[split magnet]

don't zoom in too close though, because then you will see the imperfections

[alpine pine]

I’d highly recommend setting up a custom node for your inputs. Have that node do the math for you and then connect it to ksamplers and an SDulti upscale node. I did this today and it simplified my workflow massively.

[chrome flicker]

Do you have any idea to use a input image as reference in sdxl like the same way in midjourney?

[alpine pine]

How does it work in Midjourney?

[chrome flicker]

Midjourney will generate creativate results but these results are very related from the input image, like style, object.

[boreal bough]

oh god never mind. found kohya's officialy docu and translated it

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.

I was very wrong XD

[alpine pine]

One way would be to interrogate with Vit-H using the “best” preset and use it as a caption. This gives an image that feels like the first image

Last that you may need control nets when they release

[vital ermine]

I don't actually know how and spinning my wheels getting nowhere only frustrates me. No upscaling for now anyway just something simple.

[chrome flicker]

okay, that canbe a good assumption。

[alpine pine]

Have you tried Comfy’s basic example?

That’s just inputs and ksamplers

[vital ermine]

yes

tbh, this comfyui is not for me only I am forced into it because auto's is DOA for XL, and limping for much else now.

I like to dissect things only in this I can't do that. I can't click a wire and see where it goes and most times it just a bag of frustration.

[heady vale]

on the input of a node left click and hold, the connection is highlighted

[vital ermine]

in some of the complicated ones that is highly inferior and makes one easily lost, at least for me.

dragging a box to move it out of the way of the wires only makes it worse.

Oh, and another thing I miss from other node based systems I have used is click the wire and press delete and it goes poof, and ctrl-z to bring it back.

[autumn forum]

now thats some fine tuning if your usin that lol

[split magnet]

have you thought about trying SD.Next?

[boreal bough]

literally steamrolling the sdxl model XD

[split magnet]

It's a branch of A1111

[vital ermine]

Yes

what are these?

[heady vale]

to reroute connections. some use them to cleanup the spagett but you can have multiple connections off them

[shy kelp]

https://tenor.com/view/racist-thats-racist-dats-racist-gif-5036293

[broken hare]

i think i have one.....i will find it

[boreal bough]

find the worlds smallest doggo in the background! XD

[vital ermine]

This has a bunch of them but didn't help, lol

[tribal jackal]

lol yeah I accidentally stumbled upon a way to make Where's Waldo, and the tricky part has actually been eliminating them from the output.

[spark bear]

the joys of tiled upscalers

they still give you best quality tho just need cleanup sometimes

[tribal jackal]

Oh is that the cause?

Hmm, I might just try one big tile then

[broken hare]

most simple with and without refiner workflow i have

[oak ember]

Hi everyone, so I just download the XL model along with the refiner, put it in the model/stablediffusion folder, but when I'm trying to choose the model, it processes it, but then it doesn't change the model at all but stay in the previews model I was using. Any idea what's happening or how to fix it?

[tribal jackal]

winner winner chicken dinner

[upbeat summit]

Hey Nido is back 🙂

[vital ermine]

Thank you. /fingers crossed

[tribal jackal]

just finished testing a combination of upscalers to stick on the tail end of that giant workflow I screenshot earlier

A few more prompts couldn't hurt...

[vital ermine]

That looks remarkably similar to the one I made but the image before refiner (the base image) was always funky

[broken hare]

i wont say its great, it is the very first one i used when .9 was out....i just modified the look of it. I dont use it anymore, but it works

i seem to get better results from another one i can send you if you want.....but it doesnt show the image before and after. I feel like it just saves a little time

it also has great notes that describe what is going on

[urban fjord]

Thanks, I thought the same as you that it dropped a certain portion of the captions at certain epochs.

Though now I want to be able to train captions without images so I can reduce the size of my datasets.

[vital ermine]

yeah, I was using the before and after for testing and learning

I am wanting to see what various prompts do for refiner

[broken hare]

ah....then that one may be the one to use.....and it is simple to modify

[vital ermine]

send me the other, if you would

[broken hare]

i like this one because there is a nice way to tell it total steps, and where you want the refiner to start

[fleet harness]

[tribal jackal]

2704x1544 in 98s

[broken hare]

@vital ermine I am working on getting one put together that doesnt use refiner at all because the brand new finetunes do not need it, and im going to add image to image. all very simple.

[upbeat summit]

pretty good!

[broken hare]

im learning how to do that as we speak

[hardy cipher]

I think I might have figured out a little workflow trick to improve image output quality. or maybe someone else has done it before. but I haven't esen it

[half ivy]

Guis guis guis
Is this it?

https://huggingface.co/SargeZT

[hardy cipher]

very nice

[split magnet]

controlnet on SDXL, yes please

[hardy cipher]

but anyways, I'd have to test it out more to actually confirm it improves things, but thus far every instance has been positive

[half ivy]

Not finished or not tested or summin like that
Just wondering if anyone has tried yet

[vital ermine]

I miss a viable inpaint with a brush like the one in Auto1111

[hardy cipher]

it could be cleaned up quite a lot, but it seems to work

[broken hare]

me too. I'm using auto about half the time, but the results just do not look quite as good to me, and its just slower for me

[vital ermine]

huge mem leak now too as evry single time I loaded refiner my page file grew until my 48gigs of ram was using a 128 gig pagefile

[hardy cipher]

I mean, that's a difference right?

[broken hare]

i only use base in auto....the extension to add refiner takes a shit ton of memory

[hardy cipher]

exact same prompt and settings

[upbeat summit]

the second one with your workflow? yeah, that coherence is a lot better

[vital ermine]

I went to img2img and loaded it there

[broken hare]

that is what i prefer with finetunes. thats what i do

[vital ermine]

besides I just updated auto from feb release and the new one is so sluggish but I had to havce it for sdxl

ALMOST

[hardy cipher]

took the first prompt and basically combined the negative leaving the L blank. then inverted it and sent it through the sampler upside down. so positive prompt in negative. then interrogated that and put it in the negative L prompt of the final image

[broken hare]

that image to image in auto is nice....i like to change resolutions there. just not as easy to do all in one place in comfy,

[hardy cipher]

basically had the ai make the opposite of what I want

[fleet harness]

When you sign up for a space gig but end up on a boat behind a turret

[hardy cipher]

and then tell me what tha tlooks like

and use it as a negative

and you can still put your own negative in neg G

[tribal jackal]

MFW when I realize I can copy paste nodes across multiple tabs

https://tenor.com/view/dance-lil-dicky-gif-21053087

[upbeat summit]

https://tenor.com/view/huh-gif-23918002

[hardy cipher]

lol

so just put the positive in the negative and negative in positive, create an image

[upbeat summit]

I get it - sounds fascinating. but I don't know the implications 😄

either insanity or genius! looks like the latter in your example 😄

[hardy cipher]

I would do it with a1111 and not once did it make an image worse

every single time it was better

but it was a hassle as well

basically you're getting what the ai thinks is the opposite of your prompt and putting that in the negative. and I have it set up so it's unique for each seed number and all that

it's not super exact, but doesn't need to be

[upbeat summit]

yeah it sounds super interesting. please share more of your findings and examples if you want 😄

[hardy cipher]

I just like what the opposite images are, lol. they're not what you'd think at all

[fleet harness]

https://tenor.com/view/prank-switcharoo-different-person-gif-12926194

the ol switcharoo eh?

it's fun doing weird stuff like that.

you should also try emojis, leaving things blank, using other languages, and I'm sure there are more ideas

[upbeat summit]

you can also leave the prompt field blank

[midnight shuttle]

Anything will tokenize and convert to weights, even complete gibberish. With no prompt I think you are starting with fixed weights and then the random seed is fully responsible for the image.

[fleet harness]

Have your cat walk across the keyboard. The options are limitless.

[hardy cipher]

I've done that

well didn't really want her to

but cats cat, so she did anyway

[fleet harness]

I bet it was glorious right?

[hardy cipher]

it did make a thing

[fleet harness]

lol

[upbeat summit]

that's the beauty of it 😉

[hardy cipher]

alright, another prompt incoming

let's see if it's consistent

[fleet harness]

what are they talking about? >.>

[heady vale]

"its too small"

[hardy cipher]

ugh, comfy acting rude

the ui, not the person

[upbeat summit]

I'm not saying this has anything to do with it - my workflow with SD2 was using the hires fix to enhance fidelity and details in general.

are you giving your last step kind of a "low detail" image as guidance?

[hardy cipher]

here's the negative image

[upbeat summit]

makes sense

[hardy cipher]

they're identical settings other than the negative L

and I'm using some junk booro interrogator

[upbeat summit]

so I made some guidance / reference image tests with SDXL. I think I noticed increased fidelity with some things... I know you are doing something different but you are also feeding it an image that it uses as guidance

or anti-guidance?!

[hardy cipher]

well sort of, just using the description of the negative image, although there might be a better way to do it

[upbeat summit]

so you are reinforcing what it should not do?

like a dynamic negative prompt?

[hardy cipher]

essentially. just sending the output of the interrogator to the negative L input

since its just a list of words

and then you can still put in your normal negative to be sent to negative G

[upbeat summit]

yeah, that sounds really interesting to explore. you are already showing some very promising results.

I was trying to do aerial urban photography earlier and it is very hard to get the details right - so I can tell you the details in that image are really good

[hardy cipher]

what's cool with this setup is it's dynamic, so it'll change with each seed

[hardy cipher]

nice

so this method obviously takes longer, but I'm thinking I wouldn't really need to run the negative image for that many steps. just enough to make it coherent

I also wonder if having it overlay on the initial output image would be preferable

[tribal jackal]

Alright! Got the upscaling all implemented. No missing wire connections. Time to go watch an episode of a show lol

[upbeat summit]

is there any facedetailer or restoration on it?

or isn't that image part of your research?

[hardy cipher]

lol, nah.made that earlier

[upbeat summit]

lots to explore in this one 😉

[hardy cipher]

so I can say that this workflow doesn't necessarily always make the pictures "better." but they are definitely much more well defined and more true to the prompt

[hardy cipher]

told it to make a lego metropolis and it went full lego

it's like it fills it out

[hardy cipher]

some strange legos there

[upbeat summit]

looks like it emphasized your prompt

[hardy cipher]

yes. the optimal approach might be to have the negative L from the interrogator have variable weight

so less impactful to start but then increase influence through the steps

[hardy cipher]

lots of unknowns and I'm just starting to learn about a lot of these tools. but I've had ideas bouncing around in my head for a while that I had no idea how to implement

[upbeat summit]

yeah same here. that's how you discover things 🙂

[vital ermine]

heh

[upbeat summit]

General Awareness has arrived

[vital ermine]

I am playing around with perlin noise

no idea what all these do

[hardy cipher]

well with a1111 it was a bit more daunting to consider coming up with anything because I'm not much of a programmer. but the nodes are a nice happy medium for me. haven't even had comfy 2 weeks, but learned a bit in that time

[upbeat summit]

Yeah you can build custom workflows without writing code. I enjoy it a lot as well 🙂

[trim orbit]

So all this hubbub about only using safe tensors is that attackers can't just stuff any ol data into the file format , right?

[hardy cipher]

I wouldn't mind getting back into coding, or just taking it more seriously. took some classes in school, but never really did much with it. but I've been messing with python more lately so who knows

[trim orbit]

yet safetensors still have metadata in their format, and any dangerous data could also be stuffed into the metadata, so why are safetensors a solution?

this attack would likely occur through image metadata anyways, like it always has in the past since it's not a new attack.

[hardy cipher]

what sort of things are you doing with the noise?

[hardy cipher]

I don't know, bud. I don't make the rules

[vital ermine]

those images were the same prompt and same seed only the noise was different

[hardy cipher]

how are you modifying it?

changing the offset?

well that wouldn't do that much actually

I'm just curious

[hardy cipher]

that's cool. didn't even know about that

I need to check that out now.

[dense chasm]

i've successfully rola training of realistic character with sdxl 1.0,total steps 3500,10 epochs,the base model with fine tuning is not bad,this is epoch 3,looks like michelle yeoh already

[hardy cipher]

ooh, there's a perlin power fractal? wat?

[vital ermine]

yep

[hardy cipher]

I don't know what that means but I like it

[vital ermine]

perlin is a good one

[autumn forum]

3500 steps?? tjats a lot

[vital ermine]

used that in cinema4d years ago

[hardy cipher]

something new to learn about. I've never even really heard of the concept before. I know what noise is, but not all these other things

[trim orbit]

if i'm not mistaken, the only thing the seed changes is the initial noise that's created. so changing the noise algorithm would be the same as changing the seed

[vital ermine]

still having issues with janky faces and no i2i to fix them

[hardy cipher]

I think this workflow I'm messing with would help with face clarity. probably not 100 percent of the time, but mostly

[dense chasm]

yes,sdxl 1.0 base model you can set it to 9999 total steps as well,will not overcooked

[upbeat summit]

actually this package has a very advanced face detection method https://github.com/ltdrdata/ComfyUI-Impact-Pack

it's huge, but my first results are looking promising. It has lots of settings... but it's the most advanced face fixer I've worked with yet.

I started to look into it since I have so many images that are interesting but the face makes them hard to pick

[trim orbit]

#✨｜sdxl message sytan's workflow has a 2x i2i fix to see how this works. i've also added the facedetailer nodes to it which gives that specific face pass

[hardy cipher]

that face detailer just seems like a lot, lol

[autumn forum]

i feel like that is just a long time for a character model. i can successfully bake a character in in about 300 steps probably with the right settings. how long did it take?

[trim orbit]

thats what i've put into what i've linked

[midnight shuttle]

As I understand it, .ckpt can contain code but .safetensors can only contain non-executable metadata. So it would be much harder to get that metadata to execute, as compared with code that is designed to execute.

[trim orbit]

it would be non executing data in either format. the attack is another process uses that extra data in it's execution

[vital ermine]

I haven't used a face fixer since 1.5 days so since 2.0 hit. I just i2i and done

[dense chasm]

takes around 2-3 hours,i just tested for different parameters and choose an appropriate config settings file,what's you dataset image number and batch size,repeats etc?

[trim orbit]

facedetailer is an i2i pass specifically masked onto the face

[midnight shuttle]

Probably need to put together a POC and send it to the developers so they can make .safertensors.

[hardy cipher]

I've looked at it

[upbeat summit]

yeah I don't like the look of codeformer or gpfgan at all. I never use it. but this works differently since you can use any model for face restoration. it doesn't produce generic faces like codeformer etc. but still testing.

[upbeat summit]

you will form a warp bubble soon

[hardy cipher]

bro, are those all node paths? that thing is going to explode, lol

[tribal jackal]

lol yeah

[trim orbit]

law of diminishing returns is the thing though, or in this case, return 0. a new file format doesn't prevent extra data, never will. a newer one couldn't do it either.

[hardy cipher]

and your computer hasn't fallen over dead?

[upbeat summit]

what's your fps :)?

[tribal jackal]

jobs estimated to take about 30-35 min

[tribal jackal]

Shhhh

[tribal jackal]

lol i actually get a stable full screen video player on streaming services even with this 5120x1440 monitor

[autumn forum]

36 images, 7 epochs, 7 repeats, tagged of course, batch size 3, learn rate 0.002, adamw optimizer. 588 steps. learned the character very well. but take my knowledge with a grain of salt cuz i dont know much about loras i just feel like thats a lot of steps ya know?

[tribal jackal]

but if I move my mouse around and it has to load the interface things get sketchy

without the video player everything is fine

except like dragging and dropping nodes/wires

during upscaling

[hardy cipher]

are you running an a100?

[tribal jackal]

4080 local

[hardy cipher]

I was reading about the h100 earlier. those seem pretty legit

[upbeat summit]

yeah but what's your ComfyUI FPS (fps are displayed in the bottom left corner)? 😄 Chrome based browsers give you definitely a better framerate compared to Firefox.

[tribal jackal]

oh one sec

[upbeat summit]

I mean if it doesn't feel sluggish - that's good

but mine started to lag a bit - not as fluid

[dense chasm]

number repeats should be at least 50,then it will show up more details,when batch size increase,txt2img results in higher weights becoming less saturated,but it take more training epochs to learn the target

[autumn forum]

mines 120 (monitor framerate) 😛

[tribal jackal]

its hovering around 120

[soft zealot]

which is great until someone renames the Node to suit their workflow and then someone downloads that image , tries running and asks "where is this node from" ;o)

[autumn forum]

hmm interesting, ill try that out next time i train! but then again sdxl is wayy different at training then 1.5 and maybe the repeats arnt as needed?

[upbeat summit]

ok so my node setup does kill the fps - and it's not even a lot of stuff compared to others. time to investigate 🙂 thx

[tribal jackal]

good luck!

[autumn forum]

kittehhhh

[hardy cipher]

I question whether I'll ever really understand how all these different samplers differ

[upbeat summit]

thanks. other workflows do run with 120fps... so lets delete some nodes and see who's the culprit

[tribal jackal]

The only way I'll ever "know" is if I build a workflow to let me test quickly every time I forget.

[tribal jackal]

It's probably not SD Ultimate Upscale. That's the only non-default node I think I'm using

[upbeat summit]

well with x/y plotting like in a1111 you could do a lot of parameter testing heh quickly

[hardy cipher]

I'm going to see how much dpmpp_2m_sde and dpmpp_2m_sde_gpu differ

[tribal jackal]

I think comfy can do that as well.

[hardy cipher]

these names really roll of the tongue

like poetry

[upbeat summit]

it can - but compare how it needs to be set up. not the same. but it works of course

[tribal jackal]

I may be partial to blueprint scripting lol

Just a little bit

[hardy cipher]

I haven't loaded a1111 for a while. really not reasonable to run them both on my little barely sufficient video card, lol

[upbeat summit]

I'm Commander Shepard, and dpmpp_sde_gpu karras is my favorite sampler in ComfyUI

[hardy cipher]

I really thought it was decent until I found stable diffusion

[tribal jackal]

omg no not that "curator"

[hardy cipher]

I just pick whichever name is the longest

because that's usually the best once

whoa, dpmpp_2m_sde_gpu kind of deepfrying

[royal fern]

f@ck ddim, too short

[upbeat summit]

well you need to adjust your settings of course

[hardy cipher]

yes, there have to be at leats 5 random sequences of numbers and letters for me to actually want to use it

[royal fern]

so far haven't seen any advantage in the _gpu samplers though

[hardy cipher]

well it made it more intricate. I still like euler_a tbh

[royal fern]

I rarely use non converging samplers

[hardy cipher]

heun is good for things like landscapes, but it's not all that exciting. I don't know how most of them even perform

[royal fern]

even though they seem to work better on SDXL

[upbeat summit]

I just like the noise more - probably placebo and voodoo thinking but I've been using it forever in a1111 etc

[hardy cipher]

so are the converging samplers the ones that just get more clarity over time but don't more like euler a?

[royal fern]

_SDE and _A are the weird guys, yes... where each iteration is different from the previous

[hardy cipher]

it's not an area I've focused on much thus far

[tribal jackal]

16 images at 2704x1544

[hardy cipher]

goodness. it better be a banger

[tribal jackal]

It's intended to be a production ready workflow for prompts ive tested in a smaller environment

[royal fern]

when doing img2img I find "constant" samplers are better

[tribal jackal]

But this was a trial run to ensure it all went smoothly

[hardy cipher]

what are you doing exactly?

[hardy cipher]

you made those?

[tribal jackal]

I like to automate things

yeah

[hardy cipher]

they're definitely very clear and realistic looking

[tribal jackal]

Goal!

[hardy cipher]

but why so many node paths though? I still don't understand

[upbeat summit]

because of the last few days and you posting images, I'm thinking I've subscribed to Nido's travel blog 😉 very nice

[hardy cipher]

good lord. that'd take me a couple days

I need to get comfy working on colab again so I can do more crazy things with sdxl

nice. well I like really intricate things so it's right up my alley. just need to upgrade my vram situation at some point. not sure what the most cost effective method would be though

might actually be cheaper to just use colab honestly

but it kind of sucks

[upbeat summit]

are you doing anything to the faces or is fixed through your sampler setup?

[tribal jackal]

I have to try not to be salty about the efficacy of simply renting a google colab server vs owning local hardware but I am ultimately glad people without the hardware have options.

[fleet harness]

Renting A100s aint a bad life

[tribal jackal]

Can't play Star Citizen on google colab though lol

or can you?

[fleet harness]

truth

cant wait for that btw

[tribal jackal]

Nothing manual, all through the setup.

I am anxiously awaiting controlnet for sdxl workflow though

[royal fern]

sometimes I feel people just convince themself that they don't like or need what they don't understand, because everything has to be simple and god forbid educating themselves

[tribal jackal]

The same people who will downvote you for not directly linking to a copy-pastable workflow that robs them of the educational experience they think they don't want

[royal fern]

also anime boobs

[fleet harness]

I bet its a lot of anime centric folk complaining ftmp

[glad grove]

least saturated node in blender

[upbeat summit]

https://tenor.com/view/hugh-jackman-unbelievable-hugh-jackman-gif-5281866

[royal fern]

given I have a 3060 12gb... I don't find SDXL slow considering it's generating at 1024... it's like 4x images in 1.5

[fleet harness]

all this for a generative bowl of cheerios? ||jk||

[uncut steeple]

People complaining that the newest tech doesn't run on their fridge

[fleet harness]

It wont run on my 3DS, I dont want it

[royal fern]

but it doesn't run on my laptop with a 1050!

useless!

[fleet harness]

unless it gets Celeron support I'm not interested

[upbeat summit]

my nokia 6210 runs it just fine

[hardy cipher]

these images are getting wild, lol

[royal fern]

dude when you are frantically jerking off on celebrity fanart p0rn those images have to come out fast!

[upbeat summit]

oldschool collages

[uncut steeple]

I think the user group with the biggest complains also prompts for the biggest boobs

[hardy cipher]

they took a weird turn

[royal fern]

it was like this when I found it

[glad grove]

im tired of having to move around the thousands of strands of hair that is blender nodes just to render 10seconds of looped 3d porn,i can finally move away from that installs comfyui

[late marsh]

https://media.discordapp.net/attachments/1131466616526344292/1135470936657768498/loki_madness.gif

[upbeat summit]

I have seen many many millions of images in my life, and your images trick my brain quite often thinking at first glance they are real photographs. well done 😉

[royal fern]

yeah the only thing that gives it away is

[upbeat summit]

it's the lora trigger 😄

Nido

[upbeat summit]

that post timing...

[woeful patio]

have y'all seen this yet: https://www.reddit.com/r/StableDiffusion/comments/15hag5s/sargezt_has_published_the_first_batch_of/?utm_source=share&utm_medium=web2x&context=3

can we use it in comfy?

[royal fern]

yeah I downloaded a couple of them but too lazy try to 😄

[upbeat summit]

yo - the QoL feature package https://github.com/failfa-st/failfast-comfyui-extensions by Pixelass added an option to hide all links / cables

[tribal jackal]

Thank you so much!!!

[vivid tide]

That’s right. I’m just gonna reinstall then maybe that’s a fix for it

[ionic dragon]

For real?
https://huggingface.co/SargeZT

[upbeat summit]

that's pretty good

[vivid tide]

Good morning!

[royal fern]

yeah the quality doesn't seem great at the moment

[woeful patio]

yeah but the examples look pretty poor 😦

[upbeat summit]

Hi! Good morning 🙂

[ionic dragon]

OMG each control net is 5gb

[royal fern]

yeah

SAI said they are working on a smaller version

[ionic dragon]

As he is an individual it might not be the best, but still he has room to make them better

Something is better than nothing

[tribal jackal]

The reason I went through the hassle of setting up all those nodes is easier to see with this prompt in particular. I wanted a greater variety within the same theme from one prompt.

Each image varies slightly from its paired image, and each pair of images varies more from others the earlier the stage they separated from was

[ionic dragon]

They are gonna prune ig, still waiting for them to release

[trim orbit]

they're diffusers too. it works good enough when i get it going in comfy ui but isn't super reliable. and i also can't get the annotators loading in comfy so well

[woeful patio]

absolutely and i really appreciate anyone pushing this technology forward

[trim orbit]

comfy's internal version of cnet isn't just pruning. whole new architecture so that you can use many without bogging xl down hard

[royal fern]

https://imgur.com/a/gSJ4vbN#3dRzA4z

(sorry I replied to you, it was meant for anybody 🙂 )

[trim orbit]

that controlnet in comfy? heh

[tribal jackal]

lol nah just my neuroticism

[royal fern]

train a lora with comfyui workflows

[tribal jackal]

https://tenor.com/view/that-not-how-it-works-thats-not-how-any-of-this-works-gif-15098334

[trim orbit]

just use kohya instead

tools for what they're specialized for

[royal fern]

it.. was... a joke

like take screenshots of comfyui workflows and train a lora with them

[tribal jackal]

oh shit

that's genius

I thought you meant use the nodes to train a lora, and I got a good laugh

[ionic dragon]

Yeah I have seen it, they are brilliant, but I don't think they would be out anywhere soon

[royal fern]

stop killing unicorns!

[vivid tide]

-I was able to fix it. I had to extract the folder in in the modded_nodes one.

[ionic dragon]

Yeah Joe said that they are moving to New arch

[ionic dragon]

Are we part of this unicorn?

[woeful patio]

are pictures with sunglasses on a problem for training loras?

[trim orbit]

nope. one or two in the set are fine. caption "sunglasses" on them too. should work well

[woeful patio]

cool, thanks

[glad grove]

this one is cleaner

[uncut steeple]

Thats blender

[tribal jackal]

ComfyUI, my friend!

Yep!

[glad grove]

u right,if it was comfy it would have more nodes 🤠

[royal fern]

who's a stable dog?!

[tribal jackal]

when in doubt, drag from the node you want and pick primitive node under "util"

aside from that, break it til it works

[trim orbit]

i dont think rip. auto can make moves and keep things relevant yet. i think theres a ton of benefits to auto for new users still.

[tribal jackal]

oh i have no idea

[trim orbit]

professional and prohobbyist workflows will move away from tools like auto though