✨｜sdxl

[high skiff]

this whole PC is brand new

and it happened across 2 PC's

I had this issue on my old PC as well

And the issue happens with my 3060ti, my 3080, and my 3090

The 3060ti would Idle at like 80 watts with any AI program open, the 3080 would idle about 160, and the 3090 about 130

[indigo carbon]

I think this might just be it, img+img while maintaining AIT =]

might be polished enough for release, idk

[soft zealot]

still looks & feels like ReVision despite your assertations to the contraray ;o)

which isnt a problem BTW 🙂

[indigo carbon]

I started on it, then worked my way to support AIT.. seriously, if you were to use AIT on ReVision workflow it will give you an error that discord can't fit..

it's not exactly like revision though

[spring fulcrum]

Can you post an image with that workflow? I'd like to give it a try with that drawn look.

[soft zealot]

#✨｜sdxl message

[soft zealot]

there's no issues with borrowing bits, I do it all the time 🙂

But this is essentiallty Revision adapted for use with AIT 🙂

[indigo carbon]

not really. if you would use the conditioning like with revision it won't be compatible with AIT.

also revision doesn't blend images like this can

[high skiff]

ok, I just noticed something strange

restarted my PC, idle was 24 watts

opened afterburner, and now my idle won't drop below 40

[soft zealot]

this is the basic Revision Process

I get you're then doing it slightly differently but...........

[indigo carbon]

that is a modified version of an older release of my workflow..

[high skiff]

I need to figure out whats broken

@visual glade Any ideas on why pytorch would force my GPU to always be at max clocks when doing nothing?

[soft zealot]

ok 🙂

[indigo carbon]

@soft zealot I'm sorry, did you just show that you have access to things I didn't post?

[soft zealot]

how could I have access to something that wasn't posted?

[spring fulcrum]

How do I connect my GPT account to that Simple O promt node?

[soft zealot]

look at )omar92s page for the quality of Life stuff, you need to put your api key in the config json in the Omar92 folder

[indigo carbon]

first of all- the last screenshot you sent is not how my workflow is organized, second- it seems that you had access to an older image containing a prototype of that workflow, one which I haven't uploaded. I'd like to have an explanation where did you get that file.

[spring fulcrum]

Thank you...

[soft zealot]

anything I have was at one point posted here even if it is no longer here . I have no "magical" access to anything else.

[indigo carbon]

this is not true. I never uploaded an image in here that contains even a prototype of that workflow. this is odd

[hardy cipher]

got backtraced

[indigo carbon]

honestly, I don't appreciate people digging into prototypes of unfinished PRs then claiming it's something entirely different. I did in fact start by looking at how revision was built, then saw what was holding it back from supporting AIT, adjusting the nodes to actually fuse images- then result in this.

[nimble heart]

have you upgraded anything lately? Mine was doing that for a while but either a pytorch or driver update fixed it.

[high skiff]

nope, had this issue for months and months across different PC's and GPU's

[nimble heart]

Oh so it's not new

[high skiff]

no, I just only realized recently that it was abnormal

[nimble heart]

Does it do it on text generation ui?

[high skiff]

yes

any form of AI, my GPU forces max clocks and uses 130 watts at idle

[nimble heart]

damn. Mine idles at 17w with SDXL loaded

rare AMD win lol

probably not. likely just an nvidia windows ai driver shennanigan. Been seeing lots of people having issues with drivers from the last few months in this chat

[high skiff]

mine ususally idles around 20-25 watts (Cause I have 3 monitors, and one is very high refresh)

But the second I open any AI program in CMD, it hikes to 130 watts

Which is more power than it takes for me to play effectively maxed out CP2077 at 75fps at 2560x1080

So its really stupid

[spring fulcrum]

What do you use to monitor GPU Watts? Now I'm curious

[high skiff]

it just sits there doing nothing

[high skiff]

HWinfo 64, afterburner, and GPU Z

GPU Z is probably the best for general GPU monitoring

[nimble heart]

amd kernel driver's native hardware monitor 💖

On windows I use hwinfo for everything. seems accurate enough compared to the direct hardware monitoring on linux

or was it hwmon

idk the dell one

"cpuid hwmonitor"

try that one @high skiff

[high skiff]

I am asking in the python server if anybody happens to know how I could try and fix this pytorch issue

[nimble heart]

https://www.cpuid.com/softwares/hwmonitor.html

[high skiff]

I used to use that one, same readings

[nimble heart]

hm

[high skiff]

my GPU has like 16 thermisters IIRC

[nimble heart]

idk what a thermister is

[high skiff]

and GPU Z reads them all, which is nice

[nimble heart]

you're expecting too much from someone who didn't go to highschool

[high skiff]

small thermometer basically

[nimble heart]

ah

[high skiff]

*Thermistor

[nimble heart]

mine has a whole two that are exposed via /sys/

wait

3

[high skiff]

counted, it has 12 exposed

[nimble heart]

mine has edge, junction, and memory

[high skiff]

core, core hotspot, VRAM, PWR, and 8 iCX monitors

[nimble heart]

first 3 are the same just different names

[high skiff]

thats pretty exclusive to this model specifically

[nimble heart]

core == edge hotspot == junction vram == mem

[high skiff]

I was meaning 12 different readings

[nimble heart]

Ah

yea

anyways

did you have a dual boot or no

[high skiff]

core, core hot, memory junct, power delivery, and 8 boardwide iCX

[nimble heart]

if you had dual boot I'd try it on linux see if it's a windows nvidia driver fuckery issue

[high skiff]

no, I have never done dual boot

[nimble heart]

hm

[high skiff]

I do not consider Linux for anything.

[nimble heart]

could try the pytorch 2.1.0 rc1 build that's out now

it's compiled for cuda 11.8, 12.1, and rocm 5.6

both windows and linux

[high skiff]

I have Cuda 12.2

[nimble heart]

it's under the test subdirectory instead of nightly

[nimble heart]

torch doesn't even have a 12.2 version?

unless it's statically linked and it doesn't matter

[spring fulcrum]

Sitting at pretty much idle with sdxl comfy on but not generating anything this is what I get:

[high skiff]

I have no idea, I just know I have Cuda 12.2, and its what I have used since SDXL came out

[high skiff]

4090?

[spring fulcrum]

yes

[high skiff]

makes sense, looks about right

[nimble heart]

50w is still a lot unless you have a youtube vid or something open

[high skiff]

maybe a tad high

yeah

[nimble heart]

i got a 4k screen and I idle @ 17w with just the desktop and comfyui

[high skiff]

high refresh monitor could be a culprit tho, my 165hz monitor does increase my idle power by a considerable amount

[spring fulcrum]

Oh mine is 47 inch ultrawide at 240hz

[high skiff]

yeah, in that case, makes sense your idle would be a little higher haha

[nimble heart]

so it's 100% a sytan problem

[high skiff]

I have seen posts about this issue on linux, multiple actually

[nimble heart]

I had it for a bit then it went away

[high skiff]

but no info that really helps me here

[spring fulcrum]

I honestly cant stand it. I miss my triple 4k monitors.. I thought ultrawide would be nice but its a pain in the ass

[nimble heart]

Started when something updated then went away when something updated

[high skiff]

it can be, yeah

I have an ultrawide

I have 3 monitors

[nimble heart]

update comfyui too just in case

[high skiff]

its not that, but alright

this has been an issue for well over half a year

before comfy UI was even out

[nimble heart]

ah right text gen does it too

[high skiff]

and RVC

[nimble heart]

its 3 am here brain on 2%

try the torch release candidate for cuda 12.1

[high skiff]

comfy is downloading torch 2.1.0

[spring fulcrum]

@soft zealot Thank you... This is pretty much what I was looking for

[nimble heart]

nightly or testing?

[high skiff]

nightly

I always use comfy nightly

[nimble heart]

since the testing branch was just cut out yesterday I'd use that so it's a better chance of working

Ah alright

[high skiff]

BF16 VAE is too delicious to not use nightly

[nimble heart]

testing should have that now too

since 210 is in testing

you'll still need nightly vision though

[high skiff]

I have comfy UI nightly, and it installs all the things it needs for nightly to work

[nimble heart]

idk what that means to be honest

I just do everything the normal way

[high skiff]

its a branch of comfy UI

it uses some different requirements, like nightly pytorch to begin with, which supports BF16 VAE, which is daddy

[nimble heart]

lies

[high skiff]

1.8 second 2048x2048 VAE decode on a 3090

[high skiff]

I don't remember how I did it, comfy helped me set it up

[nimble heart]

comfy just pushes everything to master

[high skiff]

but it automatically uses pytorch nightly for everything

it needs special command line args too

[nimble heart]

is the nightly version pinned or does it re-fetch the latest?

[high skiff]

--bf16 VAE

If you use that not on nightly, it crashes

[high skiff]

it re-fetches

I just updated, and it auto installed pytorch 2.1.0 nightly

[nimble heart]

yea

should say 2908 or something for the build #

updated vision today

[high skiff]

where do I find that?

[nimble heart]

assuming the command prompt has the pip log it should be in the history up there

[high skiff]

WAIT

my GPU is only drawing 38 watts with comfy open-

OWO

[nimble heart]

did the update finish

[high skiff]

yeah, I have the UI open

[nimble heart]

or did it just randomly do that

nice

so probably

whatever was going on in nightly torch around the time XL 0.9was out

[high skiff]

woooo!!!!

[nimble heart]

was bugged

and we both had it

[high skiff]

I generated, went up to 390 watts, then 80, then back to 38

[nimble heart]

but I fixed it when I upgraded to rocm 5.6 or something

and you just never did

[high skiff]

I guess that was the issue all along. So weird

[nimble heart]

yea

[high skiff]

I have had it like this for MONTHS, across all AI programs

[nimble heart]

that's why I'm not using nightly anymore lol

[nimble heart]

do you share a venv or something???

or did you just like install torch nightly at the same time on all of them

o

wait

if you installed the borked one first

pip will fetch it from cache

could be

time to update everything ✨

[high skiff]

I only started usning nightly like 2 months ago

I have had this issue since like January

[nimble heart]

bruv

[high skiff]

IDK man, but I am HAPPY

this is huge for me

[nimble heart]

no more space heater

[high skiff]

my room should be soooo much cooler

[nimble heart]

yea I literally got rid of my space heater

[high skiff]

cause 130 watts for hours and hours adds up for sure

[nimble heart]

in the winter I just let stable diffusion or something run

[high skiff]

Our electric rates here are criminal, and its hotter than Satans left nut too, so any energy/heat I can reduce is

[nimble heart]

fun fact, space heaters and gaming PCs draw the same amount of power for the same amount of heat

[high skiff]

yeah, its a standard heat formula

[nimble heart]

if you manage to build a 1500w gaming PC it'll put out as much heat as a full size space heater

formulae are for nerds

I just try random numbers until my code works

[high skiff]

yup, but it will be doing something additionally beneficial for that same heat, which is a huge benefit for a lot of people. Bitcoin miners were popular as space heaters in some regions after the crash, cause they made a little money for the same power/heat

[nimble heart]

if I move back to north dakota I'm 100% making a mining rig

[high skiff]

and they were being sold at dirt rates cause it wasn't viable for just profits

[nimble heart]

or at least setting up idle mining on my main machine

so cold there not even heat pumps work well

[high skiff]

pseudo is working with me to try and help me build a reclaimed solar grid for our house here, cause our eletric bills are actually physically painful here

[nimble heart]

our house was in the last group before they started adding solar panels to all the new homes

so we don't got any either

[high skiff]

goal is a 5KW array if I can

[nimble heart]

goal is to leave this shit state

[high skiff]

Pseudo said I should be able to do 5KW pretty easy for a few thousand

He has a 1.5KW system he made for like $1500 I think he said. He can't use the other panels he paid for cause he doesn't have anymore room

[nimble heart]

1.5kw + a few car batteries and you could basically power your computer for free

[high skiff]

we just need to offset the energy use of our AC, cause its so fucking expensive

[nimble heart]

old house?

[high skiff]

the plan he had for us, if we wanted to be serious was to buy a lot of reclaimed/damaged panels (cause we have a ton of property for solar panels, so lost efficiency from damaged ones is better than paying more for working), and then getting a forklift battery

[high skiff]

no, it was built in 2007

it just runs for 8-12 hours a day where we live

cause its so fucking HOT

[nimble heart]

nice

I can get away with it not being perpetually on if I raise it to 77F

[high skiff]

we have ours at 78

so its not even like we are living comfy cool

[nimble heart]

yea you sound pretty sticky

especially with a 3090 ftw

[high skiff]

we just have high winds, high temps, and severe UV/infrared radiation, which saps cool out of the house soooo fast

[high skiff]

luckily, I run it at 50% power, at which is pretty fucking sick

[upbeat summit]

@high skiff I always wanted to ask if you came up with the terms Linguistic (clip_g) and Supportive (clip_l)?

[high skiff]

same power use as my old 3060ti, and much faster

I did

[nimble heart]

I haven't tried undervolting my card

the kernel didn't even support it till like 3 days ago

[high skiff]

I did hundreds of clip tests and found that G worked much better when you spoke to it, and L much better when you tagged to it

my 3090 is insanely efficient with underclock. I love it

my 3090 at 210 watts is faster than my old 3080 at 390 watts

throw in AI template, and its as fast as stock without AI template

[nimble heart]

I have my XTX set to "silent" using the physical bios switch but it's a difference of like 1%

[high skiff]

yeah, I switched my FTW to its normal bios

its OC BIOS goes to like 480 watts

[nimble heart]

silent power targets 327w OC targets 339

[high skiff]

its normal goes to 420, but it never goes over like 320 watts for most things when unpower capped

[upbeat summit]

thanks! yeah, it makes sense based on what clip_g and clip_l are. I'm still trying out different setups how to use it, but I was interested how the names were created.

[nimble heart]

gotta love that 2mhz idle lol

[high skiff]

but with the full 420 watts, undervolting and overclocking, I can get this bad boy to like 2100mhz at 420 watts, which is like a 15% perf gain

[high skiff]

NO WONDER ITS SO EFFICIENT LMAO

my GPU idles at 210/405

[nimble heart]

my card hit 3200 Mhz @ 430w on the OC bios while playing cyberpunk with RT

room so hot

[high skiff]

yeah, 7900XTX is a much higher clocking card

[nimble heart]

I'll have to try underclocking it

assuming the lower power doesn't destabilize the chiplets

like I've been told it might

gpu reset fun time

[high skiff]

with my settings, I play maxed out ultra (No RT) at 2560x1080 @75fps locked with DLLS balanced for only 100 watts of GPU, and its amazing

Up until now, I chose to play cyberpunk during the day instead of idling AI, cause the AI took more power! lmfao

[high skiff]

try undervolting first

[nimble heart]

I will once the 6.5 kernel lands in my repos lol

I physically can't right now I'm pretty sure

[high skiff]

my 3090 was able to stay at 1950mhz with a -130mv undervolt. Saved like 50 watts I think

my 3080 was an undervolt king

that bad bitch stayed at 1800mhz at only 805mv

[nimble heart]

I've got my CPU undervolted as fuck so I can max out like 8 cores before my fans even ramp

hits 5.9 GHZ on a single thread

[high skiff]

I have been meaning to try and undervolt my CPU

[nimble heart]

amd?

ryzen cpus literally have an "undervolt" slider you can just move

and it adjusts everything else for you

[high skiff]

intel

[nimble heart]

oof

[high skiff]

12600k, got it for a great price

it was a way better buy than AMD when I bought

[nimble heart]

ddr4?

[high skiff]

it was 12600k for $160, or 7600x for $240

[high skiff]

DDR5

[nimble heart]

oh neat. hopefully mobo and ram wasn't too expensive

[high skiff]

mobo was $100 (bought it from amazon warehouse), and 64GB DDR5 6000mhz CL 36-40-40-40-80 was only $160

[nimble heart]

am5 it's like $220 for a micro atx board

[high skiff]

this board is normally $200, I saved 50% by going used: Acceptable

[nimble heart]

i got a good deal on my 7900x though so worth

[high skiff]

in the case of this board, the "acceptable" part was, it comes with nothing else. Just the mobo in a bag in a cardboard box

which is all I needed!

so I saved 50% lmao

[nimble heart]

ya I actually wanted to have screws and my wifi antennae lol

I stopped buying used electronics after burning through three PS2 Slims and like a dozen controllers

[high skiff]

the board looked brand spanking new. the socket still had the sticker, same with the CMOS battery

as far ask I know, it was only deemed acceptable (by their terms) because of "Missing critical parts required in order to function"

but I just put it in my current case, so I needed no mounting hardware, and I am ethernet, so no wifi was needed

[nimble heart]

i mean if it wasn't a part swap it was missing all the screws and everything lol

[high skiff]

and I already have a 1.5gbit wifi dongle if I DO need it

[nimble heart]

so it wouldn't work for a new build

[high skiff]

yeah, but for me, perfect 😎

[nimble heart]

unless you seriously bubba'd your case with some home depot screws

[high skiff]

A steal, if you will

Its a solid board too

[nimble heart]

yea I really prefer them over liquid boards

never again...

[high skiff]

B760, 3x m.2, DDR5, 2.5GbE LAN, Q-Flash Plus, USB 4.0, Error LED's, WiFi 6E

Even looks nice

[nimble heart]

what leds

[high skiff]

IO is a little limited, but that seems to be an issue with all B760 boards

[nimble heart]

there's no 7-seg

[high skiff]

oh no, 7-seg is considered "premium" now

[glad grove]

where is the rgb ?????

[high skiff]

it has a patch of LEDs that can light up for DRAM, POST, CPU, BOOT, and something else

And depending on the color and pattern, you can tell whats wrong

[nimble heart]

those fucking single light setups that just scream either "RAM" or "CPU" when literally anything goes wrong don't count they may as well not exist

terrible trend

0/10

my board has those too I hate them

[high skiff]

Mine does Boot, DRAM, CPU, POST, and colors/flashing combos for specific issues

[nimble heart]

my old EVGA board has twin 7-seg leds and a beeper

got 200% of the information you needed

[high skiff]

like my 3 flash red with DRAM was able to tell me I had the sticks in the rong config, cause this board doesn't share the same RAM stick preference as the DDR4 version

and I had no manual sooo... lol

[high skiff]

yeah, it really sucks, 7 seg displays are reserved for high end mobos now

I think the cheapest new gen MSI one for intel is like $390 or some shit

[nimble heart]

that's fine but high end motherboards shouldn't cost $700

[high skiff]

gamers nexus made a whole video about how the 7-seg cutoff is fucking stupid, cause its like $0.10 to add one

[nimble heart]

my evga FTW motherboard with literally every option was $350

[high skiff]

hell, there are mobos now well in excess of $1k

[nimble heart]

the b boards nowadays still give you all the overclock control so those X/Z boards are just excess

$1200 fuck outta here

probably a reseller

[high skiff]

interesting

[nimble heart]

msi has shitty bioses too

[high skiff]

no

its that much

I wish I was kidding

I think Aorus has one like $1600

[nimble heart]

the 7950x3d is a $600 chip so that one's not crazy

[high skiff]

hell $1200 is cheap for it now

lmfao

[nimble heart]

ew

I've had nothing but problems with my msi b650

every single bios version has severe bugs

so for $2.1k my faith is low

[sand pier]

700

[nimble heart]

I'm using bios 5 currently cause 6 causes my PC to take 15 minutes to post

[sand pier]

But still not crazy

[nimble heart]

but 5 you can't reboot into firmware

[high skiff]

this one is reasonable for what it is, IMO

[nimble heart]

so I have to reboot and hammer DELETE

[glad grove]

this one is better

[nimble heart]

server boards are different

[high skiff]

you're not even comparing the same class of hardware lmao

[glad grove]

we just throwin money around now

[high skiff]

that mobo is nothing compared to the one I posted, they are not made for the sam ething

[nimble heart]

for regular am5/intel whatever boards should be capped @ like $450

damn its 4 am

imma sleep

[glad grove]

no

[high skiff]

2x10 GbE LAN, 7x PCIE 5.0 x16 slots, 2TB RAM capacity

[nimble heart]

remember to update your torch on ooglyboogly and others @high skiff

[high skiff]

oh lord

[nimble heart]

see if it fixes the 150w idle

[high skiff]

how do I do that?

[nimble heart]

very carefully

[high skiff]

I have no idea how to force nightly on all of them

[nimble heart]

do they use conda or just normal python

[high skiff]

ooba uses python, silly tavern uses miniconda IIRC

but silly is just a frontend to ooba, so it likely doesn't have the same idle bug

[peak dove]

I want one 🙂

[nimble heart]

if it's just python and there's a /venv/ folder you can just open powershell

./venv/Scripts/activate.ps1
pip3 install -U --pre torch --index-url https://download.pytorch.org/whl/nightly/cu121

replace nightly with test for the stable-er build

[sand pier]

Motherboards that costs the same as 4090 lol

[high skiff]

you need an $8k processor for it lol

[peak dove]

'k

[sand pier]

Wait more probably if it's a threadripper

[high skiff]

its XEON

[sand pier]

Pain

[high skiff]

the minimum one that can use those lanes is 8k IIRC

GYAT DAMB thats a lot of IO

IOwO

[sand pier]

I want a 7950x3d but is 900 bucks ish here, it's a bit much lol

[high skiff]

I wonder how many TB/s the throughput of all of those traces together are lmao

[high skiff]

why x3D over the standard X? The standard X is no slower, and is faster in a lot of cases

in order to get proper gaming perf on the 7950x3D, you need to disable half the cores

[sand pier]

Part of it is temps

[fierce void]

I went with a 7950x - the x3d is compromised with additional cache being available to only half the cores

[sand pier]

But yes that core disabling is a major pain point

And probably enough for me to nope regardless

[high skiff]

yeah, the 7950x3D is a goofy ass product. Its slower than the 7800X3D in single core, and slower than the 7950x (non 3D) in multicore

its like... why?

silly goofy

[fierce void]

7950x with the undervolt settings isn't getting beyond 40c at full load with an AIO cooler

[nimble heart]

7900X is fun. 5.9ghz just using the built in boost control

[high skiff]

the 7900X3D is just as bad

slower than 7900X in multi, slower than 7800X3D single

[nimble heart]

i've got two 7900's in my system

[sand pier]

I run 5900x currently

[high skiff]

towo?

[nimble heart]

7900X cpu with 7900XTX gpu

[sand pier]

Was great CPU return it released

[nimble heart]

now I just need an additional 15 gigs of ram so I have 79 total

[high skiff]

I have a 12600k, and I love it. So much faster than my old CPU

12600k solved all of my huge single core bottlenecks

[fierce void]

I went for the best I could get at the time when I built my new system late last year, only thing I regret is going with Asus for the motherboard.

[nimble heart]

yea I came from a 6700k and it's like "holy shit everything loads so fast"

[sand pier]

Yeah I would not go Asus this gen

[high skiff]

I came from 7700k, so basically identical

And yeaahhh

especially for music prod, oh my god

[fierce void]

I've basically given up on Asus being able to produce anything other than overpriced, under supported garbage with very ropey quality control :/

[high skiff]

7700k was getting obliterated by my sound design

[sand pier]

I've read so many horror stories, and their products really have trash ratings on the sellers page here

[nimble heart]

refusing to upgrade computers for 5+ years gang 🤝

[high skiff]

not refusing, I was BROKE

[fierce void]

Yeah, I did the 5+ year wait

[high skiff]

this single sound could likely still pin my 12600k, or hell, probably even a 13900k with how severe I am with my production lol

it took over 3 minutes to render on my 7700k

[fierce void]

went from a 7700k on a z190 to a 7950x on a X670e 😄

woosh

[nimble heart]

how

[sand pier]

I went from 6700k to 5900x myself

[high skiff]

but thats what a 46 effect chain will do to a completely synthesized sound

[nimble heart]

audio with multiple filters in ffmpeg encodes at like 20x realtime for me

[high skiff]

Its a lot to explain

[nimble heart]

obviuosly synths will eat more but still

[high skiff]

that sound is insane

that sound is wasteful

[fierce void]

only rough spot in my system now is the 3080 10GB

[high skiff]

hell, I used pitchmap on it 2 different times, which 1 of those alone can almost pin an M1 CPU

[sand pier]

10 GB is a bit little vram yeah

[nimble heart]

needs a new comfyui workflow smh

[fierce void]

yeah, can still push out 1920*1200 SDXL renders though

[high skiff]

also, for anybody curious, this is what it sound like in context to the WIP demo

Its a small ID for a track on my album

[sand pier]

I sold my 3070 with significant loss but oh well

[high skiff]

very early WIP

trying to think if I have any other insanely wasteful sound design samples

oh, I have one

[fierce void]

Yeah, I had to buy a new gpu at completely the wrong time

3080 was around £1300 :/

[high skiff]

jesus, my 3090 was $700

[fierce void]

But that was somewhat reasonable at the time

[high skiff]

my 3080 was $340 brand new

[fierce void]

yeah, old gpu died at the peak of GPU insanity

I still have an 1080ti as a backup backup though, should have probably gone with that in hindsight, but oh well...

[high skiff]

here we go

[sand pier]

Should have sold it when it had great value

[high skiff]

another insanely processed sound

first is rigth out of synth, second is the processed output lol

[sand pier]

I sold my 1080 for like 500 bucks in the great GPU crunch

Which ate a lot of the pain of the 3070 I got

[fierce void]

Yeah, I sold some old trash gpu's I had laying about for a silly amount, but didn't want to part with the 1080ti

[sand pier]

I have a old 660 ti laying around for debugging that's it

[high skiff]

I got a brand new 1080 for $300 when there was the GPU scare after turing launched (from a friend of my uncle)

Sold it 2 years later for $250, bought my 3060ti for $400 3 days after it launched

can likely sell it for $400+ cause its considered special now

Got my 3080 for $340, gonna try to sell it for $400+

got my 3090 for $700

[soft zealot]

Im still using the 1080ti, doing quite well even with SDXL 🙂

[high skiff]

the 11GB VRAM helps

[soft zealot]

yup

[high skiff]

you can give it more time, you can't give it more VRAM

[sand pier]

I have a 4090 and I've ran out of vram a few times xD

[fierce void]

Yeah, stable diffusion wasn't the reason I moved to the 3080 - trash frames at 1440p were 😄

[high skiff]

yeah, next gen needs to increase, 24 is restrictive for the highe nd now

[fierce void]

I'm hoping that we'll start seeing 32GB

[high skiff]

I upgraded from 3060ti to 3080 cause it was a steal of a deal

Then from 3080 to 3090 cause it was a good deal and I wanted to get more serious with AI

[high skiff]

rumors for 5090 is 36 IIRC

[soft zealot]

but at what price?

[fierce void]

I'm done with nvidia tbh

[high skiff]

no idea

[sand pier]

That's rumoured for 2025

[high skiff]

same, fucking hate NVIDIA

[sand pier]

Long lol

[high skiff]

just too bad that AMD is a bad option for a lot of AI

[fierce void]

too big, too power hungry, too greedy

[sand pier]

I'll buy whichever GPU makes the most sense at the time lol

Both brands suck

[high skiff]

my fingers are crossed for intel to do some cool AI stuff. Cause their new XEONs have some super powerful AI accelerators on die

[fierce void]

I might nab a second hand 3090ti at some point, but that's about it for a while

[high skiff]

used 3090/ti are the best options right now for sure

thats why I went 3090

[fierce void]

yeah, prefer a ti as they won't try to melt the ram

[royal python]

Imagine if we could just upgrade vram for cheap

[high skiff]

yeah, big reason

my 3090 has great VRAM temps cause EVGA made it right, luckily

front and back VRAM is padded

[fierce void]

yeah, my 3080 is an EVGA

[soft zealot]

the biggest dissapointment recently has to be the 4060ti 16Gb

When I twas initially announced I thought " oooh that might do as an upgarde if the price is right, which it wasn't.

Then it turned out to be crippled on the memory bandiwdth as well

[high skiff]

I miss EVGA

best GPY brand ever man

[high skiff]

oh man, its a literal dogshit worthless card

[fierce void]

yeah, gutted that they stopped entirely

[high skiff]

horrifically bad VRAM, loses to the 3060ti in a lot of benches, considerably more expensive

Its fucking terrible

[fierce void]

16gb at half bandwidth or something dumb wasn't it?

[high skiff]

16GB DDR6 at 128 bit bus

[fierce void]

128bit, oof

[high skiff]

the 4060 non ti is even WORSE

8GB 128bit bus at only PCIE 8X

so if you have gen 3 mobo, its only x4

[nimble heart]

my 7900 XTX does pretty good in the places that use it properly. LLMs using AutoGPTQ its competitive with a 3090.
Stable diffusion has memory issues though.

[high skiff]

oh, and it loses to the 3060 in a lot of things too

everything below 4080 for lovelace is dogshit

[nimble heart]

if you're not in a rush could wait to see how rocm plays out in a few months

[high skiff]

4070 is terrible too

12GB GDDR6x at 192bit

slower than 3080 in raster

$600

[fierce void]

Yeah, I'll probably look for a reasonably priced 3090ti then wait it out until sanity returns

[high skiff]

I would recommend a 3090 and $10 worth of thermal pads. Its typically very easy to mod, and 3090's are way more common, and thus lower priced cause of availability

[fierce void]

I really want an EVGA 3090 though, and those are still commanding a premium

[high skiff]

I got my 3090 FTW 3 ultra for $700 shipped with tax

[soft zealot]

good products always will

[fierce void]

Yeah

[high skiff]

only downside was it was FILTHY

[nimble heart]

I play games more than I do ai so XTX was a no brainer. if you're only ai a 3090 would probably be worth the extra money

[high skiff]

but 2.5 hours of cleaning and its

[fierce void]

I'm keeping an eye out on ebay for anytthing good

[high skiff]

oh yeah, for gaming 7900XTX is fantastic

[sand pier]

People are so dumb with the prices with 3090s, great value sometimes tho

[high skiff]

I recommend the hardware swap discord. Thats where I got mine

staff are great too. Helped me massivebly with a fraud case (had nothing to do with them)

[nimble heart]

devil may cry 5 4k 200% scale 80fps. V looks so pretty

[fierce void]

Got a link handy?

[high skiff]

gotcha

have a look at this

My GPU when I got it

[uncut fiber]

what is the name of game cant run on any gpu 4k at 60fps? 😄

[nimble heart]

Jedi survivor lol

Kerbal space program 2

[high skiff]

Vs after 2.5 hours of manual labor

Dropped 15-25C on some parts. It was insane lmfao

[uncut fiber]

it is from EA

[indigo carbon]

gettin' closer

[high skiff]

star citizen

[nimble heart]

never played

[high skiff]

it can't run 60fps at ANY res on ANY hardware lmfao

[nimble heart]

yea

oh

[high skiff]

13900k pinned moment

[nimble heart]

dwarf fortress

[high skiff]

undertale

[uncut fiber]

probably

[high skiff]

||30fps cap kek||

[nimble heart]

30fps cap on what

[high skiff]

cyberpunk overdrive can only hit 60FPS at 4k on a 4090 with DLSS3.5

[high skiff]

undertale lol

[nimble heart]

undertale isn't 30 capped

[high skiff]

yeah it is

[nimble heart]

daduq

[high skiff]

first thing I noticed when I played. I am hyper sensitive to refresh rate

[uncut fiber]

which is on LEDs weird.

[high skiff]

LEDs?

[nimble heart]

I played that the week it came out so I still had my old crap monitor

[uncut fiber]

oh past? CRT. O.k.

[high skiff]

I am confused

[uncut fiber]

iam too 😄

[high skiff]

why do LED's have anything to do with it?

[nimble heart]

blinky lights

[high skiff]

ohhh

[uncut fiber]

refresh rate and crt vs LED

[high skiff]

I see what you mean

[indigo carbon]

would it be cool if I'll release a final version of this?

[high skiff]

I am very sensitive to refresh rate and flicker in general

[earnest anvil]

this is an image combining workflow?

[high skiff]

I can reliably tell the difference between 30/50/75/90/120/144/165 hz

can't tell higher than that tho

[indigo carbon]

yeah, not only that it also uses AIT so it's the same speed as my other optimized workflow

about 14-ish seconds per image with that

[high skiff]

opti-nice

[uncut fiber]

me to, i got helthy issues on crt

[earnest anvil]

can you direct me to your other workflow?

[indigo carbon]

https://civitai.com/models/133818/tdg8uu-aitworkflow

[high skiff]

one of my biggest weaknesses to refresh rate is strobe TV's

ugh, they make me feel rancid

[indigo carbon]

reference to minecraft? lol

[high skiff]

and fluorescent lighting, which we have here

[high skiff]

play on optimize

opti-nice

[indigo carbon]

I think this is better than MJ's image blend?

only part of that workflow that needs text is the refiner's negative for slightly better quality without performance tradeoff

[strong field]

everything SD is better than MJ

[indigo carbon]

I was talking about image blending capabilities

[strong field]

I know i just had to make a point

Unfamiliar with MJ image blend, but SD image blend is stellar

Revision, IPA, etc..

plus control nets, yikes

[indigo carbon]

this is built on top of Revision, I learnt how Revision functions, then made a version of the AIT workflow that wants 2 images, then it makes 1 as demonstrated here

also same speed as normal AIT workflow

also did this in a way that didn't require me to compile new modules, so big pog

I'm planning on releasing it so everyone can fuse 2 images in the same speed as normal AIT workflow which is fast

[strong field]

good stuff, i have revision/ipa and ait in my daily driver so definitely think people will go for it

its like AIT with control net in some ways

[indigo carbon]

oh did you also make a workflow like the one I'm working on? or did you mean you do those separately?

[strong field]

i have a workflow that is similar to what you go, but with the addition to IPAdapter, (can do 2 rev images also)

In fact, chaining Revision and IPA has become one of my favorite techniqes

except i would never do anybody the disservice of sharing this nonsense

so yeah, its really good for the community to have your option

i throw alot of shit at the wall and see what sticks

[indigo carbon]

this is what I'm currently working on

[strong field]

you can see multiple samplers, testing refining again with different settings, various upscale and downscale its a noodle pasta

[strong field]

This looks fantastic

[shy kelp]

so this is how old people feel when getting-out-of-the-loop. I just randomly check what's going on here, and all I see are cute doggos and unreal engine's blueprint system! 😵

[strong field]

heres what i threw together with my setup

The control is INSANE

[indigo carbon]

can you give me those 2 images? I want to see how my workflow does

[indigo carbon]

I think your IPA version did better here. when using IPA, was AIT still implemented @strong field ?

[strong field]

yes AIT very implement cant do without it hahaha

your weights will definitely come into account but without a doubt IPA is much better at constraining the composition

no two ways about it

simple hook up, 1 revision image like normal via conditioning input to ksampler

IPA > same Ksampler in model input

[indigo carbon]

I feel like if the weights of image A would be with IPA, it will be insane

[strong field]

yup you can make it just like that

I dont use control net anymore

[indigo carbon]

can we go to DM? we could make it better probably

[strong field]

hahahah

yeah for sure lets go

[eternal fog]

When you use AIT with IPAdapter does it use more VRAM

I've tried and it causes me to go over my max VRAM so it's slower.

Also did you get IPAdapter to work with batches?
When I try use batches it just ignores the IPAdapter.

[indigo carbon]

that's the thing, I didn't use IPA there, me and @strong field are trying to figure it out

[zinc cargo]

hi everyone, i'm trying to do CR CN stack but i get this error

[eternal fog]

Oh what did you use? Just the normal clip vision stuff?

[indigo carbon]

yeah

[eternal fog]

Interesting

[zinc cargo]

anyone can tell me what i'm doing wrong?
beside the regular stuff of course 🙂

[soft zealot]

@strong field loving these little cup cakes I got ReVisioning your images 🙂

[meager canopy]

I don't see the point of AIT. I get no improvement over anything, and it causes conflicts with some other nodes. What am I missing?

[soft zealot]

correctly configured and working it can give a boost to iteration speed

[indigo carbon]

what about speed? it should dramatically increase with 3000-4000 series

[meager canopy]

It takes about 20 secs per image, which is the same as I get without AIT.

I have a 4090

[indigo carbon]

try it with the AIT workflow, it's configured correctly

[meager canopy]

That is based on your workflow on Civitai. I just changed the sampler.

[spring fulcrum]

Did you make the galaxy time machine model on Civitai?

[indigo carbon]

it should speed stuff up.. I'll look int it; I have a few friends that have a 4090 and for them it almost tripled the speed =\

[thorny frost]

For 3090 up to a 50% only

[meager canopy]

Killing my performance

I had to stop the server to get control back 😦

[strong field]

hahaha, those are awesome

[soft zealot]

makes me want to nip out to the nearest bakery lol

[indigo carbon]

it could be your torch version. it should be 2.0.1+cu118

and xformers==0.0.20

[meager canopy]

It's the latest nightly build.

Doesn't do that with other flows

Maybe it's that sampler, I'll try another

[indigo carbon]

then that's your issue. AIT is built for 2.0.1

and xformers0.0.20

[meager canopy]

So no good for anyone with a 4090 not using xformers

[indigo carbon]

if you use ComfyUI, vanilla Xformers is on by default

[eternal fog]

Something odd is happening there as it's using all your VRAM

[meager canopy]

Yup

[eternal fog]

Mines been doing that when I use IPAdapter but it's fine it I don't.

I also don't use xformers

[meager canopy]

These? Yes
https://civitai.com/user/galaxytimemachine/models

[indigo carbon]

it still gave you the boost?

[spring fulcrum]

I use your SDXL model quite a bit. Great work.... Are you working on another model yet?

[eternal fog]

Yeah

[indigo carbon]

heh, peculiar

[meager canopy]

Not at the moment. I have another version, but I don't think it's enough of a difference to post.

[indigo carbon]

you probably use torch 2.0.1 right?

[meager canopy]

@indigo carbon Does that workflow I posted work for you?

[eternal fog]

Can't remember at the moment I think so.

[meager canopy]

It's in the image

[peak dove]

I'm entering the Civitai SDXL "Best Picture" Competition - that RTX 4090 is winging its way to me as we speak!!! 😉

[real jay]

not-fit-for-production

[soft zealot]

well thats an interesting result 🙂

[eternal fog]

Yeah revision does that

I've noticed that anything with muscles, even if it's on a woman tends to bias it towards making men.

[soft zealot]

in fairness I have heavily biased it towards the Bat if you look at the weight s:)

This ones quite neat

[peak dove]

My ComfyUI has rather randomly just started to crawl along --- about 5 mins/picture from about 40 seconds this a.m. ?!?!?!?

[soft zealot]

first course of action, restart it 🙂

[strong field]

Did you see my latest cake?

[soft zealot]

mmmmmmm cake

[strong field]

sometimes SD does things i didnt ask for....

[soft zealot]

as in most of the time lol

[visual glade]

update comfyui and if that doesn't work downgrade your drivers to 531

[uncut fiber]

Hi AIT worked today until i changed prompt. Any chance 531 drivers could help in my case?

[wet nacelle]

I'm back

[uncut fiber]

o.k. it seems good

[uncut fiber]

@visual glade
i get out of memory for stupid reason keep in memory was enabled in AIT. But not generating images, it still keep occupy 6GB. Is it o.k. Sorry for ping you

[eternal fog]

Restart comfy

I've noticed that if you ever cancel a gen or accidentally keep loaded enabled it gets stuck

I don't think it's a comfy issue, it's that module

[uncut fiber]

@eternal fog exactly what i did. Solved issue, but didnt know it is non comfy thing 🙂 Thank you!

[lilac wren]

ah, finally someone who understood that Afterburner is the best monitoring tool! know that you can modify the colors of the curves individually when they are "attached" together

[uncut fiber]

i think AB is some sort of swiss knife for all brand of gpu, not only MSI. @lilac wren

[soft zealot]

coming with the aid of one of your images again @strong field 🙂

[hardy cipher]

this guy looks like he eats paint chips

[strong field]

for the record i just copped that from a genius creator on civit

[strong field]

however, i do love that image and use it OFTEN

[strong field]

That is some SHEEN

[lilac wren]

Obviously. Enable "Perfcounter.dll" you will get bandwith, write/read disk and others good stats

[peak dove]

ComfyUI ERROR

[toxic moss]

Hello, 😃
I have just installed SDXL on Automatic1111.
I get poor results (images have been refined) and it takes 10 times longer than with the previous 1.5 or 2.1 models.
I must have made a mistake somewhere...
Can anyone help me?
Windows 10, Python 3.10.9
webui auto update
--no-half --xformers
cuda enabled in settings
token generated / activated
15GB stuffs downloaded
default settings in SDXL Demo tab
And I've used SD several times with different models

[uncut fiber]

What is your GPU? Performance of SDXL on A1111 is poor. You can switch to devs version, or go with Comfui. Where performance is great

[toxic moss]

2080Ti

[uncut fiber]

Maybe good or better to ask in #🤝｜tech-support
Would suggest Comfui, but very different way it works

[toxic moss]

I tested it

There is a lot of stuff I don't find, like img2img or scripts

so I tried to "go back" on webui 1111

But thanks, I post this on tech support 👍

[uncut fiber]

there is almost everything, but yes it takes time to learn.

[toxic moss]

ok great

[uncut fiber]

somebody claimed that devs version of A1111 performance is much better than normal version

[toxic moss]

If everything is in CommyUI I'll use it!
I like nodal

[peak dove]

First time that I have had this happen 531.61 is my nVidia Studio Driver

[toxic moss]

Oh ComfyUI result seems better

[uncut fiber]

i like all images. And speed?

[toxic moss]

seems very long

The flower I didn't check

Now it takes forever but i think it is crashing or something

I relaunched.
I get this dog in 20 + 7 sec
Seems ok this time!

[uncut fiber]

it looks like vendesian basset 🙂
good things to suggest, is render at 1024x1024 not 1080x1080 And Try Euler ancestrial @toxic moss

[toxic moss]

oh defauli is 1080, thanks

[toxic moss]

Euler ancestrial on both Ksampler ?

[uncut fiber]

would try both, as it can help you with performance bit maybe

[strong field]

a little behind the scenes

They arent all gems

[uncut fiber]

i made yesterday as well some knitting things 😄 But not gems

[peak dove]

I am serching for a home for this line of code in ComfyUI

PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.6,max_split_size_mb:128

[strong field]

Do you want to change that line of code or is it throwing you an error?

How much VRAM do you have? Are you using SDXL?

[peak dove]

I need to know exactly where to use it? In a CMD Window? In the ComfyUI .bat? In a MoonPig Birthday Card?

Today is the first time using ComfyUI that I have had this CUDA OOM. Most sites say use Mem Allocation in PyTorch

[strong field]

Reduce the size of your generation until it works

[peak dove]

But I don't know which file or folder to modigy? 🙂

[strong field]

You dont know if you need to modify yet, lots of settings that can make things act funny

[peak dove]

Yes, but as its SDXL, it has to be 1024x1024

[strong field]

Do you have a screenshot of your workflow?

Try 512x512 and see if it will spit something out, (it wont look good)

Just trying to track down issue

Also, are you running --lowvram ?

[peak dove]

No, I guess I could try --lowvram or even --medvram

[strong field]

From that screenshot you have 8gb which is barely enough for SDXL, you might find alot of hangs

I suggest tyring --lowvram

Just to see if it is more stable

[peak dove]

How do I add --lowvram into the ComfyUI .bat?

[strong field]

Windows?

Open notepad, drag and drop comfyui.bat

edit command line args in notepad, save, close, restart

[uncut fiber]

@strong field i have 8GB AIT is working, just must reinstall drivers for gpu to 531

[strong field]

yeah nvidia drivers acting a damn fool out here

[peak dove]

--lowvram still gets CUDA OOM

[strong field]

did you reduce to 512*512

[peak dove]

I downgraded my nVidia Driver to 531 because the latest (536) was slowing comfyUI

[strong field]

and now its throwing errors?

hmmm

[peak dove]

512 x 512 worked 🙂

[strong field]

try to add this to your command line args and try with 1024

--disable-smart-memory

(did 1024 ever work?)

[uncut fiber]

768x768 is useable, try it

[strong field]

yes, slowly move up in resolution until you break it

[peak dove]

768 x 768 works

[uncut fiber]

are you using AIT @peak dove

[peak dove]

No AiT - Base and refiner

832 x 832 wirks

[uncut fiber]

o.k. two models then

[peak dove]

864 x 864 is OK

[uncut fiber]

896x896

[peak dove]

928 x 928 OK

[uncut fiber]

oh you are using Sytan workflow? @peak dove

[peak dove]

960 x 960 Kapowwwwww

😦

[strong field]

your max res is somewhere around 1856 pixels

[peak dove]

But this morning it was fine

[strong field]

stay below that between width and height

try with 1024
--disable-smart-memory

[peak dove]

I can stay at 928x928 - but it was running perfectly at 1024 x 1024?

[uncut fiber]

@peak dove try 1024x576 it is all div 64 but not trained images, on other hand 16/9

[peak dove]

--disable-smart-memory failed at 960 x 960

1024 x 576 is OK

[peak dove]

1024 x 576 I meant!

1024 x 768 is OK

[uncut fiber]

do you have upscale in workflow?

[peak dove]

Upscale is there, but disconnected

[strong field]

768 + tiled upscale

[peak dove]

I tend to use GigaPixel to UpScale. I can do 100 6 x UpScales in GPxl in the time that 2 UpScales are done using Sytan's (excellent) w/flow

What I must do is invest in a 4090 - which has 3 x my current VRAM of 8Gb!!!

[lilac wren]

Yeah, tile methode break limits

[ripe summit]

Does anyone have comfyUI workflow for XL with lora and controlnet?

[peak dove]

Thank you for all the help! 🙂

[uncut fiber]

i am doing batch resize when needed in A1111 🙂

[wet nacelle]

mods are poopy and I dislike that. total poop mang.

rather frustrating

Who here actually has a workflow that they stay on? I just hop from one to the other when needed.

[ripe summit]

Do you understand all the nodes or do you copy other peoples workflows?

[wet nacelle]

copy while understanding only some of the nodes.

[strong field]

https://github.com/JPS-GER/JPS-ComfyUI-Workflows
https://github.com/SeargeDP/SeargeSDXL

[ripe summit]

Thanks man!

[uncut fiber]

this reminds me one con from scooby doo

[wet nacelle]

(better)

[uncut fiber]

he can wear sunglasses 😄

[wet nacelle]

inky mess on artistic frontal wound

[uncut fiber]

i can get mona lisa playing metal with band and fans, somehow cant with skeleton.

[wet nacelle]

Did I send this already?

[uncut fiber]

me? not, you? not

[wet nacelle]

Fixed it

[uncut fiber]

no 🙂

far from good.

[uncut fiber]

Mexico!

what old master used this colors? I think it is in metadata, just curious if somebody will know.

[stone fossil]

Piet Mondrian

[uncut fiber]

no, older master 🙂

[upbeat summit]

enlighten us!

[uncut fiber]

it isnt image, it is if he make some abstract picture.
Vermeer! 😄

[upbeat summit]

yeah the color palette is similar to many of his works, true

[vital wraith]

Is that done via prompting or a LoRA?

[stone fossil]

Its a fine tuned model and a lora that does not make sense to use but I could not care to unload it cause im lazy and prompting ofc.

[soft zealot]

sooooooooo I had a mad lad idea this evening.

With ReVision if instead of selecting 2 images to load what if I created the image dynamically ?

Add a little jeopardy?
just playing with it currently , Image 1 is prompted, Image2 is a Lexica sourced random prompt, negative is common.

[vital wraith]

Not sure what happened here.

[soft zealot]

this is quite fun 🙂

[strange mist]

nice experiments. I think you've really figured out revision 😄
I've played aronund with it a bit of course. but don't really know what to do with it.

[soft zealot]

be nicer if I had a faster card lol

Taking about 5 minutes per final image on my 1080ti
That's 20 steps,20 steps,30 steps,20 steps, an upscale and multiple switches between 2 different models

I may be adding in something to allow you to siwtch/mixnmatch between inputted and prompted/random images 🙂

strictly speaking as both these are manually selected I could have bypassed the first 2 ksamplers, but I didnt

Of course the ultimate would be to take 2 different ReVision feeds and use them as inputs toa 3rd ReVision feed to prodcue a Recursive ReVision lol

[strange mist]

I like how detailed and versatile your workflows are, but still too complex for me 😄
revisionception!!!
what's up with your text prompts, btw? placeholders? can't you specify some weight to it as well? 🤔

[soft zealot]

the Text prompts on the RH side are used in the 2 Ksamplers in what Ive called the Pre Feeder Zone

Inside ReVision the Positive isn't used and is simply there to avoid issues bit is zeroed out. The Negative in the middle ReVision section is used as par tof the ReVision process

[vital ermine]

Just chillin' with my bro

[uncut fiber]

sorry i am very behind in comfyUI. Do you think this can be proper result for img2img from Mona to Lennon?
Or i have it terribly wrong?

[upbeat summit]

bottom left is pretty good .)

[strange mist]

I see. so the pre-feeding part is for generating input images to fed into the revision process?

[uncut fiber]

his shoulder is front part of mona 🙂 @upbeat summit

[soft zealot]

yup but because Im lazy Im only prompting for one of them manually and then using a Random Prompt pull from Lexica (WAS Nodes Suite) for the second image

[indigo carbon]

I think I got it more seamless now.. plus still with AIT

[soft zealot]

still looks like the ReVision process with added AITness 🙂

[strange mist]

got it. complex and creative. and of course pretty cool 🤩

[soft zealot]

Just adding a note and then I'll post an image with the workflow in if anyone else want to play with it

[eternal fog]

I tried to neaten my workflow a bit, nowhere near as neat as your guys though

[strange mist]

if it's just what is visible in the img, I'd love to play with it 👀

[indigo carbon]

I feel like I'm somewhat close to having a seamless img+img workflow, but if I'd get to a point where AIT isn't compatible, I'll have to make new components

[strange mist]

hey, I used your previous version of this workflow. wasn't sure what were you trying to achieve with using revision, ip adapter and control net together 😅
I probably missed the posts where you were discussing it

[eternal fog]

You can achieve whatever you want to achieve lol

I mostly use it to just chuck stuff together and see what happens

IPAdapter is better than revision for copying 1 subject

And I'm crap at text prompting so in comes Revision

[soft zealot]

Oh it is, but I wanted to add the note first , patience my young Padawan 🙂

[eternal fog]

I'd like to use AIT, but I run out of VRAM when I use it with my current setup

[strange mist]

I get what revision is, I think. and I've used control net before.
not sure about the ip adapter part.

[indigo carbon]

what GPU do you have?

[eternal fog]

3080 10GB

[indigo carbon]

you need latest NVIDIA drivers, torch2.0.1+cu118, xformers==0.0.20

[eternal fog]

Pretty sure I have all that except xformers.

It works fine until I use IPAdapter and then it runs out of VRAM

[indigo carbon]

IPAdapter doesn't seem to be compatible; but CLIP vision is

[eternal fog]

Yeah, it's not a massive problem anyway, the new updates comfy has done to memory management have sped it up quite a bit anyway

[soft zealot]

Ok here you go @strange mist

Workflow in attached image , open in browser , download, save yadda yadda yadda

Also included the little man with a beer picture as a bonus 🙂

[eternal fog]

Any Canadians?

[indigo carbon]

plus IPAdapter doesn't blend images like it "overlays" them. I think I managed to blend images with AIT using CLIP_vision for now

[eternal fog]

I'm not using it to directly "Blend" images so it's fine. I'm mostly just messing around with it anyway.

[glad grove]

yes my qween,i love the kanada and everything it represents ( 🤠 )

[indigo carbon]

yeah, this is already miles ahead of MJ's blend thingy anyways

[strange mist]

thank you so much 🙇‍♂️
I'd have a proper go at it tomorrow.

[eternal fog]

The blending seems a little hit or miss for me, seems highly dependant on the images you give it.

[indigo carbon]

with the last prototype I made it seems to almost always blend the 2 images.. don't worry; I won't release it until I'm as confident with it as I was with the normal AIT workflow

[eternal fog]

Is it just from the weights your changing to get it to do that

[indigo carbon]

not only that, also changing the CLIP's flow to understand image blending even better

doing that is crucial to get it to even work on AIT

[eternal fog]

When you say the CLIPs flow, what do you mean. Have you made custom modules? Or do you just mean how you are plugging nodes into each other.

[indigo carbon]

I had to compile AIT modules for CLIP_vision, I also meant the nodes, yeah

[eternal fog]

Do you not have the nodes like "Load Clip Vision" -> "CLIP Vision Encode" -> "Unclip Conditioning"

And you feed each image into it's own Clip Vision Endode and then join the Unclip Conditioning together

[indigo carbon]

yes, after the conditioning is zero'ed out it does that, except the refiner's negative prompt.

it also needs a module for CLIP vision though

[eternal fog]

Yeah I zero out the conditioning, I also have a slider to average the Zero and a Text Prompt, in case you want to use a text prompt as well, but not have it use the full weight as text prompts tend to override Revision quite easily.

[soft zealot]

thats exactly how the ReVision process uses it as well, Zeroed out. I mean yo ucan renabale it (I have in option in my main workflow) but it gets a but wonky

[strong field]

Yeah it do! Just not with AIT 😦

Also, your clip vision “ hacks “ are clever

[indigo carbon]

normal CNET-ish stuff will come on its own, for now I think I'm close to mastering img+img

[strong field]

These pups are hilarious

[indigo carbon]

after more beta testing and after I'll be as confident with this as I was with the normal AIT workflow, I'll release it. and that's a promise

[strong field]

How much are you messing with weights on average?

[indigo carbon]

it depends, I like to start with 1:1 then see which gets more dominant, then changing accordingly

[indigo carbon]

in this case, it was 1.15 imageB, 1.0 imageA

and it figured out to put the dog on the motorcycle on its own

[strong field]

Interesting my revisions are around 1.0 but IPAs around .3

[soft zealot]

Just an FYI I've found out it Red Walls idf you use Vanilla SDXL Base & Refiner models most likely because I didnt use the SDXL specific conditioners in the areas highlighted.

Any other SDXL models seem to work fine though

[turbid osprey]

hi

would kissing boys be gay if i did it?

[glad grove]

kiss my pfp

[strange mist]

thanks for the heads up. I'm testing revision with base sdxl 👍

[soft zealot]

depends on your gender identification surely?

[soft zealot]

I haven t used vanilla SDXL for a while now so dint ebven think about it. Then again all these are primarily built for me , I just leave them open for others tomake use of but don't actively shove them up on Civitai etc 🙂

[turbid osprey]

i guess it depends then

sometimes ye sometimes nay

[strange mist]

you should definitely put them up on civitai 😮
maybe add a detailed walkthrough with pictures

[soft zealot]

I did with my original one and I got pissed off with being repeatedly asked the same questions even though the answers were in the credits & notes. 🙂

[strong field]

Lol

[soft zealot]

which were also on th edownload page

[eternal fog]

imagine expecting people to read

[strange mist]

if I were you, I'd put them up with a faq and ignore all comments 😛

[vital ermine]

https://www.youtube.com/watch?v=UyoXmHS-KGc

[eternal fog]

It seems to be quite easy to get the Anime in front of a more realistic image with IPAdapter

[indigo carbon]

I find it rather impressive that it made from a fire man and a pirate ship a fire pirate, so creative

[strong field]

No prompt?

[native knot]

I recognize that pirate ship.

😉

[indigo carbon]

nope, as I mentioned; the only part of the img+img version of the AIT workflow that uses text is the refiner's negative for better clearance

[strong field]

Impressive

[upbeat summit]

the workflow looks nice!

[glad grove]

real

[slender coral]

These are amazing, do you share prompts?

[native knot]

Just trying to bring all the concepts together...

[upbeat summit]

interesting

[upbeat summit]

nice!

[glad grove]

[native knot]

A bit better in the first one with the eyes and the color consistency for the character.

But this was an interesting pose.

[sterile yoke]

ok i have another computer question
https://www.dell.com/en-us/shop/gaming-laptops-pcs-and-accessories/g15-gaming-laptop/spd/g-series-15-5530-laptop/usegbts5530gjpx
i found this computer which has an Nvidia GeForce RTX 4060 with 8 gb GDDR6
is GDDR6 the same as vram? and if not, what's the difference?

and is this sufficient to run sdxl

[glad grove]

128bit bus

[native knot]

Yes, the GDDR6 is the VRAM. VRAM is essentially just video RAM; aka RAM dedicated to the video processing. GDDR is just Graphics DDR.

Now, is it sufficient to run SDXL? ...that kinda depends on how patient you are. It'll run, but frankly unless you're dropping nice money on really high-end graphics in a laptop, it's almost never the equivalent to the desktop version with the same name. On top of that, the 4060 is kinda shit on the desktop...as was pointed out, it's only got a 128-bit data bus as opposed to the 192-bit of say a 4070. That means that it's got a limitation for the amount of memory data that it can transfer at any given time. So...could it do it? Probably, but how patient are you?

[sterile yoke]

how long do you think it'd take

[native knot]

I wouldn't want to speculate on that and it would depend on what you're doing.

[sterile yoke]

are we talking like a minute? 5 minutes? an hour?

[glad grove]

id say from 5mins to 10mins

[sterile yoke]

hmm

[native knot]

Yeah...in that range.

But...

You could easily overwhelm it.

[sterile yoke]

maybe i should just say feh never mind and not get a new computer

[native knot]

Let's put it this way...I'd rather keep my 3080 than get the 4060 in a laptop.

[glad grove]

with $1k you could buy an used 3090 24gb and a pair of 64gb ram and you would be settled for years