✨｜sdxl

[hardy cipher]

esrgan battle

[indigo carbon]

Or will ComfyUI just use AIT on UNET?

[upbeat summit]

accelerated upscale_models? yes please

[visual glade]

it's going to be AIT only on unet

[hardy cipher]

damoniano being a spicy boi. wow

[visual glade]

I'm not planning to do ESRGAN so feel free to do that

it's not a priority because it's probably never going to break if you make an accelerated ESRGAN node

because I never touch that code

[indigo carbon]

I might, because that's the part that takes the most time in the AIT workflow I made. Also, will the official implementation support IPA and CNET?

[hardy cipher]

could anyone explain why I randomally get all black output images when using ipadaptor?

[visual glade]

probably not IPAdapter but controlnet should work

[hardy cipher]

or it's seemingly random. ran the same setup with different seed numbers and some were all black

[visual glade]

it's going to "work" with ipadapter by switching to the pytorch implementation

[soft zealot]

Who defines what the IImportant Nodes are.

What makes one Node more Improtant than another?

How do you prioritise Nodes?

There's an argument that Child Nodes should follow the PArent Node rather than the PArent follow the child.

Thi is the joys of opensource & git hub

This is the way 🙂

[indigo carbon]

I am not sure what do you mean by IImportant Nodes, but that wasn't an argument. Me and Comfy discussed about moving forward

[soft zealot]

I was generalising not being specific

you used the phrase "Important nodes" originally 🙂

its a general issue with anything opensource/github not specific to CXOmfy or A1111 etc

[hoary saddle]

ok, i'm done, sleepy and losing my mind 🙂

[sour obsidian]

relatable

[hoary saddle]

ok now i'm done, gn cruel world

[peak dove]

Here is another 1960's LoRA

FWIW I am impressed by Ideogram.ai - perhaps the best response for AI text yet! https://ideogram.ai/u/Torcello

[ionic dragon]

ideogram?

[dense chasm]

today is the first day of school in China,the first grade primary school students have to learn English and mandarin meanwhile,it's a big step forward,a couple of years ago,english teaching begin at only junior high school

[heady vale]

Im enjoying the Classipeint lora. Some nice results

[shy kelp]

If I check the Google Drive checkbox, will it download to my Google storage? does it mean i don't have to redownloading models every time I open Colab?

[hardy cipher]

well it doesn't exactly download your storage. but it'll have access to it

[dense chasm]

checkbox in colab is only an running procedure option, downloading model or not it depends on the code,most of time it will download to your current working folder,if you want to reuse the model,you should download the model and mount on google driver

comfyUI github python code supports not only mount on google drive but also comfy updates option

[ionic gulch]

what's your favorite upscaling method for sdxl at the moment?

[indigo carbon]

ESRGAN

[dense chasm]

ultrasharp 4X is good for me ,anyway actually SAI's API upscaling only have two options:RealESRGANUpscale and LatentUpscalerUpscale
i've changed to ultrasharp 4X since sdxl can be run on comfyUI

[indigo carbon]

Ultrasharp_4x was made 3 years ago. there is an upscaler in the upscale DB that was released about a month ago, it's actually incredible

[peak dove]

Just d/loaded! 🙂

[peak dove]

Topaz GigaPixel

[spring fulcrum]

what upscaler is that... the one that was released a month ago? What is the link for that database?

[vale eagle]

I am very interesting on GigaGan upscale but I don't know how to utilize it.

[indigo carbon]

https://github.com/Phhofm/models/raw/main/4xLSDIRplus/4xLSDIRplus.pth

this upscaler is CRAZY

easily my personal favorite

[ionic gulch]

database: https://openmodeldb.info/

do you use a special workflow for the upscaling? or just basic upscaling with the upscale image node?

[dense chasm]

yep,upscale nodes requires more VRAM it's a trade-off

[indigo carbon]

I made two workflows I use all the time. one is the normal AIT workflow, the second is the image+image version I made 12 hours ago

[spring fulcrum]

too bad there isnt something we can use to turn these images into vectors then upscaling isnt that important.

[dense chasm]

i've tested,any pre-training model out there? GAN and diffusion are two mainstream vision computing generative AI technics

controlnet is another big issue,huggingface released some small models already,optimized on RESNET and unet

[spring fulcrum]

Can you post an image with your two latest workflows? I'd like to check them out.

[dense chasm]

SD run slower than GAN main reason is noise and denoise iteration.GAN speed up with different routine,generator and discriminator,but sometimes model collapse issue

[indigo carbon]

both are on my civitAI. the normal AIT one is actually the official one for AIT atm.
https://civitai.com/user/tdg8uu/models

[dense chasm]

GAN and SD both requires huge training datasets and too many parameters already

[fringe barn]

Do we have ControlNet yet for SDXL in A1111?
I'm a bit out of the loop. I know it's working in Comfy

[dense chasm]

i've run diffusers controlnet model on google colab,have to pay attention to canny algorithm thr opencv library of python and others,such as depth or whatever,and GPU vram run highly

when i run on any webUI or GUI,it takes much more time,so i give up when you have not enough computing power

[spring fulcrum]

I keep getting this when generating an image?

[vital ermine]

roll back to two days ago

[uncut fiber]

How is called that black glass in helmet? Visor?

[uncut fiber]

o.k. got it helmet shield

[stone fossil]

Try welding mask.

[uncut fiber]

@stone fossil she actualy was welding 🙂 Thank you!

[vital ermine]

well, I actually got a dreambooth to work on XL

[uncut fiber]

would you guess it is only 768x768?

[vital ermine]

I am not sure what it picked up on

[uncut fiber]

reminds China house because roof

[glossy dune]

LoRA style?

[vital ermine]

model, not a lora. A bit strange

[glossy dune]

oh alright, trained on the style or just zombie woman ?

[vital ermine]

all the before are the same model

[indigo carbon]

that's an issue with newer version of ComfyUI, the first commit from august 28th is the last one that works with AIT

[vital ermine]

my segal lora with my style model

[indigo carbon]

@spring fulcrum merge that commit with the one you currently have and replace all conflicts; that's the current fix

[glossy dune]

I'm preparing a data set to train a style LoRA, do you think training on SDXL is better now then 1.5?

[spring fulcrum]

yes

[vital ermine]

well, it is dang harder for sure

lora is lora but making a full model is a pita

[spring fulcrum]

can you even train a full scale model on a 4090?

[vital ermine]

I just did

[glossy dune]

harder on what aspect? hardware or settings?

[spring fulcrum]

How do you do that? I have a 4090 and wanna make a model

[vital ermine]

It is a style model like this

[spring fulcrum]

I do believe I seen Jesus! Praise the Lort! (in my best southern baptist accent)

[vital ermine]

LOL

I need prompts to see what this has done cause, so far, I like the style but that is not any style I used for my training images

[glossy dune]

speaking of that lmao where do you guys go to get good prompts, used to use lexica but the generations there are not as good as they used to be

[vital ermine]

I still use it but lexica is a bit suck now 😦

[shy kelp]

i only have 4gb vram, how can i optimize comfy ui?

[spring fulcrum]

I use prompts from civitai images I like then change them to suit my liking.. I also use ChatGPT to help me get creative sometimes

[glossy dune]

gpt4? and any plugins?

or just gpt3.5

[spring fulcrum]

use --lowvram as a command line arguement in your .bat file

[indigo carbon]

I just have 2 sections when I use ComfyUI, first; I generate some cool images, second; I blend the ones I like best, it's an endless cycle

[spring fulcrum]

I dont use plugins for that directly in ComfyUI... I just go to the civitai site and look at pictures I like and with chatgpt I just explain to it the format of the promt I am looking for and then ask it to produce image ideas in that format for Ideas I feed it... Or i even ask it for suggestions

[glossy dune]

I see ! thanks for sharing

[glossy dune]

@vital ermine I think this much better then lexica atm xd

[spring fulcrum]

It works well for me

[uncut fiber]

bubbles are awful

i mean in my image 😄

[vital ermine]

I tried to get the DB extension for auto1111 to work, but no dice

[indigo carbon]

I mixed an image of a cyborg, and an image of a dog.. what did I expect?

[vital ermine]

I submitted the ticket as the file he creates his own program thinks it is pickled and to use --disable-safe-unpickle YET it is already being used.

[vital ermine]

d8hazard went nuts once on his users for reporting bugs then he came back so not sure if it will be answered or not.

His is different than Kohya

[shy kelp]

how to solve this i keep getting 'pause' and while render, and when continue it disconnect

[stone fossil]

Hmm looks like a dupe ive seen before. 😉

[stone fossil]

Meow.

[vital ermine]

Even has a sig, lol

[indigo carbon]

Oh LOL sorry, I just loaded my workflow and the seed was the same

[stone fossil]

Yep if u drop one it will gen same seed.

I was just playing SD XL memory. 🙂

[vital ermine]

I find it amazing on what it picks up on

[soft zealot]

there was an unexpected element of surrealism in that one 🙂

[uncut fiber]

reminds me this

[soft zealot]

that reminds me of pictures I've seen of an inland sea in Russia I think it is (somewhere along the route of the Road of Bones) thats dried up and left a load of ships landbound

The Aral Sea

[uncut fiber]

yes in south afrika there is something similar. Cemetery of ships. <- for sure i misspelled something.

[soft zealot]

the only misspelling I see (no pun intended ) probably isn't the one you're thinking of (and I cant really comment with my typing!!

(afrika rather than africa)

and that depends on your mother tongue

[uncut fiber]

o.k. now i see it. 🙂
what do you think about consistency. His wearing isnt very detailed described i think it can be used somehow in animation.

Yes we say Afrika, Germany maybe as well. But yes for sure in English Africa.

[soft zealot]

some minor differences but nothing I would call major

[uncut fiber]

yes i know. 2nd and 3rd got not that spike in armor, as well knee protection. And last got more vivid colors 🙂

[smoky patrol]

Love the volume and the background on that one.

[soft zealot]

You find yourself adrift in the vast expanse of the ocean, surrounded by nothing but endless waves. A radiant sunset paints the skies in hues of orange and purple, casting a warm glow over the tranquil waters. In the distance, a school of playful dolphins leaps from the water, their graceful bodies glistening under the sunlight. As you navigate the choppy waters, a lone seagull swoops down, guiding you towards a distant island silhouette on the horizon. The image captures the serenity and hope amidst the challenge of being lost at sea.

nope I see no dolphins lol

[thorny frost]

they are underwater

btw i am experimenting with IP-sadapter and get this erro: missing clip vision: ['vision_model.embeddings.position_ids']

the thing seems to work though.

am i mising something or it is OK?

[soft zealot]

I was red walling originally with something similar and then I switched to the Dev branch

https://github.com/laksjdjf/IPAdapter-ComfyUI/tree/dev

[thorny frost]

dev branch always!

[peak dove]

IP Adaptor?

[vital ermine]

Nope, that is my model I just retrained.

Civit prompts have too many loras and stuff in them

[thorny frost]

image 1 + image 2 + "dragon" = image 3

[vital ermine]

I think this one came out very well

[thorny frost]

not exactly what i was thinking but pretty funny

[vital ermine]

Oh, man

what kills me is when it blips out 100% and realism strikes. I trained on artstyles so ugh

[soft zealot]

Same 2 input images + the word "dragon" through my ipadaptor loop

[thorny frost]

1+2+"huge fire dragon" = 3

[thorny frost]

this IP thing really works

[uncut fiber]

has somebody did some short focused only on this workflow?

[soft zealot]

because I'm perverse. I've used image senders/receivers to send the output back to the input (after the first gen) and then started a fresh gen withthe "auto queue prompt" option enabled

Going down the rabbit hole lol

(temp setup ., just playing)

[uncut fiber]

Looking complicated enough 😄

[soft zealot]

its a self perpatuating loop lol

[uncut fiber]

He looks realy odd 😄

[indigo carbon]

I released this if anyone missed it

[uncut fiber]

going for it!

[soft zealot]

the next generation

Wonder where it will end up if I leave it for some time lol

will it just converge?

[vital ermine]

with and without my lora

with and without and I believe this was a success

[uncut fiber]

clip vision encoder i will find in manager, models?

[soft zealot]

4 levels deep & counting lol

[vital ermine]

2nd one is using a different model

[strong field]

I was doing this yesterday and created a nice clipspace loop

[zinc cargo]

this has no info in it

[strong field]

here was one of my favs

@soft zealot

[soft zealot]

now 6 levels deep from the original 2 images 🙂

think I'll stop there

[soft zealot]

noice

thisis my favourite so far

[uncut fiber]

@indigo carbon i cant use it getting some recursive error and as well mentioned AIT...

[uncut fiber]

maybe not enough memory.

[indigo carbon]

it's on CivitAI

[uncut fiber]

i know where was problem... 4 batches, now i need it consume less memory. Where to put in Comfui, medvram or lowvram please?

[uncut fiber]

some kitch 🙂

[ionic gulch]

new "2x Upscale - v1.0" workflow with upscaler (basic solution based on @high skiff - will be improved in the future) and option to turn off parts of the workflow (for example the upscaler) by cutting wires in the menu section - also has improved seed option and only generates two non-upscaled versions instead of four.
https://imgur.com/tltvAth.jpg
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

[high skiff]

I have a new gen high res fix 2.0 workflow for my 1.1 release of my SDXL workflow, but I am just kinda burned out. I can't seem to get it reliable, and it shows promise of being way better than anything I have seen from anybody else, and its fast, but the fuckery I use to get it to work introduces some severe issues I cannot figure out how to fix

[hoary saddle]

Dunno what that is, just used controlnet in ComfyUI

[ionic gulch]

yes, i've seen you talking about it and was wondering when you will release the new version. but already guessed there are some problems as it didn't happen. hope you can solve the issues somehow.

[high skiff]

if I could just figure out how to fix the latent issue I have with my current 2nd gen high res fix workflow, I am confident it would be one of the single highest quality, most reliable, easiest to use, and most effecient ways to do high res upscaling

like here

base image vs old high res fix vs new high res fix

As you can see, the detail and texture/realism on the face in the V2 version is incredible, but it murders all of the background information, and I am not sure why.

and I have spent hours and hours trying to engineer some fancy work around, and I just can't

I mean, its obvious that the core concept works exceptionally good

[ionic gulch]

from my own tests i think we will need different upscalers for different art styles and realism. i've tried to build an own solution for a while, but evertime i have something working well, it won't work for different styles.

so i went for your 1.0 upscaler for now, as it at least improves some of the images and doesn't take too long.

[vale eagle]

I think you might try to seperate the main character and the background to get the job done?

[steel mist]

FWIW this would still be useful af. Can always do a composite in photoshop etc

[high skiff]

the V2 is kinda faster, and improves texture monumentally in a lot of images, but sadly, it also adds texture where its not wanted

[high skiff]

no easy/reliable/fast weay to do that, especially with stock only nodes unfortunately

[vale eagle]

stock only nodes is the limitation. It should able to done with Groundingdino by prompt

[zinc cargo]

can i have a link?

[ionic gulch]

i would die if i only had the stock nodes. especially the menu would look so much worse without my own nodes

if they want stock nodes, they have to provide more options imo.

[high skiff]

All of the stuff I have done so far has been with stocky only, as thats the requirement to be on the offical wiki

and I have been able to still produce some very high quality results, but I am hitting a wall that I am not even sure custom nodes can fix

[ionic gulch]

a basic workflow may be possible, but if you want to do ip adpater, revision, switching between txt2img and img2img, etc. you will also hit a wall soon - not only because of the special nodes needed for those things, but also for the math and logical elements, like switches, to make those features optional.

[high skiff]

tons of people adapt my baseline work for stuff like that. My core workflow is made fo rthe masses that want high quality images with no extra work, and then people adapt it for higher tier use

[soft zealot]

FWIW Ive switched to using a very simple upscale that I'm happy with (not with that model though lol)

[thorny frost]

https://www.youtube.com/watch?v=30b2k1p2CiE

[glad lily]

Hey guys. I don't know if it's only my problem or not. Latest version of ComfyUI showing this error message, when I'm using controlnet models. Few weeks ago everything worked nice. Do anyone have same errors?

[thorny frost]

posable fingers!

[tribal lantern]

anyone here has a basic comfy workflow with controlnet that supports regional prompting? Like controlnet for entire image, but part follows prompt a, the rest prompt b, and there is an overarching background prompt. i just can't get that to work.

[ionic gulch]

i'm basically using @high skiff solution. it looks quite similar. will have to check for the details, if i can improve it with your solution and settings:
https://i.imgur.com/8uxWGmU.png

[soft zealot]

theres a reason its similar lol

[high skiff]

the current solution I have found for my high res fix V2, which monumentally improves textures and detail retention, also seems to be the cause of its one biggest downfall, which will likely make most people hate it, so I am considering releasing it as an advanced branch. But for now I am just kinda depressed about all the workflow stuff. I was hoping things would go a little bit better with it all, but that does not seem to be the case so far

[strong field]

You using AIT? that or outdated custom nodes

[soft zealot]

@ionic gulch essentially all I've done is simplify it by using a sstandard rather than an advanced sampler do I can just select a denoise value insread of doing match and dropped the contrast fix

[strong field]

does the problem boil down to getting unwanted texture everywhere, not only with subject?

[high skiff]

basically, yes

[strong field]

What about setting a latent noise mask prior to the sample portion to keep texture only where you want it

[high skiff]

sometimes its good, sometimes its not

[strong field]

Manual mask for testing, but if you want convenience i think you'd have to do custom node with SAM detector, or SEGS

[high skiff]

that will never be reliable enough to put out for the masses in any meaningful way unfortunately

I am sure I could get it working by manually masking and doing 2 gens then compositing, but almost nobody will ever wanna do that, and its also not likely to look very good

[ionic gulch]

maybe you can do some kind of automatic background recognition and compose them automatically.

[strong field]

this right here is the real solution, but not with stock nodes 😛

i already have 3 or 4 ideas for how to do it but not with stock lol

[high skiff]

That would be good, but again, it will never be reliable with our current technology/tools

It is unfortunate, because the method does show huge promise, especially in images that have high levels of texture all over, they really do bring out detail to another level compared to everything else I have seen.

I just wish I knew how to properly manipulate what I'm doing to get it to work better

I thought about maybe using depth map blending, but that's horrifically unreliable, and yielded terrible results

A friend proposed an idea using the cany filter and then fill tooling in the silhouettes of things, but that also doesn't sound like it would work very well, and it would be a bitch and a half to get working

[glad lily]

I've tried it with standard ComfyUI nodes too, without AIT. The same error. Both 1.5 & XL

[strong field]

can you explain what you are doing?

@high skiff

[strong field]

This looks like an AIT specific error.

[high skiff]

Not in a reliable or consistent way, not really.

I have done hundreds, if not thousands of variations on it, and I never got to a result that was consistently good enough to ship, therefore I have tons of different values that all work in quite drastically different ways, so I couldn't really explain it in a way that would allow for proper consideration across all of the potential outcomes

Once you do find the good values, it's a set and forget sort of thing, you never have to worry about it again, but the results will only be as good as the values that you choose permanently, and I've been trying to search through hundred/thousands of variations of these settings in different setups to try and tweak ever so slightly to get it closer to what I'm looking for, but I can never really get it there

[strong field]

basically image > upscale> downscale > resample > combine images?

[high skiff]

The current settings I have right now work phenomenal for things like watercolor painting, oil painting, and high frequency realism, where there's a lot of background detail

But it really starts to fall apart on more simple subjects, where it can introduce some beautiful textures to finer details, but also tends to chew up smooth/softer shapes and details

[high skiff]

If I really were to kind of explain the general idea of what's happening it's, image, upscale, downscale, first sample, pass, second noise pass, one step fractional offset diffusion, into a second sampler, combining two separate latents together with a form of mixed math, then running it through a third sampler, to then output it into a VAE encoder, and then multiplying the image by itself to make up for the lost contrast and fidelity

But again, there's still a lot of variations in that process that I have tried, and that one is not the best or the worst

[ionic gulch]

i was also trying to do the upscaling at 80%, so the last 20% would already use the higher resolution, with the guide of the first 80% (so less double objects). it worked great for some images, but after changing style it broke.

[high skiff]

In the end, even with all of those steps, it's the same speed or faster than my 1.0 upscale, which means you can get a 2048x2048 image on a 3090 in like 15 seconds?

[vale eagle]

You might introduce multiple workflow for different type of generation.

[high skiff]

It's very fast, and it goes even faster if you use AIT, however that's not exactly reliable in any capacity right now, especially not for lower dedication general usage

[high skiff]

Unfortunately, if I were to do that, then it would really only be available for a small range of styles, effectively being styles that have a ton of detail in them and absolutely no smooth surfaces, because there's no combination of settings that makes it Good for anything outside of highly textured and detailed images

The one example where this new upscale blows any other that I've seen out of the water is specifically watercolor paintings, it does an incredibly good job of upscaling and properly texturing watercolor paintings

[ionic gulch]

you could include different upscalers and make them switchable by reconnecting a single wire. that should be possible with stock nodes.

[high skiff]

That's kind of already what I'm doing, but again, you would be choosing between V1 for 95% of images, and V2 for a very narrow select few images that would benefit from it, and the average end user does not like that sort of thing

And the problem is, while V1 does look good on its own, compared to what V2 can do in some cases, it just makes V1 look terrible, which really sucks because you're forced to use V1 most of the time

In isolation, V1 looks fine, because you don't have the context of how incredible V2 can look at times

But the second you see V2 and realize how much better the potential is, V1 looks horrible

I know it sounds like I'm talking it up a lot, but I genuinely do mean it when I say it makes V1 look like dog shit when it works properly lol

left is original base image, middle is V2, right is V2

as you can see, the amount of texture and fine contrast it can bring out in water color is truly incredible

and even in photo realistic images such as this one

Here as well (tho you can start to see the issues with the over texturing in the background blur)

as well as these two

and in all of these cases, its no slower than the normal V1 version

@ionic gulch@vale eagle@strong field

Just so you guys can see what it is capable of in a vew different cases

[glad lily]

Yeah. You were right. I've disabled AIT nodes in manager and it start working. Thx!

[strong field]

the watercolor is my fav use case

[high skiff]

yeah, it really just works incredible for water color

the more texture something has/is supposed to have, the better it works

[strong field]

Lots of work going on behind the scenes with AIT, there is a plan to get it into Comfy main branch, but its a "soon" type of thing

[vale eagle]

I wonder if negative prompt could reduce the damage on background.

[strong field]

interesting

[high skiff]

I have tested that, and it has helped basically none

in fact, the damage seems to be less without a prompt at all sometimes

big asterisk after sometimes

I am testing it with a realism finetune right now to see if maybe a finetune helps it

Actually

with this finetune... 1.0 seems to look much better than with base

[vale eagle]

If you would able to use controlnet, depth map might help

[high skiff]

Already tried, it makes things way worse cause depth map is extreemly unreliable

has a huge propensity to make eyes super sunk in

in which case, if I used it as a mix mask, they would lose detail

[vale eagle]

how about after the current output, use depth map, invert the depth map and resample the background? I don't know if it would work or not. Just throw some idea.

[high skiff]

interesting

so this finetune works amazing on animals (tigers is what I tested)

and it works like dogshit on humans

tho, this finetune with my realism LoRA looks incredible, I really must say

like, damn

too bad when you upscale it it looks like a wet fart lol

[thorny frost]

tiny horse

[thorny frost]

solved

[zinc cargo]

@ionic gulch comfymanager couldng find these in your latest from civit...

oh.. having WAS problems i guess

[willow bane]

hlo

[dusty jewel]

yo

you know that ufospotted outside sandiego calif?

it was a transparentsphere, with a solid dark metalcube inside

try to make that in sdxl

icouldn't :/
neither could the others here

[willow bane]

please anyone help me out

[peak dove]

Some Underwater LoRA/1960's LoRA mix, ComfyUI

[outer ether]

I have a question about lora training, it seems like the general consensus is to NOT train the text encoder for sdxl, but I've not seen anything describing what the consequences of not training TE are. does this mean that sdxl loras can't be trained on new concepts and that you shouldn't caption images?

[hardy cipher]

where did you hear that?

[outer ether]

here and on many parts of the internet

kohya_ss even says it in their webui

[hardy cipher]

ahh. well just curious. not really an expert. I've just never heard that before. maybe that's a recent development or I misunderstood people or something.

I trained quite a few things in 1.5 and rarely captioned, but that's because I was lazy

[outer ether]

ah no problem, yeah it seems like it bashes some of the text portions of the generation, and in my testing it does seem to make a difference not training the text encoder

[hardy cipher]

I need to learn up on that stuff

[outer ether]

I just want to know the why, being told "just don't do it" is not enough haha

I would imagine captions would still be used in the unet training I just don't know what I'm losing from 1.x-2.x models by not training the text encoder

[hardy cipher]

yeah. I've been diving deeper into a lot of things. but it's like pandora's box

[outer ether]

I found a post here about it but it's still not clear to me

[outer ether]

it's too much fun

[vale eagle]

Not train the text encoder is true when you are try to train something the text encoder already know. If you are training something completely new to the text encoder, you still set the training rate of text encoder and train it.

[hardy cipher]

yeah, realize it's not exactly the same thing. but yeah, not sure how not training the text encoder would even work? I played with it a little bit in 1.5, but not sure what method you guys are using now

[outer ether]

interesting so how would the images be captioned in that case? do you still include captions for concepts it may know?

[vale eagle]

No captions also work in that type of training. Like some style training.

You might ask in #🔧｜finetune channel to get more information.

[hardy cipher]

one thing I have seen people say here is that you should name your concepts differently with xl. which doesn't fully make sense to me yet

[vale eagle]

I certainly confirm there are people tried freeze text encoder and get great result.

[hardy cipher]

like if you trained on a specific person in 1.5 you'd name it something the model didn't recognize. but apparently with xl you shouldn't do that? I'd like to find a guide that lays it all out and explains it

[strong field]

those are absolutely awful

in the best way possible

[vale eagle]

There is a training method to find the closest celebrity which already learned by text encoder and provide images and caption to train the unet close to your image using celebrity token

[strong field]

This method works exceptionally well

[hardy cipher]

taylor z'dar swift to steven z'dar seagal. the evolution everyone was waiting to see

[hardy cipher]

I tried something similar in 1.5 and whatever I did really didn't work out well for me. but I realize xl is a differnt game

[vale eagle]

It is becuase xl already learned a lot of concept

and provide decent output

[hardy cipher]

yeah. it's very interesting stuff. after the evolution of the tech in the last year, I'm very curious to see where it goes in the next 1-2 years

[strong field]

sdxl was released at the mid/end of july, its a baby

incredible to think about that

[hardy cipher]

1.5 is a baby, lol. barely a year old

[strong field]

thats an eternity in AIT time

[stone fossil]

I have a dream. 😄

[hardy cipher]

yeah. it's crazy though. I don't know how much I buy into the technologicla singularity concept in the way the media and openai hyped it, but seems like we're heading towards something. and it's going to get weird

[strong field]

100% weird, something we couldnt calculate, something different. AGI, SGAI, its going to get weird. ChatGPT 4 is Einstein level of aptitude, thats just not something any of us understand yet

What it means ill leave to the dogmatics and the simps

[strong field]

love these, fantatic work as always

[hardy cipher]

gpt-4 is ridiculous. it's just unfortunate it's being ran by openai. I very quickly soured to their way of doing things

[stone fossil]

At your service. 🙂

[strong field]

live long enough to see yourself become the villain.

[hardy cipher]

it's getting better now. but the overarching guardrails the put on it, and the memory limitations

[hardy cipher]

well sam altman seems like a scumbag with a good PR team

[strong field]

definitely my opinion of the guy, there's a recent interview, hes on stage, big crowd, someone asks, we want to do what OpenAi does, how do we do it?
His response is basically, "dont try, nobody can compete with us now"

[hardy cipher]

argued for the dire need for regulations on ai in the US. then went to the EU and argued the opposite a couple days later. because those regulations wouldn't benefit him

[strong field]

eh, ill just link, oopshttps://www.youtube.com/shorts/xHVsk7d1L-0

[hardy cipher]

I'm not going to discount their contributions to the progress of AI

[strong field]

certainly not, but doesnt exclude them from valid criticisms

[hardy cipher]

but now they're trying to monopolize it

and snuff out other innovation

[strong field]

just their name is a PsyOp

[stone fossil]

Type all your emotions into SD and make a image instead.

[strong field]

like i said, ill leave that to the people that care, im just along for the ride now

[hardy cipher]

I wonder what they have in the works. because seems to me the huge monolithic beast models might get passed up in a sense

they'll serve a purpose

[hardy cipher]

but all these streamlined open source models that work much more efficiently. just seems like an ecosystem will evolve with different specialized models working in unison. I mean, we're already doing that now

[strong field]

theres a reason Meta and Microsoft on both sides of the fence

[stone fossil]

Now thats a start.

[steel mist]

I am 100% on board with the coming ai apocalypse – “things are about to get weird” but GPT4 is nowhere near that level of intelligence

[hardy cipher]

I have to give it to meta for their approach. I know they always have ulterior motives. but they have been releasing an insane number of open source models on huggingface

[steel mist]

If they were ahead of open AI they would not be open sourcing though

[hardy cipher]

it is but it isn't in my experience. it's like an extreme savant

[steel mist]

It is very good and it is good at things I would not expect it to be good at. Like, it has an inkling of a theory of mind. Can do basic spatial reasoning?

But I do not think that if you removed all concepts of general relativity from it’s training set it would independently be capable of inferring GR

[hardy cipher]

it can do some spatial reasoning stuff. but it's very bad at some things. and it has no idea that it's bad at them

[steel mist]

Nor could it explain the ultraviolet catastrophe, contribute to quantum mechanics, etc

[hardy cipher]

I don't think it can infer many independent things outside of what's it's specifically trained on

I've really tried to get it to

[steel mist]

So like “it’s Einstein and the singularity is 2-4 years away” is just a big [citation needed] for me

It can’t even tell me what I should get for lunch grrrr

[hardy cipher]

openai pushed a lot of that "tech singularity! be terrified!" stuff

because brings them attention and helps with potential regulations

that'd lock out competition

[steel mist]

Yeah the incumbent always loves the idea of a frozen status quo heh

[hardy cipher]

literally trying to snuff out open source models

the one that perplexes me the most is bard

[strong field]

ok

[hardy cipher]

it's literally gotten way worse over the last few months

it just lies about anything and everything.

[hardy cipher]

or "hallucinates"

not sure what google is doing there

and they seemed to zap it's personality or whatever you'd call that. used to come up with interesting creative things at least. but now it's so over restricted

[strong field]

bruh, did i tell you you are one of my fav creators yet? creativity abounds

[hardy cipher]

it trips me out how real this one looks

[hardy cipher]

yeah, thaevil coming out of nowhere

[stone fossil]

Lolz thanks mate I just have a keyboard and a mouse and 2 cats. 🙂

[hardy cipher]

or somewhere

[stone fossil]

Thanks lol it means a lot to me.

[stone fossil]

I am not a social media person I just keep things private and giggle al them.

[strong field]

samsies

[hardy cipher]

I have facebook but I ignore it other than talking to a couple people

[strong field]

dont even care to argue anymore, 1+1 = 5? Ok, you right

[hardy cipher]

same

[stone fossil]

Would you be more happy if I did not strip meta data? xD

[hardy cipher]

realized a couple years ago just how futile it was. I got nothing good from it and it was aggravating me. so disengaged

[hardy cipher]

tsk tsk

if my images don't have metadata it's because I made them on dreambooth due to not being at my pc

[strong field]

ay, i respect it either way, with IPA and Revision im not sure the text encoder matters as much

[hardy cipher]

or dreamstudio?

[strong field]

Hey but if you want to DM me some prompts i wont be mad

[hardy cipher]

anyway, like 2 cents an image, lol

beans

[strong field]

this funny AF

[hardy cipher]

had a whole long prompt about an existential bean nightmare

[hardy cipher]

no love, no hope. no dreams, just beans

no need to sleep, don't need to eat, but the kids ravenously eat the beans because there's nothing else. but they never feel full

trying to fill an unfillable void

[hardy cipher]

some of my best work

[stone fossil]

Id buy it.

❤️

[shy kelp]

what model do you guys use?

[strong field]

all of them, haha,

this one is on the top 5

[hardy cipher]

I need to update mine. I'm probably using outdated versions

[stone fossil]

I like turtles.

[hardy cipher]

this ipadaptor just decides to output black images for identical configurations of nodes and parameters. I'm going to assume no one knows what causes this, but it's disappointing

[strong field]

are you using clip vision with it?

or just the model inputs/output

[shy kelp]

what does ipadapter do?

[zinc cargo]

what workflow you use there?

[stone fossil]

420

[strong field]

the right time

[strong field]

This is what it does, taking top LHS as input and then it will ReImagine it using clipspace, conditioning, whatever you choose.

I use a custom workflow that I am not willing to share

BUT your saving grace is, i made this other workflow to show the TLDR of it

[strong field]

as for what my custom workflow looks like, well,

[strong field]

i dont know why but this one is funny to me

[stone fossil]

😛

[strong field]

wow some of these were atrocious, but the bottom LHS is ok

[zinc cargo]

does the sd15 indicates that it's a 1.5 model? and to do SDXL gonna need to get a different one?

[strong field]

that is correct, the models are downloadable from comfy manager

[cyan crown]

15 hours to train a Lora with my face.........

for sdxl it takes 30 mins....

btw results are great!

[strong field]

im laughing over here

Fashion of the future

[strong field]

https://tenor.com/view/lmao-laughing-hysterically-gif-26398146

[strong field]

Hey now...

[strong field]

what is with the cute wolves idk

but im here for it

[stone fossil]

🙂

More madness.

[strong field]

lol wtaf

[zinc cargo]

kohya?

[strong field]

holy... this vibe was sick

[strong field]

this my girl

[stone fossil]

Shes all urs.

[hardy cipher]

@strong field I'm using the CLIPVision model. Is that incorrect?

[strong field]

can you show me a screenshot?

you will need clip vision model to run it, but you dont have to connect conditioning.

If you do, then yuou need to adjust your conditioning weights as well as your IPA weights

[zinc cargo]

with @strong field tldr workflow and some lora

[strong field]

nice!

[zinc cargo]

when ever you're feel like sharing the moe elaborate one, or some bits from it...

🙂

[strong field]

I can tell you what its using, Revision + IPA, IPA + Clip Vision Conditioning, Revision Only, IPA Clip Vision Only, IPA only, in no specific order, all hooked up to their own sampler, and then i just vary the settings according to how much influence i want

[stone fossil]

Lets make good stuff, chop chop.

😛

[zinc cargo]

https://tenor.com/view/national-spaghetti-day-national-spaghetti-day-hungry-gif-7317183

[strong field]

I built it for experimenting and comparing all the revision/ipa methods and fell in love with the reinterpretation machine

[zinc cargo]

i love it as well

i do wish to learn how to bring even more details to the pics though.
i think latenet upscales is what i'm looking for?

[strong field]

you want them to add detail? or just be higher res?

[zinc cargo]

add details

[strong field]

You can add some details via prompt, or lora

[zinc cargo]

running wowifier and prompting for more details usually

[strong field]

p.s. let me know if you want any of these in higher res

im literally just messing around

[stone fossil]

I can upscale them trough the roof if needed.

The pics shared here are already degraded and meta deta stripped.

[strong field]

well i mean more than a screenshot of the gens im getting from your inputs, but i gues syou can screenshot, and resample lol

[hardy cipher]

Not on my computer right now. But I'll take a screenshot when I'm able to

[stone fossil]

Just jammin. 🙂

[ionic dragon]

inpainting dropped?

[upbeat summit]

I mean ComfyUI added support for it... so...

[ionic dragon]

left(inpainted)

[echo temple]

I am trying to use the ip-adapter in comfy with inpainting. It works fine in normal text-to-image but for inpainting I am getting very weird results. Is there anything different when using it for inpainting? For instance, I want to change the face of an image conditioning with a specific face in the ip-adapter, what would be the best way to achieve this?

[strong field]

Set latent noise mask, I paint with a separate prompt and conditioning from ipa

[echo temple]

Thanks for the answer @strong field I'll take a look! Would you have an example?

[glad grove]

bro thinks he batman 😭

[echo temple]

Oh I found this https://github.com/comfyanonymous/ComfyUI/discussions/639 @strong field

[stone fossil]

Sour chips are nasty good.

[strong field]

Not at desk at the moment but I can ping you when I get back

[echo temple]

Thanks! I highly appreciated

[strong field]

I see you getting that detail

[stone fossil]

https://www.youtube.com/watch?v=6W5pq4bIzIw&list=PLaWR0UtAgqduBbazYOcekhD0ghnob4D_C&index=5

[noble shoal]

5 Fingers and see through. Now i am exhausted.

[upbeat summit]

nice work!

[noble shoal]

Thanks 😅

[zinc cargo]

moar!

how should one go for trying to create real life versions of cartoons?
canny? I2t? img2img?

[indigo carbon]

hey, I'm getting this weird error: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! what does this mean?

[ionic gulch]

Could you solve your problem? (I guess you mean github - my workflow i not on civit)

[zinc cargo]

i'll admit i gave up later and didnt persue it 🙂

[vale eagle]

This means your workflow was using different device. Restart comfyui and try again?

[ionic gulch]

have you used my workflows in the past? there shouldn't be a change for WAS nodes with the latest one.
i would love to reduce the needed third party nodes, but WAS is the only one if found, that provides image history and that many save options
really wish the core nodes had more of those basic features.

[zinc cargo]

iirc you helped me run it on my other pc, but i had to remove all other custom nodes for it...

[ionic gulch]

is that only a problem for me or for everyone? why doesn't it get fixed?

[zinc cargo]

happens to me too.
zoom out fix it.

[zinc cargo]

my WAS says ineed to update llvmlite...
how do i do that?

[noble shoal]

Can't you just change it in the requirements.txt and run "......\python_embeded\python.exe -s -m pip install -r requirements.txt" in the was folder?

[zinc cargo]

no idea 🙂 but i'm going to try it

[zinc cargo]

well, llvm is not here, should i just add it as a line?

or perhpas comment numba out

[noble shoal]

Give it a try. llvmlite==14.0.0. Commenting will probably have no effect since its already installed

[ionic gulch]

no idea. never had to manually update anything in comfy until now. auto update through the manager did everything for me since i've moved to the standard - not standalone - version and installing python in user space, so it could update without admin rights.

changing the requirements.txt to your version may help if the new version isn't really needed, but that's just a "hotfix". would be better to really update that component or find which of your other nodes messes with the versions, if that is the case.

[glad fulcrum]

Control net still not working on auto1111? 😦

[zinc cargo]

there are no other nodes, it's a clean and updated version.
just your workflow and what ever the required nodes were.

the hotfix did not work

[ionic gulch]

and you installed everything through comfy manager? or manually?

[zinc cargo]

manager

since my other pc runs it without a problem, i think i'll try to over run it

[ionic gulch]

did the same. the only thing i've installed manually is the manager.

[zinc cargo]

yea, you got to start with manually adding the manger i guess 🙂

[ionic gulch]

try to remove WAS nodes and add them again maybe. so it will install the requirements again. look at the console when you do the reinstall and check if it gives you any errors that could help to solve the problem

[zinc cargo]

on it

[ionic gulch]

was nodes is a very common addon. so more people should have that problem.

[zinc cargo]

so good

[zinc cargo]

no luck with that...

[plain marten]

Is there any downside to using a 4x upscaler when you only need 2x upscaling? (I've been downscaling by hand but 16 times is too much for what I need)

seems most upscalers are 4x so just wondering

[noble shoal]

I love that he is watching a cucumber TV show 🤣

[noble shoal]

By hand? Just use a "Upscale Image By" Node after the 4x Upscale and set it to scale_by = 0.5

[plain marten]

Well by hand yeah but it's not the downscaling I'm worried about 😆

using the Ultimate SD Upscale node on ComfyUI I can set "upscale_by" to 2 with ESRGAN_4x , but the result is a bit odd looking, with weird horizontal lines and stuff

it's not outright terrible, but with the added noise I can't say that's it's really better than just a bicubic upscale

[noble shoal]

There are 2x Models on https://openmodeldb.info .

I would upscale by the factor the model was intended to upscale by

[plain marten]

Fair enough, I've looked at the 2x models, there just isn't nearly as many as 4x which is why I'm wondering

My settings might very well just be fucked tbh, you guys use the base model as an input, right?

[noble shoal]

In this case i am out. i use sytan's method of upscaling.

[plain marten]

Hmm I think it's the ESRGAN model that doesn't behave, some NMKD model actually upscales properly (even though at 2x I should probably reduce denoise)

[void python]

Hey Guys! I'm new to stable diffusion. Is there any way that I can generate images of different camera angles but with the same background?

[uncut fiber]

@void python i think not

[void python]

If I provide a specific background in the prompt every time I change the angle, will it work then?

[uncut fiber]

i think it wont work, in best case you will have very similar results, but not identical. Maybe there can be way with controlnet, but i think not.

[plain marten]

You're looking to keep the same background (looking down the river) but have the model face another direction?

[void python]

Yes

[void python]

Oh

[plain marten]

If you want a coherent output, that's gonna be a lot of work

[void python]

How should I start then? I'm relatively new, I don't even know which models are good to start with

I've only tried SDXL

[plain marten]

I think futzing about with the prompt and see if you can get something that looks vaguely similar is your best bet, if you REALLY wanted to do that I don't think anything short of rendering the background alone, and then rendering a 3D model with matching perspective on a depthmap for a controlnet would work

truth is from my experience anything that doesn't involve a complete regeneration of an image just doesn't work aswell as you'd hope (inpainting/outpainting and such) but I'm sure a controlnet wizard could sort something out

[void python]

I'll try that out, thanks!

[plain marten]

good luck

[uncut fiber]

best resize algorithm

[glad grove]

sexy

[strong field]

Now I know I need to see a remake of fatboy slim, weapon of choice remade with seagal

[hardy cipher]

not quite weapon of choice, but you know, still classy

[crisp owl]

Hey if Austin Powers can destroy the fembots with his dancing, it is a completely viable weapon

[hardy cipher]

I already had these images on hand just in case someone wanted to see Steven Seagal dancing

[upbeat summit]

as soon as AI movies are a thing, you know what to do

[hardy cipher]

make poignant movies about Steven Seagal's journey to become a professional dancer

[glad grove]

fifty shades of grey with steven seagal as main girl

[upbeat summit]

whatever floats your boat, buddy 😄

[glad grove]

thicc hands

[hardy cipher]

mr hands

[glad grove]

the messiah

[steel mist]

looks like a recast of the righteous gemstones

[vital ermine]

Looks a bit like david Carradine

[hardy cipher]

David Carradine eating a large pile of raw meat in 1978 classic movie RAW MEAT EATER

[hardy cipher]

current workflow

[glad grove]

game of the year (all the years)

[glad grove]

he grumpy

[glad grove]

bro has the vegeta hairline

[eternal fog]

lmao

[eternal fog]

lmao

cursed image

[heady vale]

green screen revealed

[glad grove]

this is the perfect male body

[eternal fog]

definitely

[glad grove]

ranes the goat

[glad grove]

the prince of seagals

[glad grove]

he kinda sus

[glad grove]

massive chungus

[hardy cipher]

interesting eyewear there

[crisp owl]

Haha well that message gave me a chuckle

[vital ermine]

hehehe, I am taking prompts from civit

not sure why XL decided to slap me with the nake fat chick, but alright

[hardy cipher]

have you taken any from the degenerate section?

[vital ermine]

no, I have a NSFW filter on

[hardy cipher]

ahh, smart

[vital ermine]

WELL

[crisp owl]

Hellraiser remake

[heady vale]

or a Hellboy

[strong field]

Here is a quick example of how this would work.

[heady vale]

Image blending is so cool

[pure crystal]

how strong is my magic? fuhghetaboutit

[strong field]

I’ll say

[strong field]

Really like the colors here

[vital ermine]

Did comfy add this pop up when you type in the nodes with embedding choices? I disabled my extensions so not sure how to disable that?

[wet nacelle]

@strong field

[willow bane]

hlo

anyone knows how to install gradio

i am having error in infinite zoom extension regarding that

[cedar tusk]

I think pip install gradio should work?

[vital ermine]

https://civitai.com/models/138833

[hardy cipher]

You shouldn't move a model when it is dispatched on multiple devices.
You shouldn't move a model when it is dispatched on multiple devices.
You shouldn't move a model when it is dispatched on multiple devices.
You shouldn't move a model when it is dispatched on multiple devices.
You shouldn't move a model when it is dispatched on multiple devices.
You shouldn't move a model when it is dispatched on multiple devices.

oh, okay, thanks for that

[glad grove]

You shouldn't move a model when it is dispatched on multiple devices though

[hardy cipher]

sound advice

I mean, I'm not though. but something surely thinks I am

or maybe these are more cautionary warnings. and it's just really adamant about me knowing

[vital ermine]

little donnie

[hardy cipher]

he likes pears, there's no denying that

[glad grove]

bro has them polydactil hands 🤣

speakin of sexy characters, homer simpson

[hardy cipher]

cursed

we heard you liked hands

these lads are causing a ruckus

that wheelchair has a pretty unconventional wheel there

[crisp owl]

Sport version

[indigo carbon]

I'm scared

[willow bane]

hlo

how should i update this

i tried pip install --upgrade pip but it shows it is upto date

[west breach]

I wouldn' worry too much about that warning

A1111 uses it's own python environment

[vital ermine]

yep, done the proper way.

I was surprised at comfyui not using a virtual environment.

[uncut fiber]

@willow bane solution python.exe -m pip install upgrade pip
When in folder script, didnt work?
Also i think as well it isnt important

[willow bane]

why it is saying it is not able to find cuda cores i have rtx 3060

i can fix it by writing but will it limit something?

[uncut fiber]

I think you will end deleting venv and then let it recreate it again.
For this is better to ask in #🤝｜tech-support

[ionic gulch]

inpainting fully implemented in my workflow:

[snow spade]

/xena

[clever verge]

SDXL Impaiting models coming? https://github.com/comfyanonymous/ComfyUI/commit/7931ff0fd95c1842b0c8e7f5cc3a2ce5d3b88b3b

[steady grove]

difusers team released one

https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1/resolve/main/unet/diffusion_pytorch_model.fp16.safetensors

[clever verge]

Ah, 17 hours ago. 👍

[steady grove]

yeah its fresh

[zinc cargo]

is there a simple workflow for it?

[ionic gulch]

Version 3.0 of my workflow is available on Github now.

Features:

• TXT2IMG
• IMG2IMG
• Control Net Canny
• Control Net Depth
• Inpainting
• IP Adapter
• Revision
• Optional Upscaler

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

[zinc cargo]

i'm getting OOM with a 4090 on your workflow 😦

[vital ermine]

Hell, I just had a crash then another crash of a plugin due to a mem leak I guess. Not his workflow but python does have a leak I was told.

[mental lion]

Hello everyone,
where can I find a SDXL controlnet tile Preprocessor safetensor?
Is it not out yet ? I only have the SD15

[zinc cargo]

i'm not blaiming JPS of course, just venting my frustration bcz i can't wait to use his workflow 🙂

[vital ermine]

Oh, no worries as this is a mem leak while my 4090 hasn't had that happen today. :/

[zinc cargo]

having to settle for my own mess 😛

[ionic gulch]

try cutting the upscalers. they need the most vram. working fine on my 4090

[hardy cipher]

big dawg 4090

[vital ermine]

if you both have a 4090 then why would there be OOM for one and not the other?

[zinc cargo]

btw, @icy brook your Lora rocks on warp 🙂

[ionic gulch]

maybe different nvidia drivers (i'm using the old ones) or too much other open applications.

[zinc cargo]

no idea

[vital ermine]

Yeah, I rolled back to 531.79 myself

I get slightly faster speeds but more OOM

[ionic gulch]

maybe you don't use the fp16 versions of the models?

[zinc cargo]

this?

[ionic gulch]

no the sdxl models

[vital ermine]

wait, is there a fp16 version of base and refiner or is the 6.4GB THE fp16 version?

[indigo carbon]

I think it's the pruned....?

might be fp16 though

[ionic gulch]

i'm using:

[vital ermine]

I can't remember as I had them from the first day

[ionic gulch]

UNET:

[autumn scarab]

painterly output from sdxl... possible? everything looks so instragrammy for me

[vital ermine]

I'd kill to even have 500mb shaved off the base

[ionic gulch]

Clip Vision:

[indigo carbon]

also; It MIGHT be possible to stack TRT and AIT on each other. if that gets figured out, its safe to say that instant image generation is VERY close

[ionic gulch]

VAE:

[autumn scarab]

even getting the main subject to be off-center would be awesome

[icy brook]

Fun stuff!

[vital ermine]

See, I want faster training speeds myself as gen speeds are alright.

[dense chasm]

Imao

[ionic gulch]

your ip adapter looks twice the size as mine

[zinc cargo]

https://tenor.com/view/wink-wink-agnes-agatha-harkness-kathryn-hahn-wandavision-gif-22927975

[zinc cargo]

I just used the manager to download it, but can i have a link to yours?

[ionic gulch]

oh, i think we compared different files. you need two files for ip adapter. one is actually your size.

[zinc cargo]

which one is the 2nd one?

[ionic gulch]

one is in ComfyUI\models\clip_vision the other one is in a subfolder of the ip adapter custom node

[zinc cargo]

that's the one i've got

we're good on that department

[ionic gulch]

https://huggingface.co/h94/IP-Adapter/tree/main/sdxl_models
https://huggingface.co/h94/IP-Adapter/blob/main/models/image_encoder/pytorch_model.bin

how much vram is in use before you start comfyui?

~2GB on my system

[glad grove]

0.7 gb

[zinc cargo]

running only discord and chrome for these tries

[ionic gulch]

my workflow doesn't go above 16gb for me with the example.jpg - even with upscaler, but i've seen 21-22gb before with larger images.

[bronze arch]

i wonder why lvmin zhang hasnt made official sdxl controlnets yet.. the community ones are kinda bad

[vital ermine]

He swore off SAI in an open letter due to the shit, his words, TE/Clip 2.x used. I am glad we had what we finally had for 2.1.

[bronze arch]

wait what

but hes still forcing on fooocus with sdxl

[ionic gulch]

@zinc cargo maybe try "python.exe ComfyUI\main.py --disable-smart-memory" when launching comfy. the new smart memory management of comfy gave me problems before.

[vital ermine]

yep, was an open letter about it and he called them out heavily.

[bronze arch]

where

[vital ermine]

That is why it took forever for someone to make some of them for 2.1

[vital ermine]

That was a long while back so I have no idea. Check his Twitter feed.

He did say if they stuck with the same TE (clip) he wouldn't do it so that may be why. Community only with official only for 1.5

[bronze arch]

cant find anything of the sort, but he works on a webui for sdxl

[zinc cargo]

new error

[bronze arch]

so i dont see why he wouldnt work on a controlnet

[vital ermine]

No idea but he was fucking pissed at SAI over it. Maybe 6 months ago now so things change.

We who did 2.1 was sad then we got a community few of them but the really good ones we never got

[zinc cargo]

@ionic gulch you mind if i'll messege you instead of using public channel for support?

just realized what i'm doing here

[ionic gulch]

sure. i've just tried with the new smart memory management. it uses up to 22gb of vram, but no oom for me.

[hardy cipher]

lol. man, I need to get me one of those

just didn't have this stuff in mind when I got my computer

[ionic gulch]

lol. the cat turned green:

[vital ermine]

Anyone know how to get XL to do a cyclops?

[hardy cipher]

hmm. challenge accepted

[ionic gulch]

second one was better - but still some artefacts on the mask border:

[vital ermine]

Yeah, I can't even get an interrogator to work

[nimble heart]

just like in paint an eye

[vital ermine]

it should know how to do it tbh

[nimble heart]

Probably

[vital ermine]

Can't get it to do that and even 1.5-SDXL interrogators mess it up

[nimble heart]

you could spend a few hours fiddling with prompts to avoid having two eyes or you could just in paint the eye after the initial gen

[vital ermine]

LOL, I would have to fight the damn thing each time I gen that way.