✨｜sdxl

[steady grove]

But months ago he decided he knew better

[vocal stream]

either way, it sucks for me that Comfy is on its way to become the standard because I have a shittier experience with it

[crisp owl]

Then use swarm

or wait

[vocal stream]

I wish - not demand, just wish - that something like a111 was funded instead

[hardy cipher]

lol

dude, fund it then

[crisp owl]

this is all so early

that's the pain points of early tech

[vocal stream]

such a childish response

[hardy cipher]

I do my best to not complain about things I don't pay for

I beieve that's childish

[steady grove]

Voldy def has deals going on. Hosts like civit might throw cash his way to build the community up

[hardy cipher]

I hope they all make money, a1111 people included

[steady grove]

Be the change you want to see. It's not a childish idea at all. It's something children often can't grasp

Open source software really relies on people doing it themselves

[vocal stream]

Sorry, you think it's realistic for me to hire a dev to work full time just because I'd have prefered if a different thing was done

should I make my own Samsung when I prefer for them to not take the sd card away, come on

[hardy cipher]

it's an extremely lucrative industry that is going to explode with time. so hoping it does trickle down to the developers of these UIs

all of them

[steady grove]

Python coding is far from android engineering

[hardy cipher]

lol

[azure oxide]

is the base of what youre upset about that comfy has a payroll so now he's constantly updating his software 24/7 because he's being paid to do so and a1 isn't doing the same? i dont think this has any merit if so

[hardy cipher]

it's pretty high level

I'm starting to learn python just so I can make nodes that don't seem to be available

[vocal stream]

the base of what I am 'upset' about is that Comfy is likely to become the standard in part due to a push by Stability and that I don;t like to work with Comfy

that's it

[hardy cipher]

I could have come in here and complained those nodes don't exist

[steady grove]

Comfy made node API so simple an nice

[azure oxide]

compare comfy's activity on comfyui: https://github.com/comfyanonymous/ComfyUI/activity?ref=master
with a1:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/activity?ref=dev

they're both active. they're both good devs. one being hired on by SAI does not change anything for you

[hardy cipher]

man, I have to say, he really thought of everything

surely some glitches and stuff still

[steady grove]

It's just not stability's issue that voldy hasnt built auto effectively

[hardy cipher]

but the fact that it's even put together as well as it is

[hardy cipher]

confusing perspectives

[azure oxide]

voldy is actually SDnext, a1's main dev is actually a1

[hardy cipher]

you know who's a real hoot? thelastben

[azure oxide]

if youre referring to something else tho, disregard

[hardy cipher]

colab a1111, lol

[vocal stream]

whichever one becomes the standard will get the extensions first so that changes things for me. Now seperately you can argue that funding devs for a project doesnt mean more work is done on that project but that seems either naive or too pessimistic

[steady grove]

When I resize a reroute node then connect to it, comfyui resizes it back to default. Completely unusable..what am I even paying for????

[autumn forum]

Can we all just agree that development takes time, sdxl is still fairly new and it’s gunna take time to get all the qol features we need for sdxl in comfy and a1111. Ai is moving faster than we can develope software for it😂

[hardy cipher]

we live in a strange society

[autumn forum]

That we do..

[hardy cipher]

could you explain how this works? I guess I don't understand very well

[soft zealot]

those of you that aren't happy with A1111 not properly supporting SDXL whilst at teh same tme moaning about ComfyUI would be better off IMHO just getting a MJ subscription

[vocal stream]

when you hire people to spend more time to work on something, more time is spent on that thing on average

[vocal stream]

this take is even sillier. How does getting a MJ subscription make it easier for me to use things like deforum or ControlNet

[hardy cipher]

you do realize comfy isn't releasing most of the extensions and what not right?

neither is a1111

they're largely 3rd party

[vocal stream]

devs are more likely to release extensions on top of whatever UIs are most popular, typically on the most popular first/only

[hardy cipher]

they take care of core implementation

[vocal stream]

during 1.5 era a ton of things got released only for a111 and maybe later for some other ones with less hours spent on it because users were on a111

[soft zealot]

wooohooo give that poster a gold star for noticijng the abusridty ingherent in the statement lol

[hardy cipher]

still lots of things only on a1111

[crisp owl]

OK, well we all get it. You're upset something you like isn't working anymore for you.

Adapt, or stick with it.

[hardy cipher]

my favorite extension ever still stuck on a1111. but I don't complain too much about it

I've actually looked into replicating it but I'm not smart enough, lol

[soft zealot]

ComfyUI didnt exist back then , It only launched back in May or June of this year.

[compact hatch]

which one if you don't mind sharing?

[vocal stream]

I mean same, and it's because a111 was the most popular. It'll just be annoying for me if the next deforum or something else I end up liking is stuck on Comfy next time around. that's all

[hardy cipher]

well here's the thing, tenoke, with swarmui you can use comfy on the backend and their a1111-esque front end if you're not into that. you can also use the front end with comfy workflows. so essentially you could find a workflow that uses whatever tool you want, load into swarm, and then you never really have to look at it. maybe that's not ideal for you, but lots of moving parts at this point. so kind of hard to have it all clean and streamlined for everyone

[steady grove]

This is just an irrational fomo

Visual scripting is what creative work needs. Node editors are huge for this field.

That's why most studio software has some kind of node graph

[vale eagle]

A111 vs Comfyui conversation likely happen everyday. Why don't people just use whatever they love and do not need to care about another tools.

[vocal stream]

sure a better UI built on top of Comfy (whether swarm or something else) is my main hope for comfort if Comfy becomes the standard

[steady grove]

Comfy was always made to be a back end. There are a half dozen different front ends for it already

[upbeat summit]

nice https://github.com/comfyanonymous/ComfyUI/commit/76d53c4622fc06372975ed2a43ad345935b8a551

[uncut fiber]

@polar epoch

[vale eagle]

Wow, does it for ip adapter?

[hardy cipher]

more LoRAs? or do you think 3 is enough?

[soft zealot]

yeah, lets just ignore all the tools ;o)

[crisp owl]

Forgive my naivity, what is clipvision? Seen the node, haven't seen what it's for

[upbeat summit]

I currently I use up to 3 loras, but my old TI embed workflow went up to 8 :D. you can always chain more together right?

[upbeat summit]

Sorry, I don't know.

[soft zealot]

as far as I can tell its for loading standalon clip models

[hardy cipher]

well yeah. I'm just wondering if I should make a deluxe 10 LoRA version

for the big dawgs

[crisp owl]

Curious under what circumstances that would be utilized

[soft zealot]

well now theres a question..........

[upbeat summit]

afaik clipvision models are used for style transfers

[crisp owl]

As in?

[upbeat summit]

add a specific model with a specific style to your image gen

I think it is available for a1111 for some time but I haven't tried it

[crisp owl]

ah gotcha, sounds like something I'll have to do some more reading up on

[soft zealot]

so many questions

Where do the Clips come from

where do style models come from

[crisp owl]

https://tenor.com/view/chris-farley-no-idea-giddy-excited-nope-gif-5165531

Tried following that node path myself at one point and the undefined selections was as far as I got lol
"I'll look at that another day"

[soft zealot]

there is no way this could be made to work. Is there??

[crisp owl]

No clue, personally don't even know what models goes where to change those undefined selections

[vale eagle]

I don't know if it is related. https://github.com/tencent-ailab/IP-Adapter

[soft zealot]

https://blenderneko.github.io/ComfyUI-docs/Core Nodes/Loaders/LoadCLIPVision/

[hardy cipher]

is that jimmy carter?

[wet nacelle]

Jimmy Flobber and his Meatball Cavern

[upbeat summit]

there's this video by SECourses showing it as a ControlNet addon in a1111: https://www.youtube.com/watch?v=tXaQAkOgezQ

[queen adder]

Are more repeats the way the model learns more detail in individual image? My model so far still doesn't understand textures properly and im using only 5 repeats.

[hardy cipher]

how old is the comment you're responding to? lol

I remember reading that days ago

[wet nacelle]

I like this

[azure oxide]

ive went back months before when i needed information on very specific things lmao

[wet nacelle]

https://tenor.com/view/onnit-rogan-kettlebells-gif-5729101

[queen adder]

i just checked my computer today and that was the first message on my screen lol

[crisp owl]

hmmm, I do recall doing that once or twice in A1111 with style transfer now that I see that. Will have to see about that in Comfy now sometime

[steel mist]

Ok these daily checkpoint releases are killing me. Comcast is gonna be knocking my door down

6 gigs here, 6 gigs there, pretty soon your talking about a lot of data

[crisp owl]

I won't say how my bandwidth looks like then 😅

[steel mist]

Copax has gone from 1.0 to 4.2 in 2 weeks

[hardy cipher]

lol

[steel mist]

Pretty sure I’ve snagged every one of those

[steady grove]

Delivering bits luckily costs Comcast nothing

But they have you convinced it's a huge load

[steel mist]

And yet, they have data caps everywhere

[hardy cipher]

doesn't mean comcast will let you just do it, lol

[steady grove]

Corporate propaganda wins again

[steel mist]

Also there’s no marginal costs but building out the system does cost capital

[hardy cipher]

not so much propaganda as it is virtual monopolies for a lot of these companies

[steel mist]

And if everyone downloads 2-3 6 gig models nightly, they will run out of bandwidth at some point

[steady grove]

Another gooder. Listerine was house cleaner before they convinced everyone their breath is nasty all the time without it

[hardy cipher]

your choices are xfinity or dsl that's 5 times faster than dialup

[steady grove]

So now we all swish with house cleaner

[steel mist]

I can get an ATT fiber line but it’s slower than xfinity I believe

[crisp owl]

never used mouth wash nastiness

[wet nacelle]

https://tenor.com/view/jerma-dance-dance-monkey-jerma985-crazy-dance-gif-26252145

[steel mist]

I keep yelling at my local fiber ISP to PLEASE expand to my area, but they don’t want to trench neighborhood lines if att already serves

[steady grove]

It does seem unnatural but propaganda sells it

[hardy cipher]

I lived in a deadzone for a while that only had at&t dsl. I could barely watch youtube videos, lol

[steel mist]

There is some really questionable swimwear happening in the background there lol

[steady grove]

Government here paid Telus to lay fiber. I live in the deep woods and have fiber. But Telus tries to convince me it costed them money to lay

[hardy cipher]

well of course

[crisp owl]

Yeah, red shirt at the beach? I'd never

[steady grove]

Corporations love to socialize their expenses

[hardy cipher]

privatized profits, socialized expenses

[hardy cipher]

https://civitai.com/models/128456/pe-shitty-fanart

this is brilliant

[cold mica]

We're live with ComfyUI Ep3! Join us at #1005545148211527800 and #1029055412764422214 🔴

https://discord.gg/JhaP82Sr?event=1141107696532209784

[cosmic compass]

my sd keeps telling me memory error when creating above 512x512 resolotuion, is it due to my 1060 gtx with 6gb or is there a way to allow me to create higher resolutions even with this low tier graphics card ?

[steady grove]

Don't think 6gb cuts it

[cosmic compass]

I can do 512 resolution with the model i downloaded, and someome here told me sdxl is also possible but i wonder what settings i can change to compensate low vram ?

[hardy cipher]

--lowvram

[craggy ibex]

It generated 2 people

how to make it only 1?

[uncut fiber]

cant see prompt, try 1boy or 1man

[craggy ibex]

how about 2 people for negative prompt?

[uncut fiber]

or make resolution to be tall like 512*768

[craggy ibex]

Hmm okay

[uncut fiber]

@craggy ibex i think it will not work.

[craggy ibex]

My friend sent me his oc picture

he want me to use an ai to turn him to different art style (or this stable diffusion)

[uncut fiber]

probably controlnet is beter for this, dont know how it works in comfui

[raw prairie]

Yeah thanks for reminding me 6gb is low in 2023

[craggy ibex]

Idek what is the best img2img workflow

[raw prairie]

Holy shit I remember when 4gb was a lot

[craggy ibex]

that could magically turn some characters to anime style like magically

Like that tiktok user...

Idk how..

[raw prairie]

Civitai has a lot of workflows

I didn't know i2i was a thing on comfy btw

[craggy ibex]

It's impossible for me to do this

[craggy ibex]

Not work

the prompy 1man not work

[uncut fiber]

so try taller layout.

[strong field]

Set resolution lower, then upscale when you have something you like. Put (1man,solo) at beginning of prompt

Also add descriptors for the background, SD will redraw prompts with higher CFG and empty space

[zinc cargo]

is there a preprocessor for softedge sdxl?

[craggy ibex]

How to upscale?

[hardy cipher]

lol

[uncut fiber]

@craggy ibex do you have A1111 still installed? Probably with controlnet and older models.

[craggy ibex]

Webui?

Yesh why not?

[uncut fiber]

you are working in A1111?
o.k. i forgot, what do you want to add remove?

[craggy ibex]

Currently I am using Comfyui

Idk it is A1111 or not?

[craggy ibex]

Can you fucking generate better?

for fucking sake.

Why am I always have skill issue in Stable Diffusion

I don't fucking get it

[stone fossil]

Anger is never the answer, young one.

[craggy ibex]

I have 0 help for this ngl

[craggy ibex]

Yeah I know

but I'm very jealous of someone who can do this

This is impossible

[craggy ibex]

How are both of these seem accurate?

I'm very jealous

[craggy ibex]

I need to find an answer

[vital ermine]

I like this training as it has that Ralph Bakshi 1970s look to them.

Welp, my masterpiece is finally done, lol

[craggy ibex]

I wish I can do something like yall do :(

[vital ermine]

Prompting is everything and negative prompting even more than that.

[stone fossil]

Yes.

[cosmic compass]

did that but only allows 512 , how much do u need for 2048x2048 ?

[zinc cargo]

is that your own model/lora? i love it

[hardy cipher]

I have 6gb of ram and I have no problem with larger resolutions. if you're using a1111 that might be the issue currently

[wet nacelle]

naw

[craggy ibex]

How to upscale?

[craggy ibex]

Fuck you

piece of shit.

😠

[vital ermine]

I like this one

[crisp owl]

I've been playing around with this style trying to get it right

[craggy ibex]

Such a disappointment

[craggy ibex]

Do I need to know of how to coding?

[crisp owl]

no

[craggy ibex]

in order to create perfect ai generated image?

[hardy cipher]

there is no perfect, my friend

[craggy ibex]

It's 3 AM in here

[crisp owl]

It's all in prompting and matching to a style of checkpoint you find fits your style preferences

[craggy ibex]

I can't sleep until it's not beautiful yet.

[wet nacelle]

https://tenor.com/view/what-tate-power-seriously-gif-15179198

[craggy ibex]

FUCK YOU

[autumn forum]

Load upscale with model node, load upscaler node, plug the two with image input and output, done

[wet nacelle]

FUCK YOU

[hardy cipher]

one day you'll be able to make beauty like me

[craggy ibex]

Idk what to do anymore

It's all failure

[autumn forum]

Looks like your doing image to image with high denoise values, lower denoise value or do what I said

[autumn forum]

I’m telling you what to do

[vital ermine]

Stretched beyond XL's limits

[craggy ibex]

I'll decrease to 0.8

[autumn forum]

No to much. Decrease it to 0.25 or 0.3

[craggy ibex]

Seem.. better??????

It could be more

[craggy ibex]

but it's impossible for A.I. to generate perfect hands

[vital ermine]

I like that art style, btw.

[autumn forum]

Okay just ignore me then whatever😂 no help for you

[crisp owl]

often it's a luck of the seed

[craggy ibex]

I'll try ig

tysm

[vital ermine]

Has this been Harry on ketamine?

[hardy cipher]

think of it like photography. if you're going for "perfection" there will be lots of fails before you get a win

[autumn forum]

Your welcome. Next time maybe take some help from others rather than say fuck you with a picture🤣 I gave you two options to upscale and got ignored lol

[craggy ibex]

Upscale like this?

[autumn forum]

No I would use a upscaler model with the upscale with model node

[craggy ibex]

Oh I think I have something from Searge 8k stuff

That might be useful

[autumn forum]

#✨｜sdxl message. It’s like this if this makes sense

[craggy ibex]

Alright

[terse pewter]

Guys
I made a video, using the A.I images of SDXL, it's a sad story, do you want to see it?
It's a 1 minute video

[crisp owl]

If I don't cry, I'll be disappointed

[craggy ibex]

Hope I will get angry issue after this

[terse pewter]

I think you will laugh instead 🤣

but let's see..i m uploading it here

It is so bad, that I am cringing right now. I need suggestions how to improve

Here is it guys ^

Need suggestions, how to improve

It's a sad story, that I tried making using A.I

[craggy ibex]

Like this?

[wet nacelle]

awful story

[crisp owl]

https://tenor.com/view/she-is-dead-passed-away-gone-persihed-crying-gif-15033484

[terse pewter]

How can I improve?

[crisp owl]

ask an llm to write a story, then make images based off each section

stick with an art style

[autumn forum]

Yes now drag the upscale model point to the left and add a upscale model loader node. Then if you have one they will show up, if you don’t look up modeldb in google and find one you like. And add it to the upscaler models folder in your comfyui directory. And it’ll show up.

[terse pewter]

i see, you mean chatgpt for a story, and images based off each section, but only 1 art style?

[crisp owl]

Yeah, that's one option

[autumn forum]

Yup! You got it. That’s upscaling!

[craggy ibex]

Okaaay and how do I run it hmm

Do I need sampler?

[autumn forum]

You’d hit queue like normal, and no no sampler.

[craggy ibex]

Alright cool

[autumn forum]

That would just do a basic upscale with upscale model, there’s also other upscalers made for more anime style such as that one ultra sharp is a good general one though

[craggy ibex]

Hmm...

What could that be?

[ionic dragon]

@tiny walrus gotcha

[craggy ibex]

Are there any prompt to prevent this?

[craggy ibex]

It's REALLY fucking annoying

[hardy cipher]

booru life

[hard fractal]

IMMA DO IT

IMMA CLICK IT

As soon as someone responds to this with an image of a ship

We'll release our ControlNets

[azure oxide]

july 18th

wait its aug 18th, thats some coincidence

[hard fractal]

weird eh?

waiting on the ship!

then it goes live!

[hard fractal]

O

M

G

https://huggingface.co/stabilityai/control-lora/

[sour obsidian]

Send it

[hard fractal]

It's ALIIIIIVVVVE

[sour obsidian]

[tiny walrus]

YESSS

[wicked frigate]

collect all 4! or 8 if you want both sizes!

[coral orchid]

I'm getting the big size, it's got the other size in it too

[azure oxide]

Revision is a novel approach of using images to prompt SDXL.
👀

[hardy cipher]

any plans on doing something like this with comfy? https://github.com/mcmonkeyprojects/sd-dynamic-thresholding

[sour obsidian]

Wait those aren't controlnets what could they be

[crisp owl]

https://tenor.com/view/holy-mother-forking-shirt-balls-eleanor-shellstrop-the-good-place-oh-my-god-omg-gif-10603389

[coral orchid]

The Good Place was such a fun show

[crisp owl]

haha loved it

[hard fractal]

Woo hoo!! https://www.reddit.com/r/StableDiffusion/comments/15uwomn/stability_releases_controlloras_efficient/?

[sour obsidian]

I reimplemented this as a node some while back for some testing. Could prob clean it up and push it out if thats fine with you @wicked frigate

[crisp owl]

@hard fractal

[wicked frigate]

i should probably redo it and release it myself. there's been feature updates since then

[hardy cipher]

it was my most favoritest a1111 extension

[wicked frigate]

which, uh, have been largely ignored cause they've been on auto lol

[hard fractal]

heh

oops

[stone fossil]

So will this help me improve my noodle ships? 😛

[sour obsidian]

All yours, mine wasnt the best implementation anyways haha

[wicked frigate]

@hardy cipher follow this git thread https://github.com/mcmonkeyprojects/sd-dynamic-thresholding/issues/54 i keep meaning to reply but forgetting cause busy and/or sick, will post there when it's up

[sour obsidian]

Give it a try 😀

[ionic dragon]

So there are only 4 CNs or are the others gonna drop in few days?

[ionic dragon]

I think it's too Gore for this server

[wet nacelle]

no

[ionic dragon]

Unless its generated through SD

[hardy cipher]

used it to make stuff like this with 1.5 in a1111

[hard fractal]

And now some minor qol nodes: https://github.com/Stability-AI/stability-ComfyUI-nodes

[sour obsidian]

No guarantee on others atm though these base ones are pretty nice. We have some in progress but atm we're looking to switch up our training system for em to work more direct with the control-lora format to try and improve the process a bit

[azure oxide]

edit: nevermind it works now

[hardy cipher]

sweet, thanks! I understand life takes priority over such things. but, man, I really had a lot of fun with that one

[ionic dragon]

So it might take few weeks?

[crisp owl]

up now, at least for me

edited 🙂

[azure oxide]

the second i posted that, it was fixed 😅

[sour obsidian]

Can't really put a timestamp, its a bit r&d atm. Could be much sooner for some

Im the meantime it would be great to hear how people like this new format

[stone fossil]

I will for sure cheers nice release, been checking out canny / depth / pose so far for XL glad to give this a spin and compsare. 🙂

[ionic dragon]

Ah, got it
Can we expect a lora which specializes in QR code generators?

[sour obsidian]

Excited to hear what you think 🙂

[azure oxide]

these 4 are the major ones for sure, i think the only other controlnet that i used often was openpose

[ionic dragon]

Openpose is og, it can get perfect hands

[sour obsidian]

Got it on the list, love those qr/luminance ones. Some out there prob have nice data for it already so a regular ol qr one might get trained out in the community first. Focus here is trying to improve the format first (control-lora) to make this all easier into the future

[hardy cipher]

man, the cfg and mimic cfg scheduling stuff really does some cool stuff when combined with noise offset and what not. still have no idea how mimic cfg really works beyond how it impacts render. but really great tool

[ionic dragon]

So you are basically working on the algo atm

[hardy cipher]

I've messed with qr codes a bit. only some of the squares are important. that's the key

or vital rather

[sour obsidian]

Yea so these new ones aren't "controlnets", they are basically full encoder weight diffs (loras for all weights) to allow the models own encoder to be its own controlnet so you dont need a seperate model

[stone fossil]

QR codes are fun but barcores some more usefull to me.

[ionic dragon]

It's too technical for me

[ionic dragon]

People scan QR codes but not barcodes

[sour obsidian]

For users it will be the same on the interface, just means smaller files and less vram required

[stone fossil]

Embed barcode all over a object almost not visible, scanning never fails, win win?

[ionic dragon]

Barcodes are only used by businesses

[visual glade]

they are controlnets just that the weight of the controlnets are stored as a kind of lora

[ionic dragon]

That's kinda better considering SDXL vram consumption

[visual glade]

yeah the main goal was reducing vram usage

[zinc cargo]

say, where does one get this CLIPVisionAsPooled node?

[sour obsidian]

"Im my own controlnet now"

I might make a little diagram actually of this

[hardy cipher]

what sort of vram requirements? I apologize if that's been stated already

[visual glade]

use this workflow

[hardy cipher]

I'm a bit deficient in that department

[ionic dragon]

@sour obsidian which text encoder does clipg use?
It doesn't use blip/clip?

[stone fossil]

It would be cool if you added a switch to select what model to load on what cuda device and then also select a checkbox that allowd to keep the model into VRAM.

[zinc cargo]

what's clip vision?

[sour obsidian]

Clip g is the text encoder 🙂

[ionic dragon]

Oh wait

I typed something

[visual glade]

https://huggingface.co/comfyanonymous/clip_vision_g/tree/main

put it in: comfyui/models/clip_vision

[zinc cargo]

https://tenor.com/view/rudy-kobansky-profitansky-power-respect-respect-bro-gif-15464776

[ionic dragon]

Were SDXL images trained on blip/clip or was it trained on completely diff captions which were long and a bit a more detailed?

[ionic dragon]

Why is it used?

It enhances the clip encoding?

[hardy cipher]

so is clip vision like unclip? or am I misunderstanding how it works?

[zinc cargo]

does comfy looks for the controlnet lora from the lora folder or the controlnet folder?

[ionic dragon]

Ultimately generating bit more relavent images?

[visual glade]

controlnet folder

it's used like a controlnet

[zinc cargo]

thank you

so these new ones make the canny the rest obselete?

[wicked frigate]

yeah the community canny/depth are basically instantly obsolete now lol

[hardy cipher]

lol

[sour obsidian]

XL was training on clip l + clip g text embeddings made from the captioned images

[hard fractal]

https://huggingface.co/stabilityai/control-lora/tree/main/comfy-control-LoRA-workflows

workflows there

[sour obsidian]

The new revision is something new. Like unclip but far more conceptual vs literal. We are still exploring it as well

[hardy cipher]

awesome. I always liked unclip tbh. but then it kind of did some wacky stuff sometimes

[zinc cargo]

i wish model pages will come with sample workflows as well 🙂

[sour obsidian]

Oh unless you meant actual clip vision. Its part of the clip model as clip was designed to compare text/image similarly

[sour obsidian]

There should be some 🙂

(Or were you commenting that all should come with workflows, in which case I totally agree 😎)

[upbeat summit]

dropping crazy new things right before the weekend

[crisp owl]

https://huggingface.co/stabilityai/control-lora/tree/main/comfy-control-LoRA-workflows

[wicked frigate]

rule #1: always release on a friday so if there's any complaints you can run and hide during the weekend

[autumn forum]

release on friday so it gives the weekend to play 😛

[upbeat summit]

brilliant!

[zinc cargo]

mm, very strange unrelated results... 🙂

[soft zealot]

I hate it when the powers to be decide to release a new software package on a Friday for something thats mission critical (which at least this isn't)

[zinc cargo]

i've taken these to pics 🙂 guitar and burger...

[hardy cipher]

unclippin

[soft zealot]

intersting nipples for Superman

[hardy cipher]

my favorite unclip image

very top notch

[zinc cargo]

i dont get it 🙂

[hardy cipher]

well it just pasted images in

for some reason

[zinc cargo]

not your pic, @hardy cipher it's amzing

😛

[hardy cipher]

lol

[zinc cargo]

just the concept...

[soft zealot]

it helkps if you click reply on a message so people have got a chnace of understanding what you're referring too

[zinc cargo]

sorry, it's late here and im trying to wrap my head about everything

[soft zealot]

its late(ish) here as well 🙂

[zinc cargo]

@upbeat summit nice!

[zinc cargo]

any text or just that?

[upbeat summit]

I left the text from comfy's workflow in beautiful photograph 😄

[wet nacelle]

Was new stuff released today? If so can someone mass link it?

[peak hornet]

How do i get SDXL?

[hardy cipher]

internet

[zinc cargo]

download the model from civit or hugging face

[peak hornet]

And how do i use model?

[hardy cipher]

I mean

there are lots of ways, bud

[crisp owl]

#✨｜sdxl message

https://huggingface.co/comfyanonymous/clip_vision_g/tree/main

https://huggingface.co/stabilityai/control-lora/tree/main/control-LoRAs-rank256

https://github.com/Stability-AI/stability-ComfyUI-nodes

[zinc cargo]

this is just text

this is with 2 pictures and delicious guitar in prompt

[peak hornet]

Just tell me one way

[hardy cipher]

https://beta.dreamstudio.ai/generate

[peak hornet]

Okay thanks

[upbeat summit]

yeah - the one example I posted worked great. but now I'm only getting random results. not sure yet how to control this

[zinc cargo]

@visual glade are there preprocessor nodes for canny and the rest?
i can't see the ones that ar ein the samples.

[upbeat summit]

do I describe my input images like in img2img?

[hardy cipher]

if you actually want to use it on your own you'll need to download some software, and you'd need to learn some things. so might want to read up on it a bit. can't really explain it all here

[wet nacelle]

What is the clip_g for? Is it just to have it work?

[upbeat summit]

or do I need input images the clipvision model can understand?

[zinc cargo]

@upbeat summit can't wait to see your images with sdxl

[soft zealot]

hmmm she go "boom" boss

[crisp owl]

Yeaaaah that's all I'm getting

[zinc cargo]

did you update comfy after?

[visual glade]

make sure you are updated

[upbeat summit]

hmm works great here - updated comfyui earlier

[crisp owl]

oh not in the last couple hours haha

[zinc cargo]

anyone was able to run canny edge?

[soft zealot]

*SIGH here we go again, the curse and the blessing of OpenSource, everything being out of date if you blink!!!

[zinc cargo]

earlier i took the burger into community canny

[vital wraith]

I had issues with comfyroll this last update and it was because I never installed ComfyUIManager correctly. Didn’t show up as a problem until the most recent update >.> Was totally user error in the end in my case.

[zinc cargo]

and did this

[crisp owl]

redwall, just ran update .bat file

[tiny walrus]

i now want sushi burgers

[soft zealot]

Seems to be tiunning COmfys example workflow with red walling after I updated

[upbeat summit]

prompt: cinematic still

[crisp owl]

ah shiet, updated my wrong instance

[upbeat summit]

Hi KaliYuga 🙂

[unreal plover]

control lora is working with auto1111 ?

[tiny walrus]

howdy 🙂

[zinc cargo]

could you install the stability nodes?
i dont have the preprocssors

[crisp owl]

I ahven't gone through everything yet, so unsure

[zinc cargo]

oh 🙂

[soft zealot]

I dont know what I expected really lol

[zinc cargo]

context works out.

[soft zealot]

the context is that Comfys example workflow

[soft zealot]

with my images n lol

[zinc cargo]

naaa, the context is the struggle of the mind between the inner joker and the bat.

[crisp owl]

nice

It works. Tinkering and figuring out what it does will be later haha

[soft zealot]

for that I'll share them 🙂

[zinc cargo]

when it works it works!

so this should be a Blend kind of option?

like Midj'?

[eternal fog]

I stopped paying attention for a couple days and new things appear 😮

[crisp owl]

couple hours is all it takes

[soft zealot]

Well thats a better result if I dont uise the basic SDXL1 model.

Not sure wherre this little darling came from given the input images though lol

[upbeat summit]

such a surprise in this scene 😄

[eternal fog]

Do you need a model for revision?

They talk about it on the Hugging face page, but I don't see a model for it.

[zinc cargo]

dont blink 🙂

[visual glade]

the only thing you need is CLIP-G Vision: https://huggingface.co/comfyanonymous/clip_vision_g/tree/main
put it in your ComfyUI/models/clip_vision/ folder

[eternal fog]

Ok, seems like maybe an oversight on the Hugging Face Model card then?

As it tells you about it, but doesn't explain how to use it

[soft zealot]

Weeping Angels, Classic episodes of New Who

[upbeat summit]

color grade and pattern to jacket transfer

[visual glade]

yeah it's going to be added to the huggingface page

[zinc cargo]

sorry to nag, what about the rest of the nodes? we want to play with canny edge

[crisp owl]

https://huggingface.co/stabilityai/control-lora

[visual glade]

for canny there's a node in the base install

[native knot]

Now do it with her skin instead. 😉

[zinc cargo]

is there a way for us to know what the clip vision sees?
like a preview for what ever input it passes?

[soft zealot]

Answers on a Postcard as to what this will produce??

[native knot]

Manbat. 😄

[soft zealot]

spoiler ;o)

[zinc cargo]

hey 🙂

but i also had to add a lora to this

@tiny walrus

[upbeat summit]

amazing

[autumn forum]

id eat that

[upbeat summit]

no prompt

[sour obsidian]

I am excited to see all the combos you all come up with 😄

[upbeat summit]

any tips for noise_augmentation?

[sour obsidian]

ah I think thats actually just a vestigial piece from a real unclip. Dont believe it has an effect here. It was meant as a sort of feathering between clip img and a clip txt embedding made with a prior model (real unclip). Correct me if it is doing something though @visual glade

id just leave it at 0 since these are operating on their own

[zinc cargo]

anyone can help me with the settings for canny?
im trying to understand why my preprocessor can't "see" the image and how to setup the low and high thresholds

[upbeat summit]

thank you 🙂

[eternal fog]

I don't think I'm using it right lmao

[visual glade]

noise augmentation will have an effect but it's not useful in this case

[visual glade]

so it should be zero

[steel mist]

Whoa it never even occurred to me to use load a png with alpha

[crisp owl]

Generally it just converts is to a white background I believe

It was just an image I had in my downloads folder

[upbeat summit]

great experiment

[steel mist]

Ah bummer figured it would use it as a mask

[crisp owl]

Yeah I think if you use in in img2img it just white backgrounds it

[upbeat summit]

what is your prompt if I may ask?

[upbeat summit]

every time I get beautiful landscapes my settings aren't dialed in yet, right? 😄

[stone fossil]

As something simple such as: a guitar made out of PBC, and motherboard traces and computer parts
will do.

[autumn forum]

hey this is pretty fun

[stone fossil]

IF you want the van gogh one try: a very abstract painting of guitar in the style of Vincent van Gogh, starry night backdrop

[upbeat summit]

ok so you are customizing the prompt depending on your input images

[crisp owl]

That's what I'm doing

Not sure if that's the proper use, but 🤷🏻‍♂️

[stone fossil]

Thats how to use CN right?

[upbeat summit]

I guess so 😄

playing with the strength values a lot right now

and what differences are there where the image is placed and how it's all connected in comfy's workflow

[autumn forum]

honestly might be better than prompting.

[soft zealot]

I've no idea what im doing, where I'm goig etc (other than too bed) but........

[stone fossil]

AI gen in AI gen out.

[zinc cargo]

dat order though

[autumn forum]

i just zerod out the conditions and it started working without generating women lol.

[sour obsidian]

I personally found 1.75 on the unclip node when combining with text input tends to give a nice balance between em 🙂 might not be perfect in all cases though

[zinc cargo]

but which model?

[upbeat summit]

which node?

[sour obsidian]

unclip conditioning strength

lots to be tested though

[upbeat summit]

because I just looked at how it's all connected and there's of course a difference between both unclip conditioning nodes (how they are connected)

[sour obsidian]

ah this would be for single image input. so an example might be spiderman text input + starry night image as the one img input to revision. For two it might be worth bumping them both up a touch since text tends to over ride the image input at base values when not 0ing it out

[upbeat summit]

single input... gotta try this 😄

[soft zealot]

thng is you can Combine, Concat (which seems to be the default in the example provided) or Average the Conditioning .

Meanwhile zeroed out the top 2 text inputs and got this

[upbeat summit]

thanks for explaining! lots of experiments to do heh

[sour obsidian]

you could always prompt the "new" way . Revision can read text 👀

[crisp owl]

thinks about 3..no 4...no 5! image inputs

[soft zealot]

this is going to rurn into a classic Oil Type talking point about what is allegedly the best way to do something.

[upbeat summit]

https://tenor.com/view/this-is-all-too-much-i-need-to-lay-down-i-cant-even-gif-18356272

[eternal fog]

Which example workflow should we be using, there seems to be multiple

[zinc cargo]

this is amazing!

[autumn forum]

ive noticed that it takes elements in a weird way. i could definitely tell how it would be able to do this. this clip vision thing is very cool.

[sour obsidian]

yea its pretty interesting to play around with. its very conceptual, merging the concepts as a whole vs just the image features

my mind was also blown when @runic granite tested the text input the first time with something about a tree and a tree showed up haha

[autumn forum]

yeah! thats a way to put it. i honestly would have had a hard time describing it but that was well put

[upbeat summit]

so when do we get subdirectory support in /ComfyUI/input/ - this will get a lot of new content 😄

[soft zealot]

toss a coint

[zinc cargo]

a couple of rolls, but yea

cfg 9 btw

[native knot]

To your Witcher?

[sour obsidian]

yea we noticed single words seem tougher for some reason. So just dog would rarely work, but a cute dog showed up much more often

[autumn forum]

like these outputs are so insanely good its like a whole new way of prompting. with images!

[zinc cargo]

should we get a dataset of texts on images? ---- wait i've got an idea!

[sour obsidian]

haha it might help a good bit. There is likely some good potential in tuning things for revision specifically

[zinc cargo]

no idea

actually, when i write beautiful photograph in the prompt, i'll get a woman

[soft zealot]

I just loaded the lyrics in as the positive prompt and got this

[hard fractal]

Hahaha

CLIP 🤷‍♂️

[autumn forum]

zero out the conditions if you dont care to prompt

like this. and it works great

[crisp owl]

haha, sometimes 🤣
No prompt for my Woody and Mona Lisa merge

[hard fractal]

Ah. The word bubble around it makes it think it's a comic book

[sour obsidian]

yea its neat how it interpreted it

[zinc cargo]

yea, i guess i dont get it

[autumn forum]

did you zero out the conditions? or do you have to prompt? if prompting, describe exactly what you want

[zinc cargo]

i've no prompt and i zeroed the condition on both

[autumn forum]

no look at my photo lol

#✨｜sdxl message

[hard fractal]

It's not zero'd out

You need a new node

[sour obsidian]

ah yea youre just 0ing out the image strength there so its not using them. Gotta add that little node in between

[zinc cargo]

https://tenor.com/view/facepalm-really-stressed-mad-angry-gif-16109475

alright i'm back in the game

ah!

[autumn forum]

nice!

[zinc cargo]

such a nice series as well

[autumn forum]

indeed. thanks so much to the team. this is like the ultimate wildcard system. i never have to worry about silly prompting again! hehe

[zinc cargo]

lol

[hard fractal]

We can add it to the bot now 😄

Before, people would've been very confused about the new "ControlNet"

[zinc cargo]

sorry, i'm gonna do a few more in excitment of how nice it comes out 🙂

blantly borrowed from MJ's community page (sorry OP)

[hard fractal]

Pretty cool eh?

[upbeat summit]

no prompt

[zinc cargo]

can't read hebrew though...

[hard fractal]

One image in, with a prompt:

[upbeat summit]

thanks for this! 😄

[sour obsidian]

@tribal gale we need to fix this

[autumn forum]

your welcome! however i want to try and find a way to implement a prompt for additive details instead of having to prompt for the whole thing.

[zinc cargo]

changing the aspect ratio is none issue as well

[soft zealot]

Is it because I have a perverted mind??

[zinc cargo]

i wish

[zinc cargo]

supportive prompting

[autumn forum]

yeah thats a good name for it lol.

[zinc cargo]

those neat workflows make me feel so bad for my spaggethi

[soft zealot]

only @high skiff &and @visual glade quickly mashed together and straight lines turned on

[upbeat summit]

yeah, makes sense. since both images (and both unclip nodes) are connected in different positions in the pipeline I'm trying to figure out what the differences are

[autumn forum]

so i tried to implement supportive prompting. annnnd no. but there is red hair

[upbeat summit]

beautiful 😄

[sour obsidian]

im waiting to see you all mix revision and controlnet stuff too 😄

really fun to give a base with say depth and toss image concepts to mold into it haha

[autumn forum]

soo it worked???? but i can tell mid generation that its realllly struggling

[zinc cargo]

straight lines? is that a plugin?

[sour obsidian]

click the little gear on the menu and change the line render mode, should be in the base comfy 🙂

[zinc cargo]

https://tenor.com/view/blinking-eyes-white-guy-blinking-meme-what-huh-gif-26334322

[hasty smelt]

I'm considering subscribing to Collab Premium to generate a few images per day. The free plan is always disconnecting, which seems to be a common problem with free accounts. I'd like to know your opinion if this seems like a good idea. My goal is to just generate images, nothing too heavy like training models or something like that. I'd like to know if XL works well on Collab Premium?

[zinc cargo]

my life kind of changed right now...

[zinc cargo]

can only use canny right now, right?

[autumn forum]

straight lines = satisfying, but when they overlap your like. wait where is this line going? lol

[sour obsidian]

depth is pretty great too. Just need to grab one of the pre-processors out there for it

[wet nacelle]

Alright. Does anyone have a workflow they want to share yet?

[pallid path]

SDXL remix possible?

[autumn forum]

for what?

[wet nacelle]

The newest set of plugins

I'll drag that and see how much I understand it.

[zinc cargo]

i was trying to find one 🙂

[wet nacelle]

@upbeat summit Klinter did already but do you think that you could also share yours?

[autumn forum]

so i definitely got it to work guys. supportive prompting. no text = zerod out condition and works like normal. and with text adds to the image, prompt may need to be a bit specific sometimes

[upbeat summit]

for clipvision I'm using the one comfy posted earlier #✨｜sdxl message

[wet nacelle]

Also should I download windows_portable and uninstall the old Comfy Windows version? I still have the one that is just labeled Comfy for Windows.

[wet nacelle]

Alright rad

[zinc cargo]

did you try to inject lora there?

[autumn forum]

havent tried lora yet. i guess i could.

[sour obsidian]

I definitely recommend cranking up the strength on the unclip node if you are mixing revision and text. I found ~1.75 strength to be good for most cases but didnt test for too long. Text still has more inherent strength over revision in most cases

[empty orbit]

is it normal for the VAE step to take a very long time

[autumn forum]

the main thing i was trying to solve was that without the zeroout, not having text produces basically nothing. i wanted to be able to do both with no problems. so text adds details, without text is basically zeroout so it works either way. if that makes sense

[sour obsidian]

there is definitely a difference between no text and 0'ed text. 0'ed will get you even cleaner images when purely image prompting as it totally 0s the embedding where as no text still produces a non-zero embeddings. Still there is a lot to try with these things

[zinc cargo]

i'm not gonna get sleep tonight ah...

combine all the images!

[autumn forum]

agreed. its very exciting. ill definitely be doing lots of tests

[empty orbit]

comfyui?

how are you doing that lol

[autumn forum]

clip vision!!

[empty orbit]

damn i gotta try that now

[zinc cargo]

@empty orbit just grab any of these images

[zinc cargo]

with Aether?

[upbeat summit]

base sdxl

[zinc cargo]

nice

[upbeat summit]

no positive but some fidelity tweaking with this negative
photoshop, video game, ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, mutation, mutated, extra limbs, extra legs, extra arms, disfigured, deformed, cross-eye, body out of frame, blurry, bad art, bad anatomy, 3d render, photograph

[zinc cargo]

@sour obsidian can i assume that clip recognize instagram icone and prdocuve pretty girls because of it?

(also, yea, i've added a 3rd one)

[sour obsidian]

hahah its totally possible

[eternal fog]

I'm offended, I gave it images of London and it gives me the eiffel tower 😢

[eternal fog]

Still don't fully understand how this is working, but it's interesting

[upbeat summit]

latent space style transfer image mixer interpreter

[sour obsidian]

put that on the poster

[eternal fog]

That makes me no closer to understanding lmao

[zinc cargo]

on a bumper sticker

[sour obsidian]

id rock it on my car

[hard fractal]

Ooooo

[upbeat summit]

amazing tech

[zinc cargo]

had to be chirped

[hoary saddle]

is there a fix for when dropping workflows and/or images onto the comfyui webpage doesn't update?

[zinc cargo]

not everyone is sharing their info when they are uploading the image

but here is one from me 🙂

[hoary saddle]

yeah, this is the workflow @hard fractal dropped earlier

load nor drag are updating

[upbeat summit]

definitely need subdirectory support in Load Image nodes and ComfyUI/Input/ - like the Load Checkpoint node

[hoary saddle]

these arent' droppable i assume since just pure json?

[sour obsidian]

worst case just download and hit the load button. I moved the revision stuff under its own folder since its a tad sep along with the G vis model so its all in one spot https://huggingface.co/stabilityai/control-lora/tree/main/revision

[zinc cargo]

with prompt?

[upbeat summit]

nope - only negative

just tweaking the strengths

[zinc cargo]

sweet

so differnt toy, did anyone tried to do a fancy QR with SDXL yet?

[steel mist]

Does anyone publish a "what's new today/this week in stable diffusion" newsletter? The speed of releases is 😵‍💫

[upbeat summit]

only with SD 1.5 - very cool. starting to use it for clients

[zinc cargo]

yea, me too, but with the sdxl it looks so much better

i tried with canny but no good yet

[upbeat summit]

yeah - have not used controlnet yet with SDXL 😬 I'm also a total noob at it. but will learn it now

[zinc cargo]

well, it kind of came out today 😛

[runic granite]

You guys are killing it with the revision images! Nice stuff 🤩

[upbeat summit]

hey @runic granite 😄

[sonic vortex]

Hi, I need a little help with the new stuff please! I'm trying to use the control lora workflows from here: https://huggingface.co/stabilityai/control-lora/tree/main but i seem to be missing these nodes: ScribblePreprocessor, CLIPVisionAsPooled, ColorCorrect,MiDaS-DepthMapPreprocessor,CannyEdgePreprocessor

Could anyone point me in the right direction please (I already installed the nodes from https://github.com/Stability-AI/stability-ComfyUI-nodes)

[upbeat summit]

yeah - I meant ControlNet in general. But the time is now 😄 and also training. You know - always focused on prompt engineering but it's time to do some fine-tuning myself 😉

[zinc cargo]

@sonic vortex grab a glass and relax (the workflow is included)

[zinc cargo]

you're going to be awsome in that as well ah?
you can't leave something for us mortals?

[upbeat summit]

Thank you 😄 and you are being very awesome yourself, Klinter!

[zinc cargo]

hey, i've learnt everything i know about prompting from you.

2.1 was a great school for me

[sonic vortex]

Thank you so much! Now to try and work it all out 🤣

[upbeat summit]

https://tenor.com/view/kill-bill-master-gif-10355953

[zinc cargo]

lol

[zinc cargo]

but today's MVP is via with the zero prompt

[upbeat summit]

that's awesome!

[eternal fog]

Just changing the weight of each image can change the results a lot

That's with the space like image having higher weight

[upbeat summit]

I'm doing nothing else 😄

[eternal fog]

Why does it like the Eiffel Tower so much

[upbeat summit]

strong in the latent space

[hoary saddle]

what's this error mean for clipvision?

[autumn forum]

https://civitai.com/models/130726 i uploaded a revision workflow to civit lmao. i know that seems silly but it has additive prompting and i know a lot of people check there

[upbeat summit]

update comfyui

[zinc cargo]

update after

[native knot]

Bruce Leroy, the last dragon, has the glow and joins the Avengers

[hoary saddle]

gotcha

[upbeat summit]

it was an awesome time 🙂

[past harness]

any update on preprocessors missing for the official comfy controlnet workflows?

[eternal fog]

I think I can see where it's getting this from. Space Ship, Ship, Building, oh lighthouse.

lmao Long boy

[upbeat summit]

@autumn forum got AIT working with it yet?

[autumn forum]

lmao i just tried it and it gave me an error hah. but no not yet.

[upbeat summit]

Error occurred when executing KSampler:

Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

hmm okay

[zinc cargo]

it still is! 🙂

[zinc cargo]

wow

[indigo carbon]

huh, is this SDXL clipvision?

[eternal fog]

Yeah

[zinc cargo]

@eternal fog which model?

[eternal fog]

NightVisionXL

[indigo carbon]

link? I didn't know there is an SDXL clipvision thingy

[eternal fog]

https://huggingface.co/comfyanonymous/clip_vision_g

https://huggingface.co/stabilityai/control-lora/tree/main/revision

2nd one has a workflow with it, so probably get it from there

[upbeat summit]

comfy's workflow for ClipVision: #✨｜sdxl message

[eternal fog]

lmao

oh wow

I put a negative weight on the woman

Less woman more Alien I guess

[upbeat summit]

she's an alien

[zinc cargo]

these are so awsome!

[upbeat summit]

nice - I think you got it 😄

[eternal fog]

It's not what I was expecting, but it's cool

This is better for me because I'm useless at thinking about what I want to prompt

[golden palm]

Hey guys, I have been using the Searge 3.4 workflow and I keep getting the following error. I'm not sure how to fix it. Any help would be greatly appreciated 🙂

[zinc cargo]

almost...

[zinc cargo]

does the strength of the unclpping get normlized to 1?

(is that a smart question? it's very late here and my brain is not 100%)

[upbeat summit]

haha

[strong field]

Try to update comfy and then update the searge nodes, likely a version mismatch

[runic granite]

If you guys want to get really fun and confused, combine two images and put another image in as a negative 😁

[eternal fog]

That seems like a good way to get cursed images

[golden palm]

Trying that now

[patent badger]

what is this DRLX announced in #📣｜announcements ?
I'm not really that into software / coding so if anyone (who's obviously way smarter than me) wouldn't mind sending an explanation and their thoughts about it I would more than appreciate it!

[stiff wolf]

What tools are currently working for training textual inversion embeddings for sd-xl?

I've mainly been using auto1111 sd-webui.

Which doesn't have the training working yet for sdxl

[zinc cargo]

Gnight eveyrone. amazing new tool! thank you for this.

[upbeat summit]

sleep well! ttyl 🙂

[sour obsidian]

kinda fun, you can just toss in a simple sketch with no prompt and get better sketches of the same concept

[vast galleon]

In my quest to find a better color transfer method for I2I I've found 3 that I've managed to get working. Semi Discrete Optimal Transport, Linear Monge-Kantorovich, and one that uses a HistoGAN model. Working on cleaning the code up and popping it into a ComfyUI node.

[runic granite]

(Woman + Power Armor) - Santa on the beach = Anime Girl... Sure! 😄

[eternal fog]

?

lmao

[steel mist]

I missed the discussion; Do the control loras offer advantages beyond efficienct/compactness?

[runic granite]

PSA: to get the official CLoRA workflows going you will need this node pack:
https://github.com/Fannovel16/comfyui_controlnet_aux

Canny works with the build in preprocessor, just swap it out and tune the values accordingly. We found low_threshold 0.100 and high_threshold 0.200 does a good estimation.

We will add a note on the HF Repo pages as well.

[upbeat summit]

so clean! great characters

[hardy cipher]

thanks! left it running while I was gone. got all sorts of interesting things

[wet nacelle]

Hey which folder will these go into?

[runic granite]

Just posted some info on how to get it to work above. 🙂

[autumn forum]

controlnet folder

[wet nacelle]

thanks

[wet nacelle]

the revison file goes here correct?

[upbeat summit]

Using ProtoVision XL by @delicate kelp

[wet nacelle]

sweet

[golden palm]

Thank you that worked... 🙂 It took a while but it is working...

Now I have a new issue. I keep getting the following error:

Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.

Now I do already have --lowvram

Are there any other command line arguements I can use to get this to work better?

[wet nacelle]

I just switched over to windows_portable. Do I have xformers automatically enabled?

[hardy cipher]

what's a bit odd is the same setup also created these images, lol

ooh, put the one in twice

[eternal fog]

It says using xformers in your screenshot

[wet nacelle]

That's what I thought. Thanks.

[upbeat summit]

the latent space just wanted to have some fun!

[hardy cipher]

indeed. I've always liked the unclip/clip vision stuff. really takes on a life of it's own

[wet nacelle]

@upbeat summitCan you help me understand what works best when using revision?

[upbeat summit]

revision?

[wet nacelle]

Clip

[hardy cipher]

clip vision?

[wet nacelle]

The UI I think you said I should use.

yes?

I think?

I read somehere "revision"

[hardy cipher]

it basically creates conditioning data from an image

[upbeat summit]

hmmm... I don't think I mentioned that

[eternal fog]

Revision is what SAI are calling the Clip Vision G stuff on their hugging face

[upbeat summit]

oh ok 😄

[wet nacelle]

I mean how can I make it actually look the way it was meant to look? I'm trying to use the workflow Comfy sent had the tree as the reference image.

[hardy cipher]

similar to what you'd get from a prompt encoder

[eternal fog]

You put 2 images in and they act as prompts.

You can change it up and use 1 image, or images and text

[hardy cipher]

you use an image as input, attach the model to the clip vision

[eternal fog]

It's only just come out, so play with it and see what it does.

https://huggingface.co/stabilityai/control-lora/blob/main/revision/revision-basic_example.json

This is a more simple workflow that uses no prompting and just images.

[hardy cipher]

and then patch it into conditioning data from your prompt encoder

[upbeat summit]

you first need to get images... like your inputs. styles, people, artwork, film screengrabs. anything you want to play with and put those images files in your /ComfyUI/inputs/ folder

[wet nacelle]

I mean it looks awful. Not at all the way it's meant to. I'll send in an example in a second.

@upbeat summit What checkpoint have you been using with it?

[visual glade]

There's some more revision examples on this page: https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

[eternal fog]

You don't need to do that. You can just drag and drop images into the Load Image nodes.

[upbeat summit]

I used based earlier. I'm not using ProtoVision XL

[upbeat summit]

true - I just like to have them all there to switch between them

[eternal fog]

That requires planning on what images you are going to use, I just find random stuff and throw them in.

[wet nacelle]

I mean I just got this from thjese.

[eternal fog]

What are you expecting to get

[autumn forum]

https://civitai.com/models/130726 use this workflow and it fixes the problem without prompting anything.

[wet nacelle]

I don't know. I don't know how it works. From what I've seen though it should've have mixed the two styles together to make something new. It shouldn't be entirely heavy on one style or subject.

[autumn forum]

sorry i keep spamming my workflow heh. just want people to get good results out of the box

[sour obsidian]

hey Via, I updated the clip vison model onto the main repo. Any chance you'd be willing to switch the link at the bottom https://huggingface.co/stabilityai/control-lora/blob/main/revision/clip_vision_g.safetensors

[wet nacelle]

@upbeat summitOh Comfy called it revision.

[eternal fog]

It doesn't really mix, it takes what's in the images and creates a prompt from them, as far as I can tell.

You also have the added problem that you aren't using the workflow with the nodes to Zero out the conditioning, so even though your prompt is blank, there's still technically something there. And that has much higher strength that the images.

[autumn forum]

yes no problem will do that right now

[hardy cipher]

you can set it up like this

[hoary saddle]

would be cool to see what prompt the combined clip was

[hardy cipher]

the unclip conditioning nodes require conditioning input

[sour obsidian]

no prompt involved. All just image concepts 🙂

[polar epoch]

I need me some SDXXL 2048 models

[eternal fog]

I guess does it actually use "prompts" in the way we'd think about it, or does it just mess around with latent space.

[hardy cipher]

I wonder if it could be quantified into actual tokens

[eternal fog]

I get what you mean, like what prompts is it putting into the model to get the results, but I don't think it's using text tokens.

[wet nacelle]

Igga thinks I know the first thing about setting up nodes myself. No offense to you but I'd rather have a workflow that someone else makes for me (or they share it with the rest of the communtiy).

I am lacking the mental strength to learn that currently. The last time I tried a node set up myself I sweat so bad and got a headache.

[hoary saddle]

oh i thought that's what clip did was that AI would attempt to describe what's in the images, is that wrong?

[sour obsidian]

its using a shared latent space where clip compares images and text (pooled embedding space) so its not a prompt but it acts more conceptual like one

[hoary saddle]

so its not simliar to a1111 clip interrogate?

[sour obsidian]

its very different, this is directly using concepts from the image, no tokens involved

I guess it could be used in a similar way but this is direct

[hardy cipher]

I have my workflow attached here, but I'm using a couple custom nodes I made, so you would either have to swap those out or download them

[wet nacelle]

Plus a lora right?

[heady vale]

Have an issue getting the Controlnet nodes to work. I run the install.bat for dependencies and everything is satisfied but have this error appear repeatedly
DEPRECATION: torchsde 0.2.5 has a non-standard dependency specifier numpy>=1.19.*; python_version >= "3.7".

[eternal fog]

@wet nacelle Use the workflow that me or Via posted. That will be the easiest way to start without using any promptng.

[polar epoch]

Of course the mutated stray ears are perfect

[eternal fog]

The one I posted is SAI's example, Via has done their own.

[hardy cipher]

you don't really need a lora, but I'm using one currently

[heady vale]

updated torch, updated comfyui, updated controlnet nodes. but no dice

[polar epoch]

damn, name module pack please

[wet nacelle]

The thing about that is I fear that I may need to rewire the nodes if the Lora is active.

[hardy cipher]

I just made it today. . you can download it here if you want https://github.com/picturesonpictures/comfy_PoP

[polar epoch]

Neat, thanks! :D

[hardy cipher]

no problem. it's something I'd wanted for a while, so had to teach myself a bit of python to make it, lol

there's the lora stacks, but none of them have that input output that I'm aware of

along with the on/off switches

[polar epoch]

Aye, i sadly lack skills on where to even start to learn python, so i'm at dead start for eternity as i don't know any way i can effectively learn it

[native knot]

Holy shit...I just hit a baller combination.

[polar epoch]

That's what i've been waiting for for weeks. A simple switch to turn off an entire section, for instance img to img, upscalers etc

[steel mist]

Ugh, I'm running into this exact issue right now

[wet nacelle]

Could I be doing something wrong?

[heady vale]

not sure what the problem is. still ttroubleshooting here

[autumn forum]

im not sure. lol idk how that could happen

[steel mist]

I'm updating a few things now but yeah, not sure what's happening

[polar epoch]

Noice, found a somewhat highest native res i can make without too much artifacts, now to just shout at negative prompts to not screw up tails or ears

[hardy cipher]

same. it was a bit of a hassle to put together, but future nodes should be easier for me

[autumn forum]

unless it doesnt like the alpha channel on that image? idk

[wet nacelle]

Should I try to find a way to update Comfy?

[autumn forum]

have you not? lol if so you definitely need to lol. this got updated today

[wet nacelle]

Would I do it from the update folder?

[autumn forum]

do you have a update_comfyui.bat file?

[polar epoch]

Aye. If you get it all running nicely, i have a node request, and that's a model and lora combi module :P Were base, refiner, as well as lora's underneath. And if possible, a plus button to add as many extra loras as one wants with lora priority for each

[wet nacelle]

I just clicked it. It showed the latest update date which was today.

[autumn forum]

okay cool then it should be updated

[hardy cipher]

hmm. I'll see if I can figure it out. the plus button part might be tricky. but could add like 5 lora spots

[wet nacelle]

Can you link where you got your clip file?

[hardy cipher]

I'm sure adding loras like that is possible. but not sure how I'd do it

[autumn forum]

its on my civit page where you got the workflow

[polar epoch]

Indeed

[polar epoch]

if successfull, do like 15 or 20 lora's, i made a image once that had 19 lora's

[hardy cipher]

I'll make a deluxe lora stack, lol

[polar epoch]

Here's my 19 lora generation from sd 1.5 using just armor loras

[hardy cipher]

damn, that's pretty clean

[polar epoch]

You've no idea how many generations that went for that one clean ish gen xD

[hardy cipher]

well I could probably do a larger lora node later tonight. but your other request would take a bit more time. but I can do that too. good practice for me

[polar epoch]

If you go crosseyed, tail will fit /s

[strong field]

i didnt know i needed this in my life

[polar epoch]

[hardy cipher]

I'll be back here in a few hours. if you're still around and I get that figured out I'll send you a link

[polar epoch]

Ping me when done, it's 4 am, so soon heading to bed

Also, how the heck did i get gamer role lol. It's not in roles

[polar epoch]

Damn, that's clean

Casper?

[autumn forum]

did someone say casper?

"touch my chickens and you get the lead"

[autumn forum]

this would be sick if the pignose grill wasnt f'd

[vital ermine]

heh

[wet nacelle]

@visual gladeDo you think you can help me set up Clip Vision? I'm on the latest version of ComfyUI and have all the needed files to get it working. The issue is that each time I use it in a workflow it seems to only understand the prompt and not the image nodes.

I've also just uninstalled the UI and reinstalled it to see if that was the issue.

[wet nacelle]

I've tried several workflows now.