✨｜sdxl

[spring fulcrum]

ya

that the only one I could get.. you were MIA at the time I asked about it earlier tonight

[indigo carbon]

right click, then select bypass it should turn purple after bypassed

[spring fulcrum]

https://tenor.com/view/staples-easy-button-that-was-easy-easy-gif-12841577

[indigo carbon]

I can give you this if you want, but no AIT support yet

[hardy cipher]

ipadapter is magic

got an extra leg there

[spring fulcrum]

ya that sounds good

[spring fulcrum]

what in the brady bunch is that?

[hardy cipher]

put in 3 images, get a 4th one. give me 3 images and I'll put them in

first one dictates the theme and structure most, 2nd is kind of the accessorizer, third mostly dictates structure

[spring fulcrum]

try these

[hardy cipher]

lol, well okay.

[spring fulcrum]

I dont really have anything in mind sry they are pretty random and very different

[hardy cipher]

there are a couple other settings to mess with as well, but it's mostly good to go

[spring fulcrum]

Let me give that one a try... I'll download it and give it a try in the morning... I have to get up early to drop off my son

[crisp owl]

lol he says those are random...then you compare to what picturesonpictures showed

I think he wins in random pics lol

[spring fulcrum]

Ya I would like to give that a try too.

[spring fulcrum]

This is the new one? yay

[hardy cipher]

lol

have blip models looking at input images and then giving descriptions, then sending those through stylers

[indigo carbon]

also did this, it's a version that does image input, so prompting is like [IMAGE] as a.../doing.../at...../

[spring fulcrum]

I will have to play around with that to get some interesting results. Can you post the resulting image?

I want to try to keep your settings to start with

[hardy cipher]

you might have to download some nodes

[spring fulcrum]

I did

nvm I still need

are those on github? or Civit?

[hardy cipher]

do you have the manager?

[spring fulcrum]

ya and I downloaded the missing nodes and restarted comfy and that still comes up

[hardy cipher]

hmm

let me check something

one moment

[spring fulcrum]

Do I need this top one too?

[hardy cipher]

probably

[spring fulcrum]

I figured it out... I needed to download POP

[hardy cipher]

ahh

you had me worried there

[spring fulcrum]

Its all good

I am running 3 different images with those settings

[hardy cipher]

time for some magic

I just made that workflow up today

[spring fulcrum]

lmao

ok well at least it work... Ill keep trying tomorrow. have a good night everyone

[hardy cipher]

that's pretty high quality tbh

goodnight

[spring fulcrum]

It was pretty quick too for a full upscale to 4k

[hardy cipher]

slightly cursed

[stone fossil]

Prompting.

[stone fossil]

Nice haha.

[hardy cipher]

WAS Node Suite: BLIP Caption: a cat made out of yarn on a table with a bookcase in the background

WAS Node Suite: BLIP Caption: a corona corona corona corona corona corona corona corona corona corona corona corona corona corona corona corona corona

BLIP having a bit of a trip

have the first caption going to pos g and the second to pos l, then both combine with a second text box to put in various things, and then through a styler. first image is clip g and ipadapter, so it'll take on whatever style that image is

[floral island]

steampunk and denim ❤️

[floral island]

he seems terribly excited to see me 🙂

[craggy ibex]

Are there any custom node for Latent Image?

the default one seem very limited

strictly forcing width and height

[uncut fiber]

have you tried pure Empty latent image?

[craggy ibex]

pure empty latent image?

[uncut fiber]

just name of node Empty latent image?

[craggy ibex]

oke hold on

[floral island]

great fisheye lens 👁️

[craggy ibex]

I have

[soft zealot]

FYI been revampping my workflows and bundled together as a collection

My personal Daily Go to is my "Middleweight" workflow.

Features include

1. ChatGPT Assistant (disabled by default as requires an Open AI account)
2. Selectable Positive Prompts (linguistic.Random or One Button)
3. Selectable Upscale Method (Standard, Ultimate or None)

By Default loads (on a 1440p monitor) into the view in the screenshot.

Lesser used (by me at least) features such as ReVision & Controlnet are off to the right hand side

All control elements are in "The Engine Room" on the Left Hand Side

https://civitai.com/models/138397/winstons-workflow-collection

[craggy ibex]

when I typed.. 540 it forced to 544

[west breach]

has to be a multiple of 8

[uncut fiber]

must be divideable by 8

[craggy ibex]

oh

damn it.

[west breach]

you could resize after

[stuck bobcat]

Did some crazy cool ones today

[zinc cargo]

i love nightvision model

[untold dawn]

anyone knows a stable diffusion that is compatible with amd gpu

[uncut fiber]

i already sent you link on AMD version of A1111

[royal hatch]

ei guys, can xl stable diffusion API be used for commercial use? And the other APIs?

Because i have read the term of conditions of the site and the site is not free for commercial use, but maybe the apis yes? I do not know

[spring fulcrum]

@indigo carbon So I was trying the new image blend and it wont let me change the model name

[indigo carbon]

You are missing the models required, if you have them with a different name, you can change to them accordingly

[spring fulcrum]

Where do I get them and what folder do they need to go in?

It doesn't seem like I can use regular models like copax or nightvision

[untold dawn]

it uses my cpu to run for some reason

[indigo carbon]

IPAdapter SDXL goes in custom_nodes/IPA_comfyui/models CLIP vision in models/clip vision

[spring fulcrum]

It uses your cpu to run because like we said last night none of these things can use AMD graphics cards so it reverts to using your CPU

Can I get IPAdapter SDXL on civit or huggingface or github?

[indigo carbon]

Of course you can, but not instead of IPA hooking onto them.. IPA is a new component that goes after the checkpoint

[untold dawn]

oh alr

im still new to this field

[indigo carbon]

Yeah, it's all linked in the IPA node's repo

[spring fulcrum]

cool beans

[spring fulcrum]

Ok so now I have that but it is now saying the following... I don't even see a node where it is referring to copax to be able to change this

[west breach]

that's the checkpoint file it can't find

[uncut fiber]

yes second most frequently error i am getting. 1st is AITemplate enable 😄

[spring fulcrum]

I get that but there isnt a node in the workflow that allows me to select a ckpt

[uncut fiber]

can you send a image? Probably packed

[spring fulcrum]

I can send the workflow... I cant make an image yet because of the errors

[uncut fiber]

i meant screenshot only

[indigo carbon]

Switch the load checkpoint to whatever checkpoint you wish to use

It is under the batch preview

[spring fulcrum]

ohhh... lol its behind the box... I found Waldo... he was hiding

Is the refiner model the one on the top or bottom?

[indigo carbon]

Refiner is the one on the top, yeah

[spring fulcrum]

I have to say great work. Image blend is pretty cool

I'm going to try your AIT V5 in a little bit

[indigo carbon]

Also try the multimodel version, it's Img+txt instead of Img+Img

[spring fulcrum]

Is that a different workflow? can you send me that one too?

[indigo carbon]

I did, it's somewhere in this channel, I also explained what it does

[spring fulcrum]

@indigo carbon Isn't that what I'm using right now?

[ionic dragon]

damn, its great

[spring fulcrum]

thanks

It definately gives some interesting results

[ionic dragon]

can you please share the workflow

[spring fulcrum]

its from @indigo carbon I take no credit for it

I'll post an image and you can drag it into your ComfyUI setup

[ionic dragon]

ok

tq

[spring fulcrum]

np

[ionic dragon]

what is style know as?

[spring fulcrum]

no idea... I would call that lava lamp... lol

[royal hatch]

ei anyone can help me about the licence.. i can't understand if apis have a commercial user or not, because tools in the site are not.. i do not know apis

[spring fulcrum]

try to google it

[ionic dragon]

how would you google images?

[spring fulcrum]

I meant about the license

[uncut fiber]

hole in night skies

it looks like drink with ice cubes. @ionic dragon

[ionic dragon]

@spring fulcrum what is this?

[uncut fiber]

missing node?

[wet nacelle]

https://tenor.com/view/success-100-great-gif-16111463

[spring fulcrum]

idk... I think you need to install missing nodes

[ionic dragon]

i did install all

[uncut fiber]

@ionic dragon It is EMPTY LATENT IMAGE (AITemplate)

[ionic dragon]

it works

i opened it in diff tab

mb

[uncut fiber]

where to put IPadapter?

[uncut gull]

Can SDXL make a person standing in the rain? (Rain falling in front of the person?) Prompt?

[crisp owl]

Sure

[spring fulcrum]

give it a try and find out lol

[stone fossil]

Perhaps prompt for the person to face the camera that might help.

[uncut fiber]

if on me, i tried to help @spring fulcrum

[spring fulcrum]

who needs help?

[hardy cipher]

me

[uncut fiber]

also qwerty_.... was checking for missing node... I told him what node is missing.
I need help, where to put ipadapter models

[spring fulcrum]

if you go to install models on comfy manager you can search ipadapter... click install on the first one and it will put it where it needs to go

[uncut fiber]

i did it but it was missed styl

[spring fulcrum]

@ionic dragon if you get an error when trying to generate an image after you already installed missing nodes. I got something that said PoP at the end... Just search for pop in the install custom nodes and install both pop plugins

[spring fulcrum]

do you have a screenshot of the error

[ionic dragon]

solved

[spring fulcrum]

cool beans

[uncut gull]

Prompt? I'm not at my PC, and I'll never be able to find this message later in you included PNG info.

[crisp owl]

RealVisxlV10 checkpoint

[uncut fiber]

it installed it in ipadapterplus, i find ipadapter comfui

[spring fulcrum]

as long as it works

[uncut gull]

I'll try later, but nothing like this worked for me a month ago.

[crisp owl]

Probably a difference of newer more finetuned models.

[stone fossil]

Tought you ment more like so.

[spring fulcrum]

Is that even a ckpt for SDXL.... If its for SD 1.5 or 2.x you probably will not get results like us..

[crisp owl]

lol yes, it's sdxl.
https://civitai.com/models/139562/realvisxl-v10

[spring fulcrum]

that model looks pretty good

[crisp owl]

Same guy that made 1.5 RealisticVision

[hardy cipher]

massive inbreeder

but they work

[crisp owl]

I loved the work he did with his 1.5 model, and he stated this current model is only about 18% completed.

[hardy cipher]

always liked deliberate and then revanimated better

[uncut fiber]

Is this o.k. result? Credits goes to @indigo carbon

[crisp owl]

I made my own merge for 1.5 of quite a bit of checkpoints, but used deliberate as the main base

[hardy cipher]

realisticvision was always great when I was going for actual realism. but then that's not always my primary focus

[crisp owl]

ah, that's where I stay in a majority of the time

[hardy cipher]

it looks real enough. almost

[hardy cipher]

it's gonna have kitties and string and an octopus of some kind

[hardy cipher]

lol

[stone fossil]

xD

[crisp owl]

oh my

[stone fossil]

Perfect.

[hardy cipher]

some of those balls have eyes. eye balls

[stone fossil]

I no joke for these images use balls in the negative. 😛

So it wont generate to much woolen balls.

[stone fossil]

And even then its russian roulette.

[hardy cipher]

well lit was those yarn balls invading

also, for the longest time it was realyl trying to make cats surrounded by string and yarn

let's see what sort of new mildly cursed image we'll get

oh jeez

[stone fossil]

Awesome.

[indigo carbon]

it's a different version, here; this is what I was talking about

[upbeat summit]

cool - interesting experiments you've been posting the last few days!

[indigo carbon]

this is what multimodel does

[hardy cipher]

had to bring back the frog girls

[hardy cipher]

thanks!. did you see what I'v been doing here? it's got quite a few things going on but they work out alright together

[indigo carbon]

did you just blend cats with wool lol

[hardy cipher]

another moderately cursed image

[upbeat summit]

I tried to follow it a bit and it got me really interested in trying it out myself. I can't claim that I really got all what you are doing - but it does have something to do with multiple IPAdapter passes 😄

[hardy cipher]

is that frog in the back smoking a cigarette?

[hardy cipher]

the main new things I did, at least for me, was mess with pooled output before clip vision, and then also have 2 of my 3 input images go through blip models and use those results as part of the prompt

I just like doing different things to see what happens. sometimes it works out great. sometimes it's horror

[upbeat summit]

crazy and intriguing 🙂

[stone fossil]

I was told it is legal.

[upbeat summit]

yeah - a reoccurring theme with experimentation and the latent space in general heh

[hardy cipher]

but this one kind of works

[uncut fiber]

Now i see what slow gpu means :)))

[indigo carbon]

tf

[hardy cipher]

samurai frogs

[uncut fiber]

Not sure result is o.k.

[spring fulcrum]

I mean considering the images you put in... I think it came out pretty good.

[uncut fiber]

Thanks, just not sure if not publicly bad picture 🙂

[stone fossil]

Woolen animals are the most dangerous species on earth.

[spring fulcrum]

@indigo carbon I'm curious... how did you come up with your username??

[indigo carbon]

long story, I'm sure it doesn't really matter

[spring fulcrum]

its all good.. i was just curious.. no worries

[leaden fable]

hey im trying to install control net models for invokeai/A1111 and dont know where should i download them from, any idea ?

[leaden fable]

what is this ?

[hardy cipher]

input 3 images, get a 4th image

should be good

[vital ermine]

I installed tensorrt but it seems nothing is made for it, shit.

[eternal fog]

Does anyone find when using 2 IP Adapters, it has a habit of chucking the subject of the image off to the side like this?

[hardy cipher]

have you tried the solid masks?

[crisp owl]

yah, tdg was messing with doubling up the ipa nodes, not sure if he fiddled more with it

[eternal fog]

Nah i've not done anything with masking. And it doesn't do this all the time. Just sometimes.

[hardy cipher]

well the masks would help you place concepts

they essentially tell it where to concentrate

[eternal fog]

I'm guessing it's something to do with the input images being different aspect ratios and it not understanding where things should be

[hardy cipher]

probably

[crisp owl]

@indigo carbon you determine if that doubled nodes thing fixed completely the odd images being pushed to the edge?

[leaden fable]

wow, do you have a workflow for this ?

[eternal fog]

If I turn down the strength of the secondary image it stops doing it, which is what makes me think it's the mismatched aspect ratios

[cyan crown]

ipadapter?

or imgtoimg?

[hardy cipher]

both

[hardy cipher]

it's in a lot of the images I post in here. one moment

[uncut fiber]

any idea what second image could be with some good result 🙂

[indigo carbon]

I personally don't like the idea of using img2img when blending, defeats the purpose of making something entirely new by blending 2 subjects, not overlaying 2 images and using IPA for sampling. The fix to subjects going out of frame was to use UNCLIP and IPA in the same time; this is what the workflow I'm working on does

[cyan crown]

ipadapter is only for "copyng" images?

[hardy cipher]

doesn't defeat any purpose, it's just not what you prefer

[indigo carbon]

No, it's only active with the conditioning and the sampling adapter input. It shouldn't have any other effect

[true fjord]

Hey

How do I use Stable?

[uncut fiber]

😄

[eternal fog]

How heavily do you weight it. I've got unclip nodes linked up to both my ipadapter nodes

[eternal fog]

Even better

[uncut fiber]

yes 😄

[indigo carbon]

IPA can also be used for multimodel workflows, including Img+Img and Img+txt equational inputs

[indigo carbon]

I only weight the IPA nodes, I leave both UNCLIPs at 1

[hardy cipher]

I like the idea of swapping images around having them impact the output in different ways. sometimes I do use an empty latent, but that eliminates an option for having another image com into the mix

[eternal fog]

Ah ok

[hardy cipher]

I'm going to mess with model and conditioning data more

[cyan crown]

can you explain better

?

[hardy cipher]

me?

[cyan crown]

both

[indigo carbon]

this is what IPA multimodel is capable of; I won't release neither of these workflow until we figure out how to make AIT compatible with them

[hardy cipher]

well I created a workflow that took blip descriptions of two images and used those as the prompt for thos 2 images to then go through ipadapter and clipvision. worked out pretty well but then I got more black images as output than I really wanted so I messed with the conditioning data going into the clipvision and it seemed to alleviate those issues. then I added styles to the prompt inputs so I could have more flair. and when that started working I decided I'd start using a third image as input rather than an empty latent because I wanted to combine 3 concepts

I will probably add more to it. just messing around

[uncut fiber]

pitty will try make it wider...

[cyan crown]

in Auto1111 ehat is it meant for?

[hardy cipher]

no

but I mean, it works. if you want to drown out the third image which is the input latent then you just bump up denoising. anything below 50-55 percent will basically give you the third image with some mods. but around 65-70 percent it takes on new forms

that's 70 percent denoising

[uncut fiber]

nice

[hardy cipher]

this should be good

first image is described by blip in g positive, second in l positive

[ionic dragon]

Is it worthy to train a lora to just generate acrylic pour images?

[hardy cipher]

if that is your want

hmm

[uncut fiber]

a=b+c

[hardy cipher]

I like the frog girls

[uncut fiber]

artificial noise @stone fossil ?

[hardy cipher]

alright, I have 100+ moderately random combinations of things set to run while I go do real life stuff. this should work out well

I'll do about 8 of each combo. gives me a nice overview of the concept and allows me to weed out the outliers

[ionic dragon]

So would training with only these images get only these
https://pin.it/2t1KEtZ

As I don't want to messup and generate random images other than these

[hardy cipher]

man, there are a lot of variables that come into play

those images seem alright, but you have to set the parameters correctly. I'd suggest tracking down one of the training wizards because I do not know exactly how they do it

[ionic dragon]

Oh ok

[noble shoal]

Uhm, i kinda have a prompt for these

[ionic dragon]

Pls share

[noble shoal]

Ok, but its a bit...weird

[noble shoal]

photo of A living painting, created through a process of slowly injecting colors directly into the artist's veins, allowing them to bleed onto the canvas in a mesmerizing display of organic abstraction. colorful, mixed colors

[lusty wolf]

Space Pirate...

[eternal fog]

You could pop them into IPAdapter as well

[ionic dragon]

It's generating low quality images, my pc is off so I can't use comfy. Used this bot and results were bad

#1100170312106127410 message

[noble shoal]

Wow, that really bad. In comfyui, even with the standard base checkpoint, i get consistent results.

[ionic dragon]

Cool

That's better

[ionic dragon]

Damn, I like those characters

[eternal fog]

I used one of those images in IPAdapter and then used Text Prompts for a person or an animal. It's pretty cool.

[ionic dragon]

Cool tbh

[noble shoal]

Is that on the right what i think it is? 🤔

[zinc cargo]

this one is so good

[hardy cipher]

A genie lamp? Yes, you are correct.

[vital ermine]

I just downloaded ipadapter but the model was for 1.5?

[zinc cargo]

phenomenal cosmic powers.... iittty bitty tiny space...

[noble shoal]

Exactly what i thought. 😬

[hardy cipher]

Hands turned out alright

Dapper gents

[zinc cargo]

@hardy cipher lol

[sleek sun]

Cool tbh

[hardy cipher]

Brilliant

[floral island]

jungle warriors for every occassion!

[uncut gull]

As far as I can tell, SDXL can draw rain falling in front of clothing, but not in front of skin?

sheets of rain, rainfall, torrential downpour, hurricane, flooding, a woman walking in a rainstorm, drenched, soaked, wet hair, clear, background in focus, uhd, 8k, Unreal render
Negative prompt: blur, umbrella, dry, depth of field, lens
Steps: 50, Sampler: DPM++ 3M SDE Exponential, CFG scale: 7, Seed: 2522279332, Size: 1024x1024, Model hash: 0b76532e03, Model: SDXL-CrystalClear, Version: v1.6.0

[hardy cipher]

Frog kings

[uncut gull]

In general, I have trouble getting fine structures "in front of" other things.
E.g., trying to get a bird inside a birdcage is almost impossible. The bars are behind the bird 99% of the time.

[floral island]

or it's clipping through the cage

same with transparancy, that's just damn hard

[uncut gull]

Yeah, maybe it's just a limitation after all.

[floral island]

nah, i have prompt where transparancy works, but that's because it's clothing near the skin

i think it's hard to differentiate 'something behind something'

have some more jungle critters

[uncut gull]

Yes, but the rain will never fall in front of those jungle critters. Only behind them. 🙂

[floral island]

i know 😢

[uncut gull]

I have transparency problems with windows, not sure why you'd want transparent clothes... JK I can imagine, but really. Window clipping is a problem. Lots of scenes have windows.

[floral island]

prompt specific, but it id kinda work

[uncut gull]

It's actually much easier. Just spam words like "transparent chiffon tulle gauze voile" etc. But that transparency doesn't seem to carry over to a flat glass windowpane. Or else glass is just too different.

[floral island]

yeah, you can't really "force" something to be specificly transperant, or we all just haven't figured out the magic tag for it yet

[uncut gull]

Okay. At least I'm not struggling alone. Gonna get dinner and try training some LLMs.

[zinc cargo]

@hardy cipher can you shed some light on how that workflow works?
i mean, i get it that it takes caption from blip, but does it also do IP at the same time?

[hardy cipher]

Well it takes the caption for the first image and sends to clip g positive through the optional styler. Then the same image also goes through ipadapter

Then second image same thing, but through clip l positive and then to clipvision

And before clipvision I mess with conditioning data a bit

[floral island]

upgraded horse

[crisp owl]

I'll take two

[hardy cipher]

Liminal latent space

[floral island]

we might have gone too far 😮

[crisp owl]

it's further evolving

[crisp owl]

Who knew horses were so under-evolved

[vital ermine]

ipadapter just refuses to work for me, but I guess it is because it is wanting 1.5 not XL

[crisp owl]

huh? It works in xl perfectly fine

[vital ermine]

the model I downloaded it directed me to was SD1.5

[crisp owl]

one sec

[vital ermine]

these tensor errors I know is from it being 1.5

[crisp owl]

yeah

[crisp owl]

https://huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/ip-adapter_sdxl.bin

[vital ermine]

You know I spent over six hours today trying to get nvfuser and tensorrt to work for training only to find out #1 no workie in windows and in Linux Kohya never implemented the needed stuff. 😦

[floral island]

less scary horse

[vital ermine]

yeah, mine is 15. that will work. Thanks

[crisp owl]

Yeah it links to the 1.5 one. The links are wonky

[empty osprey]

https://i.imgur.com/ST0861a.png how should i go about mixing this so the "box" disaperes?

[vital ermine]

I saw 15 and instantly knew what was coming, and sure enough, lol

[crisp owl]

yeah you're not the first person lol

[vital ermine]

I am playing with it having no idea wth I am doing as no English speaking/written about it that I could find so I jumped in after the dismal findings that we need new tools as the tech is out there.

deepspeed+dynamo and once we get those it will be so much nicer we could even pull off a 2048x2048 training in the same time as current 1024x1024 on the same card.

LOL

[crisp owl]

Yeah that doesn't help for sure, I threw it in a translater, but it wasn't perfect.

[patent portal]

+1 on this error too - did you solve it somehow?

[vital ermine]

Still I have no idea what it is doing

[crisp owl]

You can leave your prompt empty, or use zerocondition node to clear out your text nodes, if you do that, it mostly just uses the two input images and blends together prominant things from each. You can do two linked together and play with the weights of each.
Or you can do prompts, but they tend to put a lot of weight into the mix, so keep it simple to see more from the images

[cyan crown]

Also William S. Loved pizza

[vital ermine]

I just don't see where it is mixing the images at

[crisp owl]

what's your flow look like?

[vital ermine]

shit

[hardy cipher]

Hmm

[vital ermine]

ngl

[crisp owl]

that's fine, I'm workin gon changing one of mine up atm

spaghetti factory atm

[vital ermine]

I just want simple so A+B

I am used to revision

it sure slowed down my 4090, sheesh

[spring fulcrum]

does anyone know of a really good LoRA for getting perfect eyes or at least eyes that aren't all blasted out?

[vital ermine]

Dang

[crisp owl]

Here's as basic of one as you can get

From these two images

[vital ermine]

cool

how is this different than revision?

[zinc cargo]

@crisp owl is this also blip and then ip?

[crisp owl]

This is just plain out ipadapter

[noble shoal]

That's not what you meant, right?

[crisp owl]

different in my own wording by:

IPA seems to do an actual merge of the two images. Direct images, like the pixels we see.

Revision seems to more grab the concepts from each image and converts those into text (conditioning) which it then uses to make an image.

So from my picture it would take like, girl, sitting, drawing, window, stool, etc.
So in the revision image you'd get stuff with those concepts, but not necessarily how the image itself looks

[vital ermine]

Ahhh, I see

[zinc cargo]

there was a channel on sdxl called img2img, right?

[noble shoal]

Neyes

[vital ermine]

why is the ipadapter only take 1 image?

[crisp owl]

My flow above does two images

But you can do one image also

if you remove the second ipa

[crisp owl]

just chain them

model to model in a second node

[vital ermine]

okay, you lost me

[crisp owl]

split the weights so it's now 1 to 1, but more .5 to .5

[vital ermine]

which node will model/model?

[crisp owl]

the two ipa nodes

[vital ermine]

See, my mind sees that as one way but my way is actually parallel. Problem is all ofthis is in series

[sour obsidian]

eh kinda. Both are using a form of image embedding. The CLIP pooled embedding exists in the same latent space and the text embedding but its not neccesarily a raw text embedding. The base XL model was not tuned with any pooled image embeddings always using the pooled txt embed but because the concepts exist in that same pooled space image concepts transfer across as well. Both methods are in fact using image embeds though, not direct pixels and both likely could be improved quite a good bit

in both cases mixing the images is mixing clip embedding concepts

[floral island]

laughing my ass off xD

[crisp owl]

Oh yeah, I know my description isn't technically right, but in it's workings it's more or less what it's doing and helps to understand how to use the two

[crisp owl]

There, a bit more inline visually

[floral island]

uhh...i think this horse comes from chernobyl or smth

LOL. just... i can't even ...

these are masterpieces!

[vital ermine]

I see the fisherman I had in my prompt

[indigo carbon]

really? I remember KohyaSS having an implementation for DeepSpeed, it does the same thing AIT does

[vital ermine]

not anymore nor dynamo

Now deepseed is actually worthless for me though

due to this

[indigo carbon]

huh, I think going forward we would lean towards OneFlow, that's currently the fastest

[vital ermine]

OneFlow looks promising

[indigo carbon]

performance wise, it has a good 10% over AIT, but idk how easy it is to implement in a user friendly way

[cyan crown]

Napoleon and his Playstation

[vital ermine]

deepspeed cant work with lo rank stuff because it doesn't have the forward it needs. I could not find a json to get deepspeed to work even for db

HF seems to say for 1 gpu it is slower as I showed. Regardless I have yet to see anyone using it with kohya

[indigo carbon]

@visual glade the AIT node seems to have gotten more compatible.. I think you should aim towards implementing OneFlow because currently that's the fastest

[vital ermine]

Now Dynamo there is for the compressing of models and stuff but sadly everyone is bitching at Titon to implement for windows as a ton of pt2 stuff demands it.

[visual glade]

does oneflow even work on windows?

[indigo carbon]

and AIT is already compatible enough, so the goal should be making OneFlow compatible

[vital ermine]

oneflow is for training and inference if I read right

[indigo carbon]

probably, but it is at least 10% faster than AIT

[indigo carbon]

nah, some people benchmarked it on SD1.5 and it was faster than AIT

[vital ermine]

deepspeed's claim to fame is that it can take excessively large models and allow lower vram usage to train them

for me I don't care about anything inference based only as I am looking at the stuff that works for training

too

[indigo carbon]

DeepSpeed isn't flexible or has any superiority to OneFlow, assuming they are both only for Linux(which I'm pretty sure isn't true) OneFlow is superior in any way

[vital ermine]

it is newer too

[indigo carbon]

also the speed is insane actually

[vital ermine]

That is why I would love to see it for training DB but will any of these work with lora types

[indigo carbon]

I think it has things in common with both AIT and TRT, didn't look into it yet

[vital ermine]

A bit sad I got it all installed but no way to make it work on windows and in linux it demanded stuff kohya had not implemented. I tried though.

that was for TRT

[indigo carbon]

OneFlow seems like the way going forward.. I heard it's also flexible

[vital ermine]

for linux it complained it wanted fake tensors and I need to implement them. beats me I didn't write kohya code so no idea

spit out a bunch of matrix stuff at me too

[indigo carbon]

OneFlow seems to take all the good properties from AIT and TRT, making for what could be the fastest inference as of now

[vital ermine]

yes, but did I read it right that it can be used for training as well?

some of these are one or the other

[indigo carbon]

I don't see any reason for it not to.. give it some time; I'm sure it will be an accessible thing soon, so far its performance speeks for itself

[vital ermine]

the best kind is the kind that doesn't need devs refactoring.

[indigo carbon]

I think I heard SAI used it for training SDXL or for the bots and ClipDrop

[vital ermine]

Probably true

Right now we need deepspeed for DB because even my 4090 BS2 is it.

so something to help with mem optimization

[indigo carbon]

I feel like OneFlow would be the solution.. we just need someone like Comfy to implement it in an *it just works™️ * way

[vital ermine]

I wonder why accelerate doesn't do it yet, or even talking about it?

[indigo carbon]

I looked in the repo, it does say they didn't pre-build it for windows yet, but it should be able to be built for windows

[floral island]

hope some of that stuff gets ported over to a1111 some day

[indigo carbon]

@visual glade so if you build the module from source, it should act the same way AIT does, you can't just go pip install oneflow you will need to do git clone https://github.com/Oneflow-Inc/oneflow then do python setup.py bdist_wheel or something like that.. basically same process when making an environment for compiling AIT modules

[indigo carbon]

getting OneFlow to be an easy use for ComfyUI will be a huge step in that direction for sure, but we gotta start from somewhere..

[floral island]

yeah, i think the modular setup on comfy should make initial implementation a lot easier

[vital ermine]

Yeah, for TRT I had to jump through a lot of hoops but on linux pip install and 3 mins later done

[spring fulcrum]

@indigo carbon Do you have any interest in adding a high res fix... The only two things I think it could use to make AIT my main workflow would be a slot for a LoRA and a high res fix.... Or something that can fix the eyes... the eyes are the only part that consistently don't come out quite right. Other than that your workflow is amazing... and quick.. I really like it a lot... I especially like the style selector

my only small critic would be to label the nodes.... have them specifically say this is your refiner model... this is for base model... this sampler is base image.. this is for refiner... It just makes it easier for people that are new to this to be able to use it easier without having to ask questions.

[indigo carbon]

I might add those to future versions, but the version I shared earlier today is a prototype. It's not a final version

[hardy cipher]

this ima

[spring fulcrum]

I really like it a lot... great work... that was just my two cents

[hardy cipher]

I really need to take the time to get ait working properly. but not just yet

[shy kelp]

mhh what would be better, 3090 or 4070 ti ?

[indigo carbon]

I'm not quite happy with the speeds yet. AIT is very nice to have but OneFlow is faster and newer, next time I work on new workflows after I finish the AIT one it will likely be already after everything gets figured out

[hardy cipher]

where are you learning about all these new tools?

[indigo carbon]

OneFlow should be compatible with IPA theoretically, so I won't be uploading any of the multimodel workflows officially until it's optimized with such optimization

[indigo carbon]

but I will probably add a LoRA feature to the main workflow like @spring fulcrum said before I go silent

[spring fulcrum]

That would be my biggest ask and would make this my main workflow at that point

[vital ermine]

some of these comfy nodes has a mem leak

[spring fulcrum]

I am getting fantastic results with AIT... Thanks @indigo carbon

[vital ermine]

no idea which one

I think I read it is pytorch itself

Damn, I restarted and got hit with the same OOM

[eternal fog]

You trying to merge models together?

[vital ermine]

no, I stuck a lora in and boom

[hardy cipher]

which nodes are processing conditioning data?

[eternal fog]

I've not used LoRAs for a while, but not had issues with them before

[brittle plover]

hello everyone, i have tried using the sdxl model, but when i want to make a simple image it's loading very slowly. is that normal? my VRAM is 8 GB and RAm is 16GB

[vital ermine]

Yeah, I switched the lora off and it ran

6 gigs free, so weird

[hardy cipher]

ipadapter + clipvision are basically on the fly loras for me

[vital ermine]

lacks a lot of control as you never know what is gonna come out

[hardy cipher]

well you can gain control of it with the right approach

[vital ermine]

I think I had a crossed wiring is why but all this wire it gets confusing

[hardy cipher]

what's your flow look like?

[vital ermine]

nope, this lora is not gonna work in this flow

like utter shit. I am never organized in comfy

just too hectic for me

Never did like spagehtti jungles.

[hardy cipher]

when it gets that way for me I try to organize things into sections

[eternal fog]

I've tried to branch mine out into sections

Still gets annoying sometimes when I have to move a link somewhere and then it doesn't connect and I lose where it went lol

[hardy cipher]

you have laughing goofed tooth hispanic guy

very nice

[eternal fog]

https://tenor.com/view/champoy-el-risitas-kek-issou-etu-gif-17837830

[hardy cipher]

he's so happy

[eternal fog]

I hear the laugh, even when it's only a gif

[vital ermine]

Why am I getting this ait error and ait is no where in this workflow?

[hardy cipher]

ipadapter is so good at taking low quality images and outputting high quality things

man, AIT is cursed when I try it

[vital ermine]

I no longer use ait so surprised

[hardy cipher]

3 times now I've installed it, and 3 times I get errors like that. not that error but similar vein. the latest round it was giving me errors even when I wasn't using the nodes

[noble shoal]

I think it was this: https://github.com/FizzleDorf/AIT/pull/44

[vital ermine]

hmmm, even when it has no ait node ?

it is weird because the workflow works without but with boom.

[noble shoal]

Currentls comfy + AIT is only running with this patch. As far as i know

[vital ermine]

none of my nodes tells me I need to update. welp, hmmmm

[noble shoal]

Its not merged

You need to manually patch

[vital ermine]

Alright. Still confused why I am not using AIT anywhere in my workflow but got it. must be always running but not activated or something

go into his folder and git pull?

or if not merged how do I update?

[noble shoal]

I think you need this commit: https://github.com/FizzleDorf/AIT/tree/ab84fac064b6fbf16b8431f8b7a34394ae72f02f

[vital ermine]

I just hit the folder and git pull

[noble shoal]

Ok, it may work now

[vital ermine]

let's see

[vital ermine]

nope

apparently I need a 32gb card

I see the error message is is a recursion issue

not sure where or how though

[glad grove]

time to buy that sweet h100

[hardy cipher]

hmm

[noble shoal]

Can you post the complete traceback? I've run into a lot of errors, maybe it's something that looks familiar

[crisp owl]

RecursionError: maximum recursion depth exceeded while calling a Python object
heh oops

[noble shoal]

What workflow are you on? Does it also happen if you have a simple Model -> AIT -> Sampler barebone workflow?

[vital ermine]

that is the point I have no ait nodes as I no longer touch them for a couple of weeks now

[noble shoal]

Oh, so, lets say you deactivate (or move) the AIT nodes for a second. Does it still break?

[vital ermine]

in a min I will try that

[hardy cipher]

I do not know, but it's saying you need at least two 4090s for that workflow, lol

perhaps an h100, or even an a100

[vital ermine]

yeah, just by sticking in a lora

[noble shoal]

WTF is that workflow?

[hardy cipher]

there's something real wacky going on either way. because it shouldn't be requiring anything remotely close to that should it?

[vital ermine]

nothing that would cause this. I will move ait in a min

no

[hardy cipher]

unless you're trying to train 2 or 3 loras simultaneously or something

[vital ermine]

LOL

[hardy cipher]

it'd work on runpod. you can rent 8 h100s simultaneously

[vital ermine]

Might as well just pull down my pants for that one

[hardy cipher]

about 35 dollar an hour. but you'd have like 640 gb vram

[vital ermine]

I moved AIT and still crash

[hardy cipher]

yeah, I uninstalled it through the manager. if completely wiping it doesn't work you might want to start replacing the files that are causing the errors? I don't know. when I see that sort of thing and it's not easily rectified I start purging immediately

[noble shoal]

Is this your workflow?

[hardy cipher]

10 checkpoints, lol

[vital ermine]

this is the problem. turn that lora (any lora) off and it works, or remove it

whoa

it is the lora

I bet I know why

[noble shoal]

🤯

[hardy cipher]

throw it in a fire

[vital ermine]

throw comfyui in a fire

[noble shoal]

Is it a 50gb lora?

[vital ermine]

heheheh, no. I am trying different loras and they are all doing it

[noble shoal]

Ok, throw the lora loader into the fire

[vital ermine]

yeah, let me try a different one

[hardy cipher]

I made lora loaders

[vital ermine]

nope

native lora loader same issue

something about lora and ipadapter

all I am doing is taking the model and coming out into the lora and from it going onward as shown

works without ipadapter that way

I wonder if it needs to go after the ipadapter?

[hardy cipher]

which clipvision model are you using for ipadapter?

[vital ermine]

https://huggingface.co/h94/IP-Adapter/blob/main/sdxl_models/ip-adapter_sdxl.bin

and clipvision I dunno

[hardy cipher]

you need to use the right clipvision model. I don't know that it'd give you that error though

[hardy cipher]

but could have something to do with it maybe?

[lofty valve]

/prompt a boy in office

[hardy cipher]

this is what I use . not sure what that one is

might be the same, not sure

[vital ermine]

I dunno it was a link someone in here gave me on the HF site

probably the same

have a link?

[hardy cipher]

it's off huggingface. one minute

[vital ermine]

I bet the same one

[hardy cipher]

you know, I think I got it from the manager models actually

it's not in my history

[noble shoal]

Isn't that the clip vision g.safetensors from the stability ai huggingface?

https://huggingface.co/stabilityai/control-lora/blob/main/revision/clip_vision_g.safetensors

[heady vale]

https://huggingface.co/h94/IP-Adapter/tree/main/sdxl_models/image_encoder

Added the dragon to one of my castle landscapes. worked out nicely

[vital ermine]

whoa, wth?

that ip-adapter is 3.5GB

[heady vale]

yep

[vital ermine]

*3.69

is that the one cause I had 1.5 then earlier given link to SDXL now this one

use it instead? grabbing the safetensors one

[noble shoal]

I think it's just the clip vision model

[vital ermine]

then I dunno. needs to be one place to get this stuff

[heady vale]

use the link I pasted

[vital ermine]

I am

how about the ip-adapter? confused cause clip vision but link is to 3.69gb ip-adapter model

[vital ermine]

alright, that is the one I have

[heady vale]

I dont have that one

[vital ermine]

Been a while since I got it but it came from HF

hopefully I didn't just automatic1111 myself with that update

[hardy cipher]

correction, it was mc dali

[native knot]

I've literally just finished up a script and workflow change that allows my Randomizer to drop the files into a directory that's monitored by my Lychee server and it automatically adds the images into my Random Diffusions gallery, archives the images after being scraped. Then it periodically prunes the archive and log files. This way, when I set the Randomizer to create images overnight, all I have to do is just look in the gallery instead of manually uploading everything.

[hardy cipher]

nice

[native knot]

Gallery is here for anyone who doesn't already know:
https://lychee.soulctcher.net/

[hardy cipher]

some of these old ai images I saved are kind of 🔥

[native knot]

Oh...I think I still need that one node from you.

Two nodes, actually.

[native knot]

ConditioningNormalizerPoP and ConditioningMultiplierPoP

I see what you're doing outside of those, though.

[hardy cipher]

one second

[native knot]

word

Metalephant

[hardy cipher]

they appear to be on the github. are they not working correctly?

if so please let me know because I don't want that to be the case

[native knot]

I just updated a few minutes ago and am getting red nodes for those.

[hardy cipher]

hmm

let me see

[native knot]

Let me triple check.

[hardy cipher]

oooh

I know what it is

[native knot]

Yeah, they're not copasetic.

[hardy cipher]

I changed the file name for them before I uploaded them

[native knot]

😄

[native knot]

That'd do it.

[hardy cipher]

so if you have one of the old images

[soft bone]

has anyone used "MediaPipe Facemesh to SEGS" node from impact in a facedetailer workflow? doesnt seem to work:(

[hardy cipher]

try that one

[native knot]

Ok, that loaded.

Danke

[hardy cipher]

changed the spelling of the base file

so broke the old ones

[native knot]

Let me... dissect this one.

😉

[hardy cipher]

you know, I can't say if it's optimal, but it works

[native knot]

Yeah, no worries...I like to play around with the stuff, too.

[hardy cipher]

thinking of messing with the model data in a similar way

also, if you want to bump up the multiplier max it's in the conditioning node file

I just set it at 3 because past that it gets very off

[native knot]

ok

[hardy cipher]

and negative seems to be nonsensical and broken. but who knows, might be a use case

[native knot]

First thing's first...pull the workflow out to trace everything.

[hardy cipher]

you know, I didn't know anything about any of this stuff a few weeks ago. so it's quite possible I didn't do everything completely optimally. so please let me know if you see anything that seems off

[native knot]

It's all good...I'm around the same place you are.

I have been messing with this in different ways for like a year, but only recently got into Comfy

[hardy cipher]

well I started with ai image stuff a couple years ago. but not text to image. that was future technology

but found it pretty quickly when it became available to the public, then SD in november along with a1111. but didn't start with comfy until around the end of july

learned a lot about python with a1111 because of how often it broke

[native knot]

Python is great.

super useful

[hardy cipher]

hadn't really touched it until then. but learned quite a bit sense then

at first it was out of sheer necessity, but now I can actually do stuff with it

trash image, lol

[native knot]

why can't I easily move the textDebug nodes? They're not locked, not pinned, not frozen...

[heady vale]

I wish Image Load didnt constantly make duplicates when you reuse the same image

[native knot]

Oh, I see...they were collapsed, but it was tough to expand them due to their want to show the true/false switch. got it.

[hardy cipher]

I think they're broken. and I think you can delete them

yeah, I don't really think they're necessary

[vital ermine]

I started over and it works as it did earlier but as soon as I add the refiner it will break

yep, breaks it

[native knot]

@hardy cipher - Your workflow spread out...now I can start to follow this version. 🙂

[hardy cipher]

nice

yeah, I had it scrunched up because I knew what everything was doing

[native knot]

I had to pull it apart for me to begin.

😄

Now I can follow the noooodz.

[dim steeple]

hey guys where can i go to download the tile controlnet?

[heady vale]

hasnt been made for SDXL yet

[dim steeple]

oh sorry i meant for 1.5

is this it?

https://huggingface.co/lllyasviel/control_v11f1e_sd15_tile/tree/main

I don't understand how I download it

I need a yaml too?

[native knot]

@hardy cipher

=

[hardy cipher]

damn

[glad grove]

i like your balls

pictures

[native knot]

lul

[hardy cipher]

set runpod up so I could do some rapid fire

[hardy cipher]

hmm

[crisp owl]

Backread a bit. @vital ermine if you don't have it, this is the clipvision model you need
https://huggingface.co/comfyanonymous/clip_vision_g/blob/main/clip_vision_g.safetensors

[uncut fiber]

using @hardy cipher WF

[crisp owl]

sexy

[dim steeple]

Is there a SDXL controlnet model for lineart?

[crisp owl]

scribble probbaly closest

oh the controllora is called sketch

https://huggingface.co/stabilityai/control-lora/blob/main/control-LoRAs-rank256/control-lora-sketch-rank256.safetensors

but don't think it makes lineart

haven't played much with it

[hardy cipher]

I don't think it hit that last image yet. I queued a few

[shy kelp]

cant figure out how to control weights, something stronger with 1.4 seems fine but doing something like 0.4 or 0.6 doesn't want to work

[hardy cipher]

in what?

[shy kelp]

a1

[static wigeon]

Hey there people! I am very new to Stable Diffusion and I am working on it with a very tight time deadline. I am trying to generate some images in a comic style. I am using a comic-styled LORA which works perfectly. The issue I am having is that the models I use which is "https://civitai.com/models/17800/cartoon-fanart-style" is generating the images but it is messing up the background characters. I am using Euler A Sampling method, 20 Steps, a CFG Scale of 7 and a random seed that I generate. This is my negative prompt:

nudity, text bubbles, dialogues, disfigured faces, low quality output, text, username, logo, (low quality, worst quality:1.4), (bad anatomy), (inaccurate limb:1.2), bad composition, inaccurate eyes, extra digit, fewer digits, (extra arms:1.2)

But the outputs of the background panels look a bit too disfigured. Am I doing something wrong? Does seeds actually matter? I would really appreciate some help

[shy kelp]

usually if i do like freckles:0.4 or 0.6 or anything really under 1, it's softer to non-existant but i can't seem to get it to work here

[hardy cipher]

sometimes the dropoff can be quick with stuff like that. you just have to experiment with things because every scenario is different

[hardy cipher]

try more steps and possibly lower cfg scale

[shy kelp]

ty

[ionic gulch]

can someone with access to an unlimited chat bot (local or payed chat gpt) help me create a json file with 250 movies? the free chat gpt won't give me more than 10-15 movies? i would provide more details for the prompt.

[static wigeon]

Ohh, any recommended values or should I just experiment? And I just started using aDetailer too, it is helping but it isn't allowing me to generate multiple characters

[hardy cipher]

hmm

[hardy cipher]

I could help out probably. just give me the criteria for what you're looking for

[ionic gulch]

@hardy cipher this prompt works for me in general, it just limits the output too much on the free version:
please create a list of 250 movies in the following json format. use movies that are known for their unique art style.
[
{
"name": "Blade Runner",
"prompt": "{prompt} in the style of 'Blade Runner' by Ridley Scott, cinematography by Jordan Cronenweth",
"negative_prompt": ""
},
{
"name": "The Grand Budapest Hotel",
"prompt": "{prompt} in the style of 'The Grand Budapest Hotel' by Wes Anderson, cinematography by Robert D. Yeoman",
"negative_prompt": ""
},
{
"name": "Pan's Labyrinth",
"prompt": "{prompt} in the style of 'Pan's Labyrinth' by Guillermo del Toro, cinematography by Guillermo Navarro",
"negative_prompt": ""
}
]

[hardy cipher]

I'm quite adept with the LLMs. give me a minute. probably won't be able to do 250 in one response. but shouldn't be too difficult

[ionic gulch]

if you can sort it in alphabetic order it would be even better, but i've noticed that will reduce the likelyhood of a long list even more

[hardy cipher]

well, worst case scenario they'll only be in order for each response

because hassle to get it to put 250 things in order. gpt likes to omit things and not say anything about it

I'll have it put everything into codeblocks for convenience

[dense chasm]

python's aria2c library runs more faster than wget or request, 5 Gigabytes safetensor or cptk file only takes several secs downloading,upload is a bit slower,around dozens of secs

[hardy cipher]

got it rolling. I have a sneaking suspicion it's not going to count very well though

[ionic gulch]

if it's more than 100, so i can have at least 100 after filtering out some bad movies or duplicates, it would be a start and good enough to release the new workflow

[hardy cipher]

having it list all the movies previously added so it won't repeat itself, hopefully

[ionic gulch]

the new workflow will look like that (prompt/style section):

[hardy cipher]

2001: A Space Odyssey, A Clockwork Orange, Akira, Amélie, Avatar, Baraka, Birdman, Blade Runner 2049, Brazil, Children of Men, Coraline, Donnie Darko, Drive, Eternal Sunshine of the Spotless Mind, Fantastic Mr. Fox, Fight Club, Ghost in the Shell, Hero, Inception, In the Mood for Love, Interstellar, La La Land, Mad Max: Fury Road, Moonlight, Mulholland Drive, No Country for Old Men, Oldboy, Parasite, Pulp Fiction, Requiem for a Dream, Roma, Schindler's List, Scott Pilgrim vs. the World, Seven Samurai, Sin City, Spirited Away, The Big Lebowski, The Dark Knight, The Fall, The Fountain, The Godfather, The Great Beauty, The Matrix, The Revenant, The Royal Tenenbaums, The Shape of Water, The Shining, The Social Network, The Tree of Life, The Truman Show, The Witch, There Will Be Blood, Trainspotting, Uncut Gems, V for Vendetta, Vertigo, Waking Life, Wall-E, Whiplash, Wings of Desire, Y Tu Mamá También, Your Name, Zodiac

first run through the alphabet. it's still going

I just dont' know how many are actually listed there and if it's lying to me

[ionic gulch]

@hardy cipher so you do it in two steps? first the movies in alphabetical order and then creating the json file?

[hardy cipher]

no

it forgets after so many tokens

but it doesn't tell you

so it'll start repeating things without realizing what's happening

so it's a way to keep it fresh in the memory

[ionic gulch]

lol the two step approach worked. i've used your list and asked for the json and it did it in the free version:

so this could actually work. get an alphabetic list of as many movies as possible and than ask for the json file :)

[hardy cipher]

well I'm still making my list, lol

[ionic gulch]

sure, if you can get more movies it would be even better. just happy that i have already more than the 15 movies it did before

[uncut fiber]

nice movies

[dim steeple]

Is there a SDXL controlnet model for openpose and openface?

[hardy cipher]

maybe

[crisp owl]

https://huggingface.co/thibaud/controlnet-openpose-sdxl-1.0/tree/main

[dim steeple]

thanks

but no lineart

[floral island]

*Muted voice* "I'm in"

[crisp owl]

not that i've seen, no true lineart stuff for controlnet, yet.

[dim steeple]

I am trying to use this workflow but in XL: https://www.youtube.com/watch?v=cEnKLyodsWA&ab_channel=DigitalMagic

[glad grove]

we will get all of the controlnet models for sdxl after they release the 2.1 controlnet models

[dim steeple]

sorry which one is it?

the 5GB one?

[hardy cipher]

https://pastebin.com/raw/FAHU6W5A

might be repeats. didn't really verify for sure

[crisp owl]

The 256 version should be plenty good enough.

[hardy cipher]

2001: A Space Odyssey, A Beautiful Mind, A Clockwork Orange, A Ghost Story, A Scanner Darkly, A Separation, A Single Man, Adaptation, Akira, Alien, American Beauty, American Psycho, Amélie, Annie Hall, Apocalypse Now, Arrival, Avatar, Baby Driver, Baraka, Beasts of the Southern Wild, Being John Malkovich, Big Fish, Birdman, Birds of Passage, Blade Runner 2049, Black Swan, Blow-Up, Blue Velvet, Boogie Nights, Boyhood, Brazil, Breaking the Waves, Caché, Call Me by Your Name, Capote, Casino, Children of Men, Chinatown, City of God, Clerks, Close-Up, Coco, Come and See, Coraline, Crouching Tiger, Hidden Dragon, Dancer in the Dark, Darjeeling Limited, Das Boot, Dead Man, Delicatessen, Departures, Django Unchained, Dogville, Don't Look Now, Donnie Darko, Drive, Dunkirk, Eternal Sunshine of the Spotless Mind, Eyes Wide Shut, Fantastic Mr. Fox, Fargo, Fear and Loathing in Las Vegas, Fellini's 8½, Fight Club, Forrest Gump, Full Metal Jacket, Gangs of New York, Garden State, Gattaca, Get Out, Ghost in the Shell, Gladiator, Goodfellas, Grand Illusion, Grave of the Fireflies, Heat, Her, Hidden Figures, House of Flying Daggers, Howl's Moving Castle, 

I Am Love, In Bruges, In the Name of the Father, Inglourious Basterds, Inside Out, Into the Wild, Irreversible, It Follows, Jackie, Jaws, Joker, Koyaanisqatsi, La Dolce Vita, La La Land, Labyrinth, Lady Bird, Lawrence of Arabia, Le Samouraï, Life of Pi, Lost in Translation, Mad Max: Fury Road, Magnolia, Malcolm X, Mandy, Manhattan, Marie Antoinette, Melancholia, Memories of Murder, Metropolis, Midnight in Paris, Midsommar, Million Dollar Baby, Minority Report, Moon, Moonlight, Moulin Rouge!, Mulholland Drive, My Neighbor Totoro, Natural Born Killers, No Country for Old Men, Oldboy, Once Upon a Time in America, Once Upon a Time in the West, One Flew Over the Cuckoo's Nest, Only God Forgives, Only Lovers Left Alive, Orpheus, Pan's Labyrinth, Parasite, Paris, Texas, Paths of Glory, Persepolis, Persona, Pi, Pink Floyd: The Wall, Pirates of the Caribbean: The Curse of the Black Pearl, Platoon, Pleasantville, Ponyo, Princess Mononoke, Pulp Fiction, Raging Bull, Ran, Rashomon, Requiem for a Dream, Reservoir Dogs

[dim steeple]

what is the difference?

[crisp owl]

I think it may be difference of pruned vs unpruned...in a sense. However that works with Lora's.
Kinda like how the StabilityAI controlLora models have both the 256 and 128 models

256 will perform better than 128, but take more memory and processing power

[floral island]

i think that coffee might be a bit hot

[ionic gulch]

thanks. i will check it and clean it up after work.

[ionic gulch]

@hardy cipher if you want to check out the new workflow. it's already uploaded to github (based on the first list of movies you gave me):
https://i.imgur.com/Q5EImJB.jpg
https://github.com/JPS-GER/JPS-ComfyUI-Workflows

you can mix movie styles and general styles (sai styles included) with everything else. (text prompt in the example screenshot from above)

example (ip adapter + text prompt + styler):

[static wigeon]

You made this using Stable or?

[ionic gulch]

@static wigeon this is stable diffusion with comfyui

[static wigeon]

Ahh I see

[ionic gulch]

everything in here should be stable diffusion somehow. some people use comfyui, some a1111 or sd.next and a few use other solutions linke invoke.

[static wigeon]

I see

I am new to Stable Diffusion so I am not familiar with anything related to it 😭

Been using it only for 2-3 days now. I normally use Midjourney, playground or gencraft. So yeahhh. Still figuring things out.

[ionic gulch]

@static wigeon do you use comfyui or a1111? if you use comfyui you can use the workflow i've provided above.

[ionic gulch]

mad max, blade runner, your name (with the same prompt and ip adapter image):

[hardy cipher]

alright, now I can try out the movie game

[static wigeon]

I use a1111

But I did save your git link just in case I decide to use ComfyUI

I am getting this error "RuntimeError: Not enough memory, use lower resolution (max approx. 320x320). Need: 0.0GB free, Have:0.0GB free" in a1111, what does it mean?

[hardy cipher]

means you have no vram available

[static wigeon]

Oh

[hardy cipher]

do you have vram?

[static wigeon]

I don't know, this is my work laptop so 😭

But it was working till now

[hardy cipher]

well you can still render images in comfy without vram. might take 20 minutes each though

[ionic gulch]

without using sai-cinematic it works even better:
your name, blade runner, akira, mad max

[static wigeon]

Ohh

[wide crane]

there is just something i like about heterochromia

idk

[static wigeon]

I think it is using my IGPU

How do I make it use my Nvidia gpu

[hardy cipher]

did you install this? https://pytorch.org/docs/stable/cuda.html

[shy kelp]

accidental discoveries today

[ionic gulch]

different ip adapter image (still mad max):

result / ip adapter image (mad max):

akira:

[hardy cipher]

where is this wildcard thing? I replaced the seed randomizer and now it's broken

[ionic gulch]

akira:

[supple knot]

@JPS Where do I find the ClipVisionLoader File? for your workflow. I need help setting this up

[shy kelp]

this new coffee joint aint it man

[supple knot]

you get your coffee there?

[shy kelp]

not anymore! 😭

[brittle vortex]

How can I keep the original shape unchanged and turn the picture into the elements of black and white chess pieces?

[crisp owl]

probably this one.
https://huggingface.co/comfyanonymous/clip_vision_g/blob/main/clip_vision_g.safetensors

[noble shoal]

I think it's easier, for instance with bing, to ask for a list of the top 50 IMDb movies with director and cinematographer and then write a python script to make a json out of it. They tend to screw up json very fast.

[floral island]

sexy skeleton/zombie is imparting us with great knowledge

[ionic gulch]

the links should be in the black info text box:
IP-Adapter Models:
https://huggingface.co/h94/IP-Adapter/tree/main/sdxl_models
https://huggingface.co/h94/IP-Adapter/blob/main/models/image_encoder/pytorch_model.bin

if you need additional help i will be back in a few hours - have to make some calls now

[supple knot]

Thanks @crisp owl @JPS Its running , But T2i is still confusing to me. I haven't been using it and I got file all over the place trying to do it

[floral island]

anime sd goddes be like "i died for your sins"

[noble shoal]

🤨

[strange mist]

is that a biblically accurate angel? 👀

[floral island]

I have no idea 😮 but uhh... this was an interesting prompt for sure 😮

i wonder what it woudl take to actually try to generate something like that with an actual sensical prompt

[strange mist]

looks cool 😄

[floral island]

the actual prompt(dynamic prompts):

As per request, the unabridged repeat, the answer to My understanding of the truth as it pertains to you We are the last of the fated the lost and unseen We reject whatever we were supposed to be
| All eyes on me, gather all who believe Though far apart, united by heart
| We cast increasingly irrelevant designs aside Make room for our incendiary cultural divide We are the last of the fated the lost and unseen We reject whatever we were supposed to be
| It is a cause for us to sing It's how we know that we will win It's true, you wouldn't know 'Cause it cannot be taught or shown There are no gates to hold us back And all your words are only that And this is who we are
| All eyes on me, gather all who believe Though far apart, united by heart All eyes on me, gather all who believe (We rise to reset the world) Though far apart, united by heart (And fight forever) All eyes on me, gather all who believe (We rise to reset the world) Though far apart, united by heart (And fight forever)
} BREAK [anime cell shading:0.2]```

[strange mist]

that prompt 😵‍💫

[floral island]

i mean, you get the most awesome stuff doing that (different prompt)

oh, and thx mods, for removing that irritating dude

[strange mist]

you've convinced me. I'll run some wall of text prompts and see what comes up 🫡

[floral island]

important detail: try to keep the "resets" to 3 maximum

tokens are sent in batches of 75, and each group of 75 becomes a weighted average

so anything above 3 groups becomes incredibly diluted and won't work very well

[static wigeon]

Yeah I did

[floral island]

that's the {2-3$$ BREAK} i'm using: get 2 or 3 lines, then split them in their own groups using BREAK, and then a final BREAK to get anime flavour

[strange mist]

so not more than 3 BREAKs? 🤔

[floral island]

yeah, try not to go above, because then the prompt becomes pretty hard to steer actually

and if you get a combination of lines that's really cool in results, you want to be able to steer it a little

[static wigeon]

Wait I am not sure Nvidia one I am showing and the one you linked me to are the same

[floral island]

right now i'm going through my incredibly long backlog of lyrics ialready once did, and will merge them in a HUGE single propmt

so the prompt will be like probably 70 kilobytes or more

but from that huge amount of lines, just select 2-3 random ones

should make interesting result 😄

[strange mist]

are you using an anime checkpoint? or the effect is from anime cell shading?

[floral island]

anime cell shading

it's protovision: a non-anime model

i'm using [anime cell shading:0.2] so it isn't there for the first 20% of the steps, else it'll contaminate the image with "generic anime shit" which i don't want -> i want it to add anime flavour wwithout influencing the initial result

[strange mist]

I'm loving the effect it's producing 😮

[floral island]

This is why i have a hard time giving up lyrics prompting 😛 (this was throught the fire and the flames)

[static wigeon]

How are your generations so flawless

Mine always has one issue or the other

[floral island]

uhh... having a great model i suppose? this model is pretty smart to be honest

i fell in love with it the moment i started using it

[shy kelp]

hmm, these aint too bad

[floral island]

lovely!

[static wigeon]

Oh I see, may I know its name? And can it generate cartoon/comic style images?

[floral island]

protovision

[floral island]

sure it can, it might need some prompting work

[static wigeon]

This is one of my generations, I need it like this

Ahh I see

I am weak with prompts so I will try

Will give it a shot!

[static wigeon]

The third one looks cool

[floral island]

most important with testing prompts is figuring out what tokens the model actually recognizes

sometimes you're trying to achieve something, but the model simply doens't know what to do with the requested token

for example, this model will not ever produce a sword slash

perhaps it can if i indirectly prompt for it, but prompting 'slash' will only get me a guitarist or a static sword

[static wigeon]

Oh I see

https://civitai.com/models/125703/p This is the model right?

[floral island]

yeah, that's the one

[static wigeon]

Okayy thank youu