✨｜sdxl

[hardy cipher]

does he have the answers?

[strong field]

no, thats why he's laughing

nobody knows how this works anymore !

[ionic gulch]

still the same result for:
https://i.imgur.com/WMWVgOD.png
I've changed input_types to cls, and did the int(largest)
imo this is some bug, if you only have one output value - because it changes behavior if you add a second one

[hardy cipher]

I always avoid the youtube links. maybe I shouldn't. I turned off the video link search results in google. and then it just showed me more youtube videos

no hate on the youtubers trying to make a buck, but it's so tiresome

[strong field]

2x speed on those videos

[hardy cipher]

you know, I was doing that for a while and then forgot about it, lol

but thanks for reminding me

[ionic gulch]

so if you add the same value twice, it doesn't create those 3 wrong outputs:
https://i.imgur.com/2T7gv5a.png
that's why i think this is a bug in comfyui

[strong field]

This is great info jps

[hardy cipher]

I just want to know if it somehow split INT up into 3 outputs, or if it's something else entirely

[ionic gulch]

the output still doesn't work for some other reason. i always get 1024 instead of the larger one of the two inputs.

[vast galleon]

class JPS_Math_LargestInt:

def init(self):
    pass

@classmethod
def INPUT_TYPES(cls):
    return {
        "required": {
            "value_a": ("INT", {"default": 1,}),
            "value_b": ("INT", {"default": 1,}),
        }
    }
RETURN_TYPES = ("INT", )
RETURN_NAMES = ("largest", )
FUNCTION = "get_lrg"

CATEGORY="JPS Nodes/Math"

def get_lrg(self,value_a,value_b):
    largest = int(1024)
    if value_a >= value_b:
        largest = value_a
    if value_b < value_a:
        largest = value_b

    return(int(largest), )

[ionic gulch]

@vast galleon what did you change? does it work for you if you input two int values? does it give you the larger one as output?

[vast galleon]

RETURN_TYPES = ("INT", )
RETURN_NAMES = ("largest", )

the commas

[vast galleon]

trying to test it. what's the best node to hook it up to and see?

[ionic gulch]

@vast galleon still outputs 1024 all the time (but display is fine now - only one output) - i use it for height and width values and want to get the higher one of those two values, but get 1024 instead - so i guess the if part is wrong too

also had if / else before, with the same result

[hardy cipher]

my brain is slow and has no idea what you're trying to do, lol

[vast galleon]

try

def get_lrg(self,value_a,value_b):
    largest = int(1024)
    
    largest = value_b

    if value_a >= value_b:
        largest = value_a


    return(int(largest), )

[hardy cipher]

I should put together some of the math nodes so you can put in any value for height or width and it'll output what the other one should be

[strong field]

sometimes the refiner makes me mad

[hardy cipher]

I'm experimenting with running the same lora settings on refiner as I am base

also, I've just been sending both my G and L prompts to it. combining them

because I have no idea what I'm actually supposed to do

[strong field]

https://tenor.com/view/oh-no-baby-what-is-you-doing-what-are-you-doing-confused-gif-16262696

sometimes im just like

[hardy cipher]

https://tenor.com/view/bubb-rubb-whistletips-gif-3559209

that's how I feel when it actually works

[hasty smelt]

@wet nacelle thanks for the prompt!

[wet nacelle]

Wicked

[hardy cipher]

very nice! I'm actually going for a moderately similar look with what I'm putting together

haven't got there yet. messing with a new setup

I don't understand these people that dislike the noodles. it's like playing with ai legos

[wet nacelle]

It's even odder when you realize that a lot of these people either have autism or extreme ADHD.

How the fuck do you not find this easier?

[ionic gulch]

thanks. this one seems to work. can you explain me why the commas are needed? i did some other nodes without any problems even without the commas
for example, this one works:
https://i.imgur.com/sW9X2oh.png

[hardy cipher]

lol. yeah. plus all the quick image adjustments. it's basically what I would do in photoshop endlessly. not as in depth obviously. but being able to just put all those blends and stuff into a workflow and have them execute in a half second

so satisfying

[wet nacelle]

Totally!

[vast galleon]

The last commas aren't needed if you have more than 1 output. As to why, that's a question either for comfy, or the guy who created litegraph.

[ionic gulch]

@vast galleon oh, so this is not a general python syntax thing, but some "special" thing about comfy or litegraph?

[hardy cipher]

I've been doing weird things. get the difference between two similar outputs, then take the result and use that on one of the images. maybe a bit of burn or dodge depending on the situation

brings out a lot of detail

[vast galleon]

Yes, it's how they defined node classes

[ionic gulch]

and the if part had a problem with executing everything after the if only if the condition was met i guess?

[vast galleon]

Or the node doesn't even show up as available

[wet nacelle]

@hardy cipher @ionic gulch @vast galleon alright. Otter time guys. https://youtu.be/kNQxjjOjF9A

[hardy cipher]

otterly wonderful

[strong field]

stuff i guess

[hardy cipher]

have you ever tried to switch the lora loader name widget to input?

lol, it did not seem to like it when I did that

[cedar tusk]

I just pushed a commit that should let you customize the time format, so you should be able to do that now!

[hardy cipher]

very excellent

[strong field]

all ideas from my prompt gen

there was a guy in here that was like, guys i want to do a local LLM for prompts , what would you ask it?

and i was like, sounds like a weekend project

[hardy cipher]

ooh, did you finetune one? create a lora?

[strong field]

just finetuned a llm

not even my data lol

shh dont tell

[hardy cipher]

I'm really curious how that worls. what data do you use?

not that you have to tell me specifically

[hardy cipher]

but in general, I don't know what I'd use for training, or how much is required

[strong field]

i used this, and its a helpful reference for anyone

https://supagruen.github.io/StableDiffusion-CheatSheet/#Balaskas-Christopher

i also have 4 or 5 custom txt files with prompt work scraped off the ol web

[hardy cipher]

I've wanted to finetune a small LLM for prompting for quite a while, but don't know where I'd start. seems yours is working out well

[ionic gulch]

nice img2img result:
https://i.imgur.com/zf58ZOL.jpg https://i.imgur.com/7BImqCP.jpg

[strong field]

no code chat bots are an easy place to start

https://www.voiceflow.com/

[hardy cipher]

neat. didn't know that was a thing

[strong field]

and this one makes it pretty easy, sign up, start the tutorial for gettings started, upload data documents to knowlege base, and run a test

[hardy cipher]

I'll check it out. I've tried with oobabooga. but it's like a111 if it was way buggier and had almost no documentation

or maybe there's documentation, but everything is so fragmented and incomplete

that styles website is great

[strong field]

yeah, textgen webui is fun and all but its not production ready lol

[hardy cipher]

I rage quit and deleted it at least 3 times before I finally got it to work

[strong field]

i also pulled from this website for styles/idea generation

https://aituts.com/sdxl-prompts/

[hardy cipher]

I was just looking at that a few minutes ago

[strong field]

its part of my database now

lol

[hardy cipher]

good. there's about 20 ads on that page for every prompt. at least when I looked at it on my phone

[strong field]

now i dont even have to think about it

[hardy cipher]

the online LLMs are terrible at prompting in my experience. I can finesse them to give me some cool descriptions, but when I try to explain the how the prompts work they still use normal conversational english

[strong field]

that is verbatim, copy and paste, stick it into the blender

[hardy cipher]

on another note, have you ever observed that when parameters get pushed too far in weird directions that the resulting image tends to take on a circular shape? like it starts to kaleidoscope essentially? maybe I'm just imagining things, but it used to happen when I'd push cfg too high with a long prompt in 1.5, and just had the same issue with xl

[strong field]

send me a screencap

not sure i understand

[hardy cipher]

hmm, I'll post a couple of the images. they're trash, lol. but they can't all be winners

[strong field]

then i used the style cheat sheet to pick a random artist, this is in the style of Emily Balivet:1.5

[hardy cipher]

I didn't put anything in my prompt anywhere that would imply it should crate something like that

[strong field]

the cartographer is poking through

[hardy cipher]

I could track down some of my 1.5 images that would do the same thing

but I just took out the weirdness from my prompt so it's a bit better now

[strong field]

high CFG on dpmpp_2m karras does WEIRD stuff

[hardy cipher]

https://github.com/mcmonkeyprojects/sd-dynamic-thresholding

I'd use that to push cfg well above 20 in a1111

[strong field]

i think thats prob most common sampler cause its consistent, fast, and converges

[hardy cipher]

and then I'd have to scale it back sometimes

after adding loras

[strong field]

woah, thats HIGH

i never go beyond 9

lol

[hardy cipher]

well man, that extension is great. but no idea how to emulate it in comfy

[strong field]

what happens when you push it that high?

any obvious benefit?

[hardy cipher]

it'd really make sharp details and everything would pop. and then I'd schedule the mimic cfg and cfg to work together in unison so nothing would break

I'd couple it with the noise offset extension

I'll see if I can find an example

the readme for that github is very incomplete. it has way more to it than it says

that's a high cfg there

not sure exactly what. probably between 17-22

[strong field]

oh, oh my thats intoxicating

did you prompt this one for circular/kaleidoscoping?

thats nutso

[hardy cipher]

that one too\

no circular in the prompt for that one

[spring fulcrum]

Does anyone know if there are other refiner models for SDXL 1.0 other than the sd_xl_refiner_1.0.safetensors ?

[strong field]

25cfg

not seeing the same effect

[hardy cipher]

well you have to couple it with the threshholding, mimic cfg, and scheduling

[strong field]

oh, duh

interesting

[hardy cipher]

as well as the noise offput, loras. I had quite a little setup going for a while

tried to do it there

[crisp owl]

Not that I've seen, but not really a need for a separate model since the refiner should only be making extremely minor detail changes. Nothing really a different model would change in that regard

[hardy cipher]

anyway, I could post hundreds of examples, but just curious about what's going on and haven't really found any info

[strong field]

yeah those are some sick photos, curious to know whats going on behind the scenes

[hardy cipher]

a bit of it going on there

these are all super high cfg as well. never had that happen without it

not gonna lie, some of these images have me missing that old workflow I had

[strong field]

this is the important bits

[hardy cipher]

ahh. that sort of makes sense to me. it just sort of breaks down to a degree? it just felt like I was trying to push it too hard and it didn't really know how to respond

I'd really love to figure out how that extension works and how to get the same functionality in comfy

I know how the settings work in application, but not really behind the scenes

[strong field]

sounds like latent stacking , pretty sure you would need a custom node

[hardy cipher]

yeah. and I'm not even sure how it would really work. if I'd have to modify a sampler, or if it could be done externally

[strong field]

damn this guy has some cool math going on in his plugin

this is the gist

so its defining the limits for every step and then using the math to stack them togther

[hardy cipher]

with those ones above I was using repeating cosine modulation, or however you'd say it, so the cfg would go high to low, high to low, several times over the course of the render

[strong field]

really neat

alright that tells me i understood some shit

hahah

[hardy cipher]

the concept of mimic cfg still eludes me. I've read about it a bit, but it's kind of hard to wrap my head around. lots of new concepts

[steady grove]

https://tenor.com/view/vindication-brooklyn99-holt-gif-11799749

[azure oxide]

😆

the creator of that plugin actually got hired on by SAI for this and mabey some other ones

thats mcmonkey

[strong field]

are you F* kidding hahahahahahhaha

[hardy cipher]

cool! good for him

[strong field]

https://tenor.com/view/spiderman-defoe-respect-gif-24551385

[trim orbit]

you know, i'm somewhat of a mather myself

[hardy cipher]

my brain is quite satisfied by mathy things. but I really need to catch up on the programming side of it

[trim orbit]

https://www.reddit.com/r/StableDiffusion/comments/15oeymz/depth_sdxl_controlnet_coming_soon_brace_yourselves/ BRAAACE

[upbeat summit]

I used it for some time in a1111. I could control the effects a bit, but I didn't get 95% of the math or logic behind it

[hardy cipher]

well since the creator is a regular on this server I think I might try to ask him how the mimic scale works if I see him

[upbeat summit]

I simulated noise offset effects with it. got really dark images out of it. it could also enhance image fidelity or details. but most of the time you would make images that were burned and overcooked

[hardy cipher]

sometimes overcooked, but I'd mess with noise offset to fix that

[upbeat summit]

prompt, sampler settings and CFG scheduling had all to be balanced each time

[hardy cipher]

oh yeah, it was like trying to juggle 10 things at once

[upbeat summit]

some SD2 + cfg scheduling experiments

so you could get more brighter and darker perceived images than SD2 was capable of before noise offset was a thing. but of course the cfg scheduling has other side effects

[vagrant wasp]

why cartoon ?

[upbeat summit]

why not?

[vagrant wasp]

wasnt like this before

[upbeat summit]

I don't know what your workflow is from the screenshot. is it the same model and prompt...

are you doing img2img etc

[vagrant wasp]

everything comes out low quality cartoon picture

yesterday it was working fine

bug ?

[fathom hound]

cannot diagnose with what little information you provided

[vagrant wasp]

let me queue once more

[upbeat summit]

it looks like all your style tokens from the prompt are missing or it's a total different model

[vagrant wasp]

sdxl 1.0

[hardy cipher]

right on. it really does amp up the saturation and contrast. yours are a lot more cohesive than a lot of the things I made, lol

[upbeat summit]

thank you! yeah, it took days to find the right settings and prompt combination

[upbeat summit]

are you using any kind of node that inserts styles into your prompt?

if you haven't updated your setup and load an old image, it should make the same image if you are not using wildcard files

[vagrant wasp]

new workspace

[upbeat summit]

your old workspace probably had a node that inserted styles for you into the prompt. at least that is my best guess

[vagrant wasp]

styles ?

tried adding cartoon in negative

[upbeat summit]

I can't say how your old workflow looked like and what nodes it used. from the image you've posted the styles are totally different. so I guess the image in anime style had a different prompt

[hardy cipher]

man, I spent so much time messing with that stuff. after a whle it almost became intuitive

like I'd see something undesirable and know exactly which settings to adjust

[upbeat summit]

yeah. I was pushing the CFG as high as I could. like CFG 30 and than mimic 6-8. it worked with some prompts

[soft zealot]

Cheers for the shout out on social mdeia platform formerly known as twitter

@high skiff

[hardy cipher]

the readme said use half cosine up I believe. but I was a rebel and I'd use cosine down for cfg

well not always, but it worked well for me. just had to adjust values differently than most people would

[upbeat summit]

hehe yeah I played a lot with those settings. I really liked this combo... 3D anime, created that style with many TI embeds and cfg scheduling (SD2)

[vagrant wasp]

finally

changing resolution and sampler

[upbeat summit]

that makes sense

[hardy cipher]

here's some naturish images I made with it

[vagrant wasp]

a refiner would make that a masterpiece

[upbeat summit]

those are beautiful! also very interesting to look at and there's lot to explore.

a perfect use case for the effects of mimicing imo. CFG scheduling does have a strong impact on structures and details. the backgrounds are probably very much enhanced here because of it

[hardy cipher]

thanks! yeah, those ones came out pretty well. I have soooo many more too. once I got the settings properly tuned it was pretty much plug and play

I was trying to make them look like paintings but translated into real-life if that makes any sense

[upbeat summit]

it does 🙂 they could be HDR nature photography but they are paintings

[hardy cipher]

and I just couldn't bring out those colors without the extensions I was using. not sure it'd even be possible

[upbeat summit]

yeah probably not

[hardy cipher]

well at least not with the way things were structured

[upbeat summit]

I have not tried it yet, but comfy made this script that is very similar to CFG scheduling: https://github.com/comfyanonymous/ComfyUI_experiments#sampler_tonemappy

[hardy cipher]

I've messed with a bit. doesn't offer quite the same functionality but it does help

[upbeat summit]

I see

[vagrant wasp]

if you have collection i can upload them in my youtube photobook channel

5 subscriber 😂

[lucid hedge]

where can i get support?

i need help :c

[hardy cipher]

lol

[hardy cipher]

cool. yeah, I could upload them and give you a link. you could put whatever you wanted on there

[vagrant wasp]

https://chat.openai.com

[lucid hedge]

imagine

[vagrant wasp]

https://tenor.com/view/imagine-that-gif-11726635

[vagrant wasp]

ask away.. the community is helpful you will probably get your answer when the nerds are back

[clever verge]

AIT implementation in the works, maybe?! https://github.com/facebookincubator/AITemplate/issues/867

[vagrant wasp]

this workflow is so advance it crash my gpu

[hardy cipher]

weird happenings in lego world

[soft zealot]

I do hope your tongue was firmly in your cheek when you wrote that ;o)

[upbeat summit]

there are some nice structures in there 🙂 love the bottom mid one with the green hair

[clever verge]

Is there any information that indicates whether the same dataset and data was used to train both CLIP ViT-L and OpenCLIP ViT-bigG in SDXL? I think I've heard somewhere that same data was used on both but not sure where.

[hardy cipher]

I like this one

[soft zealot]

Lego(ish) using my favourote swedish girl in a bar prompt (very very loosley lol)

[hardy cipher]

got a bit weird

but yeah, refiner on these might be good

[upbeat summit]

definitely unique

[hardy cipher]

hans memling meets legos and drugs, lol

[upbeat summit]

or it might be negative for the style

[hardy cipher]

that might be true. hard to say I guess. I'd like to explore those prompts again, but didn't have as much luck with the workflow I had going in comfy

[hasty mortar]

Guys has anyone here tried to get the outline of an image with img2img?

[clever verge]

@soft zealot I sent a PM to you

[shy kelp]

how do you maintain a sort of similar character in different pictures?

[soft zealot]

Luck

[meager canopy]

Can't you do this with controlnet?

[hasty mortar]

I'm asking because I didn't try yet

[hasty mortar]

I was thinking of building a comfyui flow for this

[meager canopy]

Includes controlnet: https://civitai.com/models/119257/gtm-comfyui-workflows-including-sdxl-and-sd15

[hasty mortar]

I'm a check that thanks

[late marsh]

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

[vagrant wasp]

i just need help with the craSHING part

maybe need to update

[soft zealot]

well given all you've said is that your GPU is crashing thats a bit like going to the GP and saying "I dont feel well"

Bare miinimum need before any opinions etc can be formed

1. What GPU do you have?
2. What error messages (if any)are you getting in the condole when it "crashes" ?

[vagrant wasp]

GTX 1660 6gb ram

[vagrant wasp]

no console.. just closed

[soft zealot]

if you're trying torun an SDXL model thats probably most of the problem

[vagrant wasp]

some days it runs fine

may AI doesnt like monday

[soft zealot]

aye but its probably teetering on the edge and one more chrome tab or YT video opening is enough to push it over

[vagrant wasp]

hmmm seems like it

[molten gull]

try to restart your computer, start NO other programs at all, and see if it then runs

[halcyon tusk]

came to post this and you just did! important notice, this openpose controlnet for sdxl does not work with Auto1111, just Comfy for now.

[trim totem]

Hey all,

does anyone know how to get realistic hands? I'm using SDXL in comfyUI and Searge's nodes, which is great. But I cannot believe that Midjourney is doing a better Job at creating hands.

Any idea?

[dense chasm]

https://imagen.research.google/video/paper.pdf diffusion video coming soo

[vagrant wasp]

Any tutorial to make my own node ?

[high skiff]

Love!

[hasty mortar]

Midjourney 5.1 was specifically trained for hands, I suggest to find a sdxl model with similar features

[wet nacelle]

I do.

Some even reccomend the 0.9+1.0 one.

[buoyant axle]

a while ago since i fired it up, but as far as i remember, there is an option in the settings where you can configure the samplers to show

then i dont know sry

[half ivy]

I am here
What did i miss?

[tender timber]

Awesome, I’ll check it out!

[ionic dragon]

this is realistic af

[wet nacelle]

Guys I know there's been a heatwave for the last week. Just come to work guys. Pwease?

[wet nacelle]

"Elon. Please. We don't want to be in the factory today."

[uncut fiber]

can i ask about AIT loader in ComfUI?
It makes it/s twice faster, but on VAE box it stay loooong about 10sec. So what i gain with AIT i lost with VAE decoder. Any solution to this?

[unreal plover]

https://huggingface.co/diffusers/controlnet-depth-sdxl-1.0

[unreal plover]

100% real image, nice try

[upbeat summit]

are you sure?

[unreal plover]

https://tenor.com/view/the-office-jim-halpert-hmm-doubt-it-doubt-it-doubt-gif-20978255

[thorny frost]

the one with the close-up is 100% ai generated the last one probably too.

[hardy cipher]

https://tenor.com/view/eyeball-creepy-halloween-scary-cool-gif-19806626

[wet nacelle]

My sweat fetish helps with the realism prompting a lot.

I can totes see the imperfections though.

[hardy cipher]

sweat fetish

[wet nacelle]

yuppers

[hardy cipher]

might be the first time I've seen those two words combined like that

[wet nacelle]

And it will likely be the last.

[hardy cipher]

I mean, I think there are much worse things in the world, lol. whatever floats your boat

[wet nacelle]

Totes.

Here is the upscale. The smaller nose is the upscaled one.

I think it destroyed the overall texture and facial structure of her face.

It's as if the image was ran through topaz labs.

[hardy cipher]

so I never did figure out that controlnet thing. but I did figure out how to make a node

her skin looks a bit distressed

[wet nacelle]

I watched his other video on setting up controlnet with comfyui.

It seems that I need the windows portable install for it to work at all.

I don't want to go through moving everything into another folder yet.

[hardy cipher]

hmm, I wonder what that's about

[hardy cipher]

so anyway, I'm not much of a coder, but that node seems to work

[half ivy]

this workflow hurts me

[soft zealot]

did you use workping.py to embed a workflow that generated a different image in here?

I have DLd the same Model & Lora you used plus used same seed/sampler/scheduler/prompt and got a different image from yours.

Now maybe it's me.....................

[wet nacelle]

I don't know that humbo jumbo honestly. workping.py? huh?

[soft zealot]

I know what in the embeded Json thats how I went looking fo rthe model & LORA

as I said, it may well be me doing something different (in fact theres every chandce it is as I use a Pre conditioning step)

Simply oibserving I got a different genrate dimage 🙂

[wet nacelle]

I really don't know more than just prompting knowledge really.

[wet nacelle]

Right. odd.

[ionic gulch]

my nodes are now available in comfy manager:
https://i.imgur.com/E4AVGTX.png

[hardy cipher]

nice!

I made one myself. wanted to figure out how it works

[ionic gulch]

even got a nice id (111) :)

[hardy cipher]

https://github.com/picturesonpictures/comfy_PoP

calculates any aspect ratio for any resolution square up to 4096. not sure how useful that is, but you know

so basically 1024x1024, put in any 2 numbers for AR, and it'll output values that correspond to the same pixel count

or 512x512, 1742x1742

lots of downloads as you can see

[strong field]

well done

[ionic gulch]

is it just the pixel count or do you consider other factors as well?

[hardy cipher]

such as? are there other things I should consider?

[strong field]

dont dash his hopes and dreams man!

[hardy cipher]

I mean, I can change it if it needs to be changed

[strong field]

https://tenor.com/view/whats-going-on-drenched-flooding-raining-lots-of-water-gif-14243222

[strong field]

its more a matter of how complex you want it to get

[ionic gulch]

i'm not sure, but afair the resolutions should also be devidable by some number like 16 or 64. not sure where i got that info.

[hardy cipher]

well I could have restricted it to only 512, 768, 1024 or something

but didn't see a reason

[ionic gulch]

here is some info i found:
https://stablediffusionxl.com/sdxl-resolutions-and-aspect-ratios/#:~:text=For the best results%2C it,x 896 (9%3A7)
that's where i got the resolutions for my resolution node

[hardy cipher]

and you might be right as far as it being divisible by those numbers. but that's not a hard fix

[strong field]

i believe thats right, official docs maybe?

i believe those are the best practice resolutions

had something to do with how it was trained

[hardy cipher]

meh, well I'm curious to see what difference it makes

[strong field]

let me see if i can find it

[hardy cipher]

could also just make another one that adheres to that, it'd be pretty easy

[strong field]

make a little enable disable lock button

all in the same node

[hardy cipher]

yeah, that's what I was thinking

[ionic gulch]

guess it reduces the risk of double objects, but other resolutions should work too, if you try often enough.

[hardy cipher]

I actually don't really know python. so it took me a little while to figure out wtf I was doing. but it does work as it's supposed to

[strong field]

thats a W

[lusty token]

A lot of my generations have a very slight layer of blurriness, prior to any upscaling attempts using. Does anyone else run into this?

I've tried 1024x1024 or 832x1216 for example.
base model: sdxl_base_1.0
refiner: sdxl_refiner_1.0
vae: sdxl_vae

[ionic gulch]

@hardy cipher great. after you've got that first one, additional nodes will be much easier. and it's great to have a way to adjust everything to your own needs.

[hardy cipher]

even got it to show up in it's own folder in the node menue rather than one of the random default node menus

wasn't sure what category to put it in. but it works so all good

[hardy cipher]

it could be lots of different things

[upbeat summit]

Thanks for sharing it!

[lusty token]

Look at this assclown for example

[ionic gulch]

those are my 4 main ones, that reduced the needed space in my gui a lot:
https://i.imgur.com/p1UatDd.png

[strong field]

CONFIRMED, divisible by 16 for best results

[hardy cipher]

for sure. I figured I'd post it and be informed I didn't do something correctly, but all good. I'll probably implement the option to adhere to the proper multiples of 16 or 64 or whatever

[strong field]

https://youtu.be/2Xe79Nl_6jA?t=508

video timestamp if anyone cares

[hardy cipher]

but also, would be kind of nice to veer from that and see what difference it makes

or interesting

[strong field]

that is a nice node

the top right

[strong field]

what sampler, scheduler, and step count (base/refiner/total)

[hardy cipher]

I do like the efficient nodes. and the pipes. although running sdxl through the basic pip gives me all kinds of errors about tensor sizes

not sure if it impacts the results or what

but stopped using them

[lusty token]

sampler: euler
scheduler: normal
steps: 30
cfg: 7

[strong field]

base steps?

and refiner steps?

also refiner denoise?

[ionic gulch]

most of them should be very universal, only the additional settings are especially for my own workflow.

[strong field]

im guessing the refiner is smashing things in its particle colider

[lusty token]

I'm trying to trace it... it's based on Searge SDXL.

[strong field]

can you send a screenshot of the main settings?

[lusty token]

oh sure, sorry. I thought it was in the pic I sent of the joker.

[strong field]

lol maybe i just didnt drag and rop

[unreal plover]

output is named input but it is output right ?

[strong field]

this might help, i suspect the step count is too low

https://www.reddit.com/media?url=https%3A%2F%2Fi.redd.it%2Fsdxl-1-0-grid-cfg-and-steps-v0-7dy3xg4hl1fb1.jpg%3Fs%3Dd1136fac634155b5b3dd6bb8f53071e1d3590676

for your chosen sampler/scheduler combo

try 45 and see if you see an improvement

[ionic gulch]

would try euler a and 60+ steps to be sure that sampler and steps are not the problem. 30 is not much for base + refiner

[strong field]

(or increase the CFG but thats a bit more delicate lol)

[lusty token]

Ah okay, very well. I'll move back to eulerA and bump up the steps.

[strong field]

FIX YOUR SEED when you are trouble shooting

or trying to isolate issues

[lusty token]

Ah yes, thank you for the reminder on my seed.

[strong field]

np

[lusty token]

scheduler, normal or karras?

[strong field]

i like karras personally

[upbeat summit]

let's test some dpmpp_3m

[strong field]

https://tenor.com/view/michael-jackson-turn-head-gif-11960836

[vapid roost]

hello friends, while waiting on SDXL training methods, check out post on SDXL 'implementation' in ComfyUI from scratch.
I also try to discuss different aspects of the model architechture along with basics of the tool so maybe you'll find it interesting.
It will get much deeper starting from part 2
https://followfoxai.substack.com/p/part-1-sdxl-in-comfyui-from-scratch

[upbeat summit]

https://tenor.com/view/michael-jackson-michael-kackson-gif-michael-jackson-gif-smooth-criminal-smooth-criminal-mv-gif-24082079

[strong field]

smooth...

are you being serious is there 3m?

[upbeat summit]

yes

[steady grove]

a crimmminall minddd is alll i have!

[strong field]

is it open access?

[upbeat summit]

update your comfyui 😄 https://github.com/comfyanonymous/ComfyUI/commit/58c7da3665726a75adc4d91217364d93e7f44945

[steady grove]

ooo depth controlnet is out finally. i can stop bracing

[strong field]

https://tenor.com/view/ayayay-chris-jake-beer-destroyed-gif-20740887

[upbeat summit]

thanks for sharing it! I will definitely take look.

[upbeat summit]

lol https://github.com/comfyanonymous/ComfyUI/commit/0cb6dac943a2550f6e5c5cf4f99a64b40c89ac80

stop the party

[strong field]

https://tenor.com/view/the-office-jim-from-krasinski-gif-13369858

[azure oxide]

3m SDE tho 👀

[steady grove]

https://huggingface.co/thibaud/controlnet-openpose-sdxl-1.0 thibaud coming in hot too!

sarge updated his depth model today too

what a time to be alive!

thibaud dropped it all!! scribble, hev, all of it!

wait no mb. just openpose.

[warped kiln]

Yeah other stuff is 2.1

[lusty token]

oh good catch, you mean on the nodes?

[steady grove]

he updated it all today to trix me

[warped kiln]

Anybody here found success trying to get uncentered images yet? Still struggling to get any remotely consistent control of framing at the prompt level

[lusty token]

It seems like an improvement having just bumped it up to 45 steps.
I kept the same seed from the previous image, but it generated a different result anyway. I guess because of the sampler change...

[vital ermine]

My effects lora seems to work but seems dependant on subject a bit

[trim orbit]

controlnet openpose is out now apparantly. that'll be a success

[vital ermine]

why would my effects lora need more steps to show?

under cooked

[warped kiln]

Yeah been messing with them, unfortunately they tend to fall apart fairly quickly if your prompt isn't fairly strongly correlated to the controlnet model's input. Like for architecture for instance. Works ok if all you want to do is basic human posing, but if you stray from that the prompt coherence goes out the window (this is including canny, depth, and pose)

[vital ermine]

str of 1.2

[strong field]

looks much better

definitely sampler will change how noise is handled

@rovo excellent write up here if you are interested in understanding more: https://stable-diffusion-art.com/samplers/#So8230_which_one_is_the_best

[vital ermine]

Looks to be very slightly undercooked

[upbeat summit]

nice - that looks cool 🙂

[strong field]

running those tests for comparison

100 batch lets go

[vital ermine]

Thanks, it is like a lofi

scan lines and all

I can't use it in comfy since he doesn't work with ia3 lora

[trim orbit]

sepia and scanlines seem like an odd mixture to me

[strong field]

@rovo heres another great one, less reading, more doing, pick your poison

https://sdxl.replicate.dev/

[upbeat summit]

yeah effects loras are very interesting - and glitches are great anyway.

[lusty token]

ah cool, really nicely put together. Thank you.

[vital ermine]

yep, that is the effect I was going after

glitch effect.

Yeah, this needs a bit more time in an oven

[upbeat summit]

I see I see. also a person with well organized resources. very good.

[vagrant wasp]

crash when using refiner

[vagrant wasp]

is it so much a strain on gpu ? using two loader

[trim orbit]

strain implies it could break the gpu. it's not being strained

[upbeat summit]

it eats vram, that's it

[strong field]

and heat, it does heat

[vagrant wasp]

explains the crash

[upbeat summit]

and eats dust

[vital ermine]

my lora as a neg

[trim orbit]

the sun was beaming thorugh my window directly onto my pc case for about an hour yesterday. it crashed on me then haha. the one hour of the day during the one week in august that it could happen

[vital ermine]

Interesting as a neg

[lusty token]

That's Time for ya! Always indicating just where the sun is going to strike the surface of our planet. 😄

[strong field]

really digging the asthetic on these

[vital ermine]

Top one was with my lora as a neg

[lusty token]

60 from 45 steps;
I'm seeing the light now.

[strong field]

thats base?

[vagrant wasp]

more steps more fun ?

[lusty token]

Yes, that is Base. Slightly upscaled.

[peak dove]

Love the 'scan lines' reminds me of my old 625 line TV 🤩

[strong field]

that is excellent for base

[strong field]

The man on fire is back

initial 3m SDE GPU speed tests: no change to 2m, lol

quality up next

[crisp owl]

doens't have the GPU ones that I can tell, but you can take a look at this chart for sampler rendering time

[unreal plover]

Have they used upscaled images to train SDXL ?

[soft zealot]

grrrr mutter mutter

OOM when trying SDXL ControlNet

(Yellow are debug only to confirm ControlNet Input Image aint mahoosive)

[unreal plover]

how do you hide links ?

[pure hazel]

Do you guys notice SDXL produces the same lips in every photo for women?

[soft zealot]

https://github.com/failfa-st/failfast-comfyui-extensions

[upbeat summit]

it was possible before with failfast QoL package but since ComfyUI now has native support for Link Render modes in the settings, failfast removed it from the package - so was the Link Render mode option: none.

[upbeat summit]

it was removed

[soft zealot]

ah I still have it instal;led, dont think I updated

will uncomment uit if OOTB now

[upbeat summit]

https://github.com/failfa-st/failfast-comfyui-extensions/commit/e0a4343965dec9c3d283d1c9cd281a85f95e1063

[soft zealot]

Im not disputing 🙂

[upbeat summit]

yeah, that wasn't my intention 😄 just a src

[crisp owl]

Any reason he removed those that has been discussed?

[soft zealot]

because comfy added the option OOTB so not needed ?

[crisp owl]

Ah gotcha

[upbeat summit]

minus the "show none links at all" option

[crisp owl]

fair enough lol

[soft zealot]

ah so the OOTB option doesnt have the HIDE Option?

[upbeat summit]

correct

[crisp owl]

Didn't realize it was also a native option as I still have the failfa option I've been using. But yeah no hide option in the native menu I see

[soft zealot]

leaves his failfast in place and makes note to not upgradeit

just so much cleaner with everything hidden

[forest zephyr]

question about lora and model training sdxl. 24gb absolute or can you squeeze by on 16?

[strong field]

obligatory thx

[upbeat summit]

is it back in or did you get the previous commit?

[strong field]

did you prompt different lips?

[strong field]

3m SDE GPU looks like its still there

[upbeat summit]

the only thing I heard was - the CFG is much hotter now

[strong field]

CFG is next!

[ionic dragon]

is this a1111?

[strong field]

negative, comfy

[upbeat summit]

yeah still need to try efficiency plots. installed a1111 just to do plots 😄

[hoary saddle]

anyone getting a message "Killed" after a bunch of rendering on ComfyUI? running -lowvram on a 3080 10GB

[strong field]

theres the workflow, try it 😛

[upbeat summit]

ah awesome! thank you for doing the tests and the workflow!

[strong field]

im a mathlete so nerd is inferred but forget what you heard...

[upbeat summit]

yesterday it was recommended to only run it with the exponential scheduler

[strong field]

really? hmmmm why is that?

pulled exp in to test, removed normal and ddim

[upbeat summit]

it seems to be updated now according to the last commits. I couldn'T find any info about 3m

only that there's a pull request for a1111 as well

[strong field]

wait wrong clip

well i concur, exp is producing best results there

[ionic dragon]

how did you get grid in comfy/

[upbeat summit]

yeah exp does look best here

[strong field]

https://github.com/LucianoCirino/efficiency-nodes-comfyui

[ionic dragon]

which node?

[strong field]

XY plotters

[strong field]

your gonna love this next test

sneak peak

[upbeat summit]

teaser!

[ionic dragon]

thank you

[indigo carbon]

did you use AIT on this?

[strong field]

yes, all ait

be advised, starting tests with SD1.5 for comparison later

[upbeat summit]

cfg 6.5, steps 55 could be the start of overcooking

steps 55 all look on the verge

but maybe it's the seed

I will do tests tonight as well

all run with exponential?

[strong field]

yes all exp

i honestly like sub 50 step flavor

[upbeat summit]

the steps 53 lane has some interesting details

[strong field]

running tests higher to push boundary

[upbeat summit]

sure. but it's SDE - so

[strong field]

53 was my fav

[upbeat summit]

yeah - looks most coherent

[strong field]

SDXL will come next lol, but makes sense to eliminate options before running in sdxl

[upbeat summit]

I mean I ran a couple of sessions with dpmpp_sde_gpu at 55 steps the last few days

[strong field]

sde gpu at 50 for me, testing

but this changes it for sure

[upbeat summit]

at one point you will have reached final convergence 😄

[upbeat summit]

dpmpp_sde can look great from 15-55 - depends on the rest of the settings and prompt

[strong field]

honestly 3m looks goddam good

damn you can push this far

https://tenor.com/view/mama-mía-joséluis-gioia-master-chef-argentina-madre-nuestro-madre-mía-gif-24079005

CFG: 8.25 and Step:56

[upbeat summit]

cfg 12 come on

[strong field]

https://tenor.com/view/they-made-me-do-it-forced-they-told-me-had-to-didnt-do-it-gif-15152608

[strong field]

i see you have the documentation on your side

17 is the limit lol

but man im suprised by 15/53

There it is again, 53....

[upbeat summit]

I may have read that but I knew it from testing before - because I made over 100k images with SD2

it were probably 150k - but who counts (beside my electricity bill)

[strong field]

tooo true

SDXL is a beast holy shit

[upbeat summit]

Thanks for testing high cfg

[crisp owl]

https://tenor.com/view/shhh-shush-silence-nose-gif-17895433

[strong field]

my pleasure

going to run the sdxl plots now

[upbeat summit]

and it doesn't look as burned as I would have expected

[indigo carbon]

Karras

simple

ddim

normal

exponential

[strong field]

not at all im super surpised, looks like 3m improved quality SUBSTANTIALLY

[strong field]

karras and exp are my fav

initial tests to figure boundaries

its insane that there is actually intelligibility at CFG:20

SDXL is flexing tho

https://tenor.com/view/macho-gif-8000415

[upbeat summit]

with cfg scheduling I ran cfg 30 / mimic 5 😄

[strong field]

i havent messed with mimic, hows it look

[upbeat summit]

I used it in a1111 with SD2 7 months ago. interesting effects, but not easy to dial in. you mostly get overcooked images. everything can look like wax.

once I had dialed it in with the correct settings and a fitting prompt, it changed contrast, increased details in the whole scene and you could emulate a way darker lighting environment compared to what SD2 was capable of before noise offset was a thing.

I used it for some time making SD2 images.

[strong field]

how this looks so good is beyond me

what a marvel

[upbeat summit]

15 is definitely too hot here

[strong field]

oh for sure, 12.75 look sjust on the verge

but still looks great, it has a great aesthetic if thats what you are going for

consistency is impeccable too

25 steps not enough for high cfg

[upbeat summit]

yeah you can see how the coherence isn't there

[strong field]

ok trying a different subject, lets see where it goes with it

[upbeat summit]

46 does have it's charm. 60 has more details

[strong field]

alright, lets break shit

[upbeat summit]

I'm running plots right now :/ so thanks for doing some testing

[strong field]

send pics or it didnt happen!

lol send your favs, im interested in what you come up wiht

[autumn forum]

Are y’all using like Fibonacci numbers or something?😂

[strong field]

ok last one for a bit

[upbeat summit]

I wanna see this - than I'll be back in a bit and see what I can find out

[strong field]

need to put computer in the freezer for abit

[upbeat summit]

I guess 60 it is 😉

[strong field]

WOW

this is some ish

2m performs remarkably well here, but 3m is clearly king of the details/aesthetic/vibe

compositon is better, DR is better

[hardy cipher]

I'd like to know what sort of psychopaths are rendering 5 and 10 step images

[upbeat summit]

you totally can, but I won't

[strong field]

this kills dpmpp_2m for me...

https://tenor.com/view/starwarsday-gif-5413754

[autumn forum]

Wait has dpmpp_2m been overtaken as the best sampler??

[upbeat summit]

when I used DPM++ SDE Karras with 15 steps on SD2 I was told that 10 would be enough. So I pumped it to 20 and ended up at 30. good sweetspot - great quality

[strong field]

for me yeah, its up to personal tastes.....but im not saying i didnt just render like 300 comparisons showing that yeah dpmpp_2m_sde_gpu with exponential scheduler outperforms

[steady grove]

i came up learning that anything over 25 is A LOT so when i see these 50 and 70 step workflows i'm like surprisedpikachu.gif

not saying i know anything at all. just saying im surprised

[shy kelp]

The pathway to exponential is considered unnatural

[upbeat summit]

yeah - if I can see details emerging, why would I not want that? speed? okay... is it always necessary? no

[autumn forum]

Ah okay lol, I’ll have to try it out for my taste then.

[steady grove]

huh?

[strong field]

if you know how to compose what you want then its much less about generating qty, and you can focus on quality

[upbeat summit]

yeah - I always trade speed for quality with this

[strong field]

i posted all my work in here so you can see some comparisons if you go up

[autumn forum]

Yeah I’m lookin at em!

[strong field]

is that by choice or necessity? lol

[upbeat summit]

choice. I can always choose different settings, but my main thing with this here is pushing the fidelity - see how far it can go

[strong field]

i think i misread, you prefer quality right?

same

[upbeat summit]

trade speed = for quality 😄

speed < quality

[hardy cipher]

well I totally get rendering at lower steps for some things, or if you're trying to find the right setup. but then I see people so flabbergasted when they see someone else used more steps. like the goal is to mass produce as many images as possible

and I do a pretty good job of that with my little insufficient video card

[indigo carbon]

yeah, that's exactly the reason AIT exists, crazy speed without precision loss

[hardy cipher]

if it's a super nominal decrease in quality for significant increase in speed, that makes sense, but people are just silly about it

[strong field]

i dont want 1000 potato heads

[upbeat summit]

I produce a lot of images. in the early days of SD1 and than SD2 my way of getting interesting images was pure brute force. render 500 images - get 20-25 good ones that I found acceptable. like a photographer that shoots 500 photos and gets 3 - 5 really good ones.

with SDXL it has changed quite a bit. the ratio of interesting images is much higher now - which is really satisfying and cool 🙂

[indigo carbon]

we already proved that, more consistent than Xformers- while being incredibly fast.

[hardy cipher]

yeah, much higher win frequency

[hardy cipher]

I believe it. not doubting you. I wasn't attempting to call you out or anything. I'm all about it

[strong field]

4k in

[hardy cipher]

4k image?

[strong field]

yes

[hardy cipher]

bro, that doesn't fit within the size parameter restrictions

[strong field]

upscale a la mode

[upbeat summit]

I've set scale_by by accident to 2.0 using 4x Ultrasharp - ended up with a 13000x7000 image - took 29 seconds. of course it's way to blurry for so many pixels but it was very fast

[hardy cipher]

yeah, I've made some HUGE images. well not really bigger than that actually

but they're what, 100 mb or something?

[upbeat summit]

60mb png

[indigo carbon]

that's the reason the AIT workflow I made has upscale by to 0.6, this makes ~4k consistently, without loosing any speed

[hardy cipher]

my 4k images are 30 I believe

I blew an image up to 8k in 1.5 so I could go wild with inpainting

then realized there was hardly any way to send it to people online. or host it even. almost every service will either shrink it or not allow you to send it

[trim orbit]

i give it about a month before hosting services are like "okay hold on wtf is going on here"

[trim orbit]

it's not so much the file sizes of individual generations. it's the constant deluge of uploads

[hardy cipher]

if they don't actually shrink it resolution wise they still gut it

my goto fix is to just make sure it's saved as a jpg

rather than png

[trim orbit]

discord likely has a plan

and a prayer

[strong field]

you could zip it?

lol

[upbeat summit]

yeah you could compress it, but that costs compute

[hardy cipher]

tile it. split it up into 1/9ths

[upbeat summit]

you can also make a 5mb jpeg out of it

but who wants a lossy format

[trim orbit]

at some point, we may just start swapping parameters and people will generate the entire image on client side with perfect reproduction

[strong field]

this is the way

[trim orbit]

a lot of protocols will need to be developed but i think the compression benefits of offloading it to the client is something data centers are going to be like "yup"

[strong field]

discord with its own sample decode haha, that sounds incredible, excellent data transmission with transcode process

[trim orbit]

could even imagine a way of storing a low dimensional latent version of a photo from a camera, and reproducing that in higher fidelity with diffusion. as part of a fileformat

[hardy cipher]

someone just needs to come up with an efficient way of doing that

it just seems like something someone might just pop up out of nowhere and release in 3 months

[abstract blaze]

Any idea how to make these kind of artifects tidy in A1111?:

[strong field]

now we are talking

[trim orbit]

i'm just high level spitballing and have really no idea if this is actually possible, but from the papers and research i've glanced at it really seems like this is the direction

[upbeat summit]

discord's image data storage is probably an equivalent to what youtube is for video (minus a couple of exabyte)

[strong field]

lets dream

[trim orbit]

intel probably. hyped like "MMX TECHNOLOGY"

https://www.youtube.com/watch?v=o1Rz2A5mxk8

[strong field]

oh god

[indigo carbon]

I remember there being an image format that only works with certain software that no matter how much you zoom it, it's never low res and files are ~2mb, it only worked on images created with that specific format though

[trim orbit]

real media?

[indigo carbon]

I don't remember names, but there was something like that somewhere.. I'm sure if you look for it you will end up finding it

[versed locust]

Guys im trying to make a Logo on SDXL, any tips on what to put in to make it proffesional looking?

[vagrant wasp]

Custom nodes programming possible

With script

[trim orbit]

theres been so many failed iamge formats. i like the old interlacing ones that tried to encode higher fidelities into tricks that crt could provide. it was the way konami made oldschool cool pixel art

[upbeat summit]

depends on the style you want to do

[vagrant wasp]

Like generate json data filtering prompts

[trim orbit]

start with a plan and rough sketches then use diffusion to realize that plan. That's how professionals should be doing it

[versed locust]

Any idea what model would be best?

[trim orbit]

if you haven't done at least 2 dozen thumbnail sketches for this logo yet, you haven't even considered designing a logo. imo.

[indigo carbon]

I'm sure it might be possible to make something like PNG that does something similar though. it's definitely a possibility

[trim orbit]

here's one from at&t that required hardware

[versed locust]

well thats why im using SD because it will come up with a million possibilities that i wouldnt think of lol

[upbeat summit]

done

[trim orbit]

feels cheaty but in a Kobayashi maru kinda way

[indigo carbon]

yeah, AI needs some format like that to store gens. I'm sure it will show up eventually

[strong field]

this was so good and impressive at the time, like the final fantasy 7 characters looking HD on a potato

[trim orbit]

almost certainly. we're at gen 2 of diffusion (imo) and it's making 80mb files lol

[hoary saddle]

sometimes Load button doesn't do anything in Chrome, there a fix for that?

[upbeat summit]

hey - hacking is allowed in space (or at the academy)

[indigo carbon]

yeah LOL, like, it's probably taking so much of my storage

[trim orbit]

yeah hahaha. i remember loading it in an emulator for the first time and thinking "this looked way better before" always gas lighting myself thinking i was just blinded by nostalgia. but yeah, it actually do look WAY better on a CRT

[strong field]

"professional logo design, minimalist, sleek, modern, award-winning" - add creative idea here

i had this exact reaction too!

[trim orbit]

i mean. barret's hands were cubes. there was a little bit of nostalgia

[upbeat summit]

good old barret's hands

[strong field]

is that a gun or a hand, its a gun and a hand, ok im cool with it - young me prob

[upbeat summit]

now we have more fingers than we need

[elfin flare]

I think it would've looked better to do plasma

Because plasma can exist under water

[strong field]

it obvioulsy would have looked better to do plastic bottles and oil!

[elfin flare]

Bruh

[trim orbit]

was diffusion used to create that "zygote monster" episode in oats studios?

[hardy cipher]

I'm fairly certain fire can exist under water

[hoary saddle]

does the AIT custom nodes from the Manager install work on Linux now? or do we know?

[unreal plover]

You posted this image so many times !! 😂

[indigo carbon]

yes, it was fixed a few days ago

[hardy cipher]

Tdg8uU has his bangers and he's not shy about sharing them more than once

[indigo carbon]

I just generated this.. I just really liked that prompt

[trim orbit]

no. after you share, it's done forever. keep it in the folder until bit rot. as is tradition.

[hoary saddle]

think it worked

[indigo carbon]

did you use CPU ARG? it should look slightly better

[hoary saddle]

probably not, that changable in a conf file somewhere?

[indigo carbon]

also might be new record, 4090?

[hoary saddle]

yeah, very first render on this new card (installed about 30 minutes ago after amazon shipping debacle)

[indigo carbon]

it's enabled by default in \ComfyUI\nodes.py

you should change it to GPU ARG, it makes it work better with AIT for some reason

[indigo carbon]

change line 1154 to noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="device")

[hoary saddle]

6 instances of "CPU" in nodes.py

k

[indigo carbon]

also do this while ComfyUI is not active, it might cause issues

[hoary saddle]

all this spam safe to ignore?

[crisp owl]

yes

that's it setting up basically

[shy kelp]

hello

[indigo carbon]

of course, you didn't get this before? this is how it should look like before loading models in general, not just with AIT

[hoary saddle]

yeah it came up on first launch before AIT

[indigo carbon]

@hoary saddle try to gen now, speed should be the same ~12 seconds but quality should be somewhat better

[shy kelp]

do you guys have any tips for generating consistent characters? im using comfyui and tried to to the celebrity name matching think but also read there's no alternate prompt editing, any suggestions for the prompt or the workflow? im new

[brazen sapphire]

You should look into generating LoRAs

[indigo carbon]

also since someone asked me to do plasma, here it is..

[brazen sapphire]

Search YouTube for Koyha @shy kelp

[hoary saddle]

i'm using base too instead of whatever checkpoint you have, could be it, but yeah 12 sec

[ionic dragon]

using ultra long prompts, is giving great results

[azure oxide]

but also read there's no alternate prompt editing
this is incorrect, by default there's the usual prompt editing syntax from a1 here https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing
and also the dynamic prompt syntax of randomizing prompts {random1|random2|random3}

[shy kelp]

alright but is there anything i can do with the prompt?

[azure oxide]

i use those in my usual a1 prompts so that was pretty big in making the jump to comfy

[shy kelp]

yes thats where i read about it, and issues in comfy says they're not adding it i think

[azure oxide]

unsure about that, it works fine for me

[indigo carbon]

you should try that model, it's my personal favorite XL finetune. same reason why it's there by default when using the AIT workflow I made

[azure oxide]

not saying it doesnt work for those users and mabey even you down the line but its not a hard fact that comfy doesnt support it is all im saying

[shy kelp]

the random is assigned at prompt evaluation time in comfy as far as i read and a1 allows for going back and forth between the option each step right

ahh so you suggest using a1?

[azure oxide]

no im just saying it should work, give it a try

the random prompt is assigned at the start for a1 as well, you can tell which one it uses with the metadata

[hoary saddle]

looking it up now

[brazen sapphire]

Does anyone know if the only thing holding us back from using these new xl control nets in a1111 is just the lack of provided YAMLs?

[strong field]

confirming this works, i use it all the time, using it right now

assigned at gen is correct

i learned this form detweiler who works with comfy, so....supported unsuported? he said she said, but it works

[strong field]

this is sick

[lusty token]

Any recommended comprehensive documentation of ComfyUI Nodes? Like as a table or list?
This is the closest I've found:
https://blenderneko.github.io/ComfyUI-docs/Core Nodes/

[indigo carbon]

idk about A1111 anymore, the best modern alternative for people that hate nodes is probably ComfyBox, it's just a frontend that uses your ComfyUI install as a backend

[strong field]

i think thats the best ive found

[azure oxide]

https://github.com/ssitu/ComfyUI_restart_sampling

i forgot that this existed

now i have another thing on the list to test with sdxl zzzzzzzz so much to do

[vital ermine]

Well, lol

[shy kelp]

can i train text embeddings with comfyui or do i need another tool did anyone do that?

[strong field]

not aware that comfy can train embeddings.

[hardy cipher]

I think I got it working . all those ratios are on the stable diffusion list of proper aspect ratios. I realize it's probably basic stuff to a lot of you guys, but had to learn myself some new things figure this one out

[vagrant wasp]

for every lora i need a node ?

[hardy cipher]

and there's still the option to do really dumb things, lol. just have the default values as they should be

[strong field]

everybody gotta start somewhere, at least you are taking initiative

[hardy cipher]

there's that efficiency node that loads several, but then if I remember correctly the only output by default is to another one of the efficiency nodes

[strong field]

this, or chain the loras with a node per lora

[strong field]

now thats a good design

take off the training wheels and roll down the hill

[hardy cipher]

I just chain the LoRAs normally. it's pretty quick

[strong field]

this, also you can CTRL+SHIFT+V to paste with same connections, speeds up the layout

[hardy cipher]

well I thought about having it only load the listed ratios, but then what if I want to experiment with something off the wall to see why it won't work and is a waste of time?

[ionic gulch]

i like this lora loader with on/off switch. so you can wire 3-6 loaders permanetly and just turn them on/off as needed

https://i.imgur.com/YK32Vpz.png

[hardy cipher]

I don't know if it's just my limited resources or what, but if I copy paste too many times in a short time span the user interface basically breaks and I have to refresh. not a huge deal, but it's like everything disappears

[strong field]

i like this approach

[hasty mortar]

I think you should check the YouTube channel with the last event recording

[autumn forum]

Comfyroll ftw

[hardy cipher]

doesn't happen with cloning, but cloning doesn't copy connections

[strong field]

usually i just turned down the strength lol

[vale eagle]

ctrl+shift+v

[trim orbit]

ctrlshift v wow okay

[ionic gulch]

comfyui really needs more optional nodes. like turning on/off face-fix, upsaler, etc. without changing the wires

[strong field]

yes please

[trim orbit]

since we're giving paste protips, y'all know that win11 has a start+v little clipboard history feature?

i copy 8 different things then pull them all out of that

[untold goblet]

Is there a channel/forum/discord for ComfyUI?

[hasty mortar]

Looks like this lora is trained for skin and hands (not tested) https://civitai.com/models/109043?modelVersionId=117467

[hardy cipher]

as I said, when I control V too many times in a short time span, my user interface glitches and I have to refresh the webpage

[vale eagle]

oh

[hardy cipher]

but doesn't happen with cloning

[hoary saddle]

is that model on Civi? i see one for 1.5 with similar name

[ionic gulch]

is there a good upscaler for sdxl that doesn't remove the facial details the refiner added (like the one from Sytan does by using the base model for upscaling)?

[hoary saddle]

nm, found it

[indigo carbon]

Yes, it is called Crystal clear XL(CCXL)

[hoary saddle]

scrolled right past it

[hardy cipher]

all those ultimate upscale things make my computer mad. it'll be zooming along, then come to a screeching halt toward the end. I end up having to use tiled vae everything

[trim orbit]

yeah sytan's refiner pass kind of seems to not do anything outside of the default settings it comes wiht

[vagrant wasp]

can i use refiner over SD 1.5 model

[hardy cipher]

yes

[vagrant wasp]

nice

[trim orbit]

i see no reason why not

[strong field]

asking the right question

[strong field]

everything i have used removes details at upscale, like a smoothing feature

there is a workflow with a face detailer after upscale

[strong field]

these are so badass

[trim orbit]

i'd prefer using the refiner after the hires pass if i am going to use it at all. haven't gotten there yet though. i'm relying on sytan's a lot

[ionic gulch]

i'm using face detailer and upscaler at the moment - face detailer sometimes helps (but also smoothes the image a little but), but upscaler is only good for comic style. for realistic image it removes the details and makes everything look photoshopped

[soft zealot]

did someone want different lips earlier??

[trim orbit]

has a sweet AR visor inside it

[hardy cipher]

perfection

[vagrant wasp]

for refiner clip text , empty or clone the first prompt

[soft zealot]

that when I add it, seems to work better although in fairness to SDXL mostly facedetailing is superflous .

[indigo carbon]

yeah, it's so incredible we can generate stuff like this in a few seconds at this point

[trim orbit]

i'm not talking about a facedetailing pass. i use those with lora portraits

[hardy cipher]

I was looking at stuff I was makign pre-stable diffusion about a year ago

it was amazing to me at the time

not so much now

[trim orbit]

haha same. dall-e mini, now craiyon, had me enamoured

[vagrant wasp]

node ?

[hardy cipher]

first text to image I found was wombo dream. it actually frustrated me even then because it was just styles. but better than nothing

[vagrant wasp]

i use 5 lora prompt in SD, how to do that in comfy ?

[hardy cipher]

then deepdreamgenerator text to image was vastly better

bro, you just chain 5 loras together

[vagrant wasp]

eww

[hardy cipher]

you stack them on top of each other

[strong field]

waiting for runway gen3 - gen5

take one photo and make it follow my script

[indigo carbon]

T2V?

[shy kelp]

so if you're using lora in sdxl do you need to use two one for the base and another for refiner model?

[strong field]

https://tenor.com/view/rihanna-spoiled-tongue-out-hey-hello-gif-5120241

[vagrant wasp]

in comfy ?

[hardy cipher]

I don't know, I don't think it looks that terrible

[strong field]

chain them, stack em, collapse them

[strong field]

yup

[vagrant wasp]

refiner tips ?

new prompt or empty or copy paste

[hardy cipher]

yeah, I only collapse the stuff I'm probably not going to need to use again

[trim orbit]

it often seems bitter sweet to me, that the censorers are coming chicken little reveres, hold runway ml up as the hero of nsfw opensource ml, but never admit that gen2 isn't released to the public.

[strong field]

refiner steps = 80% of base steps

[vagrant wasp]

spoonfeed me

[strong field]

we will be fighting this fight till we pass the torch to the next gen

[indigo carbon]

the main issue is that videos have sound, how would that work? it can't be a diffusion process, that's for sure..

[vagrant wasp]

gen 2 node coming soon ?

😄

[trim orbit]

comfy has that new feature where you can ctrl+b a node and it just bypasses inputs to outputs

[strong field]

eh, generate the frames, ill add the sound

[trim orbit]

so 5 loras chained doesn't feel so unreasonable

[strong field]

thx, was using ctrl+m but had to reroute

[trim orbit]

same. it was an update in the pst couple weeks he put it in, and it's not documented

[indigo carbon]

that's so easy with stable diffusion already.. the issue is that it makes frame by frame, a true video model should iterate videos somehow, this also includes sounds

[trim orbit]

eh. nothing wrong with a many experts approach. apply another model ot the video post frames to get audio into it

[strong field]

https://sdxl.replicate.dev/
skip to refiner section

[indigo carbon]

then why not use SDXL for videos?..

[indigo carbon]

there should be an entirely different architecture of models capable of iterating actual videos. I remember Google being able to achieve this, but it was sooo bad

[trim orbit]

i think it lacks temperal knowledge but i might be wrong.

[strong field]

i think there will be, they are being designed and built as we speak

[vagrant wasp]

after adding loras

[strong field]

your weights are insane in the membrane

[indigo carbon]

yeah, I wonder if there will be a point where something like "VDXL" would exist, probably a year or two from now =p

[trim orbit]

the first motion pictures didn't have audio. it was added in after. nothing about a video inherently requires audio along to it and it can always be stitched in. none of this true video stuff.

[strong field]

hoping less 🙂

[vagrant wasp]

so the weights is not in the prompt

[strong field]

correct it is on the lora node "strength", set them all to .5 and see what you get

[steel mist]

I can’t imagine the stack will look anything like a “video diffusion”. So much easier to have a pipeline that looks more like:

• AI generates script
• AI generates storyboards and concept art per scene
• AI generates a blender scene
• AI controls the rigging…

Etc, etc, at each step. Rather than doing the LLM model of “just train on a million hours of video”

[strong field]

yes, lots of specialized models doing one thing

[steel mist]

Image and text generation are much more atomic problems than video generation. It makes more sense to break video down

[strong field]

kinda like ONE NODE PER FUNCTION

[steel mist]

Sure, I just think it’s going to look a lot different than SD. I wouldn’t expect a “feed a prompt, get a video” setup in the next 2-3 years. The training alone requires 24x the data. So “are there going to be generative video AI tools” in 2-3 years, yes. Will it follow the training/prompt model of SD/DalleE/GPT… yeah I just don’t see it in that time frame

[hardy cipher]

how long until text to feature-length film models?

I'm guessing the only real barrier will be resource usage

[steel mist]

If you’re talking about adding AI to existing workflows?

[crisp owl]

given where we were at a year ago, I'd say give it another year. lol

[steel mist]

I think 2-3 years we’ll see something crappy

[hardy cipher]

doubt it'll be the sort of thing that can be done on a PC any time soon, but there will surely be subscription services for it

[steel mist]

I just think the trick is training blender to build a scene to spec more than it is teaching a generative model to “make video”

Given what adobe has on deck in photoshop, etc, I don’t think “Create a scene with the following parameters and animate it in a cinematic style” is far off, and IMO could be done on a desktop.

Other than the rendering which would still require a farm

[vagrant wasp]

now we are all artist

[steel mist]

Yeah I’ll move myself over to #philosophy for that one heh

[vagrant wasp]

my lora not working out

trying to transfer SD generation over to comfy

🤣

[strong field]

SDXL cannot do 512

up your latent to at least 768, should be 1024

[hardy cipher]

valuable information https://www.libhunt.com/compare-ComfyUI-vs-LoRA

[hoary saddle]

14 seconds at 1024x1024 with that model and square resolution

[strong field]

nice progress

[indigo carbon]

did you use my latest AIT workflow? it should work with any AR

[hoary saddle]

yeah, i just dropped in your image and changed res

[spring fulcrum]

does drag and drop with comfyui images not work anymore.. I cant get it to work from discord anymore

[ionic dragon]

@trim orbit can I dm you

[ionic dragon]

the image should contain the metadata

anyone uses linux?

[spring fulcrum]

@indigo carbon can you send me the link for the model you use?

[indigo carbon]

yeah, that should do it. it calculates what resolution it should downscale to after upscaling 4x. this way all you need to set is empty latent image to the desired AR (both x and y values should add up to about 1 megapixel)

[hoary saddle]

lol, added in a image2image node to your workflow at .9 denoise, 4 seconds

[indigo carbon]

it's not meant for I2I, its built for T2I, but it seems you inferred it properly

[hoary saddle]

honestly dont know what i'm doing but seems to work lol

denoise .75, back up to 12 sec

[azure oxide]

life with SD in general 😎

[hoary saddle]

which is weird, i thought it was the other way around, less noise for i2i = faster render right?

[indigo carbon]

that's one way to do it. if you guys want I can make an I2I version. but yeah, you used it correctly here

[strong field]

go for it, i like your workflow decisions

[azure oxide]

denoise strength is different from more/less noise

i had fun playing with this setting with controlnet tile and bumping the denoise str up on some samplers

didnt make too much of a difference either way tho

[hoary saddle]

i like that model so far, much clearer than base

[indigo carbon]

same, tbh my favorite XL model yet

[candid walrus]

which xl model?

[crisp owl]

crystalclear

[pure hazel]

What do you guys think is the current best inpainting model for SDXL especially related to photorealism?
1.5 model (obviously) 🙂

[indigo carbon]

Can't SDXL itself do inpainting? if it can, why would you use 1.5 when SDXL is a waaay better model?

[urban fjord]

SDXL can be used for inpainting though I do see some artifacts around the subject after running img2img.