✨｜sdxl

[eternal fog]

And if you change the settings enough, then the image changes a lot, which then what's the point lol, just use a different seed.

It's like, generate this image normally

FreeU does this

But I could just not use FreeU and use Seed Variation

Not really much difference

And it doesn't add a couple seconds to generation time

[supple knot]

different shirts

[shy kelp]

i have no idea what freeu thing is or supposed to do lol

[steady grove]

making it not a free lunch

[steady grove]

they claim some revolutionary restructuring of the unet, an effort to seperate high and low frequency generation processes, but i think it's a lot of bullshit weaving

[eternal fog]

I doubt it could be that revolutionary when it's only a very small snippet of code

[shy kelp]

⬅️ XL, MJ ➡️

[rustic garnet]

I wouldn't say so

but the paper reads a bit strange. In principal I find their observation somewhat interesting / their plots are not what I expected

I would have thought that high frequencies are dealed with in late timesteps and low frequencies in early timesteps

[steady grove]

yeah when the code released as a snippet in the readme file, bullshit alarms went off

[rustic garnet]

seems like it's much more chaotic in practice

but then their paper got really strange

they increase the impact of the decoder side of the unet by a simple scaling factor. This leads to washed out textures

[steady grove]

there is no real "high frequency" or "low frequency" passes to the architecture. These are marketing terms coming out of adobe to communicate new technologies to the end users. not legitimate modes of function

"retina display" doesn't mean that the screen is a retina

[rustic garnet]

so they come up with a lot of tricks, all of them super hacky . for the fourier transform I can imagine why it could work, but just applying the scaling arbitrary on halve the channels???

[rustic garnet]

frequencies have nothing to do with retina

[steady grove]

i think the fourier transform just shuffles the noise a bit during generation and effectively all you're getting is a seed variation.

[shy kelp]

https://tenor.com/view/bttf-back-to-the-future-gif-26440497

[steady grove]

retina display is just another one of those marketing terms that is a metaphor instead of a real function

[rustic garnet]

Fourier transform is a common thing in image processing and definitely not a marketing trick

[steady grove]

i know what fourier transforms are. i use them in signal processing a lot.

[rustic garnet]

and noise does affect frequencies differently. low noise only effects high frequencies, high noise effects all frequencies

so it's a valid assumption that the denoising process first fixes the low frequencies and later the high frequencies

[steady grove]

in this case, i think they're just affecting the noise in a way they claim improves things. but all it really does is act like a seed variation which also affects the way the noise is produced

[rustic garnet]

I don't disagree with that part 😉

[shy kelp]

i want to try this freeu thing but idk how lol

[drowsy dust]

@indigo carbon made something similar to your blend. Still testing to see how good it is

[eternal fog]

It's in ComfyUI under the "For Testing" nodes

[indigo carbon]

MJ didn't even get close here haha

[shy kelp]

someone told me MJ looks to GTA-ish

[indigo carbon]

the left looks like it could be an actual photograph, while the right looks like something you can spot was made by AI

[shy kelp]

lots of testing to get it that way

[indigo carbon]

idk, I also kick MJ's ass all the time, I just made a pretty good workflow and that's pretty much it

[tawny needle]

guys, is there any SDXL usage expert around? i'm in need of some helpz on it... to make the most of it really.

[shy kelp]

i will always love the skeleton designs like these

hmmm

can't replicate the hands again ahhhhhhhhh

[zinc cargo]

it came out very nice

[shy kelp]

(fire issues in one, ear issues in the other) idk

[shy kelp]

angy

[indigo carbon]

Why are people always making this fat gremlin?

Is this like a new trend like when people started making QR codes?

also, should I update my drivers? I find that when my GPU is over 90% VRAM full, the speed becomes a fraction of what it used to be, the older ComfyUI commits didn't do this, but the newer does. is there a driver version that resolves that issue?

[glad grove]

because its the funni

[slender coral]

Thanks so much, could I ask you where would I use one of those?

[hardy cipher]

Took me way too long to realize what these are

[supple knot]

these are cool, that ear has a finger nail \

[slender coral]

Hey lads, I'm wondering if this exists. I want to be able to generate a new image but keep a lot of the same features here (It's a part of the image I wan to replicate but make it seem like it's a new area (Down the block)) is there anything that could achieve this or is a image to image my best bet?

[hardy cipher]

lots of things could potentially help with tht

[slender coral]

Could you recommend them?

[hardy cipher]

controlnet, clipvision, ipadapter, various other things I'm sure

they all have their own approach to it but would help you keep different aspect of the image, shape, theme, structure, aesthetic, whatever

[uncut gull]

"Use the same prompt" is the best I can come up with. With SD 1.5 I would have recommended a controlnet model that we don't have for SDXL.

[slender coral]

Sorry I should have been a little more specific, I want to basically be able to "Walk" through a scene like this and capture a different sor of perspective.

[slender coral]

Understood I was just seeing if there was something like this, as there's been videos of AI shown moving through "scenes".

[uncut gull]

How about "fly" through the scene? Just crop a section, upscale it, and do img2img. As long as you don't want to turn a corner, it's doable.

[slender coral]

Yeah tried that, this is what that image gave me, was hoping it would be able to give me something similar, I'll play with the denoise, just wanted to confirm there's nothing like that.

[uncut gull]

8096x8096 upscale of figurines. (The pink SA die is under the wolf's paw in the bottom left. 🙂 )

[slender coral]

https://github.com/guoyww/AnimateDiff/blob/main/__assets__/animations/compare/new_3.gif

[uncut gull]

Denoise less on your img2img, I think. (I'm rendering an upscale, so I can't try anything for you.)

[uncut gull]

8096*8096 resolution. The wolf warrior's belt from the figurine on the left, and the rainbow shadow from the die on the bottom right.

[hardy cipher]

because 0.5 denoise changes an image a lot

if it changed it too much apply less

[uncut gull]

It should barely change at 0.5 denoise. Hang on, let me run one and show you what it should look like.

[slender coral]

Yes I know, that's why I'm confused.

[uncut gull]

Here's what I got, and I don't even have your prompt. The node setup you're using is wrong, but all I can do is recommend A1111.

[hardy cipher]

is it wrong?

[slender coral]

Wondering if anyone know's what's going on here first before I go back to auto1111

[uncut gull]

8096*8096 resolution figurines. (Last one, I promise.)

[nimble heart]

0.5 is in that swing zone where it either changes a lot or hardly at all. Depends on the image really, and if its a latent or pixel upscale.

if you really wanna be safe on a pixel upscale I wouldnt go above like 0.35 without extra guides like controlnet or something

[slender coral]

Testing

[hardy cipher]

it's all relative to input and perspecdtive

one person might say something was changed a lot when the next wouldn't

[nimble heart]

If its got something sensitive like a face then 0.5 can mess it up pretty fast

If its just like a landscape or something 0.5 might not even be enough

Also resolution is huge. 0.5 @ 3840x2160 is a much stronger effect than 0.5 @ 1920x1080

[slender coral]

😦

[nimble heart]

I usually do 0.3

On my 4k wallpapers

[slender coral]

Is this a bug? Am I doing something wrong?

[nimble heart]

lower than 0.3 and it doesnt enhance the details enough

[nimble heart]

is the left supposed to be your input?

[slender coral]

Yes

[nimble heart]

thats like not even the same image

[slender coral]

That's what I'm saying.

[nimble heart]

you're using some custom nodes so idk what's going on

[hardy cipher]

yeah, the workflow didn't load for me

maybe I did it wrong. that happens every once in a while

I'm not going to rebuild it from a screenshot

[nimble heart]

yea but it could be another node goofing its gafs no?

[slender coral]

https://github.com/SeargeDP/SeargeSDXL/blob/main/workflow/Searge-SDXL-Reborn-v4_0.json

[nimble heart]

I'm not about to install a bunch of shit tbh

[slender coral]

I'm not specifically asking you guys, I'm asking general what's going on.

[hardy cipher]

I use some of the searge stuff. has some useful things

[nimble heart]

try 0.35 on the default sampler

[slender coral]

You guys are just responding to me 😄

[nimble heart]

we have no idea cause its all custom nodes and stuff. None of those are in ComfyUI

[hardy cipher]

well it's hard to have any clue what's going on if you don't give us the information

[glad grove]

something wrong with your nodes thats why its changing too much

[slender coral]

What would you reccomend me trying?

We've established there's an issue.

[uncut gull]

Here's the one I got working earlier. Don't know why I'm posting it though, it doesn't do img2img. ComfyUI is a headache. https://civitai.com/models/147057/crystalclear-sdxl-workflow-for-low-vram

[glad grove]

try with another workflow

[hardy cipher]

https://tenor.com/view/true-story-amber-ruffin-late-night-with-seth-meyers-its-true-it-happened-gif-20030208

[nimble heart]

0.35 shouldnt change that much

0.5 can be a bit strong but the overall composition is the same

[hardy cipher]

lol, I just loaded it

[nimble heart]

so one of your other nodes just like totally replaces the image

[hardy cipher]

you didn't think that five of these would be at all relevant?

well, unless they're turned off I guess

[nimble heart]

wtf is with the arbitrary "data", too? I dont even know what I'm looking at

its green so is that a model or something?

[hardy cipher]

pretty sure it's a custom data stream of some kind

like a combination of things that don't go together normally

[nimble heart]

its like it puts all the output into a tuple or something

to avoid spaghetti

and in turn just makes it unreadable garbage anyways

[nimble heart]

at least name the streams what they actually are.
"controlnet input data" or something

[nimble heart]

ew

like its organized at least but ew

[hardy cipher]

well it's kind of a mystery flow from my perspective. nodes called magic boxes

which I am intrigued by because magic. but no idea what they do

[nimble heart]

If its using custom nodes anyways why not concat all the 400 boxes on the bottom into custom nodes too

lil schizo ngl

[hardy cipher]

it's a lot of different things

[nimble heart]

people move to ComfyUI cause auto has spaghetti code then proceed to try and recreate auto functionality with spaghetti nodes

literally insane

[hardy cipher]

well I feel like it might be too much for my smooth brain to fully grasp tbh

maybe a single megastream?

I don't know

[nimble heart]

more arbitrary 'data' types

its 100% a fucking dictionary with literally everything in it at once

wont that get cloned 800 times for the comfyui caching too?

or would it just clone the pointers cause dict is a complex type

latents arnt that big but when you have like 50 of them it adds up

[slender coral]

That doesn't make sense as they would be using a tuple.

[hardy cipher]

hrm

[nimble heart]

if its a tuple you'd have to access all 50 fields from indices which means whoever made that needs to be put in the grippy sock hospital

probably just did return (dict,)

[slender coral]

AFAIK it's not a key value, as the properties you have access to in the system are tuples.

[nimble heart]

could also be a custom class with 50 fields or whatever

so you can return (MyStruct,)

[hardy cipher]

well, there are ways of figuring this out, lol'

[nimble heart]

either way still garbage for canvas readability

[nimble heart]

I dont wanna figure it out I wanna be angry

Its been a slow boring day

[slender coral]

That still doen't explain why my image isn't image to imaging 😄

[nimble heart]

Ive just been sitting here depressed listening to my Sonic the Hedgehog mixtape so this is the most fun I've had in the last 4 hours

[hardy cipher]

ooh, I can make you next level sonic things with my fancy flow

[nimble heart]

thats 100% impossible to debug without knowing what types are used where in the arbitrary "data" streams

only whomever made it will know

hence why I was ranting about how terrible of an idea "data" is

[slender coral]

Understood, thanks lads.

[hardy cipher]

I'm just curious about it now

[nimble heart]

you'll either need a different workflow that actually takes its morning pills, or you need to literally just change a bunch of random things until it works

[slender coral]

Can you guys reccomend a workflow that has image to image, text to image, and inpaiting?

Something that will take at most 20 min to install

[nimble heart]

idk just try some find what you like. I made my own using a single custom node I wrote

[hardy cipher]

I don't think my current flow would really work for anyone but me and I'm not even sure it works for me

[slender coral]

I had a really funny feeling there was some sort of memory leak, I was correct. Restart fixed it.

[hardy cipher]

well

could have tried that

lol

[nimble heart]

memory leak

earlier I literally said the omega data stream probably fucks up the comfyui caching

[hardy cipher]

that didn't work out as I thought it would

[nimble heart]

From what I've gathered, comfy caches by keeping the output of each node in scope for re-runs

[hardy cipher]

that's why you don't tweak that data

[nimble heart]

so if you use your own FNs that directly mutate the data or do other fucky wucky stuff it messes it up

[hardy cipher]

unless you want it on your permanent record

[nimble heart]

cause it wont re-clone it every time (cause that doubles memory footprint)

[hardy cipher]

well it is efficient, aye?

[nimble heart]

yea

wonder if the cache is accessible from within a node

so I could cache the 1st pass of my 2 pass node manually

[hardy cipher]

why not?

[nimble heart]

so it doesnt re-run

[nimble heart]

it might be an engine thing that the nodes cant see from the flowchart

like the cache is entirely in runtime or something

[hardy cipher]

ahh. well I don't know enough to have a definitive answer

[nimble heart]

like no cache FNs, its managed entirely by the thread in-scope

maybe. I'd have to look

but no nodes I know of do that so I'm assuming there's no hooks or anything for manual caching

[hardy cipher]

you should see the node I've been playing with

[nimble heart]

thats the only downside with my principled node, if you modify just the refiner and re-run with the same seed it also recomputes the base pass. Same with hi-res-fix

[nimble heart]

im good thx

[hardy cipher]

lol

[nimble heart]

I cant even draw a spiral on a canvas

closest I got was a polar gradient

[hardy cipher]

you don't have to know what any of it means. just change stuff until it looks the way you want

[nimble heart]

why spend 5 minutes reading the documentation when you can spend 4 hours debugging

[hardy cipher]

well sometimes I read the documentation and it is almost impossible to understand until I see it in action

[nimble heart]

fair

[hardy cipher]

need a point of reference for some of it

[nimble heart]

but having 8 sliders with strange latin names will make it hard to learn the specific impact of each

was my 2¢

[heady vale]

does that node result in a noticable improvement of an image?

[hardy cipher]

not necessarily

but it can do all sorts of things

and yes, it can improve things

or destroy them

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

[nimble heart]

ohhh its that

i never saw the point of that tbh. high cfg always looks worse

[hardy cipher]

I was literally going to try and figure out how it worked myself so I could make my own custom nodes, lol. but then he did it which is awesome

well, it's not always the case though. if you tweak things in the right way it does not make things worse

[heady vale]

yeah not a fan of those result examples

[hardy cipher]

well they're not that good

[nimble heart]

there's like a sane range between 5-12 ish with XL and anything above bakes it while anything below turns soupy. Similar to 1.5

20 will never look better than 10 change my mind

[heady vale]

10 would be my max for certain subjects

[hardy cipher]

this is cfg 20. sure it's subjective whether it's good

and you could say that it doesn't look like cfg 20 would normally look which is also true

[nimble heart]

re-run it at like 12 or something

or 10

[glad grove]

then it would defeat the purpose of dynamic threshold,u set cfg to 30 then dynamic mimic to something low like 5 or 10 then run it to see wild or maybe shitty results

[nimble heart]

I'm not saying its 100% impossible for things to work at 20 I'm just saying theres never a reason to make it work over just using a lower CFG

[hardy cipher]

nah, not what I'm going for. I mean, I could, but I have several things precariously balanced. people are just conditioned to think high cfg = bad. but it's not always the case

great for pop art type stuff

if wrangled properly

[heady vale]

non photographic is less sensitive for sure

[hardy cipher]

oh god, I've made some abominations with people

pores that just too much

[nimble heart]

maybe cfg is just a bad method of guidance and we need something new

[steady grove]

sdxl for 🐜

[hardy cipher]

one of my favorite things with the dynamic thresholding and scheduling is it can make one part of the image almost glow it's so hot, then the rest of it is just normal

this is high cfg

not photorealistic, but I think it's pretty clean

[glad grove]

try with cfg 50

[hardy cipher]

oh snap. good idea

[upbeat summit]

Has anyone experimented with the new ComfyUI Free Lunch (FreeU) node yet?

@sour obsidian talked about it in the chat here #✨｜sdxl message. Thanks to @west breach for letting me know.

It definitely can improve proportions & coherence - consistently. But the trade off - at least with the settings I've currently dialed in - is a loss of sharpness, details like skin texture and background complexity.

The improved coherence might be a side effect of the increased contrast - noise offset like. But I'm still experimenting and trying partial, multi processing next.

This is the source for FreeU: https://github.com/ChenyangSi/FreeU

without FreeU / with FreeU

[hardy cipher]

oh yeah, I think I caught the tail end of that

[nimble heart]

What's it actually do?

[stuck bobcat]

I bet all of the cavemen ghosts that sacrificed their lives being eaten by saber tooth tigers and being stomped on from mammoths just to tell a story on a stonewall are pretty mad at AI right now

[hardy cipher]

I heard you like cfg

[steady grove]

In my tests, the improvements are very over stated, and it adds 2-5 seconds to generations on my 4080

[upbeat summit]

you are influencing the unet layers of the model

[glad grove]

daniel craig constipated

[steady grove]

any "improvements" i think are really just subjective differences that people are reading into

[upbeat summit]

the improvements in coherence are not overstated. the trade offs might very well be not worth it in the end

[hardy cipher]

those sparkling blue eyes. what a charmer

[nimble heart]

Is FreeU part of base comfy?

[upbeat summit]

yes

[nimble heart]

aight I'll pull master and see how it is

[hardy cipher]

something odd going on there

[upbeat summit]

also: https://www.youtube.com/watch?v=haL_tjJSAdc

[hardy cipher]

but if you look at her arms they're just rendered normally. but then everything else has been kicked up a notch or two

[steady grove]

in my explorations, i didn't see any consistent improvement. it was there sometimes and it wasn't others. sometimes a prompt would be consistent without freeu and fall apart with it. nothing consistent was happening and it just came down to prompting was the driving force behind good results

[nimble heart]

are there recommended values for XL? Paper only lists 1.5 and 2.1

[uncut gull]

I tested several glitches, and it seemed to help with 1, but none of the others.

[steady grove]

good prompts with freeu are just a variation seed basically

[uncut gull]

I have an 8096 SDXL portrait, but I can't post it because it's pure nightmare fuel.

[glad grove]

cant be worse than screaming steven

[upbeat summit]

sure like with everything here it's a lot of randomization, voodoo knowledge but it's fun to explore it and trying to find patterns

without / with FreeU

[nimble heart]

damn that just like deleted all the detail out of it. FreeU on right, left just has a really basic post-process to improve contrast.

[upbeat summit]

yeah, that's what it does, so a partial approach might make more sense. I don't really want the contrast increase 😄

and definitely not the detail reduction

but proportions especially in faces are getting improved - at least in some of my tests

but as said - could be a side effect of the increased contrast so it has better guidance

[hardy cipher]

I made something for you

[glad grove]

at least i can hear his voice here

[upbeat summit]

everything on 1.0 is default - so same as without. I'm currently experimenting with settings around here

[uncut gull]

Sometimes her skin is feathers, but not when it's eyes. I miss tiling ControlNet. 😦

[nimble heart]

it also breaks upscaling

[upbeat summit]

it also doesn't work with AIT

[hardy cipher]

neither do I

[nimble heart]

wait that one might be my bad

[upbeat summit]

well get yourself patched up

[hardy cipher]

but then I can't use random nodes?

I'll be honest, I tried a couple times and broke comfy until I deleted it. I realize I could have fixed it, but it didn't give me enthusiasm to do so

[upbeat summit]

yeah. I worked all week to get all my node features working with AIT and than something new comes along and I have to stop using it

[hardy cipher]

it wasn't that. literally nothing worked

I couldn't use the program at all

[nimble heart]

ok it worked with upscaling but yea not a fan I guess. It makes everything cleaner which works for the flat armor but makes literally everything else worse

[upbeat summit]

damn... but you got it fixed

[hardy cipher]

well I just removed it and restarted comfy and all was fine

[nimble heart]

Fellow AIT excluded rejoice

[hardy cipher]

not trying to knock it all

because I realize it's great for a lot of people and all

[upbeat summit]

I think a partial or multi-pass processing workflow might make sense, but we'll see

[nimble heart]

free U as base then highrers fix with non-u

[upbeat summit]

yeah

[upbeat summit]

create a base image for guidance with FreeU and than a 2nd pass without it. I already tried that but have not dialed in the settings yet to give me a good result - so not sure if it makes sense

[nimble heart]

a lot of extra work though. Not sure it'll be worth all extra render time per-image over just rendering more seeds/post process fixes

ig if you're already doing a high res fix you could just incorporate it directly

[upbeat summit]

yup. could be that it doesn't make sense to use it like that. I'm mostly interested in the possibilities to improve coherence since I'm always looking for ways to push image fidelity

[nimble heart]

I've found that compute is usually better spent cherrypicking seeds than trying to incorporate a bunch of hacks. I'd argue even the refiner falls into that category but thats debatable.

maybe its better as a one-off for when you render an image that's 90% of the way there. Turn on free-u and re-render to see if it makes a banger

[upbeat summit]

yeah could very well be

it's always a trade off. you want to brute force good images but also want a solid base fidelity

[nimble heart]

Now everyone can argue what the "Base" should be lol

Guess to me it's prompt, sampler, base model.

some might argue the refiner is part of the base which I could 100% understand

[upbeat summit]

yeah prompt is most important to me too - since I not do much else beside prompt engineering ; )

[nimble heart]

I typically re-render the same batch of 4 or so images to refine the prompt then just let it run for a while to collect some decent seeds for further monkeying using refiner/upscaling/etc

cause refiner wont really change the seed much

and upscaling wont make a bad seed good

[upbeat summit]

trade offs... better textures and some details (without FreeU) vs better proportions, better eyes, less details

[upbeat summit]

but it can very well change the style or a face

depending on the settings of course

[nimble heart]

that's probably a best case scenario because the overly narrow focus means the contrast boost from FreeU doesnt demolish details

[upbeat summit]

yup

[nimble heart]

exactly. "damn this seeds cool but the face is fucked up"

re-render with refiner

[upbeat summit]

or "that face is nice - what is the refiner doing? it changed it completely" 😄

[nimble heart]

I only have the refiner at 15% max anyways

[upbeat summit]

it really depends on the prompt and settings, yeah

[nimble heart]

and you straight up cant use the refiner for high res

so my wallpapers are all refiner-less

[upbeat summit]

the refiner is a tool for certain use cases

[nimble heart]

Ig you could for the first low res pass but all the refiner additions will be denoised out in the high res fix pass anyways

refiner's good at making mid-frequency details pop on base res

or close to

[upbeat summit]

the refiner can make very sharp details

that's apparent if you make an image with just the refiner

[nimble heart]

I found base XL actually has better high frequency details for things like texture

so skin/leather/dirt

but refiner does great with mid-frequency. Plants, etc

flower petals

[upbeat summit]

realism environments are really sharp with just the refiner

coherence is horrible, but the details are pretty good

also it makes of course no sense to use it as the only model in your chain 😄

[nimble heart]

also there's the problems of refiner+lora/controlnet/everything else

[shy kelp]

heh, cute

[nimble heart]

hence my conclusion that its not part of the "solid base"

its something you add when you feel you need it, like scenes with foliage

[upbeat summit]

yeah - use cases

[uncut gull]

The original, the classic, the unforgettable - Grid Game.

[upbeat summit]

Refiner only 😄

[nimble heart]

Yea. I feel a lot of people havent proper tried just base to understand it yet. Like the people complaining about waxy skin probably running refiner @ like 30%.

[native knot]

lul

[nimble heart]

close ups of skin is one of those areas base actually does better cause pores are too high frequency for the refiner

(in a general sense, ymmv)

[nimble heart]

looks like you forgot the caps lock was on when you drew Obi-wan and Anakin

[upbeat summit]

It does but there's still a airbrushy look to the skin no matter how you prompt. You can have a lucky seed but I worked with some fine-tunings under development that vastly improve skin pores and blemishes

[nimble heart]

yea its still airbrushy but its improved enough that I wouldnt default to the refiner with facials

[upbeat summit]

true

but you can see that they are improving it. the discord bot does sometimes output images that have better details in those areas. so they are probably testing different checkpoints on it

[nimble heart]

wonder if that's refiner or not

[nimble heart]

seems like concept of a refiner still isnt 100% certain.

like it was good for 1.0 but maybe 1.1 wont have one at all

[upbeat summit]

I don't think so from my experience because the refiner also does have a certain "look" to it

[hardy cipher]

skin smoother

[nimble heart]

yea. Like those people that doctor magazine photos to smooth the skin and fluff up the hair and enhance the general facial curves while muting the wrinkles

[upbeat summit]

the concept is great but I tried for days to make it work for me and I only use it for experiments or edge cases. I got it build into my default workflow but I'm using it very rarely right now

[hardy cipher]

think I found the largest node, guys

maybe someone made one bigger, but wew

[nimble heart]

yea. I have the refiner always on my workflow but I only set "refiner_amount" on my principled node to more than 0% when it really feels like the image needs it

[upbeat summit]

show us 😄

[nimble heart]

better not be my principled node

ive worked hard to make it as small as possible...

[nimble heart]

is that node fucking AI generated

looks like it was made by one of those python scripts that generates more python

[upbeat summit]

or chatgpt

[hardy cipher]

it's longer than that

[nimble heart]

yea that's 100% built with a for loop. no way someone made that by hand

[upbeat summit]

hahaha

user friendly

[hardy cipher]

whoa, it's several full screens all the way zoomed out

[upbeat summit]

that wins, pictures 😄

[hardy cipher]

probably about a mile long I'd say

[nimble heart]

base model vs refiner lol

[hardy cipher]

found the end

just in case you want to manually finetune every block

[nimble heart]

just in case

[upbeat summit]

post that on social media - get a cease and desist letter lol

[hardy cipher]

I might use it

I wonder if it's resource efficient

[nimble heart]

i found it on that instagram vs reality reddit thing so its already a deal

[upbeat summit]

don't forget to tweak the setting on widget 234

[hardy cipher]

oh damn, how did I not check those

[nimble heart]

i should make a 2nd pixelbuster node where instead of typing math ops into a text box, it just has 512 sliders like

add_a: 0.0
sub_a: 1.0
mul_a: 1.0
...

all the way down to arctangent_z

[upbeat summit]

so this as a node

[nimble heart]

yes

[upbeat summit]

perfect

[nimble heart]

in fact I should include 8 more copies of each in case the user needs to use variables

connected by a "copy to varN" after each cluster

[upbeat summit]

you could do one for each operation type - addition, multiplication etc

[uncut gull]

In SDXL all your pirates will have perfectly straight, gleaming white teeth, and all your Star Wars droids will have 5 fingers.

[upbeat summit]

facts

[nimble heart]

lower the CFG so their teeth succumb to the eldritch space distortions

[uncut gull]

I just wanted a golden tooth. :0(

[nimble heart]

"pirate with golden grillz" ez

[uncut gull]

Pics or it didn't happen. 😛

(I just started a 5-hour batch job, so I'm done for the night.)

[nimble heart]

ggez now where's my commission money

[uncut gull]

Sorry, I don't pay for pirated content.

[nimble heart]

Sir I assure you he's never stolen a thing in his life

Look at what a fine, upstanding young gentleman he is

[native knot]

Been messing with some upscale routines for photo-realistic that was to my liking. Using the CR multi-upscale stack, this ended up being a pretty great result.

Left is original, right is upscaled 2x.

[nimble heart]

Like, I'd give him the spare key to my house

[nimble heart]

so you ran 3 separate upscale models in series just to have the end result be a airbrushed?

[native knot]

When you look at it at the same size, it will seem "airbrushed", but take the original and scale it up raw 2x and you'll see a ton of artifacting. If my goal was to have 1:1 size, I wouldn't bother upscaling.

For instance, left again original, right is upscaled:

Sure, if I shrink the upscaled, it'll look like shit, but that's not the point; the point is up increase the size and not have it look like the example on the left.

Here's another example area:

The hair on the upscaled version is clearly better and not "airbrushed".

One last area to compare: the bottom where the fabric changes at the waist. Easily a cleaner result when zoomed to see the detail the upscale brings.

[nimble heart]

maybe?

could also just do a small denoise at your target res to get your details back for realsies

[native knot]

Denoising at the target res runs the risk of detail changes where the upscaler shouldn't change a thing. However, I find it very situational and your approach would work well for some circumstances.

[nimble heart]

I do both

run upscaler first then denoise at a very lowvalue like 0.2-0.3 to make some real details

[native knot]

As would I...just depends. Sometimes I really like a version of an image and I don't want to change anything except make it bigger.

Plus, I'm trying to find good combinations of upscaler stacks for various situations and will just have them there premade and ready to go for when I need them. I'll just label them for the situations they're good in.

I did think about using the stack in-between some samplers to push quality up as well, which I think could net some really interesting results.

[upbeat summit]

I did test 20-25 upscale models a couple of weeks ago and in all my tests (at least for realism) SwinIR s64w8 did give me the best results - definitely cleaner and sharper results compared to 4x-Ultrasharp - less artifacts, blurriness.

https://openmodeldb.info/models/4x-classicalSR-DIV2K-s64w8-SwinIR-M

In speed Ultrasharp still wins, but this version of SwinIR is only slightly slower on my setup but it's absolutely worth it in my opinion. I also compared it to the new DAT upscalers and while DAT can give you minor better details it's 2 - 3x slower.

[nimble heart]

Idk. I personally find raw upscalers kinda gross. They either have excessive ringing/smoothing/denoising or just dont work.

if its an AI generated image then you already have the prompt and settings so running a small denoise to fix the upscaler artifacts helps a ton imo

[native knot]

So, I actually ran a test on that specific image with SwinIR and ended up with some rough results that had some hashing where I really shouldn't have. I need to go back and retry it with a different order of upscalers.

[nimble heart]

XL doesnt blow up your VRAM like 1.5 does which makes it viable. I can run 3840x2160 on my AMD card

[upbeat summit]

yeah give s64w8 a try. I think it's currently one of the best. and I'm not even a fan of pixel upscalers 😄

[native knot]

Well, that's why the ultrasharp is in the middle of the stack, essentially...lol

[nimble heart]

shoot, I'll try it. I just use ultrasharp cause it gives the least blur for feeding into the denoise pass

is it runnable in comfy like any other?

[upbeat summit]

yeah just put it into the upscale_models folder

I mostly use it for realism so I can't say for other styles, but for me there's no comparison between Ultrasharp. s64w8 gets you much cleaner results

[native knot]

The other thing to keep in mind with the upscalers is that it very much depends on the input and what you are going for with the output.

Some people want something that's a bit more noisy, some don't.

[nimble heart]

github what the fuck is that speed

[native knot]

lul

[west breach]

I like the lollypop upscaler for a clean upscale, works for photo images

[nimble heart]

im not worried about clean mostly just cramming the most details into an image since the denoise cleans it up anyways

[west breach]

some add detail, but make it look scratchy, even with a denoise stage after

[glad grove]

woah u got the 500 kbps speed

rich

[west breach]

adsl speeds

[nimble heart]

fuckin people up in Warcraft III

[west breach]

running circles around people in medal of honor

[hardy cipher]

those speeds are reserved for github power users

[nimble heart]

its down to 40 right now so clearly

[upbeat summit]

when SD2 came out my workflow always included a highres fix to get the most details and fidelity from a gen. brute forcing good images was/is the way to go.

with SDXL I only just tweaked a highres fix / latent upscale workflow (I know beinsezii has been doing this for some time), but it's the same now.

you can't go below 0.5 (or even 0.55) denoising strength or your image gets all that diffusion halo blurriness and at 0.6 to 0.7 you can get amazing details but of course lots of mutations and duplication.

[hardy cipher]

the original starcraft was top notch

[shy kelp]

potion time

[nimble heart]

never played. Warcraft III was literally the first PC game I had besides space cadet pinball

[west breach]

can we share here?

[nimble heart]

im already halfway downloaded its good

up to 90KB/s now

pretty sure it took me the same amount of time to download the entire cyberpunk 2.0 patch from steam

[upbeat summit]

1 Mbit yay!

[nimble heart]

you jinxed it. 66.6 KB/s now

[upbeat summit]

it's only 40gb 😄

[hardy cipher]

I clicked and seemed and downloaded that file without even considering what it was because I live the YOLO lifestyle

[nimble heart]

"only"

my first hard drive was 35 gigabytes...

[hardy cipher]

mr fancy pants over here

[upbeat summit]

hehe

[west breach]

I think mine was like 40mb or something on my 386

[nimble heart]

yea im too young for those 500MB IDE drives

I started with Win XP

[hardy cipher]

I had a zip drive that took 250mb zip disk instead of just those trash 100 mb ones

game changer

[glad grove]

yea? well we used to share games using those 8 inches floppy disks that had like 1mb capacity

[upbeat summit]

hey at least you know IDE 😄

[native knot]

I look away for 1 second and I'm missing classic tech talk. Damnit.

[hardy cipher]

the only games that mattered to me on floppy disk were oregon trail, number munchers, and that conan game

[nimble heart]

worked in PC repair so I've had to recover deleted pictures from old people who refuse to move from Windows 95

[native knot]

Can I get a jumper to set this IRQ?

lol

[hardy cipher]

I found a hack to see people's ip address in icq

not like I made it up myself

[nimble heart]

like grandma once I clone this disk I'm throwing this tower out and making you buy a new one. The white has faded to piss yellow over the course of 30 years in your garage.

[upbeat summit]

sure, we want to play some quake and C&C later at the lan party

[native knot]

Nah, I'm too busy duping this stack of disks on my 1541s.

[hardy cipher]

cash registers still make that icq noise

[native knot]

😛

Gonna take it back further!

[upbeat summit]

I got a VC1540 in my shelf right behind me lol

[upbeat summit]

icq - I guess I dropped it after it was sold for the 2nd or 3rd time

[hardy cipher]

yeah, yahoo at one point I believe

maybe aol

all those tech giants that are still standing tall

[native knot]

Noice. I know this is #SDXL, but I'm gonna share a quick couple of pics. Give me a sec to take new ones. (I don't have my classic PCs any longer, but I do have a great classic console collection as well as a nice set of shelves for classic PC games. Uno momento, por favor.

[nimble heart]

"tall"

[hardy cipher]

I wonder if aol still makes money off clueless old people that still pay for it on top of their isp bill

[nimble heart]

probably the same people that have 3 amazon prime subscriptions

[native knot]

The arcade is self-designed and built. Has a multisync CRT from Wells Gardner, a Pixelcade marquee, and is running on a MiSTer FPGA for precision core behavior. The dust covers on the NES/SNES/Gen do have the actual consoles there as well. Also, the PS1, Wii, & Dreamcast have all been modded by me with modern HDMI as well as some other features.

[hardy cipher]

several years ago I found out my neighbor was pay an extra 10 or 15 dollars a month for those sweet aol extras. I told him he really didn't need to pay for it. but he was a devoted customer

sweet crt

[native knot]

27" in the arcade cab, 14" for the consoles.

[upbeat summit]

oh wow, soul. that is an awesome collection and setup! nice.

[glad grove]

something wrong here,i dont see a PC-FX

[nimble heart]

hey a snes. I still have mine back at the old house I'll have to bring it to my new place sometime

[native knot]

Nothing wrong with the PC-FX, just not in my stack of hardware at the moment.

[nimble heart]

never really thought about it cause I have everyting on retroarch now but it'd be cool to hook up the snes to our guest TV

[upbeat summit]

let's get that 3DO out 😄

[glad grove]

greetings fellow owner of a 32 inch crt

[nimble heart]

doesnt have the integrated VCR though, 6/10

[glad grove]

😔

[hardy cipher]

I want one of those crts with the fake wooden drawers at the bottom

one that's been sitting in a smoke filled living room for the last 35 years

[native knot]

The other pair of Detolfs...currently preparing for some figure movement, so vehicle Voltron isn't in the case. Also...open the drawer and you'll find the nice surprise! All the boxed versions on the shelves still contain their carts in nice condition while I keep my play carts here. 🙂

[hardy cipher]

dang, look at you with cool stuff

[upbeat summit]

very nice

[native knot]

Nerds gotta nerd, right?

[hardy cipher]

true

[nimble heart]

Got U.N. Squadron?

[nimble heart]

game was so hard, 8 y/o me never beat the 2nd level

[upbeat summit]

we are not going off-topic - here's one of my favorite SDXL images yet

[hardy cipher]

beauty

[native knot]

Not on console...but it's a decent shmup. Area88/UN Squadron is on the arcade.

[nimble heart]

I like how sonic's non-euclidean mouth confuses image models

[shy kelp]

vs ultrasharp v10

[nimble heart]

Arcade? Gotta have Metal Slug too then

[native knot]

All of them, yeah.

[upbeat summit]

just recently replayed metal slug in co-op

such a great game

what's your rarest game, @native knot ?

[native knot]

Across all consoles?

[upbeat summit]

or your favorite to own

[nimble heart]

all of them as in 1-3 & X or does that include 4 and 5 too?

[upbeat summit]

let's start with NES 😄

I have almost no physical retro stuff left

[steady grove]

i got a retro gameboy with a busted screen. still works, just some cracked pixesl

[native knot]

I don't think I have anything "rare", but I do have the Castlevania trilogy in the boxes, which is awesome. Also, my copy of Gradius is a really stellar hangtab box.

[steady grove]

i might dremmel it out and put a modern pi4 in it

[upbeat summit]

nice!

Gradius... awesome

[steady grove]

but i may as well just buy a pocket thing thats a fpga machine instead of emulation

[native knot]

Includes 4 & 5.

[upbeat summit]

yeah there are some good fpga powered machines out there

[shy kelp]

swinir, ultrasharp_v10, formula, esrgan, lsdirdat (i think lsdrirdat wins)

[steady grove]

https://www.analogue.co/pocket this guy

[native knot]

I have a Pocket, but I keep that by the bed for relax time.

[upbeat summit]

with the logo display - so cool

[steady grove]

i miss arcade vibes. i had it as a summer joba few years

[native knot]

Yeah, that's my Pixelcade. I actually helped the guy that makes that fix some things and a handful of my marquee art is included in his pack.

I was like the 2nd person to have one for the MiSTer.

[steady grove]

its hot

[hardy cipher]

yeah, this is another option

[steady grove]

some reason, it says cruella devil meets peg bundy to me

[native knot]

better than ted bundy

[upbeat summit]

ah very cool. I remember my time setting up hyperspin 😄

[native knot]

Did that as well...that's actually what was on the arcade before I migrated it to MiSTer.

[upbeat summit]

yeah I need to redo my retro setup from scratch. guess I will be taking a look at MiSTer

[hardy cipher]

do you have bad dudes on there?

[native knot]

[steady grove]

retro game that makes you throw or slam the controller the most?

its hard to pick theres so many

[hardy cipher]

most of them

for me

[upbeat summit]

not trying to throw anymore

[hardy cipher]

love and hate them at the same time

[steady grove]

they're not made anymore, we have to be kinder

[hardy cipher]

I never actually broke any controllers. but I've wanted to

[native knot]

Original Mega Man...so frustrating and not as well designed as the sequels. But I have a seriously nice copy of 2, my favorite of the franchise, and a decent copy of 3, my 2nd favorite.

[upbeat summit]

I mean I'm not throwing a $80 PS5 controller

[steady grove]

i never broke game pads either. they'd always just wear out at the cable connection. they were usually tanks

[steady grove]

i bought the legacy collection on steam thinking i'd use old gamer skills to fly through it. nope

barely got through 1

[upbeat summit]

modern games are too easy 😄

[hardy cipher]

one of the primary reasons they were so hard and sent you back to the beginning on game overs is because the space was so limited. not a problem these days

[upbeat summit]

I mean dying 50 times in a souls games... pff.. piece of cake 😉

[native knot]

Sorry for causing the whole channel to deviate. 🤣
(Sorry not sorry.)

[steady grove]

also, games being harder made the content last longer

[hardy cipher]

I'm just not good at them anymore. well super mario world I am still good at

[upbeat summit]

they also wanted your coins at the arcade 😄

[steady grove]

and ate more quarters

[upbeat summit]

exactly hehe

[steady grove]

the og mtx

[upbeat summit]

hey look a SDXL image on topic
1980 polaroid of a hacker wearing sunglasses using an Commodore Amiga 500, synthwave, led, nostalgic

[hardy cipher]

you need to find this https://en.wikipedia.org/wiki/Time_Traveler_(video_game)

wait, is that the one that was in the arcade and cost a dollar? that was time something

[upbeat summit]

I played this so much at Disneyland / Disneyworld

the tech amazed me

and it's so hard and unfair

but what else did I expect from the Dragon's Lair creator 😄

[hardy cipher]

I don't think I ever even saw someone beat an arcade game back then. I realize some of them just got harder and harder, but others definitely had endings.

probably a 6 second crude animated video clip and then roll credits and a cordial congratulations onscreen. and that was that

[steady grove]

sdxl n64 logo

[hardy cipher]

the reward for beating one of the very hard gameboy tetris modes was a gameboy level video of a space shuttle launch. never quite got how they decided on that

[steady grove]

that was where i worked. i played it so much. it's the same kind of hologram as those floating pennies in a mirror disc.

[hardy cipher]

cds the size of records

[steady grove]

laser discs were a thing. yup. they sure were

[glad grove]

have u ever played super mario rpg legend of the seven stars? now thats a good game

[hardy cipher]

oh yes, top notch

[steady grove]

looking forward to the remake

probably won't have axem rangers though. hasbro hated that

[slender coral]

Anyone suggest a solution for warping like this?

[hardy cipher]

nope

[native knot]

Inpainting if you have to.

[slender coral]

Any ways you know that would fix in the generation, or is this normal for SDXL?

[native knot]

That's not just normal for SDXL, it's normal for every AI image generator. You're just going to have some seeds that will do that.

[shy kelp]

not as cool as yours

[glad grove]

u could try to draw on top of it then img2img

[slender coral]

There's no warping on this.

[native knot]

Not every combination in the infinite set of combinations will be perfect and not all will have issues. Literally change one number in your seed and you might get a completely different result. Hell, one change to CFG could do it too...or your sampler, or your schedule...

Shall I go on? 😄

[steady grove]

it's in the 4th dimension. to see if you've got to take a step to the left

wait. jump to the left, step to the right

[slender coral]

I'm asking what would help me not get warping. So if you have a suggestion on what to play with yes.

[shy kelp]

the lights on this forklift is how i feel (i am not forklift certified)

[native knot]

Literally everything I mentioned could net you a result that is what you are looking for. But it's not something anyone could magically say, "oh yeah, just change X to Y and it goes away"; that's not how this works.

[slender coral]

Mate, I fully understand that, I'm asking for suggestions.

Not being told I can't do it or what I'm not doing...

You guys undertsand that right?

[upbeat summit]

I think he's pretty cool 😄

[shy kelp]

i sadly do run into warping issues on certain generations with this model though

[glad grove]

just keep generating m8 sometimes it just happens

[hardy cipher]

lots of people run into warping

[shy kelp]

i rarely go past 4 cfg

[native knot]

Nobody's saying what you can or can't do. We're literally giving you suggestions.

[slender coral]

The only one who gave a suggestion was i don't know just n ow.

[native knot]

just keep generating m8 sometimes it just happens
This. Exactly this.

[slender coral]

The rest of you guys are telling me to just do generations.

[slender coral]

testing thank you

[native knot]

Bruh.

[slender coral]

That's not a solution.

[steady grove]

is to

[native knot]

That's a suggestion and a solution.

[slender coral]

I'm asking about generating, not inpainting.

[hardy cipher]

I'm actually not going to offer any more advice at all

[native knot]

It's just not the one you want.

[slender coral]

Great thanks for coming out.

[shy kelp]

im about to showcase some warping though even at 2 cfg

[hardy cipher]

because you seem to have contempt for people that suggest things you didn't want to hear

[steady grove]

you could prompt better. try "no deformed"

[native knot]

No, icey, that's not a solution...unless he wants it to be one, then it is. 🤣

[hardy cipher]

well what he's not telling you about are the 600 other nodes impacting the image

he won't show them

[native knot]

Seriously, though. Everything that's been mentioned here is a viable possibility.

[slender coral]

Open photoshop and draw it yourself yes it's a solution but it's not what I asked.

[hardy cipher]

who cares

[steady grove]

actually, deformed is one of my standard negative bits. "deformed, demented, disconnected, broken, dead, old, disfigured"

[native knot]

That's why we didn't say that. We gave you solutions related to image generation.

[shy kelp]

not gonna lie

[slender coral]

Lower cfg and higher steps seels like a good direction thanks. .

[native knot]

same...it's in my neg as well.

[crisp owl]

best thing to test what's doing what is of course to do more generations.
Keep the same seed, everything else the same, and add some prompting. See if one thing prompted vs another changes things as you wish.

SD often isn't just a process of one great image after another. I've sometimes worked on getting a single image right for hours

[hardy cipher]

lower cfg to 0 and leave it there

[glad grove]

try with this negative

[crisp owl]

At least they included "latent space" in there.

I hate when my images contain left-over latent space

[native knot]

So, lowering your CFG also increases the possibility of losing elements that follow your prompt. I also mentioned a change in CFG as a possible solution above. You could literally change it by .5 and get a better/worse result.

[steady grove]

make sure bad anatomy has an emphasis of about 12.8, roughly

[native knot]

I try to keep all of my latent space next to my dark matter...just to be tidy.

[steady grove]

the model knows what bad anatomy is and is just leaning that way

[native knot]

Careful not to mix them, though.

[hardy cipher]

what I normally do is max out that aesthetic value in sdxl

[steady grove]

whats the max

[native knot]

https://tenor.com/view/jason-momoa-chair-interested-gif-9751403

[crisp owl]

over 9000?

[steady grove]

a1111 has a good feature for learning and doing your own investigations into the model. understanding how your prompts do things. xyz plotting with prompt s/r

[shy kelp]

[native knot]

lul

Now that's a prompt.

[shy kelp]

warps

[native knot]

Get all the good, none of the bad.

[hardy cipher]

running at 999 aesthetic score atm.

[steady grove]

woah slow down man. you dont wanna blow the fuse

[steady grove]

i'm not sure if that does have warping. it doesn't look like an era of precesion engineering

[native knot]

Right? 100% perfectly perfect thatch roofing. ™️

[shy kelp]

honestly, it's not that bad

[native knot]

Ladies and gentlemen...we did it. The "good/bad" prompt is all we needed. Pack it up, boys!

[steady grove]

stable diffusion solved. whats next

[glad grove]

https://tenor.com/view/drake-gif-21416115

[native knot]

Thank goodness. I can get back to folding protiens and reading radio signals from space.

[slender coral]

What was the prompt for that one so I can see if I'm at all close to that?

[native knot]

Man, what is the prompt for that?

"isometric post-apocalyptic landfill oil factory"?

[steady grove]

make sure bad is in the negative prompt so that nothing is bad

[slender coral]

closer

[shy kelp]

isometric perspective of gorgeous fantasy market village

[slender coral]

(masterpiece:1,2), best quality, highres, original, wallpaper, (extremely detailed CG:1.2), Perfect Light, isometric realistic top down, sci fi farms, walkable path, post-apocalyptic, perfect portrait, hyper detailed, volumetric lighting, octane render, 8k, HDR, RTX

```

[steady grove]

"Award winning drone photography"

[glad grove]

RTX,16k,official,IGN,CHAD,

[shy kelp]

helps

lol

[shy kelp]

[crisp owl]

"we require more minerals"

[native knot]

I've suddenly opened up Factorio.

[steady grove]

command and conquer sprite vibes

[steady grove]

i've actually laged hard on keeping up with sdxl practice because iv'e been playing a mod pack on factorio

[native knot]

So easy to get sucked into that game if you're in the right mood.

[steady grove]

most of my pc gaming life has revolved around openTTD, so when factorio came out i was all 😮 i know this!

[native knot]

Man...just look at them straight lines.

[shy kelp]

[native knot]

😄

[slender coral]

:feelsbad:

[shy kelp]

i sometimes toss random words into a prompt to see what wild shit it throws back at me

[steady grove]

wish sd could do animations like that

[glad grove]

have u tried to put good in positive and bad in negative 🤔

[steady grove]

soon

[native knot]

I've done that before, yeah...in this example though, my neg is empty. Absolutley nothing in there.

Positive was:
isometric post-apocalyptic wasteland landfill oil factory 35mm photograph

Didn't have the 35mm photograph for the earlier ones.

[steady grove]

roofs aren't perfectly square and level. clearly fake

[glad grove]

but u didnt put (masterpiece:1,2), best quality, highres, original, wallpaper, (extremely detailed CG:1.2), Perfect Light, isometric , ultra mege hyper ultimate perfect portrait, hyper detailed, volumetric lighting, octane render, 8k, HDR, RTX in prompt thats why u didnt get a perfect pic

[native knot]

fuck.

💀

[steady grove]

can see warping in the dirt

[native knot]

3 models merged at and even 0.33 for these:
7 of 9, Juggernaut, & fodda
Also using the add-detail LoRA at 3

[steady grove]

and the cloud puff looks warped

[native knot]

Just remember:

[steady grove]

the fire is warps

[native knot]

Bus bench billboard Brad sees you while you sleep:

[steady grove]

i can see warping on the hair

[shy kelp]

[ionic gulch]

Version 14.0 of my workflow, with FreeU settings, first beta of artist styler (letters A-D):
https://github.com/JPS-GER/JPS-ComfyUI-Workflows
if you prefer to use your own workflow you can still use the SDXL Prompt styler node - especially the new artist option gives some great results IMO.

[native knot]

Nice job, JPS.

[upbeat summit]

thanks for sharing it, JPS!

[native knot]

Toying with that other prompt that I was doing upscaler comparisons and got this absolutely great lighting on the left side of the hair on this particular gen. It's otherwise normal, but that hair lighting is fab.

[slender coral]

Mate, where did you get that workflow 😄

[shy kelp]

just prompting and merged models

im picky so i spot the tiniest thing that's wrong and dismiss the entire image (even though i can healing brush with photoshop if i really wanted to)

[steady grove]

the mud is warpped

[shy kelp]

lol

cool car, nonsense background though

wait a minute

[native knot]

The building is warped.

[shy kelp]

[native knot]

There you go.

Should probably put warped in there, too.

[crisp owl]

and "latent space"

[native knot]

lol

[shy kelp]

sometimes ill put something like person so it adds people. so i remove person and put in people and it gives me one person or a crowd. sometimes, not mentioning people will give me no people

[crisp owl]

don't think about a pink elephant

[shy kelp]

but it's an elephant right?

[crisp owl]

Yeah but it's brown

[steady grove]

we talkin bout the elephant again?

[crisp owl]

is it in the room again?

[steady grove]

https://tenor.com/view/zootopia-zootopiagifs-stretch-gif-5167916

[native knot]

I found it, guys.

[crisp owl]

at least it can actually fit through the sliding door

[native knot]

Shit...they're multiplying.

Bigger room or smaller elephants?

[steady grove]

the ceilng fan is deformed

[shy kelp]

elephants aren't pink what the heck

[crisp owl]

https://tenor.com/view/how-dare-you-screaming-lady-rage-angry-gif-15730816

[native knot]

These guys beg to differ.

[steady grove]

[native knot]

This is getting out of hand.

[crisp owl]

are they conspiring now?

this is what happens when you don't include "latent space" in your negative prompt guys

[steady grove]

what happens when stable diffusion starts to empower the kind of people that makee this stuff https://youtu.be/l5OR6k1fqdM

[native knot]

I really don't know what to do now, guys. Halp.

[upbeat summit]

better turn down the CFG

[native knot]

They've started organizing.

[shy kelp]

wait

[steady grove]

this is what happens when you turn the aesthetics up to 999. i tried to tell you

[upbeat summit]

oh no - they aligned! getting ready to march

[native knot]

Damn. My bad.

[native knot]

Oh snap. Their king.

They've gathered supplies.

[native knot]

They've decorated their room with it.

[shy kelp]

https://tenor.com/view/james-franco-wait-what-wut-wtf-nani-gif-25479769

[native knot]

They've gone pink elephant militia on us and now have pink camo.

These guys are here to see a special guest...

[vital ermine]

Here we go this is the one as I am testing

[upbeat summit]

that's a huge army

[uncut fiber]

here we have insted of pink elephants little white mouses.

[native knot]

The special guest speaker the elephant army was here to see:

[supple knot]

for text you write "says" in prompt

[vital ermine]

didn't work

[native knot]

The white mice army is amassing...

Oh shit.

Guys....

The mice are getting serious.

[crisp owl]

I knew all these mice used in science experiments was gonna end strangely

[native knot]

I think they've come for revenge.

[vale eagle]

Do we have Token Merging for stable diffusion in Comfyui?

[native knot]

The mice have an artillery unit.

[native knot]

Now do a pink elephant.

[steady grove]

elephant sized hippos

[native knot]

Looks like the mice have air support as well.

[native knot]

Their navy may be a bit outdated.

[uncut fiber]

they are fighting on same side with pink elephants or are against each other?

[native knot]

They're fighting against the pink elephants.

[vital ermine]

That has to hurt

[nova lion]

hi, did anyone know how to create multiple animals in one image? i can't seem to get it right, always mixed.
eg: elephant, snake, giraffe, crocodile, in prompt mixed up to multiple creatures.

[shy kelp]

could try one elephant, one snake, one giraffe, or inpaint them in

[nova lion]

thanks, i think i will inpaint them. one elephant, one snake, and so on still can't get it right.

[steady grove]

the minerals are warped

[uncut fiber]

lennon is seems to be longer time zombie

[floral island]

it's organic crystal 😮

[vital ermine]

Everybody do the Twist

[shy kelp]

special ops

[vital ermine]

Mouse Army will fall

[glad grove]

the rat pack

[native knot]

I feel like we need to make this mice vs elephants conflict a perm thing on this server now

Choose a side kinda deal

I'm going to sleep, but when I wake up, I expect military operations to continue. 🤣

[upbeat summit]

sleep well!

[floral island]

leave me out of it!

[vital ermine]

Crack head

[uncut fiber]

how those pink elephants started?

[crisp owl]

don't think about pink elephants

[uncut fiber]

because i got it in stickers things for sure before all those 😄

[crisp owl]

it's an old psychological example. Tell someone to not think about pink elephants, but in telling them that, they are thinking about pink elephants

[glad grove]

he can fight them

[crisp owl]

It was used as an example back for I think 1.5 when talking about the negative prompt, how sometimes you can put in a negative prompt term, but in so doing, you get that very thing in your image, where it may not have been created in the first place.

[uncut fiber]

yes i know 🙂 just coincidence with my sticker 🙂
@crisp owl thanks for clarifying it to me

[uncut fiber]

actualy not sure, i know elephants got fear from mouses. Because they bite their foots. Difficult to predict

[shy kelp]

is this the action lora? i like prompting in this lora with the power turned down

[shy kelp]

i dont use any lora

[crisp owl]

🤔
Anyone seen this warning message before?

WARNING:accelerate.utils.modeling:The model weights are not tied. Please use the tie_weightsmethod before using theinfer_auto_device function.

got it when cancelling a gen I queued then started another one quickly after before the other fully cancelled.

[uncut fiber]

Because SD so spoilering it. It invented "I" quite funny but original way.

@crisp owl in comfy or a1111?
i noticed big troubles in A1111 changing prompt during processing.
But never see this before.

[crisp owl]

In Comfy, first time seeing that error before.
So far seems a one off, I didn't have it happen again so far

[uncut fiber]

if it gone so it is o.k. 👍

[vale eagle]

I am started to think FreeU is usable

left is w/o freeU. right is w freeU

[uncut fiber]

From Budapest?

[vital ermine]

Freddie Mercury, the Legend.

[uncut fiber]

i know just asking on concert, In Budapest, or in forgot that famous england stadium now 😄

[slender coral]

Hey mate, I'm trying to try out your workflow to see what I'm doing wrong, any idea what I'm doing wrong with this now 😄

[uncut fiber]

wembley

[vital ermine]

Right, I was agreeing

I am trying out my new lora

I trained it from 1 image

[uncut fiber]

i am trying old CN model, unfortunately it is only for SD now

[vital ermine]

It was a serious learning experience working with only 1 image

[uncut fiber]

how long training?

[vital ermine]

5m

how long learning was all day

[uncut fiber]

🙂

[vital ermine]

It picked up the style very well, and even created its own style too so I got a two out of the deal

One is like a comic book that it created

[vital ermine]

hands 😦

[rustic garnet]

I remember you people talked about an llm + lora trained on generating good stable diffusion prompts?

has anyone a url?

[noble shoal]

I use https://github.com/Zuellni/ComfyUI-ExLlama-Nodes

If it is that what you mean

Works for me even without lora

[woeful patio]

is there a good way for getting seperate people with different loras in the same image within comfyui?

[rustic garnet]

really? cause chatgpt gave me always just descriptive prompts, but no good tags

[uncut fiber]

@woeful patio only can imagine regional prompt, but only in theory

[noble shoal]

I never cared so much about the tags. Here is an exaple prompt output: "A mysterious forest where trees are made out of diamonds and crystals. There's a clearing in the middle that leads to a magical lake surrounded by white marble statues. The sky above is filled with floating islands covered in purple flowers. In the distance, there's a castle made entirely out of ice and snow, towering over everything else."

[noble shoal]

Another example: "A forest at night where everything is glowing blue and purple from bioluminescent mushrooms. The trees are covered in fairy lights that twinkle when touched by a gentle breeze"

[woeful patio]

Does that support loras?

I dont understand how it doesnt exist yet, it would be so handy

[rustic garnet]

yeah sure, but it's not so much about creating interesting prompts but about enhancing prompts

[noble shoal]

You could feed your prompt into the ExLlama Generator and let it improve it. Will try to work out a prompt in a moment. @rustic garnet

[uncut fiber]

i think it support loras, dont know why it shouldnt. But realy dont know

[rustic garnet]

yeah, I know, if you make your prompt right that works somehow. I just thought there is a lora specifically trained on SD captions

to add the right styles and tags matching your prompt

[woeful patio]

no worries, thanks

[noble shoal]

Pinging @fierce hollow . The creator of the nodes

[fierce hollow]

Annuvin desu

did something break 😭

[noble shoal]

@rustic garnet has a question about a existing llama loras for sd. No worries, nothing broke

[nimble heart]

on the contrary, everything broke and it's your fault specifically

[fierce hollow]

There is a lora trained on captions but I didn't ask if I can share it, @ionic dragon has it

oh no, anyways

and I made a lora loader for the nodes, it works, probably

[rustic garnet]

haha, thanks, I will ask qwerty

[ionic dragon]

??

[rustic garnet]

you have a lora for llm trained on SD captions?

[fierce hollow]

they want your llm-sd lora (what do you even call this)

[ionic dragon]

yeah, its not a full-fledged one, its just for testing purpose trained with 80k prompts, you wanna try it?

[ionic dragon]

idk, didnt think of a name yet

[rustic garnet]

sure, which base model did you use?

[ionic dragon]

Mythomax

[fierce hollow]

it does work pretty well compared to no lora imo

workflow here just in case

[vital ermine]

https://civitai.com/models/150871?modelVersionId=168645

[rustic garnet]

haha, yeah, good example

llms tend to be way too talkative

I think I will first try it in textgen webui

[ionic dragon]

you dont need any instructions, i didnt train it with instructions, just input&output

[fierce hollow]

yeah but mythomax itself was trained this way iirc

or at least the models it was mixed from

[ionic dragon]

sent the lora in dms

[fierce hollow]

that said it works with chat format too

[ionic dragon]

llms are flexible

so its not an issue to use it either way

[fierce hollow]

you should just upload the lora to huggingface btw

there aren't many if any

[rustic garnet]

https://huggingface.co/Gryphe/MythoMax-L2-13b <-- this one?

do you use quantized model variant or fp16?

[fierce hollow]

GPTQ from TheBloke