✨｜sdxl

[cyan crown]

I started with DB but finally I prefer Lora with Kohya

[somber hill]

well i find DB many times better

[somber hill]

but finding correct parameters took 66 trainings 🙂

[cyan crown]

I know what you mean. 😄

[upbeat summit]

looks like the work was worth it. it looks really great!

[uncut fiber]

@cyan crown qrmonster i found is 7-8 years old. What is the name of that thing for qr codes for 1.5? do you know, or link?

@shy kelp try chaiNNer, but scaling latent space can be more effective probably.

[somber hill]

yep

i should do same for LoRA too

[uncut fiber]

@cyan crown o.k. found it on hugginface

[shy kelp]

sdxl controlnet when

https://huggingface.co/diffusers

https://huggingface.co/models?other=stable-diffusion-xl&other=controlnet

[crisp owl]

sdxl already has canny, depth, scribble, openpose, and recolor controlnets

[cyan crown]

fine

have fun

[polar epoch]

For those curious. Here's peak load speed of SDXL models on sn850 gen 4, vs SX8200 pro gen 3 budget

[shy kelp]

used rush from doors roblox as control trait using sdxl model in playground ai, results was pretty neat

[uncut fiber]

Lot of things to learn. Hope i manage it 🙂

[urban fjord]

I've been trying to create my own Controlnet-lllite pose model as the existing ones are not that inuative to use and its getting promising despite the short training.
First two images are with the input "mask" overlaid on the output images. And the last two are them seperate. Given my limited training data (8 pairs) it can't do all poses but I'm happy for something that took about 2 hours of training.

It is supposed to only pose hands, but it decides that a foot passes as a hand when needed.

[upbeat summit]

so good!

[stone fossil]

Sushi + wool = Woolshi.

🙂

[cyan crown]

is there any way to apply a color scheme ?

[stone fossil]

Yes ofc prompt for it.

Tell it to only use the colors xyz.

[cyan crown]

.....I'm trying

but can't reach the exact tonality

this is what i want

this what I get with prompt

not the composition, just tonality

[stone fossil]

It depends on how complicated the subject is.

[indigo carbon]

will SDXL also have something like this? https://twitter.com/RisingSayak/status/1705223295539519858

Sayak Paul (@RisingSayak)

How about a fresh new pipeline in 🧨 diffusers!

We now have BLIP Diffusion from @SFResearch, thanks to @ayush_tues ❤️

We can now enable ZERO-SHOT subject-driven generation and transfer style from images using ControlNets 🔥

📃⬇️
https://t.co/t310DW7Zmx

Twitter · Sep 22, 2023, 2:11 PM

[uncut fiber]

4 hours to stickers end. Shouldnt we contribute?
I will try, but i will for sure spoil everything, there should be some categories but i see only BOTOs contribution.... Help me pick one so at least two ppl contribute 😄

[stone fossil]

I like the lion cats for sure the last one cute tail.

[native knot]

Before/After fun with ControlNet

[abstract depot]

First two look super similar almost identical but the last two change the face too much

Still great looking

[native knot]

Yeah...I think face detailer kinda got too much ahold of that 2nd one, but I like it enough to keep it.

I just think it's fun to take something made in a model of one type and then pull that forward into other models (or in this case a merge of 3 other models: JuggernautXL, FoddaXL, & TalosProject

[quasi sinew]

Ayo you in here too!? Lmao hello

[noble shoal]

Am i doing it right?

[uncut fiber]

yes, i think so, post it! @noble shoal

@stone fossil thank you very much

[lusty wolf]

Here's looking at you !

[uncut fiber]

they watching you!

[noble shoal]

It says its extended to next week Friday

[uncut fiber]

i got notice it ends in 4 hours, so o.k. probably they extended it

@noble shoal #1072240143521554592 message

[noble shoal]

Ok, i have added 3 to the reaction submissions

[polar epoch]

Hai! I've been here for a year and 4 days

We've been here both for a year this month and first seeing eachother first now lol

[uncut fiber]

I think you have very nice submission @noble shoal

[noble shoal]

Oh, thank you very much

[quasi sinew]

Lmao

[indigo carbon]

J XL

copax colorful

[inland vessel]

thank you!

ill check it

i hope i can get it to work

[indigo carbon]

ooooo

juggernut V2 is my favorite for sure

[stone fossil]

Last one is a song, can u guess it? 🙂

[indigo carbon]

hairy sushi (watch closely)

[crisp owl]

somewhere over the rainbow

[stone fossil]

Yes.

[olive flare]

What do you guys think of this paper ? Can it be implemented with SDXL also ?
https://github.com/ChenyangSi/FreeU

[indigo carbon]

don't think SDXL really needs it tbh

[drowsy dust]

Did you ever release your ipadater remix workflow? I wanted to try it 🤞

[indigo carbon]

no, I won't be doing that any time soon. something new released, and it doesn't use SDXL.. I won't be releasing anything IPA related until we figure out how to achieve this with SDXL.

[drowsy dust]

I hope quality is great with this level of control.

[indigo carbon]

yeah, I did manage to get a little close to it with this:

but it's not even close to what's demonstrated in that paper

so we must figure out how to implement that into SDXL

[drowsy dust]

Don’t know much about the technical but asides from IPA weight, is there a way the features so you just get the style and not both composition and style

It’s the one thing that can make it replace Loras

[drowsy dust]

Does this use controlnet?

[indigo carbon]

no, it's goal is to do zero-shot driven gen

[drowsy dust]

I mean your Minecraft generation

[indigo carbon]

but it's nowhere near BLiP diffusion like demonstrated here

so, yeah, BLiP diffusion is open source, luckily- so we might be able to learn how to achieve that level on SDXL

[drowsy dust]

Thought about making something similar to yours but I’m worried I will spend the whole weekend failing lol

[indigo carbon]

when it comes to quality of images; nothing really gets close to SDXL, but that's it, SDXL can't do any of this yet

so if we figure that out for SDXL, we should be golden

[drowsy dust]

Sdxl is quite good. It’s the easy prompting and diversity of styles I really like. Never liked very long prompts

For SAI’s model at least

[indigo carbon]

for sure. when it comes to pure txt2img; no doubt SDXL is the best. but when it comes to stuff like zero-shot subject generation, not possible yet..

[drowsy dust]

I used to be 100% sure about that before but not that certain anymore 😅
I get the MJ magazine’s and always try some prompts from it but it doesn’t always go my way

[indigo carbon]

I'd blame your inference if MJ feels better for you.. I always try prompts from MJ and SDXL never fails to beat MJ in their own prompts..

[drowsy dust]

iC2 style chair design/silver pipe/various logos and labels/black and silver mixture/in the style of Tom sachs/low back seat

[drowsy dust]

Let me know if you can come close.

[crisp owl]

When I can run MJ on my own pc, let me know

[drowsy dust]

The conversation really isn’t about that. I will always root for opensource/ self hosting. But I said I started to realize there are cases in which I just couldn’t beat MJ in terms of adherence to prompt.

[indigo carbon]

I'm loading ComfyUI real quick, I never saw MJ be better than SDXL in anything really.. give me a sec

[drowsy dust]

I really will like to gain some knowledge out of this perhaps from a workflow perspective. I have a vanilla setup in terms of node construction. So perhaps you will actually figure this prompt

[slender coral]

Could you guys recommend a way to add more detail to a scene like this:

[uncut gull]

Yes it can and SDXL absolutely will benefit from it. 😮

[indigo carbon]

why tf is everyone saying this here?

[crisp owl]

because people are clicking on the link

[indigo carbon]

so stupid

[uncut gull]

Remove the photography terms like "bokeh", because they cause blur. Remove all the negative prompts, because negative prompts always lower SDXL's level of detail. Remove the specific video game names, because specific names (e.g., "Gears of War", "Elton John", "Louvre") always force SDXL into a narrower region of its data; they hurt quality and drop detail.
Add these words: ashes, ruined buildings, scorched earth, metal debris, concrete debris, barbed wire, dust, haze, exposed rock formations, tumbleweeds, ancient, volumetric lighting, machinery debris
(Or, in general, words describing the details you want to see.)

[slender coral]

Thank you

[indigo carbon]

yeah, big negative prompts limit SDXL's capabilities, you usually just write "nsfw, painting, drawing, sketch, cartoon, manga, watermark, signature, label" or even that's too much

[uncut gull]

Aaah! I thought this was delicious sushi and then realized someone replace my nori with felt. 😱

[somber hill]

https://twitter.com/GozukaraFurkan/status/1705330920754942354

Furkan Gözükara (@GozukaraFurkan)

A very hard full body shot via SDXL DreamBooth

And I didn't even have any full body in training set

1024x1024 native

Twitter · Sep 22, 2023, 9:19 PM

[uncut gull]

Honestly, I avoid negative prompts as much as I can, including "painting". Paintings come with compositional beauty that you don't find in photographs. It's much more effective to use things like "photorealistic" in the positive prompt. (I do have to resort to negative prompts though if SDXL insists on giving me styles and subjects I don't want.)

[cyan crown]

there is a great Lora for "artist"

[slender coral]

What's your take on (masterpiece:1,2),best quality, masterpiece, highres, original, extremely detailed wallpaper,(extremely detailed CG:1.2),?

[crisp owl]

try one with and one without and make your own conclusion

[indigo carbon]

I just made a version of my AIT workflow that has a styles feature so I can just make 4 word prompts and still make masterpieces

[slender coral]

For sure, just trying to get a more technical answer.

[indigo carbon]

eat up!

[crisp owl]

When it comes to prompting, there's not really any set in stone methodology yet.
Using "masterpiece" twice seems pointless.
"masterpiece" also assumes the images trained which were tagged as "masterpiece" were actually of "masterpiece" quality.
So, as everyone has mentioned, it seems to be a pretty limiting term

[uncut gull]

"detailed" does a lot, and CG helps remove background blur. I think I tested "highres" and found out it was less effective than 8k. "wallpaper" and "masterpiece" have strong effects, so try your prompt with and without them to see which you prefer. "best quality" does nothing. I've never tested "original", but I will now! Thank you. 🙂

[slender coral]

Thanks for your time!

[indigo carbon]

I want to punch it

[stone fossil]

Owlcat. 🙂

[drowsy dust]

Did you try the prompt?

[indigo carbon]

no, what was it again?

[drowsy dust]

iC2 style chair design/silver pipe/various logos and labels/black and silver mixture/in the style of Tom sachs/low back seat

I used IPA and put death stranding inspired in the prompt lol didn’t bother using the original prompt

[stone fossil]

Owl cats rule lol. 🙂

[indigo carbon]

yeah, I'd argue SDXL nails this haha

[drowsy dust]

I like the fourth one a lot. Did you have to work on the prompt?

[indigo carbon]

on those I just wrote sci-fi chair, on the last ones I posted I used that actual prompt, can't say I'm disappointed, SDXL really nails that

[drowsy dust]

When I tried it MJ it gives similar result to yours. I don’t know how the user in the magazine got that variant with brand logos.

[drowsy dust]

I’d like to see what your IPA workflow can do though

[indigo carbon]

comparing MJ to SDXL when it comes to pure txt2img is insulting SDXL, even parameter wise SDXL is like times 3 bigger =\

the only thing I can see MJ possibly beating SDXL in is probably zeroshot subject, the SDXL image blend workflow on the other hand; can easily beat MJ most of the time

[drowsy dust]

Post some blend, I am curious 😅

[drowsy dust]

This is the first one I saw that got me interested 🤝

[shy kelp]

hmm, real pixar looks interesting

[stone fossil]

xD

[indigo carbon]

they really are pirates haha

[stone fossil]

Yarrrrrrrr. 🙂

[indigo carbon]

and the titanic is AMD somewhere in the bottom of the ocean in that gen

[stone fossil]

AMD is all cool if you just want raster for now.

Anything else team green still rocks and they know it.

[indigo carbon]

I'd love to see Nvidia get destroyed by AMD one day, Nvidia needs to be put in their place, they grew too greedy recently

[stone fossil]

Depends.

They do quite good right now I would not want to lose this.

🙂

[shy kelp]

neat

[indigo carbon]

that's not my point. I meant that Nvidia are cucking the 4000 series because no competition. if they would have competition, prices will go down and VRAM won't be cucked

[stone fossil]

Yes i skip the 40 series here.

Main rig on 3090.

Cause it was even more fun to test dual GPU amd far cheaper to step over.

[indigo carbon]

if the 4000 series would have sane prices like Nvidia used to have and they won't make the VRAM use tech from over a decade ago, I won't be angry- but that's not the case.

[stone fossil]

DLSS 3 is the main thing, but lets see what the 5090 gives us.

[indigo carbon]

so, if AMD will get back up there it will automatically make the 5000 series better and cheaper due to competition

[glad grove]

the 4070 ti 16gb looks good but idk the price maybe 600$?

[indigo carbon]

that GPU doesn't exist

[glad grove]

but they said soon tm

[indigo carbon]

I have a 4070ti, only 12gb

[shy kelp]

one of my favorite generations this week

[stone fossil]

Yeah and then you come with number of lanes used.

Which is like a crap show.

[indigo carbon]

fuck that, I will go in their walls if they won't give a refund to people with 4070ti after they give it 16gb

[stone fossil]

Because for sure if you have like a PCIe3 bor for example.

I dont say you should use a 40 serie if you do, but people do.

And downgrade theirselfs lol.

It is is a massive crap show.

[glad grove]

i also checked 3070 ti and it has 256 bit bus,they really cucked the 4000 series

[stone fossil]

3070 ti is good but the one and only thing about it, is it has GDDR6X buts only 8 so quite eh not useful.

[indigo carbon]

they did, if VRAM on all the 4000 series except for 4090 goes over 90% speed decreases in ~90%

the 3000 doesn't get effected when VRAM gets full

why? because the VRAM doesn't use tech from 2012

fuck Nvidia for that

[stone fossil]

Best always been to buy top cards.

Say you had 4 ram slots and you fit them all with 8 GB RAM.

Then u need to upgrade and ditch 2. 😉

While in first place you could have better went for the higher budget to safe money later so it is for GPU's.

You dont wanna eject em to much eh.

Swapping them costs time and money. 🙂

[indigo carbon]

idk, I might get a refund on my 4070ti and save for when Nvidia would be more sane

[glad grove]

if only i could find an used titan rtx for like 500$ 😔

[stone fossil]

Get a 3090 instead.

[indigo carbon]

yeah, it doesn't matter how dirty it is, the chip is the same chip- nothing cleaning shit can't do

[stone fossil]

3090 is the sweet spot imo upi cam go more crazy. 🙂

Next step imo would be A6000.

[glad grove]

i thought i could find an used rtx titan cheaper than a 3090 since its older but looks like no 😔 at least not now on ebay

[indigo carbon]

isn't titan rtx from 2000 series?

[glad grove]

yea

[indigo carbon]

titan rtx 4000 was cancelled I think

or is that the 4090?

[stone fossil]

TItan and Tesla are fun but I think best way to go instaed is 3090 / ti / 4090.

These have kinda over shadowed these cards. 🙂

[glad grove]

ppl in ebay really chargin 700 and 800 $ for a titan rtx when 3090 is around the same price and faster

[stone fossil]

😉

Because I think peopel in the titan GPU age are not willing to let go to what they have spent for it.

This seems a lot more doable for ex 3090 owners.

As most when kicking them out upgrade for a 4090.

Look at the 3060 lol.

It used to cost like 2 arms and 4 legs in the miner times.

But now no longer you can buy it for like a bucket of rabbit food.

Inflation plays a big part in it.

🙂

[glad grove]

yea they prob think its a collectors item🧐

[stone fossil]

To bad for them buy something else. 🙂

[glad grove]

well if nvidia kills titan lineup it will prob become a collectors item even more expensive if its sealed

[hollow crown]

x D

that was on sdxl 0.9

[stone fossil]

Not bad tho. 🙂

[hollow crown]

not nice modular i done it on comfyui

still on comfyui

a last

some clive barker way ring on my try on comfy

exiftool is like png info tool eatch image have informations =)

[stone fossil]

I swear the air keeps drinking my beer.

[hollow crown]

me too crop !

so

by use

i prefere stable diffusion 1.55

[hollow crown]

not the 1.6

♥

[hollow crown]

ha changé le processeur

=) sd 1.55

idem

en karras

avec texture inversion d'une autre entité makima

[hollow crown]

vache c'est détaillé

[stone fossil]

🙂

[hollow crown]

la derniere de faite sur makima si je la retrouve

mdr 35 minutes de rendu

Chouette en tous cas ce partage =)

j ne sais pas comment aider les équipes ca avance tellement vite

[stone fossil]

Je ne comprends pas un seul mot de cette langue mais merci allez Google traduire des choses comme celle-ci ne sont plus un problème mdr. 🙂

[hollow crown]

pas mal de controlnet perde pied sur la 1.6, sur la 1.5 ou plus c'est du betton

[crisp owl]

man, was starting to think I lost it with my gens, past couple days been fighting extra hard to get a good image. Still could, but just seemed I lost my touch.

Turns out when I was building out my main workflow and added back in the LORA nodes, I accidentally left them on a random lora, not the offset one. So it was affecting all my images.
Relieving yet frustrating at the same time

[stone fossil]

Phew. 🙂

[steady grove]

this is why comfyui is suited towards being a backend and not a frontend

[crisp owl]

could have just as easily been done on a frontend ui, the two naming schemes for the lora looked very similar.
Was just simply my oversight

[stone fossil]

What is? 🙂

[tepid surge]

https://tenor.com/view/dancing-girl-cool-listen-vibe-gif-14077809

Thaevil is music humans ✨

[steady grove]

the way automatic and sdnext both do it is the lora is prompted and not a setting. They used to do it with a programmer UI style where you had to go into the settings and select an option from a menu, but it made no sense for actual creative work.

UI's designed to be front ends are just better for creativity. I look forward to SDFX because they're hyping actual studio UX design on top of comfy. SD Swarm is coming along well too

swarm has the follow of programmer UI invading stuff too. like selecting loras from a drop down

[somber hill]

https://www.linkedin.com/posts/furkangozukara_stable-diffusion-xl-sdxl-dreambooth-is-activity-7111131970776031233-K3GJ?utm_source=share&utm_medium=member_desktop

[urban fjord]

What are the requirements to run DreamBooth training for SDXL?

[indigo carbon]

this swarm thing is really cool! it allows me to use my normal AIT workflow to keep the speeds I'm used to, but have it be a fancy interface

any idea how to make it work with batch sizes?

[steady grove]

how'd you get the ui to use the loaded workflow? whenever i go back to ui mode it uses it's own internal one

[indigo carbon]

idk, when it uses ComfyUI as a backend it just translates the workflow to a simple UI that is identical in performance and functionality- except it gets rid of batch size

[steady grove]

might not be using ait and the speeds of it's internal workflow is just fast

[indigo carbon]

no, it's identical. I looked in my backend, and output image metadata, it's using the AIT workflow

that's how it looks like when my workflow is being used

it's that, but for some reason Swarm doesn't do batch size?

[steady grove]

i dont get the hype for ait. 14 seconds but it's using a gan for scaling. kind of a cheat and it can be seen in the image quality. working with it is just so much more of a technical challenge wrestling with quality trade offs instead of a creative process. if you don't use a gan and just go latent upscaling and loras, does it help?

[indigo carbon]

nah, AIT is just faster. if it effected your outputs you're using it wrong. trust me, I have a lot of experience with this

[steady grove]

i'm looking at yours. i can tell a gan was on the finsihed product

[indigo carbon]

?

[steady grove]

gan galore here

[indigo carbon]

this works really well. idk what do you want

[stone fossil]

Ill tell you one thing and it is: soya sauce is freaking master.

[steady grove]

sdxl do work really well i know. i just think the ait layer is kind of a snake oil and might be doing less than you think. it's always a wrestle with quality vs speed. with or without ait.

with ait it just seems to have a lot more technical hurdles in the same battle

there's all sorts of gan artifact on that image up close. i think that's where all the speed gains are coming from. the gan doing a lot of the heavy lifting instead of diffusion

[indigo carbon]

you wouldn't know AIT was used here if I wouldn't say that, it seems you're just stuck on the idea of optimizations always being a tradeoff, which might be true in some cases. but not really on this one

[steady grove]

it's the gan i see

[indigo carbon]

I guess I'm blind then.. =\ , you're seeing something we don't

[steady grove]

i don't use gan upscaling in my workflows because they produce really sloppy artifacts

[glad grove]

coolest 4090 model

[west breach]

very efficient cooling system, just burn!

[drowsy dust]

You really turned that prompt into a gem

[glad grove]

thats just a real pic of your dogg 🐶

[wet nacelle]

https://tenor.com/view/the-goat-gif-4115569673926897153

BROTHER!!!

[somber hill]

24 GB card for best settings. 16 can run too

[west breach]

whoopsie!

[rare gull]

a little extra touch

[rare gull]

am in love with sdxl

[rare gull]

these cave images are generated right?

they are good enough to make me question... so, they are very good imo

[vital ermine]

https://civitai.com/models/150046

[rare gull]

anyone have any good prompts to share?

[glad grove]

this is what i call a good prompt

[rare gull]

it's certainly striking visually

[urban fjord]

Thanks, so the average person kind of have to train in the cloud here.

[rare gull]

looking hungry

[glad grove]

tasty burger

[rare gull]

am getting hungry now

[somber hill]

yep. it would cost like 1$ to train

max 2$ - RTX 3090 is sufficient 35 cent hour on RunPod

[urban fjord]

It's mostly that I find cloud training a bit inconvinient, though the longer training sessions the less of a problem it becomes.

[uncut fiber]

almost 🙂

[glad grove]

looks like the little hand from the scary movie

[wet nacelle]

SHIT CASTLE STUDIO!

[glad grove]

my honest reaction

[glad grove]

hello ladies

[hardy cipher]

he's becoming one with the stench bog or whatever that is

[glad grove]

he looks like a manatee 🤣

[west breach]

naughty kitty!

[uncut fiber]

very rarely it add legs 🙂 Makes feeling of motion

[hardy cipher]

he looks like this is his favorite treat

[glad grove]

ill eat it after im done with shower

[hardy cipher]

he's about to pop

so much energy and virility

[glad grove]

idk what happened to him here

[hardy cipher]

input these images

got these

well 3 left ones above

as you may have deduced

\

[native knot]

How was the pull to the right fixed?

[hardy cipher]

man, I have a real goofy setup, but it's working. I'm actually not entirely sure. need to investigate more

I'm using dynamic thresholding with it which took a while to tune

[crisp owl]

Try fiddling with the clip crop numbers, that's what I've been testing out

[hardy cipher]

I messed with them a bit too, but ended up zeroing them out

[crisp owl]

I think the vast majority of the time they should stay zerod, but want to test on one's that are off to the right when zerod, then changing the width crop to 256 or something

[hardy cipher]

trying to figure out the nan errors that cause it to break somewhere in the chain. well maybe not the source cause, but how to work around it

that's the same everything with a little tweak

trying some different approaches, but this one might actually work. it'd be on a case by case basis though since it happens randomly depending on seed number and a million other things

[glad grove]

nice balls

[uncut fiber]

this pumpking reminds me very Jack Nicholson 🙂

[vital ermine]

When your neck is broader than my waist you know you are in the presence of a giga chad

[clever verge]

https://github.com/comfyanonymous/ComfyUI/commit/fd93c759e278f832b149bc5b0150a8b437c48c77

Not sure if it works for SDXL though

[sour obsidian]

it does

[west breach]

what is it?

[sour obsidian]

the operation is super simple. Works on any unet though the idea of it being good on all unets is questionable. Curious to see some experimentation with it

in the unet the decoder blocks take in a concatenated tensor of both the info from the skip connections passing along encoder info and the tensor directly from the block before it. This suggests the values are not weighted optimally in training and so applying a small multiplier to give more important to the decoder block outputs gives a better image

little more nuance to where/how its applied but that is the gist

[west breach]

https://tenor.com/view/good-beardstroke-nod-i-see-gif-14497805

I will try it out

left: without, right: with

looks cleaner

sample of 1

I'll do some more testing

[glad grove]

looks like its made out of plastic

[sour obsidian]

I am always a bit skeptical of single value changes being an all around improvement in any model. Though it might reveal so interesting stuff if it turns out to be the case, or at least often enough to merit some thought on the weightings

[west breach]

Apart from training have you tried tweaking block weights to improve the image quality?

[vale eagle]

what cfg does it used?

[west breach]

5

[sour obsidian]

seen some interesting stuff directly modifying cross/self attention weights. I have played plenty with modifying various weights in the models 1 at a time to see how each part tends to effect outputs. Sometimes freezing parts of a model and training pieces specifically can improve or help smooth out things esp when trying to teach models new stuff, though when aiming generally tweaking stuff tends to have some trade offs, usually as well if something "just works" I figure its something good to learn from and might be applicable in training to just correct for it haha

[vale eagle]

Could you try 3.5 with freeU

[bright valley]

maybe in animation as well? and if so I could be filled in at least to a point of understanding the mechanics you speak of

[west breach]

I've been playing around with block merges when adding loras I've trained, since that's the only way I know how to 'fine tune' SDXL. One thing I noticed, but not sure if true, is some blocks might be linked, e.g. if you set their weights differently you end up with weird artifacts (confetti on the face, blue face, circle on the cheek)

[sour obsidian]

I mean thats def a great way of going about trying stuff! 😄 Curious which blocks you found to be linked haha

[west breach]

gotta find the images now 😄

[west breach]

in 4 & 5

same, different, different

[sour obsidian]

now when you say 4/5 how are you going about changing these?

[west breach]

3.5

[west breach]

like this

[sour obsidian]

ah hmm interesting.. They are the main blocks on input layer 2 and should be able to take dif weights but they are processing basically the portion of the latent back to back

I guess it might make sense that the blocks accounting for each layer if shifted enough might not play as next with the part next to them

[west breach]

Yeah I noticed the loras I trained seemed to be mainly affecting the IN blocks 4 - 8 and OUT 0 & 1

[sour obsidian]

ah yea that makes total sense, those are where the transformers are which are where loras operate. Out 0/1 are the two biggest transformers at the lowest latent size in the unet (besides mid as well) making up the bulk of the composition and color so it makes sense there. You should get some effect on 2-5 as well on our but dimishing only affecting smaller details

[sour obsidian]

also 6 IN you can skip for reg old loras haha. Its just a downres cnn

[west breach]

I think the anime lora I trained was almost all in 6?

I'll have to find those pics haha

[sour obsidian]

was it a locon?

that would affect it

but still not too much happenin there besides downresing to the next layer

[sour obsidian]

4,5,6 here

[west breach]

just normal lora

[west breach]

correction, it was IN 8

[west breach]

@sour obsidian with just IN 8 being applied, anime off, anime on

[hardy cipher]

I've been going to strange corners of the latent space myself

[sour obsidian]

aha yea that makes a lot of sense then 😄 Final encoder supplying its skip connection output to the first big ol decoder block. very important block in the chain

[sour obsidian]

yea that looks awesome ❤️

really clean

[west breach]

other blocks changed little things, but the style definitely was in 8

it can be hard to add in new data without changing the model too much

[sour obsidian]

yea thats a tough challenge in general, esp at smaller scales. The models are so fluid haha

[west breach]

I was trying to get the eyes fixed, but it would make the poses all weird, random missing limbs

[west breach]

why can't we just ask the model which weights to train on? 😄

[hardy cipher]

just gotta know how to ask it

[sour obsidian]

make to use "please" and "thank you" when asking too

[sour obsidian]

yea though balancing is quite a challenge haha. In some sense the data and loss functions are the ask the response is very akin to the monkeys paw riddles, where you get all sorts of unexpected outcomes haha

[hardy cipher]

interesting how it works these things out

[slender coral]

What would you guys reccomend as the best detail processor for environments / landscape?

[west breach]

I'm still new to all the training, the loss chart in tensorboard would be related to the images it's training on at the time? As it moves through the training data there are peaks and troughs because it's not looking at all images at the same time? If so, would be good to know which images it's struggling with

[hardy cipher]

I think it just hits spots that it has a harder time working through?

[west breach]

you are definitely making some interesting and cool images with unusual textures

[west breach]

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

I've been using this one that masslevel recommended to me

[hardy cipher]

still trying to figure things out. like to mess with things I barely understand and see how far I can go before they break

[west breach]

yeah me too, I just constantly tweak things and go by gut feel, because I have no idea what any of it means 😄

[hardy cipher]

doing weird things with the pooled output in conditioning. split the stream in two, do different things to each stream, concatenate. not sure what I'm achieving exactly. but seems to add something

[slender coral]

Thank you, noob question, after putting it in my upscalers I don't have access to it?

[west breach]

did you click the refresh button on the menu?

[slender coral]

Restarted comfy as well.

[west breach]

refresh the webpage?

[slender coral]

Yes, and restarted.

So I'm confused : D

:/

[west breach]

did you drag an image/load an image you created before into comfy?

[sour obsidian]

yea though loss on diffusion models is really funky and tends to be a poor indicator. Honestly I just print many many samples and compare a mix of same prompt/inference and random prompt over time once loss is platued. Also seen various people using models like Dino to do image evals but thats a bit more complex to set up haha. And ofc there is our little bot here which we have used before with the a/b for evals as well. So much of late stage diffusion model training is by feel and preference and really understanding your data and how it will likely be interpreted, from your anime tune it seems like you already got a good feel for it. Lots of playing and figuring out what is best for your stuff

[slender coral]

Just did a full refresh of the workspace (cleared and dragged the custom workspace back in)

I have the upscaler in the correct directory, right?

[floral island]

aight, it's that time again, time to go do some highres renders so i can make a nice showcase on the model update ^^

[hardy cipher]

check that cfg

it's the number that starts with 20

[slender coral]

So in Searge workflows, I had to manually set the new detail preprocessor for it to pick up, is this a bug or intentional for from the one who wrote this workflow?

[slender coral]

Also thank you this is pretty much what I was looking for

[floral island]

three where?!

[slender coral]

The voice actor died, so they won't be making a three.

[floral island]

oh no, gordon freeman's voiceactor died?! 🤣

[hardy cipher]

gordon

[glad grove]

the va that did the voice for claude from gta also died, rip 🙏 😔

[slender coral]

One final question for upscalers, where would you guys reccomend to put this one?

https://openmodeldb.info/models/4x-LSDIRDAT

[hardy cipher]

https://tenor.com/view/sonic-i-have-no-idea-no-idea-sonic-movie-sonic-the-hedgehog-gif-19665300

[hardy cipher]

https://tenor.com/view/fire-fine-fail-burn-dog-gif-5848141

[glad grove]

my man just fused with the bed

[hardy cipher]

lovely

let's see what sort of beautiful music he'll add to the mix

really captures the essence of your image here

[glad grove]

ballsy 🤣

[shy kelp]

i want a 4090

or 4080

or 4070ti

or 3090

or 3090 ti

[hardy cipher]

well, ok

[shy kelp]

https://media.discordapp.net/attachments/100633018493837312/1081358928912330863/41231.gif

[crisp owl]

cutie

[hardy cipher]

nice

[glad grove]

speaking of cuties

[crisp owl]

ooh those pucker lips

[hardy cipher]

they're having such a good time

[glad grove]

hes not happy about it

[crisp owl]

She's indifferent about it

[hardy cipher]

hmm

didn't expert her to have this impact

[shy kelp]

does comyui support postive prompts like (daylight:-1)? or is there any other way i can get away with negative prompts in my positive prompt encoder

[hardy cipher]

oh no

[west breach]

that should work

[shy kelp]

oh man ive always wanted to do it but thought it should work 😄 makes my wildcards easier to do

[tropic forge]

does anyone uses colab for generating sdxl? also i want to use the models from civitai, is that possible?

[west breach]

I'm pretty sure it's banned unless you pay for compute credits

[tropic forge]

hmmm is it? cuz using a1111 locally, when generating it took ages for me lol

[west breach]

I tried it and it didn't work :\

[shy kelp]

yeah im not getting great success with it either

[slender coral]

I'm losing a lot of detail when I try and inpaint my image, same settings just a slightly different prompt, the area you see hasn't been masked, the one on the left is the inpainted image, the one on the right is the original. Anyone know a fix?

[shy kelp]

man that sucks, there seems to be no way to combine positive and negative prompt into a wildcard

[uncut fiber]

@slender coral how about only masked? But supposing you know...

[slender coral]

Sorry I don't understand. This is the masked area.

[uncut fiber]

o.k. and you are in comfui? Is there option to only affect masked area? In A1111 there is. Anyway it is quite large mask.

[slender coral]

The issue is it's not just effecting the masked areas, it's effecting the overall image, in turn lowering the entire quality.

[uncut fiber]

hm in a1111 it used to work. Do not you using lossy image format?

[slender coral]

Lossy would be for encoding, I don't think that's the issue.

[uncut fiber]

i dont think eighter
masked-full, what is other option?

[uncut fiber]

hm... and how about lower mask blur when it is affecting whole area?

[slender coral]

I'll try but 8 shound't be changing it this much (3rd time running the inpaint) it looks like a painting now:

[ionic dragon]

which cn model should we use to get these kinda images?

[uncut fiber]

i know. hm...

@ionic dragon qrmonster?

[slender coral]

That just looks like a image to image prompt.

[ionic dragon]

its qrmonster?

[uncut fiber]

could be.

[ionic dragon]

can we even do it using canny?

or depth?

[uncut fiber]

#✨｜sdxl message
@ionic dragon

probably but according person used to use it is qr monster best

[ionic dragon]

anything for sdxl?

[uncut fiber]

i realy dont know. probably posible with canny or depth.

[slender coral]

That's just a image to image transfer

[ionic dragon]

which depth map can detect facial features?

[ionic dragon]

no

[slender coral]

I'm 90% sure it's just converting that to grayscale and then preforming a image to image.

[uncut fiber]

it is in controlnet, and i checked in temp it create b&w image in this case Trump. It made very interesting results, so i dont care what is behind it 🙂

[noble shoal]

I made a @glad grove LoRA. Something is wrong with me.

[glad grove]

he looks so handsome 🫡

[uncut fiber]

ai robot painter is Sioux! 🙂

[ionic dragon]

can you please share

[uncut fiber]

what should i share?
I did it in 1.5 with qr monster from huggingface

[ionic dragon]

oh ok

[ionic dragon]

the output

[uncut fiber]

oh it was bad. mmnt

[ionic dragon]

oh ok

@uncut fiber can we use sdxl base as the model?

i dont have 1.5

[noble shoal]

I am pretty sure you can find 1 or 2 models on Civit.ai 😅

[uncut fiber]

i can recognize hand, problem is i put there that man

#🤝｜tech-support message
here are links to 1.5 CN

[ionic dragon]

yeah i can

[uncut fiber]

whats the difference 1.5 and sdxl? Isnt it just model? I mean you can have two installation 1.5 and SDXL and you can do anything

probably canny can work for you better @ionic dragon or cant?

[rustic garnet]

inpainting can always affect other parts of the image. To avoid that, you have to blend in the inpained image into the original one

in Comfy you do that with the ImageMaskComposite Node

[ionic dragon]

im installing 1.5

[ionic dragon]

i dont think so

[uncut fiber]

Ai is painting on blue canvas with white color 🙂

[floral island]

see, true art has no limits!

[uncut fiber]

🙂

[uncut fiber]

@floral island it is lora? i like it.

[floral island]

no, pure prompt

i'm generating images for my new model post 😄

[uncut fiber]

aha 🙂

[floral island]

i managed to fix the horns/swirlies a bit

making the image output a lot more pleasing

[peak dove]

AUTOMATIC1111 SDXL and Nik Filters

[soft zealot]

NB this is not intended to be an egg sucking granny exercise 🙂

a quick & dirty comparison ComfyUI workflow showing the differences Conditioning & noise generation methods make to the final image.

Uses common inputs to generate 3 comparison images.

May help those trying to understand why they cant always replicate other peoples images from places such as CivitAI

Requires
https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb
&
https://github.com/ltdrdata/ComfyUI-Inspire-Pack

[peak dove]

Thank you

[icy brook]

anyone got different results in inference on auto11 than comfy using some LoRA's?

[peak dove]

LoRAs in CUI are so different to when I use in A1111 - e.g. Underwater LoRA, 1960's LoRA ...

[icy brook]

ok. how so?

and do you know why?

[soft zealot]

most likely for the same reason that A1111 & ComfyUI convert your text inputs erc differently as per #✨｜sdxl message

All comes down to the method used to interpret things

[icy brook]

so how do we fix this?

how can I train a LoRA and get at least similar results?

[soft zealot]

This is Stable Diffusion, This is The Way

[icy brook]

?

[soft zealot]

ie there's more than one way to skin a cat

think of A1111 & ComfyUI as translation programs , not all translation programs translate the same way

[icy brook]

yes, but how can I then train my LoRA's and get at least similar results?

[rustic garnet]

these are similar results

[soft zealot]

oh that I can't answer, I dont bother with training LORAs . Merely pointing out that things are "different" for reasons which essentially come down to core coding choices int he different UIs 🙂

[rustic garnet]

I think it's just the prompt weighting that is differently handled in comfy and auto111

[icy brook]

nope

my glitch lora on auto11 vs comfy (I know res is different but we have tested with the same, just happen to have these two image):

auto11:

[peak dove]

ComfyUI makes the Underwater LoRA very shiny - in A1111 its quite matte and dull

[icy brook]

comfy:

[rustic garnet]

you use Euler?

[peak dove]

But don't know why that shd be

[rustic garnet]

cause ancestral samplers are always instable

[soft zealot]

have you tried alternate LORA nods in COmfyUI or just theone?

[floral island]

dat tiny leg tho

[vital ermine]

yeah 😦

[peak dove]

Ancestral Samplers change the output as and when you change the iteration number. Non-Ancestral Samplers change as and when you adjust the CFG ... not an exact science, but valuable information

[uncut fiber]

why not same resolution?

[vital ermine]

LOL

[peak dove]

I have about 20 LoRAs I use in both ComfyUI and A1111 - what is an alternate LoRA node?

[soft zealot]

its an alternae node for Loading loras lol

These are the options I seem to have from different node packs Ive installed.

TBH I rarely use LORAs so no idea if they do anything different

[icy brook]

DPM++ SDE Karras

[icy brook]

What a fun time to be alive. As if finetuning had no issues before.

@soft zealot does comfy handle token order the same way auto11 does: stronger influence the earlier they show in the prompt?

[soft zealot]

pass what I know is in one of the links in my earlier message

[icy brook]

wdym?

[soft zealot]

theres a write up here on the differing methods used by the differing UIs

https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb

[icy brook]

ok, but you have no quick answer to the word order in prompts?

[soft zealot]

no

[icy brook]

ok. thanks

[slender fable]

Looking for advice, please.
The dreamstudio sometimes blurs images it creates that it deems inappropriate.
I'm concerned that the API calls we are utilizing in our app might sometimes return blurred images to our users.
Is this something that happens? (only experienced in Dreamstudio so far)
And if so, is there anything that can be done to have these instances return an error, rather than a blurred image?
TIA!

[soft zealot]

run your own server with your own SD instance on that isnt censored

[slender fable]

Thanks for the suggestion @soft zealot That is a bit beyond us

We're using rest API, so was hoping to find a solution for that

[noble shoal]

Maybe ask in #1025467151206854736

Most of us here generate locally

@glad grove Do you think somebody would vote for that in the Sticker-Scrapbook event? 🤔

[glad grove]

i will vote for sensei 🫡

[noble shoal]

Ok, then its officially submitted

[uncut fiber]

@ionic dragon
just trying, This both are same famous painting. Without preprocessor

[ionic dragon]

i didnt get you

[uncut fiber]

we didnt talk about qrmonster today?

[uncut fiber]

@floral island my answer, but you have nice star in the middle

[floral island]

same prompt? 😮

[uncut fiber]

nope, walpaper made earlier. Got just impression it is related 🙂

@floral island in this picture should be prompt

[floral island]

ah, don't have comfy opened 🙂

[uncut fiber]

np

[slender fable]

Thanks @noble shoal

[hybrid rapids]

medieval spiderman , thunder armour , rising voltage electricity discharge

[uncut fiber]

@uncut gull lora?

[uncut gull]

No, just experimenting with prompts.

[uncut fiber]

so it looks superb then. plastical gold, pearls 👍

[indigo carbon]

why is this so depressing

he happy now

[uncut gull]

It's so detailed. (The full one on the right is a 4096*4096 image.) Really makes me want to make a game.

[floral island]

very nice upscale ❤️

[uncut fiber]

latent space upscale?

[uncut gull]

It's actually img2img with the tiling script in A1111, with the 4x-UltraSharp upscale and 0.51 denoising. (Probably any upscaler would work fine though, since you're running denoising and re-generating.)

[uncut fiber]

once denoising it seems to be latent space. I will try something if i can.

[uncut gull]

Yeah you definitely get the added detail like latent upscale. Completely different from just ESRGAN or something.

[uncut fiber]

yes

There is on steam overpriced app to let you make card game 😄 , or you can do something for tabletop simulator 🙂

[uncut gull]

I've been doing game design since middle school. How I first got into programming. But the kind of uber-insanely-high-quality resources you can get out of SDXL with just a few weeks of work have just been eating at me. You can't just make anything obviously. No way to animate this quality. But if you had the right premise... I really want to figure out something that would work.

[rustic garnet]

is there already a tiling script for SDXL?

[uncut gull]

Works. (Just found out an hour ago though.)

[upbeat summit]

some electronic flowers made with the CircuitCraft LoRA by @stone fossil (base SDXL model)

[rustic garnet]

or in general: what are your upscaling strategies? I have the feeling SDXL is not good in upscaling. Doing img2img on gigh res images often looks blurry and low detail. Maybe I do something wrong, though

[rustic garnet]

and it is using SDXL, not SD 1.5?

is there a SDXL tiling script for comfyui?

[uncut gull]

I mean, check out the results I just got. Click on the image on the right and "open in browser". It's 4096*4096, and the detail is just beautiful. It's SDXL CrystalClear.
I don't use ComfyUI unless there's something A1111 just can't do, so I can't help you there, sorry.

[uncut fiber]

anybody know if plans CN qrmonster for sdxl? Trying understand it but it is SD theme then

[floral island]

wee, new model is up 😄 https://civitai.com/models/148871?modelVersionId=168123

[indigo vine]

from the owner of the repo, August12:

No plans for 2.1, but I'm in the process of making a XL version.
The 768x768 images are made with 1.5, it works well.

[uncut fiber]

@indigo vine thank you very much. i dont search on github if there, because qrmonster i found there was 8 years old.
Thank you!

[indigo vine]

I found it in the repo on huggingface. And if you want to play with 1.5, this basic setup helped me a lot: https://huggingface.co/monster-labs/control_v1p_sd15_qrcode_monster/discussions/28

with this: https://qrcode.antfu.me/#generator and this: https://antfu.me/posts/ai-qrcode-101#multiple-controlnet

[uncut fiber]

nice

[floral island]

"i'm in a good establishment, stop horsing around" -> said the unicorn

[woeful patio]

is there a way to use 2 different characters (loras) on the same image in sd 1.5?

[floral island]

regional prompter or latent couple

[uncut fiber]

pixel perfect is something less demanding?

[uncut gull]

Moderately successful 8K*8K upscale. There's some tile banding on the soft gradients in the background rays. (But Discord limits images to 4000x4000, so you can't see my beautiful 8K render.)

[woeful patio]

does that work with loras?

[floral island]

i think there's a latent couple that supports it, and i think i read regional prompter does it too, but i never use loras ever so... uhh..

don't take my words for granted

but i do think it's possible

[woeful patio]

allright thanks 🙂

[uncut gull]

Oh wait no! It's A1111 that exports JPEGs as 4000x4000! This might work. 8K render. (Yep! It worked. So much detail! 🤩 )

[woeful patio]

stupid question but is there a way to merge loras together?

[uncut gull]

Mathematically, yes. I don't think you can do it in A1111 though. (Even if the loras are on different layers and use different precisions, you can always convert and merge somehow.)

[steady grove]

wicked good. you can see great deets in the iridescence of the pearl colors

[uncut gull]

I know right! Those little flecks of gold and rainbow light. And the seams really don't show up except in the background. (I doubt this level of detail would work for a face portrait, but I might try that in a bit.)

[polar epoch]

can comfyUI properly sort models by version? Like SDXL, and one for SD 1.5

[eternal fog]

Just put them in different folders

[steady grove]

thats a hard detail to train for. the physical lines in holographic foils

[polar epoch]

Perfect!

[steady grove]

comes out really well here

[polar epoch]

@eternal fogYou wouldn't happen to know if there's a plugin/node/extension that can show image previews of models next to the names?

As that's a feature i dearly miss from automatic

[eternal fog]

Not that I know of, but I've never looked

Has anyone had a play around with FreeU on SDXL?

[uncut gull]

It only has those lines because of the necessity of stamping though, right? I have flat / non-textured foil for my Cricut Maker right here on my shelf.

[steady grove]

yeah. manufacturing artifacts

they're the subtle kind of details that make things "realistic" ❤️

[polar epoch]

Found one for those who wants it as well :D

https://github.com/pythongosssss/ComfyUI-Custom-Scripts

[rustic garnet]

not so difficult to do, just a few lines of python code. But I think kohya_ss can already do that for you

[eternal fog]

FreeU seems to give SDXL way too much contrast using the settings they have on the github page.

Left to right is base SDXL, SD1.4 FreeU Parameters, SD2.1 FreeU Parameters.

[indigo carbon]

I don't get what FreeU does to SDXL?

[eternal fog]

Yeah I dunno. Doesn't seem to do anything good with the listed parameters.

[indigo carbon]

does it change architecture?

[eternal fog]

Guess I could just try randomly change stuff and see if it does anything that looks good

I don't understand the code enough to tell

[shy kelp]

cswađš u < 0, hy C:\Users\User\Desktop\ComfyUI_windows_portable\ComfyUI\outpute

[crisp owl]

Changes the weights or something from what sdxl was trained at

It was described a bit yesterday in this channel

[steady grove]

its supposed to be way better approach at understanding prompts and turning them into quality images by doing something to the unet. similar to the cfg scaling i think.

i'm probably wrong

[indigo carbon]

heck no. this might imply I would need to provide more AIT modules for that

or just give up on AIT to use it

[crisp owl]

Here starts the conversation #✨｜sdxl message

[steady grove]

same situation with controlnet isn't it? ait fails when extra networks tie in

[indigo carbon]

not really, CNET isn't hard to use with AIT

I even managed to use IPA with AIT

[steady grove]

weird. if ait is even in the custom nodes folder, controlnet won't work for me

i tried it last week most recently

[indigo carbon]

likely conflicts with your inference's commits

I use a commit from a month ago, so idk how it works with the latest ones

[steady grove]

yeah. it's a technical challenge to deal with. very fragile library. i tend to always update to the newest versions and ride the wave

[indigo carbon]

I first read ALL the papers before I change anything with my inference, and even have backups of the backups

I've had so many bad experiences with stuff like this

I also always use Conda for it to be even more reliable, so I don't need to handle the VENV myself

[woeful patio]

ah i see yeah thanks

[uncut gull]

FreeU explanation:
TLDR; It makes all diffuser-based AIs fundamentally better, no matter what you're doing. Sound, video, images, code, text. In all cases for all applications. With no downsides, with no compatibility problems, no retraining. Just flat out better.

Background: There's a thing called "skip connections" discovered a while back that also makes all diffuser-based AIs fundamentally better in all cases for all applications. SDXL uses skip connections. (All modern diffusers do.)
Short story: Skip connections work, but the researchers discovered they have a side effect. This algorithm fixes that side effect.

Details:

1. The researchers found a way to "look inside" the u-net while it was running. That was the first big breakthrough. (Fourier transform graphs of the noise.)
2. They tested what the different parts of the u-net were doing. (We've known for years that "skip connections" improve diffusers, but we never knew why.)
3. They realized the "backbone" was doing structural stuff (like eyes, faces, fingers, windows, etc.) and the "skip-connections" were doing textures (like skin, hair, leather, etc.)
4. They wrote a function that amplifies half of the channels of the backbone (it has 4 channels if you're not inpainting), and they wrote a second function that "sharpens" the skip connections.
5. Profit!!!

[eternal fog]

It only seems to work "better" with the correct settings though and there aren't any released settings for SDXL. And out of the box it certainly isn't better.

[uncut gull]

Did you test it? Bad settings should just mean you get "smoothed-out" textures. It should still result in far fewer glitches, better anatomy, better poses, better proportions, etc.

[eternal fog]

I'm messing around with it and it's blowing out the image contrast a lot

[uncut gull]

What parameters do you have to mess with? (I don't have it, and I'm in the middle of a render.)

[eternal fog]

This is the info they give on the github page

b1: 1.2, b2: 1.4, s1: 0.9, s2: 0.2

SD2.1
b1: 1.1, b2: 1.2, s1: 0.9, s2: 0.2

Range for More Parameters
When trying additional parameters, consider the following ranges:

    b1: 1 ≤ b1 ≤ 1.2
    b2: 1.2 ≤ b2 ≤ 1.6
    s1: s1 ≤ 1
    s2: s2 ≤ 1

The awful looking one is with the SD1.4 parameters

[uncut gull]

What's the low-contrast one? Turned off?

[eternal fog]

Yeah

And then this is the SD2.1 parameters

[steady grove]

i see a lot of hype for free u. i don't know why. it just looks liek more knobs with hand waiving explanations. Quality examples just look like regular examples but cherry picked

[eternal fog]

It's because from the paper it looks "better" with no extra performance loss

So people are interested

[steady grove]

when i use it i get more of the same and once again, prompting is king. a good prompt makes good results. bad prompts make bad results.

it's literally called free lunch. the paper seems like a lot of hype for what amounts to more of the same

[uncut gull]

Could you stick to 1024*1024? Have you tried just lowering the CFG? And could you do a prompt that tends to show off glitches:
Woman writing in a journal with a fountain pen, sitting at a cafe table, crowded busy street, crowd of people, sunlight

[eternal fog]

I don't understand the details of the paper, so I can't say much other than what I see when I try and make an image with it.

[steady grove]

not at home right now where i have it set up. will experiment more with it as its getting more attention. i do want to try to find the quality gains being hyped

[uncut gull]

My render just finished. May I ask where / how you got it?

[eternal fog]

on mine CFG is 4, so I can't really go much lower

[uncut gull]

true

[peak dove]

Where did IPXL_v2 LoRA come from? And what can it do? 🙂

[eternal fog]

It's built into ComfyUI, just look in the "for Testing" nodes

[uncut gull]

Thanks, will update and look for it.

[steady grove]

this

[eternal fog]

Left off, right on

Sorry other way around

Right looks worse to me in this case.

I do have similar thoughts to iceycold, it's just another button that might make images look better, but it might make images look worse as well

Same as changing the seed, or noise or anything else.

[uncut gull]

It's not random. It specifically improves structural information at the cost of high-frequency texture information.

[eternal fog]

Doesn't look like it has there. Maybe it doesn't work as well with SDXL and that's why they've not provided any parameters.

[uncut gull]

I'm wondering if it can be used to render a base image, then img2img that for texture.

[uncut gull]

This image is too simple for it to do anything. You need things where SDXL tends to mess up and glitch on the structure. (In order to test it. Obviously I'm not saying it definitely works. I just really want to know.)

[eternal fog]

Left off, right on

left off, right on

[uncut gull]

Well, I can see the difference. The hands are the most important thing to me, and it messed those up worse. But there are so many other changes...

[eternal fog]

They look very similar in composition to me, in this these ones the contast is better, but the structure is similar.

[steady grove]

just seems like a slightly different random seed to me

you can look at all the differences and make subjective opinions about them being better or worse. i dont see it though. it just looks like it came out slightly differently

prompting is king

[eternal fog]

it seems like it manipulates stuff in the background a bit

[indigo carbon]

SDXL, right?

[eternal fog]

Yeah

[indigo carbon]

can you test if it's compatible with AIT or needs new modules?

[uncut gull]

I JUST WANT AN SDXL WORKFLOW THAT WORKS. Why are all of these offering a million features I will never use with 5 to 10 custom nodes I have to go hunting somewhere for?

[indigo carbon]

I made a workflow that uses AIT if that's what you're looking for. it also has a styles feature

[eternal fog]

Seems to work with AIT

It produces a completely different result though

[uncut gull]

No, I want no extra features, not more features.

[eternal fog]

Left is without AIT, right is with. Exact same settings.

So looks like it runs, but doesn't work properly.

[indigo carbon]

odd

idk then

if it doesn't improve SDXL I'm not touching it

[eternal fog]

I guess it's sort of helped here

Left is FreeU off and right is FreeU on

But I still feel like you could just get that by trying different seeds

It makes some things pop a bit more

[indigo carbon]

it makes it look less realistic imo

[eternal fog]

Yeah in that case it has

It does seem like it can make the shape of things better sometimes

But it's also making it look more like an art piece than a photo

[indigo carbon]

I like the normal one better here too, right looks like it's drawn

[eternal fog]

Yeah

[indigo carbon]

you can tell it's the same seed though

[eternal fog]

Absolutely yeah

[indigo carbon]

if you flip between them, the shapes are similar

[uncut gull]

Got it working! Tests in just a second.

[indigo carbon]

it's a small yet big difference, but I can objectively say it's kinda worse

[eternal fog]

I don't know why it's making it loo more drawn

Messing with the settings a bit I can sort of remove that drawn look

[uncut gull]

Okay! This is a classic glitch: SDXL is just incapable of making people hold swords. Until now!
The first one is normal SDXL. There are 0% correctly held swords:

And now... with FreeU! We get 50% correctly held swords. 😄

[eternal fog]

But it's got overbaked contrast

[uncut gull]

Yeah, I'll work on that after I test some more stuff. I'm just really interested in what structures in can fix right now.

[eternal fog]

For this one, instead of fixing the man, it just moved him very far away lol

[indigo carbon]

again, you can tell it's the same seed

just the FreeU one looks less realistic

[eternal fog]

It looks better than before now I've messed with a few values

[indigo carbon]

it's still not as realistic as without

the bushes look more artificial

[eternal fog]

It's because it's like it's pumped up the colour of them

[uncut gull]

Glitch #2! SDXL can't do wings. It always messes up the placement.

And, with FreeU, it does just as bad. (Well, technically two of these are slightly better, but not worth the change.)

[eternal fog]

lol

[indigo carbon]

i'd say worse, the colors got fucked

[uncut gull]

Ignoring the colors, because it's not worth adjusting the settings unless it makes enough of a difference. Testing still. (The only improvement FreeU is supposed to make is to features like this.)

[eternal fog]

I think the right one looks a bit better closer up, but then the background looks weird

[indigo carbon]

I know what this reminds me of now!

remember when SD2 released? the colors also looked like this

[uncut gull]

It makes text accuracy worse, so I'm guessing text isn't on the backbone's features.

[eternal fog]

Let me see what it does to an anime model, because they tend to have a lot more weird shapes and stuff that SDXL makes

[indigo carbon]

I think the model called "juggernautV2" doesn't have those issues

[vale eagle]

kind of interesting

prompt "a warrior holding a sword fighting with dragon"

with freeU default setting

[uncut gull]

It doesn't help with a painter holding a paintbrush either. I guess I'll mess with the settings...

[eternal fog]

Doesn't seem to help with holding guns either

It seems more different, than "better" at the moment with SDXL

[rustic garnet]

for me all this images look much worse. As if you used way too high cfg

[eternal fog]

Yeah it absolutely blows out the contrast

[rustic garnet]

I also don't get the reasoning behind it

[steady grove]

"SDXL can't do" ... works fine. prompt issues

that's eular a 30 steps with base 1.0

[rustic garnet]

the weighting between skip connection and previous block is learned during training, why changing it at inference?
I mean yes, you can always experiment and sometimes there are interesting effects.

[uncut gull]

I'm calling it. This is useless. The paper's authors are full of it.

I could've been upscaling beautiful pearlescent stuff for the last hour.

[eternal fog]

I think it works better with anime models

It makes the colours pop more

I'm just playing with the parameters, I think I've made it look a little better for this prompt by using parameters outside their recommended numbers,

[random yarrow]

so, what's FreeU? i've been looking through this server for some time, I keep seeing FreeU followed by a horrible, saturated image, why do people use it?

[eternal fog]

It's a new thing that was released, so we are testing it out

https://github.com/ChenyangSi/FreeU

[steady grove]

it's new and the authors are doing a lot of hype about it. weird claims about "free lunch" and quality boosts, with cherry picked examples

[supple knot]

https://www.youtube.com/watch?v=65E53rNC1io

[random yarrow]

if these are cherrypicked i dont want to see the non cherrypicked examples

[steady grove]

lol blast processing was actually real

[shy kelp]

there is a XL QR out there https://huggingface.co/Nacholmo/controlnet-qr-pattern-sdxl

(not HIRES)

[supple knot]

whaa I was trying sd2.0 qr one, but could it going

[eternal fog]

That's cherry picked, let me run another seed and see if it still makes it "better"

[shy kelp]

right loses a lot of details

[eternal fog]

It loses some background stuff, it seems to be what one of the sliders does

This was the next seed

[steady grove]

what threw me was the claims of discovering fourier transforms to look into the inner workings of the unet. it threw up bs alarms to me.

[supple knot]

Can I use this with ComfyUI?

[shy kelp]

idk, shit is slow as hell though

[supple knot]

what do Hires add?

[shy kelp]

make it more detailed, idk

[supple knot]

have you tried the SD2.0 one?

[shy kelp]

if you mean this one then yeah

[supple knot]

do you get good results?

[shy kelp]

yeah, better than this sdxl one that is struggling

one of em

[supple knot]

Among us

thats really good

[shy kelp]

realistic pixar

[uncut gull]

Yeah, that's how I feel about Pixar's latest movies too.

[shy kelp]

where are the rated R pixar movies where it's uncanny realistic look now catered to adults who grew up on pixar movies as kids

[uncut gull]

LOL okay well, let's see what we get from AI video in a couple years. 🙂

[steady grove]

few more months

its just an unfortunate truth of the internet. ideas accelerate and they often spin off towards hedonistic tendencies

[shy kelp]

ngl, i am not a fan of the jittery ai videos

[steady grove]

i was toying with animatediff's cli tools. there is a branch with prompt travelling and controlnet

[supple knot]

I'm not going to mess with this SDXL QR till its fleshed out more, I look forward to other controlnets for sdxl also. like DW Pose

[shy kelp]

mine

[steady grove]

win

[shy kelp]

https://comfyanonymous.github.io/ComfyUI_examples/controlnet/

[supple knot]

@shy kelp its cool but 2 many changes ,if you could steady the style,maybe a fixed background

[shy kelp]

30 minutes for this hires? 😭

[supple knot]

ComfyUI would be faster

@shy kelp Can you tell me about T2i and what makes it different?

[eternal fog]

@uncut gull Give FreeU a try with these settings, do you think it's "better"?

[shy kelp]

i do not know lol
(1m34s (NOT SDXL))

[supple knot]

have you got Scribble to work?

[supple knot]

The bottom "STuvwxyz "look good they all should look like that

it not bad thou, the place is like a Quarry or something

,you could run your video thru runways rotoscope tool to get a fixed background

[steady grove]

i've got stable swarm setup here at my friends with his 4080. it has the new freeu tools. they just feel like seed variations. quality increases are sometimes there, and sometimes worse. generation times are 2 seconds longer with freeu on.

its a whole lot of snake oil over here too. its kind of a waste of time

[uncut gull]

LOL you're not doing enough to make any difference IMO. But I'm sticking to my shiny renders for now.

[shy kelp]

do i have to pay for runways or is there a watermark?

[eternal fog]

It does make a difference, but it seems to depend highly on the prompt

I do agree with iceycold though, you can get the same stuff with a seed variation, for example, can you tell which of these is the original, the FreeU and just a normal seed variation

[steady grove]

often with quality gain claims, it's just cherry picking and what the real champ of the day is prompting. Prompting is always king.

[supple knot]

@shy kelp its credits and they give you so many free 150 i think ,you have to run on chrome browser for some reason

[steady grove]

i think a big part of it is the need to defend doctorates. competitive academic spaces. maybe thats where the hype claims come from.

[shy kelp]

is this the one without freeu?

[eternal fog]

Yeah

[rare gull]

fresh off the gpu

[shy kelp]

i could tell because a lot of the finer details got deleted in the other images lol

[steady grove]

3rd image i think has the best details

[eternal fog]

Yeah I don't think I'll use it. It doesn't really add anything and I have enough knobs to turn to make it do different things

[steady grove]

the differences are so slight that i don't think it can be qualified as a distinct improvement