✨｜sdxl

[soft bone]

1 hour and 1.5gb but i consider this nailed

[inner ruin]

is that with Caith's params?

[soft bone]

no very very different actually. 3e-4 and 256 dim

with reg no captions adafactor and celebrity token for 4 epochs 20 repeats

25 images would be probably 2 hours. regularization doubles training time

[somber hill]

it found me similar to

i dont see any resemblance

i think this strategy fails :d

testing top 10 now

['Ayad Akhtar', 'Dariush Kashani', 'Daniel Zolghadri', 'Amir Talai', 'Nik Dodani', 'Yousef Abu-Taleb', 'Nicholas Kadi', 'Omar Maskati', 'Samrat Chakrabarti', 'Michael Aronov']

nope none like me

i will use my deepface script :d

[inner ruin]

my take is using the name of celebs is superstition that came from a shitty paper 😄

just use token + class and it should work

works for me at least

[somber hill]

it has a logic if you dont train text encoder

[somber hill]

ye this is proven tactic

[inner ruin]

nahh

assessing likeness is super subjective and the paper had bad methodology

[somber hill]

i am using deepface to asses likeliness. it is working really good

consider it as face recgonition

[alpine pine]

Assesssing likeness is subjective?

then how does facial recognition work?

[barren olive]

im getting "Some backends have errored on the server. Check the server logs for details." but how am i supposed to know how to check the logs?

[soft bone]

Joe Penna personally commented to never use a random token, on @somber hill 's video.

As a lead dev on the model, i aint gonna argue with him

I'm thinking we can get away with just using the same skin tone and gender.

[alpine pine]

would the wording of the skin tone matter in this case?

[soft bone]

no i mean a celeb with the same skin tone and gender

[soft bone]

might i see that process?

[alpine pine]

ahhh, okok

[somber hill]

ye full process shared here : https://youtu.be/343I11mhnXs

i will generate faces with your dataset and test with my script

[trim orbit]

this is quality peer review 🙄

while i don't use that approach, it doesn't seem like a fair assesment

[inner ruin]

see I tried it and it just mixes me with the celebrity

like yeah it has more flexibility

but it's like... some mix of me and a celebrity

[inner ruin]

and these were overbaked

[soft bone]

i look forward to it, post your findings

[barren olive]

with StableSwarmUI do i need to do anything special to get it working when i dont have a nvidia gpu?

[somber hill]

i suspect same thing

realism would be hard

my linkedin profile is exactly as me

https://www.linkedin.com/in/furkangozukara/

i made it with sd 1.5

https://youtu.be/g0wXIcRhkJk

[soft bone]

train longer

[inner ruin]

lol I can say the same for [token] person

[soft bone]

i mean, i tried that and it was still bad. clothes were very limited and had artifacts, as well as surroundings sometimes. lighting and expressions had lower variety. using the celeb name gets you to ideal convergence quicker no matter what.

[glad grove]

wdf goin on in ussr bro ☠️

[soft bone]

random token vs. celebrity name. identical training & inference params. left is so far off its a joke

likeness is 97% accurate on the right btw, doesnt resemble celebrity used at all

[alpine pine]

"-doesnt resemble celebrity used at all" so would this mean it doesn't matter what you think is the celebrity equivalent, it just matters what the interrogator (or clip) thinks is the celebrity equivalent?

[somber hill]

but it produced very different faces

looks like unet spit out all known ones :d

[soft bone]

it really matters what XL itself thinks but we cant ask it. but again i think skin color and gender might be plenty of a match to use. The idea is getting a human starting point vs. building a concept from scratch. details like eye color and nose shape would hardly save you time

[somber hill]

i have pretty much bases to compare

so we will see

first i need to find best celebirty match

keeping my tutorial repositories working really taxing. every update breaking something haha

[soft bone]

gpt can help!

[heady vale]

Nvidia keynote on AI breakthroughs in about 16 hours including content creation https://www.youtube.com/watch?v=3qSQjRaseos

[rustic garnet]

I usually use textual inversion with a real name as starting point.
However, I found for SDXL textual inversion overfits very quickly, so using celebrity names might be the easier way

[somber hill]

did you guys ever tried extra install options of kohya?

like deep speed

accelerate on windows

[hardy cipher]

I've used deepspeed with oobabooga

I've read it's more for training? but whatever it does is quite, sped things up substantially

[soft bone]

I've heard it used in context of optimizations that sacrifice quality (like mem efficient attention or grad checkpointing, and the other extra options), so I never use it

[somber hill]

someone said me that it moved a lot of thing to ram

so he was able to do sdxl training on 8gb vram

but it was super slow

anyone used multi gpu of kohya does it work?

[soft bone]

bahaha my ram is already hit with 24gb for lora training so i cant afford more

[somber hill]

ye he had 128 gb ram

[autumn forum]

what is the difference between kohyass gui vs sd scripts? do you need both or does kohya ss contain sd scripts? reason im asking if because of an error im getting when trying to merge loras so maybe i boched the install

[hardy cipher]

I have 64gb of ram just sitting around with nothing to do normally

[rustic garnet]

sd-scripts is the original one

kohyass gui is a fork with a simple user interface

[hardy cipher]

I want to modify one of these pipe nodes so it will input/output all the SDXL sampler requirements

and also seed number

[rustic garnet]

sd-scripts is usually more up to date

but branches might be different. For example, SDXL was for long time in the dev branch of sd-scripts but in the main branch of kohya gui

[autumn forum]

thats confusing.. lmao so theoretically, everything in kohya ss gui should work without having to install sd scripts correct?

[rustic garnet]

in comfy? I can help with that, but not tonight going to bed soon

[rustic garnet]

yes. It's an outdated superset so to say 😅

[hardy cipher]

nice! yeah, in comfy. I'm not really well versed on this stuff but I'm looking at the code now. doesn't seem super complex to do

[rustic garnet]

yeah, there is a nodes.py class

you basically just have to look at the implementation of the KSampler node

then write a custom node that calls the same code but adds additional ouputs

[hardy cipher]

ahh, well I'll see what I can figure out. I'll probably have to fail a few times, but I'm persistent, lol

I appreciate the info. I've been thinking about messing with the nodes for a while now. just haven't really followed through yet

[rustic garnet]

feel free to dm me if you have questions. But I'll go to bed now 👋

[hardy cipher]

all good, thanks!

[autumn forum]

it wont even let me delete my current kohya install to install fresh becausde it says invalid file handle lmao

[tribal jackal]

I keep getting distracted with other stuff, and my BestBoyNido Lora still isn't having issue, but I also have experienced this issue consistently with other models in something as simple as 16:9.

I've tracked down this old comment, and found the relevant section but I'm not sure which XY node you used to set up the aspect ratio.

also that screenshot right here is one hell of a rabbit hole

[alpine pine]

hmmm... the aspect ratio script is just a simple math node. Just takes three numbers and outputs four. The full script is one math equation converted into whole numbers.

I can make a simpler comfy workflow to more clearly show how it works, but the X and Y values should just directly input into the image width and height. The clipX and clipY go into that advanced clip node, they are not always necessary though as some workflows just use the image resolution for clip.

[echo pine]

I have a question, is Micro-Conditioning from the paper also related to training? or is it only related to generating?
If I select don't upscale buckets in training a Lora (train on the original resolution of the images) and then do Micro-Conditioning in generating in comfy does it know that it should use the quality of the images above 1024 px but still use the knowledge from the images under 1024px?

[hardy cipher]

sometimes I read comments in here, and although I understand all the words, I don't really grasp what they mean in the order they're typed

[tribal jackal]

Don’t worry about it. I have a few ideas in mind, busy with life again, and I plan to learn it anyways. Thanks bud.

The manual inputs look promising

[soft bone]

anyone familiar with bbox and sam settings?

[hardy cipher]

man

I've been trying to figure that out myself

I've just been leaving most of the default settings for this face detailer. but I would like to actually understand what's going on

[soft bone]

i just want it to inpaint at full resolution. how could that be so hard

[hardy cipher]

every time I start figuring something out I turn the corner and I'm bombarded by 50 new things, lol

[soft bone]

i sometimes avoid this chat for fear of learning i've been doing things wrong.

ignorance is bliss

[hardy cipher]

I think with a lot of it wrong is relative

in the end what looks good is largely subjective

[soft bone]

oh yeah, with training i trust no one. i used every config i could find and got crap. had to blend them all together

[trim orbit]

UGH. i hate form interfaces sometimes. i just spent 70m training a lora at 0.1 dim, because i guess while i was trying to scroll the page, i had my mouse over the dim form and it was like "Okay boss! MINIMUM DIM IT IS"

[hardy cipher]

I'm just having to play a lot of catchup because hadn't really messed with python before early this year. and hadn't messed with programming in any significant way in years. did some arduino stuff but that was mostly just copy pasting and tweaking things. and also knew little to nothing about AI or machine learning. so basically I knew nothing at all

[trim orbit]

othertimes i only hate them a little

[soft bone]

being good at training but new to code. life is hard

[trim orbit]

proof of work. its why crypto needed gpus. using gpus for proof of work is now obsolete

[hardy cipher]

I had a complicated workflow going in comfy and zoomed out to move everything around and organize. and turns out you can still adjust settings while zoomed out

that made some interesting things

I'm just hoping someone comes up with an improvement on the nested nodes. it'd be nice to have something that was collapsable but not necessarily nested

that's his thing

[soft bone]

this is why it worries me to hear how crucial the bots were to training XL. ppl cant be trusted with RLHF lol

[hardy cipher]

how do those channels even work?

[soft bone]

reinforcement learning by human feedback. main reason MJ is so good

[hardy cipher]

bot-1, bot-2, etc. I guess I should just read about it myself

[trim orbit]

well, no. we don't know any reason why it's good. only they know how they did it. people speculated that MJ uses RLHF

bruh

[hardy cipher]

ahh, cool. I'll check it out. thanks

[trim orbit]

i'm so sour that the dim was set to 0.1

[soft bone]

i wouldnt call it speculation considering it would be incredibly easy for them with the bot system and dumb not to do it. and it's the only thing the other models arent doing yet theyre not as good

[trim orbit]

well, it is speculation. even an educated guess is still a guess

[soft bone]

yes true. i'd just put my money on it

[trim orbit]

my point is is that MJ keeps everything very secret and proprietary. they could have a multitude of experts style system behind the scenes

which is most likely

[echo pine]

they use aesthetic guidance, they have the images and they all have a rating score like 1.25456 or 0.54612 , they then train a model on that scoring guidance thing, thats the reason they give away gpu hours to people who rate a lot

[trim orbit]

and yet all of that is hearsay because they keep everything secret. they're really good at it and are likely getting people to sign ndas

i like his hat

[soft bone]

new server pic?

[echo pine]

i presented david the idea of rating and aesthetic score and giving away gpu hours for training the score thing through dm back when midjourney had like 15k users he answered 'interesting' and then they implemented the rating thing

[trim orbit]

he probably heard the idea through 100 different facets since it was buzzing in the entire academic community

same way you heard about it. it was just a good idea that was proliferating

[hardy cipher]

wow, such beauthy

[trim orbit]

i always knew @hard fractal . out the airlock you go.

[echo pine]

yeah sure, it's an obvious idea on how to improve the model, but I'm just saying they would have no reason to collect all that data and give away free gpu hours if they wouldn't use it

[trim orbit]

right. but i wasn't saying "they dn't use it"

[trim orbit]

"main reason" can't be known. the main reason is likely the quality of the training data and their ability to control prompting the model are bigger reasons in my opinion. Prompting especially. Prompts are king and they likely have a preprocessor that takes what people intend to prompt, and gives it to the model in a better more effective way.

we can speculate all day and whether or not they use RLHF probably doesn't matter much. Their quality comes mainly from other aspects of their system

[echo pine]

yeah there's lots of things they can do with the stuff they collected, they can determine which image + prompt is preferred, they can train a lora-like model that they can weight towards the 'MJ aesthetic', they can use the highest rated images in their dataset to make the model understand more complex concepts, they can use the prompts of the highest rated images to train a prompt generator that automatically adds some words that fit the prompt and model to enhance the images, they can collect preferred parameters and settings etc..

[trim orbit]

point of the day: MJ is a closed system thats very gaurded, and theres not much the open source community can learn from them. They're certainly not leaders of the research field and don't share any break throughs.

speculation is really just another form of gushing over them

[hardy cipher]

just had bard write me a top notch prompt for sdxl

Base
Positive G: A cute cat
Positive L: A realistic cat
Negative G: A scary cat
Negative L: A cartoon cat
Refiner
Positive: The cat should have big eyes and fluffy fur.
Negative: The cat should not have a tail.

[fervent ermine]

飛機

[trim orbit]

i'm still pretty unimpressed by holz last flop. leap motion's point was what? expensive gimmick really.

doing metalayers far far more closed garden the facebook ever could. so happy leap motion never caught momentum

[soft bone]

this you?🤣 #✨｜sdxl message

[hardy cipher]

I believe so

[pallid path]

does gligen work with sdxl

[autumn forum]

today is just not a good day for me. lol. getting this error in comfy 'NoneType' object has no attribute 'lower'
all the models and upscalers are all in the right directory in comfy so idk whats going on

[somber hill]

my little dragon

[glad grove]

bro thinks he the dragonborn

[copper rose]

After a ton of work I finished my SDXL finetune with a very similar style to the original ReV Animated
https://civitai.com/models/124736/psy-animated-xl

Hope you guys enjoy ! 🙂

[hardy cipher]

nice!

that became my favorite SD model

[upbeat summit]

looks great. thanks for sharing it!

[urban fjord]

Is there anything new from the SDXL team regarding ControlNet that I may have missed?

[trim orbit]

hwo do i help you win your 4090 that your model deserves ?

[hardy cipher]

I need to load this immediately. lol

do we have to post images or something? I scanned the criteria and realized I couldn't win it myself so didn't retain much of it

[trim orbit]

they teased me with it last week. bullies.

[copper rose]

You're too kind, the best thing you can do to help me win the 4090 is give it 5 stars and share it with your friends 🙂

[copper rose]

I think its all about 5 star reviews, pictures and hearts! Sharing your awesome work with this model will definitely help it get to the top of the page!

[maiden gale]

@boreal bough I have trained a lora with you settings and reg images, I must say it is looking much better and flexible!

[azure oxide]

man, SAI was so excited to see what the community was going to do with SDXL when it was released 2 weeks ago.
Doggystyle Plowcam - SDXL just popped out at me as I filtered the civitai search for SDXL stuff lmaoo

[maiden gale]

Thank you, do you have any recommendation for using this settings? High cfg, refiner etc... @boreal bough

[upbeat summit]

It's great that the scene is so active. Of course you will have lots of varied content. There are already a couple of interesting and promising releases that show what SDXL can do.

People are already preparing for stock and product photography with fine-tunings like this 😄 https://civitai.com/models/124094/taiwan-chicken-cutlet

[trim orbit]

https://tenor.com/view/fallout-liberty-reprimed-fallout4-thumbs-up-gif-5417364

[trim orbit]

i've noticed that the censorship harbingers are very quiet

[astral jay]

Base SDXL Lora

[heady vale]

😜

[trim orbit]

looks just like his metaverse avatar

[rich coral]

Hehe, was just thinking that the dawgs seemed awfully quiet lately and must be preoccupied with something.

[trim orbit]

https://tenor.com/view/honey-pooh-jizz-gif-20662275

[raw prairie]

How do i use loras in comfyui?

[trim orbit]

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

best way is to open a workflow someone has setup already

[hoary saddle]

do you need any special nodes to create wildcard prompts?
for example: a bull wearing (hat crown mohawk) and (business suit tshirt jacket)

[raw prairie]

Thanks

[autumn forum]

i cant believe this is a thing.

[crisp owl]

Hey man, no kink shaming

[autumn forum]

chicken cutlets kink? wtf

[crisp owl]

I'm kidding

Kinda

[soft bone]

@upbeat summit what's your favorite sampler for XL?

[upbeat summit]

currently I'm using dpmpp_sde_gpu karras at 50 steps / cfg 4

Not the fastest. I know 50 steps is a lot with SDE. Some would say it's total overkill. But I like the outputs of what I'm currently doing.

with SD 2.1 in a1111 DPM++ SDE Karras at 25-30 steps was a great fidelity sweetspot

[zinc lynx]

ancestral variants have been doing well for me with karras

[upbeat summit]

yeah there is no best sampler. everybody got their favorite ones - their own voodoo and convictions 😉

[zinc lynx]

from what i understand gpu is non deterministic

[upbeat summit]

yeah. I chose it mainly because afaik in ComfyUI dpmpp_sde_gpu karras comes closes to the DPM++ SDE Karras implementation in a1111 and I'm very used to it.

of course you should use a different setup if you want to do studies or analysis. but beside doing lots of experiments I like the output / aesthetics of the current setup.

[runic lark]

Does concat add a comma when not using linebreak, or should you always end each textbox with ", "?

[upbeat summit]

doesn't add a comma. just a \n

[zinc lynx]

i believe you need the comma added yourself

[runic lark]

Ah, ok. Thanks! 🙂 Sorry for wrong channel, even though it's sdxl base model 🙂

[upbeat summit]

all good. lots of ComfyUI topics are discussed here

[visual glade]

yeah this is the unofficial comfyui channel

[runic lark]

And no one has understood how to use the SDXL prompt yet I guess

[crisp owl]

There is no "right" or "wrong" way really. There's the idea behind what it's doing, but ultimately, if you get an image you like, great. If not, change it up, figure out what works for you, and roll with it

[zinc lynx]

25 steps, 20 base 5 refiner is my starting point rn

then another 5 with upscaling if i like that output

its fast

[upbeat summit]

I guess no one knows 100%. there are theories and you can definitely influence an image either by combining the prompt or sending 2 different ones to clip_g and clip_l. but I wouldn't say that there's a verdict about one or the other - at least from my experiments and what I've seen.

[zinc lynx]

i would have to agree with that but im not convinced

[crisp owl]

Clip_g is a newer clip style, which tends to follow more conversational style prompts.
Clip_L is the same as the clip used in 1.5, so more comma separated key word focused.

[runic lark]

My workflow is basically first with sdxl and then refine the img with a 1.5 model. Feels like the texture isn't all there yet for sdxl.

[zinc lynx]

this is more to my understanding and how i have my current prompting setup

linguistic postive to g and supporting terms to l

and then fundamental negative prompt

can share my workflow if anyone wants it

[upbeat summit]

but you can also send 2 different negatives to the encoder ,)

[crisp owl]

they both makeup the cross attention and how it hits 2.6 billion parameters compared to the 850m or so of 1.5 and 2.1

[upbeat summit]

yeah that's how I got it as well

[crisp owl]

Basically, they interact with an absolutely mind boggling space to come up with your image

[zinc lynx]

i had a workflow i was working on with the negative prompting setup the same way but i find the negative to be less important

[midnight shuttle]

Review of the code has shown that this is not the meaning of these encoders. CLIP-G is the newer more natural encoder used by SDXL and SD2. CLIP-L is the older tag-based encoder used by SD1. When the model was trained, the weights from both encoders were concatenated. This is also how generation is handled.

[crisp owl]

drawing attention to one or another weight, and evaluating the difference between them

[midnight shuttle]

Negative seems hardly needed with SDXL. Only when you do a generation and then specifically notice something you don't want. The super long negatives of SD1 and SD2 aren't as important.

[crisp owl]

My negatives went down to a few terms, compared to the....many...of 1.5

[zinc lynx]

i often start with no negatives and work up

[upbeat summit]

It works very different to previous SD versions. it can be very powerful. in SD1 and SD2 I overwhelmed and confused the model with heavy negative prompting to get better fidelity - it worked quite well.

In SDXL negative prompts are more powerful. But since the model is so good now, I made many images without any negative prompting and when I use it I try to keep it as light as possible. I still get nice fidelity enhancements using them in some prompt builds.

[zinc lynx]

which is why i simplified the workflow to not have 2 negative prompts

[crisp owl]

I generally just add a couple very opposite terms. Then add another if I need.

If I want something liquidy, I put solid, if I want vivid, I put dull.

[zinc lynx]

if i remember that image i posted just now had no negative prompt

using negative embeddings has been interesting

[bronze kestrel]

Hi @glad flint did you ever figure out the seamless texures with SDXL? I have been banging my head on how to get it work within a model with API access and not just a webui.

[midnight shuttle]

It would be interesting to get comment from the SD devs about why they used CLIP-G and CLIP-L in SDXL. Why not only use the newer CLIP-G? Is it to allow use of older style prompts? Or is it because it works better when both types of prompmts are used? Only the devs can answer.

[upbeat summit]

yeah as solesbeedude said - negative prompting is not needed most of the time with SDXL. it's just that good 🙂

[autumn forum]

i love how 2m karras makes people old in my workflow haha

2m karras vs ddim

[zinc lynx]

ive noticed this sometimes as well

[autumn forum]

2m karras very noisy but can add detail

[soft bone]

real vs fake imo

[bronze kestrel]

Any assistance on SDXL generating seamless tiles via an API / a model that can be implemented as such? I've searched all the comments but trying to do this via something other than Auotmatic1111 etc. so it can run via API

[zinc lynx]

you guys using the offset lora?

[crisp owl]

pretty much always

[zinc lynx]

comfyui?

[crisp owl]

yah

[zinc lynx]

do you have a prefered strength for the model and clip

[autumn forum]

never lol

[crisp owl]

I see zero difference playing with the "clip" number, so I keep that default at 1.000. Strength I default that at 0.25 or 0.3. Then adjust if I feel it needs more. Most often that suffices my needs

[autumn forum]

i feel like sdxl did cara delevingne dirty, i mean her eyebrows are thick but they aint neanderthal level thick

lmao

[crisp owl]

negative prompt: caterpillars

[midnight shuttle]

Is that base model or LoRA?

[heady vale]

werewolf brows

[autumn forum]

me? base

[midnight shuttle]

Yeah, that seems like a good case for a negative prompt.

[nocturne dove]

If the lora is trained with u-net only and not the text encoder, would the "clip" number do anything? I assumed not but I also don't know how the offset lora is trained.

[autumn forum]

kinda gross how realistic this is lmao

[midnight shuttle]

CFG 20

[autumn forum]

ill fkn do it

[hardy cipher]

those brows, woo

[autumn forum]

i had to put eyebrows in the negative or else they were solid black bars man

[soft bone]

correct, lora clip multiplier does nothing

[hardy cipher]

you should see if you can accentuate the eyebrows more. maybe put "wilford brimley style eyebrows"

[soft bone]

eugene levy

https://tenor.com/view/schitts-creek-surprised-oh-eyebrow-raise-wow-gif-10754465

[hardy cipher]

good call

those are nice

I wonder fi he dyes them, or if they're just naturally like that

[glad grove]

let me ask him

[hardy cipher]

yes, please do, mr seagal

https://tenor.com/view/steven-seagal-eating-carrot-healthy-living-gif-9753152

[autumn forum]

same some eyebrow for the rest of us dang

[soft bone]

delete that

[autumn forum]

lmao

the real question is. can they go bigger...

[glad grove]

add zachary quinto to the positive

[autumn forum]

its not changing much anymore. i think ive reached peak eyebrow

maybe lets try a different seed

[crisp owl]

What is she a Chia Pet or something?

[zinc lynx]

anyone feeling cyberpunk vibes

[autumn forum]

i cant get any bigger but this is a bomb photo. moving on. sorry for cara spam

[zinc lynx]

1.0 or 0.9 VAE?

[autumn forum]

0.9

[soft bone]

i was just making some of her but eyes are just so messed up 😦

[crisp owl]

Sorry, I had to. Added Chai Pet Eyebrows. Got a bit more eyebrow action

[heady vale]

these eyebrows... Im trying to eat over here

[autumn forum]

the only way ive achieved decent eyes has been to upscale and do another pass on base

[zinc lynx]

prefered upscaler?

[crisp owl]

I prefer NMKD Siax

[soft bone]

even with face inpainting.. eyes are my arch enemy

[autumn forum]

did someone say steven seagal chia pet?

[autumn forum]

agreed, the eyes suck in sdxl. wish they were better. theres always one f'd up one

[hardy cipher]

https://civitai.com/models/118427?modelVersionId=128461

[soft bone]

saw that. more for inpainting

[glad grove]

not thicc enough

[autumn forum]

does that actually work? or only close ups

[hardy cipher]

I don't know, it doesn't say it's for close up eyes

[soft bone]

its trained on closeups of eyes so you pretty much have to inpaint with it. its not an eye helper like a neg embedding or anything

[autumn forum]

ahh eye see. hehe see what i did there.

[hardy cipher]

lol

oops

[autumn forum]

cha cha cha chia

[hardy cipher]

is that ben shapiro?

[zinc lynx]

now make a bonsai tree

[autumn forum]

no thats cara this is ben

[hardy cipher]

I made a bill murray bonsai tree at one point

I wonder if I can find that

[autumn forum]

bonsai tree is extremely strongly biased in sdxl

[hardy cipher]

orly?

[crisp owl]

playing around with my img2img stuff.
Bow down to the Raven King

[soft bone]

have you figured out how to get facedetailer to inpaint at full resolution like auto1111 does? if the face is far enough away it doesnt help to inpaint it cuz it's still rendering so tiny🤔

[hardy cipher]

whoa https://civitai.com/models/124773/morphed-muscle-sdxl

[crisp owl]

Add the facedetailer after a hires upscale

[soft bone]

ugh i never upscale

[crisp owl]

ah, that's the only ways I've been able to get the facedetailer node to look good enough

[hardy cipher]

sometimes I upscale then downscale

[autumn forum]

made this with face detailer, realistic vision face sdxl everything else

[hardy cipher]

I think I need to figure out how to finetune the face detailer. because it's made some pretty harsh looking things for me

ran the refiner afterwards and it took care of it. so there's that

[soft bone]

Oh i got rundiffusion on civit and it instantly improved my lora's eyes but not nearly enough. So i expect finetunes to eventually solve this problem (considering 1.5 base outputs demons but realistic vision did that^).

[hardy cipher]

I'm surprised realistic vision hasn't made a couple xl models by now

[zinc lynx]

im waiting for AbsoluteRealityXL

[crisp owl]

I thought for certain Realistic was gonna be one of the first

They did state their V5 for 1.5 is their last update. So I'm sure they're on SDXL training

[hardy cipher]

they kind of started flooding the market though

new model every 3 weeks

[crisp owl]

make em, sell em

[hardy cipher]

yeah, didn't that dude sell out? or people, seems like it would most likely be a single person though

[crisp owl]

I mean, profits profit. If it puts food on his table, good for him

[astral jay]

Same Lora, 5 different checkpoints including base sdxl. Votes for the best?

[hardy cipher]

oh, I'm not hating

just saying

3 does have some nice attributes

[astral jay]

3 is base sdxl 😛

[zinc lynx]

2 best overall, i like 5 as well

[hardy cipher]

nice

[autumn forum]

what was 2 kinda curious

[astral jay]

dreamshaperXL

[zinc lynx]

that was my guess lol

[crisp owl]

Guess they're doing something right

[hardy cipher]

did you use noise offset or skip steps?

[astral jay]

clip skip -2

no noise offset

[hardy cipher]

nice. I'm trying to figure out optimal approach with those. I realize it varies depending on all the things

[soft bone]

do you find dreamshaper lessens likeness? she looks the most different in that one

[hardy cipher]

least inbred model

[azure oxide]

interesting backgrounds on those, wonder if thats sdxl or dreamshaper making those multiple repeated trees

[astral jay]

I thought the one that's the most far from the true likeness was number 4

Made it look almost cartoonish I feel like

[soft bone]

true slightly anime weighted it seems

[astral jay]

Best likeness goes to talmendoXL and dreamshaper in second

[soft bone]

something important to me is lighting realism. the light on her arms is so unrealistic in most of them. 1 looks the best for that tbh. which model is that?

did you try rundiffusion?

[astral jay]

anime looking one is sdxlUnstableDiffusers

[astral jay]

lemme try

[zinc lynx]

i like rundiffusion

[soft bone]

same, night and day improvement for portraits

[astral jay]

What I really wanna try is merging lol

[crisp owl]

Just updated my img2img workflow to include random prompts and style selectors. And updated outdated MMDetDetectorProvider
https://civitai.com/models/123048?modelVersionId=136372

[astral jay]

Running on rundiffusionXL

same seed same params, to make it fair

[zinc lynx]

dreamshaperxl

[snow lotus]

how to use controlnet on sdxl ?

[crisp owl]

still being developed

[snow lotus]

is there a release date ?

[crisp owl]

nope

[snow lotus]

okay.... thx

[autumn forum]

just added face detailer to my workflow, works pretty freaken good

[crisp owl]

Yeah it does well for upscaled images

[soft bone]

no kidding. just figured out full resolution inpainting yay

not upscaled^

[autumn forum]

wow sooo much better curse sdxl for poor faces. it can do closeups sometimes but any length away its so poor

[soft bone]

yep especially if youre using a lora trained on close ups

[astral jay]

RunDiffusion is also showing some anime weights

[soft bone]

show da image

[autumn forum]

now i just gotta condence this to be easy to use, im happy with my current setup but man does it look like a mess

[soft bone]

the straight piping is almost messier tbh

[snow lotus]

I am working on changing a product backgournd. Here is the original product , and the generated one by sdxl through image-to-image. The problem is that the product changed, not the original one.... 😦 Anyidea on how to fix this ?

[zinc lynx]

i like my spaget

[autumn forum]

i switch it when i wanna see where things go lmao. straight for pleasure, curve for productivity lol

[soft bone]

i've seen some decorative workflows made of the spaghetti on civit

[astral jay]

this is mine lol:

[soft bone]

we got a spaghetti hater

[autumn forum]

ooo thats pretty

[astral jay]

no refiner, no face detailer, just stacked upscaling

lol

[midnight shuttle]

Did you forget to mask the produt so it won't be inpainted? Describe your workflow. And maybe it's easier to generate the background by itself and then use Photoshop to add the product.

[astral jay]

works the best

this is rundiffusion

Lighting is cool but otherwise likeness is no good

[soft bone]

aint no way thats the same params

[autumn forum]

yeah looks like plastic. some peoople like it but not my style

[astral jay]

same exact params

[snow lotus]

I tried to mask the product, but I would like to change the color and light on the product to have a better composition with the background. While masking the product can not have thes..

[astral jay]

of course, if you wanted a good comparison you'd need more gens

[midnight shuttle]

It sounds like you will need controlnet for this. I don't know how to use controlnet but others maybe will help.

[soft bone]

guess I'll try dreamshaper. i always stray from anything trained on art and ik dreamshaper is

[snow lotus]

thanks, I tried, but controlnet does not work with sdxl ...

[midnight shuttle]

You need to use SD 1.5 for controlnet.

[astral jay]

In case anyone is wondering how stacking upscaler iterates over a gen:

[snow lotus]

thanks

[soft bone]

how long does that take

[astral jay]

About 2 mins

On a 3080

bit less actually

[soft bone]

okay dreamshaper is lit. I'm converted

[astral jay]

Honestly I just got converted to talmendoxl lol

got me the best likeness

[trim orbit]

did you try psy animated?

some of my loras were bomb on it

[zinc lynx]

more dreamshaperXL testing

[astral jay]

on it

[soft bone]

dreamshaper environments are so dynamic and pretty

[astral jay]

Btw, any idea how they made that? Training over sdxl?

Or merging or what?

[zinc lynx]

no idea tbh but im excited for absoluterealityXL

[soft bone]

finetuning XL

[astral jay]

say more

[zinc lynx]

same everything but with sdxl base model

[astral jay]

share share share

[astral jay]

Tell me everything you can plz lol

[soft bone]

normally finetuning is with everydream2 (idk if this works with XL yet). It involves continuing training in the same way the base model was originally trained (think of it as the most deep and true form of training). Dreambooth and Lora are just accelerated and optimized ways of training, for people with limited resources.

It's the sort of thing you do when you have 50,000 images you want to train and a lot of money. training on this level can improve the actual abilities of the base model (like base 1.5 vs realistic vision). It takes an extremely long time and a lot of power. People making money from finetunes usually have a server of gpu's available to them.

[astral jay]

Yikes

Welp

Lora's for now I guess

[soft bone]

Yep. finetuning was possible locally on 1.5 but at the base XL LR of 4e-7 it would take a 3090 like a year to do a big dataset

[vale eagle]

I believe everydream2 SDXL support still is soon at this moment.

[astral jay]

bruh but wait, what did they use in dreamshaperXL?

[soft bone]

a lot of a100's

is my guess

and dreambooth with a low rate. it's all the same thing almost

[astral jay]

but dreambooth doesn't support sdxl yet, does it?

[soft bone]

on kohya, it does

[astral jay]

Huh

Interesting

https://tenor.com/view/the-wolf-jules-winnfield-pulp-fiction-thats-one-tasty-burger-gif-25336873

[soft bone]

they made a quick alpha with their 0.9 parameters. I expect a big update

[astral jay]

Not great, anime-ish, less detail, and interestingly, slightly more nsfw lol

[opaque solar]

This may have been asked but:

Does SDXL obey clip skip?

[astral jay]

yup

[opaque solar]

It does? Ok, because I just ran the same prompt with a locked seed with it set to 1 and 2 and the images are identicle...

But I may have just gotten lucky.

[hardy cipher]

this guy has some weird stuff going on

[hardy cipher]

1988?

[soft bone]

feels like it. got heavy nostalgia making these

90s in the prompt

[hardy cipher]

that shirt on the left is pre-1990

skateboard kid looks like 1992

[soft bone]

analog is peak realism

[spice skiff]

All base XL with 2x latent upscale and 2x pixel upscale.

[soft bone]

k i'll go to bed now

[dense chasm]

runway generate videos facial deformation

[molten gull]

i got a comfyUI question: i made a primitiveNode with an integer ("width") and one with "height", but i can only connect it to 5 other nodes

is there a way around this ? and if, how ?

[upbeat summit]

I ran into this a couple of times. I usually use a different node (like a reroute) than. primitives can't be connected to everything - at least right now

[heady vale]

add reroute primitives and split the connections

[molten gull]

what node do you use ?

whats a reroute primitive ? how do i add one ?

[heady vale]

search

[molten gull]

you cant connect a primitiveNode with a reroute

[upbeat summit]

you first have to connect the reroute to your destination, than it should work

[molten gull]

doesnt work

i want to do width and height

[upbeat summit]

testing now

[upbeat summit]

more than 5

[primal vault]

Version 3.4 of my ComfyUI node extension and workflow (Searge-SDXL) has been released on CitivAI https://civitai.com/models/111463?modelVersionId=136230 and Github https://github.com/SeargeDP/SeargeSDXL
The main change is, that from v3.4 going forward, workflows used to create images and embedded into them will be forward compatible with future updates of the node extension.

[upbeat summit]

Thanks for the update and sharing it, Searge! Great work. I will definitely check it out

[primal vault]

Thank you, glad to hear that you like it

[upbeat summit]

that model selection node is pretty cool!

[chrome flicker]

I see now many fine-tune anime xl mode in civital , but all of them had a bad performance , even worse than 1.5. I think training sdxl not only just train base , refiner model is also even more important. That's why until now there is not good xl model of anime.

SDXL is good, 50% is because strong clip, 50% is because refiner model. No refiner model, the result of sdxl is not surprising .

It seems that most people still don't realize how important the refiner model is.

[peak dove]

I'm in Edinburgh for the Fringe Festival

[west breach]

exciting! 🙂

[west breach]

do you use refiner with your lora?

[somber hill]

no. when refiner used you lose your trained subject

at least this is the case for me so far

[west breach]

yeah refiner seems to have a strong bias and changes faces even with low denoise value

[upbeat summit]

with some prompts and settings the refiner works a bit like a face replacement

[trim totem]

Interesting, so you recommend switching off the refiner when using loras?

[somber hill]

it depends on purpose. for a person yes

but lets say you trained a car then why not cars are generic

[trim totem]

Thanks, thats good to know. I have trained a lora for a person eating chocolate. Using Searges ComfyUI I get without lora pretty wild results. But so far using lora was not successful, but I will try using it without refiner.

[somber hill]

i am editing my video

has shown checkpoint comparison too

using automatic1111 this time

much easier

[trim totem]

Ah ok, I will watch it. Does automatic1111 have a working colab script for sdxl yet?

[somber hill]

well i never used auto1111 on colab

i use on runpod or pc

[trim totem]

What runpod package do you have?

[trim totem]

Runpod looks good I'm trying comfyUI there with rtx 4080

[tender birch]

is it possible to disable inputs on custom Nodes?

cause i dont know what to plug in

[cerulean flicker]

#1072220168534642768 a standing globe with a clock inside

[heady vale]

go to #1100170312106127410

[somber hill]

i have auto installers

using templates provided by runpod

[upbeat summit]

Not that I'm aware of. There are nodes where you can leave certain pins disconnected. But if you run a queue and it stops because the node doesn't have all the necessary links, there's not much you can do except fixing the requirements.

[elfin cobalt]

If you drag a line from the input to empty space, it should give you a few options for things that would generate those types.

[shy kelp]

is the nvidia a100 better than 4090

[elfin cobalt]

Yes and no. Depends on what you want it for.

[shy kelp]

for stable diffusion

[elfin cobalt]

The 4090:

• Has a lot more general-purpose texture/graphics units, so can be used for games.
• Has 24GB of memory instead of 40/80.
• Uses (quite a lot) more power.
• Is a little slower for Stable Diffusion in particular.
• Is a 1/3 the price.

[shy kelp]

I want to buy a graphics card specifically for image generation

[elfin cobalt]

The A100:

• Is a lot louder. (Like, jet engine loud.)
• Will fit better in server cases, so you can fit more in a single machine.
• Supports NVLink, so you can hook them up to each other.
• Basically doesn't supports graphics output at all.

For stable diffusion inference, the 4090 and A100 are almost the same speed though.

[pallid sail]

Oh darn, I forgot to turn off the watermark

[elfin cobalt]

The A100 does cost 3x as much. Basically you only want it if:

• You absolutely desperately need the memory.
• Or you need to fit 3 to 8 cards in a single machine.
• Or it'll be running constantly, so the power costs outweigh the price tag.

Given SDXL runs fine with 8GB of VRAM, point 1 can't be relevant.

[shy kelp]

Alright, I think i'm just gonna buy the 4090 then

Thanks

[elfin cobalt]

Mm. Pretty much: If you need to ask, then you don't want the A100. 😛

Also don't forget that the A100 is a really really truly terrible graphics card.

It's not meant for graphics.

Forget gaming, it has trouble running Chrome.

It is meant for workstations, so it can, but a Quadro would be better in that scenario.

(Personally, I don't want an A100 in my workstation. They're noisy as fuck.)

[shy kelp]

Well, I just got a "broken"(It has a disconnected power wire) A100 from our servers department, And I was just wandering if I should buy it from them for like 1200

That's why I was asking

I'm just a network engineer, trying to fumble with sdxl

[elfin cobalt]

Cheap!

I’d buy that one. If you think you can make use of it, great, otherwise feel free to sell it to me. :p

It’s still going to be fairly noisy though. You don’t want it in your living room.

[thorny frost]

What abot the a40?

48GB

[bleak garden]

have you tried DreamShaper XL1.0 without refiner?

[thorny frost]

4000€ "only" in amazon

[cinder aspen]

what the hell is "offset lora"?

[thorny frost]

to get more contrast in your images

[shy kelp]

yeah i'm definitely buying it and see what i can do with it, I'll put you first on the list of potential buyers

[elfin cobalt]

👍

[oblique bobcat]

Hello, Is there a way to calculate how many credits will my request spend?

[elfin cobalt]

The API docs for whatever service. You didn't mention which you're using.

Most of us here are running it locally.

[rustic garnet]

I rarely use refiner. In my opinion the base model is good enough in 90% of the cases

[smoky patrol]

https://platform.stability.ai/pricing If you're talking about Dreamstudio, the pricing is here

[cinder aspen]

are there suitable workflows for RTX 2070 kind of VRAM?

it literally takes minutes to generate a grid of 4 images

and my laptop has to slow down to do it lol

[elfin cobalt]

https://brage.info/GAN/ganbot2/7e676e5e-83f8-4c1e-b822-c2ffc78faf57.2.webp

Can't do this with 1.5

I've got to agree that the default quality of the XL models isn't up there with the 1.5 ones yet, but it's so much better at composition.

And it's easier to fix aesthetics than composition. Aesthetics just requires prompt-crafting.

@cinder aspen A laptop 2070 is about equal to a desktop 2060. You definitely shouldn't expect miracles.

Improving the speed mostly means compromising on quality. Use a scheduler that works well with as few steps as possible.

[thorny frost]

has anybody thought about finetuning the refiner? does that makes sense?

[elfin cobalt]

Absolutely!

[thorny frost]

what?

[elfin cobalt]

It just doubles the cost of making a checkpoint/lora.

[dense chasm]

wandb is much more simpler and clear,a good alternative to tensorboard

[cinder aspen]

i know that 😄

im using it for fun anyway

for more serious stuff or priority something else

[thorny frost]

why it would be like finetunning any checkpoint wouldnt it?

[elfin cobalt]

Not having done it myself, I can't say for sure. In principle yes, but you might need a different data source from just the usual images.

Since it's supposed to use the output from the base sampler, you might need to first train the base and then use that to train the refiner. Or not. I dunno!

[thorny frost]

the refiner is a bit of a mistery

[elfin cobalt]

I'm sure it's documented somewhere.

[vale eagle]

Could not train lora based on refiner now with kohya ss.

[soft zealot]

well my workflow like my gut is spreading lol

It has everything that was in it before Plus I've added in some logic switched to allow the use fo SD1.5 or 2.1 models in the same workflow (they run through the same 3 step process that SDXL does just using standard models)

One example of each using the same base prompt & seed

[strong field]

Ive tried alot of workflows, Winston, yours is one of the best

[soft zealot]

why TY kindly. I based it on @high skiff 's OG and then kept adding to it 🙂

[lusty token]

Maybe way off, but is there any possible workflow in comfyUI from SDXL to video?

[high skiff]

:>

[strong field]

Yes i recognize some of the "community modules" in there

thanks Sytan!

[high skiff]

bno problem, its really a partner effort with winston now lol

I have more on the horizon :p

[soft zealot]

not sure if anyone has done a workflow before though that allows you to switch between SDXL & SD1.5/2.1 in one place though 🙂

[strong field]

I believe "sdxl10ComfyuiULTIMATE_v24" does that

[soft zealot]

ahhh, nothing new in this world lol

[high skiff]

I do have some good/bad news about a recent finding. Some of which will be shared with the community in 1.1, and some I will be keeping to myself just because I think I would like to still have a little bit of personal information i can use to stand out

[strong field]

haha, everything is a remix

[strong field]

no one is complaining about your generosity!

[high skiff]

You'd be surprised haha. Some people think I am horrible for keeping like 5% of my findings to myself lol

but, just to sort of build a little hype as to what SDXL can do, here are some incredible new realism gens I have been able to create with some pipeline changes, and some new prompting findings

The pipeline changes will likely be shared in v1.1 of my workflow, and the prompting will likely be kept to myself, at least for the time being

[strong field]

drumroll....

[high skiff]

some people think SDXL can't do exceptional realism/detailed skin, and I am here to say that is just bad prompting

[strong field]

actually, im going to say, its been done, but not cleanly as far as I have seen, not yet at leats

[soft zealot]

But I have to say I do like the output from 1.5 being run through a Preconditioner and then the Base & Refiner using the same logic in splitting up the step as for SDXL .

This one was done with Revanimated as the Base & Photon as the refiner

[strong field]

totally agree, been experimenting with realism and details, the fact that we can do it without loras and networks/inversions is insane

[strong field]

hahah, i love how the community is a hive mind right now, I have been trying the same thing

[vale eagle]

Hey @high skiff . Do you have any findings about hands? I think SDXL has good hands inside the base But I don't know how to combine the good one to the image I prompted.

[high skiff]

SDXL can kill realism with proper prompting

[strong field]

Change the old SD1.5 checkpoints and vary them as base/refiner

[cinder aspen]

i might just have found the most complicated custom modes workflow out there lol

[strong field]

those are great!

[high skiff]

Hands are not something I have focused on at the moment, but some of my realism tests have been yielding better hands naturally

[high skiff]

thanks!

[soft zealot]

same prompt,seeds, styles etc but only using revanimated in all steps

[soft zealot]

adds to prompt " hands in pockets"

[vale eagle]

Like this. It is just base model with prompting. But the prompt has strong biase which is very diffcult to apply on other scene

[strong field]

photon still one of my favs right now

your new workflow is wild btw

ty for the share

[hexed hatch]

Any comfui expert here that could help me? Ive tried updating impact pack but cannot get it fixed.

These 2 inputs are giving me errors

and they should be like this to fix it, but updating impact pack doesnt help

[strong field]

are you using comfy manager?

[hexed hatch]

yeah

maybe because im on portable?

[strong field]

ah, I am not sure, I dont use portable and I dont know if there are differences in the way the packages are handled.

Did you try to manual install that bad boy?

Remove the folder and pull from git?

[cinder aspen]

the comfy manager definitelly breaks some custom node workflows

[rustic relic]

is that comfyui?

[cinder aspen]

and doesnt even recognize the missing nodes

it just...breaks

[strong field]

yes, been learning this the hard way

[cinder aspen]

at this point i will just do it myself and maybe look up on some tips about the values of several settings

[hexed hatch]

Found the folder, how would I pull it from git? I got the link from github for impact pack

[rustic relic]

is there something else as comfyui?

[cinder aspen]

didnt try the stableswarm UI

[strong field]

you on windows?

[hexed hatch]

yep

[strong field]

git clone <link>

sorry wait

open CMD in folder and then use that command

[hexed hatch]

thanks, was typing git pull instead

ill see if itll work now

[high skiff]

man, SDXL really gives MJ a run for its money in terms of portrait realism

[strong field]

yeah, one of the issues i found was that old workflows can break if comfy manager updates the extenions

[strong field]

Honestly, if your prompt game is right, this is MJ level good

[hexed hatch]

sadly enough they're still inputs

[strong field]

of course its not "easy"

[high skiff]

I would have to say its considerably better than MJ actually

[cinder aspen]

MJ and SDXL are not the same anyway

[high skiff]

the fidelity and realism you can get is just on another level honestly

[strong field]

touche

[strong field]

which module is that

[near pecan]

hey, has anyone got any tips for controlling the perspective when it comes to human/humanoid characters ?
I'm mainly struggling to get side view of the characters (standing sideways, facing <direction>, camera from side, etc.)

OpenPose controllnet would likely help with that....

Front and Back of the character works most of the time though

[cinder aspen]

barely anyone will switch from MJ to SDXL even if the latter one generates better image if it is true or not

[hexed hatch]

and the guide im following its supposed to be like this, with latest version of impact pack

[strong field]

oh do you just need to remove them?

[hexed hatch]

yeah

[strong field]

right click on module

covert "item"

convert to widget

[hexed hatch]

hmm no item in my list

only seed to widget

[strong field]

hahah, meant your item name

yours doesnt show up like this?

[hexed hatch]

no haha

[strong field]

hmmmm

[hexed hatch]

anywhere i right click its that same menu

[somber hill]

looks like SDXL doesnt know pegasus and horse difference 😄

[strong field]

add a new node,

and see if the new node will allow you to convert input/widget

[hexed hatch]

nop even new new nodes, its letting me convert seed and other stuff

but noise mask and force inpaint

no option for those 2

all good man ill figure it out, you helped me enough

thanks so much

[hexed hatch]

https://www.reddit.com/r/comfyui/comments/15iq7id/face_detailer_force_inpaint/ seems this guy had same problem and reinstalled pip

[strong field]

yeah looks like its related to updating

you can try pulling the older version from Civit AI

You can just download the folder, delete the updated version in extensions, and see if that fixes your dilemma

updating from new versions sometimes breaks old workflow

[high skiff]

I really need to try and find a new pixel upscale rthat maintains texture detail

cause the skin texture I get from base SDXL right now is incredible, I just need a pixel upscaler that can preserve the textures

like yeah, the upscale has much more detail, but it loses all of the texture and blemishes

its a problem with ultra sharp 4x in general

and while NMKD maintains the blemishes way more, it adds a weird static/canvas texture everywhere that makes things look super aliased

[autumn forum]

Use dpmpp-2m karras. It’ll give you texture 😂

[hexed hatch]

I guide Im checking used this, not sure if it would help u

[high skiff]

I am using super scale, and while it does maintain texture better, it also murders everything with a layer of static unfortunately

as you can see here

makes things all crusty

[crisp owl]

I use the siax, it has done the best for me for retaining textures as well, but yeah overall it's still an issue

[high skiff]

I will try Siax

[crisp owl]

ultrasharp I hated, was like pounding on foundation on their skin

[oblique bobcat]

Pricing is there but I couldn't find how can i calculate the price with different parameters. Do you have any experience with this?

[strong field]

hahah i use siax too

[dense chasm]

let's check over which steps will overfit

[high skiff]

looks like Siax has even worse static than super scale

[heady vale]

close to real, but the eyes ruin it

[strong field]

lol

[cinder aspen]

not just the eyes

[high skiff]

looks like maybe bad prompting

[cinder aspen]

its not just the eyes and none prompt can fix it to the point where its like a literal real image

more obvious than the eyes are actually the arms

[high skiff]

I have gotten some damn good eyes out of SDXL, so I am not sure about that

[cinder aspen]

yeah but i mean completelly accurate to a real photo

[heady vale]

close up eyes can be great, but step back a little and it all goes bad

[strong field]

ive gotten good eyes too

[cinder aspen]

im not sure i got what he really wants tho

[strong field]

takes a ton of work to get something convincing

[high skiff]

I mean, I think these eyes look pretty damn good

[strong field]

but thats the game right now

IRIS

[high skiff]

very very slight mishapen left eye, but yeah

even these look pretty damn solid

and they aren't particularly close

[cinder aspen]

yes

there are anomalies but its looking good

the point is not tho for it to be real image level i guess?

[heady vale]

the coffee is saving the hand there

[cinder aspen]

thats what comic artists have in their repertoire actually

hiding parts that may be hard for them to draw or paint

you are bad at feed? Hide em

feet*

[high skiff]

I am certain we are just one simple LoRA away from exellent hands

just from what I have seen in general training

SDXL is a beast with LoRA's

[strong field]

yes, i am waiting for people way smarter than me to train some awesome LoRAs

[cinder aspen]

tentacles are the hands on steroids for gen AI

[high skiff]

luckily most normal people don't generate tentacles on the regular lol

[strong field]

lol

[cinder aspen]

xD

not THAT tentacles

god damn it Sytan

xD

[soft zealot]

yarp

[high skiff]

I am collecting a dataset for a portrait LoRA

[cinder aspen]

i do tentacles fairly often

[high skiff]

a small one, but from what i have seen of SDXL, it doesn't need much to be stellar

[cinder aspen]

there is no case where i dont have to jump in myself

[high skiff]

hell, my Na'vi LoRA should be good enough testament to how well SDXL learns with a little info

[cinder aspen]

i somehow broke your (Sytan) workflow

that wasnt due to your nodes or so

it was an external issue

[high skiff]

nice lol

I can always count on the community to nuke things haha

[cinder aspen]

one day i will accidentaly open the gate to hell

trust me

[clever verge]

I'm not questioning you but if hands is an easy fix why isn't it in the base model already? 🙂

I hope you are right though.

[high skiff]

same reason it wasn't fixed in base 1.5. Its a lot of work in a very dedicated way, but its easily plausible

and yet I have 1.5 models that are very good at hands, and can reliably deliver specified finger numbers

[strong field]

SDXL is the chassis, not the fine leather seat

[high skiff]

I need to find a good repository of portrait photographs

google has nerfed their search engine so much, its blatantly obvious

[strong field]

somebody call jcpenny

and yes, its very obvious big data is clamping down on their cash cows

[high skiff]

I have a feeling they are purposely limiting image results to deter training, cause you can't look up hardly anything and get good images of them anymore

almost no high res image results anymore

[cinder aspen]

hmm ComfyUI still doesnt want to find the missing nodes for a custom workflow

for whatever reason

[high skiff]

all the "high res" images are just 400px images pixel filtered to 1200px

[strong field]

lol

poison.llm

[high skiff]

Its really quite annoyting. Its the reason I can't train a portrait LoRA, cause I can't find any good images

[heady vale]

https://www.youtube.com/watch?v=3qSQjRaseos live now

[high skiff]

I can't trick myself into giving a shit about NVIDIA anything anymore

[cinder aspen]

the savior is speaking

[strong field]

you got your amd card to work?

my 6950 xt couldnt handle the pytorch lol

[high skiff]

what?

I don't have an AMD card

[strong field]

haha, that was my facetious point!

[high skiff]

OMG, thank you so much google for this "4k" image

[strong field]

what in the wow

[high skiff]

unfortunately, AMD isn't reliable enough for me to switch

[strong field]

same i "upgraded" to a 3090 FE

[high skiff]

but I can say one thing, NVIDIA doesn't get any of my money directly anymore. I will not buy a new GPU from NVIDIA

[strong field]

all my gpus are used

[high skiff]

I just got a 3090 second hand

just a few days ago

[strong field]

haha i literally bought mine last week

[high skiff]

trying to do training, but google has neutered its image search engine so much, its insane

[high skiff]

Lets compared deals :p

how much?

[cinder aspen]

does anyone have the issue for the comfy manager to be not able to find the missing custom nodes?

[clever verge]

Getty Images and others probably have a cash cow now.

[strong field]

miss the days of public facebook, public photobucket, and flickr

[strong field]

what node?

[high skiff]

I have been considering using way back machine

[strong field]

i got mine for 700

[high skiff]

@strong fieldHow much did you get your 3090 FE for?

ah

[strong field]

I STOLE THAT CARD

[high skiff]

I got mine for the same price, but I got really lucky and got basically the best 3090 on the market

[strong field]

that FTW3?

[high skiff]

only issue with it is its filthy lol

[high skiff]

FTW3 ultra

[strong field]

ultra ftw 3 thingy lol

[high skiff]

yeah

its a beast, I love it

[clever verge]

I'm too late it seems. https://blogs.nvidia.com/blog/2023/03/21/generative-ai-getty-images/

[strong field]

yikes, mine was barely used

but your card was my second choice!

[high skiff]

my previous 3090 was a Zotach trinity OC

[strong field]

too bad it was more expensive

[high skiff]

My card was responsibly used, the guy just had a very weird living condition is all

[clever verge]

I had to add 32G of RAM, 32 was just too little.

[high skiff]

luckily, its EVGA, so I can tear down the whole card and clean it out, and still have my warranty

EVGA warranty is unrivaled

I will be doing that soon

[strong field]

great point

[cinder aspen]

you didnt know that? 😮

[high skiff]

the seller took great care of the card, it was just a byproduct of where he lives

[cinder aspen]

also Nvidia is coming with their Picasso

[strong field]

sounds like you got a great deal too

only very slightly jealous

[cinder aspen]

Searge workflow custom nodes

[clever verge]

Missed it yes. Late to the party, but atleast a made the prediction after the fact. 😄

[high skiff]

He had the card for 2 years in a laptop enclosure in an extremely small apartment in NYC, so the oils from him cooking gummed up the fans and stuff

But other than that, he said the GPU was always way more powerful than his laptop could leverage, so he always ran it at stock or power limited to save electricity, so the card has never really been run hard. Said he is selling it cause its very clunky and he bought a new laptop with a more reasonable GPU built in

[strong field]

you gotta clone the git, comfy manager cant do it for whatever reason

[high skiff]

He has a lot of reputation on hardware swap, so I trusted him, and yeah

it runs great, just a little hot (understandable with how dirty it is)

but the damn card uses 440 watts stock, so its reasonable for it to be a wee bit warm lol

[strong field]

haha they definitely run warm, i had to add external fans to move cool air into the office

[high skiff]

but damn, its a very efficient card. During the day I run a severe power limit on it, and lose not much perf

[clever verge]

Yes they are energy hungry, good thing we have negative electric prices today.

[high skiff]

50% power draw for about 80% stock performance

[cinder aspen]

they re everywhere

[high skiff]

so my 3090 at 210 watts runs almost as fast as a stock 3090, which is insane

AKA, it runs about 7% faster than my previous extremely maxed out 3080 at 380 watts

[strong field]

i havent tried that yet, what do you use to profile?

[high skiff]

just afterburner

I messed with undervolting as well, but it was a little tempermental, so all I have now is a 50% power limit

[ionic dragon]

yeah you got it

[strong field]

i could never get undervolting consistent

[high skiff]

this card is solid at undervolting

not as good as my 3080, that card was insane

stock clocks consistently at 40% lower than stock voltage. Insanity

1850mhz 3080 running at just 820mv

[ionic dragon]

xx90 series are great for OC and UV

[high skiff]

my 3080 is absolutely insane

its the most overkill and overbuilt GPU I have ever seen lol

this 3080 is bigger than some of the biggest 3090s, and bigger than most 4090's lol

it can also draw up to 450 watts, which is 130 watts over stock with its OC bios

its a massive card

especially compared to my previous 3090 lmao

its still bigger than my EVGA 3090 FTW3 Ultra in every dimention lmao

its longer, a little taller, and a tiny bit thicker, tho the 3090 weights considerably more

[strong field]

weighs more than a 3090, what are they feeding that thing TSMC?

[clever verge]

I was shocked about the size of the FE 4090.

It's 3 slots high, feels like a tank.

[high skiff]

I think the FTW3 ultra is kinda ugly 😅

But the RGB is gorgeous. I have never seen RGB as smooth and diffused as on this card

[hardy cipher]

https://civitai.com/models/65639/penis-art-sdxl

[high skiff]

lmfaooo

[ionic dragon]

looks only matter until you buy it, but after you buy a gpu, you rarely look at your gpu

[high skiff]

oh for sure

I don't run RGB software cause of the huge resource hit, so its always on

and this GPU is way prettier than my MSI. The MSI has trash RGB

[vale eagle]

Could we use certain layers of lora in comfyui?

[hardy cipher]

true gpu chads get a clear case so they can see the gpu at all times

[ionic dragon]

on how many penises did he train the model, lmao

[unreal plover]

https://youtu.be/3qSQjRaseos

he talk about SDXL

[high skiff]

the RGB on the FTW3 ultra is so smooth it looks like CGI lol

[hardy cipher]

I use tons of loras, bud

[high skiff]

I genuinely don't think I could give a shit less about what that crook has to say lmao

[unreal plover]

@high skiff Without him, no GPU to diffuse my friend

[heady vale]

Jensen runs SDXL lol

[high skiff]

I would rather him be gone lol

also, hes just a CEO, they could do just fine without him

hes not as important to NVIDIA as Lisa Su is to AMD

[unreal plover]

He have the good prompt for his team

[strong field]

the way i see it paid for 1 ten millionth of that stage outfit

[high skiff]

lmao

[unreal plover]

Jensen is a very good prompter

[hardy cipher]

I don't know, but decent results I think

[ionic dragon]

lol

[high skiff]

good for him I guess lol

Still don't give a shit about what he has to say

[hardy cipher]

I don't care about him so much that I don't know who he is

[unreal plover]

I don't know what he did to you but i am sorry for that

[high skiff]

CEO of NVIDIA

[strong field]

well thats something

[high skiff]

hes a shitty guy that takes advantage of markets, releases cash grab/e-waste products so shameless he makes Tim Cook blush, and he is massively overpaid for doing close to nothing

[soft zealot]

This is the only Jensen worth mentioning

[strong field]

now that is penix-art-sdxl

[hardy cipher]

that's awesome. I tried to comment on pns pikachu but my comment got flagged and blocked

[high skiff]

the license plate even looks like it could almost be trying to spell "Bullshit"

Even SDXL knows lol

[ionic dragon]

lol

[unreal plover]

You use a tool who was cooked on his kitchen

[high skiff]

ok, so I have to like him? Hes a shitty person, and he himself didn't make it, a company he is lucky enough to be the face of did

[soft zealot]

lol (you do know thats an actual image and an actual UK number plate

That car was auctioned last year
https://www.handh.co.uk/auction/lot/195-1971-jensen-interceptor-ii/?lot=56870

[high skiff]

yeah haha

[unreal plover]

You don't know him, he is a business leader, not a nurse.

[ionic dragon]

@high skiff how to add lora to your workflow?

[candid walrus]

you know him? wow

[unreal plover]

Only wolves can do wolf things

[high skiff]

Hes a POS, it doesn't take much to know that lol

[high skiff]

https://tenor.com/view/spongebob-oh-brother-this-guy-stinks-this-guy-stinks-spongebob-squarepants-spongebob-meme-gif-19365861

[soft zealot]

add a lora loader in and take the Model and CLip fromr he models selector to it and then from the LORA loader to wherever they connected originally

[unreal plover]

No i don't, it is why i don't talk about him and what he is

[candid walrus]

but you are in the inverse

[high skiff]

yet... you defend him... The leader of a trillion dollar corp

[candid walrus]

i'm not saying one way or another but people can have their opinions, just like you have yours

[tender birch]

also @elfin cobalt : I just had to update comfy... The node was "newer" than comfy itself which leads to this error ^^

[unreal plover]

The leader of a compagny who make stuff i use every single second of my wake life

[strong field]

i did a thing

[candid walrus]

yeah, but same would be true if they replaced him today

[ionic dragon]

ok thank you will try

[hardy cipher]

I'm running the sound offsent lora in the negative because I'm a mad lad

[high skiff]

ok? So cause you use a product that is forced on the industry as basically the only option (borderline effective monopoly), you just blindly trust that hes a good guy? lmao

[hardy cipher]

they're like the gpu mafia, lol

[strong field]

similar vibes

[unreal plover]

But not if he wasn't here 10 years ago

[high skiff]

I will never support NVIDIA as long as I can remember their massive mining card E-waste product that destroyed so much of the global silicon production over Covid doe literally a few bucks

[hardy cipher]

he has fanbois?

[candid walrus]

that's a big 'what if'

[high skiff]

I guess apparently

[unreal plover]

Oh no, he is too good to be over runned, what a devil !!!

[high skiff]

What?

I'm starting to lean towards you just trolling now-

[strong field]

true that

[unreal plover]

It's not their fault that no one is able to reach the capabilities of their products

[hardy cipher]

so I'm thinking negative sound offset might work right? it's somewhat arbitrary. or maybe it'll make garbage

aren't they a trillion dollar company?

[high skiff]

Yeah but like, that has nothing to do with Jensen, thats the accomplishment of the actually smart people at his company

[hardy cipher]

I mean, kind of hard to be a startup and compete with that

[candid walrus]

it would be if their lawyers killed them before they could make a difference

[unreal plover]

Who prompt them ? Who guide them ? Who chose to invest on AI ?

[candid walrus]

there's a whole board

[dense chasm]

diffusers has a solution run controlnet on TPU

[high skiff]

the leaders of projects, the finanical advisors,a nd the board of directors

[candid walrus]

there's experts that are not jensen

[unreal plover]

He have the last word

[hardy cipher]

well, that's what CEOs do

[strong field]

lol corporate structure doesnt work like that at all

[candid walrus]

which might just be, "yes, expert, that sounds best"

[hardy cipher]

they have the last word

[unreal plover]

AMD also have good leaders, so why they can't ?

[hardy cipher]

but they can still be overridden

[high skiff]

Please Tell me that you don't actually think that the CEO controls literally everything...

All a CEO does is a point people who are qualified to do the work that they're not qualified for lol

[hardy cipher]

they're like the president or something. they are the face of all the decisions, but they don't necessarily do the work

[high skiff]

I mean, obviously there's more to it than that, but it's not like Jensen is single-handedly creating every project and personally watching each and every project to fruition

[midnight shuttle]

Very good engineers are hard to find and cost a lot of money. You can't just hire more code monkeys and equal a good engineer.

[strong field]

why do we need to explain this right now, i should be working on more tentacle gens

[high skiff]

More than likely, as the CEO of his company, especially a company so massive, he probably has very little knowledge on what's actually internally happening across the thousands of employees that they have, as well as several different sectors across the world

The entire company is a huge trial and error, where the people who can line Jensen's pockets the most get promoted, and the people who get out competed stay stagnant

[hardy cipher]

yes, "infinity hentai tentacle" prompts

[high skiff]

I suppose the same could be said about almost all companies, but the idea that Jensen is responsible for the success of Nvidia is kind of insane

[hardy cipher]

well their job isn't supposed to be making a bunch of small decisions, they're supposed to be the decider for a few big ones. doesn't mean it always works like that

[strong field]

hes not, and yeah that would apply to most CEOs

[high skiff]

Now, somebody who is responsible for the success of their own company is Lisa Su, the CEO of AMD

And that is because not only is she a CEO, but she's an incredibly smart silicon architecture engineer, and she does actually oversee the architectural development of their new generations of chips

She's not just a CEO, she's also had architectural engineer

[unreal plover]

Sure, a waifu diffuser is way more able than the leader of a compagny who made x100 in stock market

[hardy cipher]

so turns out negative sound offset actually looks like crap

[strong field]

start talking about daddy elon next pls

[high skiff]

Before Lisa Sue was hired as the CEO of AMD, their company was worth less than 1% of what it's worth now, and just in a matter of a few years of her being CEO, they overtook Intel in terms of desktop consumer market share, which is genuinely amazing in my opinion

[hardy cipher]

who would have known?

[unreal plover]

An other genious

[strong field]

lol

i like the way he spells the best

[unreal plover]

😂

[ionic dragon]

tech CEOs are way better than non-tech CEOs
they understand what goings wrong and whats not

[high skiff]

Oh God, I can't with you lol

[hardy cipher]

this kid is a progidy

[strong field]

Cathie Wood gens pls

[high skiff]

That's true in some companies, not all of them

The reason that Intel was failing so hard for several years was because their CEO was not an engineer, he was a business major.

There was a massive disconnect between their higher management and their capable staff. The CEO would just set goals for money, not for progress, which led to a huge amount of stagnation

It's quite literally the reason he was fired and replaced by somebody who is an actual engineer, and now Intel is back to dealing a pretty severe blows to AMD on the forefront of desktop CPUs now

[unreal plover]

No one buy AMD GPU, AMD have always been the second, they are so far away

[hardy cipher]

so turns out after about -1 sounds offset it's just noise

fyi

[high skiff]

Nobody tell this guy about the fact that AMD actually wins on average in terms of price to performance ratios for raw raster performance

[hardy cipher]

if you're into that sort of thing

[high skiff]

And I hope this doesn't sound like me blowing smoke up AMD's ass, because I think there's still a really messy company that has a lot of problems, but they are certainly a lot less shitty about everything compared to Nvidia lol

[hardy cipher]

nudes?

[unreal plover]

raw rasterisation is the old world, it is why they are far away, jensen had a vision that AI will be the futur, he invested a lot in research, and won the bet

[hardy cipher]

I'm curious why other companies haven't gotten into the game. there aren't a bunch with the resources to compete, but there are definitely a few

those teeth. such beauty

[strong field]

haha, the CEO of ARK Invest, a spectacular rise and fall story