🆕｜sd3

[errant dust]

So I ran those prompts from the test in Tom's Guide through SD3 8GB and... the results were not fantastic. The prompts are also insanely long. How 'not fantastic'? Here is the retired astronaut in front of a class result for SD3 Large:

[bitter hearth]

the 2B not the 8B

[torn wharf]

The other night my subconscious just came outta no where and told me "hey you know how Jason Reitman directed the new Ghostbusters movies. Go look up Ivan Reitman director of the originals." Literally that just popped into my mind and I checked and they're father son

Brains are coo

[bitter hearth]

the 2B has had its aesthetic fine tuning but the 8B has not yet

[sterile pendant]

yeah you'll need clips. the t5 and clipL, load them with the dual clip loader and set it to flux

[torn wharf]

I wasn't even thinking of ghostbusters

[tough viper]

where do I download them?

[sterile pendant]

https://huggingface.co/stabilityai/stable-diffusion-3-medium/tree/main/text_encoders

you can use the ones from there

they are the same

[tough viper]

thank you

[sterile pendant]

i'd go with the fp8 version of the t5 encoder though

[mortal mesa]

the one thing i dont do

[sage burrow]

https://glif.app/@LadyLalita/glifs/clzvhxume000h52mw6qy5hkp2 😄

[errant dust]

Ok, well, here is SD3 2B on the astronaut:

[bitter hearth]

yeah

that's a good example

look at the fine details and colours on his face

[errant dust]

A bit of a catastrophe

[bitter hearth]

that's what I was saying
if you don't think that SD3 is better for photography
then its not gonna convince you

[tough viper]

where does the t5 go?

[errant dust]

Yes, well, 8 fingers, mutant hands, an audience fresh from a Cthulhu horror movie. We have our own ideas on photorealism for sure

[bitter hearth]

anatomy issues are to do with DiT not VAE

the way to prove that is to VAE encode and decode photographs

and what you will find is that the SD3 anatomy problems are not in the VAE

[torn wharf]

i think for the purposes of discussion, you'd select the best photos sd3 could produce and compare those.

[errant dust]

I honestly could not care then, because the image is completely unusable, and you can wax poetic all day and night on its VAE and it won't improve that image

[torn wharf]

whiel i'ts not consistent right now, it highlights differences in architecture

[rain current]

https://civitai.com/models/636355?modelVersionId=725298
What do you think of this Lora? I have mixed feelings. These are the tests with different weights: https://civitai.com/posts/5464381

[torn wharf]

at the end of the day, it doesn't matter who "wins and loses" the internet argument. The real world differences are real and will be apparant to anyone that can see them

[errant dust]

If you tell me that image I showed is BETTER than the others, well.... ok then

[torn wharf]

facts don't care about feelings

[bitter hearth]

a VAE can't improve an image
VAE encode and decode are both destructive processes
they make the image slightly worse
this can also be proven by doing a VAE encode and decode on an actual photo

[torn wharf]

i'm not bothering with loras until authors figure it out. whats the comfy converted weights? why doesn't 0.2 have them? i go get xlabs loras and theres two versions of each. their text guide lists different ones. i try to load loras in comfy. no work. it's all a catastrophe. lora training authors rushed this stuff out to get street cred and now the files that are out there are a compatibility disaster zone

[noble coyote]

Stare at a waterfall for 30 seconds - then stare at a dark, blank space - and the space seems to fall as well!

[torn wharf]

lossy is a good word in pixel contexts too

[errant dust]

Look, this talk was all about how SD3 produces better photorealistic images than MJ and Flux, and you persisted in this with some claim the reason was the VAE. What I see is a complete disaster of an image, but that's just me. So talking all day about how better its VAE is, for me, is .... absurd.

[dry wave]

the vae has nothing to do with the structure

[torn wharf]

thats a great one. i also love where you stare at a monitor screen of all weird hues, then swap to a black and white image and it'll appear in full color until you dart your eyes

[sage burrow]

SD3 8b ❤️

[dry wave]

sd3 produces extremely photorealism - but the model is fucked up and yes, the images it produces are a disaster

Flux is able to do similar degree of photorealism. It's just hard to prompt for it

[errant dust]

Tell that to him. He said SD3's photorealistic images were better, and then said the reason was the VAE.

[rain current]

I only asked for an opinion on the tests I’ve done... ??? But anyway, thank

[mortal mesa]

sooo not photo real then

[noble coyote]

As an art medium, I believe that SD3 medium is a hugely successful tool. It matters not what quintile of what quotient, which mote of SD3 makes it successful as an artist's medium

[dry wave]

the model is unfinished and basically unusable in its current form

[torn wharf]

that's my opinion. i think the effect of each is subtly working. but overalll...

[noble coyote]

But!

[mortal mesa]

it makes some pretty details

[dry wave]

ice didn't want to say that SD3 produces better images. He said that it can produce an extremely high degree of photorealism on the finedetails

[torn wharf]

you may want to unblock me so you can hear me better

[errant dust]

The discussion was with Neon

[sage burrow]

my daily flux email spam 🙂

[noble coyote]

When it comes to photorealism - there is an element of visual psychology which needs to be considered - and in SD3, what do we need to tweak in order to conform to human visual psychology?

[errant dust]

produce feet and legs?

[noble coyote]

Rick Salmon (famed photographer) heavily relies on psychovisual dynamics in his work

The realism in a photograph is underpinned by (correct fingers/limbs/toes) saturation; contrast; and sharpness

[mortal mesa]

if i take a complete trash photo on a $3 toy camera is it still photorealistic, yes

[noble coyote]

How can SD3 be prevailed upon to better exemplify human psychovisual traits?

[rain current]

It’s not better or more realistic. What you have is greater sharpness; it’s been trained to have a shallower depth of field, which makes the overall feeling more natural and realistic. But if we remove that, what’s left is this other thing (hopefully they’ll release a version that competes with Flux)

[errant dust]

Psychologically, SD3 2B should stop producing this, so I won't feel... traumatized:

[noble coyote]

Me-ness good!!!

😄

[torn wharf]

i know that name. my design teacher was all about him and his one liner pro tips for composition. we were learning scene composition at the time and he was all about how we can take wisdom from other professionals for design rules. i dug of a few of these one liners to show how great they are.

• The name of the game is to fill the frame.
• Dead center is deadly.
• When you think you are close, get closer.
• The camera looks both ways. Adapted from Freeman Patterson's "The Camera Points Both Ways."
• Expose for the highlights.
• Use your camera like a spaceship. (A Dick Zakia philosophy.)
• Light illuminates, shadows define.
• Backlight = shoot tight.
• Make pictures, just don't take pictures.
• See eye to eye - shoot eye to eye.
• Take the darn flash off the camera.

[errant dust]

But look at it this way: it has a 16ch VAE and that skin texture is perfection. Definitely the best image possible

[noble coyote]

Textual embeddings made using A1111 can fix the anatomy problems of SD3 medium

[errant dust]

The funny thing is that SD3 8B struggles too with this as seen here:

I tested this at Glif for any curious to try themselves

https://glif.app/@gliffyglif/glifs/clv488uy10000djtrx70u03no

[sage burrow]

What would be interesting would be a website listing all of the various AI txt2img, but then listing what each specializes in, and what each is terrible at.

[torn wharf]

every author is going to submit their work as the expert at everything

[sterile pendant]

in the clip folder

along with the clipL file as well

[sterile pendant]

and this is the node and way you'll use it

[bitter hearth]

at this point I am not sure if you really get the difference between VAE and DiT
because you keep making criticisms that apply only to DiT

[errant dust]

ignoring LoRA and only referring to established commercial models, I could tell you that easily, but I might not cover all special use-cases

[errant dust]

Or maybe we have different criteria on photorealistic images

[bitter hearth]

no because a VAE cannot make changes to anatomy

if you VAE encode and decode photos you will understand

differences in VAEs with these models pertain to colours and very fine (high frequency details)

something like anatomy is known as a low frequency detail

[errant dust]

I don't think so. If someone says 'more photorealistic image' or better photorealistic image' then I look at the image. Period.

[torn wharf]

i think here is that for him, the goalposts were shifted long ago from a discussion of the vae and what it brings to the model to "sd3 is bad and you just don't get it" but for you, yu're still on the original discussion about midjourney not having a sota vae yet

[bitter hearth]

but we were talking about VAEs?

[errant dust]

No, YOU said it was more photorealistic BECAUSE of the VAE

[bitter hearth]

no

[errant dust]

really?

[bitter hearth]

it was in the context of a conversation about VAEs

[errant dust]

want me to scroll up and quote you?

[bitter hearth]

the point I was making was that the image is more realistic in terms of the VAE

[torn wharf]

see it's not actually a matter of having a discussion. he has to win. there's no compromise to build a mutual empathetic understanding. no grok

[bitter hearth]

you then jumped to comparing things that don't pertain to a VAE

[torn wharf]

too adversarial. no empathy

[bitter hearth]

its fine there's not drama

its hard to hold a conversation over discord

[errant dust]

Sure?

[bitter hearth]

cos there are misunderstandings

[torn wharf]

i see human drama everywhere especially in moments of irrational discussion

[errant dust]

Ok

[torn wharf]

oh i know. the misunderstandings tennd to be forced through by some.

[bitter hearth]

I already explained that I was referring to the VAE

[errant dust]

No, you said the reason was the VAE

[bitter hearth]

you're assuming incorrectly what I was saying

[torn wharf]

see what i mean by irrational. he's insisting he knows what you meant more than you could

[errant dust]

Whatever

[torn wharf]

sometimes misunderstandings are manufactured

[bitter hearth]

also regarding the hundreds of images
this is an example of one of the sites I used:
https://images.flrty.li/
there are about 200 comparisons there

[bitter hearth]

"trash 3$ toy camera" included in the prompt haha

[bitter hearth]

ok but you're stirring the pot a bit here
there's not rly drama

[noble coyote]

Am I imagining thigs, or is Flux better at limbs and fingers?

[torn wharf]

perhaps. good check. will go make breakfast

[bitter hearth]

Flux is much better at fingers yes

ok bye iceycold

we lost a ball member

[errant dust]

I could not say. He has been on my blocked list for a while

[torn wharf]

one of my pet peeves is people who force through misunderstandings. i often see it when one starts speaking towards what someone else meant. BIG pet peeve. but yeah. perhaps you're right

[bitter hearth]

I can tell you guys had some kind of argument in the past

[errant dust]

Nah. I found him the opposite of helpful. Everything came down to, and this came up not only with me: this is all wrong and you are at fault.

[mortal mesa]

nah, 2 people right in their own ways clashing ideas

[errant dust]

nd backed this up with rudeness, which for me is a kiss of death

[torn wharf]

death. 🙄

[noble coyote]

"The text is above the page!" is often said of literature. For a photographer it would read "The photo is without the camera!" And I hope that's not too contradictory 🙂

[errant dust]

Here is Flux Dev NF4 with the astronaut. Aced it as expected:

[mortal mesa]

vlean the camera lens its blury

[bitter hearth]

I didn't really explain very well before
that I also think SD3 has bad anatomy (its does in my images too)
its the fine details that I like, for photographic styles

Sometimes I just hate the amount of blur in the gens

like come on my phone can do better

[noble coyote]

Do you get the idea that at SAI - the SD3 team 'never' spoke to the BFL team at all?! 😉

[bitter hearth]

they did cos Emad said that he helped the team transfer out

[torn wharf]

must be the suburbia silicon valley millionair neighborhood to have a classroom that ritsy

thats no inner city school

[noble coyote]

(There was a little irony there ...!)

Make your own pinhole camera

[bitter hearth]

in some ways yeah but its was nice of him

emad rly cares about open source

[torn wharf]

speaking of photography

https://huggingface.co/fal/AuraFlow-v0.3 focused on photographs

[mortal mesa]

anything noticable

[torn wharf]

damnit. all day that song in my head now

[bitter hearth]

oh wow the next auraflow

[noble coyote]

[Can anybody tell me why this w/f will not work?} TIA

[bitter hearth]

I think thats it, can't think of any other iconic songs that would have something to do with balls rolling around or something

[torn wharf]

hmm that song def over rides flow rida

[mortal mesa]

\

[torn wharf]

wait a second what does that song have to do with balls rolling around

[rain current]

Lots of DOF

[bitter hearth]

sorry its too hard to read
maybe try and load SDXL base and refine seperately?

[bitter hearth]

can't stop the ball I guess nothing much

I haven't really seen them together that often like that

who knows this one

[errant dust]

Not bad actually, and an image I can deeply sympathize with, having taken apart and salvaged many a legacy lens

[torn wharf]

but wait... it wasn't a rock. it was rock BALLL //organ solo//

[errant dust]

My personal fav is the Mamiya-Sekor 55mm f/1.4

[mortal mesa]

help im a rock - Frank Zappa

[torn wharf]

Starship - we built this city on rock and balls

[devout schooner]

Personally my take on SD3 vs Flux for what people were talking about is, SD3 struggles with anatomy due to obvious undertraining (bad results look identical to early epoch samples from Lora training), but I believe it was trained on significantly more images that were actually photographs and not synthetic than Flux. Just based on their default aesthetics respectively

[mortal mesa]

maybe some Jerry Lee Lewis

[bitter hearth]

wtf thats an actual song cx

yes its either that or its that flux was fine tuned in a way that narrowed its sampling distribution

i.e. that flux started out "better" for photography than the final flux version

and the final fine tune pushed it more towards its current style

not rly sure which

[errant dust]

I think it also has serious guardrails which may have unintended effects on the output

[bitter hearth]

yeah quite possibly

[torn wharf]

i made this one ez

[errant dust]

let me give you an example. If I ask for 'comic-style', then regardless of the specific comic style I ask for, such as 'realistic comic-style', it all looks something like this. Big bold lines, two-tone shading and coloring, and minimal fine details. No amount of prompt engineering will change that.

[devout schooner]

It's weird how they decided to fully include female nipples in Flux, but clearly undertrained them such that they're really strange and melting looking
Like what's the point lol

[bitter hearth]

maybe it could be distillation issue also

[errant dust]

However, if I drop the CFG down to about 1.5.... suddenly it starts to 'remember' some styles, including realistic comic styles.

more tonality, fine pen lines for shadows and details, etc.

[mortal mesa]

you using pro?

[errant dust]

Now this is homebaked Dev

to be able to modify the CFG in Pro I'd need to use up my Fal credits, and my interest was not that high

[torn wharf]

this is the best

[errant dust]

The other Pro offerings don't really give you that level of control

[torn wharf]

no amount of prompting changes anything.

[errant dust]

Sure, you can use Pro in other places, but not to the point of custom resolution, or CFG

[torn wharf]

biting my tongue here. i think trying to explain distillation was one reason he blocked me

[bitter hearth]

lower CFG raises image diversity in all models
so this is consistent

[bitter hearth]

ah okay

[noble coyote]

Wow! Text coherence?! It took me about 10 tries with Flux NF4 to get a long string of text correct!

[mortal mesa]

batch of 4, one came out good, and hammered guidance

[errant dust]

Yes, but that it got that much right is a huge win. The only other generator that could do that is Ideogram

[torn wharf]

memories of ocarinas

[devout schooner]

I find Pro does get text better usually but other than that it's really not a big improvement over Dev

[noble coyote]

My efforts ...

Two good ones and a near miss

[sage burrow]

@torn wharf I did actually try to create a ballz one awhile back, was with SD3 though https://glif.app/@LadyLalita/glifs/clyd621a10000yha2zalqjbb3

[silk solstice]

Love these ones

[sage burrow]

I prefer pro over Dev, but I usually try for painted art styles, which is probably relevant.

I don't like Schnell at all anymore though 😦

[noble coyote]

When I was post-processing using Photoshop, 25 steps in NF4 produced a "quilted pattern" in the processed picture.

I have upped iterations to 35 to try and deal with that quilting

[mortal mesa]

just to add i also use fp16 T5 with the NF4, i have an unscientific belief that i like it better

[noble coyote]

Yes, t5xxl_fp16 rocks

[noble coyote]

Schnell is great simply to rough-out an idea - and decide whether to expend some "real time" on making a good copy

[hollow swift]

https://imgsli.com/Mjg4MDk4
comparison flux nf4 v2 vs gguf v8

[noble coyote]

Carmen Curlers

Why oh why this mat1 and mat2 anomaly?

[errant dust]

You and me, which is why I made an embedded model with the T5 fp16

[torn wharf]

noice. did that use my lora?

Roundish monsters? ROUNDISH MONSTERS?

i dig it

[mint vortex]

Looking for bts ai models please! I found a few but only 2 that looks like Jin and jimin. I need the other members

[errant dust]

What is BTS?

[mint vortex]

A kpop group. I just need the Lora mode for the rest of them

[errant dust]

The first and foremost place would be Civit

[mint vortex]

That’s where i found the two I’m using. Not the rest.

[errant dust]

That was your best try, if not, your best bet might be training one yourself, or enlisting the help of someone to do it for you

[torn wharf]

i know this is off topic but i just want to go on record that nobody caught these free bonus points at all

https://tenor.com/view/heartsign-heart-bts-jungkook-kookie-gif-7611383

[noble coyote]

I want free ice-cream

[sage burrow]

I haven't gotten to the point of being able to get glif to use loras yet

[errant dust]

I don't think you can, since although you can build a Comfy setup, you are still confined to the models they offer

[torn wharf]

sd3 and flux know mad ballz somewhat already. my lora just helps it along, and probably not even in teh best way. it's a REALLY bad dataset still

[mint vortex]

His model is so hard to find a good one to use

[sage burrow]

OK so I used to volunteer at a male teen residential treatment center.... and I said balls one day refering to a qigong exercise... So ever since then, anytime there may be teen males int he audience I use SPHERES instead ROFL ROFL ROFL

[torn wharf]

civit has some for sd15. this is #sd3 though so i dont think we can help you much. i only ever load 1.5 models for animatediff playing

[sage burrow]

It was the most epic beavis and butthead moent ever... "she said BALLS"

[torn wharf]

what!? and take away the joy of balls?

[sage burrow]

You can enjoy your own balls 😉

[torn wharf]

one of my favorite things about this decade has been B&B made a comeback

[old falcon]

Whatever happened to the SD3.1 thing? Still no word?

[mint vortex]

It does? I’ll have to look it up later

[errant dust]

two weeks

[sage burrow]

civita doesn't have them?

[torn wharf]

https://youtu.be/Na3y1DI8bQI for scientific referencing

[mint vortex]

Just for Jin and jimin. Those are the best ones I could find

[sage burrow]

We are distracted enough with Flux for now, that they have time to perfect SD3 😄

[torn wharf]

yeah. stability shouldn't rush it lol. we're satiated

[errant dust]

oh boy...

[torn wharf]

https://tenor.com/view/abell46s-reface-sex-sexo-cigarrillo-gif-18734737

remember hearing that 16 rank lora would do nothing to a 12B model? so many experts came out to claim that immediately when lora testing was first happening

[sage burrow]

Wow, I'm surprised. You can try just typing in their names into a major checkpoint model. Most would know them by now I think.

[torn wharf]

I was honestly surprised too. They're probably the biggest act on the planet right now

[mint vortex]

Tried

Just found a jungkook one.

[torn wharf]

It amazes me that nobody has come close to Michael Jackson famous since , but BTS could be the ones. They still got a lifetime ahead of them

[mint vortex]

https://civitai.com/models/562728/jungkook

[sage burrow]

Suga (warning that site does explicit images far too often) https://civitai.com/models/139045/bts-suga

[mint vortex]

Yeah I saw that one Tbh doesn’t really look like him at all to me

[sage burrow]

https://civitai.com/models/20647/taehyung

[mint vortex]

Been using this one a lot https://civitai.com/models/139391/jin

[torn wharf]

people got to much respect for BTS to ponify them

they've somehow built a huge fanbase of non toxics

[mint vortex]

Tbh it just shows a normal photo of him. Not really showing the model they created

[mint vortex]

Tbh a lot of fans are like that. Which gives the fandom a bad name

[torn wharf]

yeh. lots of celeb loras don't do anything creative for examples. just generate the captions and post those

[mortal mesa]

young people dont usually toxify till later in life

[sage burrow]

Does this help at all? https://civitai.com/models/9006/k-pop-boy

[mint vortex]

Not really that’s good for a ooc lol

[torn wharf]

"jaded" is what i've heard . takes a while for the minerals to saturate in an turn someone's heart to stone

[sage burrow]

Option C. create your own lora. CIvitai has a lora creator on there. Just grab 200 images of each...

[mint vortex]

Wait that’s how you do it? It keeps giving me a pth to create

[torn wharf]

toxic fans i think are people who are obnoxiously obsessed without self awareness too. not just jaded people either. toxicity comes in many forms. swifties will attack people who like another star for example.

[torn wharf]

200 images for one person? wuhwuhhhhh

i thought i was doing 50 was tons

people making ones on flux with 12

[sage burrow]

That's how I do it anytime I want something specific.
You could also try the pony model and see if it h as the data. If you do that, make sure to prompt SFW , trust me o that

[mint vortex]

A lot of toxicity comes with kpop fandoms not just bts. I met a few for blackpink and exo

[sage burrow]

OK perhaps I overdo it lol.

[mint vortex]

I’ll have to try that later thanks

[mint vortex]

Flux?

Sorry I’m new to the whole creating Lora’s 😂

[torn wharf]

i dont know what blackpink is but some people seem REALLY angry about it.

yeah i've seen it a lot. "Fan" comes from "Fanatic" so that's a thing to consider https://en.wikipedia.org/wiki/Fanaticism#/media/File:Eugène_Delacroix_-_The_Fanatics_of_Tangier_-_WGA06195.jpg

[sacred jewel]

Because you lack Skillz?

JUST KIDDING 😛

Worked fine for me... first try...

[sage burrow]

oh wait that makes me think you can also use img2img workflows. On your own coputer, or glif (just cause glif is so easy)

[torn wharf]

oh man wait till you hear about flux

[mint vortex]

Blackpink is a kpop girl group

[sacred jewel]

https://tenor.com/view/wtf-gif-gif-26042600

[torn wharf]

tbh it's also a really hot color scheme

[sage burrow]

Those sorts of comy workflows give me nightmares about dozens of orange boxes

[torn wharf]

this si flux when i prompt BTS on stage. is it accurate?

long way to go before they're Michael Jackoson famous. hundreds of millions of fans is nothing to make fun of , but it aint BILLIONS of fans

[mint vortex]

Bts has 7 members lol and not really doesn’t look a thing like them lol good try though

[sage burrow]

is this a good one? It's Flux Dev

[torn wharf]

do you think anybody that's a fan of beiber or swift or bts or all these world tour acts these days, do you think any of them even consider that those idols are nothing compared to the King

[sacred jewel]

Elvis? He is still alive, you know?

[torn wharf]

MJ is too, but legally owns Elvis's estate. Son in law shit

wait i guess legally both are dead

[mint vortex]

Pretty good lol

[mortal mesa]

the estate is some hedge fund and they were just prevented from selling a house

[torn wharf]

you'll want flux dev then. Seems to know BTS really well already

[sage burrow]

https://huggingface.co/spaces/black-forest-labs/FLUX.1-dev

[mint vortex]

I created this one using aieasypic

[mint vortex]

How do I use it?

[torn wharf]

come on guys they're really phoning in those heart hands here. pop em with conviction.

[mint vortex]

They do that sometimes lol

[torn wharf]

just came out 2 weeks ago but there's already a ton of different ways to use it on your home GPU. Becky linked the hugging face space up a few. there are other sites that offer flux use like Gliphy

Gliffy?

Jliffy

far left has everything everywhere all at once vibes

i dont want to be the one that says they all look the same..... but i feel like thye all look the same

[mint vortex]

I’ll have to try it

[torn wharf]

flux does money suits though

[sage burrow]

I don't know what they look like, but I get the impression this is dead wrong

[mint vortex]

Did this one with a free anime one for Jin but can’t remember what it was lol

[sage burrow]

Try Duc Haitens Pony models

[torn wharf]

I was a clown as a kid growing up with the boy bands an pop queens in US. Brittany. Aguilera. Timberlake. Backstreet boys. Everybody. Yeeeaaah.

I'd do karaoke of it an air band shows

[sage burrow]

Flux can't count to 7....

[torn wharf]

Do a landscape one it might help for that sorta prompt. Flux is trying to compose them all in the pixels it has too

[mint vortex]

I grew up on those too. NSYNC was my jam, along with spice girls 🤭

[torn wharf]

Yeh Yeh it's all good shit. Venga boys too

BT the producer had wicked good solo albums

[mint vortex]

Will it work with a lora model? Plus i don’t remember what the prompt i used was for the photo lol

[sage burrow]

With AI art, you just have to keep trying stuff until it works. Or head over to midjourney, find a good BTS image and borrow their prompt 😉

[mint vortex]

I have mid journey but I tried to use a prompt and it says I need to pay

[sage burrow]

Oh I didn't mean to use MJ. I just mean to browse others' images

[mint vortex]

Now I’m confused 😂

[torn wharf]

Hi confused I am iceycold

[hallow lion]

2 Weeks, or as soon as we sort out hands.

[hallow lion]

Hi. Don;t be confused it;s almost Friday.

[torn wharf]

Better if they take time and get all tools ready to deploy and don't talk about it till it's all good to drop then drop it and talk about it.

Like BFL but they didn't give tooling with the release, so it's a chance to upstage

[hallow lion]

Yes. Learn from Black Forest. No hype. Then drop the goods out of clear blue sky.

[torn wharf]

Also, BFL just dropped it all and then haven't really talked about it

[hallow lion]

😏

[hallow lion]

6 Months silence... then ladies frollic in grass!

[torn wharf]

Kill hype culture. Murder it. Death to hype.

[hallow lion]

Yeah

[torn wharf]

The death of hype shady

[hallow lion]

Hype killed star wars

[torn wharf]

I agree. It's not entirely dead but.....

[hallow lion]

For sure

not entirely...

[torn wharf]

I like andor. But yeh

Rogue 1 was more sw than the trilogy

[hallow lion]

First few episodes of Mandalaroan, Andor and rogue 1 is probably the best new star wars we got.

then Kenobi came and said Well Hellow Thear.

[torn wharf]

hype kills a LOT of creative projects. Hype is the mind killer

[hallow lion]

And it should have been awesome but it was not so awesome

[torn wharf]

Kenobi worked well yeh. flash in the pan

[hallow lion]

lol

I am not sur ehow to process the Acolyte

[torn wharf]

you didn't like it? yeh there are polarized opinions about it i suppose lol

[hallow lion]

it looks like aDune ripoff

like a star wars spin on the bene geserit saga

[torn wharf]

acolyte looks like a lot of dune vibes and i'm all hardly done digesting dune yet. that one is heavy duty cinema

i haven't got to it yet

[mortal mesa]

please watch the first version

[torn wharf]

ive seen all versions many times

[hallow lion]

Whatch the David Lynch Dune

Good times

[mortal mesa]

yup

[hallow lion]

XD XD XD

[mortal mesa]

David Lynch. H.R. Giger

[hallow lion]

Weirding modules for days.

[torn wharf]

Directors Cut lynch version lol. i wish. always was rumors. people would claim to have it. then the internet came along and was like "that shit doesn't exist you liars!"

[uncut river]

I watched Inland Empire just the other day

[hallow lion]

i love lynch but inland was too ugh

[torn wharf]

Sir Patrick Stewart as Gurney Halleck wadddup. I like Thanos doing it too but lets be serious. He's no Knight

[hallow lion]

But I think it wa sjuts the crappy early digital cam technology

if Inland was shot on 35mm I woudl eat it up

[mortal mesa]

https://tenor.com/view/blue-velvet-dennis-hopper-frank-booth-no2-nitrous-oxide-gif-17415359

[hallow lion]

Ah yes.

Of course.

I'll fk anythign that moves.

[mortal mesa]

Heinekin!?!?!!

[hallow lion]

Heinken!? Fuck that shit!

Pabst Blue Ribbon!

[torn wharf]

i'm really basic so gimme the bud light lime. drink taht liek juice. but if i want to enjoy an ale i'll get one i can't see through and has more calories than a loaf of bread

that might count as a stout

tellin u

[hallow lion]

=0

Try FAXE 😄

[torn wharf]

$3 tall boys of faxe you're good to go

[hallow lion]

10% evil viking.

[bitter hearth]

this is flux dev using dpm++ 2m

[hallow lion]

Yeah

i just tried random samplers and this one was th best

[bitter hearth]

i didnt know flux1 worked with dmppp at first

[torn wharf]

works with bosh3 too if you caught the node for it

[bitter hearth]

havent heard of a sampler bosh3, is it new?

[hallow lion]

me neither

[torn wharf]

https://github.com/redhottensors/ComfyUI-ODE

[torn wharf]

if you search the server for "comfyui-ode" you'll find a lot of differetn times they get talked about

[bitter hearth]

hmm interesting, so its kinda like dpm adaptive

[torn wharf]

think @ drhead made them

this was the kpop prompt i used earlier though bosh3. ymmv. i like that finger artifact

[bitter hearth]

neat... i'll wait for it to come out of alpha version

[torn wharf]

at higher steps, it'll take a long time to work, but it works very accurately (in theory). there are other samplers in there too that work slower even. you can really make your gpu grind on images with them

slower on account of it being so damned accurate! (in theory)

[torn wharf]

this one took 90s

tell me that pose isn't a huge success. try it

[bitter hearth]

quality looks good

[mint vortex]

alot of kpop groups used that pose alot

[torn wharf]

oh yeah i know i'm prompting for it, but this sampler got it to generate the rare heart hand duet

[hallow lion]

Secret gem huh?

Like Uni PC.

[torn wharf]

unipc works on flux too yeh

it's a crapshoot still for what works and doesn't on the forge side of things. euler and dpm2++

[mint vortex]

what type of negitive prompt do i use to fix the fingers??

[hallow lion]

Forge is madness.

[bitter hearth]

@sage burrow if you wanna use flux1-dev-bnb-nf4-v2.safetensors https://huggingface.co/lllyasviel/flux1-dev-bnb-nf4/tree/main

[torn wharf]

good prompt theme. madness. balls. ball madness. mad ballz!

[hallow lion]

You can never go wrong with balls.

[torn wharf]

balls save fresh water reservoirs. true story. you drinking water? maybe balls were dunked in it. just sayin

https://youtu.be/YjDH-26Ksac balls are life

space... madness

[mortal mesa]

its a mess but i liked the idea

[torn wharf]

spaaaace
MAAAADNESSS

[mint vortex]

https://images-ng.pixai.art/images/orig/c24f5687-32e5-4f15-b998-e2baa6544ba3 oh dear lol turned him into a wolf 🤣

[bitter hearth]

Madness ?

This is...

(not sparta, its me out of free credits to gen images )

[torn wharf]

descent into madball

[errant dust]

Word in Ideogram discord is that the new model 2.0 will roll out publicly anytime now

[mortal mesa]

people mimicking AI video seems legit

[torn wharf]

love that. thats awesome. the noodles

will smith did it with his spaghetti video too.

https://youtu.be/vbWe5k4fFWE?t=40

Now I am become ball, destroyer of cubes

[hallow lion]

orbital hanging towers

advertising: MUSK DEO

[torn wharf]

one more. flux goes hard

[raven fern]

wow that is actually very nice 🙂

[torn wharf]

Would it be too political to do a ball series on mpox?

[mortal mesa]

thinking people don't conflate illnesses with politics, that just failure from the get go

[bitter hearth]

Do the ball part

[ember coyote]

balls are nice.

[bitter hearth]

Goose wtf

[ember coyote]

You heard me

My 12GB of VRAM cannot handle Flux dev 😦

[sage burrow]

Flux pro 🙂

[sage burrow]

You sure? I can run it on 8. Takes 10 mins tho lol

Have you tried nf4 yet?

[ember coyote]

I lied, takes 210 seconds for me right now. What’s Nf4?

[hallow lion]

Impossible becausee mine can. It actually uses like 6 gigs on Forge lol

[sage burrow]

That's not too bad. Nf4 is a version that makes Flux take half the time, or even less.

[ember coyote]

https://huggingface.co/duuuuuuuden/flux1-nf4-unet/tree/main is this what I’m lookin at?

[torn wharf]

this hit was runnin on ball train for years

[mortal mesa]

when you thought it was settled https://huggingface.co/city96/FLUX.1-dev-gguf

[torn wharf]

dodododododoodoododo do the ball part dodododododo dodo dodooodoo

i remember when thye tried to introduce humans to the show but the square didn't fit and no one liked it. bigger than new coke

https://education.civitai.com/quickstart-guide-to-flux-1/#train-flux-lora-on-civitai

[torn wharf]

and i guess there's some code in kohya to allow training to be split into parts and fit into vram

[hallow lion]

Best 70 Sitcom. the Ball Train.

[bitter hearth]

not bad for schnell 4 steps

[sage burrow]

Id call this an img2img flux pro fail, but some people might feel it's a great success 😄 @torn wharf

[hallow lion]

F R I D A Y.

[sage burrow]

Bosch is rolling over in his grace right now

[sacred jewel]

Modelo Especial FTW

[mortal mesa]

Flux found a woman in my word soup I'm not able to create an image prompt from that text input, it seems to be a series of characters that don't form a coherent phrase or description. If you meant to provide an image request, please rephrase the input into a brief and descriptive phrase that includes keywords for (an adjective, type of image, framing/composition, subject, subject appearance/action, environment, lighting situation, details of the shoot/illustration, visuals aesthetics and artists), I'll be happy to assist you in creating an image prompt.

[lavish vine]

Naked girl

[errant dust]

So for anyone interested, Aura Flow 0.3 is out, the third model in a bit over a month. It is still in beta, so keep that in mind. Here is the link: https://huggingface.co/fal/AuraFlow-v0.3 Just download the main file into the Checkpoints folder of Comfy. Here is an image with simple workflow.

[fading basalt]

oooo. do ya have a prompt for this?

trying to copy it 😄

[sacred jewel]

Damn. I just went to bed.

Remind me tomorrow or hit me up via private dm 👊

[raven fern]

so you telling me you cheated? jk 🤣

[sage burrow]

Flux pro via glif, with gpts' help, I cheated big time 😉

[noble coyote]

Do the Paris Olympics do medals for downright creativity?!

[noble coyote]

Including Face Detailer? With me it gets stuck at mat1 and mat2 shapes ... ?

Ordinarily it works just fine - yet when I add-in the Face Detailer - I get the above error?

[bitter hearth]

the way you lay it out is kinda the worst case scenario

in terms of how confusing it is

[noble coyote]

Change the noodles from spline to straight?

[bitter hearth]

its difficult to give advice that would apply to every workflow
but mostly
space it all out a lot more, across several screens
and consider bus nodes
also sometimes I have multiple loaders instead of loading something once and then stretching noodles all the way

a big thing is to choose the horizontal and vertical position of nodes on the basis of whatever would produce the least line crosses

[noble coyote]

It works perfectly well without Face Detailer, but stops when Face Detailer is unmuted. I agree that it is not attractively laid out! 😦

... still waiting for Aura.Flow to d/load ...

[bitter hearth]

Big Barrel

[noble coyote]

Looks like Paris

Horse Face changed on OpenArt Face Changer (free)

[noble coyote]

Any body used these at all? https://huggingface.co/marduk191/Auraflow0.3_collection/tree/main

[tough wyvern]

#artisan-1 asd

[noble coyote]

^..^<

[noble coyote]

SwarmUI SD3

[bitter hearth]

does dev nf4 version support lora yet?

[sullen moss]

It's like coming home from an IMAX theater and sitting down to watch a black-and-white TV.

[sullen moss]

Nope

[noble coyote]

Art is cool on a B&W TV (imho) 😉

But SD3 does stop me using all my favourite artists in the prompt 😦

[bitter hearth]

i see ok, there's been comfyui update yesterday or something and some ppl are saying nf4 might support lora after that

[noble coyote]

There are 5 Dev LoRAs on Civitai

But full Dev takes a month of Sundays!!!

[bitter hearth]

im curious on the development with nf4... regular dev takes too long to render on my system

[sullen moss]

I just don't really follow this format much because I use the standard one.

[noble coyote]

ComfyUI update is mainly toolbar focussed

And css/html instead of visual basic

[bitter hearth]

schnell nf4 takes 15 seconds about, and dev nf4 takes around 1min few seconds

[noble coyote]

I'm getting used to it

On my 8Gb VRAM 64Gb RAM setup, nf4 takes 35 steps and one image every 2 minutes 10 seconds

Flux.Dev can take up to 30 minutes only at 20 steps

[bitter hearth]

30 mins... 👀

[noble coyote]

But Flux.Dev can take advantage of LoRAs

[noble coyote]

Yes, VRAM too little ...

If I don't use 35+ steps on nf4 a strange "quilting pattern" can be seen ...

I'm eager to try the new ODE Samplers in ComfyUI - and AuroraFlow3

[bitter hearth]

This is the same image with SAG+PAG off

and with SAG+PAG on

im liking the comfyui new updates

UI looks good

[sacred jewel]

Yup. No hangups 🤷‍♂️

[bitter hearth]

maybe instead of using facedetailer

write the process out manually with impact pack nodes

all the detailers are just inpainting with some combination of Yolo, Grounded DINO and Segment Anything

[turbid grotto]

Guys do you know why guff q4 flux which is about the size of sdxl takes same amount of s/it as full fp16 version of flux?

like, shouldn't it be faster?

isn't llm gets faster with smaller quants

[bitter hearth]

the otter (?) 10x better than the robot

the otter is better yeah

[sage burrow]

The auraflow before this one wasn't worth the price, so I'm more than hesitant to try.

Has comfy figured out how to make the nf4 install easier yet?

[bitter hearth]

ugly

[sage burrow]

Art critics everywhere!

[sterile pendant]

Yo... Whatever City96 did with the GGUF loader addon last night... Omg... I'm using the dev q4_1 model and can generate images at 4 megapixels, without oom errors and while getting 15 sec/it, all on an 8gb gpu

Regular sized images in the 1mp range are around 3 sec/it

[bitter hearth]

Uglier

machines are taking over the girls

can you smell the sdxl in these

[bitter hearth]

very good composure and sweet girl

[noble coyote]

I got it up and running with the help of @errant dust in under 12 minutes

Install bitandbytes ver 43.3 first

Make sure you have a Cuda of 12.4

[sage burrow]

Did all that, kept getting errors after an hour finally got it to work. Then accidentally hit install the next morning and destroyed it 😦

So now I'm waiting in a method that works without 100 errors on the way 😄

[noble coyote]

This is SD1.5

[sacred jewel]

He must have worked at a Sawmill in his early days 😛

[bitter hearth]

turns out you get good color grading if you just put R2D2 as the IP adapter images

[errant dust]

YMMV. It is slower for me than NF4 on 1k images.

NF4 v1. v2 is a snail as it exceeds the confines of my 8GB

[bitter hearth]

4 megapixels?

it can do that without structural errors?

[errant dust]

Might, but I get a massive slowdown on NF4 when I exceed 1k anyhow

nevertheless, will check it out. the diff in speed is not huge between gguf and NF4 at 1k. NF4 is clearly faster for me, but once it tries anything larger than 1k, even 1280 x 1280, it slows to a crawl. Will see how GGUF model handles it. I tried them all last night. Well, all the Q4 and Q5 models

[mortal mesa]

NF4 was indeed faster for me also

[sage burrow]

How can flux not know woodcut style, even with gpt's help?! 😭

[sullen moss]

Where did all the SAI representatives go? Ever since FLUX appeared in this channel, none of them have posted anything.

[dusky thistle]

"appeared" 😂

[sage burrow]

So is it possible to run cascade via comfy? How?

[dusky thistle]

yup

[dusky thistle]

pretty easy, just need the unets from the huggingface page in your comfy/models/unet folder, etc

[dusky thistle]

https://github.com/ClownsharkBatwing/UltraCascade

need to clean up the description etc but got it up now

[sage burrow]

I'm getting really tired of that faded look Flux seeks to default to. If I were to ever print and sell one, everyone would ask if I was running low on printer ink.

[brittle nexus]

What they can do? Send us to another discord won't help them. What can they say?

[dusky thistle]

lol

[craggy crest]

add vibrant to your prompt

[sage burrow]

So this is interesting, same with sd I wonder?

[sterile pendant]

update it, it's fixed. i'm literally running the q8 gguf now at 4 sec/it and that's pretty much the same speed i was running nf4v2. this is an 8gb gpu. i'm even doing 2048x2048 images at like 16-18 sec/it. also make sure you aren't using the force clip device node anymore

[sage burrow]

This is goind to take some time!

[rain current]

nf4 or fp16?

[sacred jewel]

Full Flux.1 Dev

[dusky thistle]

you need the RES4LYF repo to be able to use those exact nodes

https://github.com/ClownsharkBatwing/RES4LYF

[sterile pendant]

rtx 2080 FE and this is with Q8

[torn wharf]

hmmmmm. 2000 civit buzz and i can train a lora for flux. so i go look in the bounties to see what i can do. TMNT lora for 100k? wtf.... okay.. oh. they want TMNT from the 80s in porn scenes.

The bountie board at civit is disgusting shit. Thats outta hand.

[sage burrow]

Oooh, did they add flux to their training models now? 😄

[torn wharf]

yeah but i got no buzz . and earning it means participating.

[sterile pendant]

it's supposed to be tiger sized, but there's nothing to compare the size to. so you just have to pretend it's a 400lb cat

[sage burrow]

click on any image, hit the remix button, click on the claim 25 buzz...but don't generate, or you lose some buzz. daily. Or buy some for, $5, costs $2 to create a lora apparently.

I have lots of buzz because I was going to create a bounty before, but everyone told me just use controlnet after I posted it.

[sage burrow]

Way above my skill level. Does all this work on Mage or glif? 😄

[craggy crest]

you'd have to ask Roi about it for mage, they'd have to impliment it

[dusky thistle]

nope

[sage burrow]

I'm glad to see he apparently doens't have bad breath 😉

[sage burrow]

Yet... 😉 😄

[sacred jewel]

The power of MENTOS

[torn wharf]

FRESH GOES BETTER MENTOS FRESHNESS

https://tenor.com/view/foo-foo-fighters-big-me-grohl-dave-grohl-gif-13350388

[sage burrow]

So only $2 per Flux lora creation now... I've been saving up my buzz (actually I was saving it ever since sd3 just in case lol). Time to experiment!!!

[sage burrow]

and the benefit of using their service is, you can up to about 1000 images last I remember!

[torn wharf]

$2 is all it costs for 2000 buzz? hm

maybe i don't have to build a pornographic 90s ninja turtles dataset after all

[bitter hearth]

only problem with online lora creators like civit is
they set the learning rate super high so that most loras don't fail

this is fine but it means it over-fits

so the lora result might kinda lock you into that subject a bit

but that's ok cos most people seem to prefer that type of lora anyway

[sage burrow]

Min purchase is $5 for 5000 though.

[sage burrow]

That is my plan 😄 Though, ty, I'll change the auto settings for my next one and compare.
I created several with SDXL before and they turned out pretty awesomely. Though I probably have higher standards now, so we'll see 😄

[bitter hearth]

a lot of the time people use Lora to create a specific character or a movie scene
so its ok if the Lora makes the model forget quite how to make teapots like it used to

[torn wharf]

fuck. no deal then. it's not even an evenly divisible amount. thats how they getcha

/chronically cheap/

[sage burrow]

2 loras for $5 seems less expensive than the other methods?

[torn wharf]

civit's allows you to define the lr. looks like the default is 5e-4

[torn wharf]

i'm just being bombastic. i might use it yet.

[sage burrow]

Your skill levels are MUCH higher than mine for such things though, so you might be able to do it for less using Vast?

Quickly making loras before they decide to make Flux ones more expensive!

[torn wharf]

i'm out here winging it most times. but i'm also a big believer that most of the skill is in the training dataset building. settings are a lot more forgiving than people think. i lean towards UI's i'm comfortable with and can affect easily. convenience.

like you know how kids will dive off the 20m tall tower at the pool, even though they're not olympic divers? meeting the challenge is a big step, but it doesn't make them into contenders. thats the skill level i think i'm at. i'm just running and leaping off and then trying to land in a non bellyflop position

civit looks appealing to me to try once or twice, because it's a stupid easy interface from what i can gather

i've also been poking around at this library from nvidia, "transformer engine". if i can get it working on windows or a docker, can potentially train flux in fp8

[bitter hearth]

people are paying for loras?

I'll do them for free

just give me the source material

hungry ahh mfs tryina monetize everything

[torn wharf]

greed is good. coffee is for closers

[sage burrow]

Now to figure out where to use my new loras. Mage doesn't have Dev yet (last I looked), and Glif doesn't add loras (quite yet).... Also, I don't think Dev loras work with nf4?

[torn wharf]

https://www.youtube.com/watch?v=VVxYOQS6ggk educate yourself

[bitter hearth]

just use prodigy

learning rates is so 2022

prodigy has never cranked out a shit lora for me

[torn wharf]

do you know why rickrolling is a thing? because the memes that have the most staying power come from the 80s. like wallstreet.

[torn wharf]

prodigy often is just adamw that finds a learn rate in the first 20min of training and then stableizes. running adamw at that stable lr is no different.

[bitter hearth]

yh its a dynamic scheduler

[torn wharf]

it's also optimized for unets and not MMDiT networks

[bitter hearth]

its not the only one of its class, but it is the best one Ive tested

oic

yh the same stuff that worked for 1.5 is prolly not gonna work for flux. I am trying to learn how to train flux after all

[sage burrow]

Looks at Prodigy pricing............ $$$$$$$$$

[torn wharf]

at this point in the game, you're better off sticking to adamw for flux and sd3. the adaptive ones will just find a lr and stableize at it. just use that lr instead

[bitter hearth]

its not a website

its a training method

so that ppl dont have to set a learning rate manually

[sage burrow]

So you don't need to dl some overpriced software?

[bitter hearth]

noooo. kohya is free

and it comes with the scheduler

if yall can help me find a tut on flux loras I can gladly train whatever yall want free

I mastered dreambooth and loras for 1.5

flux is new tech

[torn wharf]

rip prodigy ||https://vimeo.com/144850907||

[sage burrow]

Kohya is for the rich who can afford 16gb GPUs 😄

(or it can be used on a cloud too though)

[bitter hearth]

I don't understand language of people with short money. 😎

I got 24 gigibytes

[sage burrow]

https://x.com/angrypenguinPNG/status/1823064101528260772

AP (@angrypenguinPNG) on X

FLUX [DEV] LoRA Training in Google Colab ❗️❗️

Notebook and code by the 🐐 @ostrisai

Twitter · Aug 12, 2024, 6:28 PM

[torn wharf]

ugh. attitude. noted.

[bitter hearth]

hey I'm volunteering to help.

free GPU power

[sage burrow]

My computer budget was for 8gb GPU only lol

[bitter hearth]

loras take like 30mins top

[torn wharf]

uh huh. you're flexing and it's tacky

[sterile pendant]

It's all the same shit: curve fitting, derivatives through back propagation and setting the million knob directions to be scaled by the learning rate per step. Prodigy is good shit, you just mostly hear about other methods because people are scared of change and the sunk cost fallacy that goes along with them having spent hundreds to thousands of hours perfecting their settings and workflows.

[sage burrow]

Do you mind several hundred N SFW furry images? LOL

[bitter hearth]

OMG

hit the nail on the head

[bitter hearth]

do you want them all in one lora or different ones?

[errant dust]

Well, to compound the difference, I could not run V2 at all faster than say plain Dev. Only V1

[torn wharf]

what sunk cost? it's a tick box.

test it on sd3 yourself. it finds a lr and stablizes. plug that lr into adamw and give it a similar warmup on the same training seed. similar network will be produced.

[bitter hearth]

I started reading the prodigy paper it sounds good
thanks for this

[torn wharf]

or have you tested prodigy and invested too much time into it to investigate this?

[bitter hearth]

alls im sayin is, prodigy works. its made to avoid overtraining. so if you wanna train 2000 steps or 10,000 steps, it will adjust the lr so that it wont overtrain

and...it works. ive trained about 80 loras

always cooked to perfection 🧑‍🍳

I just load the .json on kohya

select the source folder and click a button. no more

[torn wharf]

prodigy can very much still over train. done it a few times with it on xl and 1.5 models

other options mitigate that. less steps. decay. etc

[bitter hearth]

and

[errant dust]

I will add this though, after testing them all at length at 1280 x 1280, only the GGUF models were actually doing them in usable times. NF4 quickly becomes impossible. V1 or V2

I say remotely usable, meaning 6 mins

[torn wharf]

nf4 on my system is 30 second gens.

Your tests are qualified by your system specs and your skill level it looks like.

[errant dust]

as opposed to.... 20 for NF4

[torn wharf]

page file swapping is a biiish

[errant dust]

NF4 1280

loaded partially 5323.275 5859.856831550598 0
100%|███████████████████████████████████████████████████████████████████████████████████████████| 35/35 [20:46<00:00, 35.62s/it]
Requested to load AutoencodingEngine
Loading 1 new model
Prompt executed in 1335.04 seconds

[bitter hearth]

how much vram is this?

[errant dust]

my laptop 4060 has 8

Q4_1 1280

Prompt executed in 352.33 seconds

[sterile pendant]

I've trained dozens and dozens of lots with datasets ranging from dozens to hundreds of images. Used every major type of training method. Prodigy is extremely hard to f up, rarely over trains beyond recovery(picking an earlier save point) unless you have a bad dataset or screw the settings up (don't use network dropout ANYWHERE else other than in the prodigy settings and never go above maybe 3-5%) and usually settles quicker than a lot of the other methods

[sage burrow]

both. One with a mix of everything, then a few with specific themes. It'll take me awhile to organize them though!
But starting with just one, with a specific theme would probably be good.

Wait, you said half an hour?!!!!!! How much vram do you have?

[bitter hearth]

24giggies. its a 3090

ya I can do that np. I am curious to test flux after all

and am gonna do some loras of mine as well

I just need to learn the settings for flux first since its all new

[torn wharf]

it works best on larger datasets in my opinion. my poorly built ball dataset is one that prodigy hates dealing with. i heard once that the authors intended it more for anime, but i've never encountered problems with it for that

[bitter hearth]

loras are usually quick to train, its dreambooth that takes like 3hrs for me

[torn wharf]

also, network drop out is a tool i use to generalize my loras a lot better. maybe that's why i have difficulty with prodigy sometimes. thats a good thing to mention. thanks

[sterile pendant]

like these were my settings for the last lora i trained, used cosine with no restarts and no warmup. lora turned out fine

the key is to make absolutely sure you dont have network dropout in any of the other settings

[sage burrow]

I did a checkpoint with everydreamtrainer2 once... it has a dataset of 15 images LOL. Took 4 hours haha

[bitter hearth]

I'm trying to learn Huggingface Accelerate
cos it lets you split stuff across multiple GPUs
I find good deals at night time for Swedish 8 GPU servers cos no one wants them lol

[torn wharf]

where is network dropout in the prodigy settings?

[sterile pendant]

in onetrainer or koyha?

[torn wharf]

either. i'm agnostic to the interface as long as it's an interface

[sterile pendant]

in koyha, you have to add it to the line with the betas and whatnot

in onetrainer, you click the options

[torn wharf]

this screen?

[bitter hearth]

Lack of balls and sd3 today

[torn wharf]

that goes to the screen where i don't see any network dropout options, the one i just replied to

[bitter hearth]

people still use sd3?

[sterile pendant]

this is in the lora tab

set that to zero

[torn wharf]

yes. why woudln't they?

I mean, if they're fixated on porn then no. but other people are still using it. i thought at first i wouldn't ditch flux but sd3 suceeds in some situations

[sterile pendant]

there are other settings that are technically "dropout" like settings, like noise and whatnot, but i'm not doing a deepdive on it all right now

[sage burrow]

SD3 is awesome if you want particular artist's styles. Flux isn't as good at that.

[bitter hearth]

Sd3 best model

[bitter hearth]

flux is infinitely better

[sterile pendant]

oh and weight decay = dropout in the prodigy settings

[bitter hearth]

better efficiency, better detail, better ANATOMY

better prompt adherence

[bitter hearth]

For anatomy yeah

better nsfw 😏

[torn wharf]

this is the key right? but the one you just showed me is the only place network dropout is.

In kohya it's only in one place too and i can't find anything in the prodigy paper https://arxiv.org/pdf/2306.06101

[bitter hearth]

Other than that.. Questionable

I went back to SDXL
but now I throw up to 50 images of R2D2 into IP adapter to get different looks

[bitter hearth]

Pony is better

[sterile pendant]

read what i just said and scroll up to my first screenshot. it's weight decay. basically does the identical thing that network dropout does. randomly forgets some precentage of connections

[torn wharf]

i'll keep using network dropout i think. if it's incompatible with prodigy then i'll just avoid prodigy

[sage burrow]

I have yet to see that... but give me about half an hour, my civitai lora experiment is running lol

[torn wharf]

weight decay is not basically the same thing as network drop outs.

[sterile pendant]

for prodigy, it is

[bitter hearth]

remember that this is still only base model flux

once we get finetuned flux.... ooh la la

consequences will never be the same 😏

[sterile pendant]

this is why people have issues with prodigy, because they have a bunch of other stuff hammered into them

[bitter hearth]

people love making stuff complicated.

I dont know why, but they do.

[torn wharf]

if we're just making things up, then i dont' think we're having an honest technical conversation. i'm trying to legitiamtely ask you for accurate information of the technical sides of things, and you're giving me loose "its basically the same" explanations.

forget i asked.

[bitter hearth]

prodigy is made specifically so you dont have to twiddle a million knobs. just load the damn .json and train!

I can send yall the .json if yall need it

[sage burrow]

I can only imagine how many loras people will have out about 1 week from now!

[dull star]

https://civitai.com/models/651715/oofas-flux-style?modelVersionId=729105

Claude Monet

[torn wharf]

"read what i said" is also very abrasive to me. We can all read here. I think that's apparant.

[bitter hearth]

Post it on Reddit

[torn wharf]

i posted my balls lora to reddit and the stable diffusion community mass reported me for spamming. my 2nd appeal just went through and my account finally got reinstated

don't post to reddit.

[sage burrow]

Ty 🙂

[bitter hearth]

mfw I have a nipples lora 👀

the reddit is remarkably bad compared to the discord

[sterile pendant]

look in the code, and network dropout is an artifical thing hacked into the algorithms

[bitter hearth]

not just for SD overall but for comfy too

and diffusers reddit community doesn't rly even exist

[torn wharf]

all the code is an artificial thing

[bitter hearth]

Lmao

[torn wharf]

i'm not sure what that means

[sterile pendant]

it just willy nilly nixes some random weights at each step and wasn't actually a part of the original research

[torn wharf]

yes. the name is self explanitory. it drops out parts of the network

[bitter hearth]

I know Reddit is a deranged sewer of people

It was a meme

[bitter hearth]

this is something I always worried about with the training scripts like kohya
that they could just put some random things in

[sterile pendant]

prodigy uses something more along the lines of alpha=(-group["weight_decay"] * lr)) to more properly decay the weights

thats from adafactor, but prodigy is likely the same

i don't feel like pecking through a bunch of code to fully explain it, but it's better than just randomly killing entire connections

since the algorithm is already adaptive and all

[sage burrow]

Better than the ones I've seen on civitai I hope 😄

[sterile pendant]

yeah and 99/100, commonly circlejerked scripts are just specialized for people making their furry anime pron with some ultra specific dataset size and expected GPU vram for things like batch count and whatnot

[torn wharf]

caption shuffling and caption dropout are artificial hacks just added in later too. so is cosine scheduling. cfg is a hack added after the fact. i'm still hung up on that point, like why even make it?

[bitter hearth]

Here

That's literally the last prodigy .json you will ever need

at least for 1.5

[sterile pendant]

well i'm not going to bother trying to covert you over or anything, you do you. prodigy works great, you just have to use it right.

[torn wharf]

like any tool . tools have proper uses and work for some and don't others. it's why there are many different kidns of wrenches for the same job.

[sterile pendant]

for instance, there are probably a dozen variables in onetrainer/kohya that literally end up doing the same thing: scaling the learning rate. like an easy one is network dimensions vs alpha. divide one by the other and boom! that's essentially your scale factor. if it turns out to be like 0.1, then whatever learning rate you're using is now going to be lr*0.1. oh and another big key thing is that for prodigy, make sure to set all learning rates to 1.0. not sure if it automatically ignores the settings or not(hopefully), since it handles it all

now you fuck with batch count

and boom

another 0.25 scaler

on top of the 0.1, so it's now 0.1 * 0.25 * lr

[torn wharf]

weight decay is actually one aspect i've really tried to read up on before. https://arxiv.org/abs/1711.05101

i've not found anything to suggest that it's incompatible with network drops or caption drops or rank drops. they're all just different tools to affect the network in different ways. here's a paper on network dropout btw. not just a "hack" but actual real research (which in many cases is just hacking with academic funding) https://www.cs.toronto.edu/~rsalakhu/papers/srivastava14a.pdf

[sterile pendant]

yeah, that's the setting you use with prodigy. network dropout isn't the same

[torn wharf]

You might recognize a couple of those names on it

[sterile pendant]

and that paper is old as dirt now

[torn wharf]

right. kidna my point. network dropout wasn't just added after the fact

if you knew, why'd you say that?

you can tell it is old because ilya's email address isn't [e-mail]

[e-mail]

[bitter hearth]

wow the original dropout paper was Ilya? that's cool

[sterile pendant]

network dropout has always been a thing, been working with AI for over a decade now(simple shit like teaching a network to win a super mario level).

decay ends up working out better in the long run for shit like loras because your aren't teaching mario that he has to move to the right, not the left

within a single pass of the dataset, while training a lora, the network already has a pretty good understanding of which way the neurons need to go

that's the whole point of having back propagation: you don't have to brute force shit now

[torn wharf]

this is weird because now you're saying network dropout has been around all along and you have lots of experience, but earlier you said this which we just established was not entirely correct.

i got experience galore too, just not in ML. I would craft algorithms for game bots and i did a ton of database normalisation and i lived in flow charts, diagrams, and vim. i'm no summer child. i'm not going to hold any of that up as a reason why what i'm saying is correct though. i rely on source info and manpages for that usualy.

[sterile pendant]

network dropout isn't normally included in the official model training papers for architectures, that's why i said that

[torn wharf]

neither is prodigy

[sterile pendant]

you're confused what prodigy actually does though

[torn wharf]

it's adamw with an adaptive learn rate. same as adafactor and dadapt. i think it's a fork of dadapt. all basically the same jist , justusing different algorithms to figure out their deltas.

[sterile pendant]

we can sit here arguing about this all day, but i've got shit to do. like i said earlier, you do you and think what you want. at any rate, if you use prodigy and want dropout, you only use weight decay with it and do not use regular dropout with it. you might be able to set weight decay to 0 and use network dropout, but i don't think the algorithm likes it as much, so i don't recommend it

[torn wharf]

ultimately a good adaptive algorithm shouln't break with network drops. it should be entirely decoupled from that and when the network drops, the adaptive learn rate figures it out. That might be tough though becaues how do you calculate loss with the networkdrop before you get there. That consideration might be the root of why adaptive learn rates don't play well with dropout

[bitter hearth]

its not rly worth arguing about, its just a case of going through the papers again

[torn wharf]

when i'm trying to figure out something, the last thing i want is a magical faith explanation.

I rely on faith a lot but i also like an accurate lens to peer through that veil with.

"just a hack added later" 🙄

sorta like http

also in my experience network dropout is valuable and not worth giving up. weight decay even works well with it. i use that combo with adamw often. i've stopped using so much prodigy and dadapt because when i look at the tensorboard graphs, the lr just levels out at one value

then it's just adamw with a static lr and a little bit of overhead to calculate the non delta adapting every step

[sterile pendant]

I wasn't going to spend 15min typing up a rant about it. Network dropout is kind of an OG carryover from early ML before we started using shit like back propagation. It was there for things like teaching Mario to run and jump toward the right and to prevent getting stuck because of some bad cluster of neurons in the network. When the dataset is known, you just back propagate everything and don't need to rely on vanilla dropout like that since it takes derivatives of every step backwards to figure out exactly what combination of +/- to apply to all the knobs to fit the n dimensional data curve and can then scale the learning rate by that factor as well. You really aren't going to end up with as many "stuck" nodes like you would if you were blindly brute forcing the actual training, that would then benefit from completely deleting the connection. We already know how wide and deep the lora will be, so it's also not going to balloon out of control either

[mortal mesa]

if everyone followed the papers we would not have any advancement

[sterile pendant]

So using network dropout for loras usually just makes shit train longer to keep plugging the deleted holes vs decay handling it more elegantly

[bitter hearth]

yeah this is true
but every time I replaced advice I saw on reddit/discord/youtube with what the papers say, it has gone well LOL

[bitter hearth]

a great example is the samplers and schedulers
which sampler to use and for what schedule is discussed in massive detail over the course of like 100 papers

[mortal mesa]

well gossip vs developed and tested theories, im not surprised

[glossy pilot]

I have a stupid question!

Is sd 3 somehow related to flux?

[mortal mesa]

has similarities

[glossy pilot]

VERY LARGE

[sage burrow]

You see a new model came out, so we are all using it instead now lol

[glossy pilot]

Has anyone tried using it via colab + lora?

I mean flux

[bitter hearth]

SD3 and Flux are kinda the only two in their class
cos they are the only models that have all 3 of the new meta:

1. rectified flow instead of diffusion
2. big 16 channel VAE
3. DiT instead of Unet

[glossy pilot]

• Can someone test my LoRA Flux bc it seems to me that it will be Very Bad (I am writing here bc for flux you need a prompt like sd 3)

[glossy pilot]

It makes sense I guess

[bitter hearth]

Kill it

Another furry

[torn wharf]

auraflow lacks the 16chan yet but i trust they'll get to that. The other two aspects are a BIG deal yet

[bitter hearth]

yeah I seem to remember either from reddit or discord that both the auraflow and pixart models in the future will have 16 channel vae

rectified flow is great cos we don't need the SDE samplers any more

its a lot faster to use ODE

well, unless you use the super slow ODE samplers of course

isn't ODE sampler still in alpha tho?

euler, heun, DPM 2M++, UniPC are all ODE

oh huh

euler a, DPM 2M++ SDE are SDE, for example

well im using euler as we speak with flux schnell

what about dpmpp 2m

oh nvm you mentioned that

there's only really one rule
and that's no SDE samplers on rectified flow models

so long as you don't break that rule its ok

well its either euler or dpmpp 2m depending on schnell and dev

dpmpp 2m is a lot better
euler is mathematically the worst possible sampler
it still does ok but its as simple as possible

yeah but euler is recommended for schnell

you don't have to follow the recommendation though
that's the fun part of the models

this is schnell with euler

most of the crazy samplers I use like Bogacki–Shampine are not "recommended"

let me try that same seed with dpmpp 2m

[mortal mesa]

deis/beta

[bitter hearth]

I love the beta scheduler yeah

I divide all the sigmas by 0.8 though