🆕｜sd3

[toxic bone]

https://youtu.be/pq1T6z-Lbxg

[bitter hearth]

I like this one from your examples

[toxic bone]

https://civitai.com/articles/5799/buzz-changes-rewards-and-botting before or after this?

[bitter hearth]

its a bit too contrasty for my taste but the water is excellent
that's kinda close to how midjourney does water

[toxic bone]

the only peoplle earning buzz on civit before they changed the system wer botters

[bitter hearth]

my personal taste in images is rly soft and pastel like this:

is partly why I use SDXL

thanks

[toxic bone]

i actually already disable adblocker on sites i regularly visit. least i could do

civit already earned that status but i noticed there were no ads. i'm not bothered there are now

[bitter hearth]

I like the jet ski one a lot

in general I like water I think

but yeah I definitely go for lower contrast and colours

[bitter hearth]

there's a big advantage of higher contrast- you can use more CFG

[toxic bone]

https://civitai.com/images/14105830 you posted it before the changes and you've got obscenely botted counts on it. and your contest win was revoked. hrm

[bitter hearth]

I have to use such low CFG for my soft images that it gets a bit janky

SDXL can't do chandeliers until quite high CFG

assassin's creed is an interesting style yeah

[toxic bone]

Unity is the best way to explore notre dam since you know, 🔥

[sacred jewel]

Michaelangelo LoRA

[toxic bone]

cathredals burning cathredals burning put your hands up! 🎶

you won the contest before the botting changes. Nobody was competing against botters then. Like the Tour de France, if you won you were doping

https://civitai.com/articles/5799/buzz-changes-rewards-and-botting

[sage burrow]

Now you can watch ads on civitai to gain enough buz, that you can make a lora for free, once every 5 days.

I can't find their ads though. I must have installed an adblocker I forgot about

[toxic bone]

before that, there were github projects to automate creation of civit accounts and give yourself votes. Civit had no defenses against it

everything before that post is the doping era of the tour de france

i believe you do know what i'm talking about

[sacred jewel]

Dan Mumford LoRA

*you're

[toxic bone]

You know very well that you wouldn't need to know coding to use one of these projects, the same way you don't for sd. you're just pretending you have no idea what i mean.

We've added additional measures to detect suspicious Buzz sharing and usage to combat Buzz farming bot networks. When such activity is detected, the user and any additional accounts associated with their IP address will have their ability to earn Buzz disabled. This system isn't perfect, and we'll continue to refine it.

bragging about all the buzz you won on civit before that post isn't too bright.

[bitter hearth]

you've been saying stuff like this for a couple of days its probably time to stop now

I don't rly see the point in continuing such conversation

[bitter hearth]

is that the inside of a lightbulb

[sage burrow]

Well he obviously hasn't botted YT yet... 😄 lol

[toxic bone]

bad actors should be revealed. thats the point.

[bitter hearth]

ok but you've got your message across now

[sacred jewel]

more Mumford

[toxic bone]

if he was winning civit contests before june he was. Same way the champion of tour de france are guys nobody heard of who just did the race for fun. Every cyclist in the top positions were doping and disqualified

• yer

[bitter hearth]

I'm not sure how much it matters when most of the top Civit models (zavy, colourful, jibmix etc) all look kinda similar
his model looks like those
so its plausible he won without bots since his model looks like the most popular models

[sacred jewel]

Hey man, it's You Are .... so "you're" ... fyi

[toxic bone]

context is about winning an image contest for 100k buzz

[bitter hearth]

oh the image contest

I thought it was about models

[bitter hearth]

anyway when I want to use a model that looks like this (high contrast, vibrant colours, semi-realistic) I personally use Leosam Hello World, but I used to use Zavy

[sacred jewel]

GREAT models indeed. That Duc Huytem one is awesome too. (I butchered his name, let me get the right spelling)

[toxic bone]

i like how civit catered to the buzz farming crowd though. Enforcing against bots really got a ton of people mad at them. but then they built this olive branch which i personally just love. Giving them a place to go that properly labels them. https://civitai.com/collections/3870938

[sage burrow]

Whatever, at least an animegirl image didn't win lol

[sacred jewel]

Duc Haiten

[toxic bone]

those civit guys are cheeky and clever

[sage burrow]

Someone donated 9 buzz to me yesterdy, and I didn't even ask!!! 😄

[sage burrow]

Duc Haiten, and his models are amazing!!!!!

[sacred jewel]

OH YES!!

I hope for a Flux Finetune

[sacred jewel]

Yes, but, you;re a sweetheart 😄

[sage burrow]

How can civitai tell bots now? I guess if the ip addresses of likers aren't from 1st world countries?

[toxic bone]

https://civitai.com/articles/5829/elemental-extravaganza-contest-results these are the winners now

[sage burrow]

Aw ty 😄
(though it's probably more because I create nsfw lol

[toxic bone]

there's lots of signatures to detect bots. the obvious ones like ips but also just patterns in the data.

before June, civit just had ZERO defenses against it

[bitter hearth]

there are bot detection AI models now, funnily enough

[sacred jewel]

YEah sure, that too but, you're a sweetheart first LOL.

[bitter hearth]

personally I enjoy Civit a lot more now that I only interact with Civit like this: cd /workspace/ComfyUI/models/checkpoints && \ wget -O CinenautXLv2.safetensors "https://civitai.com/api/download/models/607422?type=Model&format=SafeTensor&size=full&fp=fp16&token= Put your token here " && \
instead of actually opening civit

[sage burrow]

How did I not know this?! (I wouldn't risk my main account however, so prob wouldn't have bothered)

[sacred jewel]

I wonder if it's NSFW if you work in the porn industry ... 🧐

[sage burrow]

Why thank you 😮

[young blade]

is there any workflows out there for gguf schnell working with loras? been gone for a few days, playing catchup

[sage burrow]

One of my titles on DA for images is "NSFW even if you work from home" ROFL

[bitter hearth]

also I can't open Civit ai without serious lag
even H100 server was not enough to open Civit ai

[sacred jewel]

Get two then... pfffttt... LOL

[toxic bone]

they control the database they likely jsust zerod the gift you gave

[bitter hearth]

ah ok thanks

[bitter hearth]

maybe I use wrong browser too as I use Brave browser, Waterfox or Ungoogled chromium

but these are not great

[sacred jewel]

*their

[sage burrow]

Their discord tends to have a "we are down again", or we are legged again" messages all day, most days, so it's probaby not your system.

[sacred jewel]

Why? You're misspelling so I am correcting you. You do not want to spell properly? That's fine, but, it is still wrong.

[bitter hearth]

i forgot to look at the civit discord
is there anything cool there

[sage burrow]

Chrome with 1000 tabs open here...

[sacred jewel]

You should. It's important... and if you don't, it makes you look uneducated and you do not want that. I assume...

[sage burrow]

I have a 340 page book that needs editing! 😉

[sacred jewel]

How would you feel if I kept misspelling your name as
Idec-Ia-Velepomdnt

[sage burrow]

They answer your questions...

But there are other better ones...

[sacred jewel]

It has some decent channels...

[bitter hearth]

ok I'm in

[toxic bone]

The scale of botting among the common users in the internet is crazy. A report came out this week. The largest DDoS attack EVER recorded. Against any service in the history of the internet. Do you think it's china vs the government? Do you think it's russia hitting twitter?

It's some kids who were mad they got banned from a popular minecraft server.

botnets are only getting worse. They're so easy to make now

when someone is telling you its not possible or they don't know what you're talking about, it's a smoke screen

[bitter hearth]

how long are you gonna continue telling us that you think Dice used bots? 🤔

[sacred jewel]

Kubrick LoRA

[toxic bone]

we'll see ¯_(ツ)_/¯

its just a general public advisory really. awareness for others to recognize bs when it's being obscurred from their vision

[bitter hearth]

even if you are correct
the worst case scenario is that
he used some bots to win some fake currency
on a website that he got banned from anyway
and then they took the fake currency back

[sage burrow]

So the workflow is, you have 4 accounts deep.... the 4th level account gets all the botting (because who cares if that goes), then that popular account likes all the 3rd account's stuff... and so on and so forth, up until your main account.

This was from 5 years ago though, no idea if they have all caught onto this now or not.

Not civitai specific, but it prob worked there also.

[bitter hearth]

so even if he did use bots
he got banned and lost the buzz anyway
so "justice" has been "restored"

so there's nothing left to talk about regarding that

[toxic bone]

sure there is. These kind of people are still using those "accomplishments" as clout. You don't hear Neal Armstrong bragging anymore do you?

wait.. he would

i mean the biker armstrong

LANCE yeah. Neal's name is the truly memorable man

[toxic bone]

woah

[sacred jewel]

*their

[toxic bone]

my facts are you won buzz contests before june. so...

pretty straight

[bitter hearth]

this isn't rly evidence though

I feel like its been 2 full days of this
and there has been no evidence yet

[toxic bone]

hmm. is this a fact i need to have straight, or are you being a command line operator who exposes vulnerabilities the fact i need to have straight? I'm not entirely sure these are compatible facts.

[toxic bone]

i'm not trying to convince you obviously. just spreading awareness for other people.

[bitter hearth]

but why do you feel the need to spread awareness

[toxic bone]

just a feeling i have

[bitter hearth]

are you aware there is only like 10 people in this discord

[toxic bone]

why do anyone have feels?

[bitter hearth]

and all 10 people have seen your message

[sacred jewel]

ARob LoRA

[toxic bone]

99% rule. lurkers. i lurked for a year or more

[bitter hearth]

but after a full year of lurking
why did you decide to make this your one contribution to the community

[sacred jewel]

https://tenor.com/view/malcolm-x-angry-mad-upset-hit-gif-5005267

10? 😄

[bitter hearth]

LOL

[sacred jewel]

😄

[bitter hearth]

a bit more than 10 ye

[sacred jewel]

I am just busting your chops...

[sage burrow]

my 20 alt discord accounts are offended!

[toxic bone]

i'm just back again. flux is going on. you make lots of assumptions. better to stick to facts like, there were zero defenses against botting on civit before june of this year, and their highest ranking images were all botted

[bitter hearth]

but this is not evidence that he, in particular, used bots

this is what they call circumstantial evidence

about the circumstances in general

[toxic bone]

so was all the home run records suddenly being set in the MLB ¯_(ツ)_/¯

[sacred jewel]

ARob LoRA is pretty good...

[bitter hearth]

but the issue is
because your evidence is only circumstantial at best, its indirect evidence, and what this means is that it can't be conclusively proved one way or another
so the conversation can never end

[toxic bone]

your choice to feel that way about the circumstances. but civit had zero bot guards before june. nobody who won contests won them legitimately and the admins had to go correct everything. it was a big deal on their discord server and discussions

i replied to him since he replied to me.

i think only 1 person told me they don't care. and you

[sacred jewel]

Lines LoRA.... it is good but too good. It ignores the prompt beyond the Trigger 😛 😛

[bitter hearth]

I mean
its been useful to learn that Civit had bot issues
cos I didn't know that
but I think I've learnt enough now

[sacred jewel]

Putting the trigger at the end of the prompt, helped

[toxic bone]

and knowing is half the battle!

[sacred jewel]

I am always impressed by a good LoRA ... wow...

[sage burrow]

I used it as a ref image 😄

[bitter hearth]

they are managing to get significantly different styles out of flux which is a good sign

[sacred jewel]

I posted yesterday or before that I was BLOWN AWAY by how many hundreds of LoRAs in such a short time for Flux compared to SD3 which has been out for longer... amazing effort by everyone.

[toxic bone]

people loras work good, but i like that line art one. the depths of creativty have barely been tapped

[bitter hearth]

yeah I consistently under-estimate flux

[toxic bone]

you know the file name of the original video is still visible right? https://www.youtube.com/watch?v=NimnGpBEEvY dude isn't an ai developer.

that aint you. he's professional musician with an entirely diff business

[sage burrow]

Honey, we need to go in for some eyebrow shaping 😉

[toxic bone]

that's from 13 years ago

[sage burrow]

I LOVE mirror ai stuff!!!!

[toxic bone]

https://escueladebajistas.com/ dude's official site

https://www.youtube.com/watch?v=K4dY2KsdpOc he's still posting

[desert garnet]

im actually Tony Mottola too what a coincidence

[sacred jewel]

And then there are some LoRAs that are not so great at times 😦

[toxic bone]

https://tenor.com/view/not-sure-hmm-thinking-gif-5730421

https://www.youtube.com/watch?v=K4dY2KsdpOc&lc=UgwpCsEDWKgbhX0fkZl4AaABAg i made a comment. just saying.

[sacred jewel]

Renaissance LoRA btw

[desert garnet]

u made that lora?

[toxic bone]

i dont think he realized that discord keeps the og file name. you'll see all his uploads will have generic filenames now

[desert garnet]

oh well maybe he can share what settings he used for flux on civitai trainer

[sacred jewel]

Nah, I only have the one I made which is the SpyWorld LoRA (it's so so)

I am not talented in this stuff yet 😄

[desert garnet]

oh yea same,i tried two flux style loras on civit but they come out deformed so gotta wait a few weeks for ppl to post more settings

[sacred jewel]

I found that the one version I did on Replicate came out better than the last two on Civit. I just wish I knew what and how all the controls do and mixed together.

Curating a few Rene Magritte images now to see if I can try my hand at a Magritte LoRA.

[toxic bone]

if you reply to my comment with a "yes I am he" then i will eat a loaf of poop

no shenanigans. just a simple reply

[desert garnet]

yea civitai doesnt have a lot of options on their trainer

[toxic bone]

lol. bruh. you were saying it was you

smh

"i was just pretending!" 😄

literally were bragging about your bass skills to someone

[sacred jewel]

I am so confused now. I thought the same thing...

[toxic bone]

@bitter hearth

[sacred jewel]

Wait, did you not post that video yesterday saying it was you? and you chatted about guitars and techniques? Not trying to imply anything... I am seriously confused if so though.

[toxic bone]

it's sort of the same topic ¯_(ツ)_/¯

[desert garnet]

hes legit i know this guy from under the bridge asking for coins he spent all his money on nfts now he moved to ai

[toxic bone]

yeah it's some other guy. the file name led me straight to the original youtube channel

[cunning lintel]

just undertrain your lora (like 1k steps) so it doesn't leanr a subject too well, get a nice looking image and pretend that's what you intended all along 🤡 flux picks things up so easily

[toxic bone]

apparantly he only ever posted it 8 times the last couple of days to fish me in and get me to stop talking about civit bots

[sage burrow]

Darnit, I wanted to compare the two methods, but had that replicate "compressed file not valid" error!

[sacred jewel]

Oh wait.. .the video was a file upload, not a link to YouTube, but you then went to the ACTUAL source youtube video? I think I get it now.

[desert garnet]

oh wait ill just upload fake photoshopped imgs and call ppl skill issue when they claim my lora doesnt work

[toxic bone]

correct

[toxic bone]

so many fakersons trying to fake clout

[desert garnet]

make sure to smash that mf like button subscribe and donate 💸

[toxic bone]

"i'm a victim guys. come support my channel"

[sage burrow]

I was asking the wrong questions; I had shown the sitting outside playing guitar vid to a very experienced guitar player, and asked if it was real or AI ROFL

[toxic bone]

"no that's just a youtube vid from 15 years ago"

[sacred jewel]

Some images do come out well... guess it is sensitive to prompting and subject...

[desert garnet]

yea same thing happens to me,sometimes they come out good but they all have deformed limbs or deformed stuff in background

[sacred jewel]

You mean like the nice looking horse above? 👀 😛

[sage burrow]

Whatever, 90% prob not possible to have as many hobbies as he claims, irrelevant imo. Does the person have good AI skills, does te person have good advice for people looking to learn. 99% of the internet is fake people. Whether it be makeovers (both genders) pre photoshop, rented fancy cars and houses for vid shoots...

Who cares I figure.

[desert garnet]

yes 😔

[sage burrow]

I challenge everyone to creating Dali like images with Flux!!!

[sacred jewel]

Salvador Dali? I tried the clocks, very hard for AI to do it right with all models I have tried it with. Have not tried it with Flux though.

"The Persistence of Memory" ... I have seen the original here in NYC many times. It is absolutely stunning in person.

Without LoRA trigger and With LoRA trigger word
This one is not so strong with the keyword applied

[sage burrow]

I got some copying of Dali with flux, there was MUCH help fro Claude!!!

[sage burrow]

Why did I not visit the art gallery when I was in NYC???!!!!

[toxic bone]

legendary paintings are not given justice on paper prints

[sacred jewel]

Which one? The good ones are The MEt, MoMA and Guggenheim... there are many others but MEt and MoMA are awesome.

[sage burrow]

Even on gallery grade paper? That was my trick for when I was selling (my own, not trad paintings) AI art in person....

[sacred jewel]

Just like my first time going to a live Broadway show. I was frozen stiff how different it is in real person live... I had only seen Movies before seeign live shows...

[toxic bone]

i mean like the old masters of oil painting like dali and da vinci

[sage burrow]

I was going to go to the MET, but something came up (probably actually slept that day lol)

[toxic bone]

or many more

[sacred jewel]

YES... the paints and inks used and the medium and the level of skill was amazing. Rembrandt blows me away.

[toxic bone]

in person. looking at the same thing the artist touched. thats vibe.

i saw the fake mona lisa though and i still felt energy off it

[toxic bone]

maybe is more about the layers of the inks and paints and oils

[sacred jewel]

I could only get within 50 feet of the Mona Lisa when I went to Le Louvre... it was MOBBED. I was lucky to see it here in The Met when it came around ... got 10 feet closer LOL

[toxic bone]

Yea i saw the replica when it was on world tour. wasn't a mob at all.

the one thye made when it was stolen and they didn't tell anyone

[sage burrow]

I got to see several in Amsterdam as a kid. My family thought it would be a fun trip. I was way too young to appreciate it 😦

[sacred jewel]

Prompt : Salvador Dali's "The Persistence of Memory"

[toxic bone]

think a drooping clock lora would work?

[sacred jewel]

It might... a single image LoRA? hmmm...

That's be a great prank.

a LoRA that ALWAYS generates the same image 😄

[sage burrow]

I used far too long of a prompt, created by Claude #🆕｜sd3 message

[sage burrow]

Probably, they seem to work really well with Flux

[sage burrow]

But with a few hundred tags? 😉

[sacred jewel]

CRUEL... I love it 😛

[sage burrow]

Picasso anyone? 😄

[bitter hearth]

can forge run that high power dev model or only comfy?

auto cant run any though right

i think i might do better with forge. is it easy to add thing like the upscaler? i like the image i get from comfy but it doesnt have the crispness i see in some

noice thanks

you da man

[sage burrow]

Dali || #🆕｜sd3 message ||

[sacred jewel]

Mix of Bosch and Where is Waldo LoRA

[toxic bone]

oh look the file names are varieties of asgjha.jpg . just use power renamer

[toxic bone]

rename your whole folder of scraped images with 3 clicks

to non obvious renames

why do it manually. work smarter not harder

huh? i dont know what you mean

just giving advice man. you're taking notes. it's a good tip. powertools dude

[sage burrow]

Nice mix ROFL

I'll get bored of using your work as a ref image soon @sacred jewel 😄

[sacred jewel]

LOL

What the heck are the things on their heads? 🤭

[craggy crest]

don't start. windows powertools are very useful. must you try to cause a fight all the time?

i did. there's no reason for you to be argumentative about everything, is there?

[sage burrow]

I made something for my YT 😄

I cheated and used gliff for the text

[toxic bone]

yeah i'm just recommending you use powertools power rename to rename your images so i don't find the real sources you're ripping off anymore. its a really good tool and makes renaming a lot of files so simple

nothing to get so bent over

[tame berry]

imagen generar del futuro

[sage burrow]

Their cult members tinfoil hats I figure

[tame berry]

how generate images, command ?

[sage burrow]

Oh?! I didn't know wabout this, where is it found in windows? <-- too cheap to pay for adobe suite

[toxic bone]

official addons for windows called "power tools"

[desert garnet]

we already have a tool here dont think we need second one 🫠

[sage burrow]

I confess to using windows image efitor more often than I should, so that sounds good to me.

[sage burrow]

https://discord.com/channels/1002292111942635562/1237461679286128730

Or serach for huggingface FLUX dev

[toxic bone]

you need the right tools though. the plunger don't work for everything

[toxic bone]

for just renaming files?

i use gallery-dl to rip entire website galleries. that's more command line tool but its a good one if you can weild it

https://learn.microsoft.com/en-us/windows/powertoys/ oops it's "power toys" my bad

https://github.com/mikf/gallery-dl

[sage burrow]

Microsoft uses github?! https://github.com/microsoft/PowerToys/releases/tag/v0.83.0

[toxic bone]

microsoft literally owns github

[sage burrow]

For actually editing my images if it's just a small erase stuff thing.
I use gimp for the more complex stuff

[toxic bone]

i dont think powertoys will help there. potentially could batch some edits with something though

[sage burrow]

I feel more violated by that then Dice's guitar video even!

[sage burrow]

The image rename though, I really need that!!!!

[toxic bone]

they've done good by it. i was young when it first happened but they haven't ruined it

[sage burrow]

When creating loras, some things base things on the actual name of the files

[sacred jewel]

I love BRU (Bulk Rename Utility) which is extremely powerful...

[sacred jewel]

They OWN GitHub

@bitter hearth do you have any Flux LoRAs you've trained that I can try/test? Thanks... I've been testing as many LoRAs as possible the last few days.

[toxic bone]

good tip. yeah if you're downloading a lot of files that you intend to show are your own and act like you made them, this would be another good tool to rename them all at once with ease.

[craggy crest]

google windows power tools

[toxic bone]

i made that mistake too. "Power Toys". and i linked it for her

[craggy crest]

it shows up with a power tools search though

[toxic bone]

under all the power tools. i searched it and was like "oh . yeah thats expected"

[sacred jewel]

No worries, thanks.

[sacred jewel]

What does the Model Sampling Flux node (or the settings) do? How do they work "technically" ?

I've been leaving it at whatever defaults Matteo has in his node (or Comfy in his node)

[craggy crest]

i think matteo has a revised workflow that just uses a normal ksampler instead of that node

[sacred jewel]

Gotcha... I use his other node which I love because it makes it easy for me to make GRIDS and comparisons... it also has the two fields for Sampling shifting stuff...

[toxic bone]

i can't keep this guilt in anylonger.

i never wrote a comment on that guy's youtube video asking him if he was an ai artist

i just said i did

[sacred jewel]

You killed J. R. ?

[toxic bone]

few had to get that off chest . relief.

[toxic bone]

was maggy

[sacred jewel]

https://tenor.com/view/jr-ewing-dallas-jockey-larry-hagman-gif-10340875

[toxic bone]

https://tenor.com/view/maggie-simpson-in-math-gif-19543231

[bitter hearth]

but why furry

[sage burrow]

Biggest 2024 betrayal lol
(OK far berore that but...)

[sage burrow]

I was using an img to img and well, 99% of my images....

[sacred jewel]

😄

I think they contribute more to Open Source than the nearest 5 other contribuitors combined... or some crazy statistic I once read about the reason not to hate on Microsoft for buying gitHub. 🤷‍♂️

[sacred jewel]

Or as Spock would say, 98.76110134%

WARNING:

Do NOT, I repeat DO NOT, send a batch of 4 into the queue using Flux.1 Dev plus two LoRAs with a latent size of 1920x1080

YOU HAVE BEEN WARNED

[PSA] 😭

(now I am curious what the output will be so I don;t want to stop it... but it is CRAWLING)

[bitter hearth]

one hour later

[sacred jewel]

hours..... 😿

[craggy crest]

watches gpu catch on fire

[sage burrow]

Those sorts of things you set to run beore going to sleep for the night 😄

[toxic bone]

"Never Tell Me The Odds!" - James T Kirk

hard to roleplay when your die so smooth 🤡

[sacred jewel]

EXCEPT: I did it accidentally at the WRONG time 😦

[sacred jewel]

Yup... it went up to 65 degrees... highest I've seen in a while... usually goes to 55 or 56 deg tops...

[craggy crest]

65 isn't bad though. just don't let it get to 120

[sacred jewel]

These are from the 3090 which I had to fire up to keep myself entertained 😄

First time Flux Dev on the 3090 for me... 5 secs per iteration... not too bad.

[sage burrow]

I've definitely done that before. A 2 hour 1 image render yesterday LOLOL

[sacred jewel]

Mine is Air Cooled PNY model... it does fairly well with cooling... MUCH better than the older 3090 which used to cook the memory modules from the slightest use

[sacred jewel]

All that waiting and all images are JUNK .... ugh

I guess MDiT does not like going beyond the trained dimensions?

[pseudo owl]

flux has no idea what a watermark or low quality is lol

[sacred jewel]

and finally the fourth one is done

hard lesson learned

[sage burrow]

LOL

[craggy crest]

wrong scheduler/sampler

[sacred jewel]

I think it was the blank prompt, 1920x1080 latent and the Guidance of 2.5. I used my usual Euler/Beta sampler/scheduler combo :/

[toxic bone]

you sure? i've prompted for both and got results before

[craggy crest]

blank shouldn't have giving you that though, that's what I get with the wrong sampler/scheduler - but maybe. flux doesn't like really short prompts very well

[toxic bone]

the aesthetic guidance rating is almost like a prompt itself. if it's high then consider it like promting for the most ultimate photo shoot for vanity fair

[pseudo owl]

so far i don't see any watermark images and the low quality images are pretty high quality. maybe after a few more tries.

yeah another try, this is what i get for watermark

[toxic bone]

i dont know if i'm crazy or not but i have found that sometimes an img 2 iimg wiht high denoise will inform the new image a lot too. almost like it's an ip adapter at some denoise levels

[pseudo owl]

its an actual water mark lol.

[craggy crest]

coffee table ring

[toxic bone]

pollocks coffee ring

[pseudo owl]

what i get for low quality, not sure how its related to low quality at all.

[toxic bone]

just looks like he's in a low mood lol.

its in there. i promise. its just tricky cause of the aesthetic guidance. its hard to work around. i'm not getting watermark gems either tbh.

[sacred jewel]

So maybe the size of the latent?

[pseudo owl]

I do see watermarks sometime in completely different prompts but extremely rarely but it doesn't really seem to know what watermark is.

[toxic bone]

i got an llm to describe photographer samples you'd have watermarks on. it knows.

clip L might know better than t5 . consider that sd2 knows watermarks. that's clip l

[sage burrow]

How on earth did you get the half and half?

[sacred jewel]

Prompt + Wraith BW LoRA 😄

[hexed dirge]

my lora

[sacred jewel]

peruvian arpillera, yayoi kusama, vasily vereshchagin, cellular automata

stolen from MidJourney daily theme

[sacred jewel]

Only the best 😄

[sage burrow]

You don't use my loras @sacred jewel ? 😦
ROFL kidding 😄

[toxic bone]

you can do stuff like inform the composition in your prompt using the t5 encoder. describing the sky is drawn with red duotone and the landscape is greyscale

"top left of image is .... bottom right of image is ....

best watermark image i got lol but it's reallyw ashing out the actual images. this was a photographer proof sampler prompt

[sage burrow]

Horror loras! ❤️

[hexed dirge]

I love them too

[sacred jewel]

and with the LoRA enabled

[sage burrow]

What an awesome watermark idea!

[sacred jewel]

If Only I could find them... what's the AuthorName? Is it under same as here?

[sage burrow]

Here I thought I'd have to make a few hundred images using glif (https://glif.app/@LadyLalita/glifs/cm02f7rpk001168ybpj4udw1t), then create a lora....

[toxic bone]

https://youtu.be/oxXpB9pSETo

[sage burrow]

I'll message you, I don't think I created a single sfw one, and my public images are very NSFW. I was actually just teasing you, they are definitely a very narrow niche that most don't go for lol

[sacred jewel]

I am no prude... 😛

[sage burrow]

Make him a reptilian 😄

[toxic bone]

says the lil Pomeranian pretending to be the big dawg in the yard

this is a discord server. i'm just sitting here. going no places dude. dial down your delusions of grandeur

[sage burrow]

For the sake of the Discord channel lol
My un is oipteaapdoce

[toxic bone]

try not to back him up an be his ally. just a suggstion. take it how you will.

stephen king hates lawn mower man hey? the source material was better

[sage burrow]

Not his ally; he's going to have to produce that app/page for model creation first 😄

tbh, more devil's advocate on here...

[desert garnet]

he coded and merged two 4090s together show some respeck

[toxic bone]

is he mine or yours? get your facts straight

how do you expose a bullshitter? just wait they'll fuck up

[desert garnet]

only pair here is the one this fraudster lacks

[sacred jewel]

Steampunk Apparel LoRA

[toxic bone]

As a hetero man, i'll come out and say it. Gay men are more macho. Us hetties, we like those femine girls. Gay guys get down with big butch gay guys. They'd probably choose the bear too (feminism memes)

try to tell me that George Michael is less manly than you. Try

[desert garnet]

dem he switched to using gay as an insult,what a menace watch out

[toxic bone]

Most hetties struggle wiht this concept, but you gotta realize how gay guys get down. with other guys. that's some 300 spartans macho level shit.

kinda sounds like a white guy actin like snoop

not sure how it's sd3

[desert garnet]

he prob has a tattoo of a 4090 on his dick

[sage burrow]

He could have asked Claude how to do it, and then borrowed the video, or used his work computer for the vid 😉

Or not LOL

[sacred jewel]

I LOVE George Michael #NoHomo

[toxic bone]

yeh seriously theres nothing wrong with GM or Wham or Wham! or nothing

i always laugh when people try to call me gay. I'm secure enough with who i am to know the reality of the situation. Gay guys are tougher.

[sacred jewel]

And SUPER talented

[toxic bone]

faith? or whispers? i vibe whispers. https://www.youtube.com/watch?v=izGwDsrQ1eQ

[sage burrow]

You are making it really hard for me not to post some art in reply! 😉

[toxic bone]

i'm just speaking truths. Gay men know what being a man is about more than hetero men. Get a gay friend. They could teach you things.

homophobes clam right up whenever i pull this routine. its fun to watch

[sage burrow]

I'm keeping the sound muted on y'all's vids!!!

[sacred jewel]

Any!

[toxic bone]

oh geez he's relating himself to joker

[desert garnet]

well he already wears the makeup

[toxic bone]

state of the art lipsyncs are way better. i got one saved somewhere. let me find it.

[sage burrow]

Lip synching ai is awesome! 🙂 https://wixmp-ed30a86b8c4ca887773594c2.wixmp.com/v/mp4/33f80a36-07e4-4ce5-bf30-eecb5a275ae2/dhwb5iq-08159347-529f-4f49-968f-c98c1e7a1ea5.VideoQualities.res_480p.c0d0dd7b818b48ed827ec48da3509abc.mp4

[toxic bone]

not mine. from an emotive lip sync model from a couple months back

[toxic bone]

it messes the bug face a bit but i love the physical body language it puts to it

[sage burrow]

It came out with a new feature that it did animals one day, so I had to try it! There's a super cute video of different cats with several frames all singing the same song 😄

Hedra? That voice is awesome!!!!

[sacred jewel]

Steampunk Apparel + Wraith_BW LoRAs at 50% each

[pseudo owl]

thats ai?

[sacred jewel]

Blur Future and Wraith LoRA mix 50/50

[pseudo owl]

what are you using?

[toxic bone]

i've seen that one!

[sage burrow]

Flux pro 🙂
no lora, Claude helping

[sacred jewel]

william bouguereau LoRA

[sage burrow]

flux dev

[toxic bone]

riiight. you created that. my guess, it's an actual video that you didn't make, and you just face swapped with dialog you wrote. Her body doesn't match the words and tone. big tell.

and you're a bass guitarist too

Cheap Trick was a great band, but these kinda cheap tricks only work if they're matched with real craft and skill.

https://www.youtube.com/watch?v=-qgpewMCVjs

didnt i didnt i didnt i see you cryin ooohwoah didnt i didnt i didnt i see you cryyying.

[sage burrow]

I'm sorry, I could not resist, I tried 😄 lol

[toxic bone]

we'd be more masculine if we got it on and just hate * each other hey?

like if we just got all UFC oiled up and went at it like spartans. that'd be manlier than etiher of us are being i bet.

[desert garnet]

stay off the koolaid and bleach

[toxic bone]

lets just struggle cuddle it out

[desert garnet]

https://tenor.com/view/_society_-_funnyguy_-funny-guy-the-funny-guy-gif-21085029

[toxic bone]

see, i am praising gay culture. you're trying to use it as some kind of stab. This is due to me being secure in who i am, and you being insecure and fear your inner gay man

i'm able to recognize how macho an tough gay dudes are. you think it's an insult

Dont even try to tell me that you're manlier than George Michaels. No man can.

[sage burrow]

Flux wouldn't render a crack pipe for me 😦

[toxic bone]

Glass orb connected to a glass tube with a flame under the orb?

piece of chalk

[sage burrow]

Thank you for my first video making project idea! 😄

[toxic bone]

open discussions about man on man love really makes the insecure types get all squeemish.

everybody don juan till the dongers come out

closeup photograph of an item made of glass. fingers holding the stem of a glass tube. at the bottom end of the tube is a glass orb. at the top end of the glass tube are inhaling lips. inside the glass orb is a small piece of chalk. a bic lighter flame is beneath the glass orb kind of works.

scene even looks a little cracked out

[toxic bone]

freebasing something crazy in there. this chick got some FINGERS

its fine it works

[desert garnet]

needs more fatness and clown makeup

[sage burrow]

I tried to prompt a tide pod in a crack pipe, but it didn't go so well 😦

[sage burrow]

THat image definitely works

That's seriously awesome

[sage burrow]

Trying to remember that stock video DL site I might still have credits left on!

[sage burrow]

Oh wait, I took that as an image depicting people who soke crack are risking/gambling. I didn't realize it's probably an insult... Clearly less multitasking needed onmy part or something.

Teasing is cool, being angry is fine, but personal insults not so much 😦

regardless of the person being insulted

[toxic bone]

no totally its non incriminating at all it's all about taking risks yeah

why are you telling me to be the bigger man after he tried to turn homesexuality into an insult? girl....

[sage burrow]

I guess it's more obvious when in image form...

(to me)

[toxic bone]

you still think he's the guy in the bass video ? k.

[sage burrow]

LOL

[toxic bone]

"be nice to him he's attrractive and plays song"

[sage burrow]

?

[toxic bone]

you didn't catch that one did you?

[sage burrow]

That guy is way too young for me!

hm?

[craggy crest]

i do know that flux has a size sweet spot, which is rather odd. it's 672x1024 or 1024x672 - someone in this discord figured that out right after it was released

[sage burrow]

Someone needs to start posting images of cute bunnies or something

[toxic bone]

he called you a crackhead and you're defending his honor so i just figured you had a big ol crush on the guy.

girls love the badboy type if they're cute enough. just fax of life.

[desert garnet]

needs more muscular bunnies

[sage burrow]

Nah, I just think worrying about crap online isn't important.
I'll admit I looked into his vids a bit, but don't care really.
Now if my computer ends up with viri on it suddenly I might be throwing some accusations tho

[sage burrow]

Pretty cool, however it was suppposed to have 5 colours (prob a skill issue though)

[craggy crest]

AIs can't count. specify the colors

try "a sign saying "I'm an A.I." in the colors of red green yellow orange blue"

[sage burrow]

I created several tarot decks with AI, it was HARD! Would have been nice if the models were more advanced back then, but it seems they stil have trouble counting 8 swords etc. 😦

[toxic bone]

"From left to right" and other compositional language like that. t5 is really good at building a composed scene

[craggy crest]

they haven't had any training in actual counting

A.I. image gen = mad scientist with amnesia

[errant dust]

So after prepping 60 images for a LoRA, I am asking Claude to label them. would the following label be deemed sufficient? Black and white ink illustration with detailed cross-hatching and fine linework. Desert oasis scene with palm trees and a large tent. Foreground shows a veiled woman on horseback, wearing revealing robes. Her horse is intricately drawn. Background depicts various robed figures, including a muscular shirtless man and a long-haired individual. Some figures sit near baskets and pots. The scene suggests a Middle Eastern or fantasy setting with a mix of characters in traditional and revealing attire. The composition is dynamic with strong contrasts and meticulous detailing of textures fabrics and anatomy.

[craggy crest]

looks accurate with the exception that you might want to add some specific terms like "palm trees, horse, tent cloth'

[sage burrow]

it cheated!

[errant dust]

It does actually say those things

[sage burrow]

THat's why we get along 😄

[errant dust]

I get it. And I can ask for more detail, but have no idea how much is needed for a decent LoRA. My query to Claude was this: (Thanks to Clown for the last phrase idea): Please describe this image including the art style, the background elements, the people if any, what they are wearing, and any other pertinent aspects as if this were for a prompt to create this image. Do not use commas between coordinate adjectives, or any adjectives, though it's okay to use commas for separating concepts.

[sage burrow]

Seriously? Straight up insults, not just teasing?

[desert garnet]

id say you want to add not too many words but not too little either

[errant dust]

Well aware. My first attempt at asking Claude gave me a full page of description with bullet points! lol

[sage burrow]

Guess who forgot to take the word try out? lol

[desert garnet]

i just used wd 1.4 tagger and called it a day

[sage burrow]

@dusky thistle sooo how long or short of prompts are best for lora making? 😄 Also tags?

[errant dust]

I don't want the label to be the weakpoint that drags the LoRA down. As to the terms. The truth is that I want it to learn the style, not particular objects. SO I have 60 curated images in that vein and by the same exact artist

[sage burrow]

is that an SD workflow thing, or external?

[errant dust]

@dusky thistle noted that the quality of the label must not be underestimated though, and it was his suggestion to add that comment about the use of commas

[toxic bone]

i suppose becky was just laughing at me about me being the crackhead until she told me to leave you alone. good luck with this clown. i dont know why y'all buy into him so hard. milli vanelli over here.

[sage burrow]

I made 3 loras with flux. The first one used 64 images, the 2nd one used around 200, and the 3rd one used 500. The first one turned out the best!

[desert garnet]

no its in kohya/one trainer u use it to tag your images using booru prompts

[errant dust]

That is a bit of a relief since I have 60

and it was a ridiculous amount of work to clean up and remove all the damn text bubbles and boxes

[desert garnet]

better to just delete the bad ones even if u left with like 25

[errant dust]

There are no bad ones

[sage burrow]

I've used the (cheater) civitai version of it 😉

[errant dust]

Example of Before and After:

[desert garnet]

removing those bubbles looks like a lot of work

[errant dust]

You can believe it.

[sage burrow]

This is more my style 😄

[sage burrow]

As long as they are the same theme and/or position of the character

[errant dust]

All are the same theme and artist. Though the backgrounds and whatnot are diverse

[toxic bone]

i'd use photoshops content aware toolkit if i approached it

[errant dust]

Photoshop Firefly is the way to go

[toxic bone]

costs more money. is that what you used?

[errant dust]

It does, and I do. But I live in Brazil so regional pricing is a thing. As in $5.50/month

[toxic bone]

good deal! yeah its a nice tool can't deny it

[sage burrow]

I would hope so! Er nm, remembering those one word prompts.
Claude can help greatly!

[errant dust]

Becky does not want to be banned for life

lol

[toxic bone]

Flux was trained on instruct datasets too like llms . serperating the image out into title components works so well. anyone that plays with LLMs will learn how well instruct training works

[sage burrow]

ROFL ROFL ROFL ROFL

[toxic bone]

t5's really good at self attention and building words like that into more detail for the context

eh. not everyone needs absolute control. letting the ai do the creative lifting isn't so bad

take the word beautiful out. it'll change.

[errant dust]

It will be creative regardless of how detailed your prompt is. You are not drawing yourself, nor is it taking a photo, and even the items you specify can vary in shape, color, and more.

[sage burrow]

I'm still amazed at how those image describers can describe one down to the last detail!

[errant dust]

You must consider that they are the egg and the image generation is the chicken

[toxic bone]

a beautiful waterfall
a vibrant waterfall

same seed. simple prompts. easy to test. beautiful even hits harder when you throw it into a sentence in a way that recontextualizes it

• beautiful even hits harder when you throw it into a sentence in a way that recontextualizes it
• even hits harder when you throw it into a sentence in a way that recontextualizes it

what would that test?

it's a button to push the modle with ¯_(ツ)_/¯

[errant dust]

That looks painful

Unless he is just sitting on his boots, and those aren't actually his feet. 🙂

Agreed but at least he doesn't look like an involuntary posterchild for the disabled

Aside from the hand, it has a strong Norman Rockwell vibe going for it.

[toxic bone]

words work. just look at how this one changed the image

[errant dust]

Well, apparently Imagen 3 was not neutered on nudity, since without even requesting or wanting it, 3 of 4 images came with a bare-breasted chick

I was simply testing the labels Claude gave me

[desert garnet]

damn greg rutkowski that bastard

[toxic bone]

was actually pretty tough to get it to do that gag

[desert garnet]

https://tenor.com/view/mike-wazowski-mike-pog-pog-gif-154717314566640477

[toxic bone]

haha dice went to the mods and had all the stuff exposing that video as not his deleted. convenient.

[dusky thistle]

Depends on the model

[sacred jewel]

Skillzzzzzzz

[sacred jewel]

Works too...

[charred valve]

a cat

[desert garnet]

yea i tested it and works

but i have to up lora strength to 2.5 idk why

[bitter hearth]

I turned the fish into a ball

sorry fish

[hallow lion]

Ballz.

[bitter hearth]

does anyone know nice general SDXL loras that work for lots of different checkpoint models?

I got Detail Tweaker XL

[bitter hearth]

amazing 💯 im grabbing this and upscaling it for my desktop wallpaper

thanks will try this

[bitter hearth]

Pony:
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, rating_explicit,

Anime Neg:
blurry, muted colors, cartoonish, lack of detail, unrealistic anatomy, generic background, harsh lighting, watermark.

that with actual prompt feminine figure, lush floral garden, tight-fitting tank tops, distressed denim, bare thighs.

the negative prompt had good effect on the image, especially the color tone and image texture

can remove the cartoonish bit for anime

think i'm overdoing it with negative prompts?
ugly, muscular, incorrect proportions, deformed figures, monotone, blurry, pixelated, low resolution, muted colors, desaturated, harsh lighting, overexposed, cartoonish, noise, watermark.

think i'm overdoing it with negative prompts?ye cos the optimal negative prompt is empty

[bitter hearth]

there's a paper where they show using a very brief negative for only a few timesteps

instead of a negative the whole way

I wish I could remember the name of this paper cos I keep referencing it

[bitter hearth]

i see some good differencs tho

i dont have technical explanation however

[bitter hearth]

what is this.. schnell or sdxl?

[sage burrow]

Def flux, I don't recall SDXL having that slightly out of printer ink look. Am I right? 😄

Do Flux and do crappy art also.
However, it may not be crappy, everytime I go to delete an image from DA because I think it's not that great, I notice it has 17K views. Clearly people have poor taste in art lololol

[bitter hearth]

I noticed SD 1.5 does split images like this a lot

I remember that from over a year ago

sometimes SD 1.5 does funny text too, more than SDXL does

[bitter hearth]

actually that's SDXL done with dreamshaper turbo

[sage burrow]

Dammit SDXL, how could you?! lol 😄

[bitter hearth]

i was taking out my frustration over flux since flux always lags my pc

loading a 17 gb model into ram with 32gb in system memory and doing other tasks, like watching youtube or listening to spotify doesnt really do good to the system

[sage burrow]

I have been tempted to run a series of 100 sd 1.5 dynamic prompts just for the satisfaction of it LOL

[bitter hearth]

yeah... so for the sake of satisfaction i went back to schnell NF4 checkpoint

I did about 1,000 SD 1.5 gens this morning I rly like SD 1.5 now

[sage burrow]

It is seriously awesome!

[bitter hearth]

its got rly pretty designs, nicer than SDXL I think

[sage burrow]

Guess which images sell best on DA of all the models?

[bitter hearth]

pony?

[sage burrow]

1.5

[bitter hearth]

wow ok

[sage burrow]

I'm not even taking only my own account into consideration either

I've asked around since I was curious

Though Flux is fun since no one else is really doing flux nsfw on there yet

[bitter hearth]

I love the architecture you get from

SD 1.5

its way more chaotic than SDXL

sometimes it makes less sense though but I kinda like that

[sage burrow]

What do you think of 1.5, then run through Flux or other via img2img or or in the same workflow?

[bitter hearth]

yeah sounds good

my SD 1.5 images are refined with SDXL HelloWorld

you could try SD 3 refiner if you are willing to brave the Skill Issues

[sage burrow]

LOL

[bitter hearth]

might be ok if the denoise is lower so that SD 3 has less chance to mess it up

[sage burrow]

"hands behind back" Solved 😄

[bitter hearth]

literally yeah

get them to hide their hands somehow

and then do facefix

and its ok

[sage burrow]

Too bad we can't run 4b or 8b on our own systems, yet

[bitter hearth]

yeah I would love local 8B

[sage burrow]

Well I don't create humans, so all is fine 😄

[bitter hearth]

same TBH

[sage burrow]

So uhm, what's that thing on the bottom? I don't recall the AK having a, uhm, double barrel? lol Still way better than 1.5 or sdxl guns I'll admit!

[bitter hearth]

guns always have something slightly funny yeah

[sage burrow]

Remember those 1.5 ones? LOL

One of my most fave things about flux is that it can do bows!

[bitter hearth]

ah yeah bows are tricky cos of the strings

one of the nicer wheel upgrades I have seen

this one was a weird one

cannot decide if its a good R2 or a bad one

got one of the super front foot ones today too

[sage burrow]

It's R2D2's personal assistant 😉

[bitter hearth]

I always wondered if the weird R2s are in the training data or if I am just perturbing the model

[sage burrow]

I ran the code to perturb sd3 :D. But now I see mentiones of perturbing other models?

[bitter hearth]

look at this comparison here, NF4 version of the checkpoint has lot less weight on the system but produces pretty much exact quality as Dev https://www.reddit.com/r/StableDiffusion/comments/1eq63sj/just_do_a_some_image_flux_dev_fp8_and_nf4_fp8_on/

perturb just means "mess up"

[sage burrow]

I always liked the ones on the right better, except for a couple of the photo ones. Interesting

[sage burrow]

So how does one perturb the entire Flux, or SDXL model?

[bitter hearth]

well ofc, if you look closely at the detail, nf4 is producing better results

common SDXL perturbations I like are using FreeU/SAG/PAG/SEG/CADS/tonemap nodes at crazy settings

also can mess it up with noise injections or using way too low CFG

just stuff that messes the image up in weird ways

[sage burrow]

Is anyone in here running both Dev and Nf4? I mean who has a fast system, so doing comparison images doesn't take 20 minutes? lol

[errant dust]

This is also completely outdated since it preceded the GGUF releases.

[sage burrow]

hmmm

[errant dust]

it is almost a three week old post, while Flux has been out 4 weeks

[sage burrow]

Does GGUF work with Dev loras?

[errant dust]

Yes

[sage burrow]

Oh alright, I guess I'll figure out how to install GGUF 😄

[bitter hearth]

how is that outdated?

in case you dont know gguf came out after fp8 checkpoint

and fp8 version is pretty much the black forest lab standards i believe

[errant dust]

It definitely has a few steps involved, but if you follow them, they are painless and one-offs.

Let's see. the post is 17 days old. GGUF is less

[bitter hearth]

you are not getting what im saying

[errant dust]

and th fp8 is NOT standard release

[bitter hearth]

gguf came after fp8, also fp8 is the BFL standards, and the NF4 which came out later and before gguf, its producing results just like fp8 model

[errant dust]

it is COmfy

[bitter hearth]

also i have to mention again, gguf has some issues with some samplers that i've noticed

[errant dust]

Yes, issues the fp8 has also

[sage burrow]

anyone have a gguf workflow? 😄

[bitter hearth]

let me do a quick check....

[errant dust]

fp8 is inherently lower quality than any of the GGUF models or even NF4, simply because it is fp8

[bitter hearth]

gguf uses the same workflow as unet loaders you will find at https://comfyanonymous.github.io/ComfyUI_examples/flux/ just need to change the loader node to gguf version of the node

[bitter hearth]

i dont think so, not by the look of the results

[dry wave]

what? where did you hear that?

[bitter hearth]

the fp8 single checkpoint from that link i just mentioned produces better result than gguf

[dry wave]

fp8 is usually the worst approximation. It works sufficiently well for flux, though

[errant dust]

It is simple fact. NF4 and GGUF are not uniformly encoded with fp8. fp8 is literally the lowest standard. GGUF has mixed fp16 and even higher

[bitter hearth]

hang on i think i meant to say the model i used from black forest labs has better results than gguf

[dry wave]

bfl never released any fp8 model

[sacred jewel]

https://tenor.com/view/take-secretarian-xirtus-truthfairytales-winning-gif-11846192109590071243

[bitter hearth]

i tested some of the samplers between original schnell and gguf

[errant dust]

that would be obvious. Everything else is a compromise. The issue is simply how much

[dry wave]

somebody made a pull request on the huggingface repo adding an fp8 version

but that was not bfl

[bitter hearth]

great for wallpaper, ty 🙂

[errant dust]

The best compromise is GGUF Q8_0

[sterile pendant]

Fp16 = gguf q8_0 > fp8 > nf4.

[errant dust]

as shown in the charts

[bitter hearth]

right. thats not the bfl

[sterile pendant]

In terms of precision

[sacred jewel]

It's a bullet detour in case the main path is backed up 😛

[sage burrow]

So all the ones I've been using for Dev already?

[sage burrow]

Seems safe 😄

[errant dust]

https://huggingface.co/city96/FLUX.1-dev-gguf?show_file_info=flux1-dev-Q8_0.gguf

[sacred jewel]

totally 👀

[errant dust]

That shows the exact quality of encoding at every stage

[errant dust]

and as @sterile pendant has been so kind to share regularly:

[bitter hearth]

there is a single checkpoint version and unet version with dual clip, they are different in their workflows ofc, but ether of which you use, you'd have to simply change the loader nodes specific to if you are using gguf or nf4

[bitter hearth]

btw subtleone this is nf4 dev

redhead woman lounging on sun-kissed sandy beach, wearing a bright red and white striped bikini bottom paired with a matching red visor, her fiery locks tied back in a loose ponytail, surrounded by crashing waves and a lear blue sky.

[bitter hearth]

i used 15 steps

[errant dust]

I know it quite well ans have used it

[errant dust]

The numbers are the numbers. Showing some sample here and there won't chnage the facts

[bitter hearth]

thats; blind science

[bitter hearth]

you need to come up with an approximation and you have to also assess the actual results

[bitter hearth]

as far as I know there has not been a paper yet
comparing quants like this
so its more of a case of waiting for paper

[bitter hearth]

need the benchmarks of FID and Clipscore on 30k images of MS Coco, Imagenet or Lsuns-bedrooms

[bitter hearth]

💯

[bitter hearth]

good gracious.... 🔥

[errant dust]

Look, you are obviously welcome to use what you prefer, and by all means do so. But none of that will override the fact that NF4 and FP8 are way more lossy compared to the pure fp16 than the GGUF, which is actually a mix of fp16 and 'almost fp16'

[bitter hearth]

its an issue when people test stuff on like 10 images
when the industry standard for comparisons is more like 30k images

with 10 images how can we rly know anything either way

[sacred jewel]

Big ball and many small semi-balls 🤭

[errant dust]

Both fp8 and NF4 are very fast, but inherently worse quality adaptation of the pure fp16

[bitter hearth]

they should just rename the moon as Big Ball yeah

[errant dust]

Look, this is a plain reality for people in general. I developed and tested chess neural networks. Measuring the quality of one model over another in small differences requires thousands of games. And this is way easier than measuring some images, as it is simple results. YET, I have had arguments with people who tell me they don't need a thousand games, just a half dozen is enough for them to 'see' how good it is or not, and that they have decades of experience a thousand games cannot override. The fact is that even the WORLD CHESS CHAMPION would be too weak to make that distinction between two engines, which can crush him blindfolded and two arms tied behind its back from a cell phone. So imagine some beginner enthusiast who just runs them from his computer.

[sage burrow]

Anyone have a workflow for GGUF?

It's not liking my loras or something

or the positive prompt clip loader, I think

[bitter hearth]

ye I can imagine people saying that
that they can judge a 3000+ elo chess engine from 6 games only

[sage burrow]

I really like my current workflow though 😦 Just swapping out for GGUF versions isn't working darnit!

[bitter hearth]

this one is really nice, I love the blurred ship

[errant dust]

They cannot judge games even 200 Elo above them, never mind 1500 Elo

[sacred jewel]

Thank you... totally Ollama / Llama3 generated 🤭

[sage burrow]

Here's what I'm trying to convert to GGUF

[bitter hearth]

ye I use chatgpt for prompts mostly

[errant dust]

I only know that to describe images for labeling, ChatGPT is a massive fail compared to Claude

It constantly gets it wrong. Not worse: wrong.

[bitter hearth]

claude is better for language stuff yeah

[errant dust]

Guy is sitting on a horse. He is central figure to be clear.

ChatGPT: Man standing next to a horse
Claude: Man sitting atop a horse

NO, not language. ChatGPT is blind

[bitter hearth]

GPT is stronger for agentic stuff, like function calling

that's rly what its specialty is now in my opinion

[errant dust]

I even told ChatGPT it was wrong and to try again. No dice, it cannot see the man is sitting on horse. FIne, I try another image. Llterally the NEXT ONE, and man is now with hands waving in group. ChatGPT informs me he is shaking hands with someone. WHERE?

Bah

By comparison, Claude got this one right too.

[bitter hearth]

apparently one of the new geminis is good

[errant dust]

I wasn't even trying to trick or trap ChatGPT. I was planning to use it for this purpose. When it failed, I wondered if the image was the fault, not clear enough. So I creted an account and tested Claude.

[mortal kite]

Looks like a DUNE version of R2-D2

[errant dust]

and it got it right. Might be a coincidence, so I tried a second one

[bitter hearth]

ye its weird cos the other R2s were not like that

just this one

[mortal kite]

I can just use ollama locally 😄

[errant dust]

and it was a repeat of fail and success. So I wrote off ChatGPT for this mission and am now using Claude

[bitter hearth]

ye ChatGPT might not necessarily be the best for captioning

[errant dust]

In case you are wondering, here is a man standing next to his horse according to ChatGPT:

[bitter hearth]

could be the sketch style is throwing it off

[errant dust]

Oh, and those muscular arms you see? That is armor

The image is a black and white illustration depicting a medieval or fantasy scene. In the foreground, there is a figure wearing armor and a cape, standing next to a horse. The background features stone architecture, with stairs leading up to arched doorways and people milling about. There are carts and what appears to be market activity. The scene suggests a bustling town or city environment from an era reminiscent of sword-and-sorcery tales.

[bitter hearth]

this particular style of illustration is not very common, I am thinking in terms of the training data

[errant dust]

Doesn't matter. Claude got it right, and gets them all right, while ChatGPT does not

I am choosing a tool to use, not trying to make excuses for them

[bitter hearth]

its just surprising cos GPT benches better for vision

and its a common choice for (illegally) using to caption training data

not allowed but people do it anyway

[errant dust]

and the benches are to be taken with a grain of salt, as we know

[bitter hearth]

ye

[errant dust]

Ask yourself how many images with "impressionist oil painting of..." were used to test Flux when they announced it was the almighty god of prompt adherence?

You cannot fail a test if you are never tested

[bitter hearth]

I think the same thing as before, that the prompt adherence test should be Clip Score, MS Coco / Imagenet, 30k image

they over-rated auraflow too

for prompt adherence

[sage burrow]

This is the version I seem to have downloaded, is it a good one, or?

[bitter hearth]

its always some weird prompt they use to test too, like blue cube, red sphere, green triangle
stuff that is neither in the training data, nor something people would want to make

[errant dust]

It is fine, but try Q8

And do not use --lowvram

the GGUF developer told me there is an issue with GGUF and LoRAs when using --lowvram

[sage burrow]

How do I not use lowram? It auto kicks in every time

[errant dust]

Edit the .bat file

https://huggingface.co/city96/FLUX.1-dev-gguf/discussions/25

[sage burrow]

"Prompt executed in 307.19 seconds"

usually takes 120 seconds with Flux Dev

[errant dust]

with a LoRA?

[sage burrow]

"Requested to load AutoencodingEngine
Loading 1 new model
loaded completely 0.0 159.87335777282715 True
Prompt executed in 282.77 seconds"

for the 2nd run (since everything is loaded it's faster

but still not as fast as Flux Dev regular

is GGUF better quality than regular Dev, or?

It didn't say using lowvram at least

[errant dust]

what are you calling regular Dev?

[errant dust]

That doesn't actually answer the question. Is the UNET file 24GB in size?

[sage burrow]

or the Flux checkpoint by the Dice guy, or a Flux furry checkpoint I found, or a model merge I made between teh Dice one and the Furry one, all 2 mins

[errant dust]

None of those is Dev, or even close.

[sage burrow]

oh!

not even the furry one? I'll find the link

[errant dust]

no

there is only one real one

[sage burrow]

none of these?

the first one is a model merge I made

[errant dust]

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

This file is the Flux Dev, unmodified

[sage burrow]

isn't that the 3rd one on my list?

[errant dust]

It might be. The name matches

[sage burrow]

That's the one I usually use

2 minutes per image

Btw, GGUF likes Bokeh!

[sage burrow]

So how about this version? I'm going to guess it takes twice as long? 😄

[rain current]

In progress....

[errant dust]

no, but it will be longer. The reason even it will be faster than the pure fp16 is that it is not loaded and processed the same way

[sage burrow]

that detail!

[errant dust]

BTW, if you want to test the Flux Dev in fp16, you need to change the dtype to default

[sage burrow]

I think I'll see if I can test it on HF, my computer is too slow lolol

[bitter hearth]

a female standing next to a large window in the kitchen, looking outside the window, side view of the subject and scene. cinematic, detailed, sharp.

[errant dust]

https://github.com/comfyanonymous/ComfyUI/issues/4674

[bitter hearth]

added anime style at the end

[errant dust]

and that will disable ---lowvram?

[sage burrow]

Thank you 🙂

[errant dust]

Interesting. For me just running plain main.py with no flags is enough. If I want lowvram I have to add the --lowvram flag

[noble coyote]

I'm having great fun using this metadata-converter node - so I am re-doing a lot of A1111 images in ComfyUI
But the metadata has no answer to the many textual embeddings in A1111 😦 Mobius checkpoint

[bitter hearth]

adding --lowvram would enable it i thought

[noble coyote]

No, it allows lowvram plus a small adjustment elsewhere to free-up some VRAM

[errant dust]

Ummm... the goal is to turn it off

[noble coyote]

I thought --lowvram was desired? If not - forget I spoke!!! 😄

[sage burrow]

so 2 versions are needed? One for when using dev regular, and another for GGUF use?

[noble coyote]

My apologies - let me delete my original post ...

[sage burrow]

It usuall is but, apparently it can mess up GGUF if loras are part of the equation

[bitter hearth]

different models, different nodes, but workflow is pretty much the same except for those two nodes

[sage burrow]

but I didn't get the usual loading lowram mode message, so fortunately I think I'm personally OK

[errant dust]

the thing is this, when you enabled the weight_dtyp to fp8 you are no longer processing the pure Flux Dev at fp16

th strength and reason why the GGUF models work o well with low-end systems sucha s ours is that it feed the model in blocks, not as a single giant chunk

[sage burrow]

I've never used fp16I (OK maybe once lol)

because slow computer

[errant dust]

yes, for fp16, you need to set it to default

My system simply crashes

[sage burrow]

ah ok, ty 🙂

[errant dust]

well, not my entire OS, but the python run

[sage burrow]

I have a 200gb swap file, so mine doesn't crash... I just need to go out for a few hour long hike or something to pass time ROFL

[errant dust]

it really is not worth it. BFL gave it, but the rest is the community finding ever clever ways to allow more and more people to use it

NF4 was the first such attempt, though what it did in the end was still to fit the entire model in 8GB or less all at once

GGUF is a different strategy altogether, and feeds the model in blocks

[sage burrow]

So it seems, to sum up:
Dev flux fp8 = 2 mins per image (from the 2nd one on)
GGUF flux dev = 4 mins per image, but the 2nd etc. don't run any faster?

[errant dust]

It should run faster. Mine does.

BUT

that assumes the prompt is not changed

there are two steps: load the Clip model to process the prompt, and this must be done each time the prompt is changed, and then unload the Clip and load the model to now render the image

if the prompt does not change, the first step is skipped so you gain that much time

BTW, GGUF also exists for the text encoders

the pure 9.3GB fp16 encoder by BFL is really slow

they also provide a smaller fp8 model

[sage burrow]

So, Flux hyper, any good? As good ad regular flux or nearly as good? Or?

[errant dust]

GGUF has that Goldilocks option of not quite pure fp16, but almost

isn't Hyper a Lora?

[sage burrow]

I downloaded that as well and put it into the right area

[errant dust]

Yes, my workflow includes the nodes for it

[stoic cradle]

is SD3 uncensored yet?

[errant dust]

the one I shared here for you

It was never censored, just crippled

[sage burrow]

I took it to be similar to flsh for SD3? Just makes the main model faster, or something?

[sage burrow]

Just use loras.

[errant dust]

I don't see how a LoRA can ever make a base model faster

[stoic cradle]

Do 1.5 Loras work on SD3?

[sage burrow]

Why is the hyper thing popular then if it's just a lora?

[errant dust]

Welcome to the Internet

[sage burrow]

Unfortunately not. You could do a workflow where you start or end with 1.5 and also use SD3 though.

Any particular reason you aren't skipping right to flux? It's pretty awesome, and there are many loras out for it now. Checkpoints also.

[stoic cradle]

Not familiar with flux

[errant dust]

It is the Second Coming

(until the Third Coming)

[sage burrow]

It's the model that came out after SD3, by some of the Devs who used to work for SAI... It's pretty awesome. Also very trainable 😄

[stoic cradle]

Sweet, I'll check it out

[noble coyote]

It allows "full dev processing" in 8 steps I believe - but its pretty much useless when you have other LoRAs attached - as they often need a lot more than 8 steps to fully work!

[errant dust]

By the devs who created SD in the first place before Stability existed, it should be added

[sage burrow]

There's free areas you can try it online also, huggingface has one

[sacred jewel]

This is nice!!

[sage burrow]

My fave workflows add 2-3 loras...

[sacred jewel]

Speaking of SAI, any news on refreshed/updated SD3 model?

[errant dust]

Here, by the Devs themselves: https://huggingface.co/spaces/black-forest-labs/FLUX.1-dev

[sage burrow]

2 weeks

[noble coyote]

Me too - a habitual combination - metallic, rbcharc0al, blue future - or ethereal grace, liquid flow, poetic style ...

[sage burrow]

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

But it works so much better with loras!

[sacred jewel]

Looks like something Audi would do

[sage burrow]

I tend to just use my own, or perhaps throw in a Duc Haiten one as well 🙂

[sage burrow]

Make more of your own! You'll be thanksful you did!

I'm still waiting for my fave model makers to make some flux checkpoints, one already has....

[errant dust]

Soon enough, by weekend's end hopefully, I will hav emy first loRA done, based on images you have already seen

[noble coyote]

There are so many great ones at Huggingface/Civitai/OpenArtAI

[errant dust]

(60 in all, labels still being made)

[sage burrow]

I'm definitely looking forward to it!

[sacred jewel]

That's no moon....

[noble coyote]

Automatic1111 seems to have deprecated Textual Embeddings - I find its almost impossible to make any now! They are as good as LoRAs - and so much simpler to make!

[noble coyote]

A1111 prompts now into ComfyUI

[errant dust]

hmmm.... I may even try the more convoluted Vast.AI way. Out of curiosity, and suspecting this, I still have some 2 year old credit in it. lol

not much, but enough for this I think

might as well put that moldy old digital money to use

I know Vst still love me. After spending tens of thousands though their service

(literally)

[sage burrow]

You are probably the Vast expert around here 😄 hmmmm, so how do I train a checkpoint on there without maxing out my credit card? LOL

[vast condor]

there are templates on there for kohya_ss, they work great

if you're doing flux, you'll have to do some git commands

[errant dust]

I saw some tutorials on Civit

I haven't gotten around to them since I knew that with no data, there is no LoRA no matter the method

[errant dust]

Really not. I have zero info on training any of these image models, whether LoRA or checkpoint. I used them for generating data to train my NNs, which I did locally. I had some dozen rentals working 24/7

[vast condor]

I was using runpod and vast pretty much exclusively for training before I upgraded my card recently, and was able to start training local

[sacred jewel]

Beautiful optput

[errant dust]

Basically I had two expensive rigs to train here. I could train multiple iterations and test, and then test them massively as well. At the time, when 2080tis were the peak cards, I had two machines with 2x2080tis each

[sage burrow]

Flux is so good at photography!

[sage burrow]

Awesome, ty 🙂 Does it also do checkpoints, or just loras?

[sage burrow]

A dozen rentals running non-stop, you are seriously good at multitasking!
I'd have to add something into the code to have each one remind me they exist LOL

[errant dust]

Consider this, if I had that many rigs running 24 hours it's because I needed that amount of data to be able to keep my training set up also running 24/7. This was not a short-term process and for them to accrue tens of thousands from me it's because I ran this for eight straight months.

[vast condor]

checkpoints also

though I havent personally tried to create a flux checkpoint, they are so big

[errant dust]

The only thing I had to make sure was that the account was constantly funded

[vast condor]

I see no reason

[errant dust]

I did also have to make sure that the rigs were constantly running and on two occasions I had to drop a computer because they were overheating and located somewhere in India as I recall

[vast condor]

with the right settings, a lora can be a full finetuning. this whole notion that a checkpoint is better, it's fake news

[sage burrow]

What did you create with it all? Sounds interesting!

[errant dust]

This may be true but at least for me using a LoRA slows down rendering. Whereas a straight up checkpoint will run at just the speed of rendering with a checkpoint

[sage burrow]

I want to do some checkpoint merges!

[errant dust]

Chess playing neural networks. Some just for me or for interested parties and two which were released commercially

One of those interested parties was a five-time world champion

[vast condor]

ok, but you can merge loras and checkpoints too. It's fine, just be aware that you'll need more disk if you're spitting out full finetunes in the cloud footprint

[sage burrow]

Keeping creating ones which will beat them! 😄

[sage burrow]

Can you actually merge a lora with a checkpoint? I was tempted to try that today. (though tbf, I'm not sure comfy merge is a real merge)

[errant dust]

Beating them was not the issue. Because chess engines have been able to beat humans even the very best for decades now. These were used for their professional preparation and just having the best engine wasn't enough. They wanted an engine that suited their particular style of play or understanding of the game and none of the standard generic options no matter how well rated could do that.

[sage burrow]

That's pretty awesome

[errant dust]

I was uniquely suited for the job also because my background in this particular endeavor was not computing but that of a chess player. And a strong one. Which is how I've always approached this and they knew I would understand their particular needs

[vast condor]

with XL and prior 100% yes, with FLUX...I thought I had read that in the koyha_ss doc, but I might need to fact check myself on that one

[sage burrow]

What would be really awesome is if you could merge older models with newer models!! (comfy pretends to, but it doesn't at ALL really)