🆕｜sd3

[dusky thistle]

love it

i really like the style here

[bitter hearth]

what do ranks in lora mean

[dusky thistle]

you can think of it as being similar to the "B" number with full models... param count

[bitter hearth]

SEG is very strange

some blocks need it positive and some negative

and some blocks need a strength that is 300-1000 times higher than other blocks

[bitter hearth]

I looked at this gen for like half a second (inbetween opening windows on my pc)

I thought it was someone's butt lmao

[sacred jewel]

dpmpp_2s_ancestral | beta

euler | beta

euler | simple

dpmpp_2s_ancestral | simple

For those using Flux.1 Dev, what is your favorite combination of Sampler and Scheduler?

I am torn between Euler and dpmpp_2s_ancestral but definitely feel that beta scheduler is a bit better to my eyes.

Thank you, this is awesome

[craggy crest]

ancestral anything doesn't work well with flux - try simple or normal as schedulers and heunpp2

[brittle nexus]

I really like DPM2, but it's sooo slow, Scheduler, Beta

[floral gate]

First time trying flux

[craggy crest]

@bitter hearth this might get you more than 4gig vram https://www.mimicpc.com/

[bitter hearth]

https://media.discordapp.net/attachments/493156996644864001/1119657422324240445/shocked_i_tell_you_flabbergasted.gif

[sacred jewel]

the new dpmpp+2s_ancestral, does. 😉

[craggy crest]

run the same prompt, same scheduler, same everything else, but use the heunpp2 sampler and post the result to compare, please

and also Euler

[floral gate]

Is it normal for the flux schnell fp8 version to generate images in about a minute even though it's 4 steps?

I put the comfyui on normalvram mode since I think I have 8GB?

[sacred jewel]

I did earlier see above... martini glass looking scene

(I did not post heunpp

[sacred jewel]

I am running a test set

[bitter hearth]

But it's paid I think

[craggy crest]

these guys have free accounts https://www.mage.space/ with unlimited generations

[bitter hearth]

Free??? Unlimited power???

People were talking about mage here

I forgot to check it

I was busy playing game

[craggy crest]

yeah. there are several tiers, all tiers, including free, have unlimited generations. and if you buy a tier, it's a very small monthly charge - no credits, ever

[craggy crest]

very nice

[brittle nexus]

Well, forge is still giving the expected bad result using ancestral models. I don't know what comfy is doing.

[sacred jewel]

Some lora's no likey :/

[craggy crest]

not real surprising - you got the right VAE decoder enabled?

[floral gate]

What was that one emoji lora?

I saw one a few months ago to make custom emojis since genmoji was announced for apple

I was looking but I can't find the one I'm thinking of

[sterile pendant]

Comfyui had an update that lets one work with flux now. The dpm ancestral 2s sampler. It's really nice but keep in mind, it's twice as slow since it's a 2S(ingle) sampler

[sacred jewel]

Only one flux vae that I am aware of.

[craggy crest]

wonder what they trained their lora with

[noble coyote]

https://civitai.com/models/654477/redblack-charcoal-style-for-flux

[noble coyote]

Loving this new rbcharc0al look! May there be many more like it! ⭐

A superlative Flux.Dev LoRA!

[noble coyote]

Flux.Dev

[dry wave]

put it online now:

https://civitai.com/models/669703?modelVersionId=749684

it can sometimes be a bit messy - but that's due to the training data. you always can lower the lora strength to get a more general comic look. With strength 1 you get dark fantasy with dark tones, high contrast, and hopefully not too often weird limbs

[bitter hearth]

Are they summoning a fire or eating food

[dry wave]

Who knows what they did in medieval ages? This dark castle looks so cold, maybe they just wanted to sit on a warm place while 😛

the images are not cherry picked. So there are strange things here and there. I just took all images as they came out in comfyui

training on only 76 images definitely leads to overfitting like some faces or clothing utilities appear over and over again

anatomy problems are due to the training data :/ I removed images with too weird things in there, but still these old comics are quite chaotic

[bitter hearth]

You'd have to generate good ones and redo the lora I feel like

With more images

O:

[dry wave]

I could imagine that by generating images with the lora, selecting the proper ones, and finetune on them it will get better

yes

but I found results surprisingly good for that small dataset. Flux is really great for finetuning

[noble coyote]

Flux.Dev and rbcharc0al LoRA (red and black charcoal)

[bitter hearth]

well they said its black and red so mission achieved I guess 😄

[noble coyote]

The prompt was made for this LoRA I'm sure 😄

[bitter hearth]

TBH most SD 1.5 checkpoints would mess this up

[noble coyote]

I'm glad its Flux

[rain current]

the loras in flux are on another level, with few images and a rank of 2, it's perfect

[bitter hearth]

conflicting information out there because some people say it is not good for training

[rain current]

It could be people giving opinions about downloaded Loras, not about Loras they trained themselves. I've trained quite a few Loras

[bitter hearth]

no I was referring to people who were trying training

[rain current]

i'm only speaking from my experience. GGUF Q8_0 is the one I use for generating, and it works very well with the Loras

[noble coyote]

Will Dev LoRAs work with Q8-GGUF at all?

[rain current]

yep

[noble coyote]

Great, I'll give it a shot!

[rain current]

now I'm running a Lora with 140 high-resolution images of different women, as I have the feeling that flux generates very similar facial features

[bitter hearth]

yeah this is the main issue with flux

would likely apply to fine tunes as well

its good at getting a lot of true positives in its samples but it is bad in terms of false negatives

fairly low sampling diversity as a result of that

[rain current]

a Lora for general use doesn't work; it would be for specific purposes

the ideal would be a fine-tuning of the model with over 100k photos. The only thing I'm missing is the money

[bitter hearth]

that's what I am wondering
would doing a general use fine tune actually work on a distilled model

or can it only make individual subject loras

[dry wave]

lokr might work for general finetune

[bitter hearth]

I don't know if a distilled model will ever get a wide sample variety

in inference

[sacred jewel]

Awesome stuff... thanks...

Can I ask how you got the metadata into your LoRA?

I trained one on CivitAI and it came out without any metadata.

[dry wave]

I trained with SimpleTuner, but I'm pretty sure the metadata was added by civitai

[bitter hearth]

I'm working on a pytorch workflow for fine tuning diffusion models

its gonna take me a while though

starting out small on 16x16 models

when I say 16x16 I mean stuff like MNIST

one single number not the whole grid

lol

[rain current]

I just discovered that Civitai trains Flux Loras. I'm trying one out since I have quite a few buzzes accumulated, and it only requires around 2k-4k per training

[bitter hearth]

TBH I might try Civit lora trainer too, it would be nice to have a plug-and-play tool to use

Civit loras overfit a bit but they rly aren't bad

its necessary for them to overfit because they are being given such a wide range of loras to make

better to err on the side of overfitting a bit

[dry wave]

civit loras are rank 2

if they overfit then just because they are trained very long

[bitter hearth]

yeah its due to other hyperparameters than just the amount of the network that was trained

[rain current]

But you can adjust it; the parameter is 'Network Dim.'

[bitter hearth]

ah okay that's great

[sacred jewel]

Can you see if mine got the Metadata ? I can't see it on my end 😦

https://civitai.com/models/660770/yfg-spyworldflux

[dry wave]

yes, the metadata is in there

but it is A LOT

maybe you put too many images with too long prompts in there and the comfyui node has problems with that?

[sacred jewel]

Thank you...

I used about 300 images...

🤭

It's a shared dataset... link is in the model page. I used GPT to caption them. I am running a new training using more steps. My first attempt I feel had too few steps.

[dry wave]

I don't find the trigger word in there, though... but maybe I just don't find it

[sacred jewel]

I do not think CivitAI trainer for Style lets you enter a trigger. I do have it in the dataset tags tho...

[rain current]

epoch 6/15, in progress

[bitter hearth]

looks cool

[sage burrow]

Since everyone is talking about FLUX civit lora training, I made 3 so far, they are so-so. You can find them on civitai.com I'm not posting a link here since they are not sfw. I'm going to make some more using non-civit methods soon to compare.

$2 per lora isn't bad though.

[bitter hearth]

$2 is great yeah

[rain current]

enraged goku with an attack position
default <> lora
epoch #9/15

[errant dust]

How many images did you train on? Per LoRA that is

It is strange how inconsistent Imagen 3 is with text. Sometimes it gives output that is such gibberish you wonder.

[noble coyote]

Flux.Dev with rbcharc0al LoRA

For some reason, Flux.Dev now works images in 2 minutes; whereas two days ago it was closer to 20! I'm not complaining 😄

[errant dust]

So Ideogram 2.0 is out today

Hi everyone,

We just launched Ideogram 2.0: our frontier image generation model with state-of-the-art capabilities in photorealism, graphic design, typography. It’s now the default model for your creations, so you can start using it right away.

Now free for all Ideogram users on desktop… and our brand new iOS app! For developers, API access is available in beta.

Besides the model update, we also launched a bunch of features:
5 styles you can choose from: General, Realistic, Design, 3D, Anime
Color palette control
Custom aspect ratios
Ideogram Search (1B+ images)

Here is our blog post if you want to learn more: https://about.ideogram.ai/2.0
Watch our video and retweet us: https://x.com/ideogram_ai/status/1826277550798278804

Ideogram (@ideogram_ai) on X

Introducing Ideogram 2.0 — our most advanced text-to-image model, now available to all users for free.

Today’s milestone launch also includes the release of the Ideogram iOS app, the beta version of the Ideogram API, and Ideogram Search.

Here’s what’s new… 🧵

Twitter · Aug 21, 2024, 3:17 PM

[noble coyote]

Its cool - just did my 20 images Ideogram 2.0

Flux.Dev plus rbcharc0al LoRA

[sage burrow]

64, 500, 120. Though with the second two I made the mistake of adding a slightly different thene. I have a few more tries left....

[errant dust]

ok, just curious. I plan to try their LoRA tool, and being so cheap, it is more a cost in time to prepare the content than monetary.

[rain current]

It's very good, but it still struggles to generate realistic photographs on many occasions, especially when the prompt is detailed. Ideogram 2 realistic <> flux dev

[errant dust]

In time, the real strengths and weaknesses will come to the fore, just as they did for others like Flux and Imagen 3

[bitter hearth]

if they are closed source then they should be way ahead TBH

[errant dust]

and they arelady were, but it depends on what you are asking for

[bitter hearth]

I am not anti-closed source, if they are good

[errant dust]

Some things wil be more apparent than others

[bitter hearth]

I have seen some good ideogram stuff yeah

[noble coyote]

"Is closed source like red sauce?!" 🥳

[errant dust]

oh it goes much deeper. Here, let me show a concept I gave Flux Pro at LEAST 2 dozen images.

and the third or fourth image by Ideogram 1.0 in a 4-image sample

[rain current]

I expected much more from v2; I thought it was going to utterly crush Flux

[bitter hearth]

the current timeline is weird yeah

in the early midjourney days I thought no one would ever catch up to closed source

[noble coyote]

I think Flux has set higher standards than the world of Generative AI originally anticipated!

Flux has beome The Disruptor!!!

[bitter hearth]

I have mixed feeling about Flux
because its distilled

[torn wharf]

I had all these plans to train some loras for sd3, i was going to start hitting it. Then Everything Changed When The Flux Nation Attacked

[torn wharf]

ip adapter just came out so maybe we can crack that distillation shell

[rain current]

ideo2 - flux

[bitter hearth]

I think SD3 will be the best model overall when it fully releases

IP adapter will help a lot yeah

I really under-estimated IP adapter
until about a month ago
I would never make an image without IP adapter now, its amazing

[torn wharf]

My prediction is better than schnell, comparable to dev, pro will beat it. (on imgsys after a week)

[noble coyote]

Two things can make Generative AI good; 1) the cash to invest or 2) a really excellent development team who don't necessarily have a lot of cash!

[bitter hearth]

for image quality yeah

[torn wharf]

same actually

[errant dust]

This first is Flux Pro. and was the closest to what i wanted including adding the text which over half the time it pretended I had not. The next two are typical Ideogram 1.0. It aniled the text over half the time.

[noble coyote]

IPA for Flux?

[hallow lion]

Flux IP adapter is here.

[errant dust]

It is supposed to be a young man BTW

[bitter hearth]

but because SD3 won't be distilled I think it will be possible to push it to greater levels

[torn wharf]

imgsys is just a blind preference test i think. so people can evaluate on prompt comprehension or aesthetic preference, but y0ou're right. They're probably voting which is prettier

[noble coyote]

Send me link? 🙂

[torn wharf]

https://huggingface.co/XLabs-AI/flux-ip-adapter/tree/main

[bitter hearth]

what I mean is
when I say SD3 is overall better I am including that we can use more tools with it

[errant dust]

SD3 Medium you mean

[noble coyote]

Thanks!

[torn wharf]

That's a good point. It won't be SDXL or SD15 tooling at first though. Same ol

[errant dust]

SD3 Large is closed source and trained on entirely different data

[torn wharf]

closed for now. Likely will be released

[uncut river]

ENTIRELY? no, cant believe that

[dry wave]

why? I don't think that there will be more tools for SD3 than for Flux if it is not clearly better

[errant dust]

It was stated openly

[bitter hearth]

SD3 Large is getting released I thought

[torn wharf]

yeah and not entirely. lol. that's a huge hyperbole pulled out of one big ol hole

[torn wharf]

they haven't said much about that but that seems to be the hinting

[bitter hearth]

distilled vs non-distilled

[torn wharf]

SAUCE

[dry wave]

so what?

[noble coyote]

Ketchup

🙂

[bitter hearth]

I think there are quite a few implications of distillation

[uncut river]

whatever is stated in public, is not always true

[torn wharf]

https://tenor.com/view/pulp-fiction-catch-up-catch-up-ketchup-gif-22879385

[bitter hearth]

some of them are evident already

[dry wave]

I know. But many of them turned out to be wrong

[torn wharf]

impossible to train

[noble coyote]

Sauce

[torn wharf]

can't do inpainting

[dry wave]

exactly xD

[torn wharf]

loras will never work

there won't be controlnets for distilled models

[uncut river]

loras will not work with sd3 ... ?

[errant dust]

It was not an official statement. It was by one of the Devs who left.

[uncut river]

yeah, distilled models (at the moment) need each their own loras

[torn wharf]

i'm just propanda dumping.

[dry wave]

it's about Flux/distilled models. People claimed in the beginning you cannot train them or use them for inpainting, controlnets, ipadapter and so on. That turned out to be wrong, though

I have to admit I was sceptical myself in the beginning

[torn wharf]

if i want to indulge my conspiracy theory departments of the mind, then i would imagine that invoke AI invested heavily into the OMI and then along came flux and made them panick, so they've been on a subversive campaign to undermine Flux's mindshare

[errant dust]

As I recall the rumor was started by a CEO of a company with a vested interest ... in that being true

[torn wharf]

(insight face geez. conspiracy mind is dumb sometimes)

[uncut river]

best defense is a direct attack ?!

[dry wave]

nah, people always claim bullshit. No need for conspiracy theories

I just remember when SDXL came out

and some guy on reddit claimed you need 80 GB of VRAM to train it

[torn wharf]

always a need for conspiracy theories. they help me rationalize human behavior and not go mental

[dry wave]

everyone who understands a little bit of math and ml knew its bullshit, but many people panicked and repeated this nonsense everywhere

[uncut river]

people are typically corrupt, but conspiracies give them too much credit. It's usually just organic spontaneous stuff, not a actual thought out conspiracy

[torn wharf]

SD3 comes out and people talking about 8000 network alpha

[errant dust]

Having an agenda isn't a conspiracy

[torn wharf]

at the same time, any plans people make in secret are a conspiracy. businesses have all sorts of plans they keep hidden.

a hidden agenda is

[uncut river]

that is true, but usually it's not that complicated

[torn wharf]

no i know. i'm just self indulging a guilty pleasure

[errant dust]

For some reason I have Madonna in my mind singing "I'm a conspiracy girl, living in a conspiracy world"

[uncut river]

lol

[errant dust]

lol

[uncut river]

damn you

[errant dust]

lolol

[uncut river]

so, when will sd3 finetunes or loras do some nsfw ... ?

[errant dust]

well, Flux NSFW LoRAs are a dime a dozen on Civit

[uncut river]

yeah, but flux is too slow on my pc

the smallest version runs, but barely

[errant dust]

Bear in mind they work (LoRAs that is) with GGUF Flux models

[uncut river]

i have a GGUF flux model

[errant dust]

BUT, and it is fair to add, all LoRAs will double the render times

for Flux GGUFs at least

if not a bit more

[uncut river]

I mean it's nice to be able to run flux GGUF locally, and the model itself runs very fast. But then the loading and unloading of the clip models and vae stuff makes it very slow

the slight improvement in quality and prompts do not out-weight the slowness

[errant dust]

I'm sympathetic, completely. I use the free offerings of Flux Pro mostly

[uncut river]

online you mean?

[errant dust]

yes

There are multiple sources

[uncut river]

most online sources no nsfw right?

[errant dust]

Yes, no LoRAs

[uncut river]

flux does have much nicer anatomy compared to sd3, out of the box without loras

[errant dust]

you mean not mutant bodies

[uncut river]

touching the grass without exploding into body horror, yes

[errant dust]

Believe me, the grass thing was only a sign of the greater problem

[uncut river]

it was a nice focus point for examples

[errant dust]

it did not stop with explorijng your inner green

As a user of Ideogram I will have fun feeding it some of my more interesting and challenging prompts from past images to see how the new model handles them

[torn wharf]

hehe butt

[torn wharf]

can't you do that as a user of any text to image model? weird flex

[noble coyote]

I can't find Flux IPAdapter folder in Flux IPA - anyone have an idea?

Just started up Flux+IPAdapter ... at last! 😄

[craggy crest]

you should probably read twitter more often - apparently it's... not ready to be released. i.e. bad

[noble coyote]

I've got the workflow ...

[noble coyote]

This error Error while deserializing header: HeaderTooLarge

[uncut river]

meanwhile from sd3, some decent body anatomy

[noble coyote]

Flux IPAdapter - Error occurred when executing ApplyFluxIPAdapter: mat1 and mat2 shapes cannot be multiplied (1x1024 and 768x16384)

Flux.Dev and rbcharc0al LoRA

[rain current]

as long as you don't tell him 'selling weed'

[craggy crest]

all i know is that both matteo and fofr posted this morning on twitter about it and said it was rushed in order to be first and isn't any good.

[noble coyote]

I'm also getting the mat 1 and mat 2 error ...

[bitter hearth]

the current best IP adapter workflows utilise negative images as well as positives
may be tricky to get that working with Flux

[noble coyote]

Posts on Reddit suggest that the HeaderTooLarge is because of a corrupt checkpoint file

[bitter hearth]

some people managed to get negatives working a bit using funny methods

[noble coyote]

Or even the need to change .safetensors to .ckpt

[errant dust]

In Flux? Yes, I tested it and it tripled (literally) the render time. It did work.

[bitter hearth]

did you use perpneg?

[errant dust]

yes

[bitter hearth]

yeah that takes 3 times the time that's normal

so if its possible to get at least some negative prediction working
then that could be useful for stuff like negative image embeds for IP adapter

stuff like SAG and PAG also rely on pushing the model away from bad images

SAG pushes it away from blurry subjects, and PAG pushes it away from subjects that are structurally bad and with light colours

the big question is can this stuff work with Flux without either not working or producing CFG-burn effects

[noble coyote]

... wait-a-minute! It's stuck at 16% Xlabs Sampler ... !!!

But the GPU is racking-up 100%

[bitter hearth]

its in xlabs/flux ipadapter

[craggy crest]

i've got no idea. https://x.com/cubiq/status/1826226733194088734

Matteo Spinelli (@cubiq) on X

So, impressions on IPAdapter Flux: it's a rushed model released just to say "FIRST". It's good at a narrow sets of styles/images. When it works it works very well. When it doesn't it fails badly. SDXL/SD1.5 IPAdapter models are much better. Let's wait for future releases.

Twitter · Aug 21, 2024, 11:55 AM

[noble coyote]

Just hangs at Xsampler

[bitter hearth]

sorry if this gets asked a lot: can flux do image2image?

[bitter hearth]

yes its has somthing to the with lora patching ?

is consol spaming lora ?

[noble coyote]

I'm just trying IPAdapter Flux ... seems a tad reluctant to work

[craggy crest]

it's not the model that does image2image

[bitter hearth]

oh. so you can do it in comfy?

comfy is the only thing that has a chance to run flux right

[dry wave]

diffusers? 😉

[dry wave]

every diffusion model can do image2image. It's how diffusion works

differently spoken: if you do text2image then you just do image2image on an empty blank image on full denoise

[bitter hearth]

so if i can get comfy to run flux i can do image2image no problem

[dry wave]

yes

[bitter hearth]

I did some research but I can't work out how they are going to port tools and nodes that require a negative prediction into Flux, cos it doesn't work well with negatives. That's mostly what I had in mind when I said Flux may end up with less tools.

I hope I am wrong though and someone finds a way, maybe via perpneg or something

[dry wave]

this is already the case with turbo models, too

[bitter hearth]

it is yeah I also dislike turbo and lightning

[dry wave]

also it might be possible to "undistill" flux

in principal this already happens if you train your lora on cfg-scale>1

[bitter hearth]

why whats wrong about them , they are fast

[bitter hearth]

the power of the negative prediction gets lost
and this gets used for a lot more than just negative prompts these days

I wonder if running flux over a ton of steps would allow it to tolerate higher CFG better

[sacred jewel]

Then you will NEVER be pleased ... there will ALWAYS be limitations no matter what...

[sacred jewel]

in a couple of weeks ... ROTFL

[sacred jewel]

https://tenor.com/view/big-meaty-claws-big-meaty-balls-spongebob-meaty-gif-26152406

[torn wharf]

https://tenor.com/view/patrick-star-no-this-is-patrick-spongebob-angry-mad-gif-9319809

https://tenor.com/view/tryem-chef-south-park-try-them-give-it-a-try-gif-21521138

[sacred jewel]

😄

[bitter hearth]

why? that doesn't really follow

[sacred jewel]

Some LoRAs are not necessarily great 😛

[sacred jewel]

You said you had a problem with a model being distilled... all I was saying is that you will never be satisfied because there will always be some compromise.

Otherwise, you'd be happy now with something, anything. No?

What is your favorite model right now and why?

[bitter hearth]

Everyone would rather run an undistilled model if they could, its not a personal thing. If people could run Flux Pro at home for the same speed as Dev then they would do so.

[sacred jewel]

Pro is also distilled... so is MidJourney... if you follow what I am saying...

[bitter hearth]

Pro isn't distilled, but Midjourney might be

[sacred jewel]

How would you know PRo isn;t distilled? Honest question...

There is NO WAY they could have EVERYTHING ever conceived in a model, ever... think about what you're implying.

How big is pro? 25B params? 12.5B params? Infinite params?

ALL MODELS have limits... API or not, Pro or not... this isn't theory or philosophy

[bitter hearth]

Well what I would say is that they said it wasn't distilled and I don't particularly see a reason to not believe them. Also logically I can't really understand why you would distill a closed source API model. The two benefits of distilling are speed and hardware usage, mostly VRAM, and if you are serving a closed source API model then these benefits are not going to be as apparent since you will be using H100s if you are trying to maximise efficiency.

I think what you're trying to say is something like people should be happy with what they have. That's fine, but I still like to talk about models anyway.

[sacred jewel]

No, my point isn;t what you may be considering DISTILLATION. What I mean is that every model has a limit. Say PRo is a 100 Billion Parameter model. That is the biggest version... it is by the very nature of our physical universe, distilled because it is not a 101Billion parameter model.

You mean that all smaller/reduced versions of the 100B model are distilled from the biggest one they have.. .sure, that is also true.

my point was, distillation is just that. There will ALWAYS be a potential for a bigger or better or more but never PERFECTLY FULLY CAPABLE models.

So, I see the Flux.1 Dev as the VERY BEST model... I do nto see it as distilled from Pro ... I just mentally ignore that there is a pro version because what's the point? It is not availabel for me to run it locally so I put it out of my mind. In my mind, Dev is the better model and then only lesser ones are distilled.

😉

[bitter hearth]

Sd3 best

Makes cooler balls

ok I see what you mean yeah

I actually agree about putting closed source models out of mind

I don't use midjourney for example for SD3 ultra

when new sd3 releases

I want to say 2 weeks but I'm not sure

[sullen moss]

https://about.ideogram.ai/2.0

Not bad

[bitter hearth]

Iphone

[dull star]

ideogram looks like ideogram 🔥

anyhow back to open source/offline image generators

[rain current]

[bitter hearth]

[errant dust]

https://www.bloomberg.com/news/articles/2024-08-21/this-tiny-startup-is-helping-musk-s-grok-with-image-generation

[dry wave]

why should it be distilled? There HAS to be a teacher model, otherwise no distillation. If pro is not distilled, there would be another teacher model instead

[torn wharf]

paywalled spam for audiences that are less informed than we are

flux ultra

Flux Queen

Flux California King

[dry wave]

otherwise I agree ;P

[dry wave]

it's like in science. Every new paper claims to be better than what was there before - otherwise it wouldn't have been published. Everyone knows that these evaluations are bullshit

image generation is even worse because it's often not even academic. So you just evaluate until you get the result you want

[bitter hearth]

It's the ign 10/10 of models

[torn wharf]

you've heard of the theory of relativity? get ready for the theory of inlawivity

[torn wharf]

every phd has to publish and defend it if i know things right

[hollow swift]

for this accurate snd realistic is kinda crazy tho..

[bitter hearth]

yeah I feel like whatever model was at the end of the chain and was actually the teacher would just be implausibly large if even Flux Pro was distilled

[bitter hearth]

is that flux or ideogram 2

?

[regal oasis]

A stylized portrait of Kamala Harris, inspired by the iconic Obama 'Yes We Can' poster. The image features a bold, high-contrast color scheme with shades of red, blue, and beige. Kamala's face is rendered in a simplified, graphic style, emphasizing her strong and determined expression. The background is divided into abstract blocks of color, creating a dynamic and powerful visual impact. Below the portrait, the words 'Yes She Can' or 'Hope' are written in a bold, sans-serif font, perfectly aligned with the overall aesthetic. The design exudes a sense of empowerment and optimism.

A stylized portrait of Kamala Harris, inspired by the iconic Obama 'Yes We Can' poster. The image features a bold, high-contrast color scheme with shades of red, blue, and beige. Kamala's face is rendered in a simplified, graphic style, emphasizing her strong and determined expression. The background is divided into abstract blocks of color, creating a dynamic and powerful visual impact. Below the portrait, the words 'Yes She Can' or 'Hope' are written in a bold, sans-serif font, perfectly aligned with the overall aesthetic. The design exudes a sense of empowerment and optimism.

A stylized portrait of Kamala Harris, inspired by the iconic Obama 'Yes We Can' poster. The image features a bold, high-contrast color scheme with shades of red, blue, and beige. Kamala's face is rendered in a simplified, graphic style, emphasizing her strong and determined expression. The background is divided into abstract blocks of color, creating a dynamic and powerful visual impact. Below the portrait, the words 'Yes She Can' or 'Hope' are written in a bold, sans-serif font, perfectly aligned with the overall aesthetic. The design exudes a sense of empowerment and optimism.

[craggy crest]

pro isn't distilled. it's raw. it has to be

[bitter hearth]

I agree but I was just trying to be polite as they didn't really seem to know what distillation is

[craggy crest]

it's their golden model - what they'll make other stuff out of. it has to be as it trained, not after it's messed with

[bitter hearth]

they would just sell the largest model yeah, since its over API

[craggy crest]

understood. and I agree - they probably don't

[bitter hearth]

too valuable to not sell

[craggy crest]

sure. and you don't mess with it. you let your customer to stuff like deciding what they want out of it etc

[worn obsidian]

Anyone know if there is a stable diffusion ultra? If there is would this be better than SD3?

[sacred jewel]

I think you are all misunderstanding what I meant.

I know what distilled means.

Big model. A distilled model is one made from [it]

Ok.

What I meant was that any model will always be distilled because it is impossible for it to contain -[everything].

Distilled in the sense that it will always be missing billions and trillions of parameters.

You cannot have an infinite model.

Hope that makes sense.

Even if Pro is 50b Params, it isn't 51b. It's missing whatever it's missing.

[dry wave]

Sorry, but the way to talk about it sounds like you don't know what distilling means. Knowledge distillation is a well-defined term in ml literature. It's nost just about a model being not comprehensive or big enough.

the risk that distilled models cannot be trained or used in the same way as non-distilled models is real (they are trained on MUCH less data in the end!). But it seems so far that Flux is quite nicely trainable, so we can be optimistic about that.

[pseudo owl]

It’s sd3 8b with I believe more steps, and some other things to improve quality.

[craggy crest]

while i understand what you're talking about, that isn't what the word distilled means. when you distill something, you filter it, refine it. what you're tlaking about is along the lines of a small sample from a large amount of stuff. what we're talking about is a model that contains all the information from the sample that it was trained on and then a lot of that information 'filtered out'

like a barrel of beer that is then filtered and refined down into a keg of beer

[craggy crest]

flux is so huge that even with distilling it's still got a lot of data - but that data is badly off kiliter. terms that aren't nouns tend to default to fantasy concepts. there are a lot of things that are badly overfit, and this is flux dev. schnell is even worse

[sacred jewel]

Thanks fort making my point.

Again, I know how we use distilled in AI world. Large model. Smaller versions of [it] are distilled versions of that model.

You can omit that. We agree.

Back to my point. The largest model the author made, cannot contain everything. It is "distilled" as it does not have infinitum. 😉

[craggy crest]

it's not distilled. it's got everything that it was handed. we gave it a book, it read the book, it knows everything in the book. just giving it one book instead of a library of books isn't 'distilling' anything. however what was done was after it read the book, someone went in, did brain surgery, and removed chunks of memory so that it no loner has any idea what it read that was stored in those chunks

[sacred jewel]

No need to try and make yourself seem smart. I am a dumb mf ... But logic is logic. I am not arguing the use of the language. Just the facts that no model contains everything possible.

[craggy crest]

you can keep on insisting on misuing the term if you like, but it would be in your best interst to stop

[craggy crest]

we were never talking about that. we were talking about what knowledge the model had to start with, and how it had a lot of that knowledge removed

[sacred jewel]

That's the point. It is not everything. You're trying to over understand my understanding of it. I am not disagreeing with you.

[craggy crest]

no one said it was everything.

[sacred jewel]

Nor was i trying to discount how we use the word distilled models.

[sacred jewel]

I did 🤦‍♂️🤦‍♂️🤦‍♂️🤦‍♂️🤦‍♂️🤦‍♂️

[craggy crest]

well no one but you did. we were never talking about the entire universe and infinite knowledge. just the enterity of the knowledge that the AI learned from the data set it trained on, and how that knowledge was refined and distilled to make the flux and schnell models - and how pro still has the enterity of that knoweledge

[sacred jewel]

Wrong.

The top bestest best model has distilled knowledge of the entire possible potential absolute knowledge. You cannot argue that fact. If it was handed 50 billion parameters, it got distilled data. A choice was made to leave stuff out.

Period.

That's a fact.

[craggy crest]

WE'RE NOT TALKING ABOUT THAT! period

[sacred jewel]

The original point was that being upset that a model is a distilled versions of some other unavailable parent model is a waste of energy because even the parent isn't perfect. There is always somethingissing.

[craggy crest]

no. that wasn't the original point. not only do you not have a clue what distilling means for ai models, you don't have a clue what the conversation even was about

[sacred jewel]

I never needed your approval nor were you seeking mine. We are just defe NN Ding our ar the uments.

[mortal mesa]

your saying it wrong

[craggy crest]

no one's asking you for approval. we're over here talking about horses and you're over there shouting about gorillas

[sacred jewel]

Ok. You're much smarter. Wow 🤦‍♂️

[mortal mesa]

sthap

or ill stamp my feet

[craggy crest]

if you do, he'll tell you that you're discussing earthquakes

[sacred jewel]

There you go.

I do not have a bad word to say about anyone

You were just dying to call me names.

Nicely done.

[craggy crest]

give me the link to the post where anyone called you a name of any sort

[sacred jewel]

Links? What are you 12?

[craggy crest]

hover your mouse over the message or tap on it, activate message link, then post the link here. so we can all read the post where someone called you names

you can't - becasue no one did

[sacred jewel]

I am not a little child. I will not play children's games. Grow up

[sacred jewel]

Re read the part where you're not important to anyone.

[craggy crest]

you're the one that has disparaged yourself, called yourself names, insisted that the converation is about something it isn't, and that is now making it even worse

[sacred jewel]

Like I said. You're the smartest person in the world.

I am distilled.

Go make refined beer

[craggy crest]

you asked if someone was 12. these are the sorts of comments that 12 year olds make. so apparently someone IS in fact 12

[sacred jewel]

Go tell your mom you won on the internet.

[torn wharf]

i can play pedantics too. in that sense, all knowledge is distilled. you don't have anything pure in your head. maybe your mother's voice

god i love the pedantry game

[torn wharf]

unless you're going back to a hand axe period of human technology, you're probably using distilled knowledge. even then! EVEN THEN THE HAND AXE! SOMEONE ELSE TOLD YOU BOUT IT

[errant dust]

You could at least have given him a smartphone

[sacred jewel]

But that would assume the phone is in fact, smart 😛

[errant dust]

Believe it or not, that brings echoes of a hardcore luddite friend. 🙂 I explained to him that if I was going to argue with the term it would be calling it a phone, not smart. At first they were phones with some computer abilities. Today they are simply portable touchscreen computers which have phone capability.

[errant dust]

I will say this much about Flux. It really is king of logos now. Even Ideogram is not really doing it for me, but perhaps it is a prompting issue. Imagen 3 is a flop. It is stuck on the idea of exceptionally detailed and intricate work. The word minimalist has no meaning to it (to try to rein it in)

MJ might be able to, but not having it, I cannot comment. They do have a free trial going now, but I don't think I qualify

I have been pounding away at it to get a new logo, and have seen well over 100 options, with maybe a half dozen acceptable. Now I have two finalists. SD3 Large, nyet. Imagen 3? Fuggedaboutit. Ideogram? Nah. Dall-E 3? Sorry. Flux is the only one giving me real options

You'd think I was asking either too much, or being too vague, but not really. Circular logo for "Chess & Tech" with name large and prominent, on green circuit board background and chess pieces.

[sacred jewel]

I have been noticing some Flux.1 LoRAs generating poor results but after unloading the models from Comfy UI and regenerating, they start working fine once again...

These two are back-to-back, only difference was unloading the model and freeing the cache via UI buttons.

[bitter hearth]

👍👍👍

[sacred jewel]

https://tenor.com/view/i-dont-believe-you-whatever-lies-gif-3602502

😄

[errant dust]

https://www.tomsguide.com/ai/ai-image-video/move-over-midjourney-ideogram-drops-v2-and-its-a-masterpiece-of-ai-image-generation

[hidden helm]

hgmm

[sage burrow]

So I was testing some stuff out, adding flux loras to workflows. On both my own computer, as well as Glif.
However I messed up and kept adding CSBW's Cascade models in the Flux lora section. (instead of flux loras) It worked though!! As in obviously worked. Soooo cascade and flux get along, or?
@dusky thistle

[sacred jewel]

Very nice...

[sacred jewel]

it looks SD1.5 to me.

[sacred jewel]

If you look at the console, do you see errors? A lot of times the generation will run, but the LoRA adds no effect to the weights...

[dusky thistle]

With what? On what software? That doesn't make sense for that to work

[sage burrow]

Ideogram and I don't get along I don't think. And how does it not know who Pietre Bruegel the Elder is?!

[sage burrow]

Using comfyui on my own computer. But also on glif, they have a new flux lora loader option (which uses HF to get the model)

[craggy crest]

it's flux

[sage burrow]

The forest ended up with a bunch of mushrooms out of the blue is what made me think it worked....

[elfin quiver]

logo,酷旅

[sage burrow]

My own lora can also produce sfw apparently

[sage burrow]

Flux Dev: My lora, no lora, CSBW checkpoint as a lora (but prob not really working)

[prime trench]

https://www.reddit.com/r/StableDiffusion/comments/1exw2m4/sd_31_is_coming/
another 3.1 RUMOR

still hoping

fix the anatomy, fix T5 mask issues, others quality similar to 3.0. That would be great

[meager ore]

girl

!generate "A beautiful forest with tall trees and a clear stream"

!generate "A beautiful forest with tall trees and a clear stream"

[noble coyote]

(((I just heard Windows 3.1 is coming!!!))) 😉

[bitter hearth]

I hope it does come out

[turbid grotto]

thanks for hopium

i want sd3.1 so much

[lunar rivet]

I've just heard that flux1.1 is about to be released on the same day as sd 3.1

edit: it was emailed to me by some guy in a furry costume

edit 2: thanks for the gold kind stranger

[noble coyote]

Flux.Dev + rbcharc0al LoRA

[hollow swift]

can gguf model use lora?

[rain current]

yep

Q8_0 for sure, because it's the one I use. The others, I don't know

[desert shale]

The word ‘Gael’ rendered in a retro-style font, crafted from translucent soap with a rich green liquid inside. The letters are hyper-realistic, featuring detailed bubbles forming around them and shimmering water drops clinging to the surface.

[desert shale]

#📝｜prompting-help

[errant dust]

Believe it or not, before Windows 95 was released, an alpha of Windows 4.0 circulated on some 25 floppies if memory serves. I actually obtained and installed it. Worked fine too.

[sacred jewel]

Wasn't Windows NT 4.0 a real product? Or are you talkign about a Consumer version of NT 4?

[errant dust]

no not NT

This was wild west anyhow

[noble coyote]

I only recently threw out a whole range of rainbow-coloured floppies cotaining Windows 3.11!!! 😄

[errant dust]

Ok, so reading about it, this might indeed have been some beta, but bear in mind two things: This was BEFORE 95 and was in 1994, and there was no NT branding.

[noble coyote]

NT was most stable!

[errant dust]

I recently (in last couple of years or so) discovered original 5 1/4 floppies I owned of games such as Ultima V from my Apple IIe. For reasons unknown my late mother had saved all that crap in a box.

guess it was overlooked and just forgotten

[sage burrow]

lol @reddit

[errant dust]

25+ floppies! 3 1/2. You cannot forget such crazy things.

You had to go through them all to install

[sage burrow]

It's the week-long downloads ill never forget!

[noble coyote]

My first 32k memory computer stored all data on a cassette tape!!! LOL

The company which built my earliest computer has since morphed into ARM

[errant dust]

I never had to deal with cassettes, but I do remember a buddy who did, who copied a game over the PHONE, and that took bloody forever as it made its way on to the cassette player he had it connected to

[noble coyote]

Bits bytes and bauds!

[errant dust]

Truly

People today talk about choosing an OS (consumer) flipping between Mac, Windows, or eventually Linux, but back then? You'd go to some large store and in the computer department would be faced with 2 dozen machines each with their own OS.

Even Apple was split among multiple ones, from the commandline Apples, the first Macs, the Lisa, the Apple II GS, etc

[errant dust]

I liked the LOTR reference right after

the second breakfast and elevensies

[hollow swift]

idk why when im using flux q4 on the ksamplers it doubles up s/it time, am i missing something?

[worn obsidian]

are these with flux?

[sacred jewel]

Yup... Flux.1 Dev

[worn obsidian]

you running locally using API credits?

[sacred jewel]

Locally... RTX 4090

[worn obsidian]

everytime I try and run on 24GB 4090 it does not work and doesnt have enough power. are you using all 50 steps 1024x1024?

[errant dust]

Even I can run Dev, albeit at a snail's pace, on my laptop 4060 8GB Vram.

[sacred jewel]

Yeah, you have to make sure NOTHING ELSE is running on your computer. For example, I can't run Google Chrome at the same time as Edge and Firefox.

If I do, some tabs use up enough VRAM to put me over the edge on VRAM and it thrashes to RAM. It runs, and runs well, but it slows from 1.5s/it to sometimes 30sec/it

So, try it with NOTHING running except one tab to ComfyUI GUI and no other programs and see how far you get. Also noticed that the BETA scheduler can fluctuate a lot between slow and fast Iterations not to be confused with VRAM exhaustion.

[worn obsidian]

thanks!

[errant dust]

You can use the GGUF models too you know

[sacred jewel]

never gone past 20 steps on any of my images ever. I have been using the default ComfyUI workflow since day one. All I have added have been a LoRA model loader and a few text/prompt randomizer nodes.

How many fingers? 😛

[worn obsidian]

SD3 is terrible compared to flux honestly

[errant dust]

That really depends on what you are rendering

for 95% you are right

Try a prompt that starts with "impressionist oil painting of..." and add what you like. Flux cannot do it

at all

[dull star]

yeah its unfortunate

[errant dust]

you can peg a LoRA onto it, but we're talking the core model

[dull star]

but even then, I prefer paintings on the currently proprietary 8B model compared to the 2B

[errant dust]

As Cat 4GB has shown also, cartoon and comic styles abound in SD3, but not in Flux. Flux has a single 'comic style'. It will churn out 100% of the time

It is perfectly possible they exist in the model, but are blocked, but the result is the same.

This is not a random conspiracy theory, but based on Twitter's Grok, which has shown a very far reaching range, and is also using Flux to power it

[sage burrow]

It does manhwa also

And regular cartoon style

[errant dust]

Clearly the deal with Musk included removing all the censorship, other than NSFW or extreme gore.

[sage burrow]

This contradicts itself!

[errant dust]

So obviously they can tailor the models without needing to build a new one from scratch

[sage burrow]

So you perhaps mean musk added a llm?

[errant dust]

Which, in fact, is the same difference with the two Dall-E 3 versions

[sage burrow]

Running flux on glif works fine since there's always llm in my workflow

[errant dust]

It is deeper than that. You could get an LLM to filter and prevent some images or prompts, but the LLM would not be able to make Flux generate images it could not before

Grok is likely using a much less censored version, probably already trained but not publicly available

Even Imagen 3 can probably produce many of the images it refuses to

"an image in the style of Dali..." BEEP!, Sorry, that is against my guidelines

Bah

[bitter hearth]

Closed source ai

sleeps

[bitter hearth]

Every girl has huge personalities with that style

[errant dust]

That said, despite some weaknesses, there are some things that Deep Mind's brainchild does better than any

One prompt I give them, simple enough is "Satirical cartoon of an artist at his easel painting an image with chess." Three of the four samples were pure genius.

These are three of the four images it spat out on the very first try

Text? Oh boy. I think that for text, it is schizophrenic. Or they are using multiple models. I asked for a logo which said "Chess & Tech" It could not spell this even 20% of the time

YET, I asked it for a poster with a long phrase, and it actually nailed it one in four times. The others were near misses.

The new Ideogram 2.0 BTW is even stronger in text than it was before

[dry wave]

I think, though, that there will be some finetuned version of flux (a lora or lokr merge or maybe someone is crazy enough to do a full finetune) that solves most of the style-problems Flux has

[noble coyote]

Flux.Dev with rbcharc0al LoRA

[sage burrow]

How come this is in the Flux sample workflow? Did they just put the wrong name, or? I'm wodering if it's this that is messing with certain results.

[errant dust]

It might simply have been introduced for SD3, but applies for Flux as well

After all, the batch size just determines how many images it will produce at a time. So sure, if you put 2 or 3 it will double or triple the total render time

[torn wharf]

love that sub surface scattering effect

[mortal mesa]

the node is mostly generic even though it says SD3 just an empty latent with batch size like the others, but the height width is converted to input, you would get a drop down list of selectable resolutions if it weren't

i use that one with flux

because i figured out where to edit the resolutions and added 2MP resolutions

[sage burrow]

The SD3 aspect is what I'm worried about. The batch size doens't matter.

The problem is the images are coming out perfect on civitai, but not my own computer. Civitai is expensive.

Someone did mention that civitai does have some extra stuff running in the backend (no pun intended) though

Does anyone know how to fix this with Flux? "Error occurred when executing CheckpointLoaderSimple:

ERROR: Could not detect model type of: C:\Users\oipte\Downloads\Comfy\ComfyUI_windows_portable\ComfyUI\models\checkpoints\flux1Dev_v10.safetensors

[mortal mesa]

ide try putting it in unet folder and loading with unet loader

[sage burrow]

Which one?

[mortal mesa]

should be load diffusion model i believe

[sage burrow]

THat one still gave me the couldn't figure out the model type error 😦

[low stone]

sd3 8b

[mortal mesa]

i haven't used that type since the beginning, i settled on the NF4 the many versions has me lost, but gonna try the GGUF's

[mortal mesa]

ya sorry i dont know, do you know where you got that version

[sage burrow]

version of?

[mortal mesa]

the one thats giving you issues

[sage burrow]

comfy?

[mortal mesa]

the flux your trying to load

[sage burrow]

HF

[torn wharf]

BALLS

[torn wharf]

heh hehehe heheheh eheheheheheheheuhuhuhuh huh huhuhuhuhuhh

[junior peak]

all you need are balls

[sacred jewel]

a bit, phallic 😛

[bitter hearth]

Needs a lora

To make everything that shape

[fleet meteor]

I read its possible to train flux loras on a 3060 now ❤️

[sage burrow]

oh? 😄

[bitter hearth]

Every day something new is possible in the AI world of things

[torn wharf]

monuments of humanity always are. Wells and obelisks.

[torn wharf]

i think it was before because the pieces of code were all available. but now it's all beginning to get packaged together for us non codemonkies to use

[sacred jewel]

The Flux.1 Dev Phallus LoRA

LOVE IT 😛

This LoRA is cool...

[bitter hearth]

What's that lora for

I see spaceships

[sacred jewel]

It's called Old Space Station.

[bitter hearth]

Gen something inside a planet

Like a stationed spaceship

[torn wharf]

https://tenor.com/view/austin-powers-radar-flying-penis-weirdly-shaped-aircraft-gif-16399784

[sharp osprey]

/credits

[vast condor]

that looks like a giant....

[torn wharf]

https://tenor.com/view/woody-disney-toy-story-hello-i-just-arrived-gif-17348320

[zenith hemlock]

Samsung Galaxy TAB3 Tattoo

[hallow lion]

Quick! Help the cat with 4GB vram!

[zenith hemlock]

ahahah lolll

[dusky thistle]

clock even has the correct numerals

only 30 steps per stage with a pretty heavily modified version of cascade

[bitter hearth]

2 VI

[dusky thistle]

oh, woops, well close enough lol

[bitter hearth]

:p

Lighting is super gud

Reminds me of game engine showcase stuff

[dusky thistle]

yeah

what i'm excited about is the coherence

that's a really complex scene and there's very little "impossible" optical illusion shit

base cascade would have a meltdown

[craggy crest]

someone send the cat some vram

[sage burrow]

I no longer have any anatomy problems with Flux Dev 🙂 All it takes is a lora or 2.

Next up, Dali paintings in Dev (the other thing it hates doing)

[hallow lion]

🤔

[slim root]

does sd3 work with forge or automatic1111 yet?

i've been not creating anything for a bit just realizing sd3 is out

nvm just got to the point in the video where it says you can just use it in the models folder of forge/automatic BEAUTIFUL

[craggy crest]

hand an image of the painting to claude, ask it to describe it in the format of a prompt for stable diffusion. then copy that and hand it to flux

[sage burrow]

Claude is pretty awesome! 🙂

[craggy crest]

yes. And ... see what you get by doing what I suggested

[sacred jewel]

I tried this method. First time using Claude. Do you find it better than others for a reason?

Here is the result of the test

[craggy crest]

something matteo explained about how flux likely was trained

[solar sorrel]

/Samsung Galaxy TAB3 Tattoo

[sage burrow]

The style is good, but unfortunately the clocks aren't very melty! 😦

I think I like Gemini Advanced' prompting better, but still not very melty 😦 @craggy crest

[dusky thistle]

what on earth is that lump/bottle of human meat on the bottom left lol

looks like 2b snuck in there with a woman on the grass

[sterile pendant]

Just some Salvador Dali shit. That's one type of prompt sd3 2b could handle very strongly 🤣

[sage burrow]

SD getting creative/artistic? ROFL

[bitter hearth]

Okay but what is that under the tree

Some cursed time to look up generations in this channel

[sage burrow]

Dali posthumously dabbling in horror? 😉

I thought I'd liven it up a bit 😉

[sage burrow]

Flux has been partaking again I think!

Flux is good at roman numerals!

[sterile pendant]

I should also make note that I'm not completely doggin on sd3, I still like it a lot, just not for people(not that I make many things with people anyways)

In a lot of ways, sd3 still stomps flux for creative stuff like hybrids of animals and humanoid hybrids

[sage burrow]

Does anyone know the max character length where Flux just gets bored and gives up?

[sterile pendant]

Should be 512 tokens for the t5, but a token is usually ~3/4 of a word

You could put in novels, but it will just scale all the weights and turn into an averaged pudding

[sage burrow]

I think perhaps the prompts aren't as relevant when I use three loras in the workflow!

[obsidian junco]

if the architecture wasn't strange on that it would be convincing

[obsidian junco]

do you have the ability to import image maps so you can draw out or base that on a real interior?

[alpine summit]

only based on the prompt.

[alpine summit]

Flux

[bitter hearth]

I need to switch to Flux lol
I've been making Rococo/Baroque palace images for like 2 months and this image you made just beat all of them 😄

the lighting in flux is so nice

[alpine summit]

That image is from SD1.5, only the image of a person on horseback is from Flux, the rest are from SD1.5.

[bitter hearth]

wow that's SD 1.5? amazing

[sacred jewel]

Is that a size 17 foot? 😲

what could go wrong?

[bitter hearth]

the skull is rly cool

[proven citrus]

show me your attitude

cats

[bitter hearth]

A scool you could say

[errant dust]

The best Dali imitator by far is Midjourney

[hexed dirge]

https://civitai.com/models/658958

[sacred jewel]

Damn bro.... killing it... this is awesome! Your other one is one of my faves already... thank you!

[hexed dirge]

thank you!

[hexed dirge]

sometimes it need 1.2 or more of weight

becaus e it's only rank 2

lora

[sacred jewel]

Bro, it JUST WORKS OOB 😄 ... just lovely stuff.

Can you explain to me (like I am a 3 year old) what Rank 2 means and that CivitAI training of LoRAs is a Rank 2 etc.? I never quite got a grasp of the topic. I have two loras I trained on CivitAI and three I trained using OneTrainer locally but I have NO IDEA what I was doing other than try and get a well tagged dataset.

[hexed dirge]

network dim is 2. So with some prompt you need to go up with the weight

i.e. 1.2

[sacred jewel]

OK, I did remember seeing this setting ... but I thought it was something we could change? So if I change the network dim to 1, does that make it a rank 1? Does it only impact the strength of the effect(s) ?

[hexed dirge]

yes

[sacred jewel]

😄

Nice explosion...

I could never get this type of explosions in SD3

[hexed dirge]

prompt?

[signal shuttle]

"A arab riding a camel in the desert" Flux dev with lora

[hexed dirge]

which lora

?

[signal shuttle]

https://huggingface.co/nerijs/dark-fantasy-illustration-flux

[hexed dirge]

thanks