🆕｜sd3

[alpine summit]

Sd1.5

Sdxl

[sacred jewel]

When you say SDXL, you mean the Base model or a finetuned version?

[bitter hearth]

Bad for anatomy, better for everything else including balls

[bitter hearth]

Siamese cats are funny

[bitter hearth]

Furry

[alpine summit]

Sdxl

[pseudo owl]

hyper flux and flux dev(not schnell) can even kind of generate gif's which is really amazing

[sacred jewel]

What? How? Do tell!

[pseudo owl]

the prompt is: "A consistent side by side four frame image showing consecutive stills from a looping gif moving from left to right. The gif is of a {what the gif is about}"

Nothing else, just a plain prompt.

[sacred jewel]

WOW... OK so then I guess I can crop each panel and paste it in to a GIF? I must try it 😄

[pseudo owl]

Yeah the only problem is that sometimes flux might format it differently like this.

[sacred jewel]

Maybe my "subject" is too complex?

This may be a good way to do Triptychs

[pseudo owl]

try 1024x1024 res? it might just not work at complicated characters, i didn't test them yet.

[sacred jewel]

I think you're right. My subject is too complex

It starts out with the four panels but then it goes off the rails as it converges

[pseudo owl]

try this: "A consistent side by side four frame image showing consecutive stills from a looping gif moving from left to right. The 4 frame very consistent gif is of {prompt} There are 4 frames in the gif image and each is separated by a black line."

This helps me do a bit more complicated characters.

[sacred jewel]

My original prompt and the adapted prompt into your suggestion

[pseudo owl]

Nice, I got this too, just a man laughing while water is pouring on his head and the same format.

[alpine summit]

Sdxl

[alpine summit]

Sd1.5

[noble coyote]

Florence2/Flux img2img

[mortal mesa]

with and without LongClipL

[alpine summit]

Sdxl

[sage burrow]

Wow! 🙂 I have so much trouble with teeth sometimes.

[sacred jewel]

So, I created a single huge checkpoint that includes Flux.1 Dev, the fp16 T5XXL encoder and the CLIP_L encoder as well as the VAE... 32GB compatible with the Checkpoint Loader Simple core node and a standard ksampler.

I was surprised it worked at all 😄

Here is a result from it.

[mortal mesa]

i did a bunch of that grafting in my preferred t5

than all the other versions came out GGUF ect

[frail shoal]

Any news on 3.1 being testable ? Been a some time now.

[mortal mesa]

no, but they got a new CTO last month

[sage burrow]

url?

[sacred jewel]

https://civitai.com/models/717680?modelVersionId=802542

[noble coyote]

So what do you then use for Clip and VAE at all?

[bitter hearth]

if it works in the checkpoint loader node then you just take clip and VAE from that node

[noble coyote]

"No File"

[sage burrow]

2 weeks probably 😉

[sacred jewel]

Built in to the Load Checkpoint Simple 😉

[sacred jewel]

I know, somewhow Civit futzed the upload 😦

Re-uploading now.

I have a 1Gb/s FiOS but the upload is going at 200Mb/s 😦

[sacred jewel]

Why is CivitAI sooooo painfully slow with large files?

[turbid grotto]

too much merges of the same models per second being uploaded)

[bitter hearth]

merges of the same merges

to do the same thing

[bitter hearth]

2 weeks

[bitter hearth]

gib prompt

[sacred jewel]

impasto oil painting, dark depressing fantasy art, a medieval knight hugging with arms around a large angry squid, running through town,

[sacred jewel]

BALLZ

[sacred jewel]

Try now...

https://civitai.com/models/717680?modelVersionId=802542

[hallow lion]

lol 31 gigz

[sacred jewel]

32.6 to be exact 😄

It makes life MUCH simpler for me... two nodes replace like 7 other nodes before.

I always use T5XXL fp16, full Flux.1 Dev and the VAE so... works for me and others have also liked it so far.

My bad... it's even bigger 😛

[sacred jewel]

Also uploaded the combined model to Huggingface as an alternative

https://huggingface.co/gonzalu/flux_dev_combined/tree/main

[sterile pendant]

I wonder if there's a node to make checkpoints that can let you use ggufs for the transformer and t5. You could cut that file size in half and have 99.9% the accuracy for the bf16 versions. Would save a shitload of sysmem as well

[bitter hearth]

thanks I love these
so easy for cloud to just do 1 download

[sterile pendant]

actually on second thought, scratch that. you'd still need the gguf addon and the addon would have to be expanded to include hooks into comfy to detect that it needs to be in gguf mode. would probably take a bit of tinkering to save a few node's worth of clicks

[bitter hearth]

the typical prompt for Flux Dev

[bitter hearth]

for comparison, an SDXL image from earlier

[dusky thistle]

I wonder if it uses the correct clip L

Apparently there was a reddit thread about a clip L finetune

Im not sure anyone realizes flux has its own trained version of clip L.........

They didn't exactly announce it or draw any attention to it

Just kinda stuck it in the diffusers version and said nothing about it but it's there and it's better

Idk, they never said, but it's different and you get better results with it

Presumably it was frozen and the weights trained against it

So def makes sense to use that over anything else imo

It's def clip L still

The differences are small

They tuned it but didn't overhaul it completely or anything

Yep

In the text encoder folder

You'll see the hash is diff from. The openai one everyone is incorrectly using

They def should have something on their main page about it

I know even comfyanon was using the wrong one lol

Cool yeah you're good then

Yeah def

You want your stuff to get the best possible visibility

Why work hard as hell to make an incredible model and not spend 30 sec telling ppl to download the correct text encoder lol

Tons of images ppl are posting are degraded cuz of using the wrong te

yeah, no doubt

lol

check out those keys

my own

yeah pretty nuts what this thing is capable of with a bit of tuning

[craggy crest]

or what it's capable of doing right stright out of the box

[dusky thistle]

i'm pretty tempted to slap together a pair of 3090s and go nuts with training experiments

yeah, no doubt

the blur shit though... ugh

[craggy crest]

i don't get blur ... 😉

[dusky thistle]

managed to pretty much compleetly get rid of it

[craggy crest]

unless i want it

[dusky thistle]

agreed

watch any movie... you're not going to see 297 consecutive bokeh'd shots

[craggy crest]

pull up dev, drop in a prompt, and stick the word yesteryear at the start of the prompt

[dusky thistle]

it's like writing a song where the entire thing is a drum breakdown with guitars building up to more buildup lol

[craggy crest]

pro is overfit for fantasy, dogs, women, and anime catgirls. dev suffers with the same thing - for both of them, if there's no clear subject, you're goign to get a subject that is one of those 4 things, and sometimes a mixture of 2 or more

[craggy crest]

schnell does't have nearly the problem with that

[dusky thistle]

lol, damn

[craggy crest]

all single word prompts. workflows in the image

[dusky thistle]

Yeah and bokeh to hell and back

[craggy crest]

sure, i'm walking through latent space, finding out where things are at

[dusky thistle]

It basically reduces flux to a single subject model when it does that

Yep good stuff to check out for sure 🙂

[craggy crest]

have found some interesting effects. some interesting areas.

[dusky thistle]

generating directly at these resolutions

no inpainting or upscaling

the ceiling is really high for this thing

[craggy crest]

Xenolithic rock

and some really strange terms that i have no idea why it knows, but it does. this the term Xanthate - and it looks almost exactly like what the chemical actually is

[dusky thistle]

can confirm, i'm a chemist, that does look like it lol

[craggy crest]

now why in the world does it know that, but NOT know centaur?

[dusky thistle]

w hat does it do with "olefin metathesis"

yeah that's odd lol

[craggy crest]

once second, let me run that

[dusky thistle]

i can't tell if it's smart or glitchy

first one reminds me of this

just a tiny bit

[cursive frigate]

Does anyone have a Good Flux Workflow with LoRAs, and Upscaling that they would be willing to share. Or even a link to a page with a good workflow? I could really use the help. Everything I have found is either way to basic or way to over the top.

[dusky thistle]

now this is my kinda garden

and house

[craggy crest]

there should be some in the workflows channel on the L2 discord

[craggy crest]

that's your house?

you're not invited over anytime soon

[cursive frigate]

L2 Discord? do you have a link for that?

[craggy crest]

i'll DM it to you

can't post discord links here

[craggy crest]

what happens if you add the phrase "Yoshiyuki Tomino anime" to the start of the prompt for this one?

[dusky thistle]

will do once i figure out wtf that one was lol

[sterile pendant]

Because there are tons of science images in the dataset. Almost every wiki page for elements or compounds has some kind of image usually. Centaurs are much more likely to be hit and miss because they are mostly illustrated by artists and BFL probably went the safe route of only(mostly) using whitelisted artists

[dusky thistle]

not these...fukcin around with too many experimental custom nodes with halfass implementations, just gonna cause confusion

[hallow lion]

plain flux dev?

[dusky thistle]

with a realism lora i'm working on

[hallow lion]

well you're doing great

[dusky thistle]

thx

there's one thing unrealistic about that last image lol

though i guess maybe she was patient enough to just stand there during a snow squall for an hour or two

wearing layers lol

[muted dove]

https://tenor.com/view/all-a-bunch-of-tree-hugging-hippie-crap-eric-cartman-south-park-tree-huggers-hipsters-gif-21521965

[noble coyote]

Anyone have a random prompt generator w/f for ComfyUI at all? 😄

[muted dove]

DM'd you

[noble coyote]

I have d/loaded the 32Gb Flux1.DevSingleFile checkpoint by YFG - Flux.Dev output (4.2 s/it) and Flux.DEVSF output (5.7s/it)

I think the larger ckpt just about wins - although there's very little in it!

Florence2/Flux.Dev.SingleFile img2img

[noble coyote]

I'm getting a "No module named dyanmicprompts.samplers" error

[muted dove]

Are you using a dynamic prompt capable node?

Like the ImpactWildcardProcessor?

What I gave you is just a dynamic prompt.

[noble coyote]

https://github.com/adieyal/comfyui-dynamicprompts

So I could use something in the Impact Nodes?
Let me see ...

[noble coyote]

It works with ImpactWildcardProcessor
How to get it to reference that list of prompts at all?
They're in the Wildcard folder ...

OK, its working 🙂

[noble coyote]

Random Automobiles (Mobius checkpoint)

[sage burrow]

With flux?! If so, AWESOME!!! 🙂

[muted dove]

Dynamic prompt are independent of the model type.

[sage burrow]

Here i thought sd3 onwards didn't accept weights and dynamic prompts?!

[muted dove]

I still use them 🤷🏻‍♂️

[sage burrow]

Time to dig out my old prompts! 🙂

[muted dove]

First thing I do with every new model...go through all my old prompts 😄

[sage burrow]

Weights didn't seen to make any difference; perhaps I'll have to ramp them up

[sullen moss]

Some more upscaling tests

[winter kettle]

girl

[sacred jewel]

I added the clip_l i have been using since SD3. Seems to be fine. I saw a mention of using ViT Clip_L yesterday instead ... Have to look further.

[sacred jewel]

Do you have any side by sides you've done? I am still in bed. 🤭 But will look in to it as soon as I have coffee 👌

[sullen moss]

https://www.reddit.com/r/StableDiffusion/comments/1f84qs6/comfyui_advanced_live_portrait/?share_id=bWEJhLg6lkAppiJkX2lE1&utm_content=1&utm_medium=android_app&utm_name=androidcss&utm_source=share&utm_term=1

[alpine summit]

Sd1.5

[errant dust]

https://www.tomsguide.com/ai/ai-image-video/i-put-7-leading-ai-image-models-to-the-test-with-the-same-prompt-heres-the-winner

[sacred jewel]

How do I use the two encoders from the Flux.1-Dev Repo on Huggingface? Unless they got compiled as CLIP (sorry I do not know the proper terminology) Comfyui compatible checkpoints. I do not know how to use diffusers via Python or otherwise. I have vaguely used them before but I forget how.

Thanks

[sacred jewel]

She is a bit, necky 😛

[errant dust]

Different than what? It is the same as the SD3 Medium one.

[bitter hearth]

hash is different though
it could just be re-uploaded and it is otherwise the same

what I would say is we need to make sure it is definitely a quality improvement rather than just having an effect that is the same as changing the seed

[noble coyote]

One of Flux's many foibles - longunek

[sacred jewel]

🤣

[sacred jewel]

https://tenor.com/view/seinfeld-not-that-theres-anything-wrong-with-that-gif-10732604

[noble coyote]

Moebius checkpoint using Searge SDXL nodes

[sacred jewel]

Ok so i downloaded the versions (encoders that is) linked in the comfyanonymous Huggingface that he links to in his samples page. I assumed he compiled them from the BlacolkForestLabs flux repo diffusers.

@dusky thistle do you have any links to this info and a rabbit hole I can dive into? I am always after top highest quality so now you have me in intrigued 🤭

[bitter hearth]

gonna switch to the one of the official huggingface space just in case anyway

does anyone happen to know if you need auth token for wget to their huggingface?

in general github does not need auth token and civit does or does not depending on whether the uploader ticked a box requiring login

[sacred jewel]

Flux knows what my old room looked like back in the 90s 😛 😛

[sacred jewel]

Flux is gated... so yeah, you need to use API I think.

or some sort of token mechanism

[bitter hearth]

ok yeah will need token then

[dusky thistle]

it's not

[dusky thistle]

those are the wrong version

https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/text_encoder/model.safetensors this is what you want to use... just name it clip_l_flux.safetensors or something

[sacred jewel]

CLIP_L from ComfyUI HF Repo

[sacred jewel]

Ahh... okie... let me go play 😄 Thanks as always... 👊

ViT-L/14 replacement for CLIP_L

and now for my testing prompt: CLIP_L first (the one from ComfyUI repo)

[dusky thistle]

try having it write a full sentence of text on something

that's one of the better coherence comparisons

the weights were surely trained against the frozen clip L i just linked you to

[bitter hearth]

does anyone know good nodes to use kolors these days

[sacred jewel]

ViT-L/14
Is it better? Looks better to me...

[dusky thistle]

better to test with a lot of text

a cat holding a sign that says "i'm not sure if we've ever been using the correct version of clip L"

try smoething like that

[sacred jewel]

https://imgsli.com/Mjk0MDA3

[dusky thistle]

you just can't tell very esaily from an image like that

there's too many things that end up being subjective

text is right or wrong

[sacred jewel]

Just downloaded the Clip_L from BFL HF repo as proposed... here are the hashes for the ComfyAnon version vs the BFL version if you want to know 😉

OK here goes...

CLIP_L ComfyAnonymous Repo

[bitter hearth]

oh no

[lunar rivet]

bfl repo has both clip_l and t5 in bf16 as opposed to comfy's fp16

[sacred jewel]

CLIP_L from BFL

side by sidehttps://imgsli.com/Mjk0MDEx

[sacred jewel]

but the t5 is in diffusers format, no?

[dusky thistle]

the comfyanon one is the sd3 one

[lunar rivet]

it's the same thing, just split into a couple parts

you can merge them all together and use in comfy

[dusky thistle]

i'm gonna guess something's not ideal with your sampling/scheduling/shift

[sacred jewel]

Ah ok... how can I do that?

[lunar rivet]

lemme see if I still have the script for it

[sacred jewel]

Here is my workflow (the key parts)

And here is the ViT-L/14 version

ViT_L/14 is certainly getting long text better, no?

This is the repo for the finetuned ViT-L/14 for Flux

https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14

[lunar rivet]

I can't find the script but this model matches the sha256 of mine
https://huggingface.co/city96/t5-v1_1-xxl-encoder-bf16/blob/main/model.safetensors

[bitter hearth]

how did you verify that they are the same?

[lunar rivet]

I just looked at the keys

[bitter hearth]

I'm confused cos Clown said the hash is different

[lunar rivet]

I'm talking about t5xxl, not clipl

[dusky thistle]

yeah the t5xxl you can grab whichever

[lunar rivet]

hashes will be different between fp16 and bf16, yes

[dusky thistle]

that ones fine

the results with the clip L from the flux repo are different

[lunar rivet]

technically bf16 should be a bit more accurate in both cases (clip and t5)

but the differences are tiny

[dusky thistle]

it's pretty visible

i don't know if they've retuned it or what

there's not exactly any discussion or mention of it anywhere

i just grabbed their stuff day one and didn't look into it any further

a few ppl have done comparisons and gotten significantly better results with the one from the flux repo

[lunar rivet]

I'm honestly not sure cause I've been using their checkpoints from the start too

never bothered checking fp16

[dusky thistle]

what i can say is with cascade, i can't even see a pixel level difference between bf16 and fp32

but here it look spretty different with their official one vs the sd3 one comfy links to

[sacred jewel]

Downloading now... let's see which one I got 😄 I bet I got the one from ComfyAnonymous repo.

[lunar rivet]

you can just load the model you got with comfy and it should show precision iirc

[sacred jewel]

Is that anode that displays this? or on the console?

[lunar rivet]

in the console as it's being loaded

[sacred jewel]

ComfyAnonymous CLIP_L

BFL CLIP_L

ViT-L/14

[dusky thistle]

one way to think of it is like what language and accent you grow up learning

someone who grew up in minnesota will still be able to understand someone who grew up in ireland

[sacred jewel]

all three above using the t5xxl from ComfyAnonymous Repo.

Will test now with the one from city96

[dusky thistle]

but it's better to use the language and accent they became familiar with as they grew up

wrong language = wrong file entirely

wrong accent/slang/whatever = gonna get reasonably close, but things will be a bit off

[pseudo owl]

what flux model is that? i get this with flux.1 dev took from their repo:

[dusky thistle]

whatever encoder they trained their model on is the one you should use

finetuning clip doesn't really make a lot of sense imo unless you then freeze them and tune the weights again

[lunar rivet]

download t5 from google's repo and extract the text encoder layers if you want to be 100% sure you're getting the best of the best x)

[sacred jewel]

Flux.1 Dev from repo. Using Diffusers Loader... BFL repo that is... my seed is 1234. 16:9 A:R, Euler/Beta/3.5 Guidance

My current CLIP folder ...

[dusky thistle]

yeah there's no visible diff from their t5 and the google one

[bitter hearth]

LOL I wish I could have a nice full clip folder like that
on cloud I always start from empty

[pseudo owl]

sadly diffusers doesn't support euler beta, but with flowmatcheuler, seed 1234, 3.5 guidance

[dusky thistle]

just get a sacrificial 4090 lol

smear some syphilitic elecrtons on it to break it in

[bitter hearth]

I can't, cloud is too addictive

today I can see 8x A40 45GB for under $2 per hour
its such crazy value

[sacred jewel]

Not sure what you mean by "it's not compatible" because clearly mine is working... unless I missed something?

[pseudo owl]

no, I'm using diffusers python library, not comfyui diffusers loader.

[dusky thistle]

that's gotta add up

[sacred jewel]

Gotcha

[dusky thistle]

i'd be broke as f haha

i got shit running probably 20 hours a day here

[lunar rivet]

btw this is the new finetuned clip, right?

[sacred jewel]

25 hours a day for me 😛

[sacred jewel]

Yeah, someone posted it on Huggingface... I linked it above...

#🆕｜sd3 message

[lunar rivet]

was looking at it earlier too but it's kinda hard to say if it's obviously better

[dusky thistle]

it just doesn't make any sense to tune a text encoder after the weights unless you're fuckin around

[sacred jewel]

yeah, so far it is just different, not sure if better.

and here are the two versions of T5XXL ... ComfyAnon on the left, City96/BFL version on the right

https://imgsli.com/Mjk0MDMw

[bitter hearth]

IDK if this text is the best test

you want something with clear lines

[dusky thistle]

the best thing to do is to have like 10 different things where there's a clear right way for it to be generated, and a clear wrong way

then spam out at least 10 seeds for each

[sacred jewel]

seed now 9999
Above is CLIP_L from BFL and T5XXL from BFL

[noble coyote]

Flux1.Dev / Flux.Dev.LARGEFILE

[dusky thistle]

what clip L did you download?

[noble coyote]

Slightly less detail in the LARGE FILE version!!!

[dusky thistle]

https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/text_encoder/model.safetensors this is the official noe

one

are you using that

[noble coyote]

t5xxl_fp16/clip_l

[sacred jewel]

and now with the ComfyAnon versions of the T5 and CLIP-L

SLIGHTLY different...

[sacred jewel]

I only see a few pixels difference...

[dusky thistle]

kohya's sd-scripts is working really well for me

[noble coyote]

Realistically, its not got a lot going for it

[dusky thistle]

onetraner appears to support it now as well - i had horrible results with my one test, but i'm sure i screwed something up

[bitter hearth]

there is a new sampling method on arxiv for faces
that no one has ported anywhere yet
it might be amazing combined with existing stuff like ip adapter

[noble coyote]

What do I rename this to at all?

[sacred jewel]

Photoshop says this is the pixel diff between them

[noble coyote]

The trouble with art - both look pleasingly good!

[sacred jewel]

I renamed mine "clip_l-BFL.safetensors"

[noble coyote]

So there's no right one and one wrong one

[dusky thistle]

i renamed it to clip_l_flux.safetensors myself

[bitter hearth]

https://arxiv.org/abs/2408.02078

[sacred jewel]

Yeah, for me it has to be a glaring quality difference ... otherwise it is somewhat subjective.

[dusky thistle]

with text encoders, there are right and wrong ones

the weights are trained to interpret the results of the text encoder

if you use a different text encoder, it's going to be working with something it wasn't trained on

the difference will usually be small, often be subjective, but it will be worse as a whole

if you want to see how big of a deal it can be in a more obvious scenario, try training a lora where you train the encoders separately (which you should, with sdxl for example)

[sacred jewel]

Can you sahre your side by siudes shwoign the differences you mention? I see no clear differences PRO CLIL_L from BFL vs the version from ComfyAnonymous repo.

Maybe your tests are more obvious than mine. I only see minor differences but not yet always better differences

[dusky thistle]

i don't have em handy

other ppl ran tests a while ago

but you don't need tests for it

try training a lora where you train the unet/transformer/weights first

then, in a second run, train the text encoder

[sacred jewel]

So how can I judge otherwise? I mean, I have not seen the major differences you are suggesting 😦

[dusky thistle]

then, do it the other way: train the text encoder only, then train the weights against that

the results are quite a bit better

[noble coyote]

Running with clip_I_BFL

[dusky thistle]

well it's not the way any model is trained to train the text encoder after the model

[lunar rivet]

it's not really a different text encoder though, it's the same text encoder in different precision, the differences will be minor at best, it's like a rounding error

[dusky thistle]

idk, i never saw a difference between bf16 and fp32 that's even close to that

if they didn't tune them somehow, then something weird is going on

it's not impossble that for some crazy reasons i don't understand this new tuned version really will lead to better results across the board, but i wouldn't trust that without hundreds of test images

[lunar rivet]

neither clip_l nor t5xxl were trained along with the flux transformer, it's using the default models from other repos

[dusky thistle]

ya know what... i might just go manually recast the openai clip L to bf16 and see if that looks identical

[noble coyote]

clip_I / clip_I_BFL (both Flux1.Dev.SingleFile)

[lunar rivet]

make sure you don't just cast everything tho, some of the weights are in fp32, I mean in fp32 even in the fp16 model

[noble coyote]

More detail using clip_I

Subtle difference to chin and mouth

[bitter hearth]

yeah putting Deepfake in the name of your Arxiv paper is wild
I actually didn't like the results in the paper
but I am very optimistic because these are the results before IP adapter has been applied on top

[noble coyote]

Her left earring larger and brighter in the clip_I_BFL version

[craggy crest]

that's what is expected - that's where clip_it's strengths lie... in the fine details, the ambient and background stuff, the artsy stuff

[dusky thistle]

yeah i'll just recast everything to bf16, for both versions

quick n lazy and if they really are different it should be obvious

[craggy crest]

it was published at the beginning of august, when "the word deepfake is bad" sentiment hadn't really gained much traction. but with the california stuff in such high profile, now you don't want to even think that.

[dusky thistle]

deepfake in the title... omg lol

that word has had a really bad reputation in my social circles for a while at least

[noble coyote]

DeepFriedMarsBar

[craggy crest]

texas state fair

[noble coyote]

TexasDeepStateFair

[craggy crest]

yeah. they'll fry anything and sell it at the snack bars there

[sacred jewel]

LARGE one has better defined circles on her dress/chest area IMO. Or whatever those black round things are lol

[noble coyote]

Yes, when it comes to arty output - each choice has its strengths - there is never really a right one and a wrong one

Its better testing on the photoreal

[dusky thistle]

where you seeing these prices? i'll prolly have to go throw some training jobs on that at some point, that's perfect for fp16 loras

[bitter hearth]

vast.ai

ye that's the thing, local diffusion inference can be done locally but I would be pushed onto cloud for inference of other types of model, or most types of training, anyway

[dusky thistle]

oh, small gpu?

[noble coyote]

SingleFile and clip_I_BFL

[sacred jewel]

I accidentally selected two CLIP_Ls in the Dual Clip Loader instead of T5 and CLIP_L and got this LOLOLOL

[craggy crest]

SD3 uses clip_l, clip_g, and t5xxl - flux just uses clip_l and t5xxl. I wonder what you'd get if you switched out flux's clip_l with sd3's clip_g

[sacred jewel]

hmmm

[sacred jewel]

Error...

[noble coyote]

Trying that now ...

[bitter hearth]

I love how even botches look good now

[sacred jewel]

😄

[bitter hearth]

some of my favourite images I made a full on botches

[dusky thistle]

wut? it worked at all?

i wonder if it just used the t5 you had selected before

[craggy crest]

or nightmare fuel perhaps

[noble coyote]

Clip_g - Size Mismatch error!!!

[craggy crest]

drat. well, worth a shot

[noble coyote]

🥳

[sacred jewel]

Hmmm... I cleared the caches...

[dusky thistle]

either way if it runs it's the equiv of random prompts lol

[sacred jewel]

True...

[noble coyote]

That's a lot of cheese that error generates!!!!!

[mortal mesa]

feizc/FluxMusic Text-to-Music Generation with Rectified Flow Transformer

yup

[sacred jewel]

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

[errant dust]

Ok, well, they are the exact same size in number of bytes, and I get identical results, whether I use the one downloaded from BFL's Hugging Flux page or the SD3 page.

[sacred jewel]

What about the one in the ComfyAnonymous HF Repo?

https://huggingface.co/comfyanonymous/flux_text_encoders/tree/main

BFL and Comfy one are different SHA256

[errant dust]

Couldn't say. I always used the ones provided in the official Hugging pages. Ex:

The only clip I use that is different is the 32-bit T5 clip in GGUF format

https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main

[sacred jewel]

As per Clownshark, the one from BFL is a slightly better choice as it is different SHA256 than the one from SD3

[errant dust]

Well, I will make 3 images, and if I see a difference by even a pixel, I will look. My first sample is identical.

[errant dust]

I'll try one of those ultra detailed ones in that TomsGuide article

[errant dust]

My use of the 32-bit and 19GB T5 file is questionable. Meaning it is unclear there is any real benefit

[sacred jewel]

19GB T5? Oh boy, I want that one LOL.

[errant dust]

go to the link I shared

it is HUGE

[sacred jewel]

Can you link the link here?

[errant dust]

You need to load it through the GGUF Clip loader

https://huggingface.co/city96/t5-v1_1-xxl-encoder-gguf/tree/main

[sacred jewel]

How the heck can I fit that AND the 23GB Flux.1 Dev ? lol

[errant dust]

They load separately. I mean, if I can use it with my pitiful 8GB Vram, I am sure you can handle it

[sacred jewel]

OK so I then assume the t5 is loaded, used, dumped, Clip_l loaded, used, dumped, Flux loaded, used, dumped?

I thought all had to be in VRAM at once. Or are T5 and CLIP loaded in regular RAM?

[errant dust]

I must assume so, and even then, the GGUF is loaded in blocks, not the entire file. Otherwise it would be quite impossible for me to use it with this laptop 4060

That is the power of the GGUF models

[errant dust]

Heh, reminds me of that famous scene from Close Encounters

[errant dust]

Spielberg was in full genius mode then. Don't know if you know the story, but while he was filming Jaws, every night he was writing Close Encounters in his trailer. One night he shows the 2/3 done script to Richard Dreyfuss, and he was so blown away and enthusiastic that Spielberg rewrote the character (originaly in his 50s) so Dreyfuss could take the role.

[sacred jewel]

I did not know that.

[errant dust]

Yeah, you'd think filming the first blockbuster in history would be enough, but no, the guy has to be writing a new brilliant script at the same time.

[hexed dirge]

https://civitai.com/models/707321

[bitter hearth]

I'd like to think the sign has little paws for people to put their cats behind them and take pictures

:P

well It could be just a sign on the ground but that works >.>

make a little sign like that and put a bowl of food behind it

so when the cat is eating you see him holding it lmao

[coral sable]

Any news about SD3 2B refresh?

[sacred jewel]

In a couple of weeks

The Matrix, 100 years before 😛

oh my... 🤤

[errant dust]

in a fortnight

[sacred jewel]

When your date gets mad at you

[sterile pendant]

Honestly, with how models like flux have shown how well quantization works with dit models, even all the way down to Q4 or Q5, along with fp8, int8, etc, I'd almost bet that SAI will put out a much larger model like 6b or 8b now and just give a bunch of directions and options for quantizations so people can use the version that works with their hardware

[sacred jewel]

I doubt it. SAI might be done for good. The silence is very loud IMO

[sterile pendant]

Transitional periods do that, they might have started a new model from scratch to compete with flux

[sacred jewel]

Remember Blockbuster? Remember RIM/Blackberry? Remember [insert blank] ?

[mortal mesa]

well, new CEO, new CTO, trust and safety person gone, there is action

[desert garnet]

i cant wait to play with the new model while i wait ill go play Concord with the boys

[mortal mesa]

awh i cant i already got my refund

[desert garnet]

its not dead is just a transitional period

[sacred jewel]

This generation got SOOOO many things wrong 😛 😛

[errant dust]

when exactly did this new CTO and CEO take office?

[mortal mesa]

CTO last month, CEO like 3-4 months ago, new board members ect

[desert garnet]

very stable

[errant dust]

Diffusing the situation?

well, the point was more: were they part of the botched release, or trying to fix it, or moved on and will try to best forget it.

[sterile pendant]

Yeah, it will come back out in six months as a f2p game with overly aggressive mtx and BP, and still flop again lol... That or their cave to the degenerate gooner crowd and put waifu characters and skins in the game

[errant dust]

Myeah, say what you will, but the 2B model has serious issues

[desert garnet]

they did had waifus,its just that they were the 600 pound inflation fetish type of waifus

[runic tusk]

Not as stable as a system update from Nintendo.

[errant dust]

and even without being hung on NSFW, I found my images with people completely useless

[sterile pendant]

Ahh I didn't know that. Never played it or even watched a trailer for it, but I heard a ton about it in the news

[desert garnet]

well it was an overwatch souless copy with generic characters and it wasnt free2play so that killed it

[coral sable]

they might think they have to do this, but this isn't true. I like SD3 much better in many situations then any other available model

[sacred jewel]

What issues?

How the heck did the AI figure this out? WOW

[sacred jewel]

oh heck yes!! Imagine the parties

[sacred jewel]

Cascade?

[dusky thistle]

tuned flux

cascade remains the best for artistic styles so far, but the photo abilities of flux are next level

just needs some tuning

[errant dust]

If you look close, you will notice the grass is not the right shade of green

[sacred jewel]

https://tenor.com/view/ace-ventura-pet-detective-jim-carrey-comedy-duh-gif-11333313905875935459

Is that what it is?

This has SOOO much potential for signage and memes 😄

[sterile pendant]

I didn't say whether I was true of false, I was just speculating and throwing ideas out. Sd3 is a solid model in a lot of ways, but since the bulk majority of people are just going to use dev/schnell, SAI will need a model that can actually compete to draw users back. Even huggingface was completely blown away at how quick flux took off.

[sacred jewel]

By exact same, what do you mean exactly? Do you have some examples?

[dusky thistle]

her hair is combed against the part in an unnatural and unconvincing way

[sacred jewel]

You're kidding... 👀 😛

[dusky thistle]

the main issue i've noticed with flux that sd3 also has is a grid-like pattern of lighting artifacts at resolutions beyond 1024x1024

[sacred jewel]

my date is still mad at me 😛

[dusky thistle]

yeah

two differences imo though...

[dusky thistle]

one, flux is actually useful for most things, whereas 2b just has no idea wtf to do most of the time and makes images that are so flawed they're unsalvagable

two, flux is much, much easier to train

[dusky thistle]

i haven't had that experience

nor any real major success with training like i have already had with flux

put a lot of time into it... it's just so stiff

flux is just a lot bigger and/or better trained

flux has zero issues with hands, 2b is a disaster, just as one example

[errant dust]

and text

Absolutely. And he devs who left and reported they were instructed to not deliver or train a good model and to stop development that was well underway. They were all lying

[dusky thistle]

yeah no doubt that whole release was a messsssss

[errant dust]

and they deliberately were told not to deliver a good model

[dusky thistle]

honestly, if they'd just release 8b, they'd be fine

but they gotta do it before the ecosystem develops fully around flux

good controlnets apparently cost like $40k to train

it's gonna be hard to convince ppl to bankroll that for a lateral move from one model to something that's at a similar level

[errant dust]

No, that is not what the devs who left said

[dusky thistle]

so if they wait too much longer, those tools will exist for flux and not for sd3

and it'll stay that way, maybe for good

[errant dust]

I'm guessing I am talking to myself?

You recall Comfy?

[dusky thistle]

yeah idk... matt3o was going to train a higher resolution clipvision/ipadapter for sdxl

and then funding got pulled cuz of flux

[errant dust]

You think Flux happened by accident?

[dusky thistle]

sai would have to help make it happen, if they release their updated sd3 after a lot of good cnets etc drop for flux

yeah i don't think they'd say to trash their reputation on purpose lol

i think the whole thing was a trainwreck of bad communication

[errant dust]

Which project?

I know exactly what he said

He said the SD3 4B was cancelled.

[dusky thistle]

yea, i think 4b was his

[errant dust]

Sorry, let me extend that phrase. The SD3 4B model HE WAS TRAINING

[dusky thistle]

plug got pulled for whatever reason

[errant dust]

I wonder who these Flux devs are. They are awfully clever. Stability should try to hire them

[dusky thistle]

lol

yup

[errant dust]

The 4B model was not cut because of money

It was cut because of a change of direction

The SD guys who founded BFL and made flux and who left Stability, were university researchers. They did not leave because of lack of money either. In fact, the very fact they produced Flux, from scratch in a handful of months shows how ready they were.

[desert garnet]

mama told me they stole the model from sai and published it under diff name 😔

[errant dust]

lol

It is not a little telling that after having all the good public open source projects to be delivered to the community being cancelled, and only a broken 2B model, that they deliver a monster 12B model

It is saying "That broken 2B model you guys got? Wasn't our idea. We were against it and hated the idea of a gated 8B model to just let you wish. Here is a 12B model, the best we can do. Go crazy."

And a few weeks later they close a monster deal with Elon Musk

Side?

[bitter hearth]

this grass is white

terrible model

👍

[desert garnet]

pony side

[errant dust]

lol

Did I now?

[sacred jewel]

It is the absolute opposite for me. SD3 is good, flux is great... no question. Nothing at all fragile in my experience.

[sacred jewel]

too late ... unless it is orders of magnitude better than flux, why would anyone spend all that time, money and energy on a SD3 that is just similar to Flux?

[errant dust]

Variety is the spice of life. But honestly I'd enjoy playing around with SD3 Large/ultra if it weren't so overpriced

right now, the only commercial getting my hard-earned pez is Ideogram 2.0

[dusky thistle]

that's the issue

right now, the big advantage sai potentially has is developing a small model that can do most of what flux can do

but it's gotta be able to outperform the quantized stuff

[sacred jewel]

I got this from that prompt

[errant dust]

BTW, this LoRA is quite nice: https://civitai.com/models/705130/claude-monet-flux?modelVersionId=788725

[bitter hearth]

my hair irl

[errant dust]

In my 20s, my hair had that weird pyramid look like one of the Jesus images. No matter what i did, it would spread to reach the ends of my shoulders in waves

[desert garnet]

wait a min thats not anime

[errant dust]

Oh crap. My Bad. Worse: the people have clothes on

[desert garnet]

https://tenor.com/view/outa-here-gif-19102905

[bitter hearth]

No cat ears, no Steven Seagal

[errant dust]

(lowers head in shame)

[sacred jewel]

By itself, Flux knows Monet... but not his style ...

[errant dust]

yeah

[bitter hearth]

I liked when my hair was at shoulder length

[errant dust]

not quite there

[bitter hearth]

It is now at butt

[sacred jewel]

I figured you were hairy all over... but maybe you're one of those Sphynx cats 😛

Not quite. Actually, not even close... There are things Photoshop does that no other program does.

If you mean SAVE AS JPEG, sure, MS PAINT can do that.

[bitter hearth]

oh my god

@sacred jewel you gave me a terrible idea

now have the image

[sacred jewel]

There are also licensing constraints. AWS has their own Slack/Teams/Zoom thing as does Verizon... they all suck dirty ass but hey, someone HAS to buy it because they can;t use Microsoft or whatever due to licensing or conflicts

[sterile pendant]

Nah, it's not... It likely has quite a few major changes in the architecture and was also likely trained from scratch. You aren't a part of either dev team,

Don't spread BS like that lol

[sacred jewel]

https://tenor.com/view/cat-hmph-pft-pfft-pffft-gif-4447083553279693800

[sterile pendant]

Testing waifus != Knowing how they're training a model

Again, making some images with an API doesn't mean you know anything about what's going on behind the scenes

[sacred jewel]

LEads me to a question I've had for a bit:

Are there any such things as Easter Eggs in any Diffusion model these days? Would be awesome to have that. I miss Easter Eggs.

[sterile pendant]

so you had a local copy of the model and intimate knowledge of all of the training methods/scripts/dataset/etc to be able to boldface lie and say that BFL's flux models are SD3, just "bigger?"

you're trippin bro... calm down lol

[sterile pendant]

so you used an API...

if you didn't have direct access to the files and data, without being gated, you were using an (A)pplication (P)rogramming (I)nterface to use them...

that you had to access...

Yeah, you're describing an API... Your comfy client would be granted access to the files for use, that weren't locally accessible

[turbid grotto]

And being faster!

[sterile pendant]

So stop lying and acting like you're in some kind of "know" about the actual things/training processes that went into the models behind the scenes and stop claiming that Flux is just SD3 but bigger and that they just took it and ran with it lol... It's BS and making up that kind of BS does more harm to the community than good.

[bitter hearth]

I am not paying a single gram of attention (1 cubic hair length of toes in the air for americans) to discussions in this channel anymore, have more ugly cat

[sterile pendant]

Well you can keep deluding yourself however you want, but it's ultimately up to others reading these messages to formulate their own opinions. I'm just calling it out and dispelling BS rumors/conspiracies as I see them...

Flux != SD3, so stop saying that.

well you sure seem hot and bothered by it, but you can't powertrip here and it likely drives you nuts now (saw someone on the subreddit mention that you're a mod there now and that you've been powertripping or something)

[sacred jewel]

Ollama Bin Cattin'

[sterile pendant]

calling someone out on lying isn't power tripping. you're just embarrassed that you got called out. so again, stop lying and trying to act like you're in the know of how these models are made and stop claiming flux is just a bigger sd3.

touché

where's the alleged proof that flux = sd3...

and i already proved that you didn't have access to the models/training/etc etc

cool, salty devs make false claims all the time

[desert garnet]

well i saw the mod drama on reddit and i can confirm it was juicy

[sterile pendant]

i didn't really read any of it, but i know it was a thing

[errant dust]

so here are two images, same prompt and seed. The first is plain Flux, and he second is with the Claude Monet LoRA.

[sterile pendant]

not bad, i like it

[sacred jewel]

that's pretty cool.

[sterile pendant]

surprised how well it works even on all the shit in the background

mimics traditional art where you desaturate subjects further away and all

[errant dust]

Yes, it is a well-done LoRA. I was impressed. https://civitai.com/models/705130/claude-monet-flux?modelVersionId=788725

[bitter hearth]

that's really nice

Flux can do soft styles now

[balmy kettle]

donald trump is fighting with narendra modi and both have punching gloves

[sacred jewel]

On twitter? or on live TV?

[balmy kettle]

on facebook

[balmy kettle]

not good they both are fighting with each other

[desert garnet]

that was yesterday,today they kiss

[desert garnet]

https://tenor.com/view/tyson-smile-gif-10959502

[sacred jewel]

https://tenor.com/view/trump-pizza-donald-party-gif-23969352

[balmy kettle]

this is more powerful

add that modi has hit trump's face badly and now it is bleeding and his teeth are broken

[sacred jewel]

John Kenn Mortensen LoRA

[bitter hearth]

Looks like don't starve

[sacred jewel]

Movie Poster Style LoRA

[alpine summit]

Sd1.5

[errant dust]

CivitAI LoRAs for Flux just exploded. In the category of Style alone there are already thousands of LoRAs. I don't say thousands as a euphemism for a lot. I mean literally thousands.

[bitter hearth]

I managed to hide the seams lol
this is 3 images

seams happen in the background so the first pass uses depth maps to blur the shit out of the background

now the seams are gone

[bitter hearth]

wow there are thousands?

that's amazing

[errant dust]

No joke. I was trying to see them and got tired of the endless flow. So to simplify, I asked it to show me all LoRAs with the letter A, in the category of Style

it told me thousands

I was quite shocked

[bitter hearth]

I had no idea

[errant dust]

This isn't the sum of all the types, just kinky, or anything. No I asked only for the category Style

it was ENDLESS

[bitter hearth]

I tried to skip flux and focus on SDXL but I think flux is too good to skip

cos it clearly trains well

[errant dust]

And this is after only a single month. Less if you mean the time it has been trainable as it is now

[bitter hearth]

I was thinking
if the model gets smart enough with finetunes and tooling
we might just not need negatives

after all negatives do divide your speed in half

[errant dust]

Myeah, there are things you might want to be able to control. Background blur, or 'bokeh', is one

[bitter hearth]

there's nodes I like that need negative
SAG blurs the subject of the image and pushes the model away from blurry subject
PAG does something similar but with degraded subjects

but if the model is smart enough we can get by without these

[errant dust]

Yes, you can try a positive with camera lens references, but Flux has been known to add blur even in non-photo scenarios

[bitter hearth]

IP adapter takes negatives as well
is probably a better example

of negatives that are not prompts

yeah checkpoints will have to fine tune in more depth of field

I assume it will be the same situation as with SDXL where Jugger funded by rundiffusion.com and Realvis funded by Mage.Space are dominant

[errant dust]

But this also shows just how appreciative the community is for the gift of this incredible 12B model. A greater demonstration of love I cannot think of

[bitter hearth]

yeah lora is legit the best way to show appreciation

it kinda shows that people were hungry for something too

[errant dust]

exactly

All the quantization models, the LoRAs and so on really highlight how much it is appreicated

[bitter hearth]

I'm looking at some of the character loras
its learning faces better than any model I've seen
and it doesn't even have a proper IP adapter yet, its just straight out of model

[errant dust]

When it came out, the naysayers were desperate to paint it as a bit of a unicorn: untrainable, requiring an insane amount of memory, etc. Now look at it

[bitter hearth]

its a bit of a complex issue
there are still some costs to it being distilled, namely the worse guidance and the lower image diversity
and then the second issue that the base model was slightly overtrained on one aesthetic

but since the image quality is very high it is still a useful model for now

[errant dust]

I guess we will just have to make do

SIGH

[bitter hearth]

I think I'm staying on SDXL and waiting for the next Pixart for now

was a good sign that they are backed by Nvidia now

[errant dust]

A wise decision. It also protects you

[idle maple]

паз лейбл

[dusky thistle]

i think it's best to use multiple models

they all have their strengths

[bitter hearth]

I started using TCD lora

its a bit like lightning, its really good

its the latest one, as far as I know, possibly hyper came slightly after

TCD can do both low steps as well as high, which is cool

[bitter hearth]

yeah I worked out today that with SDXL you get at most 5 passes
before the VAE compression kills you
but 5 passes is loads, that could be 5 different models

[dusky thistle]

oh i don't mean within one wf

i jus tmean, in general

good to use multiple models, mutiple samplers, multiple methods, multiple trainers, etc

[bitter hearth]

ah yeah that's what I thought you meant

[dusky thistle]

nothing wins at everything, at least not yet

yea

[bitter hearth]

I started doing composite so each panel could be a different model like this image

just have to use nodes that help hide seams

multidiffusion, tiled ksampler or mixture-of-diffusers help

I always hate when people post Ultimate Upscale images and you can see the grid lines

[dusky thistle]

best thing to do is use overlapping gradient masks with differential diffusion

seams are impsossible with that tiled approach

[bitter hearth]

oh thanks will try that

my method still had seams it just hid them under a plant lol

[dusky thistle]

lol

yeah the idea is to do something like... for example, say you are upscaling to something that is 1024x1536

you'd have two 1024x1024 tiles

half would be solid white masked, the other half of it would be a gradient to black

then you flip it around and do the other side

so the total mask value is 1.0 across the board and no more no less

[bitter hearth]

Poor rocks

They all died because of the low sea level

[muted dove]

These are all Flux, and all without severe bokeh

[bitter hearth]

ah this sounds great

I've been trying to learn noise injection properly cos I've been doing it "wrong" the whole time

doing noise injection using k-sampling, with correct sigmas, is better than what I do

if you add noise in pixel space and VAE encode it doesn't quite work properly

blepping showed me on the civit discord. VAE encode cannot make a latent with high enough values

[dusky thistle]

Ahh yeah... I've always done that stuff in latent space so I hadn't checked on that, good to know

[muted dove]

That river dried up real quick!

[alpine summit]

Sd1.5

[noble coyote]

Florence2/Flux img2img with SD3 input images (SD3 first, then Flux)

Florence2/Flux img2img with SD3 input images (SD3 first, then Flux)

[alpine summit]

Sd1.5

[noble coyote]

Florence2/Flux img2img

[alpine summit]

Sdxl

Flux

Sd3

[clever bone]

TEST

help

[pearl forum]

TEST

[noble coyote]

Florence2/Flux img2img

[errant dust]

https://venturebeat.com/ai/stability-ai-looks-to-grow-stable-diffusion-text-to-image-ai-usage-with-amazon-bedrock/

[sage burrow]

Too many ads, didn't make it past the first paragraph

[dull star]

excellent, so SD3 8B weights are never seeing the light of day

welp, time to wait for more flux painting loras 😴

[errant dust]

https://pocket.co/share/c9f0c061-b4f6-427c-b3a4-0d63e4665555

Try that instead. What I do when I encounter this problem of too many ads is I save it to my Pocket account, which is free by the way, and it removes all the ads

[errant dust]

You mean more than the thousands that are already available?

[dull star]

I need to check them out one by one and I cannot choose

I might use one of Oofa's

[errant dust]

That wasn't a euphemism. There are literally thousands already available

At least on Civit

[noble coyote]

It came across that SAI aren't completely sure of the way forward; and have a touchy-feely attitude to the way forward?!

There was a certain amount of reserve in the tone of the article.

[errant dust]

One interesting tidbit was that they won't add SD3 Medium, but rather SDXL instead

[bitter hearth]

that is noteable yeah

[sage burrow]

Make your own 🙂

[errant dust]

Or just flux them down the toilet

[noble coyote]

If ClipDrop is 8B weights, then its photorealistic output is phenomenal

[dull star]

good idea, I'm gonna train my own 8B MMDiT

[sage burrow]

Bc most people only want to read articles woth drama, so it's always spun that way.

[noble coyote]

Its the lack of drama in the tone of the article which is more surprising - perhaps the journos are aware of SD3's history of false-dawns?!

[errant dust]

More like they have no idea what to make of it, and what it might mean towards a larger better business plan. So they just report the facts.

I know I have no clue.