🆕｜sd3

[torn wharf]

Mr Tumness. I'm old enough to know that reference.

[noble coyote]

I've just tried a Satyr - he ain't half got a massive amount of public hair!!! Perhaps not best for on here ...

[torn wharf]

styles

[mortal mesa]

is the comfy all in one flux any better on ram or same as the fp8's separate, probably same ya?

[sage burrow]

Wow, flux does taurs way better than any other model! 😄

too lazy to fix the img, but it wouldn't take long

[noble coyote]

Female Satyr - not really!

It's forgetting to add goats legs - the horns are present ... ?!

[icy drift]

I was doing painterly-style cards too, but I gave up. They were just turning out way too steamy.

Perfect! 😊

Monster Card Series V2
23 - Centaur

[noble coyote]

My female satyrs are quite racy!

[sage burrow]

are loras and checkpoints even going to be needed anymore?

[craggy crest]

what ever he posted, he deleted it

[mortal mesa]

satyr only has to have a tail and a hardon, no horns hoofs ect thats all newer depictions

[icy drift]

I've been thinking about it since yesterday, and I still don't have a solid usecase for either. Definitely controlnet though.

[noble coyote]

Flux does not realise that a pair of goat legs are needed to give the complete Satyr!!! 🙂

[icy drift]

I remember Faerie Queene mentioned "backwards-bent knees" so no too recent. Maybe you mean Greek pottery paintings. Don't remember those in detail.
But a girl satyr will not have that one thing in particular. 😛

[sage burrow]

perhaps some Claude 1st to help?

I jprompted a satyre and got a nude lady with horns (ps flux an't do nips to save its life)

[noble coyote]

I'm beginning to understand that a Satyr has a prodigious amount of tumescent flesh!!!!!! Mr Tumness was tame 😄

[sage burrow]

ty glif and gpt4

[torn wharf]

Been thinking of Legacy lately. Stability legacy now roosts at BlackForest Labs. I thought that the new CEO of stability had Weta legacy, the studio that worked on chronicals of narnia. But he was CEO of weta digital. A software offshoot of weta workshop that he built up only to sell off to the highest bidder. Unity bought it and now are squandering it. Typical enshitifcation. He came on, puffed up the value then had the company acquired, and left. Now he's brought on to stability.ai . hrm. legacy.

[icy drift]

Where is BlackForest's Discord???

[torn wharf]

good question

[sage burrow]

Flux is interesting lol

[bitter hearth]

Not even penguins escape the furry

[sage burrow]

LOL

[torn wharf]

interestingly enough, last week the antarctic had a catastrophic warming event. small pockets were warmed to 30 degrees. nobody knows why.

The sun felt hotter that week imo. sometimes i think global warming isn't entirely because of CO2. But that makes me a climate denier or something and people lock me up

i'm kind of a sun worshipper and accept that it is the supremely dominant entity in our world. 99% of the entire solar system's mass. right there in that big hot thing.

[bitter hearth]

Can flux do a cow abducting an UFO

[uncut river]

crazy colored chaotic cartoon style, cow being abducted by an alien spaceship, blue volumetric spot light on the cow

flux-schnell

[bitter hearth]

No

The other way around

[uncut river]

ohhw, wait. I didnt read that correctly

[bitter hearth]

[uncut river]

crazy colored chaotic cartoon style, spaceship being abducted by a cow, blue volumetric spot light on the spaceship, cow floating on top, spaceship on the ground

[torn wharf]

That is a solid cow abduction illustration though. applause

[icy drift]

Monster Card Series V2
24 - Pegasus

Done for now. Might make more later.

[sage burrow]

can be both

[torn wharf]

could be. But the sun is one big bitch.

[icy drift]

I'm gonna get these printed. 😊 This is the first time I've been sure the quality was good enough to warrant it.

[torn wharf]

An illustration of a farm landscape at night time. A cow hovers in the air at the top of the image. Emitting from the bottom of the cow is a volumetric blue light cone. On the ground within the light cone is an alien UFO with landing gear extended. it looks as if it is being lifted off the ground. @bitter hearth

[sage burrow]

I need a workflow, Flux main character, then SD3 background

[brittle nexus]

Missing a Lamia. Pretty classic monster girl

[torn wharf]

An illustration of a farm landscape at night time. A cow hovers in the air at the top of the image. Emitting from the bottom of the cow is a volumetric blue light cone. On the ground within the light cone is an alien UFO with landing gear extended. it looks as if it is being lifted off the ground. crazy colored chaotic cartoon style threw that style vibe on it

[torn wharf]

its so dumb how entertained i am by this

[bitter hearth]

Lmao

10/10

[craggy crest]

sure. let's say you want to be able to generate images with your cat. the AI knows cat, but doesn't know YOUR cat

[mortal mesa]

maybe there is a world where you add in your company branding via lora

[torn wharf]

And now for something completely different. My hen after a hawk dove at her this morning

[sage burrow]

A grasshopper centaur? Flux: I can do that

[bitter hearth]

Now THIS is my type of prompt, forget all these llms and techniques, behold:
Super powerful cow exodia

[bitter hearth]

i have enough nightmares already

[sage burrow]

Would you like a link to my horror DA? 😄

[craggy crest]

let me guess, it's packed with sugary sweet fluffy bunnies?

[uncut river]

slight warning here, think twice

I once forgot, and got hurt. saw things.

[sage burrow]

True story, I also have a chibi account 😄

[uncut river]

images and characters can be very flexible on the bio-gender, let's put it like that. Lovely style, but not really my content...

take that as a compliment becky93!

[sage burrow]

You are thinking of my main account, my horror account is actualy horror (and mostly sfw)

[uncut river]

ohh, ok. nevermind then.

[sage burrow]

Anyone that does DA, the trick is to keep different genres to different accounts, even if it's to not traumatize viewers lol

Btw, horror is extremely popular for views, but no one buys it ever 😦

[bitter hearth]

[uncut river]

I use to love horror, but im not an active fan anymore

[brittle nexus]

Lovely "tired of your shit" face

[sage burrow]

I really really like surrealism 🙂 Also post apocolyptic odd creatures is fun 🙂

[bitter hearth]

I'm eating sliced sweet potatoes, metaballs and eggs

they are all ball-like

[sage burrow]

It's pretty awesome that we can create our own custom art, thousands of it! (then print it out for $1 per page and frame it, if desired)

[noble coyote]

Mr Tumness! (Tum e s ) centaur

[sage burrow]

it doesn't know what rollerskating is 😦

nor rollerblading

[brittle nexus]

First model that doesn't make a nightmare fuel from that

[icy drift]

Flux brand piano. 😉

[icy drift]

Gonna use a card-printing service. I think I need more cards though... Only have 24.

[bitter hearth]

He's so angry

how dare you put me in the snow

[brittle nexus]

To be honest, cascade did a good job too

[rain current]

Even better, because those hands are used to play the piano, and to scare the children

[bitter hearth]

Just like fighting games

Hands are important, make it bigger

[torn wharf]

i bet she likes rusty spoons if u know what i mean

[lucid swift]

flux

[frail shoal]

this seems to be a very useful feature for schnell, it's already fixing anatomy in the default settings. I have to explore a bit these settings.

[brittle nexus]

That was upscaled and the prompt wasn't quite the same, but yes, it's pretty good

[frail shoal]

the hand was missing fingers, and as the steps went, it fixed it. But it changed the bacground a lot. Using euler btw which should not do a lot of change in each step. I need to find the sweet spot of fixing anatomy but not changing the overall picture a lot in each step.

[rain current]

Channel dev? I only have the shift on the default channel

[frail shoal]

idk what you mean

i'm using schnell

[rain current]

comfyui

[frail shoal]

still don't

[brittle nexus]

There's a flux official discord?

[rain current]

👇

[torn wharf]

struggling with this prompt idea i had. crayons melting into a painting. i'm hitting notes but , its not coming together. mmdits sure are fun to prompt though. and its magical when they work.

lower steps kind of helps

very pretty. i'm noticing it rarely understands what a melted crayon is supposed to look like in this particualr case.

[bitter hearth]

for painterly I always like lower steps euler

with overly low CFG

you have to perturb the model a bit

if CFG++ is working for Flux then that helps a lot with softness

[dry wave]

the issues with paintings is my biggest issue with flux so far. Hope lora fine tunes will solve that

[torn wharf]

am using dev. i hvaen't really tried other workflows yet. i saw one on civit called flux street.

i hooked up cfgguider for negative pormpting and cfg, but any cfg ruins prompt comprehension

[dry wave]

flux don't have real cfg and I don't think it works well with cfg (haven't tried yet, though)

[bitter hearth]

we rly need the paper

[torn wharf]

i saw an immediate issue using cfg so i swapped back to the basic guider

[dry wave]

dunno if there is much new in a paper

most stuff is known I guess

[torn wharf]

15 steps low as i can go for eular. in general really

[bitter hearth]

is it not novel that it doesn't take CFG?

[dry wave]

from sd3 paper, the distillation paper and the source code

they distilled the cfg out of the model

[mortal mesa]

got these

[dry wave]

similar like turbo models distilled the high step count out

the model is basically finetuned to produce images that look like images with cfg

[bitter hearth]

I don't think this is it
necessarily
I think it has a different block structure

[dry wave]

the source code is open

it's block structure is not that different

they have mmdits and dit blocks mixed as in Auraflow

they use parallel transformers - what sounds fancy but is a very simple trick to make it faster

in contrast to sd3 they don't use CLIP embeddings anymore

instead they only use the pooled embedding of CLIP as micro conditioning together with cfg and timestep information

I'm very sure the pro model still uses cfg

[bitter hearth]

ok thanks I haven't looked into it yet

[gusty trail]

guidance-distilled variant

[bitter hearth]

with SDXL I've tried a few different things like
injections into self attention, into cross attention and into key matrix (out of the K V M matricies)
maybe some of these would be useful

[torn wharf]

https://github.com/comfyanonymous/ComfyUI/commit/ce9ac2fe0581288d4a24869dae2e04a3c2b67061#rssowlmlink did anybody catch this? We were using flux in a haggard half working state for a few moments on the first day unless you updated after this happened

[bitter hearth]

I am thinking that
if the model has worse guidance due to distillation
then attention map injections could be a really great idea

i.e. dense diffusion / Omost style

[dry wave]

I don't think so

it's probably just a variable naming thing without effect

[torn wharf]

it wasn't prompting the clip layer just the t5 layer

[dry wave]

you cannot accidentally switch clip l with clip g

[bitter hearth]

injecting noise into the key matrix also works for some reason
because the CADS node has that
but I don't know why it works

[torn wharf]

flux has no clip g

[frail shoal]

i get very good quality and softness in schnell, i'm just trying to fix the missing fingers with the model sampling node. Basically each step tries to follow prompt more form what i understand and based on the settings is how much the image changes in each step. This is what i understand this far.

[torn wharf]

iit was just going no where so the code didn't fail

[bitter hearth]

there is a trade-off for soft images
in my limited experience with softness, every technique that makes it soft lowers prompt adherence and structural coherence

[torn wharf]

🔥 fire 🔥

[icy drift]

Monster Card Series V2
25 - Satyr

[frail shoal]

make it soft with artist styles in clip l

[bitter hearth]

this is my favourite soft image I have made
it was Euler-ancestral-CFG++ with a CFG of just 0.1
and then not given enough steps to fully converge

not giving it enough steps helps a lot for that soft wispy quality

this is the same reason why, unlike most people, I run realistic photos for like 300-1000 steps sometimes, depending on solver. I have seen things get fixed at over the 300 step mark.

[acoustic kite]

use fp16 if you can

[grand wyvern]

is that sd3

[bitter hearth]

its zavychroma sdxl

[torn wharf]

same prompt using replicate's flux pro interface

i think they run it full precesion

[robust junco]

very nIce

[torn wharf]

its a really popular sport

[lucid swift]

flux

[rain current]

Schnell

[craggy crest]

https://x.com/EMostaque/status/1819776928108409080

Emad (@EMostaque) on X

Now @bfl_ml has released Flux.1(dev) 12b @StabilityAI should release SD3 8b

It’s the benchmark model in the paper that performed so well (including laying down on grass aha) & is much more tunable.

Glad to see releases, onwards we go..

Twitter · Aug 3, 2024, 4:46 PM

[dull star]

fr 8b should come out

good paintings

easier to run

etc

[craggy crest]

emad just tossed down the gauntlet. we'll see what they do

that thread is interesting

[pseudo owl]

Yeah 100%, and a turbo variant would be amazing too. It could probably redeem them.

[dull star]

if 8b is faster than flux 12B then I don't even want turbo, but I can understand the amazing advantage for lower end users

[craggy crest]

Flux PRO via Repilcate

[pseudo owl]

Yes 8b is going to be faster. I still dont want to wait like 1 min for 1 image so I would definitely like turbo.

[rain current]

Schnell

[proven pecan]

SD3 (it's on purpose)

[torn wharf]

if i can run 12b that means i can run 8b right?

been trying out some imagen prompts. A single beam of light enter the room from the ceiling. The beam of light is illuminating an easel. On the easel there is a Rembrandt painting of balls.

imagen version

it's class collapsing on dragonfruit, but it kinda knows pitaya and then really nails it with pitaya fruit. prompting is king.

imagen prompt is A dragon fruit wearing karate belt in the snow.

[pseudo owl]

Yes and it will use considerably less vram.

[torn wharf]

so far the messaging about 8b is that you'll need over 24

[dry wave]

only for 32bit which nobody needs

bf16 should be 16gb

[dull star]

I remember lykon saying that it's not that much slower than 2B, but we'll see

[torn wharf]

flux vs imagen

[dull star]

yeah imagen imagens 😔

[torn wharf]

dog looks in a mirror and sees a cat

aminals!

[bitter hearth]

the 8B isn't enough to match a 12B necessarily
however that also isn't necessarily the case

[pseudo owl]

Flux dev is better, schnell with 6 steps is similar to 12b.

8b is going to be faster tho and save vram as well.

[bitter hearth]

yeah I'm probably waiting until SD4 until I use a Stability model again

there is definitely a need for low VRAM models

in the same way as there is a need for 3B LLMs

but that doesn't interest me that much

[pseudo owl]

3.1 might be decent but they are still trying to compare how good sd3.1 images of dogs are.

They should really test humans first.

[torn wharf]

humans are proabbly a big part of their testing procedures. what they published is just a small lens on it all

new leadership is at the helm now so just hope better decisions are going on

[bitter hearth]

they could compete on license

cos that's a way to make use of a model that is not quite as strong

you can release a more free license

not open source but less restricted

[torn wharf]

jump to sd5. 4's a dumb number. not even prime

[bitter hearth]

or they could just market it as a lower VRAM alternative

Mixtral got extremely popular by being lower VRAM than 70Bs

Mixtral did way better than the average 70B

[pseudo owl]

Schnell is Apache 2.0 which is basically the completely open. Dev is definitely a lot more closed

[torn wharf]

transformers change everything imo. it's a whole new ball game

there are developments to be made

[dry wave]

transformers are there since 2016.
Even SD 1.5 is a transformer architecture - but we had this discussion already...

[bitter hearth]

sadly Schnell looks too bad to me relative to Dev

I tried to like Schnell but I can't

its better to talk about things in terms of individual layers anyway
because that side-steps the debate about what is, and is not, a transformer

[torn wharf]

its a pedantic debate.

[dry wave]

models get better and better since the initial dalle 1. But these are incremental updates, no revolutions

[pseudo owl]

You have to try 6 steps, it’s a lot better then. But dev is still better but it’s close.

[bitter hearth]

ok thanks will try varying steps

will also try adaptive solver on schnell, which I have not tried yet

I use adaptive solvers like 99% of the time aside from quick test images

partly because it saves the mental effort of having to optimise scheduler or steps

[pseudo owl]

4 steps is the right one, 6 steps is left. Prompt: “An image of a woman holding her hands up in a room”

[torn wharf]

2020 i couldn't do any of this. 2024 we doin it. that's very revolutionary. french revolution took 10 years. American Revolution took 18. Sometimes revolutionary periods last a while. This is a huge shift in digital graphics things won't settle down for some time.

[bitter hearth]

ah thanks thats big change

[pseudo owl]

Random seed tho, but 6 is still considerably better most of the time.

[torn wharf]

One day we'll have the true revolution. Canada will take over it all and then you'll all be sorry

[bitter hearth]

its mostly that Schnell is the opposite type of model to the one I need
Schnell is about making images fast
but I want models that try to make good images slowly

[hallow lion]

What a time to be alive!

hold on to your papers.

[torn wharf]

now imagin, two papers down the line!

[pseudo owl]

Yeah dev is definitely better then.

[torn wharf]

has 2min done something on flux yet?

all the youtubes i found so far, the guys just read verbatim the bfl blog post

[hallow lion]

yeah not very cretaive lol youtube is like big media now just reporting the news basically

in whateve rniche they operate in

[bitter hearth]

the popular youtube channels for machine learning are ridiculously bad sadly

both for LLMs and for image gen

[torn wharf]

papers! get em!

[bitter hearth]

the amount of youtube videos pushing brand new multi-agent frameworks is over the top

having said that, some of my favourite comfy workflows came from youtubers so it does vary

[torn wharf]

oh NO it's colappsing into a singularity!!

[bitter hearth]

memes are so aesthetic now lol

[torn wharf]

thererres tooo many pappperrrs aborrrt

[fair summit]

someone know about codes in collab and maybe can help me fix a problem?

[pseudo owl]

What problem?

And also move to tech support

[fair summit]

oh ok

[torn wharf]

guess this is it. we're in the papers now. gotta just push through

now the big bang. thats it. thats the new paperverse.

what a time to be alive

[fleet meteor]

Epic

[craggy crest]

which of those 2 vases of roses is flux, and which is SD3

[brittle nexus]

I would say first one flux but it is too small to be sure

[torn wharf]

the future is now

[craggy crest]

click on the image, then click on open in browser, then you'll get the full sized image not the discord compressed version

[brittle nexus]

Still a 331 x 579 image

[torn wharf]

so the uninitiated dont think i have a demented imagination. just demented inspiration. here's the og short from don hertzfeldt's rejected https://www.youtube.com/watch?v=UcwfEMdV-aM

[dull star]

don hertzfeldt is the best

[torn wharf]

his new stuff is phenomenal

[bitter hearth]

im trying out single checkpoint version of Flux ... not using any clips in the workflow

a 16gb safetensor which is fp8 version of the dev model

20 steps and its kinda slow on my rtx 3060 gpu with 32gb ram compared to schnell with 4 steps

[acoustic kite]

i swear there's a huge difference in amount of samples when using fp8_e5m2

[bitter hearth]

what do you mean?

[acoustic kite]

fp8_e5m2 is closer to fp16 but needs more samples to look good

[bitter hearth]

won't that be incompatible with the t5 clip?

[acoustic kite]

no it works fine 😄

[bitter hearth]

interesting, im gonna try it out then

[acoustic kite]

okay

but what i meant is that any amount of samples over 20 actually helps the image a lot

lots of people were saying that anything over 20 samples is not needed

[bitter hearth]

oh, im using 4 step version of the flux model

schnell

[acoustic kite]

ohhh

[bitter hearth]

but im rendering an image now with that setting you mentioned

[acoustic kite]

i'm gonna upload an image with 120 samples in just a sec

[bitter hearth]

ok

[acoustic kite]

and then another one with 22 samples

for comparison

120 samples

22 samples

40 samples

[bitter hearth]

the top one of the 3 looks kinda more vivid in tone

[acoustic kite]

and the fence on the second floor is completely fixed

as well as other small issues

but idk if 120 samples is worth it 😂

steps

[bitter hearth]

if you got a beefy gpu sure

[acoustic kite]

i'm gonna upscale it 4x and send it here

[bitter hearth]

im on rtx 3060 so can't use the dev version

[acoustic kite]

ohhh

how much vram?

8gb?

[bitter hearth]

12gb

[acoustic kite]

nice

[bitter hearth]

120 steps is fine I often do more than that

[acoustic kite]

i wish it was easier to run

[bitter hearth]

it takes from 3 - 4 mins to render with dev

[acoustic kite]

here is the difference between 60 and 120 steps btw

[bitter hearth]

yeah sometimes stuff gets fixed at higher steps

[acoustic kite]

1st is 120 steps and 2nd is 60 steps

[bitter hearth]

I have seen stuff fixed hundreds of steps in

yeah details are much better on the left one

[acoustic kite]

wait my bad

right one is 120 steps. really makes you wonder

[bitter hearth]

left one is better

[acoustic kite]

60 steps

i'm not even kidding

[bitter hearth]

the right one misses some of the detail with blur

unless you are confusing which is which steps lol

[acoustic kite]

no i checked again and it's definitely the less detailed that's 120 steps 😄

[bitter hearth]

yep

[acoustic kite]

but the less details, the less room for error as well

when upscaling

[bitter hearth]

the left one and right one .. see the shade on top.. left one is pretty distinct

[acoustic kite]

that's where i like the 120 step one better, the 60 step version looks like it's trying to make chimneys on either side which wouldn't make sense

either way it's pretty interesting how the saying that "anything above 20 steps doesn't matter" isn't remotely true

[bitter hearth]

lol i see what you mean, but still left has more detail but a bit extra haha

for quality tho, the right image has blurry spots

[acoustic kite]

this one is 22 steps btw

just for the hell of it 😂

[bitter hearth]

60 steps looks better

but 60 steps would also take a bit of time

can you share that prompt?

[acoustic kite]

sure man

[bitter hearth]

i wanna see what i get with 4 steps

using schnell

[acoustic kite]

facade with apartments, outside view, viewed from the front

[bitter hearth]

i mean just the prompt

not the workflow 🙂

i got my workflow with schnell setup

[acoustic kite]

the prompt is above

like in the message

facade with apartments, outside view, viewed from the front

here it is again 😄

thought you would want the seed which is why i shared everything

[bitter hearth]

nah just the prompt for the heck of it, just wanna see how schnell handles it

not trying to replicate

[acoustic kite]

haha fair

pls share, now i'm interested

[bitter hearth]

yes but i can't find the text prompt lol

i scrolled up

can you link it again ?

[acoustic kite]

facade with apartments, outside view, viewed from the front

[bitter hearth]

ahh ok

[acoustic kite]

i'm upscaling it 4x with sdxl lightning

i'll share soon 😄

[bitter hearth]

lol ..

[acoustic kite]

oh the tree is very conveniently placed 😂

here is 4x upscale with sdxl lightning, now i'm gonna try sdxl normal

[bitter hearth]

i see you are upscaling the images too

you must have a pretty good gpu!

[acoustic kite]

it's an rtx 4070ti super 😄

[bitter hearth]

my system would probably freeze if i attemped to upscale above 1500

[acoustic kite]

16gb vram but it almost isn't enough lol

[bitter hearth]

ah nice 👍

[acoustic kite]

yeah that's honestly completely fair

my pc freezes for a moment when trying to upscale

[bitter hearth]

a blonde girl sitting on a bean bag, indoors, potted plants, window, sun light.

the keyword girl generates little girls

[acoustic kite]

oh

[bitter hearth]

the thing about longer steps is
you might need a different sampler to normal

[acoustic kite]

female is not age specific

try female

woman is more age specific as well i would say

female is more neutral just as male is

[acoustic kite]

oh i only saw this now

maybe i'll try a couple hundred steps on the facade pic

[icy drift]

You can specify the age: "21 y.o." or "34 y.o." etc

[acoustic kite]

the thing about upscaling the image is that my roof tiles are getting messed up

[acoustic kite]

does that work?

epic

[bitter hearth]

trying

[acoustic kite]

ok 😄

[bitter hearth]

whoa, nice call

[icy drift]

Monster Card Series V2
No. 26 - Kappa

[acoustic kite]

looks cozy lol

[acoustic kite]

oh wow you're making a lot of cards

is it for a card game?

[bitter hearth]

her feet are kinda fused lol

[icy drift]

I'm aiming for a full deck of 55. (That's how many the printer service wants.)

[bitter hearth]

but cool thing to know about female instead of girl 🙂

[acoustic kite]

i wonder if male works as well

[icy drift]

Not a game, but I am going to print them. Could make it go fish or something, but nah.
Once AI is good enough that we can consistently describe everything on the card, I'll try a game. 🙂 (And Flux is very close to that level.)

Monster Card Series V2
No. 27 - Phoenix

[bitter hearth]

definitely should, lemme try one

[acoustic kite]

would be super cool to eventually make a hearthstone inspired game using generative ai

and llm's for the programming

[icy drift]

Monster Cark Series V2
No. 28 - Basilisk

[icy drift]

And the playtesting balancing!!! That's the hardest part of card game design by far.

[acoustic kite]

yes!! don't want the cards to be too good

[bitter hearth]

slight glitch in the anatomy but face came out alright

[icy drift]

(This one turned out mostly just a lion, not a girl... Still cute though.)

Monster Card Series V2
No. 29 - Chimera

Monster Card Series V2
No. 30 - Jackalope

[acoustic kite]

i don't know how to fix these tiles :(

they're all jumbled

[icy drift]

Is that Dev? I thought this model had beaten the line soup once and for all. 😦

[acoustic kite]

it's the upscaled version of the image using sdxl lightning

[craggy crest]

the roof?

[icy drift]

Merrow is an Irish mermaid with a magical red cap.

Monster Card Series V2
No. 31 - Merrow

[acoustic kite]

yeah if only i could upscale it and still have perfect roof tiles then it would be perfect

[craggy crest]

that's how those roofs really look

they are jumbled. they're not in straight rows

[icy drift]

Maybe try upscale with Dev? But larger images take forever .

[icy drift]

Monster Card Series V2
No. 32 - Troll

[acoustic kite]

my pc is gonna hate me lolll

[craggy crest]

okay, the tile details aren't right, but i don't think the AI is going to fix that. you'd need to train a lora on real roof files

[craggy crest]

try magnific

[icy drift]

I still haven't been able to get a good 4K render. I think I need to upscale 1K -> 2K -> 4K. Hopefully I'll remember to leave that rendering tonight.

Monster Card Series V2
No. 33 - Cthulu ...(Isn't this just Kraken again though? Oh well.)

[acoustic kite]

maybe that approach could improve my roof tiles as well. thanks for the idea

so far it seems like 512x512 has the worst tiles, 1024x1024 is better and 1536x1536 is the best. i'll try to set each tile to 2048x2048 and upscale it 2x, then 4x

[icy drift]

Don't use Latent upscale. It's never worked for me on FluxDev. Use VAE decode to get pixels, then 4x-Ultrasharp or whatever your favorite is, and then use a scale node to get it to the perfect size, then VAE encode, then back into the KSampler. (Actually it worked okay for 512->1024, just not for bigger stuff.)

Monster Card Series V2
No. 34 - Selkie

(I absolutely love how the sphinx turned out. 😊 One of my top 5.)

Monster Card Series V2
No. 35 - Sphinx

[icy drift]

Very kreepy. 💀

...But not as creepy as... This!

Monster Card Series V2
No. 36 - Wraith

Have a whole hoard of undead.

[mortal mesa]

Hieronymus Bosch

[icy drift]

Monster Card Series V2
No. 37 - Lich

[icy drift]

Artist?

[mortal mesa]

ya, old

[icy drift]

Still not too sure about this one. I guess it's cute enough... Has a very Tim Burton look.

Monster Card Series V2
No. 38 - Skeleton

[bitter hearth]

im trying out negative prompt with flux .. i wonder if that makes any difference

[icy drift]

How??? Well, first tell me if it works.

[acoustic kite]

nice approach, i just tried it. it looks like it's actually the base image it's trying to upscale that has the problematic roof tiles. to fix this maybe i can render it at a higher resolution at the start 😄

[icy drift]

Hmm. Okay.

[bitter hearth]

i increased the cfg to 1.5 and added clip node for negative prompt

[icy drift]

Monster Card Series V2
No. 39 - Zombie

[bitter hearth]

uses these keywords for it ..

ugly, distorted, deformed, conjoined, fused limbs, watermark.

let me try render an image

[icy drift]

Just render a landscape and put "blue" in the negative. Does it come out red-green?
Then try "red" in the negative. Does it come out cyan?
Then try "green" in the negative. Purple and red?

Also upweight the negative maybe?

Monster Card Series V2
No. 40 - Ghoul

[acoustic kite]

thanks though, it's definitely better

[bitter hearth]

clever straight foward idea, but gimme a min im rendering an image with character

[icy drift]

At first I thought I couldn't do multi-heads, but then I realized it could be like sisters arguing. Definitely cute-able.

Monster Card Series V2
No. 41 - Hydra

[bitter hearth]

P: a female with red hair, ponytail, light makeup, choker, mini sundress, bare thighs, half body, blue background, oil painting.
N: ugly, distorted, deformed, conjoined, fused limbs, watermark.

[icy drift]

Yeah I'm not sure. Maybe.

[bitter hearth]

should have used portrait canvas

[icy drift]

Monster Card Series V2
No. 42 - Chupacabra

Is Anubis really a monster? Yes. I need 55 of them, so yes it is.

Monster Card Series V2
No. 43 - Anubis

Monster Card Series V2
No. 44 - Kumiho

It was really hard to get these working. 75% of them would render with Japanese / Chinese characters instead of English no matter how I prompted it.

Monster Card Series V2
No. 45 - Kitsune

Monster Card Series V2
No. 46 - Bake-Danuki

[bitter hearth]

[acoustic kite]

wait a minute... does this upscale something 6x?

[craggy crest]

from yesterday's discussion on flux: 1344x768 gives the most reliably sharp images.
1536x896 can be good but you get occaisional blurry images.
the higher the resolution you go, the higher your chance of blurry.
sticking at 1344x768 (or some other 1 megapixel total image) gives the best results

[icy drift]

I always use the "Upscale Image" node and just set a size. It will scale it down if needed too.

(This is what a tanuki looks like, which the bake-danuki is based on. It's a raccoon-dog thing, and the bake-danuki version is when it's possessed by a spirit.)

Can you spell "ouroboros"? Because Flux can.

Monster Card Series V2
No. 47 - Ouroboros

[acoustic kite]

will try that, thx

[icy drift]

Current collection. Just 8 more to go! 😊

My top 5 favorites. The three on the right came out with such perfectly fitting personality in their expressions.

LOL basilisk. "Eh. Guess I'll poison you. Oh you died. Okay."

[bitter hearth]

wtf

[craggy crest]

hi cat

[acoustic kite]

i can finally fix the way the roof looks. it comes at a cost though 💀

10k image

but i'll downscale it again

[icy drift]

LOL watch out. Don't want to come back after a half hour render to see a big gray rectangle full of red text.

[icy drift]

Can't get over how real these things look. Every wrinkle, button, stitch, pleat, strand of hair. Everything is solid. There's no vague line soup anywhere to be found.

[brittle nexus]

And without a second pass, adetailer, upscale, inpainting...

[icy drift]

Yeah for real. There's an AI saying (by some famous guy whose name I would know if I were a real AI guru), that goes something like, "Self-improving software will eventually eat everything else in the stack."
Meaning if you have AI plus other code and processes, eventually the AI will eat that other code and processes. It'll be AI all the way down.

[acoustic kite]

this is an 8k image i just took a screenshot of so it's scaled down

all roof tiles have been fixed

[bitter hearth]

nice!! looks very good

[acoustic kite]

this is a zoomed in version of the hedge

[acoustic kite]

thanks 😄

[icy drift]

That is legitimately amazing.

[acoustic kite]

all roof tiles have been fixed unless you're literally right up and inspecting the picture

i'm gonna do some forest images now!!

pretty happy with the result 😄

took 40 tiles of 1024x1024

i don't have the computing power to make it better but if only it was easier to run it could be improved a bit more

hopefully i can afford something better one day

[brittle nexus]

Another prompt that SD3 failed. The bump that snout

[acoustic kite]

it's pretty close though

downscaled from 8k to 4k 😄

[brittle nexus]

Both are flux

[acoustic kite]

i'm sorry but i thought you meant you wanted the dog's snout to bump the camera so i said it's pretty close to the camera. i might have misread it. my bad 😆

[brittle nexus]

My bad then. As I posted two images and mentioning SD3, it would be rational to think it was the result of each. I deleted one so as not to create confusion

[craggy crest]

https://x.com/Birchlabs/status/1820280953429787065

Birchlabs (@Birchlabs) on X

Flux VAE preserves way more high-frequency details than SDXL 0.9 VAE.
left = Flux-roundtripped (nice and scratchy)
right = SDXL roundtripped (smooth and blobby)

Twitter · Aug 5, 2024, 2:09 AM

[bitter hearth]

yes but testing it with a pencil sketch is a hilariously bad example

just VAE encode and decode a high res photo lol

[craggy crest]

birchlabs is the inventor of hourglass diffusion. i think they can test on what they want - they were looking to see how it handled those sorts of details

[bitter hearth]

I'm not saying the test is technically invalid
its just that I would have used a photo of some sand or something

[sterile pendant]

No it's not, it makes it super easy to see issues. High frequency noise is what gives your skin texture with things like pores vs it coming out like a smooth airbrushed doll

The contrast of the image used is excellent for benchmarking

[craggy crest]

sand wouldn't be good. it's too uniform. the pencil shading texture on rough paper has extremely fine details that are chaotic and perfect for what was being tested

[mortal mesa]

zoom in more, every piece is different 🙂

[torn wharf]

no no not sand. this is fine grain

[bitter hearth]

this image is not sharp enough but this is the sort of sand I mean

[torn wharf]

this is also fine grain

this fine grain too

sand isn't grain at all like not even a bit yuck

[bitter hearth]

somehow I can tell that this is Dev lol

[hallow lion]

Tony Montana moment.

This channel now officially renamed Flux.

[craggy crest]

now that's pasta before it's pasta

[sage burrow]

Until something new comes out and we are bored of Flux 😄

[craggy crest]

this channel is in ... flux...

[fast ermine]

#🆕｜sd3 message

[craggy crest]

when you typo the prompt in flux -and write elle instead of apple

[icy drift]

Just got a 25%-50% speedup in FluxDev! 😲

All I did was open ComfyUI on my laptop. The Comfy server is running on my desktop.
Opening the UI on my laptop for some reason made it way faster. Your mileage may vary.
(To clarify: The server is still running on my desktop, I'm just making the workflow and clicking queue on my laptop.)

[sterile pendant]

What it probably is is the random thing that will happen where it the model is too big for your vram, it has to offload to your sysram. Well it seems there's an issue where sometimes the CPU doesn't help at all and sometimes it does. For instance, my 2080 and 13600kf will have 15sec/it if working correctly and if the CPU doesn't help, it's like 30sec/it

I've tested every combination of Nvidia cuda fallback and comfy --lowvram --highvram. The only thing that fixes it for me is to start a generation, wait for the models to load and for it to hit the clip prompt, then cancel it and try again. Or restart comfy

I know it's working when I see my CPU at like 50% usage

My guess is it's a windows memory cache or management issue or something

[icy drift]

Yeah the VAE runs on my CPU if I use the UI on my desktop, and it takes 15-30 seconds per image.
If I switch the UI to my laptop, the VAE starts running on the GPU instead, and takes <1 second. I'm using Firefox if that makes any difference.

[cunning mesa]

Do you have hardware acceleration on on your browser?

[icy drift]

I assume it's on by default in 2024.

[cunning mesa]

Maybe try turning it off and see if it is the difference.

[icy drift]

Not worth it. (My dev work is browser WebGL.)

Also, I still have my browser open, but only Discord. Not using the ComfyUI. It might be the ComfyUI itself is using the GPU.

[cunning mesa]

For what it is worth, I have similar speeds as you with the speed-up running it locally.

[sullen moss]

https://github.com/bghira/SimpleTuner/releases/tag/v0.9.8

[icy drift]

What hardware / model / steps are you using?

[icy drift]

Here comes Pony Flux.

[cunning mesa]

3090, dev Flux. Bit less speed than you, but I am also running undervolted.

[icy drift]

Steps / Resolution?

I'm 4090(24GB) / 64GB sysram, 23 steps, 832x1216 (same as 1024x1024)

Getting very consistent 31-32 seconds.

[cunning mesa]

25 steps, 1024x1024, 16Gb sysram, Euler. About 35s.

[icy drift]

Heun / Beta. I'm surprised you're running that fast. Let me try Euler.

[cunning mesa]

I think Euler is quite bit faster than Heun.

Like, twice as fast.

[icy drift]

Yep. 16-17 seconds now.

And the quality difference isn't worth 2x. This is awesome! 🙂

[cunning mesa]

Yeah, there's differences with the samplers but unless you are shooting for absolute quality for a single planned picture, I don't really feel like it is worth it to use the slower ones.

[proven pecan]

He's a bit weird 😉 (SD3)

[icy drift]

Finished! 😊

Full set of 55 cards, ready to print.
Well, still need to size them correctly.
And a few more fix ups here and there. I'll check each one carefully for problems and see what turns up.

Final 8 monsters were:

No. 48 - Qilin

No. 49 - Will O' The Wisp

Cerberus turned out exactly like Beta from Emminence in Shadows. 😛
Also just noticed the line on her paw/leg/haunch is wrong. Gotta fix that.

No. 50 - Cerberus

No. 51 - Yuki-Onna

Headless horseman. Another practically impossible task for FluxDev.
No. 52 - Dullahan

Gremlin's are a very recent fantasy monster. Invented by airforce pilots as the supernatural entities that caused plane problems.
No. 53 - Gremlin

FluxDev also doesn't like doing shamrocks apparently? It wanted to do 3-leaf clovers. Maybe my prompt was wrong though.
No. 54 - Leprechaun

And finally, the last card in the set!
No. 55 - Lamia

Now just gotta do the backs and get them printed. 🙂

[covert sluice]

test

[sullen moss]

Hmm, I’m really surprised

[warm sonnet]

Was this image generated by Flux? SD3 can never generate an image with such perfect toes.

[sullen moss]

Yes, it's FLUX. For me, SD3 doesn't exist at the moment. The more I test FLUX, the more I fall in love with this model.

[proud geode]

How to use the drawing function of Stable Diffusion

[dull star]

MIDJOURNEY IN STABLE DIFFUSION SERVER 💀

[cursive sandal]

Under the sunset, a beautiful girl is standing by the lake watching the sunset. The plot of a Chinese girl movie is --niji 6 --s 250 - @compact snow (relaxed)

Hey @unkempt matrix
Due to extreme demand we can't provide a free trial right now. Please </subscribe:987795925764280351> to create images with Midjourney.
[Please check the bottom of the channel for more information.](#🆕｜sd3 message)

[icy drift]

Three finger hands. Flux is better.

[dull star]

nice

[bitter hearth]

you go away

[brittle nexus]

SD3 placed feet on the list of censored sexual organs. So all foot images were properly lobotomized

[brittle nexus]

Ironically they can show feet

[sullen moss]

Total CENSURED 😂

[fresh cloud]

/1girl

[radiant ledge]

ERROR: Division by zero is not defined

[brittle nexus]

This prompt photo of a woman standing against a solid black background. She is wearing a matching black bra and panties. Her long dark hair is straight and falls over her shoulders. She is facing the camera directly, with her arms relaxed by her sides and her feet slightly apart. The lighting highlights her toned physique and balanced posture, creating a sharp contrast between her figure and the dark backdrop. The overall composition is minimalistic, focusing attention entirely on the subject. will give you blurry images with Euler unless you use "Beta" scheduler. Any theory?

[sullen moss]

Hmm...

[proven pecan]

Schnell

[sage burrow]

Flux Schnell vs Dev
I definitely prefer Dev for comics!!

[sage burrow]

vs sd3

vs SD3 large (glif)

[sage burrow]

vs Meta
All the exact same prompt for everything

I'm sorry Flux, but it seems Meta is definitely giving you some competition!
"A close up of a bearded man with kind eyes, wearing a lab coat, looking through a magnifying glass at a small vial containing a green liquid. A speech bubble which says "Throughout history, humans have sought to change matter into more useful forms." Manhwa photorealistic art style, by Junji Ito"

[uncut river]

For comparison, sd3 local + copax lora, first 2 gens with this prompt above

[sage burrow]

Next I tried a grasshopper Penguin woman.....
Meta...

SD3

and Flux

and Flux via glif with GPT4's help lol

[sage burrow]

Is that Flux or SD3?

[alpine summit]

Sd3 + Sd 1.5

[sage burrow]

I knew 1.5 was still awesome! 😄

[uncut river]

does chatgpt assistence make the images better, or just make the program easier to use?

[pseudo owl]

it makes the prompt better which can make better images. Also, you dont have to write a full prompt so it can make it easier and faster.

[rain current]

Motion shots (schnell)

[uncut river]

Has chatgpt been trained to create prompts that give higher quality images? Seems hard to do

Comfy question. Can I use the Latent output of 1 sampler (using sd3) connect to a next sampler (using sd15) ? I gives some errors. Do I first need to convert to image, then back to latent space .... ?

[pseudo owl]

most models need a bit of different style of prompting. Most dit models like auraflow, flux, sd3 work good with natural prompts. You can just do "enhance this prompt: {simple_prompt]. Keep it 2-3 sentences." and it should probably work.

[uncut river]

ah, so its more like sd3 and other models are build for language model prompting

[pseudo owl]

sd3 and sd15 do not have the same latent space so you need to convert it to a image. There are some things that can let you convert sd3 latents directly into sd15. vice versa too. For example
https://github.com/city96/SD-Latent-Interposer

[sage burrow]

It just improves the prompts to be more descriptive and/or more in line with whichever specific txt2img you are using

[signal shuttle]

Flux

[torn wharf]

Til that these are a thing and that they're called an interposer. Tyvm

[craggy crest]

https://x.com/nrehiew_/status/1820404554795802684

wh (@nrehiew_) on X

Flux's Architecture diagram :)

Don't think there's a paper so had a quick look through their code. Might be useful for understanding current Diffusion architectures

Twitter · Aug 5, 2024, 10:20 AM

[torn wharf]

What I find with the t5 encoder is that since it has self attention, you don't just do style tagging with it. Give it a little lore. Backstory. Jazz it up. Talk about all the context around the events of the image.

Anyone know of a workflow for flux like for sd3 where you can prompt all the encoders separately?

[signal shuttle]

I think there is a node in comfyui called cliptextencodeflux, that does that

[torn wharf]

Beans

[uncut river]

sd3+sd15 workflow, using sd3 first for snappy prompt2img then refining with Cute Lunar for the style and details. seems to work fine.

[wispy spade]

dam nice work

[wispy spade]

how does this work?

[uncut river]

drag the image into your comfy canvas to see the workflow and all settings. almost all my posted images include the whole comfy workflow

[craggy crest]

keep in mind that the image you see in discord isn't the original image. so to get to the one with the workflow, first click the image, then on the larger one that opens, click the words open in browser. and then save THAT one.

[wispy spade]

gonna try that one thanks brother

[wispy spade]

anyone know why flux schnell cant reference certain art? for example i testing to create image about some character from arknight mobile games, but schnell model dont undersyand at all what is arknight mobile game that i referein to

[craggy crest]

probably nothing in it's data set that had the lable you're using

[wispy spade]

but how can that be? sdxl with way less parameter and data sample can do that. but schnell have 12B param.

[sage burrow]

It's cute and all that people are making SD3 nsfw loras, but WHY are they all kinda BLURRY?!

[wispy spade]

pony

[craggy crest]

what has that got to do with anything? it might ahve all the art, but if the label yo uare using wasn't used, how's it going to know what you're talking about?

[wispy spade]

what do you mean by blurry? never see that

[sage burrow]

Pony isn't blurry? Do you mean they are using Pony ref images?

[sage burrow]

Check civitae, all slightly blurry

[craggy crest]

civit probably has something set wrong

[wispy spade]

only nsfw one? or all image on civitae

[sage burrow]

People are posting finetunes they made of SD3 and they post images to highlight their models, but all those images are blurry. However, the SFW SD3 finetunes seem fine

[rain current]

• motion (schnell)

[craggy crest]

oh, well make sure your profile settings haven't changed

[sage burrow]

Only the promo images for the SD3 nsfw models

[sage burrow]

Not that blurry, more like 80's disposible camera kinda blurry

[wispy spade]

i dont see that, can you provide an example? mine look fine

[sage burrow]

off to crop some images, brb lol

[pseudo owl]

use sd3 to generate the image and a finetuned sd15 to refine it

[wispy spade]

never try this before, prob tomrrow im going to sleep rn. but that brother image look stunning wanna try it myself

[torn wharf]

should do a final pass back to sd3, enough to refine the colors out while keeping composition. get that sd3 vae

[pseudo owl]

mostly because of the sd15 model, the sd3 is just for the composition. You can do it with any other models as well.

[torn wharf]

did something happen? is sd3 cool again? what'd i miss? i thought this was flux now

[wispy spade]

will it work with model that have build in vae?

[pseudo owl]

no, sd3 just still makes ehhh images. but sd15 helps since its a lot better in humans.

[sage burrow]

cropped for obvios reasons, but that's about standard

[sage burrow]

What can 1.5 do to an image that SD3 cannot? I'm wanting to try this out, but curious what I'm after specifically

[wispy spade]

dam look sad

[sage burrow]

Sometimes flux isn't all that, so we have to revert back to SD3 lol

[sage burrow]

They aren't all this blurry, but even slightly blurry isn't acceptable these days (IMO)

At least SD3 ladies with 1 arm and 3 legs laying on grass aren't blurry ROFL

[torn wharf]

but what about the new flux. didn't you hear about the new new?

https://huggingface.co/sayakpaul/FLUX.1-merged

[sage burrow]

Using different models for different things in comfy is fun, since each has its strengths and waeknesses.
PS I need more vram

[pseudo owl]

schnell is better and faster imo but at least that supports negative prompt

[torn wharf]

its diffusers so it confusers me

[sage burrow]

How come this one would be used instead of just Dev? I hear Schnell is better for txt, but so far I haen't seen that

I just want an SD3/Flux merge 😄 (I know, I know, not possible lol)

[torn wharf]

its the two merged together. dev an schnell. i get great text from dev myself. schnell has great speed tho

[pseudo owl]

its a merge of dev and schnell. so low steps like 4-10 and cfg as well is possible.

[sage burrow]

Schnell is only twice as fast though 😦

[pseudo owl]

schnell is like 1-10 steps and dev needs like 20-25 steps i believe. 6 is a good spot for schnell and its like 3-4x faster

[craggy crest]

we had a huge discussion about this yesterday, on this discord, and the concensus was that negatives really make no difference

[pseudo owl]

they just remove some objects if you want but you cant increase quality really.

[pseudo owl]

I have no idea what prompt but i think not too bad with schnell, 8 steps

[craggy crest]

also not real sure how merging schnell and dev gets you anything. you're just undoing the distillation process

[dry wave]

they are both distilled

[craggy crest]

and probably not undoing it correctly

[dry wave]

for sdxl, merging turbo with base also had interesting effects of resulting in a better model

so why not trying it. Might be good, might be bad

[craggy crest]

yes they are, and when you merge them, you mess with that

[sage burrow]

So which one/s? 😄

[dry wave]

they are one

[craggy crest]

this is vastly different. turbo is optimized for speed. so when you merge it with the base all you do is cut down the speed optimization really

[pseudo owl]

both, diffusers shard them so you dont need to load the whole thing in ram. You require less ram bc of that.

[dry wave]

itln diffusers the models are often splitted in multiple files

[craggy crest]

just download all the models. and they go in your Unet folder

[dry wave]

turbo is also a distilled model.

[sage burrow]

Thank you 🙂

[craggy crest]

i know. optimized for speed. when you take an optimized anything, and mix it back in with the base - you destroy the optimization and you don't really accomplish anything. it's sort of like brewing beer, distilling it to make a fine, high quality beer, then pourin that into a pitcher and pouring some of the unfiltered stuff in too - and then thinking you've got something great

[dry wave]

it's not that easy

[craggy crest]

who said anything about easy. but tha's essentially what you are doing.

[dry wave]

first, merging back to base is similar to what ema is doing. It often improves quality

[craggy crest]

in who's opinion? and in compareison to what

sticking a fork in an electric socket might improve your hearing aids too - but it might do a whole lot of other damage

[dry wave]

second, turbo is restricted to 512x512. Merging it back to base gave it the possibility to generate 1024x1024 with cfg will still being extremely fast

so yes, it's no guarantee that it works but it's not a dumb idea trying to merge the models given that schnell is better for certain tasks (emg. paintings)

[mortal mesa]

guys stop doing things, someone who doesnt have a high level of knowledge doesnt like it

[dry wave]

right xD

[sage burrow]

is there a fancy workflow which puts a separate prompt into each? 😄

[craggy crest]

if you study, you'll have a higher level of knowledge

[mortal mesa]

captain obvious

[dry wave]

easy to do

[craggy crest]

since it's so very new, instead of running around trying to break it immediately, maybe it would be good to learn it first?

[dry wave]

just use the flux prompting node

[dry wave]

dude, experimenting is part of learning

[mortal mesa]

you learn by breaking when you dont have an instruction manual

[uncut river]

sd15 has more various styles, also fixes most poses

[craggy crest]

you've missed your calling in life

[mortal mesa]

wrong again

[sage burrow]

Half my fave checkpoints are 1.5 😄

[uncut river]

sd3 + sd15 can do female laying on stones, in the grass

it's just the grass!!!!

😄

[craggy crest]

you should be doing stand up comedy

[pseudo owl]

Humans are a lot better, better styles, celebrities too. Its very small and really quick as well. prompt following is very very low and might mess up your text in images but apart from that its great as a refiner.

[sage burrow]

Have you tried Flux + 1.5 yet? 😄

[dry wave]

it's just pointless talking with crystalwizard 🤷‍♂️ he has no clue if anything but starts arguing with everyone, although arguing means insulting and flaming for him

[pseudo owl]

hands are ehhhh but still good

lot better then raw sd3

[mortal mesa]

another pretty far off opinion, im not good

[craggy crest]

you're very good.

[mortal mesa]

not at comedy

[sage burrow]

wanders off to mix Pony with Flux...

[craggy crest]

you're very good at comedy.

[uncut river]

does flux has online img2img ... ? surely not

[mortal mesa]

can you go to off topic to talk to yourself

[craggy crest]

and then the ponys will flux in and out of reality

[craggy crest]

just following your lead

[uncut river]

next year: flux based pony video models

[sage burrow]

Mage does, but also a glif of that is probably not too difficult (adds onto to do list)

[dry wave]

please not

[mortal mesa]

you lead with a dumb statment of guys your ruining it

[craggy crest]

though you might work on literacy

[mortal mesa]

everybody sees

[uncut river]

,,( * .. * ),, hi guys, everything ok?

[dry wave]

only a few minutes and my first trained flux lora is done 🤩

[mortal mesa]

im not blocked by anyone teehee

[mortal mesa]

why must you attack all the time, low bar for you, my literacy is fine, you are the one often doing doublespeak and gaslighting

get help

[uncut river]

get drunk, get some of that natural green stuff, both of you. You might feel better

[signal shuttle]

Your training a lora for flux?, how much Vram is needed?

[uncut river]

cropped frustration never work out well

yeah, let's all just hug. And then cancel anyone who doesnt join! 😄

[mortal mesa]

ya the cancel culture is strong in some

[dry wave]

I'm at 18gb currently

[uncut river]

im cancelled

[mortal mesa]

mee too

[uncut river]

not here, civitai

[dry wave]

maybe you can reduce it a little bit, but I guess you will need more than 16gb

[signal shuttle]

18gb? interesting, last time i heard the minimum requirement to train a lora for flux is 24gb

[dry wave]

You have to train it quantised

[mortal mesa]

i deleted my civit account, then just yesterday i wanted a particular workflow and i was suprised when it wouldnt let me download it, made an account again....

[dry wave]

but that seems to work very well and is not huge issue - in the end we all run it quantised anyways

[uncut river]

oh wut, we're in sd3, not general-chat

[torn wharf]

what'd u delete for? i know someone else did too

[signal shuttle]

Is the quality affected at all?

[dry wave]

I don't think so. But too early to say

[torn wharf]

text probably does a lot better with higher precession. most things maybe not

[mortal mesa]

dont want to be associated with some of the smut there and the anime level was personally anoying to me digging through it, just not gonna post images

[torn wharf]

thanks. just curious. seems to be a common reasoning

[signal shuttle]

Are you using simple tuner?, i want to give training a lora for flux a shot

[dry wave]

yes

took me a while to get it run, but with every commit and bugfix it gets easier I guess

[uncut river]

I think my collection is cute, only up to X rating for some images. Can't see why I would ever delete it, but just in case, you might wanna check out: https://civitai.com/user/youhnr5/images

[sage burrow]

last article I read said 48 😦

[dry wave]

yeah, turns out it's trainable on a 3090 🥳

[torn wharf]

you think sd3 is better at text then you'll always be wrecked TOP THAT

[sage burrow]

You can disable anime I think

[mortal mesa]

ya still irked me, less anime wqas still like 80% anime lol

[uncut river]

I think you might wanna do some OCR and then let language models check this piece of text ...

[torn wharf]

nobody's appreciating my dope ass teen witch deep cuts

[sage burrow]

that image along with that text ROFL

[uncut river]

then re-route back into the image AI until it get's it perfect

[torn wharf]

full prompt

On the right hand side of the image is a large black margin with the lyrics printed out. Text print in arial font-face. Very legible letters.
("I'm king, and they know it
When I snap my fingers, everybody say show it
I'm hot, and you're not
But if you wanna hang with me, I'll give it one shot
Top that, top that
You can give all that you can, but you will never top that
Top that, top that
TOP THAT! TOP THAT!
You can dream until you're blue but you can never top that, huh-huh!")

On the left of the image is a photo.
In a dimly lit underground venue, a mad scientist takes the stage, his wild, unkempt hair and stained lab coat contrasting with the urban setting. He wears a pair of joke spring eyeball glasses, with the eyeballs bouncing humorously as he moves. The crowd is packed tightly, swaying to the heavy beats of the hiphop music. Graffiti-covered walls and flickering neon lights add to the gritty, energetic atmosphere. The mad scientist holds a microphone, his expression one of manic excitement as he delivers his rhymes, blending science and hiphop in a unique and captivating performance.

[sage burrow]

would it work better if you made the text in 4 sections? Would that be easier for the ai?

[torn wharf]

https://youtu.be/fAWJdXJvngU love these kinda plans

[mortal mesa]

Civit wouldn't let me post this but i was seeing questionable age pron all over, camel straw back

[sage burrow]

how to say your over 40, without saying you're cover 40 😉

[torn wharf]

i've tested writing lyrics in sections like that on other stuff. it doesn't seem to care.

i'm over 40 an love anime 😦

But not like LOVE like filthy frank loves it

https://tenor.com/view/eu-gif-18523165

[uncut river]

I just threw this prompt at my sd3-sd15 workflow ... aaannd it's a mess

[torn wharf]

i squeel anytime i see a new version of that scene

[sage burrow]

I like Manhwa, but don't like anime at all 😦

[torn wharf]

looks like a good show still

[uncut river]

its painfull to see the sd3 preview inside the workflow have better text, and then it's gone 😭

[torn wharf]

not even ghibli? i just figured everybody loves ghibli. maybe i'm projecting

[sage burrow]

I just googled that, no, definitely not

[uncut river]

lol, next project. Full prompt, I kid you not:

waifus = ["Asuna", "Mikasa", "Rem", "Hinata", "Zero Two"]

for waifu in waifus:
print(waifu)

[torn wharf]

hahaha the previews are often a curse. like the aesthetic distallation of the scene kicks in and suddennly blammo. all that detail you could see shaping up shifts to aesthetic pulp

with flux

[turbid grotto]

why using 1.5 to destroy hands and let 4ch vae decrease quality

[uncut river]

why? cuz style and poses. yeah, I know it destroys some of the good sd3 stuff, like also text

[mortal mesa]

i even made anime, not a hater, just dont seek it out as much

[uncut river]

look good

[uncut river]

always nice to see 2d and 3d combined in a single image

[turbid grotto]

oh ok fair and I thought first pass is flux

flux is quite good at anime

[uncut river]

yes, it has a nice anime style. one of the very few downsides of flux that I've seen is trying to get grainy photos. As soon as persons are added to the prompt, the style becomes smooth and "perfect" again (the default realistic flux style)

made with flux-schnell online yesterday

flux-schnell anime style

hm, maybe looks more cartoon or common 2d

[sage burrow]

@uncut river what am I doing wrong with your workflows?

I wanted to turn all your wiafus into zombies! 😭 🤣

[uncut river]

hm, not sure but I've seen a similar error

the error I think I made was connecting the latent output straight to the input of the next sampler. that did not work. it needs to go through vae-decode to image, and then back to latent space for sd15 with vae-encode using the sd15 stuff. But if the workflow is included in the image, it should work (without alterations, expect maybe the chosen models)

[torn wharf]

dev is great at dali's style but i can't get it to do the drooping clock faces. its neat though, in the early steps, i can make out the drooping shapes. but then as the detail comes in it changes to another form.

[sage burrow]

I didn't change anything except the models and the vae (though perhaps I used the wrong vae)
I hadn't even gotten to the prompt part yet

[uncut river]

hm, could indeed potentially be a vae mismatch or something. Probably the wrong vae connected to the vae-decoder after the sd3 generation. When did you get the error?

[craggy crest]

@sage burrow https://civitai.com/models/624564?modelVersionId=698253

[torn wharf]

my instant concern with those two pass workflows are that they do final output with the old vae

[craggy crest]

yes, but i'd rather use something like this, than merge the two flux models

[sage burrow]

OK now I give up lol

[torn wharf]

i'm not saying they're a bad approach. i'm just thinking, do a third pass

get that sota vae

[craggy crest]

maybe. have you tried running both models in one workflow yet?

[mortal mesa]

refine flux with flux, call it reflux

[craggy crest]

ouch, but good pun

[mortal mesa]

maybe i am a comedian hah

[craggy crest]

said you were good at it

grainy image with dev

[uncut river]

lol

[sullen moss]

well...

[uncut river]

nice rings!

. . .

👊

[craggy crest]

note - loading dev and schnell in the same workflow may lag your machine

[blissful raft]

can flux write words it doesnt know?? like "omgwtfbqq" ??

[uncut river]

sure

well, almost

took a few tries, but flux-schnell is schell ja!

[craggy crest]

both models - schnell into the BasicScheduler node, and dev into the BasicGuider node

[uncut river]

try it yourself at for example https://replicate.com/black-forest-labs/flux-schnell

[sullen moss]

It's hand cream, not what you thought. 😁

[uncut river]

seems realistic

can you do face lotion ?

[sullen moss]

for hair strengthenin...

[mortal mesa]

cinnamon roll icing

[uncut river]

seems like shampoo instead of something for the face, or she is just crazy

flux gives me this on:

Discover the secret to a radiant, clean face with our family-style face lotion. Just like our anime heroines, your skin will shine with confidence and grace, no matter the weather. Our lotion's advanced formula keeps your face moisturized and protected, even in the harshest conditions. Embrace every day with a glowing, flawless complexion!

[craggy crest]

that's melted vanilla ice cream and what you don't see is her younger brother behind her dumping it over her head

[uncut river]

ohhw, you cropped the full image

you ...

[torn wharf]

i've done multi model workflows before. comfyui aint my first nodegraph ui and i've done stuff so much more extensive than these boys can be. the worst is just figuring out the names for all the nodes. ever managed database normalization?

[uncut river]

After creation of the first prompt, I simply asked the AI to "now make it real" ... 1st gen

1st was better

[sullen moss]

Time to shave. lol

[torn wharf]

through strict scientific analysis i have noticed a descrepency in the data. when prompting for the letter q sometimes these modesl will use the letter p instead. from this correlation we can hypothesize that diffusion models aren't capable of minding their ps and qs. || EVERYBODY COME ON FHQWHGADS ||

[signal shuttle]

I didn't give flux a prompt and this was the result

[torn wharf]

nice scene

[signal shuttle]

I got really lucky with the seed

[torn wharf]

all this checks out for a futuristic cyborg sumo match.

[signal shuttle]

Nice human alien hybrids

[sage burrow]

Yall Bern hanging around civitae too much ,)

[torn wharf]

hmmm. i'm running these not in fp8 mode. it's filling my 16gb vram but not over .

[frail shoal]

have you guys ever gotten these weird artefacts on the red fog that you see there. I just got it in schnell and one time on the hugging face space with dev. What is that about ?

[torn wharf]

same seeds one is fp8 em4 turned on an one is off

[icy drift]

Did you know you can specify measurements in your prompts?
three cats, from left to right: a one foot tall cat, a two foot tall cat, a three foot tall cat, in the kitchen

[uncut river]

wtf put gloves on cat when they got razor sharp nails??!

make 'm claw&hit eachother

[torn wharf]

https://tenor.com/view/cosmo-kramer-seinfeld-grin-funny-gif-10495570501246051470

[uncut river]

like a real female cat fight

yeah!

[icy drift]

It doesn't get the measurements right, but it gets the differences (big little small).
three cats, from left to right: a three foot tall cat, a one foot tall cat, a two foot tall cat, in the kitchen

[bitter hearth]

I got an idea for a comfy workflow where each step gets its own ksampler

and each ksampler only moves the ODE by one step

then in between each one you select CFG, SAG, PAG and control net etc for the next step

[icy drift]

That's why they've got one paw ungloved. Get both bludgeoning and slashing damage.

[uncut river]

oh damn, im honored

seems like a full front page article 😄

[torn wharf]

yeh size prompts work good

[bitter hearth]

this text is really good WTF
is that flux?

[fleet meteor]

This is stolen from reddit, anyone else tried images about gaming on flux? 😄

[signal shuttle]

yes

[torn wharf]

this has potential to be a ridiculous discord server bot. the mod can just prompt breaking news anytime someone is stupid.

[torn wharf]

yeah one of the first things i did. it works good for gaming. had a lot of screenshots in the data surely

[bitter hearth]

is Flux text better than SD£ or about the same?

[uncut river]

much better

[icy drift]

Flux is the best text of any model I've ever seen, by a very large margin.

[bitter hearth]

ah ok nice

[uncut river]

yes! too bad its too big, and cant fit my vram

[bitter hearth]

I don't do text but that's good to know

[torn wharf]

sd pound it

[uncut river]

that is about one of the only down sides, its big size

[fleet meteor]

Its way better, the text is not just "pasted on top" like in many sd3 generations (of course this is not all the time)

[uncut river]

therefore, we NEED ... .eh, would like to have maybe a MINI-flux

[bitter hearth]

calling SD3 as SD£ was actually a typo
but it looks like I am insulting them lol
was just an accident

[bitter hearth]

oh yeah I hate the pasted on top thing

[uncut river]

flux ....

[bitter hearth]

WTF it can do a whole NYT paper with the newspaper name correct
that's so amazing

[torn wharf]

not bad for a rough lorem ipsum level layout

[icy drift]

True. Design instincts.

[bitter hearth]

sometimes I just read lorem ipsum instead of reading real human-written writing TBH

[uncut river]

I forgot and lost the prompt, but tried to recreate from mind in flux-schnell. here it is.

2d anime style, hot mom standing surprised in the kitchen, large fluffy white dog sitting on the sink, simple anime style, playful text "SHIRO" written above the white dog, lazy dog

lol, the sd3-sd15 workflow can somewhat keep up! it even adds a bit of japanese (I guess...)

[bitter hearth]

how much worse is Schell?

[icy drift]

It can't do much text, but it's a lot faster.

[uncut river]

the time-quality balance is about the same, it just very schnell

schnell means fast in german, for who doesnt know

[bitter hearth]

do you mean its the same quality?

[uncut river]

no