🆕｜sd3

[dull star]

SD3.5 and Flux

[robust thunder]

Hello! I have questions regarding the "commercial use license" for SD3. Where would be the best place to ask about that?

[sacred jewel]

shorter token quantities.. still not as good as I expected. Workflows embedded

[lavish sparrow]

nah, it's certainly better on paintings, try watercolor, or pastel acrylic

the actual composition of gemstones -> all things that are lost on flux

[dull star]

oh I remember watercolor on the service, that was good

I just remember making regular oil (impressionist) paintings on the service and it looked better than offline

that's what I'm mostly after

[raven fern]

if i personally was serious for commercial stuff, i would contact SAI directly to answer some questions i guess, idk

[lavish sparrow]

strange trojan horse.

[raven fern]

Big Ben is going down

[dull star]

damn

[dusky thistle]

lol

[lavish sparrow]

and. sd3.5 does blood. flux doesn't want to.

[raven fern]

nice

[robust thunder]

fair enough, I was hoping someone around here had some experience and understanding right off the bat :v

But thanks for replying!! I'll be looking around their website then

[raven fern]

i mean maybe someone can answer you idk :3

[dull star]

I remember sd3 large being very good at gore and flux being almost nonexistent

so its nice to see that still working

like a store made out of guts or some stuff like that

[lavish sparrow]

did a spider-pig earlier, yeah, it's good at that. for certain better than flux

flux feels very neutered

[sage burrow]

If you aren't making enough from your ai art to buy a new Farari, prob don't need to worry.

[lavish sparrow]

no idea what i put in as prompt. modern architecture is weird man

[raven fern]

i like the mana, health and poison potions on the table haha

[noble coyote]

Aortic tomato ketchup spurts from an open wound, where smashed and jagged candy bones protude thru the skin...

[raven fern]

(the example pic)

[sage burrow]

What? Flux does blood just fine? Though I never did try lakes of blood...

[lavish sparrow]

nah, it didn't do it fine at all

[noble coyote]

Nice!

[lavish sparrow]

at best a few splotches, but nothing in the realm of the thing you'd want, if you were gunning for it

[dull star]

I just can't get it to make a good impressionist oil painting

[spark quail]

holy hell

[raven fern]

that looks like a cover for some movie, nice

[raven fern]

white background 😮

[dull star]

eh I give up on paintings

[sacred jewel]

a mix of 70, 100 and 200 token prompts

[dull star]

I swear there's something wrong with the VAE or whatever, the image have a speckled look to them

[dull star]

CCTV is okay

[raven fern]

kek

[robust thunder]

🤣
I figured as much but I wanted to excercise due caution so it doesn't bite me in the ass later on.

Essentially I wanted to make graphic t-shirts but I have little physical artistic talent. So I've been having SD draw things that I want, and I'll mash them together using GIMP and edit them to my liking then slap them on a shirt.

I'm not sure of ANY legal jargon so reading the "https://stability.ai/community-license-agreement" makes my head hurt....

[sage burrow]

Not as good as MJ, but not too bad. Flux Dev lol

[lavish sparrow]

the part where it just didn't hold up for me -> it didn't want to apply the blood anywhere

it's just splotches, and non-applied blood

[raven fern]

you have to do some "prompt engineering"

[noble coyote]

Septo-prompt

[lavish sparrow]

aight, lets overload the prompt

[lavish sparrow]

thomas got weird when he snorts his coal

[dull star]

As between You and Stability AI, You own any outputs generated from the Models or Derivative Works to the extent permitted by applicable law.

[raven fern]

i like it

[lavish sparrow]

ugh...i REALLY should go to bed

[sage burrow]

Perhaps you should create some loras 😉

[lavish sparrow]

... i never make/use loras. pure prompter only 🙂

[raven fern]

that can only get you so far

[lavish sparrow]

nah, gets me exactly where iw ant to. just want to have some fun and post few images

^^

[raven fern]

nice

[hallow lion]

hey where is cat with 4gb of vram on this glorious day?

[robust thunder]

yea, "extent of applicable law" is ambiguous enough that idfk what limitations that entails 😭

[raven fern]

yea i tried to contact him lol

[dull star]

Photo of Criminal in a ski mask making a phone call in front of a store. There is caption on the bottom of the image: "It's time to Counter the Strike...". There is a red arrow pointing towards the caption. The red arrow is from a Red circle which has an image of Halo Master Chief in it.

Flux does this first try

or definitely consistently

[raven fern]

lets be honest here man, you are most likely not gonna generate anywhere near 1 million dollars :3

[dull star]

https://tenor.com/view/gman-hl-hl2-half-life-half-life2-gif-25189631

[lavish sparrow]

he lost his account, it's @rapid pivot now

[dull star]

I appreciate the texture on everything though, flux is made out of plastic

this one is made out of scratched plastic

[raven fern]

plastic booba

[robust thunder]

that's what I figured :v

[raven fern]

dont worry man, il give you a small loan of 1 million dollars

[raven fern]

holding swords is usually hard for AI, they dont hold the swords correctly

[bold ember]

Hey Im not really a technical user or anything, Ive been running a1111 SD 1.5 on my AMD gpu since last year
Will SD3 work on my rig? Or should I just stick to 1.5 until I get better hardware ( wont be building a new PC until end of 2025 )

[raven fern]

ye video game stuff is always fun

[lavish sparrow]

well, time to go to bed ^^ cyaz

[raven fern]

gn sweet prince

[dull star]

gn

[dull star]

yeah flux got this right

[raven fern]

i only played cs 1.6

[dull star]

oh look no bokeh

https://tenor.com/view/thanos-impossible-marvel-shocked-gif-15104180

okay what the fuck comfyui you are destroying my space

what are you offloading

[compact forge]

bruh at least say what card it is and what vram it has 😂

[raven fern]

(bokeh:9999) in the negative prompt :3

[sage burrow]

The free HF SAI image generators for 3.5 won't generate certain images. Though their Flux HF generation areas weill do the same prompt just fine.

[bold ember]

its an rx 590, 8gb of vram
its not powerful at all, my SD 1.5 caps out at like 900x900 ish

but i get nice results

[raven fern]

are you not generating locally?

[compact forge]

i think you need 12gb for flux and sd3. so untill then you should try out sdxl or pony

[bold ember]

Oh damn okay, Ive seen a lot of pony so I guess Ill switch to that. thanks!

[raven fern]

12gb vram and 32gb ram is like the minimum these days for the cool stuff

[noble coyote]

Flux and SD3.5L work fine on my 8Gb VRAM/64Gb bRAM

[compact forge]

oh damn, thats cool

[raven fern]

nice

[compact forge]

i often get vram errors on my 12gb card though 😅

[bold ember]

yea i built this PC in 2019 and with win10 ending in late 2025, my pc doesnt meet sys reqs for win11, i was just gonna wait and upgrade the whole rig at that time and probably get a 12gb vram nvidia card

[raven fern]

what is the generation speed? @dusky thistle

[sage burrow]

Mage just added 3.5!!!! ❤️

[raven fern]

she works at hot topic

[dusky thistle]

30/30 [00:55<00:00, 1.85s/it]

[raven fern]

kk

[rapid pivot]

https://tenor.com/view/kok-cat-gif-25438805

idk who that guy is

[raven fern]

did you make a new account? :3

[rapid pivot]

its only the 5th

Don't worry

[raven fern]

haha

[raven fern]

lol

[robust thunder]

Anyone know a good link to educate me on installing and using SD3.5-Large-Turbo in ComfyUi on my PC?

like, what files exactly and where to put them? PC requirements to feasibly run it? etc?

[raven fern]

the comfyui github readme has enough instructions to get you started, especially with example workflows and instructions where to place each file

[robust thunder]

thx, I'll check it out

[dusky thistle]

\

[raven fern]

now that is way better

[raven fern]

i like the blend of colors

[dusky thistle]

seriousyl, if this thing is anywhere near as trainable as cascade, it's gonna be amazing

[raven fern]

i wish they would bring back the Cascade feature of mixing pics

cause you can easily blend styles

[dusky thistle]

that was great stuff... the clipvision? it's actually a lot like using ipadapter

[runic tusk]

When Alioth finds The Mask.

[sacred geode]

I can't get SD3.5 to do widescreen as I did flux

[errant dust]

How so?

[sacred geode]

gives weirdness

such as this

[raven fern]

im disapoint man, where are the Thanos pics? :3

[sacred geode]

the workflow from matteo and comfy both refused to do 1920x1024

[errant dust]

I haven't had issues with alternate resolutions:

[robust thunder]

So, all these pictures you guys are sharing here, I could save them to my computer, load them into ComfyUI to see the workflows used to make them? Am I understanding that right?

[runic tusk]

Gone; reduced to atoms.

[raven fern]

wait did you specifically prompt for that kind of style?

[dull star]

can you tell me the prompt

[errant dust]

My prompts are all in the images. I never strip of meta

[dull star]

if it wasn't a glitch this would be very cool

[raven fern]

i also assumed glitch at first

[sacred geode]

@dusky thistle My, God, man, I don't normally grab image WF but Holy Toledo. Damn, lol.

[dull star]

impressionist oil painting

how come it works so well for you

maybe my sampling is bad or im just giving it a certain context that makes it snap out of the style

[errant dust]

I dont understand the comment. How come what works so well? Is it not supposed to?

[dusky thistle]

https://github.com/ClownsharkBatwing/RES4LYF nodes are here just fyi

[dull star]

it doesn't for me

I'm glad it at least does for someone else

[dusky thistle]

using stochastic sampling with third order RES, CFG++, all custom implementations

[sacred geode]

I was going to train some SD3.5 stuff in a bit, but not getting 1920x1024 sucked

[raven fern]

ye, clownshark finished a PhD on sampling, he knows his stuff

:3

[sacred geode]

HAHAHAAHAA

I wouldn't doubt it

[dusky thistle]

lol

[errant dust]

why 1920 x 1024? Strange resolution

[sacred geode]

this is going to need a lot help

[raven fern]

LOL

welcome to comfyui

[sacred geode]

must be evenly divisble by 64

[dusky thistle]

those are the nodes in the link for RES4LYF

they're my samplers

[raven fern]

welcome to comfyui, we have spaghetti

[sacred geode]

not the red ones I cared about it is all the ones I am scared to touch that are there but not connected and some are others not with all thew text

[dusky thistle]

lol

those disconnected nodes serve an important purpose

[raven fern]

i mean you can start with an easier workflow and work your way towards more complex stuff

[sacred geode]

Yeah, Straight outta Compton, yo

[dusky thistle]

my spectator nodes be chillin

[sacred geode]

Yeah, just hanging out, ready for action

[raven fern]

they ready to snipe a fool from afar

[errant dust]

They are known as sleeper agents.

[dusky thistle]

sleeper nodes

[errant dust]

"These are not the droids you are looking for"

[raven fern]

im looking for Android 18

[compact forge]

that is nuts

[raven fern]

must be heaven

[sacred geode]

oh, well I did a pip install -r rquirements.txt all good. DId a manual install of opensimplex. All good.

I do an install missing nodes it says all good I have none. :/

[errant dust]

It might be the size and not the resolution

Here is an alternate made from 1600 x 896

[sacred geode]

not worried about it I am going to go train it

his workflow is missing stuff

and it can't find them. Not a game killer

of course I am in legacy mode as I want no part of the new comfy way.

[errant dust]

Workflow and image info is in the pic

[sacred geode]

I can't wait to train sd3.5 as Flux bent me over

[raven fern]

what stuff is missing?

[sacred geode]

#🆕｜sd3 message

[errant dust]

I suspect the issue with 1920x1024 is that the res is too high

[raven fern]

ye but thats the thing he told you to instal

[errant dust]

at 1440 x 816 it is already hunky dorey

[sacred geode]

no no no, this is his defaults and what his workflow wants

not mine

[craggy crest]

the trainers are being written as we speak

[raven fern]

i know, but you need to install his custom node

[errant dust]

Already I can declare SD 3.5 the king of comics

[sacred geode]

I did. I guess you missed that.

[errant dust]

(default, no Lora)

[sacred geode]

As I said, not worries

[raven fern]

oh so you are saying, its some extra stuff even aside from his custom node?

[sacred geode]

okay, I updated all and NOW it works

[raven fern]

kk

[sacred geode]

no missing stuff. I already git his but something in the backend and I already updated right after comfy added 3.5

[dusky thistle]

that's weird... those are in that repo

[sacred geode]

I know.

[dusky thistle]

oh it works now?

[sacred geode]

I even closed it down and restarted same. I then said to update all and it worked

yes

yep

[dusky thistle]

oh sweet

[sacred geode]

that means I updated 3 times today

[dusky thistle]

lol

[sacred geode]

1920x1024

scaled down

[sacred geode]

@dusky thistle your nodes in the cli is doing this

[dusky thistle]

oops, that's a debugging thing, lol

forgot to remove that

[sacred geode]

I know

🙂

[dusky thistle]

should be gone now

[sacred geode]

1920x1024

[raven fern]

some of these can be nice wallpapers

[raven fern]

what is that border?

[compact forge]

prompted for vhs tape

[raven fern]

huh

[compact forge]

the color depth of this model is perfect

[sullen moss]

At the moment, I’m really pleased with the variety within a single prompt, especially the abundance of different faces. Yes, there are issues with anatomy and complex interactions, but I believe these can be resolved with fine-tuned models. Overall, the model turned out well. It’s only fair to praise the SAI team for this—the criticism paid off.

[compact forge]

no boring realism needed

[dusky thistle]

i've been sayin it for a while: please, god, let SAI release 8b

[raven fern]

i can hear the popo already

[errant dust]

Quite impressive variety of comics and cartoon styles

[raven fern]

i like it 🙂

[compact forge]

popo is interested in fake family photos? xD

[raven fern]

haha

[errant dust]

Same prompt each time

[raven fern]

they are coming for the VHS tapes

[silver sluice]

guys how much time does it take to generate an image using SD3.5 at 1024x1024px with an 8GB gpu?

[silver sluice]

i really liike the shadows on this image, the 3d effect is really trippy

[weak trellis]

a bit slowish, its bareable. Its loaded to ram for me

on 3070 btw

[craggy crest]

i dunno. but it's about 3 minutes on my machine and i have 16gig. you might consider using turbo

[sacred geode]

the sky 😦

[craggy crest]

prompt for clouds

[sacred geode]

why the filth in the sky?

[dusky thistle]

lots of leftover noise in that one

[sacred geode]

yes

[craggy crest]

artifacting it looks like. prompt for stuff in the sky, you'll be less likely to get that

[dusky thistle]

prolly the noise schedule

or something weird with the sampler

[craggy crest]

is your workflow in that image?

[sacred geode]

nah, it is Clowns

[dusky thistle]

sgm_uniform can do that sometimes, it drops a lil too fast

[sacred geode]

ddim

[craggy crest]

switch to beta for the scheduler

[sacred geode]

k

[dusky thistle]

yea hook this one up instead

[silver sluice]

lol that says nothing, like how much is slowish? give me solid numbers, how much ram does the 3070 have? im guessing 8gb bc you responded? for me sdxl is like 15-30 seconds and flux is like 80-200+ seconds

[dusky thistle]

to be precise, it will be more than sdxl and less than flux

[craggy crest]

give me a prompt, i'll run it on my machine and tell you how long it takes me

[weak trellis]

30-40 seconds for 15 steps i havent stopwatched in second

[silver sluice]

30-40 seconds on 8gb of vram?

[craggy crest]

the workflow in this uses the 3 encoder node. you prompt each encoder to its strengths

[sacred geode]

beta it is there but barely

[weak trellis]

yeah and 64GB of ram

[craggy crest]

what's your prompt?

[sacred geode]

a very high quality extremely detailed 4k cinematic photograph, a very sharp 4k 8k masterpiece, of a snowy mountain trail captures the stark beauty of winter in the high altitudes. The trail winds through a dense forest of tall pines, their branches heavy with fresh snow that clings to every needle, creating a soft, white blanket that muffles the sound of the world. The ground is a thick layer of snow, the surface untouched except for the clear imprints of bootprints and the occasional small animal tracks crisscrossing the path. The sky above is a brilliant, cloudless blue, the cold air crisp.

[craggy crest]

to start with, dorp the 4k and 4k 8k terms. they do nothing. let me see what i can get out of th erest of it

[sacred geode]

don't tell me tell Clown

that is his prompt

[silver sluice]

i have a laptop 8gb gpu + 32 gb of ram so I doubt ill be able to reach those 30 second times but ill be happy with 40-50, flux can't do faster than 80 on Q_8 flux-d variants

[weak trellis]

keep in mind I measured it now when everything is loaded in ram. First time loading in ram is limited by your drive(in my case 500mb/s ssd)

it slowly chugs and I can litelarly see my ram filling and then its 30-40s

[weak trellis]

I downloaded Turbo which can run on 6 steps but I tried to bump a bit samples

[dusky thistle]

which sampler are you using there

sd3.5L has a bit of trouble with that scene

i tried throwing in a few different quality terms and it helped a bit, but not enough

[raven fern]

3 minutes with a 16gb vram card? that seems strange, for example on my 12gb card, with flux, its like a minute MAX at those resolutions

[silver sluice]

good point, indeed it does run faster on subsequent attempts, so it's 30-40 after it's been primed let's say

[weak trellis]

yeah exactly, 3.5 is almost same like 3

[sacred geode]

beta

[dusky thistle]

that's the scheduler

[craggy crest]

i make it a point not to argue with sharks. especially sharks with clown makeup

[sacred geode]

euler the default

[raven fern]

kek

[dusky thistle]

euler kinda sucks tbh

[raven fern]

look at this fool dissing the great Euler

:3

[dusky thistle]

it's the simplest but it's just not very accurate

[sacred geode]

I think. not sure I closed it down as I go to linux and train 3.5

[sacred geode]

a lot of discussions for trainers and a lot of funky stuff in 3.5

the SAI decisions made are wonky at best

[craggy crest]

@sacred geode

[sacred geode]

3.5 also shows they didn't throw a lot of money at it (for obvious reasons)

[craggy crest]

workflow is in that

[dusky thistle]

this is what i'm currently using for my sampler

[silver sluice]

has anyone tried SD3 with LongCLIP yet?

[sacred geode]

it worked

[dusky thistle]

res_2s in the lil drop down is almost as good and 50% faster

[craggy crest]

i would love to know how you know this, and some of th eother stuff you've stated

[sacred geode]

I kind of thought the result was a tad better

[raven fern]

it depends also what you are going for i guess. realistic stuff is usually better with the dpm samplers, for cartoony stuff, euler is fine, but then again im no expert, just speaking from some experience i had. now if you want to use custom samplers, then thats a whole other thing. not to mention noise injection techniques and so on

[sacred geode]

would be nice, right?

[craggy crest]

i have several hundred images that show you what it can do. you have to learn how to talk to it though

[craggy crest]

yeah. since you're wrong

[dusky thistle]

i've come to believe that sampler selection is more objective than i used to

[weak trellis]

where is this node?

[dusky thistle]

https://github.com/ClownsharkBatwing/RES4LYF

[craggy crest]

hiding in his shark tank ;)

[sacred geode]

opinions and what we see in the code. You have yours and we have ours. Yes, I said we not just me.

[dusky thistle]

has like 20 samplers all rolled into the same framework

[craggy crest]

oh trust me, mine is not an opinion.

[sacred geode]

okies

[dusky thistle]

once tehy're implemented in a similar manner, you start to see it's really just a question of how accurate it is

RES is basically a patcehd up version of DPMPP

to fix some issues with the math

[craggy crest]

have you tried matteo's node he created for the flux blocks/

[dusky thistle]

nope

[craggy crest]

same neural network

[raven fern]

no wonder it's called res4life

[raven fern]

that is easy to understand

[craggy crest]

so his node should work with SD3.5

[dusky thistle]

what node is this?

[weak trellis]

yepp I memorised that chart right away

[craggy crest]

https://civitai.com/models/878387/stable-diffusion-35-large

[raven fern]

the test is tomorrow morning

[craggy crest]

matteo released a node several weeks back that allows you to adjust all the blocks individually

i don't think it's in his essentials pack though.

[dusky thistle]

ahh, gotcha

[raven fern]

matteo is one of the stable diffusion gods we have

[dusky thistle]

he's a good dude too

[craggy crest]

and major comfy node programmer :) - they jsut don't pay him

[raven fern]

😦

[craggy crest]

course he doesn't work for them ...

[raven fern]

i learned a lot of cool tricks from his videos

[craggy crest]

@dusky thistle your prompt - workflow's in the image

[raven fern]

lol

[silver sluice]

so t5xxl is only 256 tokens?

what is that part about 77 /256 tokens coming from the CLIP encoders? anyone care to elaborate? I understand that the smaller CLIP L is limited to 77 tokens and then G/t5 are supposed to be the bigger ones

[craggy crest]

there are 3 encoders that sd3.5 uses. not just t5xxl

[silver sluice]

yeah i get that

[craggy crest]

you still dont' want to go into ramble mode, even with t5xxl - and you do want to give it your rich detailed prompt. you want to give clip_l the ambient, artsy, background, fine details. and clip_g the 'just the facts mam' black and white text. then they don't battle each other

[silver sluice]

yeah i understsand that much, just wondering if that chart is saying that t5xxl is limited to 256 tokens

[raven fern]

what chart are you looking at?

[silver sluice]

this one, top left, i thought t5xxl had way bigger context width like 1024+ tokens

[craggy crest]

the way i read that chart is that it's 77 tokens for clip_l and t5xxl + 77 for clip_g -

clip_g is your workhorse. it does most of the heavy lifting

[bitter hearth]

there are longclip fine tunes out there
that might work with this, they work with the other models

[dusky thistle]

some things are better, some aren't... still tons of artifacts/structural issues, sd3.5L just doesn't understand this scene well

[craggy crest]

it would be worth trying at least

[bitter hearth]

this guy makes them https://huggingface.co/zer0int

[bitter hearth]

its a nice way to get a variation of an image also as the new clips interpret prompt differently

they don't like to put bins (garbage cans) as much for example

[silver sluice]

yeah i use the longclip finetunes from zer0int for my flux / sdxl workflows

[craggy crest]

there's also a better clip_l out there with better text understanding

[raven fern]

if you read from their huggingface model page, that 77/256 is for training stages, but t5xxl should be around 256 i guess. the others should be 77.

[dusky thistle]

issues are just in the branches more than anything, it has trouble resolving them

i think a finetune will clear this up

[bitter hearth]

the autoCFG guy made the best text encode node I have ever seen
with 856 configuration examples lol
https://github.com/Extraltodeus/Conditioning-token-experiments-for-ComfyUI

[silver sluice]

yeah i was using that one until i learned I can use the longclip finetune over the clip_l finetune, they say they manage to improve the context width from 77 tokens to something like 248

[raven fern]

last time i used the longclip, it produced some very bad results tho... idk

maybe it's better now

[bitter hearth]

its tricky
using alternate clips

my main checkpoint these days is New Reality on SD 1.5 and the zer0int better clip does worse on that model

[raven fern]

well that's the thing with components, people like to experiment and mix and match

[dusky thistle]

i think it makes the most sense to always use the encoder your weights were trained against, and not to swap them out post-training

[bitter hearth]

probably yeah its an ablation

[raven fern]

still using 1.5 eh? :3

[bitter hearth]

I could put a clip switch node in the workflow maybe
and then toggle it every now and then

[bitter hearth]

yeah my main is SD 1.5 now

[craggy crest]

1.5's the only one that understands certain terms correctly

[bitter hearth]

now that Blepping ported DiffuseHigh to comfy
you can make SD 1.5 image at 2048x2048 or more without any upscale, all in one pass
https://github.com/blepping/comfyui_jankdiffusehigh

this is currently the best method on Arxiv so its awesome that we have it now

[dusky thistle]

Lol at that name of that repo

[bitter hearth]

haha yeah
his hidiffusion one is also called janky

[raven fern]

interesting.. i actually didnt know about diffuse high. i know there was something called ultra diffusion or whatever, that is kinda similar, but il check that one tonight as well and see how it works out for some of my 1.5 stuff

[bitter hearth]

ultra diffusion yeah that's actually clown's repo
https://github.com/ClownsharkBatwing/UltraCascade

[raven fern]

no no, not that one lol

let me try to find it

[bitter hearth]

whoops

[bitter hearth]

there's this but probably not what you meant ```Ultra-Resolution Cascaded Diffusion Model for Gigapixel Image Synthesis in Histopathology

[craggy crest]

that could be extremely useful

[bitter hearth]

yeah my main interest in diffusion models is actually what they can do outside of making images
making images with them is just a great way to learn

[raven fern]

shit i cant find it 😦 il try on github

[craggy crest]

yeah. there's a wide range of things they're being used for across all industries now

[bitter hearth]

don't worry I forget the names of good papers for weeks at a time lol

diffusion language models is a funny one
the output texts are very different to LLMs

[raven fern]

ah i got it... https://jingjingrenabc.github.io/ultrapixel/

[bitter hearth]

ah that's the same as Clown's repo

[raven fern]

is it?

[bitter hearth]

yeah his repo is basically getting that model to work well

[raven fern]

well it all makes sense now lol

[bitter hearth]

I really need to try it some time

its a bit tricky for cloud as it needs quite a few files in the right places

[raven fern]

stable cascade was awesome when it came out, but then they announced sd3, but sd3 was terrible... LOL
lucky for us, flux arrived not too late and then we have sd3.5 i guess

@dusky thistle lol, it's really a small world eh, didnt know you did the ultrapixel repo as well

[dusky thistle]

the ultrapixel repo is a port

i did a full rewrite https://github.com/ClownsharkBatwing/UltraCascade

runs faster, more vram efficient, integrates natively into comfyui

[bitter hearth]

yeah I mostly skipped flux
I was looking through my download folder the other day and I really didn't do that much of it
gonna just focus on SD 3.5 now

[raven fern]

you skipped flux? how dare you sir... :3

[dusky thistle]

the original code for ultrapixel was honestly a complete mess

[raven fern]

yikes

[dusky thistle]

deleted tens of thousands of lines of unused code prolly

[raven fern]

lol

[dusky thistle]

all kinds of crazy issues with weird hacks etc that degraded quality, undocumented behavior that diverged from the paper etc

which is why i just did a rewrite

it looks fantastic though

was worth the trouble

[raven fern]

yes awesome work man 🙂

[silver sluice]

@weak trellis 82 seconds on the second run, 8GB vram, 32GB ram, 20 steps, 1024x1024

[dusky thistle]

what really broke things through to the next level was implementing PAG for cascade

[raven fern]

i remember pag, dont really use it these days

[dusky thistle]

i'm not a fan of it with sdxl etc tbh

but with cascade it was transformative

[bitter hearth]

PAG is the biggest image quality upgrade in the last 2 years in my opinion
but only when it is suitable

[dusky thistle]

oh, the other key thing was i trained my own version of stage B lite

results are much better than with the full stage B

images like the above would look like butt

the training for stage B got fucked up i guess cuz it did need patching up

[raven fern]

wait, was PAG before or after SAG?

[dusky thistle]

cool thing though is now you can generate directly at 2560x1536 or even higher without even clearing 11gb vram

it came after

[raven fern]

ah

[dusky thistle]

i got SAG implemented with cascade too, it's helpful as well

also invented RAG (random attention guidance lol) which is great for some photographic styles

[raven fern]

what compression factor you use the most for cascade?

[dusky thistle]

PAG was nuts though

i don't use compression factors

it makes shit way too confusing imo

[raven fern]

yea

[dusky thistle]

the key is getting the dimensions for latent C nailed down

B is just a superscaling model

the best resolutions for C are 24x24, 18x30, and 24x40

then you can kinda pick whatever you want for B

[raven fern]

would you say B is superscaling or refining?

[dusky thistle]

superscaling

[bitter hearth]

there is SEG as well which was a sequel to PAG
its less powerful though but somewhat prettier

[dusky thistle]

and refining

[bitter hearth]

SEG likes to put curtains up and PAG doesn't for some reason

[raven fern]

so wait, what VAE are they using for sd3.5?

[raven fern]

curtains? it's trying to censor the booba man 🙂

[dusky thistle]

cascade doesn't understand much more than a 10-15 word prompt, but it's unebilevable with aesthetics

it's one of the last models that was actually trained on artist names etc

[raven fern]

cascade 2: electric boogaloo

[bitter hearth]

its also a colour thing where PAG likes orange/yellow/red contrasted with black/purple
whereas SEG likes pastel colours

[dusky thistle]

just oozes style

[raven fern]

but i remember cascade having problems with eyes, like it makes them heterochroma, or however the hell you spell that word LOL

[dusky thistle]

lol

[raven fern]

I remember me and @icy drift making some TCG waifu cards or whatever with cascade LOL

[raven fern]

oh no, i think it was like animals

[raven fern]

can you link me SEG? i didnt see that one

[raven fern]

that is the clownshark style

[dusky thistle]

hahah

[raven fern]

you made some loras or checkpoints if i recall haha

[dusky thistle]

yeah, which somehow have more downloads than real models get

[raven fern]

LOL

nice

[dusky thistle]

some weird botnet has latched on and is repeatedly downloading my garbage in an infinite loop i swear lol

[bitter hearth]

there are a few nodes but I recommend these ones for PAG and SEG
they have a really good rescaling called SNF
https://github.com/pamparamm/sd-perturbed-attention

[dusky thistle]

my HF is basically a trash heap

there's no reason it would be some popular destination

[raven fern]

i mean dont lower yourself like that man, i think you are doing some cool stuff

[dusky thistle]

and now some real SD3.5L

[bitter hearth]

the Anthropic botnet is crazy apparently

[dusky thistle]

oh i just mean in terms of how i organized it

i just uploaded shit without bothering to name anything carefully or organize

every few days, just heaved another lora into the same folder with some partially descriptive name

https://huggingface.co/ClownsharkBatwing/CSBW_Style/tree/main nearly 2k downloads last month 🤔

[silver sluice]

• image on the left using default CLIP L, t5xxl fp8, triple encoder node
• image on the right using the fine tuned CLIP L (by zer0int), t5 fp8, triple encoder node
• I can confiirm that LongClip doesn't work with SD3, i think zer0int will have to make a new node for it to properly merge it

[dusky thistle]

so weird

[raven fern]

ah "Smoothed Energy Guidance", SEG, just wanted to see what it stands for lol, i mean il check it out tonight too.
shit i have tons of stuff tonight, sd3.5, seg, ultrapixel, diffuse high, and prob like 10 other things i forgot as of this moment

[dusky thistle]

quick find 10 more

[raven fern]

lol

[hallow lion]

We need some Raid

[raven fern]

wait where is the SNF you mentioned? i dont see it

[hallow lion]

so is segg better than pag?

[raven fern]

shadow legends?

[hallow lion]

smooth energy guidance?

[bitter hearth]

I think SEG is better yeah

however PAG is a bit stronger

SEG looks more attractive to me

[raven fern]

where is snf

[bitter hearth]

full - takes into account both CFG and Guidance.
partial - depends only on Guidance.
snf - Saliency-adaptive Noise Fusion from High-fidelity Person-centric Subject-to-Image Synthesis (Wang et al.). Should increase image quality on high guidance scales. Ignores rescale value.```

[raven fern]

ah

[bitter hearth]

if you are upscaling PAG is better I think

or fixing broken stuff

the original SAG is also very good, its the best for clarity or sharpness still

[raven fern]

i would like to see an anime in that style haha

[abstract nymph]

so how's 3.5? any improvements?

[raven fern]

from 3.0? yes

[hallow lion]

damn pyramid flow, videocogx, tons of flux loras and accessories, sd35, ultrapixel, sana and omnigen coming... madness...

[mortal mesa]

its .5 better

[raw shadow]

sd3.5 large still bad in hands . prompt:a man show his fingers with ring in the street.

[raven fern]

only .5 haha

lord of the ring

[bitter hearth]

yeah face fix and hand fix are needed for now
the ones from impact pack are fine, they do it all in 2 nodes

[raven fern]

impact pack is awesome for a lot of things

[bitter hearth]

could do flux for hand fix
then a second pass with SD 1.5 for hand fix 2nd pass
face fix I would be okay doing 2 passes with SD 1.5

[raven fern]

but wait... there's more 🙂

[silver sluice]

is there any negative prompts for bad hands?

[raw shadow]

flux is good for the sam prompt

[raven fern]

dat moment when flux 1.1 weights release tomorrow, i mean that would be funny considering we just got sd3.5 LOL

[silver sluice]

i kinda daont want flux to change bc new weights would kinda invalidate all the existing work/loras/

[raven fern]

but what is the price of invalidation compared to greatness 🙂

[raven fern]

man that bokeh is too strong

needs some work

[bitter hearth]

it can do really small details it seems

[craggy crest]

yeah, depending on how you prompt

[raw shadow]

why sd3 , sd3.5 or flux do not have inpaint model?

[craggy crest]

in which interface?

[bitter hearth]

flux has an inpainting control net

I agree though, it needs inpainting models, or patches like the fooocus ones for SXDL, or a brushnet

[craggy crest]

flux is never going to get those. but SD 3.5 will soon as someone comes up for air from everythign else they're suddenly developing and creates one

[raw shadow]

this one ? https://github.com/alimama-creative/FLUX-Controlnet-Inpainting It do not work well

[craggy crest]

yeah, cause flux isn't trainable. and you have to train controlnet. give people a few days to work with 3.5 and someone'll start on controlnet for it

[bitter hearth]

that was the one I was referring to yeah
I haven't tried it yet

I personally use the powerpaint v2 brushnet for inpainting

or the SDXL union pro max control net

[craggy crest]

or maybe krita is a good solution as it's a paint program and also runs stable diffusion

[bitter hearth]

krita looks great but I need it automated

[errant dust]

Where is this Clip found?

[bitter hearth]

they are here https://huggingface.co/zer0int

[kindred mica]

sup

@craggy crest how many images do i need to get a controlnet trained? For SDXL it was like 100k or more

[bitter hearth]

any idea why this prompt isn't generateing anime
retro anime style, a woman with long flowy blonde hair, wearing bold makeup, in her casual t-shirt and denim pants, navel cutout, indoors, solid red background. dramatic perspective.

[craggy crest]

i have no idea, i'm not the one that trained any of the controlnets in the past

[bitter hearth]

the estimates for Flux union control net cost were in the tens of thousands of dollars sadly

will be similar for SD3

some are coming from SAI though, which is good

[kindred mica]

where do you get this estimate from good sir

[bitter hearth]

can't remember, either reddit or discord

[kindred mica]

is that like, 2 weeks of H100s?

[bitter hearth]

H100s are about 2 dollars an hour or so

[kindred mica]

nah maybe 4

[raven fern]

first came Alibaba, now we have Alimama :3

[bitter hearth]

no you can get them for 2

[sage burrow]

flat comic line drawing...

[rapid pivot]

I love digimon

[bitter hearth]

a little close but not quite

[sage burrow]

what is your prompt?

[bitter hearth]

flat comic line drawing, a woman with long flowy blonde hair, wearing bold makeup, in her casual t-shirt and denim pants, navel cutout, indoors, solid red background. dramatic perspective.

[raven fern]

i had digimon cards when i was lil, but never knew how to play, so i trashed them

[rapid pivot]

good

joins us on the yu-gi-oh cult

lets burn inside churches together

[raven fern]

but i did enjoy the digimon anime

lol

[bitter hearth]

and btw im using large turbo model

sd3.5 has fingers conveniently tucked away, but so far i like the aesthetics

[sage burrow]

@bitter hearth

[bitter hearth]

nice.. how did you manage that?

[sage burrow]

flat anime style comic line drawing, a woman with long flowy blonde hair, wearing bold makeup, in her casual t-shirt and denim pants, navel cutout, indoors, solid red background. dramatic perspective, by Katsuhiro Otomo

[bitter hearth]

oh the artist ref

so that has to be mandatory?

[sage burrow]

that can definitely narrow it down! Or use a reference image of the style you like

[bitter hearth]

ok

[craggy crest]

controlnet is a huge model so ...

[bitter hearth]

did you know flux can't get this prompt right? a woman in large white t-shirt, bare legs, at a beach as the sun is setting. the ocean water is deep turquoise with orange hue in the sky. but sd3.5 did it at one go

[sage burrow]

I took the artist name off: "flat anime style comic line drawing, a woman with long flowy blonde hair, wearing bold makeup, in her casual t-shirt and denim pants, navel cutout, indoors, solid red background. dramatic perspective"

[raven fern]

hair touching ass :3

some very long hair

[sage burrow]

Oh wait, I was using SD3.5

with Flux, you have to mention anime a couple of times, and definitely want to add an artist reference

[bitter hearth]

no, im talking about sd3.5 for anime related query

[sage burrow]

it cheated, she has a nose and lips, NOT anime lololol

[bitter hearth]

to render anime with flux i usually use lora

like this

[sage burrow]

With flux you can get most things with prompt only (but not nsfw)

there that's better, no nose nor lips 😄

[raven fern]

hot

[raven fern]

is she single?

[sage burrow]

nope lol

[raven fern]

damn, no wonder she went for that dude

:3

[raven fern]

nice

[gusty trail]

My first sd3.5 lora

[kindred mica]

how are you doing anything anime

[craggy crest]

fairly easily

[gusty trail]

Civitai: https://civitai.com/models/880134?modelVersionId=985261

[craggy crest]

you are fast.

[craggy crest]

@bitter hearth @bitter hearth

[bitter hearth]

wow awesome

that was really quick yeah

[rapid pivot]

No

https://tenor.com/view/lysol-spray-spraying-disinfectant-keep-it-clean-gif-16558939

[kindred mica]

these are both the same prompt

[turbid grotto]

cats with no fingers, great choice 😁

[turbid grotto]

what is your opinion about current trainability?

[gusty trail]

I think it is better than flux. 24GB able to do fp16 training with bs1. 100 images, 4rank, 2hours

[turbid grotto]

[spark otter]

I noticed in the sample images that hashtags were used in the prompt. Is the hashtag list extensive? Like could I assume if I wanted a steampunk feel for my picture that I can put #steampunk? what does the hashtag do?

[errant dust]

I'll say this much: after a ton of flack, and deserved, Stability stepped up and delivered pretty much what was hoped for after their initial announcement in March. They had teased the 8B and it was what everyone had been pining for and now they delivered. So a big kudos to them.

[rapid dagger]

SD 3.5

I told it to put the made with AI thing to test something btw

[errant dust]

I'll even add that although it cost them in more ways than one, the stumbles that led to Flux were a bonus to the users. Perhaps not to them, but certainly to the community at large.

[bitter hearth]

Stability have smashed it yeah, amazing model

[errant dust]

We all now have two large and stellar models

[bitter hearth]

I really like how flux looks for some things I just don't like the distillation

[raven fern]

it's good, but im waiting for the usual stuff, controlnets, finetunes, and also training loras

[bitter hearth]

there's one lora 😄

[errant dust]

I like choice, and competition, and we have it in the spades now

[bitter hearth]

someone here made first one very fast

[raven fern]

no im gonna train my own lora :3

[turbid grotto]

it can have more potential than flux in the long run, if it is actually better at tunning

[rapid dagger]

I'm looking into making loras again

[errant dust]

It used to be the competition of the mini 2B models. Not anymore

[turbid grotto]

we are getting updated 2b too btw

[bitter hearth]

some kind of hybrid can be good
make image in flux then second pass in SD3 then upscale with SUPIR (which is SDXL)

[errant dust]

Now we have commercial level quality

[errant dust]

Well aware. Heh

[turbid grotto]

all we need is to fix anatomy and this is dream model

[bitter hearth]

sd 3.5 doesnt treat the keyword girl with immature toddler

[rapid dagger]

it does if you mention 3 yo girl

[rapid pivot]

Can you hires with 3.5?

[bitter hearth]

i dont want to

[errant dust]

I actualy thought the announcement image witha girl on grass a hugely funny choice

[rapid dagger]

I just mean its flexible lol

[turbid grotto]

probably no, tilling is the only way

[bitter hearth]

yeah flexible i like it so far

[rapid pivot]

Why no hires

[bitter hearth]

it goes funny after a certain number of steps

[raven fern]

[errant dust]

I rendered 1600 x 896 images with no issues

[rapid pivot]

Stupid

[bitter hearth]

DiTs don't scale as well as Unets in latent size due to the positional patchwise embeddings
so tiling is the way with Flux and SD3

[turbid grotto]

I managed to fit both, text encoder and sd3.5l into 12gb vram but don't see speed benefits compared to when model was offloading into memory

[bitter hearth]

whichever UI you are using might still be automatically offloading

[turbid grotto]

I am not sure

same was with flux

literaly have no reason to not use fp16

however, it is super fast to load and prompt calculation instant too

[bitter hearth]

how about

wow you just made 8k upscaler

no itsl 4x the size of latent

[turbid grotto]

scary numbers

[bitter hearth]

then i reduced that to .50

so thats like 2x the latent

oh I see

yeah 2x is fine

I thought you were joking by making low quality 8k lol

I'd recommend lanczos for up and area for down

or bicubic if the same node will sometimes be up or down

default size any of these before upscale #🆕｜sd3 message

[turbid grotto]

What will stability do after sd3.5?

[bitter hearth]

live happily ever after

what i have loaded into comfyui ... the model is 16gb, the clips are 6gb+ so that's about 22gb but unlike flux i don't feel my system is carrying a lot of weight... and it takes me roughly 10seconds to render images with sd3.5

[dusky thistle]

i heard the plan is to use the massive windfall to make a hostile takeover bid for microsoft

yeah, it's a much, much smoother experience than flux

[bitter hearth]

how is that possible ... i like it btw

[dusky thistle]

yeah, i'm liking sd3.5L quite a bit so far, i think this is a really promising base model

can't wait to get going on some training projects

and i'm really happy they released it

sai's 8b was the model i've been hoping all year would be the one we could get our hands on

[bitter hearth]

i anticipated the model would stress the system memory but to my suprise its so smooth i dont feel a thing

im feeling this inclination to put flux aside

[dusky thistle]

its' like how in the winter, if your water heater craps out, the best thing to do before the inevitable ice cold shower is to go roll around in the snow in your underwear for a while, cuz then that shower will feel nice and toasty

[errant dust]

Not me. I don't play favorites. I will happily use them both.

[craggy crest]

Use ALL the tools!

[bitter hearth]

there are few reasons i say that, flux is bad at full body images, and also anime with flux is not the same as sd3.5, and the big factor of memory management that sd3.5 is handling pretty well

you also cant use flux for nsfw even for aesthetic styles which you can with sd3.5

[errant dust]

I hear you. Except since I don't think I have ever produced an anime image... It is not a terribly big factor

[craggy crest]

Flux has a very narrow range, yes, but its a good tool for what its good at. Dont toss out a paring knife cause its not a chefs knife

[errant dust]

My thoughts exactly.

[bitter hearth]

yeah, good thinking, but sd3.5 feels like a nice change from heavy memory

[craggy crest]

Even windows paint has a use or two

[errant dust]

I was able to produce a great new logo for my YT channel with Flux, and am infinitely grateful. And don't think I did not try them ALL out there. Commerical and open alike

[craggy crest]

It is :) learn it.

[dusky thistle]

yep... got three trainers on my system here, cuz why not? none of them can do everything that the others can do

[craggy crest]

Link to channel?

[errant dust]

Laugh all you want, but I use it quite often to add a nice red outline square or cirlce to my tutorial articles

[craggy crest]

My comment was serious

[errant dust]

Same

https://www.youtube.com/@chess_and_tech

[craggy crest]

Oh i need to watch you, im lousy at chess

[errant dust]

lol

well, tis me job

[craggy crest]

Playing chess?

[errant dust]

writing about it, photographing it, and teaching it, yes

[dusky thistle]

way cool

[timid delta]

does stable fast 3d model go here as well?

[turbid grotto]

This guy already planning another dish
https://huggingface.co/SG161222/RealVis_Large_V1.0

[dusky thistle]

loving the aesthetics of sd3.5L

[craggy crest]

you should post that in the L3 discord's weekly challenge

[bitter hearth]

Ideogram V2 are firing shots

their new inpainting thing looks like the best inpainting

[craggy crest]

let them fire :)

[bitter hearth]

weirdly the best open source inpainting control net is on Hunyuan-DiT

I'm gonna try it but it needs unpickling first

[dusky thistle]

i believe if you're on pytorch 2.4.0 or higher you can safely load pickles

depending on how it's done in the code

[bitter hearth]

ah okay will look into this

[dusky thistle]

do i place the carnivorous blobplant into that pot

[craggy crest]

you can try, just use gloves

[mortal mesa]

how OLED's grow in urban decay

[rapid pivot]

those looks like something else

[craggy crest]

prompt: on the left side of the image is a dark skinned man wearing black jacket. On the right side of the image is a white skinned man wearing tshirt. both men are running away from an explosion that can be seen in the background

[bitter hearth]

that hollywood leather jacket

[craggy crest]

and special effects explosion

[bitter hearth]

sd3.5 doesn't take negative prompt or does it?

cause their workflow example sets cfg to 1 and has conditioning zero out

[dusky thistle]

it responds pretty crazily to neg prompts, as in, not well

[craggy crest]

it does, yes

[dusky thistle]

you can neg prompt just fine with flux but for whatever reason, it's very touchy with sd3

[craggy crest]

but best practice is ALWAYS dont' use them

[bitter hearth]

afaik cfg = 1 is no effect of negative prompt

[craggy crest]

that's flux only

[dusky thistle]

there very well may be a way to use it effectively and consistently so, but i haven't found it (yet?)

[craggy crest]

for SD3.5, i set cfg at 4

[rapid pivot]

how many steps / seconds per image you guys getting on the turbo model ?

[craggy crest]

8 steps i think?

[dusky thistle]

so for now i'm just leaving it blank mostly

[bitter hearth]

4 steps as you can see in my sampler

[craggy crest]

i have yet to find anyone that uses negative prompts correctly anyway

[bitter hearth]

on my rtx 3060 takes 10s about / image

[turbid grotto]

it is turbo variant's workflow

[bitter hearth]

that's for large model i think, but large turbo uses cfg 1

[craggy crest]

where did you find that?

[bitter hearth]

their workflow example on hugginface

[dusky thistle]

most ppl do just dump word salad into theirs, yea

[craggy crest]

and even those that don't, show by the way they use them that they don't really understand why they exist and what they are for

[bitter hearth]

without negative prompt

[craggy crest]

try my workflow

[dusky thistle]

yea, prolly

[wispy epoch]

https://civitai.com/models/880060?modelVersionId=985214

ggufs

3.5 turbo

[bitter hearth]

im using the same workflow

except you have some parameter values different

[turbid grotto]

smaller than sdxl 😁

[craggy crest]

probably. and are you prompting each encoder seperately?

[bitter hearth]

no, my prompt goes in one positive txt encoder

[wispy epoch]

ill try in few min

[bitter hearth]

doesnt' fp16 require ram about 32gb and up ?

[craggy crest]

my workflow has the triple encoder node, and you're supposed to prompt each of them to their strengths

so - you're not using the same workflow at all

[craggy crest]

no. i'm using that and i have 16 gig vram

[bitter hearth]

hmm i thought you converted them into widgets

[craggy crest]

nope

download the workflow i just posted and open it and look at it

[mortal mesa]

there is no supposed to, but you can

[bitter hearth]

i have it

[craggy crest]

i was talking about my specific workflow i posted to him

[bitter hearth]

btw this scaled t5x has better result as the author says

[craggy crest]

probably. i have a better clip l, don't have that t5 though

[bitter hearth]

oh hold on.. wrong link

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

[bitter hearth]

when you mean beter clip l you are using this? https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/tree/main

[craggy crest]

his latest, yes

[wispy epoch]

sad

[bitter hearth]

see when im using large turbo its not a unet loader, its a checkpoint ... along with triple clip loader

gguf uses unet loader

[mortal mesa]

i ran the q8 sd3.5 gguf, if your doing something like that the node needs to be updated

[craggy crest]

three's a little tiny problem with that. SD3.5 does not use unet - so i'm not sure what that's doing

[noble coyote]

Super SD3.5L (3 x KSampler Advanced - 0-10, 10-20, 20-30 steps)

[bitter hearth]

this anime quality is effortless with sd3.5

and they have good eyes

Sana

not sd1.5 mellow eyes

[noble coyote]

Link for Sana please? 😄

[bitter hearth]

its only demo but they launched it here https://nvlabs.github.io/Sana/

[craggy crest]

3.5 is fun :) isn't it?

@bitter hearth try this prompt: by artist "Chris Van Allsburg":cute, big-eyed, flaxen haired by artist "Jasmine Becket-Griffith", by artist "Cyril Rolando"

[bitter hearth]

the Sana paper examples are a bit better than the demo

I can't get the demo working at 4k, only 2k at most

but the paper has 4k examples

[gusty trail]

because the demo is 1024px demo

[bitter hearth]

ah okay that makes sense

[wispy epoch]

oh okay yes i try to ran same 18 gguf

[noble coyote]

Sana Demo

[noble coyote]

Sana Demo

[wispy epoch]

but where vae?

[bitter hearth]

this is nice

[craggy crest]

baked into the model?

[bitter hearth]

Sana can be decent for quite a lot of prompts

[bitter hearth]

you'd have to load vae node seperately

[wispy epoch]

i mean how to load from node

[wispy epoch]

okay thanks

[icy drift]

OmniGen just dropped! https://huggingface.co/Shitao/OmniGen-v1

[craggy crest]

@fossil pagoda

[bitter hearth]

double click and type in load vae

[noble coyote]

Sana Demo

[wispy epoch]

same error idk why

[noble coyote]

Long necks this SD3.5L

[icy drift]

They released training code, and OmniGen is literally MIT license 😱

[bitter hearth]

yeah there are so many good drops

[icy drift]

OmniGen first try (HF demo)

Quality is surprising. I did not expect that.

[noble coyote]

Yemeni influencer!!! 😄

[turbid grotto]

update extention, it does work

[icy drift]

(OmniGen) Okay, less impressive. Failed the text and the feather in his helmet. Still haven't tested physics or long prompts.
This is an anime drawing of an armored knight holding a sword. The knight has a large red feather in his helmet. He is riding on a giant chicken. In the background, a castle is exploding in a large fireball. The smoke rising from the exploding castle spells out a word in the sky, and it says: "OmniGen".

[sullen moss]

This year has been really productive, but unfortunately, no one has managed to catch up with DALL-E 3, which is a real shame. I’m still amazed at how OpenAI achieved such results and why others can’t do the same

[turbid grotto]

model there has "1024" label

[wispy epoch]

how to update? in node manager i dont see any updates for Unet loader gguf

[bitter hearth]

yeah but their 4k images in the paper were done with the Sana-1.6B1024px model

its probably not optimised inference anyway so we will find out when it drops

[turbid grotto]

check installed extention, press "try update" and restart

[turbid grotto]

a bit weird things, but I am looking forward the weights

howewer, not sure if the are going to release this week due to sd3.5L and M 😁

[fossil pagoda]

🤖

[craggy crest]

:)

[wispy epoch]

there is no update button so ill just reinstall

[icy drift]

I haven't tried Dall-E. Can you give me examples where it beats Flux? I know the leaderboards all have Flux ahead of it.

[turbid grotto]

yea, dalle is unbelivable

[craggy crest]

you can use it free https://www.bing.com/images/create

it's what bing uses

[turbid grotto]

it follows prompt better, quality worse, especially now with crazy quantization

[craggy crest]

it's also heavily censored

[muted dove]

SD3.5 Large Turbo

[turbid grotto]

it wasn't back then, but people were generating crazy stuff, so it is now 😦

[wispy epoch]

noisee 3.5turbo 6 steps

what i did wrong hmhhm

[turbid grotto]

it is just 4 steps?

[craggy crest]

it's from openAI, it's always been heavily censored, and got more so once microsoft got involved

[muted dove]

Normally, but I used 5 or 6 for these

[turbid grotto]

forgot to write prompt?

[sullen moss]

Not always

[turbid grotto]

well, thats fine too, there is also merge between both models on civit

[craggy crest]

wrong sampler. don't use that

use Euler

[turbid grotto]

that is negative

[craggy crest]

i've used it since day one. always

[turbid grotto]

i was getting great results with exact that

[craggy crest]

that's not what was in your screen shot.

[wispy epoch]

just changed

[wispy epoch]

oh okay

[muted dove]

Yes, I made my own with Flux Dev/Schnell and some Loras that I always use. That cuts down on VRAM usage and I get Dev output with Schnell number of steps

[craggy crest]

are you using the same VAE he is?

[icy drift]

Trying to install OmniGen locally... Failed again.

[craggy crest]

that's better. take your steps up to 8 and do NOT leave those other two encoders blank

[turbid grotto]

I think they heavily used synthetic data, like automatically generated cgi with perfect captions, so that is why it is so smart

[craggy crest]

you need to give each encoder a prompt - t5xxl gets the detail rich prompt. clip_g gets the shorter, black and white description, and clip_l gets the ambient, background, artsy, fine details