#πο½sd3
1 messages Β· Page 66 of 1
Oooo
That cat look a little too round, is it ball?
π Iβm such a blonde
Shh itβs just Furr not a ball π
I made it look like a regular txt2img, but I snuck anthro furry into everything lolol (same with my other img2img one π
Big Round Head cat strain I reckon
Nothing wrong with that at all
He may have ate ur ball
Oh shoots
Darn, a human got neutered by a cat, tables are turning at the speed of the light
I replaced all the furry references with Madball. It doesn't work all the time, but sometimes: https://glif.app/@LadyLalita/glifs/clyd621a10000yha2zalqjbb3
It works especially well with monsters π
I made it a bit better: https://glif.app/@LadyLalita/glifs/clyd621a10000yha2zalqjbb3
Monster balls is difficult to talk Claude into!
Telling Claude it's an exptert in art througout the ages probably didn't help
infected by fractal gremlin dna
That sounded like a good prompt to me:
balls
I think my balls has autism
should i be passing SDUltimateupscale a lower shift in general? just had a thought
use my balls lora on the same prompts. it's highly effective
balls lora? for sdxl then?
I'm testing glif to see how much prompts alone can do... I'd say about half of what an actual lora can
uh, what is a glif?
webgen that has SD3 license
lucky i had my glasses on for that one
https://glif.app/@LadyLalita/glifs/clyd621a10000yha2zalqjbb3 for example. Basically you can input various things, such as text2img in SD3, or img2img, then add additional aspects such as Claude helping with the prompting, upscaling etc. You can create your own on there as well from base glifs. It's all online versions of SD3, is why I started using it. My poor GPU lol
prompting is always king. especially now that we have a t5 on a dit network
i thought glif only made realy cheap low brow memes. i didn't know they had actual models hooked up
i always see their glif logo on the lamest of the memes
I started a Monsterballs one, if you want to remix it π
I can't find any for inpainting, nor for lora creation though. So they only do so much.
whats the best way to generate images from some group picures?
Yes
Huggingface has some as well
im happy with my sd3 medium running local on 12gb vram amd gpu
it's just great in flexibility on styles and prompting different subject and compositions. still not great with hands, bodies and poses.
how long now? 1,7 weeks or so before the update ?! π
Hands are the only problem I've been having lately. Fortunately most poses the person just happens to have their hands sideways
I tried "holding a jar" but that was a fail
dont see many eyes, but still two eyes is cute!
and those many eyes ... sorry not cute
I think cute overrides many because to be cute the eyes must be big.
hmm, yes. probably
SD3@ClipDrop
2.5D anime style ... or too realistic?
Can you add Loras and/or Custom models with Clipdrop's SD3?
Depends on your personal preferences
it's a weird scale between 2d anime and full realistic 3d
somewhere in between, the anime style get's totally lost
but where? around 2.62D or maybe 2.95D ?
If the style gets lost, try to add some artist's names which specialize in the style you like
Try Rembrandt within Pony sometime π
wait, ill just add some pony stuff to the prompt of that last image
sometimes sd3 is just lazily pasting text on top of it π
Nice clock
pure luck it got numbers right
NOEDEL
Sd3 does the numbers correct most of the time

Moo clock
No, u just can change resolution, SD3, SDXL1.0, SDXL0.9, Style; and Negative Prompt
Clowck
Can't cut the 'L' away from there
these balls don't lie
oh no, not banana phone again
he was like that at birth so you know
look for "The Original Banana Phone Flash" if you don't know this
ring ring ring
ball sees a waifu
generate a image
that's a lot of banana phones
Can't Wait! π€©
oh no, "coming weeks"
its was "2 weeks"
"still researching"
hm hm, okay
lower all your expectations! β
I'm ready to wait weeks, even months. as long as the model is better than SDXL.
The longer it takes them to fix the problems, the better.
Take ur time
Nah, 2 weeks seems about right
Re-re-re-release in 2 weeks?
or is it re-re-release
50% chance
either happens or it doesn't happen
but my question was not on the timing, but the quality. What will it solve? what issues will remain
Everything will be solved
It will be the best model ever!

that is super great!
like more better than the best! I cannot wait.
damn, sd3 doesnt really know luis royo style π¦
Luis royo, not the style but a slight approximation using sd3. subject and composition are ok, style lacks (probably skill issue, it must be somewhere in sd3!)
sorry, it's sd3, so I need to censor the bottom. Because of bad hands. π π
1 2 3 ALLright! guess the characters!
1 red hair
2 lond blonde hair
3 brown hair
angel wings are just an addon, not included in the characters
frozen, frozen, and ?
no and no. thats pretty 1d
ariel, rapunzel lara?
I'll give a 1/5 point for disney
ok, 2 hits out of 3 !
only 1 disney
oh wait sorry, its still disney
though not a classic disney princess
looks like anna and elsa
if two isn't disney i have no idea (and if one isn't Arial, you got a look-a-like :p)
ur obsessed
you asked. to ME that's who they look like
well, you got a look a like π€‘
lool, sorry crystal wizz
shrug - a lot of characters look very much the same, so if you want people to guess them accurately, don't change them so much it's hard for someone not reading your mind to tell what you are thinking
disney reuses characters and even animation scenes
yeah yeah, its ok
ok, i must admit 1 was a stretch. so, we'll go for another visual hint
she looks a bit older in this image, the show was some time ago
jessica?
i have no clue
reference image
New Kolors model...
no hands and legs, cheating π€‘
Image by Luis Royo
i just hope ip adapters-style or controlnets or something get really really good
I meant use one when prompting.
Or better yet, just create a Luis Roya lora π
For SD3? How ?!
was one of the first things i hoped that worked well with SD3 and now the model are out even more, not being able to ref artist sucks
For SDXL I have one
Wonders how long he had to search for a sfw Luis Roya image to be able to post it on Discord π
lol
(and for kolors too, it's more stupid about artists than sd3, if that's possible :p)
last visual hint for no 1
Find a comfy workflow, or a Glif that includes it. Orrrr, show it to Claude and have Claude describe it to SD3
I thought you meant how with the ref image. For lora use Kohya or onetrainer
or just wait 2 weeks π
hmm, were to talk to claude?
no, I don't think the next SD3 update will include artists
hey, did you ever try my fractal lora?
https://civitai.com/models/404093/yhnrfractalxl i think you can at least use it online on civit (using some buzz)
Those images on the civit site are awesome! π
So hurry up and make an SD3 version of it π
hey, you're not waiting for me!
strategically hidden in her hair π€‘ nah i'm being difficult, i played with kolors myself and an hour ago i was like "wth, all these hands have 5 fingers!!!"
The model should understand what you want from it right from the start, not immediately start talking about skill issues like some do
I still hope that 8B will be no worse than Kolors... Unfortunately, the current version doesn't work for me personally.
not really sure, but that might be nudity π€
omg 2b so wrong, I have n ot used 8b I dont think
You'd think if a chinese model can do good anatomy, SD3 can as well, eventually
Oh, the dream of everyone who understands the full power of DALL-E and what it can do.
It is beyond me, for sure.
I use SD3 to make good morning stickers for my boyfriend 
If you had to pay a monthly sub for sd3 would you do it?
I think I would
Not at the current api prices :p
And blurry pictures...
π¦
aww I miss buzz
What is the back end tech? Unet? MDIT? "German Sausage" named one?
If 8b and ultra are the same, I LOVE it!
I wonder what the original fractal checkpoint was
sdxl's unet adapter to use chatglm as text encoder
and it works sometimes better when prompted chinese π
Yes, to achieve perfect prompt understanding, you need to translate it into Chinese
so now i want the chatglm llm, which you get anyway, to translate my prompts to chinese first, but i have no idea how to do that :p
That's easier said than done. My bf is Chinese; it does not always translate.
either way
just use deepl
yup, that works
what matters isn't if the human understands the translations, it's whether the AI understands
Well, we wait and hope. I believe the SAI team will come together and release what everyone expects from them, not this half-baked model.
Just suspect that since chatglm is kinda made for english and chinese, there'd be a way to translate the prompt with it before it sends to embeddings to the unet
but translating works, it's far from ideal for me, i like to iterate on prompts, that's hard to do when relying on translations, subtly differnt wording and all gets lost in translation
nice for the chinese though, it must have been a bad experience for them to translate all to english for stable diffusion
Deepl doesn't mind NSFW apparently
that gives me ideas π€£
Even upside down, it understands anatomy perfectly.
and neither does kolors too much; win/win :p
Using chinese as a trick to bypass censored sites/etc. doesn't work btw in case that's your idea lololol
Is it better than SDXL with loras?
I am so navie. I do not know any censored sites. But I need to!
This all has left me feeling disappointed... Everyone was expecting a breakthrough from SD3, but the real breakthrough came from an unexpected place.
The funniest part is that the foundation of Kolors' code is actually SD's code.
Isn't that a model that needs an A100 and it's not even particularly good?
Kolors are prudes lol (same with my English version just in case the chinese was the problem)
Well, if it's really necessary, you can always rent servers.
My lady in a red dress standing by a lake, by Rembrandt worked fine though π
maybe the hugginsspace is filtered/moderated
Perhaps that particular huggingface is, but for other models on huggingface, no censorship whatsoever
I just test them for fun, no worries, I have SDXL and SD 1.5
but kolors is far from perfect, it's a bit heavy handed on aesthetics, struggles a lot with styles (maybe prompt issue) and knows no artists. If it weren't for pixart i'd think not using clip kills styles, but pixart does ok-ish
Lol
Try them on grass!!! (kidding)
Well, what can I say, 2B from SAI is much worse...
I tried Kolors today and it doesn't feel like a "game changer" Model.
Honestly, I still don't understand how it happened that the creators of the technology did worse than those who took their base code as a foundation.
#πο½sd3 message let's hope so, with the kinda takeover not a takeover they got resources again
Is Kolors from SD3 or SDXL?
It doesn't look as crisp to me π¦
sdxl vae and unet (but unet trained from scratch)
all this talk about the new rebrand to kolors got my head like https://www.youtube.com/watch?v=rYbrhAk_IQs
"Colors" is a song by Ice-T, issued as the title track for the soundtrack to the film of the same name.
The song was released as a single in 1988
I DO NOT TAKE CREDIT FOR THIS SONG
#ICET #COLORS
Kolors
whats this? how do i get it and does it work in Comfy?
SD3 is still better than adobe firefly? Dayum adobe
colors, colors colors, colors
I'm sad that SD3 doesn't know the word anthromorphic π¦
kolors must be fresh i fidn nothing on in on youtube university
its because the actual word is anthropomorphic ?
Kolors is so good with bodies (but in reality a slight step up from sdxl (prompt is based on old sdxl bot gens of mine, those were ok as well)
Damn typos π¦ Though usually it corrects any if I make them
this looks better than ideogram!
those pics seem like bella is missing. common name but i think you know the one
I triple checked my spelling this time, but still π¦
Gracias.
thats a pretty fast looking hedgehog tho
Porcupine (in theory lol)
I better brush up on my chinese.
colors! colors Colors! Colors! colors colors
So it turns out that when a model is good, it's really good. But when 2B was released, even the most ardent fan realized that it was a mockery of the community
my lora unlocks the model. it understands anthropomorphic now. just needed balls.
see how good it is
gimme ur money!
this one completely perfected the model
I see no NSFW π
I was going to say nor grass, but I seem to recall some balls laying on grass
This entire conversation reminded me of these two. Apparently SD3 has heard of them
need tp for my bunghole
I am the great cornholio
heuheuehhehe balls
looks like a blursed version of emil from nier automata
(i'm not an anime fan, my nephew and brother in law hounded me to play that game for like a year, it was worth it btw... didn't expect to feel the urge to platinum an anime game, but i did)
yeah i'm with ya on that
i've greatly enjoyed some anime series or movies but it's only a small part of what i watch
nier was amazing though
π
SD3 makes the BEST reference images ever!!!!!
fair question
All mine are sd3, if that helps narrow it down a bit
mine were SD3 also
so kolors is good.. how can some chinese dudes make a model that beast sd3 by 78 miles. they dont have millions of dollars laying around i assume..
prompt: reference image
whats so great about that >.>
hmm
still the best SD3 lora in the world
lil known fact, palantir is actually elvish for balls
sd3 only needs balls. its the perfect shape
ok only balls π
little known fact. krang is actually interdimensional alien language for balls
a girl
maybe dodecahedrons cause they're basically low poly balls if you think about it
After the release of Kolors, they are in a difficult position. If they release a new model and it's worse, it will be an epic fail. To rectify the current situation, they really need to release 8B, which must be better than Kolors.
kolors is sdxl modified. they will release 8b when it is ready, and not when the community tries to force them to do so
it's a modification of sdxl
Yes, it doesn't matter. Whatβs the point of new technology if it delivers worse results?
it doesn't. your argument is invalid
With LoRA
It must be just me? Colors tp me looks not as vibrant, nor precise.
I prefer Dali or Pixart
Without LoRA
What tech are they using? MDIT or UNET?
pointless technology is balls too
You only like REALLY BEAUTIFUL things... you're no fun π π π
An sd3 lora? π
spytech! i ahven't played much wiht it
Er, I should know this, someone mentioned it earlier today.... ??
Just uploaded three new versions of recent training sessions.
Use at 50% seems to do well. But YMMV. When it works it looks cool
kolors is a sdxl architecture model trained from scratch. so a unet
secret agent ball π΅ secret agent ball π΅ no one knows hes a ball cause he's a super secret agent ball πΆ πΈ
Happy Accidents?
little known fact the larval stage of fairies are actually balls in the woods
That is an empty prompt
If you just hit queue with emty prompts you do get some interesting seed results time to time.
Ballz in the woods? I raise you Spies in the woods
declassified ballien evidence. how can you explain any of this!?
factual facts
Also kinda hard to get ball ufo
prompt:
we are working hard to discover all the secrets of ballz
SPCB will not be pleased with this!
β€οΈ
Lmao
sd3 at work
creature of the depths
its evolving
the dog after it drinks your beer
Yo why this actually so good?
its because
SDXL had issues with dark and black tones
as well as colours
that issue is fixed in CosXL as well as in SD3
Did my best at recreating it but how come the guy's face on the left is so weird? I feel like everyone's sd3 images look so much better
I'm not using SD3, but has anyone tried SDXL embeddings with SD3? Since the Clip encoders should be the same they are supposed to work
Know someone was making embeddings but donβt remember if anyone tried it. I imagine that even with the t5 not running the architecture difference wouldnβt make something productive.
Try it and tell us.
/Please follow the structure of the house in the uploaded picture to generate the exterior decoration renderings of the house. The decoration style is modern and fashionable, not rural style. A half-walled courtyard is added to the floor, and there are flowers and plants in the corners.
I can't even run SDXL on my computer, and to be frank I don't like at all SD3, I was wondering if anyone tried it and whether they had better outputs
What will the follow-up "full" SD3 model be called?
its gonna be a while
because its not just a case of waiting for training time
its a case of working out how to train DiT
Is SD3 ever coming back to Civitai?
You may be able to run that setup via glif
I'm glad to see my glif working as it should π
Prompt: "3 ladies skateboarding"
Results may vary π€£
Can you tell I got boredom face only, as well as just standing there modeling poses?
These sorts of things are just gonna get worse. SAI's most recent post is about introducing even more censorship specifically to protect "AI Children" π€£. So expect more mangled bodies, less youthful faces, generating families and any other unexpected consequenses of this
Eh, SAI will probably fail before of that
Yea, but, think of the chillren.
I need sdxl inpainting now π
I changed my comment on it,it wasn't as funny being completely sfw
Lol

Getting spicy lol
Cat being catcalled and immediately reacting (this was the prompt 100%)
This looks more like a cat that catcalls other cats
F you wolf! lolololol
If anyone wants to try my txt2img sd3 glif π
https://glif.app/@LadyLalita/glifs/clycy3ugs0006dfk4a72wklco
Try such prompts as "a beautiful woman" π
My oh my lol
It's heard of all the good artists!
SD3 seems 100% complete to me π
more furries than balls
Well once civitai and SAI come to an agreement, perhaps we'll have both π
Man, where's my balls, darn
It's alright
Are you stuck in spring?
where are you at?
Balls heaven
This is where marbles go when toddlers swallow them
|| π§π· || 
Current sd3 model, I see no problems here π
we call him RoseFist
Civitai needs to fix their site β¦
My furries are just too sexy for glif π¦
is civit broken again?
If it hates urs itβs really going to hate mine
When is it not broke the site works like one time outta the month π
its like a rolling broken at all times, someone somewhere is affected
This is why I refuse to pay for buzz itβs a waste of money
Broken still
ya i hear ya, i got pissed off over assorted things and deleted my account
hah got em
@odd basalt 
I use their site to make loras
I donβt blame u I just went in lol to claim my daily buzz π
Their front page images have drastically changed lately!
Yes same β¦ I did that with chat gpt making Loraβs and using it for JSON jailbreaking and gaining data that got me in chat time out I go deep into their server as I lost meta on fb for getting information they should have not given me in the first place β¦ and why meta has accesses to locked hacking tools is beyond me.. but shhh π€« I did not share anything they donβt want anyone knowing π
Yes I see that we are not introduced with pron hub
BTW do you know of any other online lora makers? π
i clicked the show less anime and it went from like 90% Anime to 88% Anime
Free ones on hugging face or u can just train model Nightcafe Leonardo chatgpt etc
I mean without using my own gpu deprived computer. Are there any huggingface options?
What are u using atm???
Iβm on pc right now and can find something copadible with ur device
cloud gpu are surprisingly cheap
Vpn?
8gb nvidia gput, 16gb ram, i7
why would you need a vpn?
I'm worried I'll be on it for 10 hrs just trying to figure out how to setup Kohya or Onetrainer π¦
So it mask ur ip and devices β¦
but what does that have to do with cloud gpu?
I've never had google drive, nor anywhere else mind all my explicit nsfw futas, so haven't bothered with vpn yet
Cloud gpu gives u free but limits u to spend more and I have a gaming pc I donβt use cloud gpu
you can't get a very big cloud gpu for free
there is colab and kaggle
but I think its 16gb
Thatβs why the vpn
Is important key if u donβt wanna spend real money or bit coin
I do pay for 2gb google drive each year, however th e free collab that comes with it, all I've read seems to say that lora creating needs more gpu than the free they offer π¦
Create 10 more google accounts? π
Def get VPN Ull thank me later⦠and I have too many google accounts
Nailed it when ps was free trial ha loophole back then
Like I said donβt waste money on new machines or paying money into aiβ¦u can do it for free with loop holes π
Has anyone created loras using the google collab free GPU?
This is how I see it if u can run after effects smoothly u can run anything talk about graphic cards u have a good machine
Or I should say has anyone created SD3 loras. CIvitae creates them so very easily for just buzz
No I have not tried too
I have a few thousand images ready for when SD3 releases the final info π
Or before that if I get bored
Is there anything I can use on my own system? 8gb nvidia gpu, 16gb ram, i7
thats alotta images
Looking now
?? you don't like the replicate trainer?
I just lost 5000 images the other day π¦ Fortunately I've nearly make up for it π
Oh? Would that run on my system? (8gb gpu)
Can that train SD loras?
i gave you the link in DM. you just run it on replicate - it's what i used to create hammo
π§
Their site has worse navigation than civitae!
Interesting developments in the past week. I just saw Friday's announcement.
Url ?
wow
i borked upscale on it, extra faces in the black in tiles haha
this is the image of all time
I thought civitae was worse least on iPad now that I understand it π and civitae is still broken π site
good if you're a swimmer though
When you grow up in areas with that kind of heat, you know exactly how hot that pool will be be on a day like that. Shits like a hot tub.
also, everyone else is going to have the genius idea to go swimming too, then you got too many people in the pool and they're probably dirty
Everyone is dirty though, including you and I. But yeah, some more than others, that's for sure
If a pool smells especially strongly of chlorine it could be cuz ppl are peeing more than usual
Urine reacts with hypochlorite and makes chlorine gas
welcome to my ool, notice there is no p in it, lets keep it that way
Yeah but it makes a smell all on its own anyways, with our without anyone touching the water.
(I'm saying this as I check our hot tub water levels lol)
it does, but it's more fun to associate the clean smell of chlorine with the filth of even more urine lol
Forbidden anatomy
OK I've got 5 days to use up 500gb of data, SD3 checkpoints and lora links please π
turn on netflix and watch a couple movies π
A lion and four bulls living on the savannah, with the lion and the four bulls being enemies, the scene is full of tension. the four bulls is nearly, face the lion, the bulls number is four, the bulls number is four
big
not sure I want to run SD when it's so hot, make my office too hot then too
guessing that is the 8b?
the structure is really good
ella off sd 1.5. π
stuff keeps getting better.
ah SD 1.5 with kolors refiner
do you use Hi-diffusion or Koyha Deep Shrink? I find they are nice with SD 1.5
the reason I use them is they can push the resolution up a bit cos SD 1.5 starts small
i'm not, just regular generate, upscale again with the same model by 1.5x, then a 2x upscale with kolors
ah okay
if that works then that method is fine
sometimes I get artifacts
Hidiffusion and deepshrink both have the bad side effect that a smaller latent can make the structure of the image not as good
yeah every time i've played with those, I've always ended up switching back out of it
manga style, closeup of an anime woman with long white hair and black eyes smoking a cigarette, wearing a tightfitting dark top, pale skin, beautiful face, long eyelashes, luxurious fabric, high detail, beige background, soft lighting, halfbody shot by Luis Royo
they are maybe not best for most people who are targeting a 4k final image after all upscales
the reason I use that stuff is I am targeting 8k-32k final image size
so its a bit different
interesting. let me try that.
I kinda mash a bunch of workflows from reddit and github together
generally its this:
deepshrink -> DAT2/ATD/HAT-L upscale -> SUPIR upscale
that's really nice
have a prompt? I'll try it
"A detailed illustration of R2D2 navigating through a lush, vibrant jungle. The scene is filled with dense foliage, towering trees, and exotic plants. R2D2's metallic surface gleams in the dappled sunlight filtering through the canopy, and he appears to be on a mission, rolling over uneven terrain with vines and roots underfoot. In the background, colorful birds and curious animals can be seen, blending the world of Star Wars with the natural beauty of a tropical jungle."
oh wait its SD 1.5 this might not be good prompt
nah that's a good one
I only do image AI to make sci-fi
so that's the sort of thing I like to test with LOL
haha R2D2 went to the gym!
its really great
"A realistic photo of spacecraft engaged in a dramatic battle in the foggy atmosphere around a towering space elevator. The scene is filled with dense fog, with the enormous structure of the space elevator stretching from the ground up into the sky. Bright laser beams and explosions light up the fog, casting eerie glows on the sleek, futuristic spacecraft. The spacecraft are maneuvering around the space elevator, creating a dynamic and intense combat scene. The background hints at the vastness of space, with distant stars barely visible through the mist."
only half the models I try have space elevator in the training data though
wow this is so good
Models should understand the concept of a space elevator even without many clear examples of them
sometimes they do and sometimes they don't
Dalle 3 is by far the best for sci fi
Skill issue runs away
lol
its ok, maybe it is skill issue
yeah no question.
in the last year I spend 99% of my efforts on upscaling
and only 1% on actually making good images
so I am pretty behind on that area
well, the modesl that have been coming out just in the last 3 months have shot us all forward by a lot.
Space stuff
kolors looks really really good
it seems to bake in a midjourney look, which is good for me I think
yeah your one was better
the thing about dalle is that
it takes a ton of cherry picking
the top 1% of dalle seeds are amazing
also bare in mind there is dalle HD mode in the API, its a bit better
yeah that's what i have it set to for that image above. 1792x1024, hd
A towering space elevator stretching from Earth to space, nestled in a vibrant forest with glowing trees. Shimmering silver cables glint in the sunlight. A sleek, glass-walled capsule ascends gently, reflecting the azure sky. Clouds swirl around the structure's midpoint. Astronauts wave from inside, smiling. Futuristic control panels with colorful buttons. Birds fly nearby, curious and bright. A holographic sign displaying "Welcome to the Stars". The Earth below, lush and green.
yeah I like forest sci fi a lot
hmm
Ball space elevator 
its kinda a cute style
its still valid though
Where's my nature 
can anyone do this one:
"A dynamic and detailed illustration of TIE Fighters flying over Maz Kanata's castle on Takodana. The iconic castle, with its ancient stone architecture and lush surrounding forest, is depicted in vivid detail. Above, TIE Fighters zoom through the sky in tight formation, their sleek, menacing forms casting shadows on the landscape below. The scene is set during the day, with the sun casting a warm glow on the castle and the forest, highlighting the contrast between the serene setting and the imminent threat of the TIE Fighters."

thats really nice with the trees in the foreground
these are sd3 8b
ella/sd 1.5 didn't know tie fighters, but this was hilarious


the flying storm trooper is amazing
these castles are better than the real one TBH
got one more if someone wants
"A realistic photo of a massive space station crashing into the ocean, with a vivid, colorful nebula visible in the background. The scene captures the dramatic moment as the space station breaks apart upon impact with the water, sending huge waves and spray into the air. Debris and fire trail from the station as it descends, and the surrounding ocean is turbulent and churning. In the sky above, the nebula provides a stunning and otherworldly backdrop, its swirling colors contrasting with the chaos below. The overall image is a mix of destruction and cosmic beauty."
From a prompt on the artisan channel
Haha it made it a festival
Hah very cool. Have a prompt for that? It's a neat style
This is based off sd 1.5 level4 v5 with Ella for text encoding, with Kolors upscale
"wire art" and neon whatever
It's just a simple prompt
ah I used the SDXL version of level4 a lot
didn't know there is an SD 1.5 version
Yeah they're both fantastic
can someone try this one
"A dark and atmospheric scene of a colossal starship emerging from a swirling, luminescent wormhole above an alien planet's horizon. The starship, adorned with intricate details and eerie, pulsating lights, casts a menacing shadow over the landscape below. The alien planet's terrain features jagged, crystalline structures and dimly glowing bioluminescent flora, contributing to the ominous ambiance. The sky is dominated by turbulent storm clouds, with flashes of lightning illuminating the scene intermittently. A vibrant nebula and multiple moons are faintly visible through the storm, adding to the celestial drama. The entire scene captures a moment of foreboding and discovery, blending the marvels of advanced technology with the haunting mysteries of an alien world."
Neon wire art for everything
has anyone had more luck training anything? i see kohya bmaltais ui has an sd3 branch now
Did some over in #β¨ο½sdxl message
Balls
A giant, pink, dragon-like creature with enormous eyes and sharp teeth is opening its mouth wide. An elderly woman with long, white hair is wearing a dark coat. She is holding a small plate and feeding the creature using a fork. The background is soft and misty blue, giving a calm and mysterious atmosphere. Both the woman and the creature are illuminated by a soft, diffused light, highlighting their textures and expressions.
an angry gray yeti walks into the mountains leaving footprints against the forest and turns around. gm
Are there any SD3 devs here who could confirm if this is the correct loss formula for finetuning this rectified flow model? Diffusers has an example using a different formula, but I think that might be for the EDM approach which the SD3 paper mentions as giving worse results
model_pred = transformer(noisy_latents, timestep, conditioning, ...)
loss = (model_pred.float() - (noise.float() - latents.float())) ** 2
loss = loss.mean()```
Bitte erstelle ein Bild fΓΌr dieses MΓΆbelstΓΌck der Zukunft:
Der Stand.In besteht aus zwei senkrecht ineinander verschiebbaren Korpussen, welche aus Tischlerplatte mit sichtbaren Winkeln und Schrauben zusammengesetzt sind. Mit Hilfe einer manuellen Kurbel ist der obere Korpus in der HΓΆhe verstellbar. Die Form des Stand.In sieht vor, dass es eine ArbeitsflΓ€che um den Nutzer:in herum gibt. In der ArbeitsflΓ€che sind ein winkelverstellbarer Bildschirm, eine Tastatur und ein Trackpad eingelassen, welches sowohl digitales, als auch analoges Arbeiten ermΓΆglicht. Durch diverse AnschlussmΓΆglichkeiten kann man seine eigenen GerΓ€te zusΓ€tzlich anschlieΓen. Alle sichtbaren FlΓ€chen sind mit Naturharzlack beschichtet, um eine widerstandsfΓ€hige OberflΓ€che zu kreieren.
es gibt hier nicht viele deutschsprachige, fragen sie auf englisch
are people still using sd3, or is it dead?
Create prompt keywords
tons are using it. though most of the people that aren't, are the people that want to make lewd waifus all day
yeah
Image
depends which SD3
the API one is more baked
until there are fine tunes I would run a second SD1.5 or SDXL model over it as a refiner
apparently the new one, kolors, makes a good refiner too
I wouldn't stick to just using one model, you get slightly different feels and effects from different ones
SD3 has no trouble making lewd waifus! π
Oh wait, you probably meant something else lol
I have a collection of mayo in my basement, somebody wants to join my feast?
Game Ready topology.
looks like SD3 Large
What is that? 6B? Or 8B?
8b
Kolors -_-
Damn
kolors has been kicking ass for me
because eveyrone was expecting soemthing like kolors when sd3 came out :/
we only released the 2b model (and that was supposed to be a beta)
Was this made with 2B?
no, 8b
you can try 8b on api if you want
(for now)
π€·ββοΈ
fact is kolors is here it runs on whatever less than 10GB and it has a straighforward apachi 0.2 license
current SAI license is not too different from Apache 2
despite what people say
it's worded differently, but conceptually it's apache2 until 1m$
But Kolors only works really well in Chinese.
you can already train something similar to Kolors using SDXL and adding a vlm input.
||it's basically what it is||
I think the dataset is also publicly available
yes bascially it measns do whatever until u make 1m a year not sure why but ppl are still scared
i promt in english and chinese and results ar eok both ways - different but ok
still thinking about the old license maybe, gonna be hard to counter the first impression
the previous license was just.... bad.
u need to convince ponyman
also it's easier for bad news to spread comapred to good news
Ponyman chekc out the new license!
also civitai wtf... maybe they make more than 1m a year
its still banned i mena ugh archived
if civit wotn lift the ban thats almost a deathblow coz hardly anyone will search for anything anywhere else its like tha monopoly
Question is [WHEN] ?? π π
2 weeks.
sure, let's do another rushed release
Maybe hype was not a good idea after all
I had a serious question though... when... I don;t care if you take 2 years to do it ... question is still relevant don;t you agree? I did not say "when in less than a day" I said when... So, you could always say "to do it not rushed, we are looking at 4 months" or something like that. π
We have no idea is the reality
it's a business decision
I'm no biz guy
DISCLAIMER:
I
KNOW
IT'S
HARD
AND
EXPENSIVE
TO
TRAIN
A
MODEL
END DISCLAIMER
the ai world moves crazy fats tho so no one knwos anythng
literaily any day a game changer model can drop out of the blue
whatever I might say to you now might be contradicted by biz, and then people will make reddit threads about how I'm a liar 
No harm no foul. I believe you are loved in the community, myself included. YOU are not the target... I appreciate [you] Lykon.
lykon is king
I hope the 800m model will be done, it will be way easier to unfuck and will require less compute to fine-tune
lykon and ponyman will save the earth, it slike x men
800m makes very little sense with the current architecture.
Disclaimer: in my personal opinion
and it's sdxl
Next mission: unfuck the sd3 models.
Is their another virtual gpu service similar to google cloud that allows stable diffusion use?
current mmdit arch works well at hig param count. Still figuring out low params.
go look at the code. i'ts a modification of SDXL
Or do I need to connect my S.D. as an apache or tomcat?
ye i know its based off sdxl.
what are the main things to unfuck in sd3? I haven't used the model very much since the first license was very bad
so what you really want is SDXL - no need to wait or ask for anothe release
in ym headlore crystalwizard is emad
2B at current state is worse than SDXL, the 8B model will be trained only by people that want to make money out of it even before starting to train
kolors is not just sdxl its feel more like a more refined cascade - they did improve on the sdxel architetcure
only if you don't know how to use it. lots and lots of sd3 images in this channel that are much better than sdxl
it's much superior to sdxl for some use case, it's inferior in image coherency.
It is much better in prompt understanding and fine details
far better upscaler, far better interactive model
Colours has muted colours usually though π¦
but unet, at 2b params, has better coherency, for sure
hmmm really - get very vviid colors with it..
(but you can't really scale unet to 8b params)
It's a very undertrained model though, and making a 3.1 model will not start over the pretrain but rather fine-tune on top of it
here sa trick question - whats better: sdxl or ella with sd15? π
for which use cases?
just base generation
depends on the finetune
promt following mostly - i refine with whatever anyway
better is a comparison word and means nothing. it's subjective and you have to have something to compare
Comfy workflow link? π
I personally prefer sdxl over anything 1.5 just because Turbo exists
sure you give up the negative prompt
but you generate 10 times faster
it aint subjective - sdxl alone is better at promt following than ella sd15 or not - im justa skign becauspixart sigma is good but slooooooooooooooow
are you still trying to figure out how to refine SD3 output with sdxl?
the word 'better' is a comparison word, and it IS subjective. what is 'better' in one person's opinion is 'worse' in another's opinion.
Fortunately I got over it in about 10 seconds flat π
bad idea, sdxl vae is too much inferior
might work if you then refine back with sd3m
geeez man ar eu a robot or a human lol
someone said they did that, and she's been trying to figure out how since
also can we please stop saying "sd3" when referring to 2b?
it's sd3m or sd3m beta π
i'm sorry, but as an AI I am unable to answer that question
i knew it!
Looks like they would. Esp if before expenses.
revenue is before expense
When making a million is a problem, small money small problem, big money big problem.
fyi, enterprise license it's not a problem
the fact itself that requires private negotiation means it could have any custom condition
(it's a contract)
yes thats the part that sucks
that weird individual personal situation stuff like yeah ok so ill ge tin touch but who knows what happens
Dodge needs an enterprise license that'll let it go make crypto
how? At the moment SAI is the company spending millions to make free models, companies that make revenue should contribute to that
just makes a better environment for everyone
you're arguing with someone that doesn't know how to run a business and doesn't understand what an enterprise license even is.
im dodging the coins my man
Lots do, with already made workflows. Huggingface lists a few (I'd trust that source more than others)
i do runa business and no the problem is what is said - "contact us if you make more than a cool million" is not very encouraging
what if they dont reply for a month or between eahc email a month or two passes what if they just dissappear
not onyl that even the enw licens eis revokable at any time so i mena wtf is that?
the license is not "revocable at any time"
they cna pull the rug whenever emad has abad day or whoever the lord boss is atm
Emad? π
lol
It's better at works by Rennaisance and VIctorian artists!
he is the SAI CEO of my heart forever i dont care whos there now
I have a feelign well be arguing and bickering about this while the Chinese who dont give a flying this and that about copyright and licenses will take over the AI world π
Is there a kolors channel you have been posting them to? I want to see :). Though it prob doesn't run on 8gb gpu lol. Also it doesn't do NSFW π¦ π¦ I'm still curious though lol
if you make over a million, you don't want a cookie cutter contract, you want one that is designed for what you do. that's why you contact them. and if you don't make over a million, then you don't need to worry about it.
hmm lets see if kolors does boobes!
It knows Peitre Bruegel the Elder, Waterhouse, and Bosch. It's perfect, it'd done! β€οΈ
Kidding lol
Maybe
What is ella?
you raised 2 points:
- the license is revocable
- what about a grace period after I reach 1m
Point 1 is the same of apache2 in the current license. Worded differently but same meaning.
Point 2 will be addressed in the faq (tl;dr it's in our best interest to have you pay more than 0, so to give you the license, and it's not revoked in the instant you make 1m. Plus we are no police and all of this is self-reported by the company)
I keep seeing people saying the turbo puts out lesser quality, is that true?
there is gonna be a huge difference between a company that just hit 1m, has contacted us and is waiting, and a big company that makes millions for a long time and at some point is found out that they're using sd3 without ever contacting us.
yes lighting and turbo degrades quality in favor os speed
Have you ever tried hosting stable diffusion using apache or tomcat before?
I'm thinking on how I should do it. And host it myself for a friend
I have yet to find a SDXL model (non turbo) that's better overall than DreamShaper XL Turbo 2
despite the fact that I haven't updated it in months
I have that figured out fortunately, but my GPU is thinking of striking π
funny thing about that - i just realized 2 days ago you wrote that
Hmm, ty for the suggestion
Kolors everything + SD3 text rendering > Ideogram.
kolors has very limited range compared to other sdxl models
hmmmmm
good thing about Kolors is good image understanding for Unet
and Chinese text
but range is because of the dataset being a small synth one
anyway whatever as far as generating stuff i think we are there - please. 100% consistency when?
I haven't! My skill level is more along the lines of asking Gemini advanced for step by step instructions for how to go about it π lol
Kolors is kinda amazing. Whoever said it only works well in Chinese has no idea what they're talking about.
if you want consistancy, you should be using scenario
Thank you π
Kolors ... where can i download that ?
It won't do π though π
donwload the comfyu stuff and run the workflows - it will automatically dl the models
i got a nude the firts try...
thanks
Did you download a model from civitai yet?
More kolors stuff as a refiner for various things.
about 60 or so π
Awesome
I'm looking into paperspace to host and run stable diffusion
I'm not at home to play with apache or tomcat π«
The fun thing about having SD hosted elsewhere is that you can use it anytime with your cell phone.
Though I cheat and just use Mage
it's a very pleasing MJ aesthetic, since that's the dataset they used, but has its limits. No traditional art knowledge (basically 1 style only, which is "AI generated pseudo-realism, or "pseudo 3d render"), samefaces everywhere, and essentially can't do anything that can't already be done with sdxl + vlm (which is a cool concept).
Disclaimer: personal opinion
My fave aspect about SD is the fact that there are hundreds of checkpoints and loras for it. And I can easily make my own as well. The SD3 ones should be amazing.
I used to just run it with collab..
Until it git banned
typical mmdit issue, funny to see unet make this mistake π
It's a site where you can use SD (1.5 or SDXL) models and loras. But they also have some pretty awesome extras integrated that make it come out good even for newbies. It's how I learned to use SD.
But it doesn't have SD3 (yet at least)
Does it allow in and outpainting?
not good enough and i wan tlocal tools
that's his friend who is laying under the table
then you need daz studio
omg i used daz befcore no.... thats not what i need
sure it is. you create ref images with it to use with comfy
People would get excited about SD3 if they could see some decent community finetunes. The previous base models weren't all that exciting either. At the moment though the community can't be confident of how to correctly finetune it, and any published info from SAI would probably help with that. e.g. I'm not confident if the diffusers SD3 finetuning script example is correct (they forgot the VAE shift factor and near 50% TE dropout), and think they're calculating loss using a method which the SD3 paper rated as the worst. Creating the conditioning in various dropout combinations is also fuzzy, since comfy seems to drop the zeroes as well in some cases, or shift the position of clip_g embeddings if clip_l is missing
perhaps the problem is that it's 2b and unfinished?
yeah the basemodels where never perfect... i cant believe its been about a month and nothing sd3 related is out- it is very sad, i really think sd3 is a great base model to work from.
the model can be finetuned, it just runs into issues and it's hard to know if it's from finetuning it incorrectly or the model not being easily trainable
Daz AI studio - consistency
For me, SD3's VAE is the first which can encode my art style without losing the details on the characters' eyes, since the 4 channel VAE didn't work well with flat shaded artwork, so I'm really keen to try finetune it, but can't seem to quite make it work like previous models
the model is unfinished and missing things. hard to fine tune that. not what i would even try to do.
Inpainting comeing soon, they had it on their old site... not outpainting though unfortunatley.
I already know it can be finetuned on tiny datasets and have done it reasonably well, it's just a struggle with larger datasets and it's unclear if there's an error in training that just needs to be fixed, or if it's due to the model
could also be settings
any published details about SD3's training would help, particularly about how to build the conditioning with various dropout combinations and the loss calculation
I agree
that's why you have sites like https://imgsys.org/rankings with 10s of thousands of real world A/B ratings, using prompts ranging from good to garbage, rather than the biased results that commonly end up in the papers
which is still opinion and subjective.
right, but when you have very large sample sizes, you end up with what's known as a trend
