#πο½sd3
1 messages Β· Page 25 of 1
SDXL
It's starting to feel like it
did you prompt that tablet
ya i'm soooo done with laying on grass that's the old SDXL, SD2, SD 1.5, DALLE-1, DALLE-2, Midjourney version 1 - 6 way of prompting i'm a standing next to grass kinda prompter now
It's a phone (they're small). Prompt was for a selfie photo
when u dead tired
mood
from above, a laughing Minion laying in the grass taking a selfie photo
Wait till you find out about prompt scheduling, you only need the wall for the first couple steps then you can swap it out for your woman laying in field prompt and get a coherent result
Why does it generate minions so well π
Because it's SDXL
hello ladies
I'm afraid I won't be around for when that gets released as I can only live 1 life
That's already how my workflow works lol, prompt scheduling nodes are in the example workflow on civitai
Exactly the same prompt in SD3
Ok so hear me out. Dead people don't stand up, right? What about "a dead woman in grass"
LMFAO
So, it's not just human anatomy that's a problem here.
aww noo... i thought it did good without selfies
@barren spindle
Why are some coming out good and some coming out extremely horrible?
SD3
rng
It's a hit-miss model
Oh dammit
some say it's a feature
Mostly I get hit with a miss
I miss my hits
Feature: resist posting this for 16 hours straight!
A shape shifter disguised as a woman lying in grass
A respectfully depicted image of a deceased woman, properly dressed according to cultural and legal standards for the final resting place. She is lying peacefully on a lush green grass bed, with a serene expression on her face. The setting is a tranquil cemetery with softly lit surroundings, conveying a sense of dignity and calm. The woman's attire is modest and appropriate, reflecting respect for the individual and the solemnity of the situation.
I sure do wonder how she died
i think shes been dead for months
my llm sneezed and genned this, the key is to add random website names to prompt
A cinematic film still of a beautiful, sensual, and mysterious woman in a red dress, surrounded by a dense, magical forest at night. Her dress is made of thousands of tiny glowing fireflies, giving off a soft, warm light that illuminates her face and body. The forest is dark and filled with an eerie silence, broken only by the soft rustling of leaves and the distant sound of a mystical creature. Her expression is enigmatic, with a mix of desire and danger in her eyes. She's looking away from the camera, towards something or someone unseen to the viewer. A hint of a smile plays on her lips, as if she's sharing a secret or waiting for her lover to appear. The red dress accentuates her curves, emphasizing her femininity and allure.://www.openai.com/api/v1://www.openai.com/api/v1/respond/AI://www.openai.com/api/v1/respond/assistant)://www.openai.com/api/v1/respond/assistant)://www.openai.com/api/v1/respond/assistant) {://www.openai.com/api/v1/respond/assistant)://www.openai.com/api/v1/respond/assistant) {respondents/respond
o she dead
Looks like standing against a wall of grass is as good as it gets
Never gets old
can it do laying down men?
Got one!!!!
Seed - 4153053972
neat but its really uncanny lmao
It was a laughing prompt
"a grass lying down in a field of women"
vader what happend
Try the seed and see if it works for anyone else (4153053972)
Guys what's happening here
ill wait either my sagemaker studio lab timer to reset
im not sure what i see here?
It's a contest to get the first visible image of a woman lying in grass
you dont recognize darth vader?
If I change the prompt and keep seed the same...
what's it supposed to be?
Oh okay
oh nah
Can I run SD3 without ComfyUI?
Viable*
Look more like Darth melty have sex with himself π
failed diffusions
lol
thats SO RIGHT
Where can I run SD3 outside of Comfy
Can you share y our prompt?
flop
oh my god that's pretty crazy
Can you share how you got here?
I can run SD3 1024x1024 without TiledKSampler
here you go
I shared the workflow earlier, it is prompt scheduling where the first prompt is a woman leaned against a slimey green wall and the second prompt is a woman laying in grass
you can try StableSwarm
ez game first woman in all of mankind to be depicted laying on grass with no assistance from any other models or photoshop
a field.
That's one the worst I got so far. Not NSFW I guess, but spoiler for the sake of it.
you can run it with diffusers
yeah, gore
the human centipede
What is that
look like my uncle
Body horror
Definitely body horror yeah
Oh okay
skinwalker
It is simply incapable of rendering a human horizontally
apparently it can do elon musk? how about anyone make elon musk lying on grass?
lmaooo, if this is the way we have to prompt SD3 to get anything good then that's a major L.
I was thinking maybe there was some natural language trickery going on that people weren't figuring out but this is just a bandaid that won't solve most use-cases.
what's another word for laying i'm having a hard time thinking in ai
sleeping on grass
Someone else posted a prompt that did it without trickery, I can't say if it was cherrypicked or not though. But yeah my workflow is kind of just for laughs, the model should be capable of this OOB
In a horizontal position?
placed horizontally ..
always if legs and arms are in play SD3 dont know what to do
lounging, reclining, crawling, sleeping, relaxing
here you go
would you mind linking it? it's not coming up in serach for me
that did not look like elon musk looll
I had a couple of good ones, but they were rare.
sleeping fails for me
Here it is
because they removed also promis (not only NSFW stuff)
omg it's slimey wall i knew it
promis?
that look really good
how about reclining, lounging, couching
you forgot the final step, flip the image 90 degrees
prone
known public persons
Yep, the celebs I've tried, that SDXL models render decent versions of are not recognisable in SD3 versions with the same prompts
more Elon A realistic photograph capturing Elon Musk in a peaceful moment of rest on a grassy area, perhaps in a secluded park. He is lying on a blanket, face slightly tilted towards the sky, showcasing a sense of serenity and contemplation. His expression is one of relaxation, with a gentle smile and closed eyes. The background features lush green grass, a clear blue sky, and scattered flora, emphasizing the tranquility of the setting. The image conveys respect for Musk's need for quiet reflection, without depicting him in a state of sleep.
think of words that some automated wrongthink removal script wouldn't touch maybe 
Might help to sell the illusion yeah
I don't think it's a wording issue. I think it's the case of the model not knowing what a human looks like in that position
oh i thought i can do trump
a fish laying on grass yup they trained more fish on grass pictures than a woman laying on her back
1st try too
SD3 is sexist confirmed
ofc it's not a wording issue but it's funny to try anyway
peroprl: stabiklgtiy can you pls no censorshitp
stabilityai: fish
he works
Ok. How does it handle traditionally horizontal creatures standing up?
better censor that,i can see the fish skin
that's it's fish clothing
but the fish is still naked
"a dog standing up on its hind legs" maybe?
im sad that sd3 cannot do pufferfishes, i wanna generate pufferfish chumping carrots
we all are naked under our clothes are we not?
woman sleeping leaning against a grass wall
nah im born with integrated cloths
trump at the park
noclip
If it can do a dog standing but not a human lying down, I'm done
he is a cheater π
hello ladies
wait til you find out it can do a puffer fish laying on grass
weird, what i generated did not look like either puffer and blowfish
woman with her eyes closed leaning against a grass wall with her arms behind her head. The wall is made up of tall grass
But can it do a pufferfish with boobs lying in grass
try women lying on grass underwater, that seems to be working alright for the fish
a dog standing on two legs on grass
the arms are still touching the grass tho
SDXL
but he also standing though so it's right and wrong
it works
goddamit the SDXL fuckin wins
beautiful...
Overlapping prompts are amazing π
Try flipping the aspect ratio
Cake-ism
share prompts? π
good lookin pup though
Cakes!!!
Almost, those arms are a little skinny
@ruby marlin sir you are a genious! thank you! I'm going to mess around with this more
Friend me and I'll DM u the prompt
That's it... It can do a horizontal animal but vertical, but not a vertical human but horizontal
its so over
It does dogs standing better than it does humans laying
its just too hard to imagine a human lying down
- Can do dog standing
- Can do minion lying
- Sexy Shrek
maybe I'm in the minority on this, but with each release of stable diffusion i feel the community expectations get higher and higher. The base model doesn't match the community
I thought even sd 2.0 was better than 1.5 in many ways, though it lacked fine tuning
maybe we are jumping to conclusions, perhaps there are more pictures of dogs standing on their two back legs than of women laying on their back
just community with lacks of technical oversight
with all the criticism sdxl got on release, i thought it would be another 2.0 model and never get adopted, but nowadays sdxl finetunes seem almost more popular than 1.5 finetunes
nvm it can do dogs horizontal too, SD3 is sexist
and just straightaway with "because SAI remove too many nudes from their dataset"
it looks like ther are 6 legs
Still better than human lying down
just a thick butt
ass
I'm a few minutes late, but the word "supine" would be laying on one's back if that matters any.
Same prompts
community looks at dalle and how sai is hyping up their model to be the end-all and sets their expectations accordingly, I don't think you can expect anybody to be positive when the model doesn't deliver
my first SD3 image
a horse standing on its two hind legs on grass
Make her lay down
2GB VRAM 1024x1024
as a person who lost an arm at the recent antartica war i can say that i feel proud to be represented in SD3
yeah this one irks me a bit, this whole time they've been marketing the 8b model?
different model, different client, probably different seed and settings too
sd3? whats the prompt?? i cannot get puffers
I hope I'm wrong, but it seems to me that this version of the model is headed for the same fate as 2.1. Regular tuning won't be enough here, and I doubt many people will want to spend the time and effort to make something decent out of it
yes, SD3.
a photograph underwater of a pufferfish chomping a carrot. The pufferfish is covered with spikes. underwater lights, coral, and algaes on the seafloor. The photograph is taken for National geographic, and shows the pufferfish in all its glory. The pufferfish is eating a carrot
Any chance you could post the comfyui workflow json? (My DM's are also open if file uploads are prohibited here). Cheers!
puffer fish eating a carrot
damn really my prompting is bad
Was it near the site of the volcanic erection?
they've been hyping up both prior to release, claiming that the new 2b is better than the old 8b and such...
i think we probably need "staged prompt conditioning"
yes the big erocktion
Hi, I made this workflow to use both SD3 and SDXL https://civitai.com/models/511790/sd3-sdxl-sdxl-sd3
They can salvage this but they may need to work on licensing a bit.
without the smart prompting it looks like this
this is why you fail,you need more words and to respect the model,like this A photorealistic image of a properly clothed human being who identifies as a woman, depicted in a respectful and consensual manner, sleeping peacefully on a grassy lawn. She is wearing a comfortable, casual outfit, such as a loose-fitting t-shirt and shorts, suitable for a relaxing setting. The woman's expression is one of serene tranquility, and her body is positioned in a relaxed posture with a gentle, restful sleep. The grass is softly depicted to emphasize the comfort of the setting. The entire composition exudes a sense of privacy and respect for the individual's personal space and consent.
this is SD3 there is no smart prompting
I mean 20$ for 1 million annual revenue business and 6000 image per month is like, fine.
thats not what i mean ofc there is no function like ideogram do!!
I believe what they meant is 6000 image TO be commercialized, not generated
THe issue appears to be that they are makign it difficult to get enterprise licenses, which is what fine tuners would need as per my understanding
it's obviously generated.
I tried 
i mean, 1.5 fine tunes can handle women laying on the grass fine, but base 1.5 cannot
yeah.. nsfw in training = not get accepted to get the license
Yeah but we won't be getting tunes
SD3 images are not bad its actually a feature and also a reference to the old 1.5 model
Yeah price can even be $100 or more itβs just difficult it seems to obtain. Price is non issue
do you mean deriviation work, because hobbyist finetune is different from enterprise license
I wonder when we'll see some training data for SD3
If you mean people don't wanna finetune for free, I don't know. They did back in the 1.5 days so why not now? Like I can see it being too expensive maybe?
You may not use the Software Products or Derivative Works to enable third parties to use the Software Products or Derivative Works as part of your hosted service or via your APIs, whether you are adding substantial additional functionality thereto or not. Merely distributing the Software Products or Derivative Works for download online without offering any related service (ex. by distributing the Models on HuggingFace) is not a violation of this subsection.
1st LORA: on Grass v1.2_backed_vae-fp16.safetensors
you need somebody with big compute to fix the model first, then maybe people with smol compute will train some loras or something
*consensual lora training
definitely not on stolen instagram photos
duck smoking a cigarette on grasshttps://up.reselltek.com/u/bIR6Bg.png
goes hard
A field, a woman prone ...
If you distribute or make the Software Products, or any Derivative Works thereof, available to a third party, the Software Products, Derivative Works, or any portion thereof, respectively, will remain subject to this Agreement and you must (i) provide a copy of this Agreement to such third party, and (ii) retain the following attribution notice within a "Notice" text file distributed as a part of such copies: ...
seems kind of foolish for folks with a lot of compute to invest in fixing it given how they are positioning themselves with the license. I feel like some sort of rug pull is getting set up for the future
Almost everyone has given up on SD3 by the looks of it.
speaking of hard
nah, probably not
They can change the terms of the license each month and you wonβt be able to do anything about it
female sniper prone on grass
They should release the training code then.
Correct. It did not live up to their marketing hype at all
dalle in the mud
that's only happened for small business to Civitai-identical site
Are you here when SDXL first release
Lykon knows this. Barely anyone is using SD3 as we speak. It's over for non-existent SAI.
I think she knows we are trying to fine tune her
I'm not sure what you're getting at, if you finetune sd3, the finetune shall have the same license as sd3
Yeah, I used the bot in fact.
its better not to fight back,instead embrace the prey
that's some "back room" kinda video
Diffusers library is already merged with SD3 training code it seems
well yeah.
However, anything that does not involve humans gets you pretty decent results.
I'm setting up SD3 i2i PiXart-Sigma+SD15+SDXL+PAG+LoRA+FaceDetailer π
What about "a woman pretending to lay down on grass"
i bet it makes genmoji's really good
almost like they designed and trained it to be safe for companies to use in their products
So I'm fine with SD3, after all. I like the quality I get from my outputs.
SD3 respects me it laughs with me not at me
Here I uploaded it to civitai: https://civitai.com/models/512696?modelVersionId=569780
Not sure that "pretending" has a 'discrete horizon' - it is too vast and imponderable for SD π
really loving SD3 for now. For a base model, it's miles better than what I used to get with previous base models. I'll have fun finetuning it ^^
The trick is to make the human float π
Fair enough lol
Cheers!
I'm not sure if you're talking to me or not, but what I'm getting at is that people fine tuned for free back in the days. It's sort of evolved into a business now, but sadly that's more on the community and not stability ai. I think people can still toy around with the tech as a hobby
genmoji of a woman laying in the grass
rip
it doesn't behave like real base model - it's more like it was poisoned or destroyed on purpose:
Gen AI safeguards... my friend.
A photo of a woman that has fallen in a field of grass
she is floatin
those hands...
sai said themselves that the model's undertrained, coupled with nuking every nsfw image and artist names, that probably does the job of severely crippling it
You might need to remember that some people who bashed SD3, are those who sell their AI art onto the web, which is mostly NSFW. I mean sure it is a big deal for finetuners who as well wanted to declare "real free AI" on SD
its a reference to old 1.5 models
i think i may have cracked it you have to take out any possibly censored words so the model doesn't think you're trying to have sex with the pixels
a ragdoll photo of a woman with broken limbs is enjoying her time after having being hit by a truck full of SD3
nice prompt
"a woman getting isekai'd by a truck on grass"
"woman grass horizontal"
why does her hair look like a cat, though?
it works
because she's purrfect?
Ragdoll is a common breed of cat
horizontal is prohibited when using the term woman
Quack-Smoak
woman grass non-vertical
Did anyone read the model card?
Out-of-Scope Uses
The model was not trained to be factual or true representations of people or events. As such, using the model to generate such content is out-of-scope of the abilities of this model.
Horizontal simile = Prone
Woman grass in a position rotated 90 degrees from standing
Ah gotchya, we're meant to be making body horror and LSD-dreamscapes.
The model is perfect then!
Woman grass in a position rotated 90 degrees from standing
consensual horror like consensual slashin
Well shit
it's just ignoring all the bad words and keeping safe words like standing
consensual body snatchin
Try putting standing in the negative?
Unstanding?
non non-consensual body pixel generating
here u go both horizontal and vertical
Parallel with grass
maybe try emojis, these should be pretty corpo friendly, π© π π€ΈββοΈ ποΈ
We need to be more inclusive. TRANSWOMAN laying in grass
Prone, flat out on the deck ...
omg it worked
... laying in UniSex grass
unstanding with negative standing
But that's sitting
Be Unstanding for our prompt success!
Close. Add sitting in neg
Put Standing positive, Unstanding negative
i mean sometimes you gotta crack a few eggs to make an omlet but we figured it out
Grass Poncho
She in the grass alright
these ar elike Octoberfest images
speaking of grass
not horific at all xD
From above, a woman with grass behind her
A hawaiian woman in a grass skirt, smoking grass, while on a grassy knoll
has anybody tried drawing a stick figure and using that as i2i for the pose? 
Looks like her parachute didn't open!!!
just go lay in some grass and it will feel better guys ^^ ok the model seems to have some safeguard against woman laying in the grass, that seem quite tested now x) I don't get the obsession about it
What about sun tanning?
I think the obsession is to wrangle SD3 to the ground - to get it to do exactly what's asked!
You don't like a challenge?
too much skin, definitely not kosher
... yeah I get it now. last I did was "a horse riding on a cat". it was fun
i gotta goto bed, i hope you guys figure it out I really want to use SD3 for all my grass pictures
Someone try woman grass bed
Clearly shows anyone that sits on the grass is mangled π
Maybe it's because "ass" in is grass
ΠΊΠ°ΠΊ ΡΠΎΠ·Π΄Π°ΡΡ
Woman + bed is too sexy
They can try sod or lawn
"womxn in meadow"
Woman + Flower Bed/River Bed
It's not just grass, it's any prone/laying/sitting position on any surface as I demonstrated here: #πο½sd3 message
"she/her folx in floral surrounding"
no its not SD3 fault,its humans fault and their weird 3D anatomy
...and not just women
I am using sd3 with diffusers on kaggle but when i try to use sd3pipeline i have exception saying that the sd3 repo is private an i need to logging
hmmm well it's something I guess
But i am logged with notebook_log() function
I have been using SD since the SD 1.4 alpha and SD3 has the same anatomy knowledge as 1.4 alpha
At least the model is safe and Christian right
I think you need to register before using SD3
well, yeah, the use case they seem to have safeguarded is "sexy position" mostly, man or woman laying down on something, so that the base model could not be used for pornography. I mean, it's been the same for most base models. 2.1 was even putting the body in gray if it was not clothed enough. I'm not really surprised here, and I'm sure it will be trained back in by anyone/everyone in less than a week
some of the abominations it produces don't feel very christian
What about "a woman on a surgery operating table"
who cares about frollicing in grass anywayz
I registered with login() function
"a woman lying cross in front of jesus"
try ponies on grass. XD
More Christian than, God forbid, seeing a bo*b or maybe even a n*pple π
alternatively, "inquisition burning a witch in a field of grass"
christians dont use SD3 so well be ok
Woman placed gently on grass sleeping?
FFS how do I use an asterisk in discord
SD3 is the devil
escape it \*
\*
look at all these possessed women on fields of grass
Ah, thank you
I have been big fan of SAI since its beginning, but with how things are playing out i have been mad with SAI recently
A hawaiian woman in a grass skirt, smoking grass, while on a grassy knoll
Cannot access gated repo for url https://huggingface.co/api/models/stabilityai/stable-diffusion-3-medium-diffusers.
Access to model stabilityai/stable-diffusion-3-medium-diffusers is restricted. You must be authenticated to access it..
Will try to load from local cache.
code : ```python
import torch
from diffusers import StableDiffusion3Pipeline
from huggingface_hub import notebook_login
notebook_login()
pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16)
pipe.enable_model_cpu_offload()
you need to agree to the terms on hf first and then log in with hf-cli and a token
Ngl you wishing for unsafe and unethical uses of the model seems very problematic
#π€ο½tech-support would be more appropriate
Maybe fingers are worse because the censoring mistakes them for...uhm... the male genitals? Kek
Use a hf (Write) token as a login password
I like the universe diagram there
A man in moon
ha sorry
Wouldn't it need to be trained on dicks for that?
Looks like he could "re-fuel" in mid-air!!! π₯³
No, as far as I know there are techniques to remove elements. Basically inverse training. And maybe that cut off some fingers.
too many people with skill issues lol
It's so funny how this cutting edge tech is held back by outdated standards disguised as something progressive
hey i just woke up. Is SD3 out? i cant wait to try it out
Don't bother
Local is shit
sd3 was collectively cancelled
rlly? why did they cancel it
I mean you would need to for wait to get better so it can actually produce good images
Go back to sleep, just don't lay down when you do it or you may become [REDACTED].
The API SD3 is so much better, it's not even funny.
Like what happened to being sex- and body-positive
the outputs aren't christian enough sadly, god bless
Fascinating. I have an interesting update on human anatomy:
Check otu the two "art" images above. Simply describing a scene and using the word art gave me these with descent anatomy.
However by using the exact same promt and changing it to photograph, or photograph of, and this happens
Fascinating indeed
yep each model can be "unlocked" and molded into whatever π
so it can do proper anatomy
Btw how much vram would I need to run sd3 medium approximately?
Hopefully
I tried generating an angel but I must have sinned
try doing black or white ppl
just not for photographs
Holy shit I cracked the code guys. You can't lay on grass, you have to become grass and lay. π€―
Full diaper
"jesus but black and chinese"
Why are the photographs asian? Did you change the prompt other than "art" and "photograpy"
Now give it a grass dong
extremelty large DDD boobs but only in clothes
do the chinese elon musk
@tranquil sand can you share the prompt? You've just given me an idea. DM me if you need to π
nope, I only changed art to photograph of. THe other promt words were pin up, girly pic
The prompt was actually a failure, I was trying to meme and inverse the "human laying on grass" with a grass mannequin laying on skin, so forgive the cringe prompt:
Pos: A mannequin made of grass is laying on a field made of people and human skin. The field is made of sweaty human skin with a tan.
Neg: bad quality, poor quality, doll, disfigured, jpg, toy, bad anatomy, missing limbs, missing fingers, 3d, cgi, grass field, grass background, dirt,
6gbs of vram minimum
i guess when they meant its undertrained they werent kidding, they really will just let the community do the heavy lifting this time
Oh cool I can run this mess then
yeah 6GB vram very nice
That's only if the community adopts it
cinematic shot from above of a beautiful girl who has become one with the grass, she is grass and grass is her, the grass calls out to her as she sinks into it
I hope it will and we wont pull another Cascade and not bother with medium coz anywayz 8B is coming so whateverrrrr.
Or lobotomized it
The fact that it works π
@lavish osprey SD3 body anatomy for sdxl lora is released - https://civitai.com/models/512653
The twisted mangulations of SD3!!
i will make some side by side sdxl sd3 comparisons - same seed same promt same everything and then ill know the truth
humans if they were good
yeah, the level of bigotry is just ridiculous. GA censored models shouldn't be called base models
Damn
i'd rather they focus on alternative models this time tbh
Can we revive the #1207078178510872636 channel now?
@agile glen thanks! here are the results from my experiment! I tried to use prompt scheduling to get a better result! The results are....enteretaining!
that bad huh, nah come on there is potential here..
I changed the promt from manequin to man after about 0.1/0.2 to get the above results
i made some people standing and it looked ok earlier
lmao not bad
tbf as soon as i try different angles and positions other than plain standing and looking at camera i get a hot mess in all models
what's next 2 hours thing?
U think they worked overnight to get stuff done and will replace the version by a better one for humans?
within the next 2 hours ponydiffusion pixart will release and save us all
Because standing is the only pose left. And CogVLM is great at describing poses.
this chat shutting up about shit the model is doing xD
someone should probably start KS campaign, and then train unbiased model with just physically accurate and properly described pictures - based on SD3 architecture, but without all side effects caused by "safety" aka bigotry teams lobotomizing the models. like seriously - every one of us sees the breasts since we're born, why the hell censor that?
so close.... I added that he should wear a tropical shirt π¦
Not from what they've been saying in Discord. They've said that they won't be making an update because we have a skill issue.
pony ain't coming. THey said they can't finetune SD3 because of the license, and apparently they are not getting a response when inquiring about enterprise π¦
This is what happens when companies are taken over by religious zealots....
bro thinks he walter white π
Leonardo and Mona Lisa on a date
thanks it works
yeah i meant ponydiffusion pixart, pixart is a different model from sd3
oh right! is that happening?
"safety" stuff on SD3 are really frustrating... 
Pretty sure Lykon was on fire yesterday, the way he wrote showed that clearly. I dont think it was his idea as he pushed expectations. Now got back in the office/talks and there MUST be a huge discussion on how that company will survive
you guys can have it, as good as your sense of humor 
pony guy said he's considering it... but nothing is confirmed. fingers crossed!
psyduck
pony will have version 6.9 before the SD guys settle with a good license thing
no joke is funnier than the current state of sd3
if you recall unstable diffusion, this is how you get a bunch of people pocketing money and delivering nothing
Is there a chance for us to escape this safety censorship issue in the finetuned models to be done by the community?
can't seem to get it to generate these photorealistically at all. Probably a skill issue, but fuck me. I'm stumped. Looks like bad drawing. Nice coherent shapes though I guess
Guys, what do i need to put here?
Apparently that's what they were drinking while they were torturing the dataset.
what is your prompt and negative
SAI is dead (well, it is already) after this disastrous release.
Introduced my dad to my fiancΓ©e
just connect this to your prompts
just use checkpoint with clips into "Load Checkpoint", and use CLIP output then from that node
skill issue yall never seen a real naked woman XD XD thats the reason these generations of women so bad
Also looks like Lykon/someone else with mod permissions deleted all of the instances of Lykon telling people they have skill issues or just generally trolling/shitting on the community. His posts all look kosher now when you search for them. (But there are a lot of replies to 'deleted message' of people not so pleased with what he was saying).
works now i think
No. It is worse than SDXL. SAI... is... dead.
girls arent real
I've tried a lot of different ones... that was
positive: cinematic shot of a futuristic assault rifle, side on orthographic view, plain grey background, beautiful space rifle made from dark gun metal and burnished steel, strong, intricate design, highly detailed metal, the gun is covered in natural wear and tear finger prints and scratches
negative: blurry, bad quality, poor quality, jpg, toy, missing limbs, missing fingers, 3d, cgi, video game, cropped
skill issue from him
Don't bother man it's horrible
Guy
no way Really?
if you have subpar quality of image from SD3, tried changing sampling to default ( DPM++ Normal ) or Uni_PC
dont use any combination of Euler or SDE
If you don't intend to output humans with your prompts, you're good. I like doing wallpapers with SD3 Medium.
Grass laying on person.
(Idk if this is nsfw or not, it's honestly just terrifying).
unipc is a hidden gem i used thta and i got nice stuff in fact when i refined it with sdxel sdxl mad eit worse
What I want to generate isn't even porn. The impact of this 'safety' issue is very apparent in the generations of models currently being generated; we're experiencing an SD2.0 case right now
needs more bellybuttons
SAI is being so ignorant. They forget the need funding / money bad. I'm currently at a AI company that is providing AI imaging solutions to well known brands around the world. We where hyped for SD3 but seeing how it's released and the complete incompetentce of some of their devs on twitter we are now talking to other company's instead and don't wanna go down the SD3 rabbithole.
people still thinking Lykon = SAI as a whole
Can try this 
professional photo of a futuristic assault rifle in 4k, side view on a white background. Futuristic rifle made dark metal and burnished steel, with a strong and intricate design. Detailed on the metal of the gun, the gun is worn with fingerprints and scratches.
Don't tell me it's SD2 all over again π WHY DIDN"T THEY LISTEN THAT NOT HAVING NAKED PEOPLE WILL JUST MAKE BAD ANATOMY
thank you. will try
Ret... Western investors.
"m-muh safety!"
I love stability ai
stupid humans and their weird poses lying on the grass,what a bunch of protein based beings
In Stable Diffusion 2.0, even in non-pornographic generations, it behaved quite differently, and it was very likely to see anomalies. It was truly ironic that its previous version, SD1.5, produced more successful results than it did
SD1.5 will forever be their greatest invention.
I hope i'm wrong but....
I'm guessing the intention here is to drive people to Ultra by giving us a taste of SD3 but leave us wanting more, and seeing what Ultra can do.
Unfortunately I think they are too late in the game to do stuff like this. Competition is fierce with MJ, Dall-e, and now pixart, adn SD3 needs a major USP when the competition does things just as well or better.
SAI's USP was the open souce component, by h aving the model open sourced the community developed amazing tools like Instant-ID, Controlnet, Ipadapter and so on. SAI's issue was bad business practices in monetization.
In my opinion this can be easily solved with a well thought out licensing program. Give people the freedom non-commercially, but treat commercial applicaitons like partners.
A simple approach would be to basially charge an SAI tax on sold compute on models and fine-tunes. You could go further by working with "resellers" such as replicate and fireworks to become authorized partners, so that SAI gets a cut of any commercial application
Furthermore, develop a program for consultants and companies giving on-premise solutions to get certified, and reccomend them via the website for a fee
There are many successful for profit open source companies, my favorite of which being Odoo. They are an amazing ERP system and a company I worked for used it for free self-deployed for many years before eventually using a reseller because it was easier. There's much to be learned here. Huggingface is another example as a for profit company that's a stalwart of open source (They should totally buy SAI btw, the ceo tweeted it a few months ago)
Finally on the topic of safety there is enough technology around to enable SAI to setup a taskforce to track and trace bad actors misusing the technology
Honestly this approach is not going to get investors what they want. Leaning into SAI's strenght as an open source company and following the business model of other successful Open source companies
ain't reading all that
is the path to success
tl;dr: sd2.1 2.0
speaking of grass
Hey I actually can understand him. Putting immense effort, heart and time in these free to use models and then everybody is not pleased. Probably not by his decisions and you just see how the work of your life gets destroyed in minutes, after everybody was so eager and curious. I just hope SAI will take the right decisions now. It is obvious, that it is NO skill issue (Although a lot of users definitely have problems to use the new style of prompting) but that humans are just out of scope
at least the background anatomy is almost right.... xD
I got 4 limbs intact
Witchcraft if you ask some people
would you mind sharing? π
Actually, even in its current state, SD3 is still very, very successful. If StabilityAI listens to the community's pleas instead of stubbornly going their own way, all text2img models on the market could pale in comparison to SD3 like a kitten next to a lion
oh nice!
honestly the model isn't bad, has its purposes, clearly if someone wants to do anything that involve other stuff therere are other models...

I mean just tell me a model that can do this π€£
clearly 60% of the people who complained it is porn guy
I agree, I think the model is not THAT bad overall. It's just the safety features and licensing that have taken it from this could be a great base to.....is it really worth finetuning...
it's obvious that is related to porn, model can do pretty much anything else decently, let's wait some finetunes
or people who just like portraits, not always fully clothed?
Def agree on that. They tried to hit the ball and missed by a landslide with this approach. Our company is one example that will move on to a different supplier possibly losing them millions in the long run just from one company and there are many many that follow and will move somewhere else after this disaster
Now make her lay/sit down
I can't fault anyone for being upset about the community reaction (although it's only a reaction to what they received compared to the hype), still I don't think telling everyone they just suck when clearly it's the model that's at fault was the right move lol.
Deleting all the messages also doesn't solve anything other than silently admitting defeat, frankly he should apologize to all the people he shit on since the messages wouldn't have been deleted if he was right about calling them out.
sex makes the world go round
or people who likes to sleep on grass
SDXL couldnt do porn either and pony was made from that. people are so clueless how this stuff works
or people who like 10 fingers in two hands?
or people who can sit on the grass
i mean people truly have some skill issue with how to run SD3 and change its setting
There probably should have been a lot more people in swim ware in the training data and this issue I feel would have been resolved β¦ but thatβs speculation
SD3 + Luma
how is sdxl base compared to sd3
that too, but of course you still cant do sex and nipples etc.
or people who can lie on bed
the nipples is just smoothed out
Why do you even need SD with this much talent
But personally, I wanted portraits, face only. Same exact prompt as the api, I expected something similar. What I got was absolutely nothing like it.
and classic Civitai user just upload their workflow to adding back nipples
one thing when some random guys saying this in chat
other thing when SAI guy saying it instead of helping to community they depends on lol
better anatomy for simple poses, better skindetail, better realism. all out it's looking good for future nsfw finetunes
thank you kindly, I've been drawing stick figures for 3 decades
how long does finetuning takes. like prolly a month or two right
not me, that workflow creator
@vital crag Did you try the prompt, did it work and can I see the results xD im curious
2 color splash assignments
melancholic monochrome close-up portrait photo of a melancholic cheeky goth woman with goth makeup and short punk blue color splash hair wearing a black fedora, (slightly smiling:1.4) red color splash lips, designed by rockabilly vogue
Already??
I expect that we will start seeing great finetuned models within a couple of weeks
Not monochrome
awesome 
yeah but the best sdxl models too half a year to come out. but you should see stuff slowly evolving in coming weeks. especially since they say it's easier to train than previous models
That's a big if considering the licensing model
who's "they" here? the pony people?
still 
it absolutely isnt an if π
1.5 - forward, 2.1 - back, SDXL - forward, SD3 2b - back...
That wonβt stop as many people as you think if we are being 100% honest
So SD3 8b = forward
some would argue that we haven't really seen forward since 1.5...
first off you will have simple lora's that improve NSFW prompts
Maybe, finetuning costs money and computing power.
I don't think people sharing their finetuned models trained on CivitAI will care much about licensing issues. Hopefully, StabilityAI won't impose unnecessary restrictions on this matter and cause trouble for everyone
lmao only recently I started liking SDXL quality, always looked too plastic on anime and boring on realistic, but prompt understanding was great 
It took a while to be "forward"
SAI legal team is large but is it large enough?π€
Yeah, but we're not gonna see it.
hopefully humans learn how to sleep properly on grass
imo cascade was a bigger step forward than xl but everybody seems to have forgotten about it
Yup, SAI is pretty much dead after this disastrous release. Not many people are using SD3 Medium as we speak.
Cascade sux
SD3 need a fine-tune version
True though. I think the best solution would be to say something about finetunes as it is obvious what many people want and in the end the other pcitures here look great
it doesn't suck as hard as sd3, true facts
is it rlly that shit compared to base sdxl
for what? π
Cascade is just a research model though that never intended to make into popularity
agreed
I have been harping about it for months... So seriously at this point you think we dhould train cascade or SD3?? :)))
SD3
I think we'll see little to no community finetunes, I'm afraid. I'm sorry, but keep your expectations low.
LOL π€£
lora won't help much if the base model is crap... it will require a full finetuning which will be very resource intensive
I mean you didn't remember those less popular base model did get finetune like SSD 1B
lol
did you try training a lora?
most popular creators on civitai say they won't train models on sd3, so... not looking good, tbh
I trained a bunch of loras for cascade and I like them but it seems to have some issues with quality, weird noise and such, sd3 outputs don't have that but that anatomy is a bit ehhh
They were pretty much long time ago, but this time it feels
why wouldn't there any SD3 finetunes. Let's just wait how well it respond to training
That's just Pony, some people who dont like Pony / hate Pony dont even care and still tried to move their way on SD3
Come on guys pep up! theres hope, this is a solid base :))
to add to that, cascade can at least draw boobs out of the box
Creator of Pony just have long opinions with SD3
SDXl + Luma
Can you show us?
I start hating pony because there is a bunch of people who think it's the center of the universe
no... but i'm quite sure that's how it works. teaching complex concepts like anatomy will require a full finetune
wutss wrong w pony?
who cares if one guy on this civitai porn site doesn't want to train on SD3
You're right, but most people use Stable Diffusion for p*rn (or anything that involves NSFW content).
Whats luma? :0
"Has she got a wedgie?!" π₯³
Exactly. I was in the top 10 base model creators last month on civitai. But I definitely won't be working with SD3, and neither will the rest of my colleagues.
if you want pony anatomy then sure, but idk why you expected that from SD3
As whack as that is, it's the only reason SD is popular
it totally destroyed the base model, is not able to respond properly to prompts, does only understands this weird booruu tags and in total it cannot do anything you couldn't do with base sdxl except for porn - which is probably the reason people use it
luma looks so much better than all the sd video models it's honestly crazy
how is she so consistent tho?
thats the point. This is so fucking stupid and annoying.
i think i have a solution for that, dont download the model
wtf is luma whats going on??? XD
there is a world beyond porn and many SD users that are not interested in porn
it's an image to video model like oai's sora, was just released yesterday, and it's free (for now, but not open source obv)
text-to-video?
SDXL can do decent porn but not like 5 wieners in one photo with coom all over
Maybe a hot take but:
With all the developments that have been made, and incredible progress of free open source projects freely available and ripe for the picking, a 'solid base' just doesn't cut it anymore.
At this stage in the game we should be getting professional grade functionality out of the box.
so please stop crying about SD not able to do porn after every release. It's just tiredsome
indeed theres plenty of models for you out there
pony is one of the best prompt following models + it can do a lot except of porn but fine
Comparison between different text encoders pls ???
very high skill issue, also, pony anatomy is 3000 times better than anything right now, and it can do the soo wanted nsfw stuff that people using sd3 right now want :p
i dont get how it effects you or everyone though.
Pony has the best hands out of any model I've used tbh
Number 1 reaso: Its free. Number 2: Lack of restriction. Number 3: Customizations although only few go further with this one
booru tags are peak efficiency, like it or not
the tags are like mini loras, they work for what they do, can't really warp them like natural language but if theres enough of them people are happy with it
I was iffy on it at first but even just for inpainting the hands are way better
CLIP - old one that used by SD family, could understand both Booru or natural language ( which often called CLIP or BLIP )
T5 - an improved text encoder to improve model's literacy on user's prompt. Should be much better when you tried to prompt objects in specific position
yes, I use other models. I'm just annoyed by all these pony fan boys crying that there will be no pony finetune on SD3 and acting like that would mean "there are no other finetunes anymore!!!!!1111 because pony is the only model on earth". Like just go in the pony discord
there are still other finetuners right
Whats so special about Pony
they have 30 gen / month for free, but for now you'll need to wait few hours for the queue
Porn has nothing to do with it. SD3 doesn't handle anything at all, and especially not anatomy. I do SFW finetune SDXL models. But if you don't teach a model anatomy, it's simply incapable of creating characters.
I think it's not just anatomy. SD3 has severe issues with composition at all. Doesn't matter what you prompt, you often get weird composition. We will see if this issue can be solved by either finetuning or different parameters
So which one should people typically use ? Clip only or t5 only or t5 + clip ?
The reason CLIP is still there because the hardware consumption of T5 is larger
pony cant do realism whatsoever either, even the most realistic models have a very clear anime influence.
that they are like mini horses, so its cuter ? wait not
true,this channel was better when we were making girls on grass images
wuts with the grass girls meme
now speaking of SD3 hands For the close-up of male hands, we'd focus on the intricate details that define a character's personality or background. The image would feature a male high school student's hands, potentially in the middle of a task or gesture that reveals something about him. For instance, he could be holding a musical instrument like a guitar, showcasing his interest in music. His hands might be slightly calloused, indicating hours of practice, and the light catches the texture of his skin and the strings of the instrument, adding depth and realism.
Anyone wants potatoes?
Finetuners are pretty rare, 99% of the checkpoints are just merges
I dont use pony but tbf there's few finetuners that use quite as much resources as them, and people like them being pushed away by licenses and everything else kind of points towards some of the others ending up pushed away for the same reason
Pony cant never do this π
who said you cant teach the model anatomy at all
better than Dalle
Im sympathetic to people being crazy but lykon and co have been doing so much unnecessery damage on top of the more understandable issues they are facing
I belive you can do refiner into another model to make the realistic pony models out there look very good, also idk much about them I used one or two only a few times to test, so not sure if they improved since them and if I even used the best one
Dalle is the standard amirite πΌ
wtf 
The comments by him were unnecessary
It's going to be very hard to do that now that the model developers didn't take care of it.... We'll have to create huge datasets
sd3 works good ontop of it btw lol
using it as a refiner for other models, good textures, faces etc
Make SD3 great again.
shes learning how to walk!
Stability AI lost the most important customers to Adobe and partially Midjourney, thats it
Dude being insulting to Pony creator when he was just asking about licensing issue 
If the creator of pony can get a arangement with Stability than we good
yeah you can, but it's a cheap workaround.
yeah, that's a bit sad. SD3 is good at finedetails, but hope was it will get better at composition
Yeah lol, sounded like bitter comments because ratioed by pony guy
Also it's not just the pony, with the licensing issues, nobody is going to touch it
that's not gonna happen.
nah, anyways. I'm soo fucking tired about pony fan boys X_x Ciao
He did
one can pray, maybe the creator of pony might purchase the enterprise plan
it should as we'll have some good finetunes, I mean sd3 have really good potential as I see
well then make a better trained realistic one ? but that would take time and effort... or take the workaround for now since there aren't any available and use it
I just hope people still feel incentivized to build controlnets and other tools on top of SD3, but given licenses and the general state of things make it less likely quite as many people will bother sunking a lot of work
luckily, the world is more than pony and things will go on
apparently he said "I think they have no knowledge of the importance of my Pony"
bro the anatomy is not worse than SDXL and the model should be easier to train π€£
Just trust me, it's not going to happen.
Ohh really? what's the situation? PonyV7 still happening?
yea like how SD 2 and Cascade moved on right?
not with sd3
Lol. The anatomy in SD3 is much worse.
i honestly dont know who to trust here. so many contradictions
Why are sai shills just making it about pony lol, people are mostly mad about the horrible hands and anatomy
As of right now we're getting Pony v6.9 on SDXL instead of Pony v7 on SD3.
Said v7 may come if he can get the licensing sorted out.
A detailed photorealistic portrait of male hands, captured in high resolution. The hands are shown from an up-close perspective, with intricate details visible. The skin texture appears realistic, with subtle veins and fingernails that have a natural sheen. The hands are in a relaxed pose, with fingers slightly spread apart, emphasizing the delicate features and the muscle definition. The lighting casts soft shadows and highlights that create depth and dimension, showcasing the hand's contours and the intricacies of its surface. we dont need to fix anything,this gucci
I would not believe anyone who say that, but you got some reputation man
The situation is super complicated and yeah... it is hard to know the context now
nope
i hope it's pixart and not sdxl again
pixart has so much less ecosystem surrounding it tho doesnt?
ill just wait for like 2 weeks and come again to see if everyone is cooled down. too many agendas going around
Why do you keep bringing pony? people are criticising SD3 for its own shortcomings π
People dont want to admit that they failed hard with their overhype of SD and SD3 especially. Dont forget what people told months ago like how MJ, OpenAI and Adobe will succumb to SD
like I dont care which is the best model as much as I care which one has the best tooling support
the dude hates my little pony
this is cool, mind sharing?
me, too. I just refer to all these guys here saying "there won't be any finetune, because the pony creator don't want to finetune it!"
Yeah, and the earth must be flat? π
I think they're starting to realize/contemplate admitting it since all the messages accusing people of having a skill issue and shitting on thier attempts to use SD3 were deleted.
absoultely
beautiful hands
literally nobody said that
we had the same discussions with SDXL by the way. A lot of screaming and crying because you couldn't do porn with it
Yeah if SD3 was good nobody would say anything, all the negativity I see on the sub is about the terrible anatomy
forgot the (perfect hands:1.6) 
can MJ or Dalle do this? i dont think so mr nice guy
skill issue. someone already prooved you can make images like this with a 75% successrate yesterday
another victim from the horrors of war
Magnificent!
people were a lot more impressed and less pessimistic with sdxl release. it's not even close to this
good to know that it happens with sdxl
the whole channel is full of that lol
I'm happy to discuss SD 3 shortcomings, I just don't want to discuss about this pony model
He wants to but the licensing and the dev being shit to him doesn't help 
Why brag about a 75% success rate for one of the most simple things you could generate
oh tell me you so skilled one, how can i, a puny user of sd3, get a good image?
u have an excessive hate against pony though? why?
don't worry 8b will fix everything (please ignore the fact most people won't be able to run it, much less finetune it)
This is not goku this is coku
you need to respect the model like this The image captures a close-up view of male hands, emphasizing the texture and details. The hands are clean-shaven, with visible veins, and slightly roughened from daily activities. The hands are positioned in a relaxed manner, perhaps resting on a table or in a pocket, showcasing the skin's subtle variations and the intricate patterns of the fingertips. The lighting is even and soft, avoiding harsh shadows, to highlight the natural appearance of the skin and the hands' features.
75% is pretty bad if you want to create a ton of frames with animation, means it will basically never look good
it has literally never been the most simple things for AI to generate π€£
dude who said, there won't be finetunes beacuse pony creator won't make it? Like who, give me a single example
just scroll upwards
this prompt on dalle
wtf
because at least it can make actual humans unlike SD3
Yeah lol, 3/4 is horrendous for just generating someone laying on grass
Lykon, we know it's you.
everyone is saying that beacuse of the License, I couldn't find a single example of your claim
SDXL/ SD3
@agile glen @foggy cloak
I really dont see right now how they are supposed to get serious investors at this point.
cope
sdxl / SD3... hmmm worrisome
yes, the licence doesn't forbid finetunes. It does not even forbid using them commercially, although there are limitations in the amount of money you can make. It's just the pony guy who is currently pissed of on the licence and people act like everyone would have the same issues
Unless 8B is crazy good, I don't either. It would have to be a better value proposition than MJ and other competitors as well
i hope ur correct. and the licence doesnt limit finetuners
Would SD3 8B even be runnable on a 12vram gpu?/
HOLLY
As strange as that may be, I never used SD for p*rn lol. I got website and IRL for that, I don't need AI to gen some more. Really not sure why people are thinking this is the only use of SD all around.
But for anything creative, transforming pictures, training on my artstyle, brainstorming logos, change the setting on family or my pets' pictures, ... I do use SD a lot.
So a new version, that is capable of doing texts, that understands multiple subjects, positioning in the frame, longer prompts, and is doing 1M pixel with a low VRAM usage ? yes please !
I mean, it's just better than what we got before IMO, and it's free for non commercial use.
Of course it will get immensely better with more things on top, be it LoRA, controlnet, ... or by training specific things in it (if you need something it doesn't know, or was safeguarded against). No animation that want to be good would stick to the base model after some point
Anatomy being wrong isn't the biggest surprise. it's always been a weak spot of any base model, and training it back in has always been detrimental to the rest of the model. It's a balance of what you want in with a given number of parameters, and base models can't be good at everything at once (especially for concept close to safeguards)
Once again though, this is In My Opinion. We are all talking about creative work and art, this has always been subjective. And to be frank, the release week of every previous model has always been exactly like this. People were just saying SD was dead when XL got released too. I'm really not surprised about the reactions today
wut did they say
I CAN'T BELIEVE THIS NO WAY BRO THATS WILLD AF
Dude was not pissed he is a paying customer and wanted more clarity but the dev was just insulting him left and right
The issue is also that even when it worked good they failed to get investors and serious paying customers in. Thats why i also said Adobe and partially Midjourney but maybe also OpenAI nailed the coffin on SAI
can link?
"It's safe bro" but racist π π
Bro is not flying π
Superman in his iconic costume flying trough the city, the view is dynamic with lots of motion blur, and superman's face should be the main focus,
SDXL/ SD3
Thats true, release days and the following days have always been controversial to say the least
anyways. We all can agree on that the current 2B model is extremely shitty in composition. We will see if this is just some unfortunate use of parameters or prompt, or if it can be solved with finetuning. The model as it is is definitely a downgrade to the current SDXL finetunes
dude you just claimed people said stuff like nobody will make finetunes beacuse of pony, I asked who and now you switch topic, and on that matter yes the license doesn't forbid finetunes, but it limits commercial use. It's a dead end, no body is going to pour in resources for big fine tunes.
guiz
its so strange actually
Because it's a realistic model and Superman doesn't exist.π€£
should be fixable with finetuning
You should have compare SD3 base with SDXL base anyway
it was not as bad with sdxl. Some negativity but a lot more optimism than now
actually I did some research on this previously, adn it seems like they did have customers, but were so focused on research and r&d that the business side of the business couldn't handle basic customer service, and so all the customers ditched them.
Emad was great at building the company to create the models, but did not attend to the business side of the business
he wanted bunch of people to agree with him now when it didnt work, guys lets move on
I'm running sd3 rn and it generated a woman without clothes just fine
With normal anatomy
Just with no details whatsoever
ok yeah its not sdxl base but still lol
I did. SD3 is better in almost all cases, but it has definitely more problems with composition than SDXL.
Looks really odd
as long as the girls arent lyin down its all good,we so back
can someone try this prompt on its own guys? It generates black people on me π
a photo of Thinking Monkey on branch, one finger of him on his mouth, he is wearing a white shirt and text on shirt reads"SOLCU GAZETE" and revolution fist icon on shirt. documentary photography, grainy look
I remember people were underwhelmed but not really that negative for sdxl release tbh
at this point Im not even bothering to set it up until people have had a few more days to find the best settings, encoders etc. I reckon it will look better once it's more figured out
WE ARE SO FUCKING BACK WE JUST NOT GONNA MAKE DYNAMIC POSES!!! π₯π₯π₯
They did, but compare that to Adobe for example. SAI as you said did bad business traded for glorification by open source AI community
hey boto. have any good (long) prompt?
I'm currently generating a bunch of megagrids for different prompts, to show the difference in parameters. I remember you always had really damn good prompts. was hoping you got 1 or more for sd3 :3
with sdxl there were also a111 bugs initially that made it look worse at first
just hide the hands/feet and focus only on face
the good ol sd1.5 days
its so funny it generates nipple-less women
ore penisless man
you have to get the nipple license for that
SD3 will be canceled for racism π€£ So focused on banning nudity they forgot about everything else.
they look like barbie dolls
Kot
cacr
omg now way who would have guessed π
needs more corn
π yeaahh
Kit π
is it this one? https://civitai.com/articles/5671/towards-pony-diffusion-v7-i-mean-v69
theres no explicit mention of denied licencing only vague responses. am i missing something
sniff
https://imgur.com/a/rP0KYeY
My results so far if anyones curious
the current situation is vague, astralite will have to make a another article
did you use ultimate sd upscale for this?
your link 404's
lol 404
whatever
Ill come up with smth
ah ok. so thats the most updated one right. im new to everything here
Sorry in advance
wtf
You know funny thing is, I had a prompt where it's a black guy shouting at a white guy and both dalle3 and sd3 (8B api beta) did the reverse lmao
So those are his results - the lack of results π€£
PonyXL is currently the most supported model, so yes
I havent looked too deep into prompting so far, I was trying to figure out how to increase the quality first. But no success so far 
this is base res from api
I've just reported you to stability ai's safety abuse email. This is a gross misuse of SD3
https://ibb.co/qxy6rLW
https://ibb.co/gRHcvk9
Here they are
These arent terrible somehow
do you mind sharing how you got this far?
SD3 better To create an appropriate and respectful representation, one should focus on the natural elements and the human aspect. The image would depict a well-dressed woman lying on soft, green grass with a serene expression. She's wearing comfortable, outdoor-friendly clothing like a casual t-shirt and shorts, and she's positioned next to a large, artistic rendering of a giant corncob. The cornstalk could be depicted as a whimsical, non-intrusive element to the scene. The background would show a peaceful, sunlit park, emphasizing the tranquility of the moment. The depiction respects the individual's privacy and personal choices, avoiding any inappropriate context.
mannequin!
you could probably fix this with an inpaint and a segs
you guys will dream of these image 
Stock sd3 medium from huggingface
Their clip models from the same place
Stock workflow from there
And just added 'naked' to the prompt
wuts with that gpt language prompting. is it mndatory for sd 3
I mean, could you share the prompt?
LLM promptin
if you're lazy its good!
DALL-E 
Doubt finetunes will be out in a month
Give me this prompt 
oh but adding 'lying' to the prompt makes lovecraftian monsters
Can someone give some insights on text comprehension? Are we getting closer to Dall-E? The announcements says
": πΌοΈ Prompt Understanding: Comprehends long and complex prompts involving spatial reasoning, compositional elements, actions, and styles. By utilizing all three text encoders or a combination, users can trade off performance for efficiency."
yes, closer, but absolutely not DALLE3 level
i entered this whole paragraph as a prompt on SD3 and this is the img it gave me
Might be overly optimistic
wait
@desert garnet you'll be fired now that sd3 is cancelled by the people in this chat 
no more sd3 ceo 
SDXL...
@bitter hearth
photograph of a cat holding some computer ram between its paws, and a sign that spells "send help". The cat is very cute and seems sad
wow amazing
i give you a few reddit golds and upvotes
Euler, 30 steps, cfg 5, upscaled by 2x with 15 steps
its a bit long so:
https://pastebin.com/wBcV49dK
Pastebin.com is the number one paste tool since 2002. Pastebin is a website where you can store text online for a set period of time.
Still better than SD3 
wtf, are you writing a book 
wuts the best way to use comfyui non locally
not if you get a based standing pose
getting prompts from LLaVA or LLaMA adds a lot of details that im looking for
so i cant put in 8000 negative tpoeksn as it complains about memory running out so that needs optimising but i think lots of negatives improves it
btw read the prompt while looking at the image. you'll understand the prompt coherence power
who wants to make the first "woman lying in grass" lora xD
π
sd3 is so good to generate envirronements
ill give you the first trainin img To create a respectful and consensual portrayal of a human being who identifies as a woman sleeping on the grass, it's crucial to avoid any implications of non-consent or disrespect. The image should reflect a serene and peaceful moment. The woman is dressed modestly, with a comfortable outdoor attire, perhaps a loose-fitting top and shorts suitable for a grassy area. She is lying down, gazing at the sky, with the giant corncob silhouetted against the horizon in the background, symbolizing a connection to nature and tranquility. The composition would be simple and uncluttered, emphasizing the woman's relaxed state and the natural beauty of the scene.
same for me and a bunch of finetuners I know. π¦
its really hard to get rid of the artifacts. and as kai has mentioned, composition is even harder to get right.
maybe its back to SD15 days where tons of negatibve promts mattered
i wonder if we could generate nipples by describing them lmao
Dark days
almost good lol
i can see her ankles,no good
woman lying in grass is not something impossible to do. Just need more attempts and prompt skill.
hell yeah
hello β¬ββ¬γ( ΒΊ _ ΒΊγ)
not christian, i agree
thats going straight to civitai xD
yeah, its really nice 
agreed. it great most of the things
But why should it require several attempts? And what skill? Writing a paragraph?
breakthrough of the century, we did it sd3 bros
trying to make an Eevee, using only tchatGPT to build the prompt and SD3 and not interfering with the prompt
notice how it is lying on grass
Because the models trained with vlm generated caption. It doesn't like previous model which use alt text
So are girls lying on grass like a way to test a model or something?
it's a feature
it's just something basic that sd3 struggles at
My idea right now is using PixArt as a base and refine with SD3 
No, but you would think it could do it no problem
Getting serious
nono, the model doesn't struggle, it's a skill isue, ask allmighty Lycon
should work like a charm
we beating dalle3 with this one
Post it on Civitai https://civitai.com/models/497255
I already got something similar with SDXL as a refiner but doesnt work super well right now with SD3
already did
Any A1111 support for SD3 yet?
we all obviously have dunning kruger effect here
this is pretty cool 
No way π
exactly
someone made a lora for sdxl π https://civitai.com/models/512653
sd3 says touch grass to reddit users
not as cool as this img that hamuz made
you're right! my apologies, should have, us,ed mo,re commas,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,

what I like is that sd3 doesn't try to make iamges with drawing lines
nooo I picked up from google
ima steal then
1girl, real photo, 4k, 8k, ultra, masterpiece
you may want to play with shift parameter for sd3, lower it to add more details, make it bigger to gen volumes and forms
so if you want to use it purely as refiner you can try with shift 1.5-2
Absolute lack of understanding of anatomy even in such simple scenes, not to mention complex compositions. This is not what people expected to see in mid-2024, which is why there's so much outrage
sd3 is also good with faces, but to refine them better to use higher shift before adding details
can you explain with visuals
its a realistic depiction of an old man who lost his fingers at korea war
well i get it that you need to shift your way of creating prompts, but those basic prompts should do SO much better
prompt: "nipple"
how dare you
π
game starts
1 more
No Bro, everything Is fine i swear
see thats why everyone loves nipples
wholesome 100 reddit chungus ahh model
I got it setup now, my only issue is the tiling with tiled diffusion and the seams
ermmmm....... that's because it's a base model? as we all know base models are SUPPOSED to have horrible anatomy, that's how it works!
can you share the prompts?
when I used the base model sdxl some parts of the images looked like they were drawn, as if it were a realistic drawing
Just take a deep breath, and don't let a prejudice develop. Otherwise you'll be stuck on 1.5 once everyone else has moved on to the SD3 finetunes.
(Base models are flexible. There's a very deep understanding of anatomy in their hidden layers. Fine-tuning brings that out, but destroys the model's flexibility.)
which artist did you use for the above images?
dalle in the mud
1990 picture of living room , night , low resolution , jpg artifacts
Spy Kids finger goons be like
I'll use sd3 anyway, i'm just getting on the meme train: surely they'll fix It in a couple of months
o damned, amazing, how do you have it setup to do that? do you midn if I dm you?
I tried an artist name (public domain artist), and SD3 signed its generated sketch with the artist's name. π
I think the issue is not the state of SD3, more like the communication with the community and some white knights trying to defend anything without the ability to reflect and admit those issues
this reminds me of the DiscoDiffusion times = downgrade of epic proportions π LOL
I have already decided not to use SD3 until finetunes come out. Just hanging out to see what people can do with its present state. Convince me of my inability to admit issues. Go on.
zombie breakdancing at the party x3
I appreciate it, you've given me enough to work with! I didn't even know that was possible! I'm always learning something new every day! Thanks!
Are you using OneButtonPrompt too?
Im guessing SD3 will come to Automatic1111 WebUI probably soon
I dont have any problem with sd3 because I lose nothing if it's good or bad. Can just go back to SDXL, only problem I have is the restrictive license for finetunes.
most luxurious aquarium ever
Unless a lawyer wants your money, you don't have to worry about licenses. And lawyers won't lift a pinky for less than $5K.
yess

zombie fish
I use OneButtonPrompt in comfyui to select a specific kind of art, like "Dark" in artists selector, then it will run a lot of prompts automatically, with artists who usually paint dark art
I have a workflow that has been working fine, now the ksampler suddenly errors all of the time. Error is too long to post, bt ends
RuntimeError: mat1 and mat2 shapes cannot be multiplied (1x1280 and 2048x1536)
I didn't change anything from when it was working π€·π»ββοΈ
Oh i'm sorry if this sounded like i'm insulting you. engl is not my native language, if i sound rude it's not my intention, at least i didn't want to insoult you.
I'm doing the same, i'm watching and waiting for refines. What i'm saying: there are mostly two factions: one bashing SD3, the other bashing the users inability. i'm mid ground but i think the base model should do much better.
Maybe try restarting?
I have, and redownloaded models
80's anime style, a girl smoking in a train, she looks like a typical goth girl with black lipstick and black hair, wearing fishnet arm sleeves and a black croptop
Bruuhhh
Oh okay. I'm just here to have fun, make art, and hope everything improves month by month. Let me find an image from yesterday that shows why I love SD3. Oh it didn't save LOL. Oh well.
i was hoping that sd3 could dish out some midjourney level anime...
smoking is haram
I think we all had a lot of hopes
cementery suburb
what the dog doin?
cany you try with Comfy's own workflow? https://comfyanonymous.github.io/ComfyUI_examples/sd3/
the hell is that lmao
guys i think sd3 cannot generate a pencil with an eraser
funny one
a stereotypical yellow pencil with pink eraser
Quality/Anatomy aside, just being able to prompt for this and getting everything into the picture is something that gets me hyped π prompt: "a candid photograph of a blonde woman wearing a pink hoodie and yellow baggy pants. Also wearing black sneakers while standing on a blue towel at the beach. a huge giant ship is floating on the ocean in the backround while the sun sets in the evening. the image has a moody atmosphere"
like this one
Hmmmm...that works ππ»
I gave it too much power 