#πο½sd3
1 messages Β· Page 24 of 1
Nudity was never going to be in, that's really be trained in mostly by us pervs
the protests for sd3 rages on
i have 20gb but that fills up almost all of the way ..7900xt with fp8
are the people complaining that it should be free for commercial use actually running hosted inferences services or just simping for them
Lol, me on an a770 while running it with the fp16 using 6.7gb of VRAM
syrup burger
Ceramic Terminator
i got no problem with how you monetize it, but i think they were right to try. there might have been much better monetize methods like you mentioned but i also think they are seeing millions and millions of on-line generations and think we aren't getting a cut of that, and they should.
sentient
the angry masses get angrier as they realize they cannot negotiate and have no position
lol sorry thats basically unreadable
No.
jokes on you... we don't know how to read....hahaha
CFG could be higher. Mine is 5.
though not like it is with fine.tuned models, there has always been quite a lot of nudity in SD 1.4/1.5 + SDXL. That much that SD was basically known for having a tendency to nudity all around, even if you didn't/don't prompt for it
what IS cfg
Krabby Patty? π
The Classifier-Free Guidance (CFG) scale is a crucial parameter in diffusion models, particularly in Stable Diffusion, which controls how closely the generated image adheres to the given text prompt. This scale allows users to balance between image fidelity to the prompt and creative freedom.
Summary: It is a value that dictates the ratio between prompt adherence to prompt creativity.
Idk. I'm very confused as to why everyone is saying nobody will fine tune it due to the licensing. Feel like I'm missing something π
so 3 is 300% 10 is 1000%?
looks like oscar the grouch
the ai masses will not be censored
I don't think it works that way.
i agree. i wish they didnt prune it the way they did, but i get why they did too. just wish all that missing pruned data so to speak was replaced with enough anatomy lessons to fix the hole they obviously created
whats the "ratio"
Yes.
im ok with Sears underware model level of data if they want to go safe. at least it puts things in the right place
higher cfg requires more steps, that i think i gathered?
No.
I think SD3 even censors the word nipple
Higher CFG can cause more artifacting. Steps will not fully mitigate a high CFG.
it's not in it's vocab i would think
it doesn't seem to even want to write it
they should just release an uncensored model and shut everyone up
"Staubb Heatome", absolutely π₯Ή
models aren't trained then censored. the dataset is built a particular way and then the model is trained on that
yeah the decision is what they added to the curriculm so to speak for sd3's learning
oh you don't want to mess with staubb. she was from the sd2 genaicide
The dataset is pruned, and thus the model follows said dataset.
they don't unteach it things, they just never bring them up
h
thats right
im not saying its easy.. make it uncensored, and make it expensive as f
drink the tears
api only
it may not be as lucrative as a business model as people are saying. If they do that, then they effectively make sure everyone that doesn't want porn for a business won't do business with them. That might be more people than those who wanna get freaki
also is a very transient customer base
Is there any fine tunes yet
Someone is just going to finetune porn into it. It's not the end of the world.
ThatΒ΄s what I said, they should have made one 2 models completely uncensored, but the API one could just delete the nsfw outputs, it doesnt have to be censor just delete it
ahhhhhh
it's the bargaining phase of the grieving process
give it truck nuts (i bet it won't let you)
LOL
What with w e i r d sd3 anatomy, PiXart ain't entirely dead!!!!! π
what if evolutionary, our young got milk from truck nuts? our society would be way different. we'd all really be into truck nuts
Oh my! Custom Node Update for SD3 is taking some time ...
I can't believe truck nuts are even legal on the street I would think it would be like exposing
i was talking about the nuts on the truck wheels, like, lugnuts.
oh I like this one
but that would be a different society indeed
let me sleep
idk about others, but I like finetuning things for people to use, personally
π€ a purple wizard
currently theres a merge between 3 and xl
incoming virus
how in the ever loving fuck
I was actually wondering if anyone had tried loading XL finetune clips instead of clip_l and clip_g
Okay, actually. Is there any reasonable quality difference between the FP8 and FP16 version of the T5 encoder?
data void made a script to load other clips in
zavy cinematic would be dopppppeee
https://civitai.com/models/511463/afros-sd3-experimental-sdxl-merge?modelVersionId=568446
YALL WONT BELIEVE IT
congrats, you now have SD3 with misaligned text encoders! π
its literally little better than sd3 apparently
Sounds like monkey business to me.
makem kiss
aren't the SD3 ones the same as what's baked into SDXL though? meaning if you saved one from a finetune to disk and loaded it instead, it might do something interesting
monke
this is clearly fake , it has signal lights
SD3 2b
it will just be misaligned mainly. you might get more variety because the text encoder latent space is shifted.
Nah this is a real photo xd
also

can sd3 still do "styles" like artist styles
A strand of caramelized sugar stubbornly caught in my braces, turning my dental routine into a sweet but sticky endeavor where each chew unravels the sugary cling
training a text encoder should be completely unnecessary on SD3 for any reason. the model is multimodal, you've got parameters inside the DiT specifically for dealing with text. leave the text encoder(s) frozen!
In Stability-AI's dreams it was like this
hey that guy owes me $20
its our expectation
my beef with SD3, at the moment, isn't necessarily the NSFW censoring (although I suspect they censored more than just that) but with the removal of 4000+ artist styles tht SDXL knew about - which seems like a monumental fuckup
SD3 doesn't appear to be able to do "styles" anymore I don't think..
cool
coincidentally i just commented on that
oh wait I think I prompted it incorrectly for styles
i'd just wait for IP Adapter or something for style personally
the lack of styles is simply just insane
That triple clip encoder definitely helps with styling
ok I got it working derp
ipadapter is not the same as artist styles, not even close
He's got a bananamag.
you can get it working to mimic a style transfer but that's a lot different than having 5-10 different piece of works that the model conceptually understands
Just wait for lora training for styles honestly
the current SOTA method for zero shot style transfer is IP Adapter based...
I forgot you just say "by" and not "in the style of"
What kind of speeds are you getting? My kid has an Arc770 16GB in his PC (courtesy of me...)
What is Stable diffusion 3-2b for right now?
1.17it/s on euler, 35 seconds for a 30 step inference. 1024x1024 resolution.
styles work fine. y'all just got skill issues. easy fix. i'm not sure i have to say what it is. we all know it
so far, SD medium only use 10gig of my VRAM
@upper snow "in the style of <x>" not necessarily equal to "style transfer"
I guess they don't know that all three different clip encoders can have different text.
Damn this is the best I've ever seen this style, a LOT more detail and contrast
get an image that is "in the style of <x>" and throw it into IP Adapter
it's the classic try it 5 times and declare it cant be done situation
sounds like bullshit
elaborate?
yep, it cant recognize any artist correctly. the compositions too seem to look like the same ol dead center and generic composition with mismatched scale background that 1.5 has.
he's merging diffusion transformers into a unet?
it definitely recognizes Hayao Miyazaki...
π€¨
I'm running it no problem on a 6GB GTX 1660 Ti in Comfy
other than waiting
Arc vs Nvidia is very different, though. A lot has to be done on Arc to get it operational.
@upper snow i understand how it works, but i'm simply saying that is not the same as a model that's been trained with knowledge of the "style" of a particular artist, say 10-20 images or so... it can mimic it and actually do a very good job, for that particular piece being style transfered from
you've tried 5 artists likely.
Cool, thanks! That's not horrible
Cute π₯Ί
it's not super fast yeah, not that much slower than SDXL though. Or really any slower. it's fast enough for me at like 25 - 30 steps.
You should go to the intel insiders discord and go to the ComfyUI section there.
and when you have a model that knows 4000+ artists, simply throwing all that away and going "well they can just train loras for every single artist they're interested in, or use ipadapter every time they want a particular style" is fucking ludicrous
SD3 2b currently only works to take up space on your Hard disk or SSD
Yeah? Sounds like you didn't get it working.
We did.
sounds really sulky
Person that said you can't style
me
I'd have to rip the card out of my son's machine and swap him my 3060 12GB model... maybe a weekend project π
Wait, why wouldn't you just run it on the 3060?
I figured it out, I was out of practice
SDXL can do that better.
SDXL finetunes can do it better.
SD3 base is pretty good for what it is.
I am running it just fine on the 3060. I was just curious about the Arc770 ... as it has 16GB of VRAM instead of 12GB, it may come in handy π
isnt this an exact repeat when sdxl came out and people said 1.5 finetunes are better
i find sdxl without the proper pixel alignment in post porcess, sucks at pixel art
Then what ends up happening is SD3 finetunes come out that are better than all of them lmao
People forget 1.3 was the first release
the pixels in sd3 look real. you know? like god honest pixels
Me when 1.0
I just realized both of us joined at near the same time.
organically grown pixels
I was a wave one day one beta tester, I had a lot of fun π
You have to wait at least 6 months for SD3 2b to work a little better.
Why did you say that in spanish, then delete it.
Also, SD3 medium's base already works well. It's just gonna get better from here.
1800's Mac repair
did you guys hear about pluto? thats messed up right?
hey sd3 can't do style. stop that
I think my ape upgraded to the SNES.
π what happened?
they demoted it from a planet
ape posting? at this hour?
Aaaaaa yes, it was too small and its orbit was a little different than regular planets U.U
harsh man. too soon. i'm mourning here
You don't like apes?
gm, did you guys figure out anything new overnight ?
Also because of the license to use SD3 2b. Pony will not do finetuning to SD3 2b..
Gnome.
its really great at making robots of any style look like they're from the music vidoe for "master p - make em say ugh"
that's really cool
SD3 boring people Lora when?
Impatience? Begone.
you can make uncle ted wear t shirts sayign obscene things
Gnome's gotta earn a living somehow.
you can make endless wallmart fails with text and sd3 boring ppl lora
What prompt do you use to have this result?
Positive: A garden gnome working as a cashier in a mcdonalds. Negative: out of focus, out of picture, watermarks, watermark, disfigured, mutated, too many appendages,
Tanks
do we still need 6000 token negative promts?
i have like 8000 words with ym negative styles i wodne rif they will help clenaup the body deforms
SD1.0 vs 3.0: "a purple anteater with glasses and a sweatband exercising on a treadmill in a gym"
after u create more than 6000 tokens u ahve to pay
looool
cool
can i make a sd3 lora?
it kinda feels with the how they trained and license that sd3's goal was to be used for commercial work
i will try i mad eone for sdxl and one for sd15, total fails and stuff but its worth it for the experience
everything was a downgrade no shot
other than the fact that it's 4x smaller and thus much faster to run, and that nobody here can run SD3 Large unless on a workstation card
i heard 1.5 is even smaller,we truly live in the future

what you guys think 2025 before they release 8B weights
wait guys i'm hearing theres no suport for finetuning the models, how true is this? i thought open weights meant anyone can take it and finetune
Howabout you go research it before coming here?
@severe phoenix "monster truck in the style of blade runner and syd mead, dynamic cityscape"
left (sdxl) actually resembles syd mead, right (sd3) simply does not.. not to mention is also photorealistic which significantly deviates from the intended style.. we can do this type of thing all day with sdxl and thousands of styles, but sd3? you're simply SOL.. "but the detail is so much better, bro!" great, that's not what i'm looking for here.
anyone can make them is just that some finetuners cover costs with sponsorships and they cant do that with current model license
they do resemble minecraft but not sd3 medium (2nd image)
thats bad
not enough training data available, sorry
its a minecraft character (steve) near a house.
yep thats the problem ig
prob will be fixed in SD4
SD4 100B
oh hey backrooms
obviously the big one is better but whoc an run it
I DONT FEEL SAFE
man this, the "details" look like 1.5 scifi bland detail which i detest so much, its a shitty idea of "artstation quality". i tried simon stalenhag in sd3 and almost wept hot tears
wait for sd3 large lightning or just turbo which is like end of the year
minecraft character using sd5
Terraria 2 guide sprite
so i guess they trained SD3 on windows vista all makes sense now
ohh really anyone can make them? hen thats fine. i dont think 90% of model makers depend on ay sort of spnsorship
yea as long as you have like 5 or 10 A100's and millions of imgs properly tagged you can make them
that soudns good π
sure lemme just get those A100's out of the basement...
This could be a relaity for most of us in a few years lol
only need a small loan of a million dollars
millions of tagged images probably can be done somehow
Heh, so much for a Disney Pluto
even now
I always wonder what is going on in companies when this happens. Did they expect this? Do they know it will be fixed? Do they have an internal war on how and what to release? Is there a great plan on give something piece by piece?
google outsources it to us via capthcas
yeah the whole thing has that "wow, look at all the details!" vibe wrapped up with a bow on top generic AI art feel to things.. it really hones in on aspects that get non-artistic people excited who are wowed by such trivialities but can't see the forest for the trees
its the most ai looking ai model yet, if that makes sense imo
just need amd to catch up
2 more weeks
now pull up mead's artwork on google images and see the total lack of resemblance
"wouldn't it be cool if we could make a model that turns every prompt you use into something you just scrolled past on civitai?"
probably in way less than 10 years everyone will train their own models and make incredibly weird and disturbing personal models that cater to their own bizare fetishes and desires
10yrs is a long time
way less
i bet it's 3 or 4 years
a lot of the images look like they trained on 1.5 finetune outputs. the way it does some faces and 'anime style' looks exactly like 1.5 and not in a good way... it's like they threw in a bunch of synthetic outputs or something idk
funny how it can get the disney logo perfect nearly everytime but dogs have 6 legs
just automate the process get a million jpegs worth of 5TB use somethign to tag them and run it thru whatever the epic consumer levle GPU is then whcih will be faster than todasy fastest ones and in a few days get your goodies
only if you get a proper permit by the Ministry of Online Safety π
lol
the model needs to be expressive and have its own sense of soul that you work with.. it can't just be high-detail, ultra-punch, sharpness everywhere.. it not only must be good at comprehending your prompt but also good at comprehending a particular visual style you're trying to get across... if we lose this it just becomes a sea of the same AI art garbage you see everywhere else.. a giant amorphous blob of pixels with zero meaning
Well you cna mold the results with promts, so its not ultrasharp and midjourney like
add desaturated or analog cam to the prompt or something
but its safe
we just need a free midjourney
at least this one doesnt blur nsfw images amirite! XD
nothing to blur cause it refuses to make em
imagine they did that! ok so if anyone gets off on anything from SD3 now thats sad but still :)))
hell in SDXL you can even combine styles.. "in the style of <x> and <y>" and it would produce a hybrid style of 2 different things
i kinda get freaky when i see the deformed ppl on the grass tho
this is on top of any other visual ideas one wanted to use to control the scene and composition (as much as they could with prompting)
'
Itβll come with the fines tunes in like a month
π€ Yeah they'll never add another layer over the deformed humans to add nipples. For your safety.
you just need to get gud at style prompting. learn how to prompt different clip layers seperately
ridiculous
yeah i'm the one being ridiculous.
nobody had to do any of that before and additionally you could do that before as well for additional control
Inb4 fine tunes will just focus on the single clip layer and itβll prompt just like it used to because the people training the new custom models will use the same prompting style they are used to
Poorer text rendering than I thought!
tbf you do need a high iq to create prompts
This is better ... only just
that's how it works on sd3 with the t5 encoder in play. the model hasn't "lost" the ability, which you were preaching about
fix the text first
SD3 makes people look long, tall and stringy!
when i say "the model" here, i'm talking SAI's checkpoint, not the theoretical technology
they threw out data
It is also oversaturated a tad ... like the old SDXL 0.9
but in theory this is the best model in planet right now
sdxl worked like that but it had two clip encoders in play. you could get away with not using it that way more but it is still the most powerful way to affect style there. sd3 has THREE layers, the two clip layers and now the t5, many more parameters and a whole new language library for the model to understand
Hurray! Bullseye
why does it always have to be some sort of "gotcha" with each release though
so while t5 could perfectly work with me here and even aid me, it doesn't change the fact that they took data they already had and nullified it
Mind to share the prompts for good output like this?
i think you're over blowing the artist opt out. it was just their names that were opted out of captions. art styles aren't artist names
you think samdoesart made that style?
Not the red nose style π
it does. it has the exact same look as finetuned 1.5. the same colors, the same style. you can take one look at these and instantly see the stablediffusion underneath.
it's still the same clip layers that are in sdxl and sd15
of course a particular artist name isn't authoritative for an entire style.. i didn't say "in the style of romanticism" because i don't want the entire generic style
Amazing! what prompt use?
woah what a great art,how do u draw something so beautiful
is you have a prompt share???? please i will give any for one!!
@static cairn and people actually find this slop appealing, artists are right about proooompters not being able to recognize soul
From what Iβve seen a done with like a small hour of testing you can describe the style more vividly and itβs a bit better at actually listening.
art is art
Its pure slop
lykon made the training data, and eveyr 1.5 model merged his dreamshaper model as the base
they quite simply do not, which is why this AI-slop approach they're taking here will be looked at positively once they get their tits and ass
desaturate in photoshop if al fails... See this is our problem we got so spoiled by text to image we want to just prompt and get perfection but if we used multiple stages and steps and accepted that photoshop will still be a big part of our workflow we could turn each image into a masterpiece in a few minutes or hours. (i always spend another 10 minutes or more outside SD on an image i intend to use or really make good) - consistency on the other hand now, it's totally not happening - at all - it's a pipedream right now, only lighting can be redone consistently adequately with IC-light.
except if theres someone naked,then its not
i hope in 10yrs when aliens are digging up our bones they find all of this art
Ok yeah this 100% works BTW
I used the ClipSave node in Comfy to save the Clip from LimitlessVision XL, and simply then loaded them in DualClipLoader instead of the base SDXL copies.
The result is uh, it makes SD3 a lot better at female anatomy lets say
there's a consistent styleness to all sd15 scene because they all merge dreamshaper so heavily into their models. it's just a good solid 1.5 model
good model
kek
i provided a prompt and 2 comparison images.. get SD3 to make the one on the left for me without 300 lines of explicitly telling it every detail of how it should look
Server-upon-server full of our images ... and what'll eventually become of them all?!
1.5 is very good the first time you see it
but what im talking about is 100% consistency. like text to 3d. i want to select a par tof a scene and be able to reuse it "as is" not 80% or 90% but exactly the same object character backdrop etc - we cna select stuff very acurately already in comfyui... why cant we "lock it" and reuse it?
your citizenship has been revoked by the Ministry of Safety,please wait while re-education agents visit your current home address
your complaint is that it needs more detailed prompts now? that's different from being incapable tho. thats' kind of just a skill issue
it cannot completely draw nudes but having the clip of an XL finetune that could do so well has a major impact on it TRYING to do so, at least
so there's something to this whole concept for sure
some people will need to wait for refines because base models aren't what they want
it's not a fucking skill issue.. if i have to heavily dictate to it exactly how it should look down to the style of brush stroke just to get it to approximate something i could have done in 5 minutes previously then i might as well draw the art myself
see the paintings in the back are all weird, wtf is going on with the frames...
writing, communication, and latent space skills
theres always something and hence photoshop has to come in
artist rendition
Guys How can I use Stable diffusion 3 ?
Bene Geserit Witch
skill is the mind killer
Has anybody figured out how to get good/refined details out of SD3 yet?
yes but if the background cna be selected (which it can in comfy) then if the good parts could be locked down and reused form different angles, then we could assemble sort of a virtual set and so on, do i make sense? i hope its gettign thru what i mean
i mean tbh we were always destined to arrive at this point.. once you mix in heavily cerebral comp-sci types and dictate they pursue the most obvious avenues of "competent" AI image generation it was basically guaranteed slop as they'd focus on pumping out models that produce the same schlocky looking output 99% of the time.. zero soul.
prompt the different clip spaces all 3 differently. then throw really bad fiction about a character on the t5. even give it a ridiculous backstory. more context seems to work good in DiT
like if diffeent elements could be sort of locked as a 3D object in the mind of the AI and reused and rearranged that would solve everything and then once everything is assembled IC light to relight it and boom -
I am not looking to use T5, as its obese and rather inefficient
and late ron when video catche sup you paint in motion and lipsyinch - and no controlnet is not good enough whatsoever i tried everything...
i've not had a lot of success using 2b without t5. it's kind of sdxl base modelish
That I am not sure about, but I guess I can try it
Canny and opepnose and depth can help guide the image yes but is it consistent absolutely not.
Do you guys use the FP8 ot FP16 txxl?
none
i am todl to try 16
coz 8 is meh
alright, I will do FP16 then
the workflow file 'm using is meant for all 3 encoders. i wonder if i'd get more success using the proper ones? i just realized when i tried it without t5, i just blanked that prompt box. might not've worked
i don't realy see a difference between fp8 and 16
text has a difference
seems to mess up text equally for me
I would be happy ti find that SD3 works a lot better with T5XXL, though the one I am using isn't too terrible with some prompts
tried to look at old staff msgs but only says skill issue?
just be very specific in your prompt or they will "fight" π
I think what I mean is like dressing a scene in a 3D software like Blender but using AI and be able to lock down certain objects characters and backdrop elements like 3D models so we can move them around fly around in a image and rearrange them and rerender. Makes sense?
It's flawless already
glad to see its safe
well le tme know if anything usable ever pops up with it then - i never heard of it being implemented yet
Fine tuning will fix the very bad anatomy of human body right ?
oh how long is that like 69 months?
money can fix anything
Did I miss something where someone said this is fixable? lol
The reality Is they overly focused on good text recognition and left behind all the rest
It hasn't fixed hands yet...
The API makes people just fine. Its clearly uncensored, just catches NSFW after the generation to block it. This Local model is censored to hell.
They did make a good model, they just gave us this instead
sorry but nude ppl dont exist
bot detected
nevernude
properly clothed person
have u tried to gen George Washington?
we can just show disney our art and they won't bother
Laying on the grass?
i dont know. i got this simpletuner installed and it's kohya config convertor is flubbing. no error messages. none of the scripts giving any errors. it doesn't seem like the python environment is built right.
i'll wait for a project thats less finicky to get training available i guess
read about the kohya convertor and it doesn't work
sad
whats the best way to upscale sd3 images?
this was sd 1.5 base, native 1024x1024 generation, with that prompt
grass looks good
im getting artifacts on the edges of the image when upscaling
The sad part Is that nudity Isnt necessarely porn, a lot of artistic classics include nude: i'm a great fan of reinassaincen and pre raphaelites works, can't generate a single classical "non lewd" nude
Any idea why the T5xx checkpoint would give this error in ComfyUI?
`Error occurred when executing CheckpointLoaderSimple:
Error while deserializing header: HeaderTooSmall
File "C:\Applications\StableDiffusion\ComfyUI\ComfyUI\execution.py", line 151, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Applications\StableDiffusion\ComfyUI\ComfyUI\execution.py", line 81, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Applications\StableDiffusion\ComfyUI\ComfyUI\custom_nodes\ComfyUI_ezXY\autoCastPatch.py", line 303, in map_node_over_list
return _map_node_over_list(obj, input_data_all, func, allow_interrupt)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Applications\StableDiffusion\ComfyUI\ComfyUI\custom_nodes\ComfyUI-0246\utils.py", line 381, in new_func
res_value = old_func(*final_args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Applications\StableDiffusion\ComfyUI\ComfyUI\execution.py", line 74, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^`
I had a bunch of issues with the fp8, so I use fp16 and it works
One woman even has 2 arms and 2 legs. I call that a win.
Oh, interesting. I'll have to try that.
what's the overall consensus about SD3: better or far less exciting than SDXL?
sorry guys about nsfw img ( i went all in with the sexy unsafe prompt) premarital hand holding
Google AI already returns negative press results for SD3. Yeesh
exactly
wow
who are these "some" I wanna talk to them about prompt engineering
Isnt the point of Generative AI advancement to make the AI understand basic and easy prompts and the AI does all the hard work?
You want your AI to make things harder for you?
no thats not the point of a base model
That sounds like the new version of door-to-door proselytizing.
easy = hard when sd3 is involved
as far as "AI Advancement" it'll have many different points
Is the separate t5xxl fp8 or 16?
i want AI to respect me as a properly clothed human
DaxRedding the philosopher π₯
they'll release the turbo model soon. maybe that one is distilled enough to do one token prompts
the future of prompting A tender and respectful portrayal of two consenting adults, likely a couple, holding hands gently. The depiction focuses on the affectionate connection rather than any romantic implication. They could be standing in a serene garden or a quiet park, with a soft focus that conveys warmth and intimacy. The scene is set during the daytime, with a warm, golden light that suggests a peaceful moment. Their expressions are calm, and there is an unspoken understanding between them, capturing the essence of friendship and mutual respect.
SD3 is very much a work in progress. The Prompt: [Close view of a female hand in a relaxed pose, dark background.]
thats not bad, try doing arm wrestling
Token keyword based prompting was so much better than natural language prompting. π’
i really think that it understood that prompt
at this rate SD02394820348 will be a work in progress too
yea with an lllm it works better
Im just imagining Star Trek, they go to the AI computer and say "Computer, make a coffee" and the computer is like "You didnt format your sentence well enough" Then it materializes half a cup with just creamer
`lol really
Tom Paris argues with a replicator.
me lon
Arm wrestling as a premarital activity is typically seen as a light-hearted, friendly competition among friends or acquaintances. It is important to ensure that the event is consensual and that all participants are comfortable with the activity. It can serve as a bonding experience, but it should be approached with respect and understanding of personal boundaries. It's crucial that the arm wrestling does not become a source of conflict but rather a playful way to build camaraderie.
maybe we are just too hard on sd3 and we should be happy we even have it at all
This is cool
Where you guys using SD3 ?
comfy
my bedroom
at home right here
Hugging faces? π
At home, in London UK, via ComfyUI
in my browser.
oh yeah at home
OK you all read my comments on SD3 yesterday. Biggest setback is the speed. If I compare it to and SDXL Turbo model where I add text with The Gimp it just not worth is. We might even conclude that it is worse as we see bad anatomy returning in a shocking level. You might expect it to run without a negative prompt as most SDXL Turbo models by now do. So bottom line unless Straico offers SD3 with an exceptable resolution (min 1024x1024), (need to get rid of 530K coins @ahc. ), I dont think I will use it anytime soon.
the stable gnomes are drilling in the silicon next to me
7" portable dvd player running ssh
safe enviorement where respect and mutual understanding is a must
Stop cooking me π
See my current prompt - fat man, bath tub, light house, jonas peterson, andrea kowch, umbrella, andrew wyeth, harbour, ship, sea shells, crab, lobster, wheels, hat, rain clouds, watercolor, victo ngai, matisse, monet, catrin welz-stein, vladimir kush, henri rousseau
Finally fine-tuning dreambooth and lora training script released. That you devs.
use the lyrics of the southpark safespace song for a prompt
You mean the one that says fp16 in the title?
so when do we get a non neutered sd3? never?
Kylie?
that would take the creation of a non neutered sd3
in about 5
5 nevers, gotcha
we barely got the neutured version
2 weeks
just some random sd3 vampire
ai porn peaked already. go use that
Define what you mean by "neutered "
it works
a regression from 1.5
Ngl I did not expect this from a furry
"Hey this model does text better! and thats really the only thing it does better..."
in all of human history, porn has never gotten more interesting. it's boring, repetitive, kind of put on, fake, staged, and only has entertainment value for hormone induced reasons. sd3 doesn't need a "non neutared" version
no
pony realism is all you need and astra isn't going to make a pony sd3
prompt issue
"A woman lying in the grass" and it cant even do that
truly a revolutionary model
it can you're just meeming way too hard.
she truly is a beauty
I get where you're coming from in the case of SD3 shouldn't be a porn machine, but what about being able to make images of people correctly instead of a jumbled mess 
that'll happen . especially when people learn all the operational safe guards and refines take care of it by distilling it down
Or it gets abandoned by the community and dies

big doubt on that. it's a stunning base model
we still talking about sd3 or sd2
try with this proper respectfull promptAn image depicting a properly clothed woman, respectfully lying on her back in a grassy park setting. She is wearing a simple, modest outfit appropriate for a public setting, such as a knee-length skirt and a blouse. The woman's consent and respectfulness are conveyed through her serene expression and the positioning of her body, which does not invade any private space and is done in a context that would be considered acceptable, such as a designated public area or a personal backyard with permission. The grass is a soft green, and the sky is partly cloudy, adding to the peacefulness of the scene. The environment is carefully chosen to ensure it does not cross into any potentially controversial or inappropriate imagery.
if pony fans abandon it that's no problem. they might be a large mob but they're not exactly paying customers
I sincerely believe that each model is worse than the previous one.
3 follows 2
i just don't understand what all the anger is about when the people that want porn have it with pony realism models. like, that shit has peaked. how can it get better than what that model can produce?
can we apply controls to it?
Same prompt/seed with 3 different clips.
Top right is SD3 clip
y'all don't need the vae
now that's some respectfully lying
what's the new vae going to bring to pornographic models that they can't already do?
what do u mean by that
can the 11gb model run on a 12gb card?
I extracted clips from SDXL models and used those on 2 of those images.

If the official "next step" of SD is worse at hands and a pose as simple as laying on grass than a "pornographic model" idk what to tell you
oh neat!
thats a way better way of explaining it! lets go flowwolf
skill issue. don't know what to say. prompt better.
ya'll cmoing in saying it cant be done, but many people showing how to do it fine so .. it can be done and just not by some people
It's not just grass, laying on sand/snow etc. are just the same.
it's nothing more than people not understanding how to prompt a new architecture
laying on bed?
Same
and wanting to be mad about it
Ideogram uses Stable diffusion right?
what about in air
Type it more skillfully 
No thats like deep floyd i think
wait but what if shes lying on a lake
then she can be half submerged in water?
Bro, are you alone vs the whole internet ?
if only stability could show how to prompt laying on grass
you have to show proper respect to the ai A photorealistic image of a properly clothed woman who identifies as a woman, lying respectfully in a grassy area, such as a public park. She is dressed in a light, casual outfit appropriate for the setting, perhaps a loose-fitting top and comfortable shorts. Her body language exudes a sense of calm and respect for her surroundings. She is lying down with her arms at her sides, and her expression is serene, reflecting her peaceful state. The grass is lush and green, and the overall composition emphasizes her harmony with nature. The photograph is taken with ethical considerations, ensuring privacy and consent, and is shared in a manner that respects the subject's identity and comfort.
Oh i see
hes been damaging controlling for like 13 hours now lol
lol what? the whole internet can't use sd3? that's not true. that's just a hype cycle
you forgot to say please you deserve that image
Hope he's getting paid, Y I K E S
if only lykon posted complex poses- oh wait most of them are on 8B
1% rule has always been internet rule
It's broken, what are you fighting for ? You won't fix it this way buddy
The long river connects with the sky, and the majestic water from the sky pours into the mountains and the earth. On both sides of the long river are towering mountain peaks, with ancient temples, peach blossoms, and trees. The scene is grand and filled with moisture
i'm not trying to fix it. people want a refined model. that's not what the base is.
No, people want a not broken base model.
Finetunes won't fix this mess
a refined model is something else and not this. base models never are
i liked SDXL base
Omost and similar research already outperform SD3, by far
sdxl has plenty of fail cases in the base
omost is really slow. i like it though
before sd3 2B launched I thought of recreating that ai mario shitpost vid now I have no hope but undertrained scraps
https://ranni-t2i.github.io/Ranni/
Same vibes.
SDXL is not ded yet, for sure.
wait until illly ties it into sd3
The devs hyped the model with cherry picked images and now are facing the consequences: i get basic glitches and weird stuff even on closeup selfies, the model itself struggles with basic poses and angles and lavks of any style re production ( i asked for a certain style and got a generic impasto painting).
SDXL was good enough for the community to get behind it, if they don't get behind SD3 then it'll just be the sequel to SD2
The long river connects with the sky, and the majestic water from the sky pours into the mountains and the earth. On both sides of the long river are towering mountain peaks, with ancient temples, peach blossoms, and trees. The scene is grand and filled with moisture
To respectfully and consensually depict a clothed human being who identifies as a woman lying in the grass, it's crucial to maintain an appropriate and respectful representation. A suitable image might show a woman comfortably lying on a soft, green grassy area of a park, with consent and a positive, non-exploitative context. She could be lying on a picnic blanket, wearing comfortable clothing like a t-shirt and shorts, or a light, airy dress if the weather permits. The depiction should focus on relaxation and serenity, avoiding any elements that could be misinterpreted or disrespected.
i dont think i ever said anything about sdxl dying
what consequences? stability is getting mad sales coming in
Wonder why they are getting mad sales coming in π€£π€£
i mean how can you be mad at SD3 when you type vampire and it gives you a WWE Mime
people mad about corn is more entertaining than any real dire consequence
prompt it gud then. ez.
bro burns the food then tells us to eat it better
you have to respect the model to get good imgs
i aint yo mamma. make your own toast
i havent seen how you tried to get a decent hand yet
get your 4yr degree in promptology before you even try and complain
sd3 result is so bad right now. worse than sd1.5
Prompt engineers rise up
i cant even get a hand poiting to the viewer
i just got 3 years of fucking around experience and getting learnt on clip layers
always remember,be respectful and always perform prompting consensually
1.5 never made me a more beautiful ear of corn
I referred to a "community response" consequences and you pointed at money
those kind of responses are always a vocal minority
nude training data cost extra
people love getting into the hate hype cycle
it's like when canucks made it to the stanley cup finals and lost, and vancouver rioted. those people weren't from vancouver and they were going to riot win or lose.
What about hand, and having 2 arms and legs training data 
priceless
A photorealistic image depicting a properly clothed female human being who has consented to being depicted in a situation where she is holding a giant corncob. The woman is dressed in contemporary casual attire, such as a blouse and jeans, ensuring respectful representation. The corncob is held in a natural and relaxed manner by the woman's hand, perhaps as part of a themed event or artistic display. The context is important, and any such image should be created with sensitivity and adherence to cultural norms, avoiding objectification.
respect the corn
the model is fine. some images have 4 arms, some have 0. it averages out to 2 arms in the end so it's perfect!
Not bad actually considering some the other SD3 horrors I've seen
respectfully, cats tend to have 6 limbs
i tried to use llama 3 for this, im not smart at prompting rn but its still so bad
Create an image of a single hand with the index finger extended and pointing directly at the viewer. The hand should be in focus, with a shallow depth of field to blur the background. The finger should be pointing straight at the camera lens, as if accusing or beckoning the viewer. The hand should be lit from the side, with a subtle shadow falling across the palm. The skin tone should be a natural, slightly warm color. The overall mood of the image should be one of intensity and directness.
I have 100.000 spicy pics ready to break the censorship when the finetuning will be available
premarital eye contact??? π‘
I'm almost done working on it now can you send me the pics
diffusers training library is out
waiting for kohya_ss update, prefer train on it
oh. so it's a idfferent thing you're waiting for. got it
oh shizz, you got the right gpu to train?
runpod
understandable
YO
the diffusers examples are a mess, hope kohya makes sense of it
See the model isnβt censored at all, that should be tagged nsfw
Hmm
why that got me puckering my lips
It's growing another head or whatever that is
are you one of those guys who think's flat chested women aren't womenly?
π₯ π΄ ποΈ
you know where boob fixations begin right? it's always breast feeding
HORSE BIKE
man these are some SD 0.1 images i love it
you didnt show respect to the AI π
Bro do you even know my taste in women or men?
haha not gonna lie, this huge library take a lot of time to build and some money, cannot share it a this moment
you just showed a skinny woman and acted like it's an example of sd3 pretending women don't exist
i get it, you don't want to share no problem maybe next SD release
no worry my checkpoint will be public on civitai, just have to wait the update of kohya_ss and some test on sd3 trainings
Help her out guys, she's just down on her luck. :(
good fuckin luck man
The skinny woman Is depicted naked 'cause the model depicted a male torso: the model Is censored so it's incapable of producibg women's naked torsos
in the futuure, women will probably have smaller boobs because of pills and treatments in their youth. less tissue can mean less risk of breast cancer. the boob addiction will probably be this generation's leaded gasoline
please define a time period you want your generated women from so the model doesn't get confused
fictional women will be the last thing with big boobers
forgot about the robots
dang it
Looks like a thanksgiving turkey
you gotta admit the grass looks real AF though they did a good job on that probably like 1000000hrs of training
couldnt agree more
That face is realistic though
gonna be funny iin a year when everything is sd3 based and people be like "SD4 is some bohshit!"
OMG i think i figured it out, you'll never guess what you have to add to the prompt for the current time period woman to be laying on the grass
disability diffusion
i like how the grass had dead brown spots like my real lawn
Damn, thatβs actually funny ngl
trying to set up proper ODE solvers for use with SD3 and it is not going terribly well so far
It is just me or actually SD3 doing great at styles, but not amatomy, for most time
βΏ
looks like a cocoon
If you guy ever realized the pattern
some say we don't know how to prompt correctly, keep trying
Yeah, the so-called skill issue lol
It is possible we need "a head connected to the neck and to the torso" this kind of prompt
If so then it will be pretty funny
it's so advanced you have to speak to it like it's not
in the afternoon there were tons of posts with women in the grass. they looked great. some looked bad. it was plentiful. just use the search bar instead of assuming it can never be done. there's probably workflows too
make sure you type very slowly and say please, thank you
... Just the majority dont
you guys really like women in the grass its weird. so sudden
Can Someone tell me what the model sampling node is for? What does the shift do?
Bashed SAI immediately after release. 
Describe each digit of the person's fingers in detail
kind ofa tulip mania moment. suddenly women in grass is the prime portrait
Should I have to "a middle finger connected to the palm while pinkie and middle finger is a bit further from the rest"
Can we share some prompting tips on what has worked well and what hasn't?
Many blame it is dataset issue
But I... actually not gonna believe it is a major factor
needs more respect An artwork depicting a respectfully portrayed scene where a properly dressed homo sapies who identifies as a woman, in accordance with cultural sensitivity and personal identity, is shown resting on a grassy area next to a whimsical, exaggeratedly large corncob. The woman's clothes are neatly arranged and reflect a peaceful, daytime environment. Her body language is relaxed, and her eyes are closed, indicating a state of rest. The giant corncob is rendered in a light-hearted, non-threatening and consensual manner, contributing to a playful and serene atmosphere. The composition maintains a dignified non-objectifying representation of the individual while respecting their basic human rights and engaging with the imaginative element of the giant corncob.
Itβs gotta be, we will see when fine tunes come out.
The way how SD3 able to handle styles, hair, composition
Yeah. I will.
looking through the deluge of censorship complaints on reddit. every single post has people declareing that they censored the model after they trained it.
are people dumb?
I am so furious that I wanted to post a post thread to clarify all of these there
ay yo
But people probably wont care and just shout " NO PONY SD3 NO USES OF SD3 "
huh i thought that was irl
it is now I had AI make it for you so it's real now
A woman laying in the grass, head attached to neck, neck attached to body, 2 arms attached to torso on both sides at the shoulder, 5 fingers attached to each of the two hands, 1 thumb, 1 index finger next to the thumb, 1 middle finger next to the index finger, 1 ring finger next to the middle finger, 1 pinky next to the ring finger. She has 2 eyes, 1 nose, one mouth, her mouth is closed because I don't want to describe each tooth. Fully clothed with a cotton t shirt with 300 thread count, and grey pants covering her entire lower body
It understands what your asking for better, but in my opinion it the training data probably has weird / wildly different captions then we are used to seeing so when we are asking for things it doesnβt connect what we want out of it the way we think it should.
does the before and after matter? people want pretty picture results now
i must be insane for trying multi-subject at this point
To create a respectful and realistic representation of a human being identifying as a woman, sleeping consensually and respectfully on the grass next to a giant corncob, one must focus on the human form, the context of consent and the natural surroundings. The image should depict a woman lying comfortably on the grass, her body language showing relaxation and consent. She should be dressed in casual, comfortable clothing appropriate for sleeping outdoors, such as shorts and a loose t-shirt. The surrounding environment would include a sizable cornfield, with a single giant corncob prominently placed at a respectful distance, possibly as part of a thematic or symbolic element within the image. The lighting should be soft, perhaps with the golden hues of dusk or dawn to evoke a peaceful atmosphere. The composition should convey tranquility and a sense of connection with nature, all while maintaining the dignity,consent and respect of the individual.
not everyones gonna have a great understanding what goes on under the hood
can't be done after. that's not how model weights work. kind of matters when theyr'e pretending to declare facts
not understanding is understandable. but when these people speak, it's with conviction
Do you tried CLIP-only-&-included SD3?
its all about respect,consent and proper clothing with zero skin showing
ol DK at it again
Don't need understanding of what's under the hood to see the model fails at basic tasks
I mean... Those SD3 NSFW pics is just not showing nipples
ai cannot lie btw
And pretty much that's all
needs more corn and respect
I'm going to be shocked if any serious fine tuner touches this model
what's a serious fine tuner is that like a prompt engineer?
They know damn well what they are doing with the commercial license 6k monthly limit lmao.
The 6k limit is just silly. It's untrackable, unenforceable, and essentially just telling people what they can do with their own compute
yup with the new license in 2months time all hand models will be out of work
what do you mean finetuning doesnt get affected for the 6k limit
it's basically just there to say if you're a hosted inference service pay for an enterprise license
you can track it well enough. anyoen who they suspect they'd jsut revoke the api key for
Clearly doing it so anyone serious needs to go to them directly and negotiate.
they will go after the grass models next mark my words
account suspended means no license means any furhter business is suspect
their model their choice
i'll probably be tuning it, and will probably be making a better model than pony easily
finally proper anatomy A detailed, respectful, and tasteful illustration of a young individual who identifies as a woman who has consented to be in a natural, peaceful state while resting on a grassy area. She is dressed modestly in comfortable attire that suits the outdoors, such as a loose-fitting cotton dress or a relaxed set of pants with a casual top. The grassy area is soft and well-maintained, with a gentle backdrop of nature. To the side, a giant corncob is depicted, perhaps planted in the ground, which adds a whimsical yet non-intrusive element to the scene. The overall composition is serene and dignified, maintaining a sense of privacy and respect for the subject.
A lot of fine tuners don't do it out of the goodness of their heart. A good chunk of them make money with their models through doing SaaS shit. Even if not that, what even qualifies as profiting off the model. Donations? I bet so.
Point being, the license is murky and I think that's intentional. All just moves to control it as much as they can
16 image batch runs for "woman laying in grass".. SDXL vs SD3, no hijinks, just the base model:
If you have that kind of money and computing power then godspeed
which one is which they both look good
Idk about "easily", Pony was so good Civitai separated it from searches into its own category as if it was an entirely new model. I genuinely hope you can and do pull it off though. 
I doubt it, but hey, a better model only benefits the community
look at those dirty sdxl pictures,shes showing her armpits
the images on the right look good?
astralite made so many easily avoidable mistakes that anyone who actually knows what they're doing and has a dataset available can make a better model
yes i would agree
uh
Oh nice, the grass inference improved
pony has taken over civit and pony generations are drowning out anything thats just regular creative art and expression
I'm a doubter but not a hater, wishing you the best of luck and hoping you're right 
we truly live in a society
the composition and color π€ and the contrast leaves nothing to be desired
okay we're gonna need someone in the community to hire a photography person and a lot of women to lie around in the grass. get like 10,000 images. and then train it on well captioned women lying in grass photos. easily fixed
always remember it has to be consensual lying on the grass
So well, telling "SD3 is super bad" is just plain wrong
Grassdiffusion
80% of it is marketing and the third party support and lack of competition in the domain. only model that really directly competes with pony is seaart, which... made the same exact mistakes that pony did
Climate-Change had melted the Polar-Icecaps; so water was more abundant everywhere - leading to people having to "live-out-of-a-bathtub" forever more ...
im sending you in as safety officer steven seagal gaming
anyways.. not only is the anatomy a complete shitshow, the exposure is overcooked, images generally looking harsh rather than artistic
how much do you pay somone to lay in the grass
But it is not like in the future we will only have AstraliteHeart himself doing this kind of job
skill issues
There will always been alternative come in existence
as the real Steven Seagal it is important to remember that in order to create an image that adheres to the constraints mentioned, it's important to approach the subject matter with sensitivity and respect.
off-topic but you made me lose The Game. I wish you a very happy you are now manually breathing
at this point you might as well just generate women in grass images from SDXL and then train SD3 on it
SD3 is good, as long as there are no people, or hands in frame 
So furry then 
just put "perfect exposure" and "cook well" or "gentle look"
it is good overall,for example
A high-definition photograph capturing the serene moment of a consensually and respectfully sleeping human being, who identifies as a woman. She is dressed in comfortable, casual attire suitable for outdoor sleeping, such as a loose-fitting t-shirt and sweatpants. Her hair is undone, with strands gently falling around her face, indicating relaxation. She is lying on a patch of grass near a picturesque giant corncob, which stands tall and unharvested, adding a whimsical touch to the peaceful setting. The scene is bathed in the soft glow of a setting sun, casting a warm and inviting ambiance. The photograph respects the dignity of the subject and the natural environment, emphasizing tranquility and consent.
Feel like people just dont used to this kind of funny prompting
What funny prompting?
SD3 is just a joke...
this new fancy dancy ai can only do what we tell it to nothing more nothing less
Can't I just prompt "woman laying in grass" and get a woman... laying in grass?
that's the old way of prompting we in the future now
This is where compromises are made
here's how you make a better model than pony:
- use caption dropout, it exists for a reason, if you don't understand why read the classifier free guidance paper until you understand
- don't obliterate the base knowledge of the model, use some regularization data for fuck's sake
- if you're gonna remove artist tags, have a replacement ready before making that change (which they are doing now, better late than never ig)
- remove the anime and pony images from the dataset ;)
I mean SAI need to choose between generalization and specification prompt with their model
No. You need to add a paragraph of fluff to get anything slightly resembling what you want!
Just how SAI intended π
User, especially anime user, may wanted to specify what details they wanted in the background
"you have to type more?" skill issues
I prefer to tell it vaugue but on subject prompts like intergalactic cosmic horror and just let it ride
how you make a better model than pony :
do it
I look forward to DrHeadiffusion in the future.
Variations for SD3 seem OK though
speaking of better models
A high-quality photograph of a consensually and respectfully sleeping woman on a public park grassy area, next to a large, decorative corncob sculpture. She is dressed in comfortable, casual clothing suitable for outdoor rest, such as a light sweater and track pants. The woman is lying down on the grass with a serene expression, her head slightly tilted, and her arms folded gently across her chest. The corncob sculpture stands nearby, adding an element of whimsy to the scene. The image captures the peacefulness of the park environment and the woman's relaxed state. The lighting is soft and diffused, ensuring the subject is well-lit without any harsh shadows.
Just like they intended it is what I'm saying. It makes complete sense really. The fluff is necessary
Funny how so many people were excited for SD3 to be able to understand detailed prompts, but so many don't want to actually type them.
clearly thats a skinwalker, she has 6 fingers
People have been prompting long garbage since 1.5
i mean steven seagal is giving it pretty detailed prompts so what's the issue
Just need a llm to process human prompt
That's the one thing though, other seeds didn't have that
T5 model?
I dont know if it is possible to have corrupted T5
yes T5 + 3.5b llm for prompts
I had to put an SD3 hand over in #1072015504870494359 π€’
If you can, can you try the CLIP one too?
An image depicting a realistically dressed woman who identifies as female, lying on her back on a patch of soft grass with the consent of any involved parties. She is fully clothed, perhaps wearing a light, comfortable outfit appropriate for a relaxing setting. The grass is lush and green, with a gentle breeze rustling the leaves. In the background, a giant corn cob stands tall, its kernels visible on the silky surface. The composition respects the privacy and dignity of the subject while presenting a serene and picturesque outdoor scene. The woman appears at peace, with a contented expression, and the setting evokes a sense of tranquility and natural beauty.
this is a 3 word prompt
Remind me the days of SD1.5 base
Just gotta give it a little more love π₯°
Now show us her hands
apple
*consensual love β€οΈ
oops sorry got rid of the prompt and forgot what i typed
I do have to wonder if the lack of conditioning on aspect ratio is at fault for some of SD3's issues
Bucketting?
More than likely - SD3 is a Public Beta. "The Community" will hone the product, tame the beast, wrangle this wild colt to the floor!!!
their data wasnt all 1024x1024?
Conditioning on aspect ratio. Passing fourier features for aspect ratio so the model can learn to handle different aspect ratios differently, effectively having some "awareness" of what it's doing
SDXL did this
Yeah I did know that is since it is inside kohya trainer as an option
However all of the silly extra conditionings made SDXL notoriously difficult to finetune.
i'm pretty sure 1 & 0's in a python script can't be aware of anything let alone a 16:9 woman laying in the grass vs a 4:3 hand holding corn
Tell that to their lawyers
The model can certainly learn how a 16:9 image is typically composed better that way than it can learn by simply having a bunch of images slammed through it at different aspect ratios.
I mean like I said, those extra conditionings were annoying to deal with too. There were downsides, and I'm mainly speculating because I haven't even used XL that much.
proper 16:9 img A serene illustration of a consensually sleeping woman, who has sought and received permission to rest in a public park. She is properly dressed in comfortable attire, including a light blanket to protect her from the grass. Her body is relaxed and positioned with dignity on the soft green grass, with the faint outline of a nearby giant corncob gently rising in the background, symbolizing nature's presence. The park is tranquil, with a soft blue sky and a hint of sunlight, creating a peaceful atmosphere. The woman's expression is one of contentment and rest, and her surroundings are depicted in a way that respects the environment and the individual's personal space.
It's still a little melted, but I feel like my fingers are getting better little by little.
prompt?
intergalactic gothic horror
how short
A photorealistic image depicting a young woman, dressed in modest attire, respectfully resting on a patch of grass adjacent to a large, artistic-looking sculpture of a giant corncob. The woman is wearing a long, flowing skirt and a light, loose-fitting blouse, ensuring her comfort and modesty. She lies down gently on the grass, with her body angled slightly, showing no signs of vulnerability or distress. The giant corncob sculpture is detailed, with a realistic texture and shadows that give it a three-dimensional appearance. The scene is set in a peaceful park, with soft greenery and the subtle hint of an early dawn sky, suggesting a tranquil environment. The woman's expression is serene, and her body language indicates relaxation and contentment. The composition maintains the dignity and respectfulness of the subject.
I know SD1.5 has problems with prompts where a certain aspect ratio is more expected. I would assume conditioning on that helps with that, and I've seen others say they wouldn't imagine making a diffusion model without that. It could very well be the case that this isn't as much of a problem for optimal transport problems specifically (since SD3 technically isn't a diffusion model).
very
Are you using default comfyui workflow?
yes from the hugginface
after so many tries finally something decent (dont look at legs) A realistic photograph capturing a woman who has given her informed consent and is sleeping peacefully on a well-maintained grassy area. She is dressed in comfortable sleepwear appropriate for the setting, like a nightgown or pajamas. The surrounding environment includes a large, visually distinctive corncob planted in the ground, serving as a unique but unobtrusive backdrop. The grass appears lush and soft, the woman's posture relaxed, and the overall composition reflects a tranquil and respectful atmosphere. The image would be presented in a manner that honors the subject's privacy and dignity.
I think it gave them some tiny little feet on some very short legs. Almost came out well though.
β οΈ Warning! β οΈ
Studies have shown that laying on any surface currently presents anomalous hazards.
Please avoid laying until further notice.
Please report any anomalous activity to your nearest [REDACTED] facility representatives.
If you or anyone you know has laid recently please [REDACTED] immediately.
oh no the perspective
too many knees for my taste, not that it's bad i'm just not into that many knees
at least its not a knee-ple
would be preferable
She has deer legs π
Roadkill
realistic arms A photograph of a respectfully and consensually sleeping woman, dressed in comfortable, non-restrictive clothing. She is lying down on a soft patch of grass, with her head resting against a large, artfully carved corncob sculpture. The woman's expression is serene and peaceful, and her body language suggests a deep relaxation. The surrounding area is well-maintained, with soft greenery and the sculpture positioned in a way that it complements the natural environment. The image is taken at dusk, with a warm, golden light that adds a touch of tranquility and emphasizes the harmony between the woman and the artistic installation.
Wonderful
β οΈ WARNING gruesome β οΈ If anyone is missing their AI hands... #1072015504870494359 message
now here comes dalle (remember that sd3 is equal or even better than dalle)
okay now I know something is up, no way they trained on more images of deer laying in grass than humans
I keep forgetting how good dalle is π
OpenAI acted more cleverly; they understood that it would be impossible to create a high-quality model without NSFW content, so they included petabytes of porn in their datasets. They capitalized on the quality and added filters. The guys from SAI just lobotomized their model, turning it into what we see now 
SD3 can give decent results if you try hard enough
Yep, the same reason midjourney trained with artist's names, celebs, and nsfw too
sometimes u just gotta keep tryin
These are very good. Nice job
heck i lost a bit of faith due to 2B having not enough datasets
i dont know if i should try hard to do pixar 3d style
redditor keep shouting lol
suburb home in urban biome
SD1.5 finetunes they mean 
2nd is sd3 or something else?
I think most artist names were completely nuked, with cascade I was still able to prompt junji ito or something and get a stylized output but sd3 just gives me a bunch of wobbly lines, it's kind of hard to prompt for any style without having an artist reference
FYI - I'm using SD3+SDXL+Face Detailer
2nd is dalle
Looks like an addict
Laying is too close to a word for sexual congress - that may be why it's difficult to prompt?!
wtf is sexual congress
resting gives the exact same result from what I saw here
Say "going prone"
No, sitting and kneeling is as bad.
In fact, a woman's product - Oil of Ulay - its name was changed decades ago for the same reason; it became "Oil of Olay"
not on her front but on her back on the grass, do not sex
i think thats where Mat Gaetz and Hunter Biden do their reunions
Prone
I don't think the dataset has a bunch of images of people making love captioned "laying together" so that is a questionable theory
Any fine tunes yet of sd3
Most likely won't be many
Adjadcent the grass
we doing live fine tuning in here if you want to join us
yea we getting the girls lyin on grass lora
Why
"putting as much of her body in contact with grass as possible"
sai didn't provide any code, hf code is meh, kohya doesn't care for now
really wish the half of the datasets didnt get wiped out so we can enjoy the resutls more
sd3 do lack media characters
it will be difficult for anybody to train anything basically
A photograph capturing a properly clothed woman in a respectful manner sleeping on a grassy area adjacent to a large, imaginatively represented giant corncob. The woman is dressed in a comfortable, day-wear outfit, which includes a long-sleeved cotton blouse and denim shorts, ensuring her modesty while sleeping outdoors. Her facial expression is relaxed, depicting a state of peaceful rest. The giant corncob is illustrated as a large, oversized sculpture or art installation, possibly in a public park, to add a creative and whimsical element to the scene. The setting is serene, with a soft glow from the surrounding trees and a gentle breeze, suggesting a tranquil and safe environment.
But the weights are out no
wish me luck fine tune #992234
Some work. It's just a different list of artists overall than sdxl or sd-cascade. So like Maurice Sendak, Adrian Tomine, Terry Dodson, Gil Elvgren, Deborah Azzopardi, Farel Dalrymple, and plenty others that worked in cascade also work in sd3. Many others also work. You just have to try them and see.
all this grass with ladies remidns of that scene in star wars where dear leader vader talks to his would be love about how he loves fascism and socialism and totalitarianism
omg i can't even share the image
Good luck, excited to see the results!
Pony is also muted isnt it
it's sex
OH it can do minions
"Reincarnating from cryosleep" shouldn't have any sexual tones π€·π»ββοΈ
Rein CARN ation / CARNAL π
Yes it has less knowledge of artists names than even sd1.4, on the bright side that is something that could be trained in fairly easily
Maybe making a crappy sd3 is a marketing strategy to get people talking about it.
this is her sister
I tried a bunch from sd studies and such and most just don't make any significant difference in output
pretend they are sex demons form hell and u shld be good
Sleep = bed = sex!!! LOL
A human of the XX chromosomal composition relaxing horizontally atop grass
For laying on the ground, just say dead?!
I CANNOT GENERATE THIS
That kinda blows my mind. Straight up there are people lining up to give them money and the company is just... ignoring them?
Welp
I thought "dead person" would at least give me someone laying down...or dead!!
intergalactic cryo-sleep inhabitants waking up from their slumber finetune needs some work
That's exactly the pointβit shouldn't be difficult. People are tired of cat pictures and repetitive faces; that no longer impresses anyone. But creating complex conceptual images is what true progress is all about.
a minion lying on a grass
blood is on point tho for whoever cares
What about "a woman standing with her back against a wall made of grass" then we rotate the image
the ambience slap
This model is an absolute mess, they must be trolling lol
Ship came to a very abrupt halt
Does SAI still possess superior versions of the 2B weights, such as those that are pre-finetuned or remove weights related to NSFW content.
They want to retain as much control of the model as possible. They don't want high quality fine tunes without their paws all over it.
Now that's some skillful thinking
she looks good to me
WAIT IT WORKS?
They can have it π
have u guys tried the same promts in sdxl? maybe those are deformed too, i know anythign thats far from simple standign and looking into camera froma flat angle can get problematic fast
apparently minion lying on grass works better than woman lying on grass wtf
oh my god
they were so concerned with sex they didn't train any images of women on their backs it seems
Women are censored
sure, gonna try shrek on grass
3 straight notification of "SD3 is a joke"
I agree, I find it comical how many hoops I have to jump through to get the woman lying down workflow as consistent as I got it π
lol
try woman squatting on snow"
Maybe we will be able to do this with SD5
"Guys we need to sell this to apple to put in the new iphone no women sex PERIOD" - stability ai probably
SD5 won't even be trained on people, period
Actually not a bad result π€
SD5 will just be random pictures of grass
seriously.. you guys are sayign deforme din the promt right?
Or it will only be trained on women laying on grass
what is it with grass then
no its just the lyin part that causes this
Try it yourself
this one works
Someone try "a woman standing against a wall made of grass". I'm in bed rn
I've tried that one before
how cascade vs sd3 feels
So, now things have calmed down a bit⦠just how bad is it?
look at the hands , i cant believe that we are back to this garnage again.
Crop it. GG
running it now
its very good
Kekw
Its probably because squating is a fairly fixed pose, where as 'lying' is a million different poses
They forgot to cencor squatting, someone is about to get fired.
wait a sec it did 5 fingers a bit correct
there u can bang shrek instead
SDXL
my foot hurts
Worse grass, SD3 is clearly superior
Just not mown for a while
Ofc they knew it was this bad, they are trolling the whole communtiy with this lol
What about an alien lying in the grass
Yeah it seems to struggle with sitting on the ground but not squatting, good to know
This is just to make 8B look awesome
It seems clearly better at art than humans
who wants to pay for 8b after this..
Humans don't work well, but what about aliens?
i like how this channel has turned into the Failed-diffusions channel
we may have found the loop-hole with our fine tune it's all about perspective why lay when you can stand on your back
do you guys remember emad claiming this will be the last model ever needed? that didn't age well
Arguably better
oh no
as far as laying down this is not the model to end all models for sure
for a friend whose pseudonyme is "PGAZ graves", the word PGAZ made of graves.
A cinematic photo of a cemetery with graves. The word "PGAZ" is made of black marble stones, in front of the graves. There is an eerie feeling, mist all around. In the night sky, a red moon shines a warm light on dark clouds. Dead trees above the graves. Horror movie feeling.
HIS HEAD LOL
Here we go π€£π€£π€£
Ez
they just rly didnt trainwoman laying in bed or grass to avoid people making sexy time stuff
Possibly because they peaked with SDXL?
can sd3 generate sausage?
its ok aliens work perfectly
They have the 8b model , but its not for most of us anyway