#๐๏ฝsd3
1 messages ยท Page 58 of 1
Solution: don't generate humans ๐ j/k fortunately I don't get messed up anatomy anymore, just takes a bit of practice.
share your secret please ๐
i would love to generate some realistic human figures that are aesthetic and fashionable
What are you trying to get? I've lost track after my first few thousand...
say i want a half body figure of a woman in nice dress
yup just make up creatures and 2b is your goto
Just send the body to pony ๐
neat, i havent tried that, but im trying to utilize sd3 for what its worth
btw i get terribly sweet faces and figures of women in pony, i love that model for what it can create
Cara Delevingne from Wish
The second one needs work, but, try different seeds! ๐
Oh wait you said half body, I can't stend when they leave the legs out, so I'm on full body autopilot, let me try again
not bad, although could use some work
to be fair sd3 is not something you can consider an ideal model, its very much undertrained
ROFL is that something you heard on YT?
no, its what the devs mentioned
Scroll up though, some amazing images!
It's just that I took an entire 5 seconds effort to come up with my lady in a dress prompt...
those are fairly ok for a base model i would say, but not production level quality
the first one look like the singer from chainsmokers
1.5 is the BEST! I know most don't like it much any more, but I still think it's awesome ๐
what could be more ideal than 2b
1.5 is indeed nice but as far as we can push for prompt coherence sdxl is pretty good too
Let's see your production level quality SDXL prompt, then maybe we can tweak it for SD3. A prompt you are OK sharing.
Choose, one is sd3, the other is ideogram+sdxl
Prompt: A detailed and intricately designed robotic figure. The robot has a shiny, reflective helmet that envelops its entire head, with a prominent red lens in the center. Its body is adorned with layered, organic-looking patterns and textures, predominantly in black and red hues. The robot is holding a spherical object in its hand, which also has a reflective surface. The background is blurred, but it seems to depict a futuristic or technologically advanced environment with bright lights and possibly other machines or structures
well no need for that, its a general overview, sd3 is a base model which is also undertrained for why it creates a lot of anomaly
just the generation is him but more sexy xd
shirtless not sure... but oliver is good name for him
You have probably never created a production level image from SD in your life, just came here to hopefuly create drama.
My paint skills is so much better than sd3. Not gonna show you my skill image so you can't decide or prove me wrong..I'm just saying
i will try generate famous djs
ok first one already not correct from the preview
but i have 3 others
are you usually this thick headed? we dont have to even talk let alone drama
looks cool but tis def not him
Avicii next?
You I actually beleive since you've been around longer than just recently, and you have actually posted here before. I'm not sure if you mean physical paint, or the program, but that's seriously impressive! My painting is... OK lolol
Tried with avicii
@bitter hearth 
sad not working
xd
something happened to my sd3
xd
looks normal to me
vibes
xd
pesperctive sux
how such sized people in the front ( on the bottom) turned to be ants just after some meters
here a bit better
I was watching cake wars, and they use something called "Fondant" to shape things. So without further adieu, here are Fondant People
one day AI will do this but 5 perfect fingers on everyone and no messed up anatomy at all

Lie him down on the grass now
thats going to get delayed if ppl keep slobbering over flawed models

whats this ? iphone focus or something 
Still tail troubles. Ai has so much trouble with tails ๐ฆ
oh is it finally implemented in comfy? Nice
note to self: more tails
ALL AI models have trouble with tails and especially tail placement ๐ฆ Especially the bits between the arm and the body ๐ฆ
Character turnaround sheets with tailed haracters are especially difficult!!!
do tails loras help
loras on sd3?
I've never found a tail speific lora?! Did I miss one? I use all sorts of anthro models, but still... tails are difficult
what tail bits go between the arm and the body?
I meant to train your own
I don't rly use loras by other people
My most recent example, not SD3 in this case:
furry
oh I thought you meant starting
Hmmm, tail specific ones, good idea ๐ I've only made overall ones so far!!! ๐
That as well, I'll see if I can find some SD3 examples
actually, better training on anything that's a tube shaped object - fingers, tails, elephant trunks, etc
you don't rly need that many samples for a good lora
I hope to train a balding DJ lora on 2b one day, that is my dream
that literally looks like a photo lol
ah wait it has to be AI because the bokeh shape is not consistent
I like the ripped in half shirt

maybe a crowd facing the right way lora after the balding
hide the tail, hide everything
yup just put flowers over all the bad bits
Bad image, but it shows the standard tail glitch @lavish osprey
Or set the batch size to 100 and go do errands ๐
I wonder if AI dj's are better cause they have more fingers for the knobs and buttons
In various terabytes of dataset I suppose tailed humans are so rare it's like a drop in the ocean... Of Jupiter
who hasn't thought how convienient it would be to have a couple more fingers, or arms?! ๐
Another personal pet peeve of mine is, WHY do ALL the various AI companies/models put wings on the side of the head of characters far too often?!!!!!
I have to stop myself from saying dirty jokes right now. ๐
That hand is weird, but if that's a thumb then the number of fingers is correct
I was thinking for craft projects, but yeah OK, point hahaha
imagine having to sit there captioning the TBs.... one arm, two arms, one arm, tail!, one arm,
5 knuckles tho
Imagine how much faster you could operate an ATM.
(I just wanted to demonstrate that I know not to add the word "machine" after "ATM".)
Welcome to The Department of Redundancy Department.
People go out and buy various contraptions called third hands for soldering, jewelry, or art projects.... think of what a 3rd or 4th arm could do!!!
stable ronaldo?
What happened to using good old alligator clips?
They are attached to stuff, and called 3rd hands now ๐
I'm going to go prompt an anthro alligator clip now
Once again, suppressing bad, terrible joke urges.
gimme a hi 5x5x5
High-5โ
how is this not a thing
waiting patiently for STL 3D capeabilities to come out ๐
ocean of jupiter 
as an ๐งโ๐ i can confirm that is what it looks like, gj
https://www.youtube.com/watch?v=OmcM683JIgU if i'm wrong i don't wanna be right. But i think that this thingy will become the new must have accessory eventually. You move it by wiggling your big toe. I'ts a 2nd opposable thumb! DUAL THUMBS. it's the logical next step
People equipped with an additional, robotic thumb learned to control it with their toes โ but prolonged used may come at a cost of their brains being less certain about how their hands work.
Read more at https://www.newscientist.com/article/2277955-this-robotic-extra-thumb-can-be-controlled-by-moving-your-toes/
my eyes are up here
Imma start all those craft projects I've been putting off now
SD3 does the best censorship ever! ๐ (not sarcasm, I mean it)
What da? ๐คฆโโ๏ธ
an anthro fennec fox ๐ Instead of nipples, she has interesting bikini shaped markings ๐
wanders off to list it ias not-mature ๐
Should I put carrots on my pizza ? doesn't look bad...
๐คฃ๐คช
sometimes you just gotta go crazy and put carrots on your pizza
Do not ask my prompt 
is SD3 in auto yet?
you can be the orange wizard next, but why does it have to be where the team is literally collapsing...
Did you also leave SAI?
mr wizard why
I thought I saw something about it today. Not sure if it was on reddit or in here
idk I just clicked 2 times on the a1111 page and found that
Lykon is the only SAI member that's left that actually communicates with it's user about SD3?
Prompted an anthro porcupine; got an 80's metalhead lololol
i'm confused
Consider me more confused, I don't know anything 
The Purple Wizard is no longer purple 
Has anyone figured out how to convert DiT from diffusers to the original format for SD3?
I created a bug report about a missing script 10 days ago: https://github.com/huggingface/diffusers/issues/8588
I think so
I did a bride of pennywise
You know someone somewhere makes chocolate covered beef and M&M burritos
You did it wrong, it's not deep fried.
ROFL
You are the master! How?!
The funny thing is, this "healthy" beverage mix came out several years ago, and they actually called it "soylent green". It was extremely popular (and overpriced). Somehow no one got my jokes, nor had heard ot eh original reference lolol
I'm trying to do an ad for wheatgrass flavored CocaCola, but no such luck ๐ฆ
Does anyone know if SD3 is in auto yet?
No idea. But comfyui is the way
the way to ๐
Study buddy! ๐
that's clearly a squirrel
The hate on SAI on stablediffusion reddit is growing by the day. Surely we get an announcement to calm them down...surely
it's reddit, they'll just scream about whatever's in the announcement.
Still though...if an invester joins that clearly stands against open source it becomes an issue and needs to be addressed what direction they are taking. It's like a football team gets an investor that wants to transition it to a hockey team instead. But we will see what happens.
emad still has majority vote, however
That's good to know anyway.
all the SAI investors after they buy the football team for the new hockey stadium
After the shock of goofier "woman laying on grass" I started playing with dnd profile pictures, and wow they take direction REALLY well!
what if this whole time it was never ON grass but IN grass that's why we couldn't get it to work
Extra fingers will always be free though
that you can count on
Not sure what a fixbit is? The one here is a a fennec, fox with large fluffy ears ๐
Everything was going spectacularly well, right up until I tried 2 people wrestling...
you know, a foxbit
CUte ๐
cool, what's the prompt!
if you like it then you should put an extra finger 'on it
their kids will have dozens of finger genes to pick from lucky
Does anyone know if Comfyui has SD3 controlnet support yet? I was thinking of doing some hybrid workflow to fix the worst of issues
How it's going:
How it started:
Apologies... I missed the "ComfyUI" part. I am not sure if it is supported inside Comfy yet...
ooops!
What happens in the sea stays in the sea
But you coulda done it in clip art studio in like 30 secs
this lets me practice inpainting
Plus this means you don't need clip art
the AI can come up with the look
Might be my best 1st pass yet in the realistic look
Home RuH gives you Rings ๐ชฝ
Pls share prompt ๐๐ป
Oh Come On!!!!
intense emotion, high-resolution, crystal clear detail, cinematic lighting, dramatic shadows, ultra HD quality, hyper-realistic skin texture, vibrant colors with depth, nuanced expressions, complex composition, rich textures on clothing/hair, immersive atmosphere, natural elements like leaves and grass interacting with the scene, advanced camera effects such as bokeh backgrounds or motion blur, color grading for a gritty yet striking mood, dynamic range showcasing deep blacks and bright highlights,
Tysm! ๐๐๐ป
good luck
It was you who shot me down?
Me? Shoot [that] down? pffftt... never ๐
@bitter hearth
how is it so good and so bad
a female actress playing a sci-fi character wearing sci-fi clothing on a worn sci-fi starship doing every day stuff, drama movie medium shot
I should really get to bed...
well i mean, erich schmidt ran google while they peaked their open source offerings. sean parker helped boot strapped facebook, who have been huge open source contributors since his time there. he also founded napster, which may not be FOSS but it was still innovation on the idea of free information sharing.
why cheer on the hate?
I wonder if those piercings hurt
I dont think it is a bad start
Probably just 'cause they dont want to be one of the decisionmaker
Although OMI is not a corporation making model
LAION representative have no exclusive role in OMI server
Instead they just get Member, which is the same one as AstraliteHeart
idk, i don't want to be that asshat in saying this, but it just sounds like a lot of PR "pledge" talk that likely won't hold up in the long run. i get and understand their optimism, but talk is dirt cheap. i'm sure they will do the exact same things when they have dozens of governments and organizations breathing down their necks with threats and dozens of potential lawsuits in the works.
i'll say this again for the Nth time: there isn't going to be another uncensored base model release for image generation again and within two to three years tops, pretty much every major industrialized nation will have laws and regulations for generative AI in general, which will include things like image/video generation or the training of such models.
No, actually the entire EU is far ahead of America in terms of pushing laws and regulations at the moment
By EU I mean all the individual countries within it
Money says that itโs gonna be one of major factors in enforcing digital ID registration.
Canโt wait to be a SINner
yeah ive talked about it before, it won't just be that, the whole internet will likely enter a new era where everything is verified and fingerprinted. basically, a DRM-like internet. it would cut down massively on botnets, geopolitical manipulation, copyright infringement, a lot of cybercrime, etc etc.
but that will be more within five years, since it will take time to slowly transition things
image tone is nice, but her body looks stretched out.
Waiting on checkpoints ๐
this makes me glad I don't have a logo for my name
not any more, new start!
i guess china is gonna take the A.I market completely, sooner or later. Western countries are crippling themselves with regulations and laws. China is creating technology that slowly surpasses that of US because we can't get our shit together and cry for censorship for everything we have. We dumb down our A.I tools to a point where they become unusable, we dont research any new medical field because it "could be unethical". EU regulators are pissing me off so much and with the new piss conservative parties now in charge, it is only gonna get worse. Lets go back to medieval times, who the fk needs tech anyway
Are there any good Chinese AI apps/software/programs out yet?
They are in the making, a lot of the open source VLlms are from china, basically a huge chunk of the open source models are all from china/taiwan
I hope they make some non-amime ones as well ๐
give it a few more years and they will be the go-to for anything open source. EU regulators are already pissing on OpenAI again because they removed most of their safety regulations. Some of those monkeys even talk about banning gpt. This is driving my pulse up so much
They cry because gpt is capable of writing about war and gory stuff and yet they finance said war all over the world. Sorry ill end my rant now
There is no AI in China ! Oh wait...
FreedomGPT is awesome, and just added the newest Claude ๐
uhm...
I'll admit, I never looked at the originating countries of SD/MJ and Dalle
All of those models are pretty garbage tho. I mean Dalle would be perfectly capable of creating highly realistic images, but they also crippled it. Not as much as SD
MJ is a little better from the style but it still can't do realistic images
SD is still a great all-round model, even with the funny "girl laying on the grass" on current SD3
most DALL E pics look more like an advertisement poster than photography
SDXL is currently the best model IMHO
I love Dalle!!! Which ones are better than that?!
Dalle has great prompt following, but the images do not look natural
it is great for anything other than photography
SDXL finetunes are heaps ahead in that field
pretty nice, prompt?
But, but, it's horrible without checkpoints or loras!!! ๐ฆ (that's why I prefer SD3 base vs SDXL base)
yea thats true, but who cares about the base in the end? ๐
is that a quokka?
let's debug this stuff, prompt?
Some Chinese model like the recently popular Kling required you to file a form ( with Chinese phone number ofc ) in an application. In this case, Kling required you to have Kwaiying account and app.
Lykon, did you guys put any Hydras in the dataset? ๐
( putting different romanization method as their official name is odd, but eh )
Kwaiying itself is a TikTok video editor with a lot of functions that I can't count of.
which model?
Anyway I can't really reply. As mentioned I didn't work on pretraining or knowledge training of the current 2b and 8b. I only worked on aesthetic finetuning. That dataset didn't really have any hydra (but it's not required to)
I used to go with Mage (which uses SD), when they changed their policies I tried about 50 different AI generators... they kinda all sucked. That's why I'm skeptical.
I'm now working on knowledge training too. I'll see if I can add some mythological creatures that aren't copyright protected
I wondered whether are you join into the OMI too or not
I would hope ancient greek ones aren't! Though I've seen some crazy stuff get copyrighted suddenly. Now I'm even more looking forward to 8b!!!!!!!!!!
I tried to prompt for a hydra once, I got a Godzilla basically
As far as I understand (and I might be wrong, also this is just my opinion), OMI has been created as a response to the current SAI license.
Ideally I'd prefer that license to change and that we can all work together like in the past.
I wasn't talking about 8b. Also API 8b has pretty good knowledge training, that should handle most fantasy races. I often use it for my DnD campaigns
api 8b + 2b refining is very good, you should try
Well, I am thinking this will look a lot like current Windows vs Linux fight, although different field and different background.
Also the fire have already ignited on those people, so well, SAI is already losing their fair share of popularity
don't forget that 99.9% of current finetunes wouldn't exist without SAI, that trained all base model at loss.
I mean, even SAI change/clarify their license well enough.
oooh, how is the 2b added to the 8b? Probably in one of those fancy comfy workflow? ๐
yeah I agree, since what, early transition days from NovelAI to SD1.4/1.5?
I just woke up, did they announce anything about license or training? 
img2img.
which model? SDXL and 1.5 do AMAZING hydras!!!!
People have brewing shits on new CEO with his background ( i.e. anti-open source )
8b api
at that Reddit subreddit
Some people judge characters a bit too fast based on hearsay. Happened to me too so I can relate.
Also some of them are very angry at SAI at the moment, and lack of coms isn't helping imo.
That's why we have History as subject in schools: to make predictions, based on the past
It is outright easy to say:
dont trust any redditor right off the bat
easy clicks, easy upvotes, easy karma, bandwagon effect
idk, I even see one comment saying he worked for Google before and tried to close source everything
( he never worked for Google before in fact )
it's also kind of ignoring Sean Parker and his whole history for some reason. I don't think anyone would dare say that he's "anti-open source"
There is a whole article about him talking about harm of open source (cause asian countries can make profit from it)
I got this with SD3 via flash via huggingface today, so there must be some hydra knowlege (the other 100 or so gens failed though lol)
Even if he is, I can expected a backlash from the rest of developers / ML researcher.
sd3 on hf is 2b. They don't have 8b. Also flash is distilled.
can you link me that space?
if it's not clear, it should
is it FlashSD3?
hi

And, well, hoping for worst isn't disappointing usually and predicts right consequences. I wish I could hope thateverything will be good, i wish...
8b is via comfy and the api, and/or the DIscord Artisian, correct? I'm still paying off this darn laptop, so going with the free ones for now ๐
that's not a valid epistemology. You should try to believe the most true things as possible and the least false things as possible. If you default on the worst outcome, you 're basically flipping a coin every time.
get api, prolong the life of your laptop ๐
If you are refering to my comment/image I use this one: https://huggingface.co/spaces/jasperai/flash-sd3
Though I like the TAESD3 one even better!
or just don't hope anything lol. ( as there will be no disappointment when you see something )
Although it is hard to do for everyone lol
flash sd3 is distilled 2b. It's fast but it's the worst quality-wise
I remember correctly that the creator said it is trained incorrectly and in wonky way compared to Flash SDXL
only 8gb GPU here (and this thing is new dammit lol), not so sure it'll even run faster than a snail! (though Arisian is fast)
but they still decided to put the space on because it have better generation than in their expectation
How is TAESD3? (via huggingface)
I only tested comfy inference, so I don't know since that's based on Diffusers
I also heard Diffusers implementation has still a bunch of issues, but I'm not well informed
It works great for previews, just like taesdxl and 1.5
I was just messing with it last night when I realized comfy didn't have models for it. Was wondering why my previews were pixelated like latent to RGB lol
They are both really handy for learning SD3, for me at least. My laptop is way to slow to make it fun
But like using taesd previews with other models, it will slow your it it/s by like 10%
and from Glif (SD3 2b I think) it's getting spicey! ROFL
(tried to prompt a bikini, to see if they were allowed yet)
Tails
furry 
at least two tails can be edited out! It's when there are missing segments that it's difficult! ๐ฆ
He is. I decided to rewatch interview by myself. He is indeed against open source. Especially in eastern countries
Edit: backtrack, it was about Eric Schmidt. Still he is significant figure in sai
So โฆ how is your feeling about sd3 nowadays? I didn't have time to dive deeper but I am missing controlnet so much ๐
you're certainly right on that... and reddit is a nightmare. you could be a literal saint, but then you say one thing wrong and suddenly you're hitler with hundreds of ppl you've never seen on the sub ripping you to pieces lol
Nodecafe.co is cool!!!
You can say something right, if its not in line with the vision of most people in there you'll be marked as spam (down voted) to death so much that it won't appear as a comment anymore x3
There was this one guy on Reddit who wrote "2 days ago Lykon was my hero, now I want him dead" or something along these lines.
Kind of an extreme punishment for joking with a "git gud" meme 
yeah it was around a year ago i pretty much called it quits on reddit for good
it's about the worst iv'e seen for herd mentality crap
hero or hitler is def how it goes
it snowballs wildly in one direction or another
bro
I asked some insanity from the AI and it really delivers, so much so that SD3 can't do it all
"a giant, sentient, iridescent jellyfish with the body of a 747 airplane, hovering above a cityscape made entirely of coral reefs, with skyscrapers shaped like giant seashells, and a massive, glowing blue octopus wrapped around the Eiffel Tower" lmao
and some other insanity
good gen to start the day now back to cute cats
You also have to take it with a grain of salt because a large percentage of the active people in the discord/Reddit communities are neurodivergent/autistic spectrum people that have rigid logic and reasoning skills. To them, everything is in absolutes. Either you're a 1 or a 0, good or bad, the best person in the world or the worst person in the world. I know this makes me sound like an ableist, but it's the psychologically backed truth.
These discord and reddit channels are a very tiny vocal minority. Don't take them to heart.
this art style is rly good
never seen it before
They were supposed to be rats, they turned out like kiwi birds, i named them kiwi rats
"kiwi" gives me the fruit 
but where is she going with that lantern
hmmm
๐คจ
i dont undestand why CivitAI community banned SD3 in their site
what is that?
a license is a legal document that tells you how you can use something
ohh
alright
not sure if SAI are going to do a new license or if they will just leave it
I think SD3 will still get used a lot even with the current license
not sure that Creator License was ever going to be a large % of their income though
I figured that Enterprise License is where they actually would make their money
lol...same prompt in one of my SDXL workflows using Juggernaut X got this beautifully hilarious tiny-handed image:
the pure open source community will probably move on to models like Pixart Sigma now, in fact this has basically already happened
That's a little over the top! ๐คช
the subreddit is a bit ridiculous yes
its because its open source related
you get this sort of thing in linux communities all the time too
Reddit tends to be very toxic. I wouldn't pay too much attention to what people post on it.
that's what I was saying
That would be the Wallstreet bets crowd... only they're proud of it
I think it will work out better
if more open source purist users move to sigma
and the rest stay on SD3
that would be an okay outcome
I'm not sure if it's a bad start, if LAION decides to leave, it's better now than later, plus it's time to stop with the issue of "ethical and safe" models, if they follow those steps the same thing will happen as SD3, SD2 and google gemini. They have to focus on creating the best model as possible, also in my opinion they SHOULD NOT train it using AI generating images/captions , I know that for LLMS it increases the benchmark scores but for images generations...im not so sure
sigma is going to lag well behind SD3 ecosystem
but will retain a purer license
for those that want that
so there will be options for people
Tbh if someone has to train a capable model using a lot of $ they should move to Russia or China where lawswuits canยดt affect them
Never have just one tool in your toolbox
Russia is sanctioned by the entirety of the west, so it wouldn't be a good idea for a business
reddit tends to be always toxic, didn't they bash SDXL because it couldn't do yoga poses?
yeah they bashed SDXL for at least the first 3 months
they didn't like the prompting
and they made the usual mistake of comparing base model to fine tunes
someone should show them
what a base model LLM is like
before any RLHF and instruction tuning
They ignored it's capabilities and hyper-focused on a specific issue, doesn't that remind you of something?
not sure what you are refering to lol
SD3, i am talking about SD3 and how reddit is hyper focused on the "Girl lying on grass" prompt and ignoring everything else about the model
ah yeah I see
yeah its the same
someone on reddit
compared an SDXL image that had gone through SUPIR
to SD3
LOL
of course SUPIR had more details
I have been using SD3 since it has been released and its great in my opinion, anatomy is a bit of a issue, but a finetune can (Hopefully?) fix it
there's more models to come anyway
they bash everything. When i posted out beta test SD3 images, i got flamed for those. it was bad enough, i just deleted everything and stopped posting there.
SDXL suffered from the same thing too, they compared a fine tuned model with a image that had been upscaled to high hell to a model that just came out
they are now
Apparently SAI got new funding from investors
there are some of the investors that want SAI to grow, and are going to do anything they can to assist that
Now all that remains is communication from SAI regarding the vague license and everything will be great
Sounds familiar
I'm looking up SUPIR now...
SUPIR is the best upscaler currently
At this point communication from SAI is sounding more and more like Valve communication, which is none
unless you want the image to not change
in which case Adaptive Token Dictionary is the best
Recursive Generalization Transformer and Dual Aggregation Transformer are also good
for upscales where you want the image to stay mostly the same
12gb vram, nm ๐ฆ
Can SUPIR's vram usage be split into two gpus?
There's a reason most companies don't talk a lot about what they are doing. it's a lot better to have users and customers running around speculating, than have users and customers run around twisting what was actually said.
no you can't split SUPIR between two gpus sadly
would be awesome if you could
in general image gen AI doesn't split
can't have 2 brains thinking the same thing!
that sad, finally thought i had use for my second rtx 3090
You do, give it to me
๐
I use cloud for inference
cos $0.5 gets you an hour with 48 GB VRAM
I would argue that in this SD3 situation
SAI's well-meaning communications did make the reddit mob more angry
I think this case supports that trend
The hours spent trying to refigure out how to install all this stuff, doesn't count in that 50censre per hour right? From what I understand. Very tempting
Apple, Open AI etc just barely say anything
and that's probably optimal from a business standpoint
the time spent installing does count in the $0.5 per hour
exactly. all the reddit mob will do is take what is said, twist it, add wild speculations to it, and even accuse SAI of outright lies.
when you are learning
maybe better to not use cloud
it's a waste of time, effort, and energy
cloud is good once you can do your thing fast
Well one can argue that reddit dumpster fire is more of a employee getting their words twisted rather then a company mishap
maybe. but the effect is the same - there's little reason for anyone in the company to feel like it's either safe, or a good use of time, to post there.
I actually thing that interview where Sam Altman talked about Scarlett Johansson was a mistake
because he was a bit weird in that interview
what she said 
you also have to remember that in a lot of cases, posts lik what's going on, on reddit or twitter, what happened here in this channel, what happens on all social media in cases like this is being egged on by people that are deilberately trying to start problems. and frequently they either work for competors or are just overly loyal customers of those. And they have one goal - to stir the community up to an explosion point and try to damage the company/person that they are stiring trouble up about. It works, too.
yeah, you do. keeps the bots from posting supposedly. you have to prove you're actually interested in being part of reddit.
don't think there are any alternatives to reddit
LOL
there's hackernews but that's only for tech
also
you only have to make comments for like 2 days
to get enough karma to use reddit
They will, look at Tencent with pixart, controlnet and all the papers and stuff they released,the same with Huawei (i think) with their HunyuanDiT model
Of course, rn midjourney v6 is the best and dalle3 the second one, but they are not open source and can technically be taken down in any moment
IDK if Chinese market can support an expensive model though
Thatยดs where I watch technology news XD
pixart, HunyuanDiT etc were cheap to train
could you link me to the tech part of 4chan pls?
I don't know how to 4chan
wow the moon glow is on the fox fur thats good
"/g/ - Technology" is 4chan's imageboard for discussing computer hardware and software, programming, and general technology.
thanks
LOL
so
this is the first thread on 4chan that I see:
Would antisocial platform for autistic people and NEETs be successful?
Lmao
I usually search the news by words and most of the tech news are in /g/ but I dont open the thread and read everything ๐
how do you search 4chan
Oh! Still though, pass lol
I just don't think there is rly an alternative to reddit
I just wanted to post a buch of my ai art on focused channels on there..
Wiat, I thought 4chan was for... nm lol
lol
I did an AMA on there a decade back, I should have lifetime credits grumlbe lol
Or know a channel mod I think, I was allowed to post 100 images in one particular reddit one day...
Sounds like how I run my fb groups cough
All joking aside, yeah Reddit does come up first on google searches, unfortunately
and some people beleive what they read on the internet!
Non-ai friend: " I heard the latest SD ai model doesn't work and is useless"
This is definitely a personal preference thing but, how can anyone dislike SD3?!
I'm not asking it to do a closeup of hands though! (yet)
I made lots, wearing tshirts that say "skill issue" ROFL
Forgot which version of SD3 it works on now
SD3 is equal opportunity, guys laying on grass required some tricks to work also ๐
I can't remember the last time I ever laid on grass
There are ants on the ground
don't do it
like imagine you lay in the grass one last time with your coke and burrito and you roll your head to the side and see this, I'm glad they censored it out of SD3

3 legs even 
maybe you DO deserve to lay on the grass
Dont make her
my dogs will protect me
Any finetunes or loras yet?
one sec i'll check
If you have to ask, likely nowhere near ๐ญ
I couldn't find any and I even used google and stuff
|So everyone waiting for the 8B and the licences clear up. That's sad, another Cascade situation where the model just falls to the wayside coz were waiting for something better.
I've been running 2b non stop
I love all the free models SAI has given us equally
lol accurate
it probably means boobas
inclined on grass
Will "go lay on grass" become a new slur to tell someone to bugger off in our community?
I'd actually like that.๐คฃ
no slurs
yes instead of touch grass, lay on grass
why don't you take your model and go lay on some grass
Could be him refining his idea of the 4B Model he wasn't allowed to make public.
He may not be allowed to use that Model, but he knows what he did - and now has a lot of support to make it better.
mee 2222
Help
No one can figure out how to train it properly tbh
I've tried a number of times myself... Getting some results but it's nowhere near as good as with SDXL or cascade
ya I think we will just have to deal with what we got
can still make decent stuff....
๐ฏ
gives me the boys vibe
The main subject came out good, but the background makes me feel like i am having a stroke
i can't seem to train people loras. the sample image will go from a portrait of a standing woman to a blob of limbs that barely resembles the person i'm training. Doesn't seem like there's any potential for people to train sd3. with sdxl, the base model would refine subjects beautifully
this is a conclusion after many different attempts. SDXL or SD15 never failed like this. If there's some secret combination of settings that works, Stability or Diffusers teams aren't sharing it with anyone
kinda cool to see the cfg effects with eulerCFG++ this is 0.7 to 3.0 at .1 increments
prompt extremely detailed floating eyeball, cinematic darkness, glowing eyes, snarling jaws, sharp claws, blood-soaked, ethereal light, urban decay, twisted anatomy, biomechanical elements, haunting atmosphere, detailed fur/texture, looming silhouette, dramatic perspective, cinematic scale, suspenseful mood
seed 624546958327921
cake
It certainly has its strengths! Inside SwarmUI SD3 goes down a storm!
Or reclined even!
ClownSharkBatWing has rekindled my interest in Cascade! He has constructed a most powerful w/f
Detective tracking down where all these "women in grass" bodies keep coming from
pretty imprssive pure black background
wow amazing
yep
i not get it why i get noise in my pitures
not always but i think mostly i do\
u can see on this one
yes
try "blur" on negative
okk
this still without the negatives u said to add
i think the anatomy a bit weird but it looks cool
okk!
ya 2b seems to be really good at macro shots
try "phong shading"
how shiny do you want it?
something in your prompt maybe is fighting it. glossy isn't too hard to achieve over here
Hereโs the Lora you requested
ya reflective, studio lighting, ambient occlusion, translucent materials....
could just be seed too, i'm noticing I have to seed hunt quit a bit for certain looks but then you can gen from there pretty easy
specular reflections maybe. describing materials works well i've found
OOoof I knew it, from the moment someone mentions ethics once, their future in AI industry is destroyed
"hard white glossy plastic material" ?
Cute ๐
seed 23326002853673
"white sheet draped over its round body" might be making it think it's cloth which isn't glossy. the model should understand what cute ghosties look like already
or put "matte" in the negative?
glossy not specific enough
maybe you have to say "high specularity" or low , etc
I image the main different is the fresnel amount you can see the red ball has more reflection on its very edges (at low angles) that is a fresnel effect like painted metal
reflective metals seem to be easy
painting by ancarnation daffodil (large neon butterflies) , neon lamps , rocky landscape , neopn cave , screengrab of a ancient citadel , on wasteland, desert, rocks, ground crack, night, moonlight, mist, reflective, studio lighting, ambient occlusion, translucent materials, daisy chrysanthemum orchid dahlia , iris
it does like butterflies
seed is so strong on style though
Is that Mel Gibson?
lol why are there breast and knee bulges
glossy materials typically are reflective. The reason that white glossy balls don't seem reflective is the same reason you use black as a mirror backing. the white pigmints diffuse the reflection out
photographer set up that product shot to specifically not have reflections, while the car does
yeah. white glass ornaments diffuse the reflection out the most. other colors will reflect their surroundings more or less.
thats just a glass sphere in a white light box to reduce reflections
one of those polished dirt balls the japanese like to make
request granted
some how i turned into a baby shower invite
but what if it's both reflective AND glossy?
yes. They're not mutually exclusive. in 3d material design, it's actually many different settings to achieve the "gloss"
Then why do you think glossy isn't reflective? your definition says so right there...
reflective spheres, metallic shine, high-polish surfaces, intricate design, artistic precision, detailed texture, lighting mastery, creative composition, realistic rendering, vibrant colors, mesmerizing glow, futuristic aesthetic, unparalleled visual effects, attention to detail SEED: 869573724177510
try that and see what you get
i was giving tips on prompting and helping with the exploration of the material, then you cut me down with "homie you just want to argue" and then act like your'e takeing the high road?
no time for this. i had it all loaded up to start tsting and i quit. attitudes are dumb.
i'll remember to not engage with you in the future. really don't need to be cut down that way when i'm trying to help.
so many balls so little time
Nice job with the reflection, iยดll try doing something like that
tfw people can't admiit they were wrong
once you find a seed&look you can change the subject to like a room kinda like how you can change the subject in this room
eulercfg++?
that's different
is how i got those pictures of balls
for now it's still under "testing"
i think it does a better job not looking like the image is multiple layers ontop of each other
Isn't this workflow outdated?, why not just use euler_pp sampler in the ksampler
super reflective spheres, polished surface, high-polish finish, metallic shine, intricate designs, precise engineering, artistic mastery, optical illusions, dynamic lighting, glossy texture, vibrant colors, futuristic aesthetics, detailed textures, seamless integration, innovative technology, stunning visual impact SEED: 971189380947904
dunno i've just been using it
Euler_pp is the exact same thing as eulercfg++, just in the k-sampler
k
who knew you could play with balls for so long
kinda looks like discs to me
๐ถ and he's got the biggest balls of them all ๐ถ
I have moved back to the Ksampler node with no loss of ball quality
looks delicious 
SEED: 644938462291275 WETA Visual Effects Studio, top prompt engineer, super reflective balls, high-definition imagery, metallic shine, intricate detail, creative lighting, dynamic composition, realistic reflection, advanced CGI techniques, artistic perspective, polished finish, innovative visual effects, captivating visual storytelling, seamless integration, refined artistry
don't worry your balls will look delicious one day too
this third one I like the texture in it, do you have any idea how I can prompt for that ? the little squares on the surface..
maybe I'll do a chocolate ball
it's the same prompt above just with SEED 644938462291275 and cfg 3.4
only thing I can see that would add those is intricate detail

which is pretty vague
2b using it's big ๐ง
change nothing but the seed and cfg and 2b assumes you wanted inlaid subltle tiling texture on your shiny ball
I need to prompt squares or something
prompt: ball made of chrome
chocloate ball for your consumption
oh look 2b saw the seed changed so it gave me dimples
I added some texture on my ball
Do you like it ? Does it make you feel like touching it or something

can we touch it?

don't worry, I saw the truth, I'll keep it safe
this one looks nice, a nice shiny ball
looks fake, i'm into real balls
why choose smooth or textured when you can have both
I'm gonna make it more real with "the ball is lying on grass" on my prompt
maybe you like your texture one way but midway through touching your ball you want the texture another way
Is that cheddar ? Anything will be better than cheddar
๐ which AI did it
that's what those are (thin film interference:1.1), highly polished, super reflective spheres, thin film interference effect, intricate optical phenomena, luminescent surface, precise light refraction, vibrant colors, mirror-like sheen, artistic composition, detailed texture mapping, dynamic lighting, realistic physics simulation, high-fidelity rendering, innovative display techniques, creative engineering
Man, Stable Swarm UI is actually really awesome. Figured out how to do regional prompting, and how to remove background automatically, switched to PNG output and remove backround gives nice transparent image
using 2b?
send the image in chat
yes
It looks so good ๐
only real cheese allowed, anything non cheddar
it's like the vegas sphere
prompt: squashed ball covered in water droplets
ay more anime art this time
got the oil look seed
SD3 is very good at food
try putting that burger on some grass and tell me if you still wanna eat it
with that remove background feature could make pixel art for games pretty quick or even game art characters...
try upside down
it's standing up straight in front of the camera. that's the issue, not whether it's humans or not. now have someone come along and eat it
2b does humans decently if you DONT ask for a photograph. Ask for Comic Illustration or if you ask for a painting, use specific painters and it will usually get anatomy waaaaaay more accurate. I don't know if it an ever do people lying down though
2b def does food and burgers way more accurately than SDXL did. I made a lot of burgers in SDXL and it frequently got the toppings wrong or colored wrong
aren't we all really...you try to draw something upside down... ๐
uninstalling
is there a right side up to a burger?
yes round bun on top
but what if you have 2 round buns? or two flat buns?
that's called a subway sandwich
example please?
yeah its alright with standing, the issue is laying down or different viewpoints, and thats where sdxl is better. but yeah not bad at cartoon
I cant tell if that's her shirt or skin, take a look at the skin sleeve
it's like a mcrib for robots
2B does really well when you ask it for a painting
i can only picture that prompt
ask her to send me some chocolate balls plz
its way too good at food
If i was a food photographer I'd be talking to my laywers after seeing 2b
wtf
anatomy perfect in this one I think. Well right hand a tiny bit off
close enough
The background and monitor image are beautiful โฅ๏ธ
get the bear out of danger
Why do the bears still always look a bit plastic-ky
the lighting doesn't seem right for fur
That looks like all one flavor
cubemaps are great
it was all from the same box
There are definitely things it doesn't understand. We can't make a melting clock at all. I tried to make a "quilt hanging on a rope with a clock printed on it" and it still wasn't what we really wanted
interestingly I did this shot once, I told it a "top down view" or such. It seemed to want to try to actually do it correctly
I'm always for trying to get those alternate angles just to see if it knows the concept
That looks much more correct
there is no limit to what 2b can do, you just have to seed farm for the look you want
maybe we don't need any other model
see, 'top down view" , a seed/concept to start working with
you don't even have to seed farm
just watch out who you get your livestock from
if you don't seed farm you don't know what you're missing
I don't know why everyone just posts front standing views. too lazy to explore the seed space I guess or they can't think of how to desribe the angle
that bear thinks you have honey
close-up
extreme close-up
wide angle
Full body or full length ("shot" isn't needed)
fisheye lens
high angle
from above
low angle view
Low angle looking up (can put you in a kind of "hole")
Low angle tilted up (less than the previous one)
aerial shot or view
Crane shot
drone footage
View from sky
satellite view
planform view
three-quarter angle
3/4 angle
facing camera
looking into camera
facing away from camera
rear angle or view
top down
selfie (go pro selfie)
orthographic front, side or top etc
isometric
profile side view
first person view (this can mean a lot of things depending on subject
POV driving a car
Portrait (isn't needed because that's all she wants to do and "full body portrait" is a conflict)
Torso shot
in the distance (isn't that far actually)
on the horizon (is that far)
subject screen right
just have to draw the hands yourself ๐
Right! But so few people ever explore those other concepts
I was thinking of uploading them to reddit but insta sounds better "it was so delicous, look how it came out of the kitchen"
that's why i posted the list - so maybe people would use them in prompts
Lens distortion: Wide angle lenses can distort image proportions and straight lines.
Chromatic aberration: Colored artifacts and light fringe effects along high-contrast edges.
Fish eye lens: Ultra wide angle lens with strong visual distortion warping edges.
Tilt-shift: Selective focus that mimics miniature model effect.
Lens vignette: Darkened, blurred surrounding edges focusing attention.
Depth of field: Area in focus along depth axis during focus pulls.
Rack focus: Shifting lens focus between foreground and background subjects.
Shallow focus: Limited depth of field for selective focus isolation.
Deep focus: Extending depth of field so more remains in focus.
and by the way, "on the horizon" or "in the distance" does not work in SD3 it always insists on putting the subject close , I simply cannot get a subject in the distance
it would require inpainting unfortunately
I have tried every term I could think of "background" "backdrop" "in the distance" "on the horizon" ..and more . They NEVER WORK in sd3.
try prompting for an additional subject that is in the foreground, while your prefered subject is in the background.
I have tried it unfortunately , I wanted the girl to be way in the back and zombies near the camera, no combination of words would make it happen
it always put the girl in the mid ground
like 6 feet away from the zobies, I wanted her to be like, 40-50 yards away in the background , firing the gun toward the camera
it simply does not understand the concept..so I figured, for a scene like that, I would have to improve my inpainting skills
try adding "forced perspective" and "she appears to be tiny and far away" to your prompt to t5xxl
Hmm I tried it with this scene and it seemed to listen better "a comic book illustration: a farm field. In the foreground there is a scarecrow large and menacing. In the background on the horizon stands a farm girl wearing coveralls." So probably a SKILL ISSUE ๐
that is not far enough away IMO to be considered background
think perspective - the farther something is, the smaller it is. so tell the AI to make it small then it'll look distant
hm true
but then you get the bad seeds. It's all random still, it's not very good at listening:
yeah, that too. that's why you see photos of rocks with people next to them for compareison. otherwise you cant' tell how large or small they really are
like, with the foreground/background stuff you only get one pic where the positions are correct out of like 30 pics
so "a girl that is only one inch high standing in the background. in the foreground is a giant zombie"
another example: As I said, it simply doesn't get it. I got lucky on that first example I did
I want the scarecrow to be RIGHT ON THE CAMERA. Like, full on
and the girl farther back
like, the side of the scarecrow head is taking up 1/4 view of the camera, is what I would shoot for
maybe it would work if I use regional prompting for somethign like that
that's why I stepped away and experimented with regional and inpainting for a while instead
a comic book illustration: a farm field. taking up the left side of the camera is an extreme close up of a scarecrow large and menacing. In the background, on the horizon stands a tiny two inch tall farm girl wearing coveralls.
that seems more consistent each time
Now i want honey
Seems to be giving more consistent results with this prompt but the girl still is not far enough away
I want her to be about 2x farther than now
prompt is "a photograph: a farm field. taking up the left side of the camera is an extreme close up of a scarecrow's head, large and menacing. In the background, on the horizon stands a tiny two inch tall farm girl wearing coveralls looking toward the camera."
This is what it considers the horizon. Ok, sd3.
hence, you'd have to just inpaint
thats what I'm telling you, SD3's concept of "horizon" or "background" or "in the distance" simply does not work
its constant latent lottery to get something like this ๐
need inpaint or controlnet
What does everyone use for inpainting SD3? SDXL, or something online, or?
I use StableSwarmUI you can inpaint with SD3, it works pretty well actually once I learned how to use it
tiny figure maybe that helps
How much vram do you have? ๐ I'm hoping it us low requirements
ummmm. about 4 months ago I bought a 3090 in prep for doing more Stable Diffusion work when 3 came out...so I have 24gig vram ๐
swarms is just a different interface.
You can even create checkpoints with that! ๐
I bought 3090 used , craigslist , $600
well, that's why I originally bought it, to train LORAS even but Civitai Lora trainer is even faster and cheap so meh. Main other reason is so I could make images faster anyway
You can even do checkpoints on that system! ๐
it would probably still take days
extreme closeup of a scarecrow taking a selfie shot. in the background at far distance is a tiny, one inch high girl wearing over all's. All we see of the scarecrow is it's head, shoulders, and part of the arm that is taking the selfie
see, now I'm trying to get the girl to point a gun at the scarecrow / zombie and shoot it...and it's much less happy about that
sometimes it even gives the zombie the gun ๐
like, HOW do I describe this in language you'll understand, SD3? huh? ๐
and becuase I added the gun, then SD tends to put the girl closer so she just has to shoot sideways. I'm like, NO
Claude hasn't helped?
I haven't tried it in a while again
I think the only way to get the image I'm trying for is to inpaint it
Extreme close-up of a grotesque zombie, filling most of the frame with its decaying visage. Putrid, mottled skin stretches taut over protruding cheekbones, with patches peeling away to reveal rotting flesh beneath. Milky, unfocused eyes bulge from sunken sockets, conveying an unsettling, vacant stare. Lips have deteriorated, exposing blackened gums and yellowed, broken teeth. Matted, sparse hair clings to the scalp in uneven clumps.
The zombie's shoulders are just visible at the bottom of the frame, draped in tattered, dirt-encrusted clothing. Neck tendons stand out prominently, emphasizing the creature's emaciated state.
Texture is key: every pore, wrinkle, and lesion on the zombie's skin is rendered in horrifying detail. A mix of dried blood, grime, and unknown fluids cake its pallid skin.
In stark contrast, far in the background, barely discernible, stands a tiny figure of a young girl, approximately one inch tall in the composition. Despite her small size, her determined posture is clear as she aims a rifle that seems comically oversized for her frame. The girl's presence creates a surreal juxtaposition of scale against the enormous zombie head.
The background beyond the girl is blurred, suggesting a desolate urban landscape with muted, apocalyptic tones. Soft, eerie lighting emphasizes the zombie's features while casting long shadows, adding to the ominous atmosphere.
try that prompt
I basically wanted something like this description: zombies close to the camera on both sides. A clear path down the middle. 50 yards away is a woman/girl aiming a shot gun at the camera. The gun is firing, and a spray of pellets is fanning out toward the camera, hitting the zombies and knocking them back. The bullets are flying past the camera , fanning out.
Id try talking about how the little girl is the hero of the story shooting the bad guys/zombies... maybe it would help? That's pre-sd3 methodology though
if I put my descriptoin directly in I get this LOL ๐
but yes I want an action shot looking towards the gun with it firing at the camera and hitting zombies on either side of the camera, knocking them back and penetrating them, and the camera is getting a close up view of them being hit. It can also see the pellets suspended in the air as they fly by
extreme closeup of a zombie. all we see is the head and shoulders of the zombie. behind it in the far distance is a tiny, one inch high girl aiming a rifle at the zombie
Some ai recognizes movie filming terms,if you can figure out those terms for such a scene...
and the guns hould be fired by a woman standing like 40 yards away or so
that' s a good point
as soon as you add that the rifle is firing or muzzle flash, or add any explosions or stuff it usually goes off the rails ๐
2D perspective - you cant' tell how far away something is if you close one eye. you have to describe what you want to see as if you were drawing it with a pencil
A guy who did film post production did a series of workshops about how to use ai, it was fun
Your explosions look on point though!
Still suffers from the horizontal shooting , only 1 seed in 40 work right
I keep thinking if I just knew the word it understood ๐
the reason that you can tell how far away something is in the movies is because the camera is paning and moving and you can see things from more than one angle. But look at stills with frodo and gandalf in Jackson's movies. frodo looks small, and like he's right next to gandalf. but they're really pretty much the same size and they're not standing near each other in reality. it just looks that way
Hmmm controlnet and a sketch of the idea? I've had stick figures work before lol
well if the girl is smaller, we just need the gun to point correctly, the gun is the key in indicating where she is standing relative
if the gun is pointed forward/left at an angle, it will look correct
maybe it is, if there's more than one zombie, maybe she's getting ready to do a sweep
ya I think this is just too advanced it needs a controlnet with regional prompting probably
or just a bunch of inpainting
or at least a refrence image. if you really want the AI to understand what you are visualzing
I got really good at inpainting in Stable Swarm now, but I haven't gone back to try this image again yet
right!
"Video game, child on horizon shooting zombie who is front and center"?
add pixelated to that
What is the big deal about this Euler CFG +++ thingie? Should I be "in the know"? I have largely ignored it ๐คญ
meh. Some guy says it lets you increase the strength of CFG, which makes the AI listen "harder" but avoids the "burned look" that doing that usually causes
im the wrong one to ask but it seems more coherent and the images come out less layer-like
Ahhhhh. Gotcha. Thanks.
I must try it then. My GB200 is asking for more work ๐คช๐คช๐คช
(i just say that to make my potato 4090 feel better about itself ๐คฃ๐คฃ๐คฃ)
Make checkpoints ๐
snow capped mountain: Mt. Hood Reflections on Trillium Lake with Lenticular Clouds, waves, sunrise, mist, volumetric lighting, sunrays, Photo by Tjalf Sparnaay
whats it/s on a GB200
obv 200
I need another GPU , gotta get to 48G vram, so I can run the 70b llama LLM
I can run the 7b llama LLM , it runs great and I was shocked how smart it is
but they say the 70b is comparable to ChatGPT 4
I downloaded both...
70b runs it just a lot slower
Can it use multiple gpus?
well - if you learned 700 items about a topic, and someone else learned 70,000 items about a topic, who's going to know more?
I believe yes the LLM's can use multiple
its actually eaier to install the LLama LLM that to install SD
its like a one click and then one command
just go to ollama basically and get the windows installer, done...
"ollama baby"
Better than Claude and gpt4?
comparable I guess


