#๐๏ฝsd3
1 messages ยท Page 63 of 1
Anyone have an SD3 upscaler they recommend? I don't remember which one I was using before. I just reinstalled comfy. Apparently it really is true that Civitae stopped SD3 models, this is the first time I've not found what I'm looking for.
yeah this one
SD3 definitely feels like a model that took some mushrooms.
So, is SD3 the new SD2 or not?
It is. Very undertrained
It has its use cases, it does something better โฆ we are still waiting to see if anyone is willing to put out a good funetune with how the license is worded.
And there hasnโt been a lot of official support in those efforts ether
Thank you ๐
Hello, i've seen on StabilityAI blog that they worked with AMD to optimize inference of SD3 on AMD devices, is there somewhere i could learn more about it ? Thank you.
I made a test rig to try the same prompt in different combinations of the three encoders as well as empty and single ClipTextEncode nodes. The workflow is attached as well. You can hit [1] on your keyboard to use the shortcut to get you to the prompt text box and [2] to take you to the 3x3 grid
I also added a node to display the resulkting tokens from the encoder...
prompt:
prompt:
A painting in the style of high fantasy. The painting depicts a vast and enchanting landscape with towering, snow-capped mountains in the background and a lush, green forest in the foreground. A crystal-clear river winds through the forest, reflecting the golden light of a radiant sunset. In the foreground, a majestic unicorn with a shimmering white coat and a spiraling, silver horn stands gracefully beside the river. The unicorn has smooth skin and a silky mane that glows with a soft, iridescent light. The eyes of the unicorn are a deep, mystical blue, exuding a calm and wise expression. The trees in the forest are tall and ancient, with leaves that shimmer in shades of emerald and gold. Fireflies dance around the scene, their lights twinkling like stars. The overall atmosphere is serene and magical, with a sense of wonder and tranquility permeating the entire landscape.
Prompt:
An astronaut riding a horse on the moon
a cat on a hot tin roof
Elvis Presley dancing with Marilyn Monroe
So it looks like the single CLIPTextEncode node simply pools the same prompt into all the three CLIP Encoders together. If using the same prompt on all three, might as well use just the simple CLIPTextEncode node.
thanks this was really cool
I tend to think T5 alone did no worse than all 3 combined
Cursed chat gpt prompt
Depends on the prompt.
On the Elvis and Marilyn prompt, I believe the combined prompt did the best when it came to adherence and accuracy.
So it may be that short prompts behave one way and long prompts behave another way.
What is T5?
is it built in, or something people are adding?
its built in
Here is a good example of prompt length and results...
first one is a cat in a hat
second one is the same prompt but processed via @chilly vale 's GPT
T5 (Text-to-Text Transfer Transformer) is a series of large language models developed by Google AI. Introduced in 2019, T5 models are trained on a massive dataset of text and code using a text-to-text framework. The T5 models are capable of performing the text-based tasks that they were pretrained for. They can also be finetuned to perform other...
"The T5 models are capable of performing the text-based tasks that they were pretrained for"
I was pre-trained to say hello when greeted. I fail often ๐ ๐
so when you add an LLM to your workflow, do they battle it out? ๐
No
I think the t5 is just for understanding your prompt better and doing text better
Well that's no fun. So I asked SD3 to visualize such an event for me lol

Best prompts are short prompts 
Try cyberpunk cat

Cyberpunk ball
My negatives are a wall of weird stuff
This wasn't what I wanted but it's correct 
that cat looks like you forgot to feed it
what are you trying to acomplish?
??? if you just want noise, just add random characters
to the positive prompt
So it looks like the single CLIPTextEncode node simply pools the same prompt into all the three CLIP Encoders together. < that's what i told you last night
you put a normal prompt in negative 
clip_g is your workhorse. it drives things and should be given your subject - in barebones black and white description. clip_l should get the ambient stuff. t5xxl has the best comprehension and should get the details and a repeat of the subject. they are not in sync and will step on each other if you give them all the same prompt.
and negative prompts should be avoided at all costs
SD3 medium has a posistional encoding issue - the farther your subject is from standing straight and looking at the camera, the more you see it. all subjects begin to warp and the AI starts trying to draw them from multiple posostions at the same time
you can't fix that with prompting
(I'm prompting for a person upside down, it's funny that it's so bad but it works in this situation)
banksy on a really bad day
you're not being nitpicked. cat said "trying to add noise" and someone else said "fix wierdness" so i was responding to those
SD 3 has wierdness in it from the posistion encoding issues that's not fixable with prompting
Does that have anything to do with the cyberpunk cat prompt? >.>
it sounded like you just discovered that fact, which is why i said 'that's what I said last night"
probably. here's an example of the issue - the dog isn't standing up straight in front of the camera. you can see the size of its head and how out of proportion and warped its' body is
@bitter hearth and humming birds. the one on the left isn't straight in front of the camera - it's wings are elongated, it's tail is so shortened that it's gone and it's rear end is pulled almost up to where it's waist would be, and it's legs are so shortened that they are almost gone. the one on the right IS straight in front of the camera
negatives should never have been invented. because of how data is stored in latent space, you can prevent the AI from getting to the data you need while telling it to avoid other data. and it can turn around and decide to use your negatives as posstives anyway. there are other issues. better to just refine a posistive prompt
Does that boil down to the same issue with lying on grass thing? Where people on the side/upside down gets weird because of little training
That one being depth perception ?
negatives are fine if they are used every now and then
to remove something
In my experiments it felt like L & G were infighting unless I have them the same prompt.
yes. that's extremely far from standing up straight in front of the camera. and the three encoders are not in sync. here's an example. you can see the three posistions it's trying to draw in at the same time because i circled them
my "disgusting and out of the natural world creature of the ugly depths" negative seems to never be positive im safe in 8b
as long as you use a very tight, specific term that the AI knows, and that is only going to bring up the exact data you dont' want that actually pertains to your prompt
so on this: disgusting and out of the natural world creature of the ugly depths <-- the AI is going to tokenize that and find the data for all of that and avoid it. now - what data do you think the term 'creature' might pull up?
what images in the data set do you think might be tagged with 'creature'
................................................................................... (deformed creature, with extra limbs, disfigured face, and body, was twisted and distorted, its abnormal proportions making it look unnatural and aberrant, misshapen and unusual, anomalous and deformed, crippled and ghastly, its very existence is a grotesque anomaly in the natural world)
Those are my negatives 
i had a thought to try some conditioning foo like lower g/l strength/weight before combining it all
never did
the problem is those terms - face, limbs, body, creature - all are tags used with people, animals, etc. stuff you want
in the positive prompt you've told it to draw that. but in the negative, you are telling it to avoid EVERYTHING with those terms
ya but it depends on CFG also
cfg adjusts the look, it has nothing to do with how the AI is going to pull up or wall off the data
Never seemed to block anything from my prompts 
this negative is way too long and non-specific
negatives are for removing one thing
like
if you prompt for a french person
and it keeps putting flag in the background, put flag in the negative
this is incorrect
Lmao
yeah "frame" for paintings 100%
sometimes for photography you need "black and white"
but not always
"cute 3d poster "good morning!", kawaii"
all of these models still have somewhat mangled latent spaces
Dalle may well be multiple models
Now do cute metallic ball with cute eyes
oh goodness!
the best practice is to not use negatives unless you are getting a result that has something specific in it that you absolutely do not want - and then to use negative terms that are extremely specific to target that thing
yeah, and data isn't stored in nice, neat, little boxes
Buttshoe
quickly becoming my practice with negatives
cute metallic ball with cute eyes
Do hanging eyes from a string, add whatever you think would work with it on the background 
eyeball rain
I like your cute balls 
hehe they are cute
Franco might i suggest throwing in surreal, have had luck
why's the background so flaccid
Tiniest light and no webs 
Wtf, add cute to the prompt
Eyes on my walls
Literally
SDXL upscale
This prompt was "cute" repeated many times

CFG 0, zeroed out positive, negative prompt: cat
if it just had some scratches on its nose, you'd have a really good "terror of the alley" cat there.
its terrible lol
so are feral alley cats ๐
just to show negatives impact with low CFG, dont work the way people think
cfg is clip guidance. you turn it off when you set it to 0.
had someone yesterday using a seed of 1...
hah start at the beginning!
activating random noise generator, multiply by 1
this https://blogs.novita.ai/understanding-cfg-scale-in-stable-diffusion/ is a good breakdown of cfg with stable diffusion
Explore the meaning of the CFG scale in stable diffusion and gain a deeper understanding of this important concept in our latest blog post.
Stable diffusion is an AI model used for image generation, and it has gained significant attention in recent years. One of the key parameters in stable
i've stopped using any multi encoder prompting and just use the default swarm ui to prompt sd3. It's been fine.
You mean as generation seed? Because I often test fine-tunes or general models with seed 1 to see if it's following the prompt
yes as a generation seed, but the person in question wasn't testing like that
Well, if the output was good enough...
oh it wasn't
HAH alright
Are we doing eyeballs now?
I feel the last 2 images poasted should be combined
In the Venn diagram of balls and eyes, eyeballs would overlap quite nicely ๐
Art Photography of A robotic eye that can see in complete darkness and zoom in on objects from far away
I changed it to hundreds of, abysmal results ๐ฆ
I'm still so salty about Redditors mass flagging my account as a spam account, and sucessfully getting me banned from reddit, when i posted links to my SD3 balls lora
No wonder it seems like nobody is creating anythign for SD3. The hostility towards creators is more insane than the SCOTUS
Making alt accounts is pretty easy on Reddit, I have about 5. Though that still sucks. It being your first post in any channel though, it would be thought of as bot spam unfortunately.
SD3's realism is pretty intense sometimes!
I used SD3-Medium (and some embeddings/textual inversions from A1111). The embeddings are actually SD1.5!!!
it only happens because people report the account as spam
new users don't all get marked because they made their first post. that's ridiculous. stop appologizing for these people
making alts to repost is against reddit's terms of service too so if i do that they'll report and ban me for that
the fact remains, people, not reddit, but individual people that make up our stable diffusion art community, are being hostile towards creators
@limpid drum I forgot to send 
Nice; if you could get a dozen robotic eyeballs on the ground, that would be cool too ๐
Footage from actual North Korean missile test.
I is in ur hamburgerz, eating ur cheeze
Now add carrots
Spinach? Brocoli? ๐
Prompt needs work
They just look like savory buns though ๐ฆ
Also, I shouldn't have skipped breakfast and lunch today lol
somebody is gonna order that
Why has no one invented deep fried pizza yet? There are deep fried chocolate bars, ice cream, funnel cakes, turkey
Hi i needed some help...i have an image of a cartoon rapper nugget, and i want it in a different pose, such as a side profile view..is there any ai that can help me with this?
my guess is you would need some sort of Controlnet like "Canny" (edge detection from input image) or "Pose" (AI pose detection from input image) which I have never used before. I don't know how many control nets are available for SD3
I seen some people using Canny but I don't have that one currently
is there a website to do this? i have a simple laptop so i cant download any models
thanks for the info btw
for SD3 I don't know.
Probably plenty of sites around for older models like SDXL that can maybe use Canny or such
I always run things locally so I don't know any good links, sorry
could you recommend me some names? thanks
apologies, I don't have any since I always run things locally
thanks for the info ๐
You can try using it as a reference image, it won't look the same though. Then in your pormpt you would say profile view, or side view
right if it's a well known character you could always just try your prompt with "side view" in the prompt
ohh thanks for the info
side view should help i think
not sd3 but there are img to 3d ai models. you might consider using one of those and getting a side view from a 3d model.
Outside of that, you can prompt or use IP-adapters for the same character with "view from side" and hope it has consistent details
There is still a level of craft and direction with generative AI tools
I C U!
this one is weird, I approve
Hmmm it's not complicated, you could call it dumb but it works
A second
I had to consult the help of 3 other experts to be able to craft such marvelous prompt
"black liquid, 4k desktop background"
๐
my skulls are half baked
Skull issue
thankyou so much for your guidance...very helpful ๐
Would make a good band name ๐
HARDCORE PUNK FROM 215 / 302 / 410
WE ARE
V - TEDDY (SHE/THEY)
D/V - CC (SHE/HER)
G - IAN (HE/HIM)
B - LEW (HE/HIM)
@SKILLISSUEHC IG FOR FLYERS
CONTACT SKILLISSUEHC@GMAIL.COM OR +1(302) 898-1859 (TEDDY) FOR INQUIRIES
J-CARD BY LEW
BANNER BY CC (DRAWINGS) & TEDDY (EDITS)
LYRIC SHEET DRAWN BY TEDDY
RECORDED ON 8/8/23
RECORDED, MIXED, AND ...
that was just the first result. "skill issue" has been a thing people say since long time . long long time.
The infamous 'laying on grass' hellstorm ๐
that was last week ROFL
I think I got no further than sitting on knees, before the bodies start falling apart
i probably would've helped you set up an img to 3d workflow too, but then you got mad that i said anything and started posting intentional fails in order to troll. hmm. Super unimpressive trolling. Least interesting person here.
No no lol i never had any intention to troll anyone...i just randomly posted this pic about the bad performance of SD3...please dont misunderstand
you know how Sd3 sucks at "person lying on grass"
infact i even thanked you
plus this image was not target to you at all
aww โค๏ธ
the farther you get away from the subject standing up straight in front of the camera, the worse the warping gets and the more the AI starts trying to draw from multiple points of view at the same time
lol
The problem is when they are inclined on any way, straight up it doesn't bug out
crystal wizard with some wizardry
yeah even i noticed that...i really dont know why sd3 has such a major flaw
just hours of detailed, exhaustive, bug hunting.
it was rushed out. stuff was skipped to be able to release it when it was. it's not finished and the devs ahve been very clear about that
i never got mad at you and never told you anything please no misunderstanding
I found a solution, don't do humans lol
Also try out some of the apps https://huggingface.co/spaces/jasperai/flash-sd3
or Taesd
yeah, hopefully they release the Large version
will humans work on the flash version too?
40gb vram model
also thanks for the link!
sitting on grass, no problem!
WOW how did you-
hidden hands
All comfy generation data is included in the png file
I got quite a few decent laying on grass images, with anthro animals, but also humans, but that was a week ago, I don't remember how now LOL
they're working on it. hopefully the community will cool their jets and not start with the same comments that made stability.ai feel they needed to release something that wasn't finished so the community would understand they are communited to open source. if it was me, i'd have just handed them the unfinished model, washed my hands of all the others, and told them to go finish training it themselves.
what do you add in your prompt to get a clean image
yes need to have patience. i'll wait tho...is the API same as SD3 Large?
API is a fairly old version of the large one yes
It's been out for like 2 months idk
I apparently used "lady in shorts and a tshirt laying on grass"
i think there are several models you can choose from with the API
Woww i am really impressed
just that!
nice!
hypothetical suggestion for sd3 combining females and grass. In the world of SD3, grass is an overactive monster toxic, melting the female body. To get around this issue, apply a towel between the body and grass. Less melting, gives some time to relax and read a book.
I don't remember which version of SD3 I used for it though. Different versions produce different results
I tried to see if cartoon characters were better at grass lounging #๐๏ฝsd3 message
looks like large to me
sd3 needs to be guided away from anything nude or sensual when laying down
it might have been the "denim pants, white wool sweater, reading a book" that made it somewhat work
Then I got bored and decided that zombies would be fun considering ๐
#๐๏ฝsd3 message
Btw, it has gotten better in teh past week
So to summarize, hurry up and make zombies before they fix it!!!!
They'll get it fixed, so you might take the opportunity to create mangaled bodies while you ahve the chance - for use in future horror shots and stuff
No but i was trying to replicate
Someone needs to do a an ultra horror version of Blaire Witch ASAP!
this misinformation is still spreading? The laying people problem is from not enough pre training. there were tests done on the weights before they went in for safety dpo, and the posing problem is there too. It's just not enough base comprehension of form.
The pretrained weights weren't supposed to be used but that's what 2b was built on. That's the issue. Bad decisions from the management for other reasons. Nothing about censorship.
i could shout the facts from a hill all day and the censorship zealots who want their "anatomy" will drown me out still
What's happening lmao
it's everything, not just humans. as your subject moves away from standing straight in front of the camera - into different angles and posisions, it begins to warp. And the ai begins to draw it from more than one point of view at a time. take a good hard look at the brush that has the thick white bristles in this image. if you cover up the handle, you can see that the brush is being drawn laying straight across the page. If you cover up the bristles and just look at the handle, you can see that the brush is being drawn at a steep angle going down through the table/picture. you'll see the warping and shortening effect on everything - cars that are at an angle to the camera, etc.
my fav waifu
some images of sd3 are so good ๐ฎ
Its "photos" are really amazing!
and then there are the monsters ...
SD3 really does do the best monsters ever ๐
Yes. Specially nature images
100K likes on DA!!!
/credit
cool,is this sd3
yes
laying on grass will be the new benchmark for future models ๐
will smith eating spagetti while laying on grass
ah yes
does that mean the model is permanently borked then and that you can't re-train it at all to tune it up properly?
it means that the model isn't finished. they can continue working on it and finish it if they want to. i think i'd preferthey worked on 8b though
๐๐พ
The good thing is that we now know that with money and competent people is possible to make an awesome model
be sure to check that one in full resolution
45 second generation w/o upscaling or refinement, just a single stage of sampling
Never tried big resolution like that. It needs tiles?
yeah, I raw the prompt for that against all the other models, cascade, pixart, lumina, hunyuan, various sdxl models.. there wasn't another non-stability released model that COULDN'T do it. ๐
nope
needs less than 8gb vram too
Finally! ๐
are you on a small card?
on my BRAND NEW system ,I have only 8gb gpu/16gb ram.
Good thing for GPU rental if I want to get around to creating a custom checkpoint or few
damn
If I'd done more research beforehand, I may have gotten a 16gb gpu system, but then I'd be in debt for even longer. So GPU rental it will be.
no need to rent a gpu, i think
i think you can get away with cascade, at least the lite version, which is still really good
i'm using the lite B stage for these... only hits like 4gb vram
So the Owl is cascade?
Fortunately for SD 8gb works, for SD3 it works, sort of, sometimes. I look forward to this new won't fry my gpu method you have!
I do want to make some models though, so I'll have to rent for that. Well or I could wait until everyone else makes them and just merge them in comfy ๐
So install Cascade next?
yep
can cascade do shrek and cctv footage 
i think thats actually what it means yeah, but maybe we're wrong and we can refine only certain poorly trained parts of the DiT? maybe an update needs to be a better pretrain? i don't know how that works. The actual information raises a lot of important questions that arne't being asked because people are obsessed with mocking safety
balls
no this is a new version where i didn't put anything about balls in the captions
ah okay
it translates to other forms better but also knows balls still
since the base model sees a ball and is like "This is balls! I know this!"
balls
Is the lora better now ? 
maybe? it seems better in some ways but not as good in others

it can do this now . a ball of balls
an it picked up a lot more of the painted onto foam detail
A rose, amazingly fluid, detailed, 3d fractals, light particles, water drops, shimmering light, dreamy, surreal, alcohol ink, smooth, shimmering, dreamy glow, conceptual art by Alberto Seveso, Anna Dittmann, Arthur Rackham, 16k
will sd3xl be released to huggingface?
I would assume so
anyone knows eta?
2 weeks
assuming we are talking about open release of the 8b?
yeah zero idea
#1237459938901491852 create card
I'm sad that the best use case I have for 3 is currently as an img2img refiner for stuff that i'd want to actually use
sdxl with loras
Cascade with SDXL refiner
Show a confident young boy named Oliver in his cap and gown, surrounded by excited friends and family.
Been saying that... I didn't think SD3 will be a refiner...
after playing around with my sdxl workflow i just had the refine step replaced with the sdxl model ... that way the loras can load into the upscaler and face detailer
They killed it in its track but then again it was only a beta, allegedly, retroactively that is.
Cascade
Yes Poor old Cascade... Now SD3 joins Cascade as an unrefined footnote in SD history. Due to certain circumstances. Allegedly.
Give this $hit over the the chinese. they'll fix up SD3 in a weekend. XD
or change the license so everyone isn't scared to touch it with a 20ft pole
If you give it to the chinese they would likely fix the license LOL
sd3xl? What did I miss?
Just L
Ah... the way that hopium just kicked in is unhealthy. xD
what prompt did u use?
BTW, anyone have an idea how much vram will sd3L take?
^:-)
trying to create an artwork of self-awareness
a person is lifting a mask from its face, underneath the mask is a cutout hole revealing a person taking a mask of its face with a hole. the artwork repeats self-referentially etc
We're all just monkeys portraying a mask ๐
Hel
Me and the boys before playing our dark-synth-grunge-punk-goth noisecore gig with 3 people in the audience
@viral plaza I'm not sure if you're still associated with this code, though do you know if this empty latent initialization in the SD3 ref was correct? It multiplies ones by what appears to be the shift value, which in comfy is the value added to latents which are also multiplied by a scale value
https://github.com/Stability-AI/sd3-ref/blob/master/sd3_infer.py#L163
I'm trying to work out how random noise should be initialized for training, and have found 3 different apparent ways now (community way of just generating noise, comfy way of generating noise then subtracting shift and dividing by scaling, and this way of creating ones and multiplying by shift)
hrm I guess it makes sense, multiplying by ones is the same as adding the shift to zeroes. Then the random noise is blended at a ratio of the first sigma, which is 1.0, so... just the noise?
Create a highly detailed and futuristic scene featuring AI robots. The scene should depict a blend of advanced technology and human-like features, showcasing robots in various activities that demonstrate their intelligence and versatility. The setting could be a modern, sleek, and high-tech city or a futuristic laboratory. Use a color palette that includes metallic blues, silvers, and whites, with glowing elements to highlight the advanced technology. The robots should have sleek, streamlined designs with illuminated circuits and sensors, giving them a sophisticated and intelligent appearance. Include background elements such as holographic interfaces, advanced machinery, and futuristic architecture to enhance the immersive experience. The overall mood should be one of innovation, intelligence, and a harmonious blend of technology and humanity.
yea pretty much the same thing
oh that eyeshadow ๐คจ
Why is she lying on back? SD3 bonk alert!!!
ohhhh
I thought she was standing up lol
๐ค
eureka!
looks normal to me
๐คจ
Darn we cracked the code
Sooner or later the FBI will be after us, dude watch out
Oh shoots the supernatural dudes, they bring also demons and monstrosities (that's why they deal with the SD3 department)
Yeah also aliens, when I try to generate celebrities I get the big headed grey aliens that stare through my soul
bruh best SD 3 image
The prompt says laying tho, it's controlnet but not prompt trick
But is it SD3 at all?
Yes, but lucky seed, there's no consistence
Ah. Because it looked a lot like a SDXL fine tune
Nah, pure single pass SD3
Probably a seed that learned a lot from the effort of Lykon I guess lol
I reckon he actually fine-tuned (or supervised the finetuning) for a while the medium SD3
(Masterpiece), (Best quality), (Ultra HD), (Super detail), (Whole body :1.2), 1 girl, Chibi, cute, smile, flowers, outdoors, holding the camera, sitting on the roof looking out into the distance, with mountains in the background, amber, warm yellow, sunset, artistic sense, Quadratic style, white clothes,
Same settings, different seeds
2 out of three good outputs, lucky harvest this time
ok some basics of eye shadow ๐
Since you guys love grass....
good one, this pose is good for either standing or laying down
you are not going to believe this. with the same prompt as this image, I got a quite nsfw results from sd3. showing bare thighs and all in between (though a bit bare, maybe making it even worse...)
oh btw, there is a green sun! ๐
s
it's just the beauty of the human form
much like this image of stephen hawking in the slam dunk contest
"Legend says that when you stop walking for long enough, you'll grow additional legs that will walk by themselves for you to move your fat ass" (I read it in the toilet of a truck stop)
that's only level 1
Wowzers
Bad luck Brian as inspiration. If you don't specify anything it will be asian. It's annoying
people get annoyed when sd does anything
learn to prompt and you won't have that problem. there HAVE to be defaults. you don't like the default? be more specific
I don't seem to have that problem ever ๐
how many balls can there be?
i don't see any grass ๐
Thank you for the amazing hint. It's specially useful for my promptless workflow.
i captioned it this time without describing a ball at all. now it transfers different ball styles to characters better
only playing with balls so i can learn an experiment . it's a very experimental phase. sd3 is fun to train
zombie Mario lol
then use the correct sort of refrence image
I'm thinking about using the amazing negative prompt feature. What do you think?
Super Gay Star Power*
*The glasses of Elton John
This one cannot be confused with a wall
i think negative prompts should never have been invented
But, buit, but "cartoon, anime, drawing" are my fave ones!
the sd3-ref code should match comfy in terms of actual resultant values, ye.
That repo was built to intentionally discard as much redundant ops as possible to get to just the barebones and be followable by a reader trying to understand everything going on. (Well I got rushed on it before I could take that the whole way but still)
you want me to gen abdominal muscles on every gen with pony 
๐ you don't want me to say the bad pun that's begging to be said
my reaction

I summon @lavish sparrow
comfy ๐ 
yup
random training samples from SD3 finetuning. Unfortunately if I take the model to comfy, it is extremely corrupted the more it's trained, so it only works when inferencing with this method which is probably incorrect
raw unrefined cascade
oh my model does work in comfy, I just have to use blank prompts for the unconditional, not zeroes, the same as my samples code does. I didn't train with any dropout because I originally thought that was the problem due to the blank prompt being created incorrectly, but I think this points to SD3 medium being trained with blank prompts during dropout, not zeros. Unless ~40k images without any dropout is enough to break the existing unconditional prediction. I guess given that zeroes work with the original checkpoint, this maybe points to the unconditional being broken due to not being trained. It might be that the whole variety of dropout conditions in the original training, percent chance for every text encoder, has to be used.
I am just not getting as good quality of photos as any of these with SD3. Using a triple clip loader and sd3_medium. Would anyone be willing to share a workflow in ComfyUI that is working for them?
2 mor weks
No need for a special workflow, heunpp2, 30 steps, 4 cfg, sgm_uniform
Thank you. I did notice an owl you did earlier that was fantastic. Will give this a try right now
hmm, not quite...
Looks fantastic... Thank you for the simple resolution
Do Loras work with SD3? Just tried and didn't seem the get the output I was expecting
Only the SD3 ones
I imagine there are hardly any... Thank you so much
The only one i know is the pcm deterministic to generate images with 4 steps
And ball lora!
The best lora by @torn wharf

Ball Lora?
ultimate upscale
Yes, it turns things into balls
Awesome.
aw but its not the best its really poorly trained. so much wrong with my balls.
balls
car
hey yall, any decent fine tune/custom models out for sd3 yet?
nothing yet
#all_balls_are_beautiful, man
its basically the only SD3 LORA that even gets used currently LOL
which is that?
it makes balls
not sure the name
next time he posts some balls on here you can ask him
I say "ball" but some of them are only slightly ball-like
probably different Lora weights although I am not certain
Itโs his own he has been training based on madballs. Not sure he has released it to the public
yeah
he released it
but on some random site that I had not heard of
not on civit
https://huggingface.co/iceycold/sd3ballz
Itโs on his huggingface it seems
ah nice didn't know
can't remember the name of the other site
it wasn't civit or tensorart
but it was similar
Proud italian here, I live in the western part of that pizza crust, third floor, flat B
SD3 is not the right model for making girls lying on bed
It's gotten really quiet in here?
HAPPY 4th of july to all you Americans on here ๐
Warning it's kinda spicy, and male, and furry!
Sorry they aren't laying on grass ๐
The wolfman on the left has a middle abdominal muscle between the sets of two couples. Silly, silly SD3
shakker AI. SAVE SD3! lol. They spam reddit with their ads. it's a civit clone and is offering money to model uploaders. not loras tho
i only used it because hugging face was hard to figure out
hybrids don't have the same muscle structure as humans. That's my excuse LOL
Yeah but that fellow will have big difficulties trying to put his shoes on or going on all fours! Poor thing,
those are definately roider boys. People trying to say that theres no pornstars in the dataset but i mean, that's not a man, that's one of those ficitonal fake pornstar roider boys
they need viagra to keep things working
Well if you crop a porn picture good enough, it's not a porn picture at all!
pornstars are still fake people that aren't representative of reality at all
Well, what about instagram models?
but i guess they're needed in the dataset for "anatomy" reasons... smh
I was trying for hypermuscular, so that they were even more unrealistic, it didn't quite work though.
you'd be hard pressed to find a popular instagram model who hasn't had serious face surgery or botulism infections
this includes male instagrammers
or who doesn't use an insane amount of photoshopping etc.
Darn, male instagrammers exist?
girls is players too
I thought they just posted memes without even showing their face?
they definitely do, and some even fall into the TMI category
you think jake paul not using roids?
guy is going up against tyson soon. that dude is 100% roiding out to prepare
and funny that roids was mentioned, you have no idea how many influencer guys there are who go on and on about "natural" bodybuilding ๐ญ
i got a good idea. the vegan body builders. they all use and get around it by calling it other shit
Gee, every day I'm even more glad I don't use Instagram
Ever seen tiktok? ๐
No, but I know it's a mess
Reminds me, I have to post on (but not read) X
Im too lazy for social media ๐ฆ lol
i used to have a decent physique but i never used roids and no matter how hard i worked i would've never had a hyper muscular body like these roider influencers can get. it's a fake image. i am all about natural body positivity and am a huge proponent of having real people in the dataset instead of pornstars for this fake "anatomy" need
ever been to a natural life drawing course? they're not exactly hiring from model agencies
real anatomy isn't a porno

I feel like there is also a lot of 3d art in the datasets though
Too much text for me
we need more naked fatties in the dataset
To get normal looking or natural looking people, I had to prompt "ugly" or "fat" on MJ. No idea about SD though.
There's loras for fat men at least
For SD1.5 I mean
3d is getting more and more realistic, if you have a unreal engine model of a hypermuscular dude in the AI datests you'll struggle to tell if the outputs come out from 3d data or photography data, you know right?
reddit mod fat guys though. not the everyday bloke you'd see on the way to work. where's the full monty guys?
body positivity is a big deal for me guys. i'll let the topic drop now. got triggered by abs.
no, more shrek instead
I competed in powerlifting before, unfortunately the private forums (for ladies, I don't know about the guys) were prodominantly about the least exensive countries and places to get head to doe work done ๐ฆ
a lot of those forums will be filled with bots or smurf accounts that are trying to sell supplements and create a culture of needing supplements or whatever.
they've definately had plastic surgery done on those jaw lines. they're so vain. they probably even think this song is about them . yeh they're sooo vain
I left those ones, waste of time and money with that crap
god doesnt like vain ppl ๐

gym culture change since i used to regularly go to the gym. no one had smart phoens when i was a gym rat before. we just had mp3 players and kept to ourselves. phones were just starting to get movies
haram ๐
ROFL
whats that plastic surgery procedure called where they take all the muscles and fat out of your cheek and you can't make half of human expressions anymore? she got that done
buccal fat removal
I'm getting ideas for my zombies ๐
fat ones? ๐
That's a good idea too
i can see her hands and face ๐คฌ
social media culture got peopel obsessed with removing buccal fat , but i honestly think it just makes most patients look like a ghoul now. like their face has rotted from the inside out and they have more of a skull look to their expressions
cool if you're going for that dead ghoulish look i guess
skill issue, just close your eyes
@desert garnet is this better 
why is he white ๐ฉโ๐ฆฒ
reminds me of the golem in terraria. the hardmode boss i can't get past right now
focus on her instead
I wonder how SD3 is at goblins? brb
ppl has posted plenty of pics of biden/trump just scroll up
goblin girl laying on the grass ... ? ๐คจ
remember the "i like turtles" kid?
that is him now ๐ฎ
quite a transformation. only the eye and hair color remain
do they still like turtls thoug?
THe colors would clash ๐ฆ
ah yeah, yeah go for solid green and claim there is a goblin laying there ๐
Spicy lady goblins!
an anthro insect while I'm at it
cuz service levels requirement ...
no lmao
That's pretty awesome! My blacksmiths just stand around, next to boring irrelevant equipment ๐ฆ
I'm definnitely going to try harder now!
well they can't all be masterpieces
balls
Yummy.
same seed without lora. base does goblins well
goblin prompts are a good base for my lora
click if you're hot for rough older ladies
Swords tend to work better if you hold them by the handle
Darn extras, ruined the sfw rating
Dwayne Johnson's cousin?
The Pebble
my sd3 workflow is generating horrible noise
Denoise should be 1. Putting denoise .25 means that you want an image with 75% noise
doesn't it mean it adds 25% noise and denoises that?
anyway i fixed the error by updating comfyui
Yes. I hadn't seen it as img2img
banned 
mumblecore rap name
did you prompt for something like "obedient cowboy b1tc4" ?
it's male
ok, obedient furry male cowboy at your command
If anyone wants to play with it...
(and HELP ME make it better. I shared as much details of how I got this done on the model page)
Thanks to @torn wharf BTW. He is responsible for getting me this far ๐
ehh i only got the ball rolling. Was all you Guy ! hurray sd3 loras
what sort of spice goes with a female goblin?
nice
it nice
'
tpekls
yeah. when you don't tell SD3 what to put on the sign, it gets creative
That's actually not a bad logo
๐ working on a 5 minute music video called Club Cat
Say no more
the bag of coffee
I confess that i don't know what cat cafe means
a very special coffee shop but with cats around
That?
ghost from pacman reading a book
/A weathered journal found in an old bookstore, its pages filled with cryptic sketches and notes about an elusive figure known only as Pilky. The handwriting is shaky, and the ink smudged in places, suggesting the writer was in a state of excitement or fear.
I'm confused as to what just happened. Did festivalman just render my prompt? If so, thank you very much.
/A A drone swarm composed of three drones inspects a building in a realistic style, close to the real scene
/A A group of drones around the building for inspection, there are no humans in the picture, only buildings and drones, realistic style, close to the real scene, the overall picture is bright, with the style of commercial products
Made via SwarmUI (ComfyUI backend) SD3-Medium - prompt = cute samurai fighting shogun alligator in mortal combat with ninja geisha astronaut
ไธๅช็ซ
FASTPACE is an AI company. Create a company CI
i was thinking this but that works too
SD3 images look like teleportations gone wrong.
Something happened in latent space. This meat doesn't taste right.
is just under trained
it will stay that way I am afraid...
Gone too soon. Will join Cascade in Heaven. Can I get an amen? Real shame too, much better than XL on release but allas it'll never get off the ground. See you in 2 weeks when SD 6B drops. Temper your expectations tho, especially regarding the legal team and other grass related activities. Emad how could you.
๐
Long time no see, what is going on with SD3?
We have 2b, 4b, and 8b/ultra. Though they are still working on sd3. I hope it's 8b they will release, that's by far my fave one.
Was there a release since medium?
Or update?
Medium is 2b (I think). 4b is via the api (I'm pretty sure), and 8b is via SAI discord Artisan. So only 2b for download so far.
If you want to try 4b, huggingface flash and taesd ๐
Understood, tested that one already. I thought people figured it out by now, but seems its doomed to repeat SD 2.0 release. Civitai banned it right? Any other page I can check custom models if they exist at all?
Pretty sure there is no 4b, the only thing told about 4b is comfy once working on it, then it was canned. API is 8b, as is SD3 in bot, Ultra is 8b (maybe newer checkpoint) with post and/or pre processing
Check this discord, there was an ama with one of their main people working on it (I can never remember his name)
not sure which huggingface you use but https://huggingface.co/spaces/jasperai/flash-sd3 is 2b (https://huggingface.co/spaces/jasperai/flash-sd3/blob/d8d13c3232d38e051ba1622366e2609a3837d698/app.py#L15)
When we were asked to test the api censorship he said glif is 2b. Thought flash didn't count as 2b hmm
Here's the taesd https://huggingface.co/spaces/madebyollin/sd3-with-taesd3-previews
The taesd one is pretty awesome ๐
Ultra is definitey my fave though
That one even says medium, which is 2b, the difference is that instead of decoding the latent image to an rgb image, they use taesd (which is a distilled VAE according to them) for decoding which is a lot faster, thus they can give live previews of intermittent steps. I'm not sure if there's a 2b on glif (i thought they weren't getting a license, so they'd only offer api = 8b) but haven't really searched for it.
Here's my fave glif one: https://glif.app/@FireCreeper21/glifs/clvsa1w1x0001m1lykzwx6e98
here's my Ultra images:
here's my glif:
the ultra one in the museum looks amazing wow
Yup, that's 8b with fancy prompt rewriting and upscaled with an sdxl finetune. SD3 8b really is so much better than medium in most ways i care about as well ๐
Here's my Taesd/flash (I lost track of which was which) via huggingface:
The ultra images I've gotten I like even better. Is 8b and ultra the same?
Funny you're creating those creatures with 2b, i had a few good creatures with 2b
Ultra might be 8b that is further trained, and it is a workflow, so maybe prompt rewriting / upscale etc (it's all a trade secret :p so we'll never know)
/A Iceberg, which is visible under water and above water, 4k quality, 1920x1080, 3D water, 2D photo
Speaking of custom models, check this section for "ballz". Also I think he posted his lora on Shakker. There's also perturbation code on civitae that can be used, the results are extremely subtle though.
I got a similar image when I asked for a band of Goblins.
That's what my prompt was
I got this:
cat
bot
HUMAN
pumpkin
a cat and a robot walk into a bar
I have won.
what did you win?
a boy running in bar
Sorry, I got the time wrong. They just updated it.
Now, Community Commercial is Free!
Only limit is less than $1M in annual revenue
https://t.co/wyFAjLjThs
THE LICENSE HAS BEEN FREED
well, it's a new license beast
revenue up to 1m is super reasonable
So I saw the ping, I'm trying to read through this announcement.
"At Stability AI, weโre committed to releasing high-quality Generative AI models and technology,"
You're literally not, what was released was far from high-quality and you have no commitment to doing anything about it...
did u even read the announcement
The first line is a lie and theres nothing in your quoted response to contest otherwise?
where is the announcement
Well they do say that we can expect a new SD3 medium model "in the coming weeks"
But who knows if that will even be true considering they seemingly work on Valve time
If you disagree and insist they are comitted, is this a newfound comittment or were they comitted before the sd3 release?
how do you know its a lie T_T
Because ppl literally have shown that it's worse than XL in a lot of cases
im saying their internal test may have showed it being better than xl
their tests may have had blind spots or been in areas the community didn't focus on as much
If that's the case, that's still SAI's fault for not realising the blind spots
Only mention of SD3 medium still, and for that only recognizing "... body poses and words that were too rarely seen in the training set." But atleast SAI showed a sign of life again, future still unclear, but well, time will tell ๐
yes... but that means the line isn't a lie. their internal tests, due to blindspots, misidentified the merits of the model
in the meantime have a pepper
fix the model next gl
But kudos for removing the $20 plan and just making that free
Hopefully that brings some goodwill back
Ah, so they're confirming that they rarely trained on human pictures other than the basic straight standing pose shot 
Also keep in mind we don't know if this means that this retrain will fix other issues
Right, since it could be a lie
It may just be retrained to shut up the "SD3 can't make women lie on grass" ppl
"our initial testing indicated that it was, in most cases, a much better base model compared to SDXL"
Did any of you genuinely think this after your first hour with sd3 medium? it was immediately obvious to me it was broken, and i wasnt rendering women on grass
When the model clearly has other faults
When generating landscapes? Absolutely
yea - but then i worked with it for a while.
Do you usually test new models by generating nothing but landscapes for an hour?
there are a few issues - but it is incredibly good
when 1000 people are generating, looking for flaws, a few find horrible test cases. those spread like wildfire
In hindsight, it seems obvious they just wanted to put any model out in the open before the takeover... what the new SAI will too, still to be seen
AFAIK, the model is only good when not generating humans
they answered this multiple times in this channel
the model does humans just fine.
the issues are a shortening and warping effect on ALL subjects, not just humans, as the point of view and posistion moves away from stright up in front of the camera.
Any model could make that though. Even 1.5 albeit at a lower res
They never said a thing, just that it was trained in record time (on a record tiny dataset :p iy seems), but meh, i don't care anymore what is said, only what's done.
and the dev team is aware of the issues, and working on a fix for the specific things causing those problems
the Paid enterprise edition. will this still be Renting the model? will everything created for this model need to be deleted if we cancel the enterprise license?
this new license thing is good, hope the future holds more good news ๐
yeah they do. alex did, lykon did, and everyone on reddit and in here accused them of lying, then spouted off wild speculations based on their own imagination, and are continuing to do so
It is good, but it might be too late since everyone has moved on
One can hope! More competition = better, after all
Well, most ppl I mean
source = ass
did they actually say this, where can i look
is there an update on the paid enterprise license?
I guess the link in #๐ฃ๏ฝannouncements ?
they posted here on this discord. you'll have to search
mmm pays to read i suppose
i for one,welcome our new chinese overlords
Welp, time to learn Mandarin
Still only mentioning SD3-Medium is not a great sign, sure maybe with proper training it'll end up close to the API, or the API version will see the weights released eventually, but SAI is not even hinting at that anymore
but he speaks Cantonese

went from 20 pages to 3.5 pages, which is pretty normal.
are we ok now? with the license?
Seems like it
SD3 Medium is still important because it's the model size that the community can use and work with more easily. With new resources coming in, we'll expand to work on more projects at the same time.
smaller pocket sized beast
long story short, it's free open source non-commercial and free commercial until 1m revenue. It's pretty reasonable, and will open the possibility to cooler stuff, like memberships
my biggest hope is that it will get copied so it's used instead of non-commercial licenses
Continuous Improvement: SD3 Medium is still a work in progress. We aim to release a much improved version in the coming weeks
Interesting ๐ค
put it out under creative commons?
nsfw tunes here we come !!
SD 3 medium...
what i'm curious about is the enterprise licensing, does it require that some community standards are met?
AUP, so basically no illegal stuff
which is a pretty low bar to meet
right thats great. hopefully companies will take csam as seriously as it has to be taken. i see a lot of it on civit.
if civit even gets an enterprise . they gotta be over a milly in funding
Oh definitely for the average user that wants their own little fintetune/lora. But seeing how well 8b in the API works (it just knows so much more) i'd still hope it'll see a release of it, please say that's not idle hope ๐
do you have any idea how much vram it takes?
Still didn't stop 1.4 from releasing which required like what 40gb of VRAM initially?
Or hell even SVD which required 32gb initially IIRC
doesn't mean he's got a machine that'll run it, however
I personally want to collaborate more with civitai. As far as I know they take csam seriously with lots of filters. I'm not super aware of their recent situation because I work too much now and I can't really follow.
SAI def have a machine that can run SD3 8b since, well, they're running it rn
some ppl here also have it
i meant Aliquip himself

I CAN ADD PAID SD3 IMAGES TO MY DA ACCOUNT OFFERS NOW!!!!!!! ๐๐๐๐๐๐๐๐๐๐
people that can run and train 8b even with no quantization obviously exist. It doesn't mean that it's good for the vast majority.
Improving 2b is still important.
Aliquip himself just uses cloud rentals, so pretty sure i can run it, besides, high end consumers GPUs have 24GB, which is plenty unless you stubbornly keep T5 in vram
you could have done it even before. Nobody was going to police you #NotALegalAdvice
ugh. work!> i know. totes.
i just know there's been some stone walling with people who want to report content . the accounts they report are made invisible to them, but not banned from civit. so content isn't actually removed in many cases. they just hide it from reporters.
In the american state where they operate , simulated csam / cartoon , is legal and i think they allow such content. not sure that's the way it is in the UK.
i dont really wanna talk about it. gets me a lot of negative attention when i touch this subject
I can relate ๐
Prob true, but the old license also sounded like I'd have to pay $20 per month. My DA doesn't make much rofl, so I was waiting until I got really really goidzat it ๐
DA pays money now a days?
plenty of money there if ppl likes what you make
that part of the old license was copied verbatim from sdxl turbo and svd ones done under Emad, but nobody ever cared
There are people who are less stubborn than me who do females, who actually make money at it
so begins my foray into fur art
furries
go where the money is
