🏞｜general-with-images

[clever oar]

neural network for trying on clothes

[shut sinew]

Which one are you using?

[jovial tiger]

#SD3 a kart racing track in the Harry Potter style, in a cave

taking requests if anyone wants me to run prompts.

[shut sinew]

Ah I tried to get this to run today in ComfyUI but gave up after multiple hours

[clever oar]

try online

[cyan shoal]

wait wha

how

[shut sinew]

API

[cyan shoal]

ALREADY

DAMN

okay so uh

[shut sinew]

I did but I wanted a local install with it

[jovial tiger]

api. I got it configured with stability core, so it was ready to go for sd3.

[cyan shoal]

Team Fortress 2 gameplay screenshot, Heavy Weapons Guy character is holding a minigun and is shooting with furious intent. He is screaming.

iirc yoinked said it looked similar to DALLE3 in a recent model version

so if it gets to that level I'm impressed

probably wont but still worth a try

[shut sinew]

Thing is the API runs a pre release model, just keep that in mind

They are still working on it

[cyan shoal]

oh yeah I know

[cyan shoal]

yeah generic shooter

[jovial tiger]

final release, max quality. got it.

[cyan shoal]

^ LMAO

[cyan shoal]

So guys, we must judge it with super harsh critisism and treat it as a final model, good idea. I hope the Subreddit will do that.

[shut sinew]

I found it has issues with a lot of characters that the model doesnt seem to know

[cyan shoal]

the subreddit will ONE HUNDRED PERCENT do this

like no joke

[shut sinew]

Sure will, its reddit after all

[cyan shoal]

people are gonna shit on it

I am going to expect that

@jovial tiger A hilarious meme featuring two distinct halves. The top half showcases a dark, gritty scene of a Darksouls-style warrior engulfed in flames, wielding a massive sword, and looking determined. The caption above reads: "This would be a great analogy of...". The bottom half of the meme features a blonde Leon Kennedy from Resident Evil, holding two large pistols, with a mischievous grin on his face. The caption below reads: "YOUR MOM!"

if it's like free or whatever then I would like some more test with stuff like text

if its paid then this would be my last request

[jovial tiger]

so it's worth noting that of the 6 i send back, half are expanded through an llm i have. and 3 are raw as you typed it.

[cyan shoal]

wow this is pixart-sigma quality prompt adherence so far

like I know its preview

but thank you for the images

[nimble mason]

that's SD3

it's great so far

[cyan shoal]

see that's nice

I'm just testing stuff that worked great with ideogram

memes work well with ideogram

[junior sky]

Those remind me of SD 1.1. And not in a good way.

[cyan shoal]

me fr fr

[jovial tiger]

nah it can do text fine. just have to do a simpler prompt

[cyan shoal]

my dream was that it would be as good as ideogram tbh

so far preview doesn't seem like it

[jovial tiger]

ideogram can't do 37 subjects either. 🙂

[cyan shoal]

37 subjects lol

dw I haven't turned into an instant hater or anything

[jovial tiger]

the more subjects with ideogram, the most simple it gets.

[cyan shoal]

I just don't know what to expect from the final model

this is from ideogram for example

I would like to state that I am still aware that this SD3 is a PREVIEW model

[jovial tiger]

Intense anime-inspired digital art battle: towering mecha bears with glowing red eyes clash against tiny warriors in an epic struggle to breach a majestic mountain pass, sword clashes and energy blasts illuminate the chaotic scene with dramatic lighting, massive stone arches tower overhead.

[cyan shoal]

that's nice

[shut sinew]

Make some anime stuff for #🍥｜anime

They surely wanna know how it fares

[jovial tiger]

it's also worth mentioning, that this is default settings. so no adjustment for step count, upscaling, hires fix etc. all the normal stuff we'd be doing.

[cyan shoal]

Photo of Leonardo DiCaprio holding a floppy disk with "SD3-800M" written on it. He is examining the floppy disk up close with confused expression. He is puckering his lips.

[digital hornet]

that's the case for midjourney and dall-e as well tho

[jovial tiger]

In the center of a vibrant canvas, a fierce battle scene takes place against a backdrop of lava-covered terrain. The left silhouette character is a majestic ninja, holding wolverine claws and the right behind it is a japanese samurai, holding a katana. The black background is reminiscent of a vivid landscape, with the lava mountain contrasting sharply against the stark white background. The painting exudes an aura of calmness and serenity, making this an unforgettable representation of martial arts in the most unexpected way.

[cyan shoal]

YES

EXCELLENT

even if cherry picked

amazing thank you

this is so... epic

[digital hornet]

this one is pretty good yeh

[jovial tiger]

definitely not. midjourney and dall-e are doing TONS of processing. they're a workflow of lots of steps, hardly just a simple diffusion and send.

[cyan shoal]

DALLE does prompt modifications for example as well

and ideogram has it optional

[digital hornet]

oh wait which sd3 is the one the API uses? biggest one? you choose?

[jovial tiger]

it's either regular or turbo.

it's been mentioned this is like version 0.8.

[cyan shoal]

amazing thank you

[jovial tiger]

#SD3 a stone giant wrapped in chains and tattoos is crouching in front of a gate looking down at a small group of diminutive people who are trying to gain access to the cloud town

[digital hornet]

can someone do something with large groups of people, which normal SD often starts messing up, e.g.

"A party of many groups of people in the style of The Garden of Earthly Delights"

[cyan shoal]

pixart but without the overbaked look

very nice

[jovial tiger]

it's also doing that giant crouching correctly. pixart is pretty good with that, but this is better.

no squashed calves

[jovial tiger]

i'll only post one. that kind of art often has naked people, so i won't post more of that here.

[cyan shoal]

hmmm

[digital hornet]

thanks it does look a bit better than sdxl but many of them are still deformed as I feared

[cyan shoal]

yeah they are ken barby dolls

makes sense

[jovial tiger]

ok last request? have to get back to work

[cyan shoal]

Screenshot of Steam Store page website, multiple game pictures listed in grid pattern, Games listed: Team Fortress 2, Left 4 Dead 2, Titanfall 2, Half Life 3

thank you 🙏

[shut sinew]

This wont work

[cyan shoal]

yeah I know

its not going to but I want to see what it does

[cyan shoal]

nice, thanks

half-lied 2

I love errors

[digital hornet]

doomer take but as far as I can tell sd3 is a bit better in general but most of the issues are likely still there, and if this is the last open source architecture we'll get in the near future, self hosted is going to quickly fall behind

[cyan shoal]

^

[jovial tiger]

screenshot of a steam store page website, multiple game pictures of cute furry animals, games listed: monster furry 1, monster racing 3, monster jetski park

ok that's really cool.

i'm very impressed that it could just make that up.

[cyan shoal]

nice

[jovial tiger]

i'm not concerned about visuals, that can be cleaned up with more passes and steps

[cyan shoal]

(broken record - I know its a preview) but I expected a littttttlllee better from prompt adherence

like 90% of ideogram

it sadly couldn't do the meme I wanted so that dream's gone

[jovial tiger]

yeah... I think that research paper was "looks better than dall-e and ideogram" which for regular stuff, I think it does. but prompt adherence may be an evolving thing.

[cyan shoal]

or the benchmarking had terribly simple prompts

[jovial tiger]

i put a lot of other stuff against it. people sneaking past a sleeping tiger and it did it, whereas no amount of sdxl seeds could.

[cyan shoal]

yeah thats great

[jovial tiger]

but yeah, simpler non-expanded prompts seem to do better.

[cyan shoal]

aw

[jovial tiger]

it's so hit or miss though. i've done a lot of the same against pixart, and it's completely 50/50 as to whether the llm expanded prompt will come out better than raw.

so i just have my automation run both per job.

#sd3 a flamboyant K-pop singer on an elevated platform singing. Down below, hordes of zombies march across a modern city street.

yes, those are cherry picked.

out of 12.

[cyan shoal]

yeah lack of highresfix but nice

[digital hornet]

damn the face is so bad in both

[cyan shoal]

yeah that's what no highresfix gets you

[jovial tiger]

pixart is worse. it's only good after I've put it through an sdxl refiner.

i assume this is a lower end model for sd3.

not their top

[cyan shoal]

if it is I'd be happy

cause if the lower models can perform this good then its all good

[digital hornet]

they might want to generate data on all of them yeh

[shut sinew]

Close up faces are better for sure

[jovial tiger]

this is an example of what things can look like with the refiner.

of sdxl.

so i could definitely clean up those ones above easily.

[cyan shoal]

yeah we see prompt blending

[jovial tiger]

^^ is from pixart, with my model merge as sdxl refiner.

[cyan shoal]

oh

[jovial tiger]

the 3 a few posts back is just raw sd3 api

[clever oar]

hybrid half life and l4dead

you twube 😃

second its for china

[jovial tiger]

#sd3 Menacing cyberpunk Squidwarrior brandishing a towering rocket launcher, its sinister eyes alight with merriment. Tentacles grip the massive cannon firmly; fiery blasts engulf the nightmarish scene in smoky orange light.

Holding of weapons in sd3 is much better

[hybrid hemlock]

#🏞｜general-with-images

[jovial tiger]

#sd3 emperor sitting on throne in ornate vestments, drinking a soda from a can. Anthropomorphic Turtle minions raising another can of soda on a felt pillow towards his face. In a Decorated throne room.

[cyan shoal]

wow diet coke

[junior sky]

Did you also tested other ratios?

[junior robin]

Anyone got access to sd3?

[river quiver]

api is available now

[junior robin]

Where can I use it?

[river quiver]

use google

or check announcements

[junior robin]

So isn't available yet?

[clever oar]

for vip maybe

[river quiver]

it is…i named options where you can can find the answer to your question

[junior robin]

But there isn't any site that got the API and put it publicly available?

[jovial tiger]

Yeah I can run a prompt if you want. Apparently my work will suffer today.

[river quiver]

https://bit.ly/3xHrtjG

[junior robin]

Where and how?

[cyan shoal]

bruh why bitly

[jovial tiger]

API. I had written automation for it with their core model. So changing the endpoint took a few seconds.

[cyan shoal]

https://stability.ai/news/stable-diffusion-3-api non bit.ly link

[junior robin]

Interesting, but any other site did what you did to make it available for anyone?

[river quiver]

a vintage black and white photograph of an old woman in the mountains holding up large thorny plants, her face is hidden by them as she holds on to one branch for support, her dress has tattered off with leaves falling from it, the sky above is dark clouds and misty, there's some mountain range behind her

[jovial tiger]

I haven't yet. Let me know if you want me to run something for you with an alternate AR

[bright harbor]

Maybe, but there will always be a demand for the open source
source solution. gimp may be behind Photoshop, but it has a huge following, same with blender, same with Linux itself

[junior robin]

Prompt: brie Larson, with White sleeveless blouse, dirty and wounded, blonde ponytail hair, angry expression, holding a Flare gun, with old castle in the background, snowy, dark ambient, cinematic, 35mm film

[digital hornet]

it's not really that simple, as things get more expensive. there is no open-source nvidia for example

[main dew]

base models cost millions to train

[crisp stream]

Worked pretty well if you ask me 😄

[river quiver]

thanks!!

can you try these out:

glass sphere head by germaine krull, pictorialism

hand colored stamp, family heads in row, pictorialism

pictorialism print,(pictorialism), by helene schjerfbeck, featured on flickr, orphism, film grain, provia,portrait,Jules Bastien-Lepage, postcard

[nimble mason]

Those are sd15 style prompts

[river quiver]

i want to see the difference

A photo of an eerie scene from the early 20th century, featuring two antique wooden women hanging in dark woods. The focus is on their ethereal and haunting presence amidst the dense foliage, illuminated by soft moonlight casting long shadows. Shot with Kodak Portra film using a Leica M3 camera to capture intricate details and textures. in the style of an early 20th century photographer.

[bright harbor]

Hard to argue that point given the reported funding issues facing sai, but some solution will present itself, whether it's distributed computing or something else

[river quiver]

@jovial tiger

[main dew]

I dont think distributed computing works with ai training

it would be incredibly slow I think

[jovial tiger]

it's kind of weird. many of the images i'm getting black are blurred...not sure if that's intentional or not. none of the "squid doing this!" were blurred.

[bright harbor]

SETI for AI 🙂 I'm not saying there's something out there right now that's ready to go, I'm saying it's a problem in search of a solution

[river quiver]

victorian era human sized doll body, human face, in rural kitchen, balancing on an rope and holding flowers while looking to the right where the door is, in the door is and old man balancing three cups of water on his head and shoulders, pictorialism, soft focus, julia margaret cameron, alfred stieglitz @jovial tiger

wondering how coherent the scene would be

[junior robin]

Very interesting, looks like a Worsened version of midjourney

[junior sky]

Now thats something in my style

[junior robin]

It's not really live action, and the hands are not that good

[river quiver]

interesting is the variety

[junior robin]

An enchanting artwork titled "Os Sonhos de Madoriya" featuring a ginger-haired, short-haired young woman with blue eyes. She holds a glowing yellow supernatural stone and wears elegant Greek-inspired white clothes adorned with golden armor features. In the foreground, a mysterious young man with long, black hair and a cape stands, knife in hand, also in a dynamic pose. The overall atmosphere of the scene is intense, with a blend of light and shadow, creating a sense of adventure and intrigue. @jovial tiger try this one, is very specific

[river quiver]

haha 🙂

[main dew]

why does it look so uncooked

the steps are really low huh

or maybe turbo model

[junior robin]

Not very coherent actually, I expected more

[jovial tiger]

They have the turbo model as an option. This isn't that. I have no control over steps or anything though.

[main dew]

like its really really bad

[junior robin]

Yeah...

@jovial tiger try this, it is a very difficult one, If he can do that, that's good model

red dead redemption realistic concept art, realistic art, horror wallpaper vibe, the title says "os sonhos de madoriya". in the up front a ginger short haired young woman with blue eyes and short hair, with greek white clothes with golden armor features in a dynamic pose holding a a glowing bright supernatural yellow stone, a young man with white skin and brown hair and dark clothes holding a knife in a dynamic pose in the right, an evil old king with white skin and hair and beard in the back, a young woman princess with a cocky face expression and white hair and purple eyes. in the back a kingdom in the background with army of soldiers, epic, cinematic composition, action vibe, dynamic vibe

[glad narwhal]

I have to imagine this situation is similar to all of the Stable Diffusion releases.

Their models just struggle, I would bet some finetuned versions look much better than this.

[clever oar]

why i get this result with promt girl holding a large domestic hornet

[junior robin]

Yeah, probably

[main dew]

no idea XD

[shut sinew]

Just as a reminder, this is the preview Version of SD3, it's still being worked on

[bright harbor]

What if you say "hornet insect"

[junior robin]

Not coherent at all, dalle 3 has a very better prompt understanding and ideogram too

[glad narwhal]

But lets also remember that SDXL 0.9 was basically in the same ballpark as SDXL 1, things are not going to radically change, they will just be a bit more clear or concise

[junior robin]

@jovial tiger But thank you very much for showing us how the model really looks

[shut sinew]

That's just speculation tho, we dont know

[shut sinew]

Also prompting changed again, people gotta adjust and not just throw old prompts at it to compare models

[nimble mason]

four bottles lined up on a table. from left to right, they are numbered "4" then "3" then "1" then "2". from left to right, they are red, blue, green, and orange. the setting is a magical forest from a tolkien novel. in the sky is an alien UFO abducting a cow with a ray of light.

[glad narwhal]

We have like 5 stable releases where this is true though.

[bright harbor]

Nice, haha

[junior sky]

Really looks...in this aspect ratio. @jovial tiger Still don't found a prompt thats worthy wasting your credits.

[main dew]

certainly not worth 20x sdxl

[sterile kiln]

You did just with prompts?

[main dew]

who is even looking to make finetunes of sd3

[jovial tiger]

Waste away. 🙂 I'm curious to see what'll do

[junior robin]

It is just not that good

[errant yew]

same whining we got with SDXL lol

nothing ever changes

[glad narwhal]

Its not "that good" as a final step. But the composition, consistency, and variety are ace.

[errant yew]

then everyone comes around once all the finetunes are out

just in time to whine about the next stability model 😄

[glad narwhal]

Just pass it through SDXL and we are at midjourney level already, SD3 doesn't need to do everything

[junior robin]

Sdxl is not good as well lol

Dalle 3 and ideogram rules

[nimble mason]

yup, that's SDE3

SD3

[glad narwhal]

This is sounding spiteful lol

[junior sky]

Lets see if it has the classic cave entrance problem. => A analog photo of a futuristic metropolis build inside of a dark cave.

[nimble mason]

that was the prompt and that's all there was to it

[junior robin]

I know the models are cool because they are open source, but in an aesthetic sense there really is no comparison

[jovial tiger]

#sd3 Illustration of the DMV headquarters, showcasing a sleek and futuristic design. The building stands tall amidst other skyscrapers. On its exterior are massive digital screens that are unmissable. One screen prominently features the 'DMV' logo while the adjacent screen displays an alarming 'OFFLINE' message in vibrant red. Below these main messages, there are warning signs and error codes flashing, suggesting a major system disruption. Pedestrians on the street halt in their tracks, some pointing at the screens and discussing among themselves, while a few are capturing the moment on their devices.

[junior sky]

F that

[nimble mason]

got an example of the kinda image you're lokoing for

?

[jovial tiger]

He didn't skip leg day, and neither did his arms. 🙂

[errant yew]

I think Stability needs more creatives. They have brilliant researchers, and people in the AI bubble. But they need more artists.

[junior sky]

Clown, you know that SD is putting openings with light leaking in in every cave picture.

[glad narwhal]

Dalle is good for single pass processing. If you want to develop prompts, it might be a better way to get a clean result.

But SDXL is objectively superior considering how the pipeline can be augmented.

If Dalle was open, this would not be true.

[junior robin]

True

[errant yew]

How are people generating images, with Kijais SD3 comfy node?

[nimble mason]

gotcha, so you're wondering about the light? as in having it be dark

[jovial tiger]

I have my own api automation.

For now anyway

That would be neat to integrate it into comfy though for more processing

[junior sky]

Surrounded by stone, so to say.

City is emitting the light

[nimble mason]

k

A analog photo of a futuristic metropolis build inside of a dark cave. the city is plunged into darkness during a power outage. the entire image is nearly pitch black.

[junior sky]

darkness during a power outage 🤣

[nimble mason]

is that sdxl output or something?

[clever oar]

1.5

[nimble mason]

haha

[clever oar]

my model 1.5 base)

[nimble mason]

desperate measures haha

[deft bison]

the AI pushed the muscular apparatus to an extreme mutation!!! ...

[shy eagle]

Try so hard to get 4 bottles.

SDXL ft

[glad narwhal]

That styled text gen is so nice, can you use the "minecraft" text style to spell other things?

[jovial tiger]

@nimble mason @junior sky

[cyan shoal]

reuploaded image

[shrewd elm]

so API is using a very old model?

[cyan shoal]

why would they purposefully do that shit

they sabotage their image because of a possibly inferior model

[junior sky]

Isn't that just plain stupid?

[jovial tiger]

To make Boto nervous about our responses to it. /s

They knew he'd have to do damage control on here and wanted to play a prank on him

lol

I'm fine with it. I'm happy there's something to play with as they continue to grind away on it

[junior sky]

That must be it. Would have been better to open the API access on the 1st of April

[jovial tiger]

No more "release sd3!" Posts now

[junior sky]

Release the current SD3 model!!!!

[shy eagle]

No SD3 now

[jovial tiger]

Sd3

[nimble mason]

Bizarre lol

I won't pretend to understand SAIs decisions

[junior sky]

Soul would say: Skill issue

Anyway. Now i finally can go on with my life. Nothing happened.

[cyan shoal]

what I also don't get it that they almost never go out of their way to debunk myths and rumours

things that make the community disappointed

they just let it happen

[eager wasp]

Or this new tweet is damage control and they now postpone launch a few months trying to fix things

[cyan shoal]

@jovial tiger what interesting thing I found is that SD3 might be making random letterboxes the same way as Pixart-Sigma? lmao

[eager wasp]

But when sdxl was introduced on clipdrop, it was also inferior in every possible way

[cyan shoal]

at least I got one for now

[cyan shoal]

a dev said that they wont delay SD3 instead they might make SD3.1 and SD3.5 which will introduce changed or further finetunes

but idk

[junior sky]

I have trained loras to purposely reduce the quality of SDXL to what this model outputs.

[jovial tiger]

I feel that I have failed you. I can not adequately express the sheer amount of laughing that I just did at this message. I'm sorry.

I need to frame that on the wall

[eager wasp]

That'd be good. As this feels very much like sd 1.4 leak. Nowhere near done

[deft bison]

well done and thank you for this very useful information! ... 🤝

[cyan shoal]

keep in mind this is also a month older version and this has no highresfix applied to the images

lykon got amazing images but those probably had highresfix

[eager wasp]

This ideogram vs sd3 makes it look so bad. Text was supposed the big thing. Never really valued it much personally, but this is just sad.

[junior sky]

Shit. It knows how i look.

[cyan shoal]

idk what Stability is thinking rn lol

[digital hornet]

first attempt? or did it need a few

[eager wasp]

which again begs the question: why release it like this this :/ I'm glad there will be a new model, but right now i'd not even be surprised if an ella version for sdxl actually will be just as good.

[main dew]

going off what one of the devs said on twitter: the model they put on the api is a months old initial version of sd3 that is really bad

and that the sd3 devs have no control over release or pricing

[jovial tiger]

These were all 6 from that run

[nimble mason]

yup

[jovial tiger]

But it did a good job

[junior sky]

SD3 does a better job hiding the hands to not embarrass itself completely.

[cyan shoal]

hehe

[glad narwhal]

Why bake a model forever?

If stability did that with 1.5, how long would it have taken them to match the quality the community made?

[jovial tiger]

Sadly the rest were much worse. Not doubled fingers, just not 5, so I guess that's an improvement

[junior sky]

So, no word about the dental situation that is going on?

[eager wasp]

Never get a second chance for a first impression

[jovial tiger]

It's worth mentioning that everything lykon has posted in the last week or so was MUCH better than any of this. So this version certainly shows its age

[shy eagle]

what prompt of it? I want to try it by sdxl

[cyan shoal]

https://twitter.com/Lykon4072/status/1780641513862512983

Lykon (@Lykon4072) on X

@el_mejnun I expected something like this was gonna happen. That's a model from months ago (basically the paper one) and the inference is not on Comfy but on some custom made script by Fireworks.
I wouldn't rely on that to assess the model quality.

Twitter · Apr 17, 2024, 4:56 PM

I know this was posted btw

[junior sky]

The gigantic elephant in the room is still: Why would somebody do that?

[eager wasp]

But how much of what was on twitter was cherrypicked (or even postprocessed/upscaled), that's why i wanted to try this api. And going by "i'll belief it when i see it" what i see now isn't very good

[stark anvil]

what sd3 looked like in feb (plus sd upscaler)
If i had to guess, management fumbled the release and opted for a cheaper to inference version? or something similar?

[eager wasp]

Dark panorama with a horrific green-skinned, hyperrealistic, life-like, zombie with rotting flesh and maggots crawling out of its eyes on a graveyard, large dark forest below, extremely detailed, 8k, intricate, warm summer night vibes, eerie silence, undead, dramatic back lighting, by by emil melmoth. just a normal prompt. It's a mess of noise

[glad narwhal]

SD1.5, SD2, SDXL0.9, StableCascade.... all had middling presentation from the get go.

They have been used by the community in ways that were convenient. Same will be true here.

[digital hornet]

maybe with the final version + some finetunes and tooling and figuring out what settings work best it will be more obviously cooler

but going from 1b 1.5 to 3.5b sdxl, the now 8B sd3 doesnt look as better as the parameter increase would make me think

[glad narwhal]

SD3 should be super responsive to the IpAdapter style dynamic step weighting, I think it will look phenomenal once we get our hands on it.

[main dew]

sd3 isnt out yet

the api version is the paper version which is months old

[jovial tiger]

/sai an angry man with a name tag that says "aimingfail" holding his hands up to the viewer so you can see his fingers.

[main dew]

wait for the full release to make ur opininos

[glad narwhal]

Source?

[digital hornet]

isdk Stability AI made a post that literally say 'Stable Diffusion 3 API Now Available' so it is sd3

[main dew]

#🏞｜general-with-images message

[digital hornet]

it just might get much better

[cyan shoal]

just to clarify, this is what is "soon" to Emad:

To clarify weights will be made available soon (always API first, then a few weeks later weights).

lol

lost in translation

[shy eagle]

Thanks

From SDXL ft. Just for fun.

[glad narwhal]

That is not a source, that is speculation. Unless I am mistaken.

[main dew]

they are a dev for stable diffusion

[cyan shoal]

^

[glad narwhal]

Well I am happy to be wrong lol

I still think it is good enough as is, and expecting one pass perfection is a fools gambit.

At least right now, the tech is still too young to want it to do everything all at once.

[main dew]

well the point is that its a botched release and we should wait until the weights come out

[glad narwhal]

Its not botched.

[digital hornet]

almost everyone here was going to wait for the weights to come out either way

[glad narwhal]

Its like saying sd1.5 is botched because the base model is crap

[main dew]

no, this isnt the finished sd3 model

and its not even being inferenced properly

https://x.com/Lykon4072/status/1780634581990776961

Lykon (@Lykon4072) on X

@itaybachman @StabilityAI As I said earlier today and also yesterday, this is a very old (and broken) build of the model. Also I learned the backend is not Comfy so it's not using my current (or even my older) workflow, nor the workflow that's on the Early Access bot. Not sure why it was marketed as final

Twitter · Apr 17, 2024, 4:29 PM

[cyan shoal]

which is embarrassing from stability's part

[glad narwhal]

Calling this release "botched" is going to set expectations at an unrealistic level.

That is probably semantically correct, but we are in the ballpark.

[cyan shoal]

idk why they'd sabotage their image like this

[main dew]

lykon is saying the backend is not even correctly inferencing sd3 I count that as botched

[glad narwhal]

That is fine, and I love the work Lykon is doing, but how people interpret things is a different matter.

I do not like the API rollout, but I do like what the model is doing, even in a hobbled state.

[shy eagle]

"Membership"

[nimble mason]

it's free for non-commercial

[shy eagle]

which version? the api one or?

[nimble mason]

the membership

for commercial it's $20/mo

[junior sky]

Oh, there's a SD3 channel now. I wonder if somebody will notice it if i just post upscaled v1.5-pruned-emaonly outputs.

[cyan shoal]

#🆕｜sd3

[unkempt mica]

This is what I am currently working on. Its using Command-R tools in combination with Stable Diffusion. It works almost the same how ChatGPT and Dall-E integrate with. There are no hardcoded triggerwords or anything like that. The model itself decides when and how to generate images.

Text above the image is the bot's reply. Text under the image is the generated prompt for Stable Diffusion. Look how Command-R knows exactly the prompt format AUTOMATIC1111 uses despite me not describing it. I only told it to use the Stable-Diffusion prompt format when generating stuff. however sometimes it still doesn't close all parentheses it opens if not instructed further

This is gonna be open source and will be an extension to text-generation-webui

[glad narwhal]

I use the Faraday app, and load 7b parameter models onto my CPU and then do text conversation prompt engineering then just copy it over it comfy.

What you have looks very very nice though.

[unkempt mica]

This is a bit different than prompt engineering, its more like GPT-4V as its a normal chat bot, not something just for image generation

you can talk with it and then randomly request image generation, it will also keep your previous instructions in mind

e.g. if you somewhen said to use only cartoon style, it will only create with cartoon style going forward

and its multilingual as well, not limited to English

[rapid current]

SD3 pretty good

[glad narwhal]

I have used LLava models to "look" and integrate that into my pipeline as well. But when LLava and LLM are not using the same data backend, it definitely causes some incoherency.

[glad narwhal]

WizardLM2 will let you do conversational prompt engineering with text alone, definitely not multimodal, but it is very good.

Using the system command as a place to inject example prompts is a really good way to get it to output proper prompts.

[graceful bolt]

Guys I don't wanna say nothing but it's so over for diffusion based image generators

[cyan shoal]

excellent

[glad narwhal]

Diffusion models were always just a transition. We will end up multimodal

[cyan shoal]

To ruin the joke: ||https://github.com/CompVis/zigma||

[main dew]

zigma balls

[main dew]

is it good tho

idk how many 30k iterations is

oh its 1 second per iteration so the training only took them 8 hours or what

[wispy nest]

Any recommendations on extensions that can accurately segment the product from a photo?

Attached photos are the before/after from using the following tool: https://www.photoroom.com/tools/background-remover

How are they so consistent with their results, any extensions that can help with this?

[crisp stream]

per image?

😄

[languid pebble]

This SD3 Release is pretty strange ...

Like selling cookie dough with missing ingredients saying: "See how to get along"

[south temple]

Did we get a release of SD3?

[languid pebble]

Kinda Beta you've got to pay for

https://discord.com/channels/1002292111942635562/1230206273451069540

[lyric storm]

@grand walrus yes, this is most definitely not correct, this is what happens when using the Karras scheduler for example -- same thing that happened in A1111. The only way that the user has to control refiner switchover is by sampling steps, which aren't always aligned with discrete model timesteps, and because of that it is switching to the refiner four steps later than it should have. The refiner is going to do practically nothing under this configuration, and a user isn't going to understand why

[unkempt mica]

Done: https://github.com/Trojaner/text-generation-webui-stable_diffusion/releases/tag/1.6.1

[lyric storm]

and the image differences. in order: no refiner, refiner switch at 20/25, refiner switch at 16/25 (which is the correct point for switchover on this scheduler)

[grand walrus]

why wouldn't they be aligned?

[lyric storm]

the different schedulers like Karras change the sigma schedule for sampling, and call the timestep which is closest to that sigma

I inserted a line to print the timestep at the point where that conversion takes place

[grand walrus]

I don't see the issue

I'm not sure why you would use karras on SDXL

[lyric storm]

🤷‍♂️ i can't say I've ever had it not generate more coherent images on any model (normal left, karras right)

also it's not technically a training-inference gap, since the model was trained on all of the timesteps being used, just not uniformly to how you are using them -- but this works somewhat in the refiner's benefit overall

[grand walrus]

you are wasting steps on those lower timesteps that do pretty much nothing

[lyric storm]

That's not at all what is happening or why this schedule is used. The point of the Karras schedule is making it so that the tangent of the solution trajectory for any value of sigma will always point towards x0, which you absolutely want. This is covered starting at page 5 in the k-diffusion paper: https://arxiv.org/pdf/2206.00364.pdf

[sick yarrow]

Hello everyone good day, I have a question, why in A1111 with colab, the DPM++ 2M Karras sampler has been removed....?

[languid pebble]

Maybe deselected in Settinjgs?

[lyric storm]

They separated the scheduler selection from samplers, so there should be a spot for you to select "Karras" separately. I don't know where it is because I have had that in quick settings for a while.

[languid pebble]

Sounds like I have to install A1111 again 🙂

[grand walrus]

you should try stableswarm instead a1111 is deprecated

[sick yarrow]

nop, I checked it

[grand walrus]

just try chopping off those last steps and you'll see

[lyric storm]

🤷‍♂️ skipping CFG on them is fast enough if I feel like being stingy about it (and apparently helps image quality too, https://arxiv.org/abs/2404.07724), but even then I end up seeing somewhat substantial changes to style so I often won't bother

[sick yarrow]

you were right, here is the new list, I didn't noticed at the beginning, thanks a lot

[thorn scroll]

Guys look what I did

[quaint quarry]

I got the pixel art models and stuff but how would I get it to make pixel art to a similar style of this?

I also got other platformer assets which I would like SD to style on

[arctic laurel]

that level of granularity would likely require a finetune of some kind, probably all on commodore 64 / intellivision / atari 2600-era sprites specifically

otherwise you tend to get generally 16-bit and early 32-bit era sprite art

before donkey kong country dropped and 3d pre-rendered sprites became the dominant aesthetic

[minor halo]

what is this used for?

[clever oar]

styles

promt saver

[crisp stream]

1.5 btw

[south temple]

Working on a project

[crisp stream]

I find the astronaut helmet frog transformation thing is pretty interesting

[unkempt lance]

craar

[pallid ruin]

cool!

[pallid ruin]

Wow it looks nice

[mild jay]

anyone know of a budget gpu/device for stable diffusion/ai and what are the bare minimal is for stable diffusion?

[crisp stream]

thank you 🙂

[mild jay]

is this sd3?

[crisp stream]

SDXL, pure base model

[mild jay]

How did you get theses images to look so good?

[crisp stream]

high-res fix for one 1.5x, rest is prompting (along artists)

[mild jay]

someday i hope to run sdxl locally.

[crisp stream]

it's quite pleasant 🙂

[mild jay]

Your making me a bit jealous, here i am stuck in 1.5 due to tech.

[crisp stream]

I had that until several months ago

[mild jay]

how did you overcome said obstacle?

[crisp stream]

buying a new Video card

🙂

[mild jay]

lucky, i wish i had money but i dont really because im in highschool.

[crisp stream]

yes, I see, good luck on your quest anyway 🙂

[arctic laurel]

lol

[crude zinc]

would this image of me be bad for making a lora?

[nimble mason]

just be sure to post it on civitai

[crude zinc]

im training a lora though kohya it's fine for it right? im asking bc of my hood and my arms are a tiny bit blurry

[junior sky]

I would kick this image out of the dataset. It can ruin the whole lora.

[crude zinc]

oh ok thanks ill do that tbh i did a lora already and included this image and my hair sytle (Color was right) and skin tone isn't even close to right lol

the rest are more clear

also i used booru this time to help with the captions even tho i already did the captions from kohya idk if that's gonna make much of a change tho

[nimble mason]

your lora will only be as good as your worst image

look at it that way

also, use onetrainer

[crude zinc]

onetrainer?

[hazy warren]

need more curry

[hazy warren]

needs aliens

[deft bison]

It looks delicious! ... 😀

[deft bison]

good idea ! ...I didn't pay attention! ... 😀

[graceful bolt]

@arctic laurel

Looks blurry to me

[hazy warren]

😄

[weary light]

[tame leaf]

When can SD3 be downloaded?

[shut sinew]

In the upcoming weeks, no date

[crisp stream]

Looks like the example workflow for Refiner (if I remember correcty) 😄

[languid pebble]

That would so 100% fit to one of my series ^^

Bad luck to much white ... otherwise print on Aludibond without white color ...

[crisp stream]

didn´t you use levels (Tonwerkorrektur)?

[quasi furnace]

@unkempt mica sent you a friend request

Does anyone know any professional Ai designers?... im looking to hire one i would pay $300 to who ever can put me in contact with someone and they get the role please 🙏

[shut sinew]

The bot doesnt work right now, please dont spam every channel

[hardy thicket]

kv design

[languid pebble]

Preparing for Dune 4:

[heady anchor]

I'm doing something like a social experiment. It's a collective world generator and in the future, I want to make a game based on it.

[terse nebula]

Stable Diffusion 3

[thin sable]

Yo

Is Stable Diffusion 3 good at full body shots?

I'll appreciate if someone can share some images with full body shots.

Obtained with SD3

[thin sable]

Thanks

[crisp stream]

any painty stuff?

[crisp stream]

looks interesting

[languid pebble]

Not yet 🙂

[languid pebble]

Couldn't find a picture of Discworld yet ...

[languid pebble]

Good nite! 👋🏻

[spiral summit]

What prompt was this?

[stoic lark]

I thought this was going to take forever, but I pooped it out quite smoothly. No wipe necessary.

[nimble mason]

Here is the image you requested.

[nimble mason]

Here is the image you requested.

[south temple]

lol

[nimble mason]

Here is the image you requested.

[south temple]

ok... Soooo the battle...

Im gonna make a thing... then the next image needs to defeat that thing.

Fight

[heady anchor]

https://discord.com/channels/1002292111942635562/1230212453288378418

[noble halo]

/pr

[foggy plaza]

Hello, I hope you can help me. I want to make images like this. Please tell me which model is appropriate. My idea was to make superheroes as vehicles for comparison. For example, the pictures I sent are the spider boat and the venom boat

@shut sinew Thank you

[shut sinew]

Juggernaut(XL) probably, or you wait for the SD3 release or use the API

[foggy plaza]

Please give me prompt to create such a picture, thank you

thank you

sorry I am new to searching about Steady Diffusion

[shut sinew]

Nope

[clever oar]

how do you make sure that objects in the distance are not distorted?

maybe all xl models good for that

[ivory glacier]

hello guys, what should i type in prompt to get a half face image please ? i tried (half face) (left side) (partial face)... none worked. is there a way to do that ? Thanks!

[wise field]

what mny desk draw looked like in the 90s

[random yew]

maybeeee try close up photography then describe the eye, ear, to let it "focus" on that side...

[shell sleet]

trying my hand at scenery

it really hates pumpkins

this is fucking gorgeous

[shell sleet]

on forge

using hires fix 4xldsir at 1.55

threw it through usdu and got this

It does the shadows better, but it makes the pumpkins lose detail

[quartz ravine]

hey, I thought SDXL is not supposed to have fucked up fingers anymore

what's with this shit

[nimble mason]

Are you some kind of five fingered mutant? This is normal

[nocturne oak]

Whoever said it was perfect with hands? No model is always perfect with hands. Period.

[quartz ravine]

:((((

why hands are wrong every time though

[nocturne oak]

Hands are incredibly complex. It's not like your nose where you have one of them and it's generally in the same spot with limited positions. You have 8 fingers and 2 thumbs to deal with, often times in various positions. That means that across many many images, machine learning isn't nearly as likely to figure out the huge variety of common patterns that represent hands in comparison to a nose, for instance.

[quartz ravine]

No no, I'm not asking philosophically

I'm asking more or less: I thought XL fixed that issue largely. Am I wrong? Is SD3 the one that fixes it?

[nocturne oak]

I'm not answering philosophically. I'm literally telling you why it is not solved. SD3 will also still generate hands with problems. Probably less frequently in the same way that XL does it less frequently than 1.5.

[quartz ravine]

I thought I didn't need to stuff like 100 word negative prompts filled with "bad fingers, too many fingers, way way way too many fingers" etc

[nocturne oak]

You shouldn't. That' doesn't actually help.

[quartz ravine]

Hmm

now how am I going to share pictures of two women eating mcdonalds icecream with my friends

[nocturne oak]

By going to a McDonalds with a camera and paying pairs of woment to eat ice cream while you take pictures of them.

Let me know how that goes, btw.

[quartz ravine]

these Midjourney chads with their "two women eating mcdonalds icecream" are breaking my balls about my fingers

How am I supposed to compete with this???

What kind of McDonalds are they getting their images from?!

The right number of fingers while they eat their ICECREAM?!

It must be AI generated because the machine isn't broken!!

[unique condor]

Stable diffusion doesn't actually "know" what hands are and there are a million and one different poses they can be in. It would probably take a gigabyte's worth of network space for hands alone to do them right and under all kinds of various environments, skin tones, etc

Might even take a handful of gigabytes

[nocturne oak]

Likely

[wild sorrel]

left girl's hand looks fine, yea, but if you'll look closer to the hand of the right one...I'm not sure what's happening with her hand

[hazy warren]

need the black perl ship

[hazy warren]

a Mcdonals I want to be at

[hazy warren]

needs a Maccas meal next to it

[pearl loom]

hei hi , sorry to disturb may i ask u how can i replicate the same concept what can i write? how to describe to have a similar design and colors etc etc ?

[mighty patrol]

got some free labor for one of u, im too new at this to be able to do it anytime soon, so this is for anyone skilled enough to do this, but itd be a good idea for someone to make a lora with excessive training on english letters, so teach it each individual letter by having an blank white image with the single black letter, and describe those images accordingly, and then give it more complex versions with images of strings of letters and sentences, and put words and letters in various regular non-blank images

[severe roost]

Where to send prompt?

[arctic shuttle]

/d painting with flowers that come from the background surrounded with color

[white pivot]

One of my longest video so far. Add everything I know so far about ai animation.

[limpid lichen]

yup. this is the biggest reason why the simpsons have 3 fingers and a thumb. easier for animators to draw then.

hands are a difficult art problem. fingers can fing all over the place.

https://tenor.com/view/the-call-them-fingers-but-you-never-see-em-fing-otto-the-simpsons-stoner-stoned-gif-23967439

[hazy warren]

😮

[chrome nebula]

How to recreate this image: https://civitai.com/images/1519380

Tried copying into SD UI, but getting different image.

[ripe bluff]

@chrome nebula person in question:

[clever oar]

how many lines need connect oh is not for me

[spark finch]

You can compete with that by feeding her more icecream, getting her bikini bottoms that aren't 3x too large, and sending her to the hospital to address that abdominal torsion. And maybe remove the second arm she has from th left girl, in the darker blurry area behnind them, unless that is right girls dislocated elbow leading to that interesting angle.

[spark finch]

really though, what they don't understand, while acting as such, is that this... is not "two women eating mcdonalds icecream" this is "two women eating mcdonalds icecream" + prompt negotiation and extension by midjourney. Midjourney is not "two women eating mcdonalds icecream" -> make image. it is "two women eating mcdonalds icecream" -> write a better prompt for the user to match sentiment and intent with a more complex input before sending it to generator

[spark finch]

for example if you try ideogram with prompt magic that original prompt becomes - " A cheerful scene of two women enjoying McDonald's ice cream cones in a casual setting. The women are seated on a park bench, basking in the warm sunlight, and they appear to be good friends. One woman is smiling, her ice cream covered with colorful sprinkles, while the other, with a more serious expression, enjoys her chocolate-dipped cone. The background shows a lively city park filled with families and children playing. "

properly using more descriptive prompting or LLM to negotiate your initial prompt , you will find you can get more "MJ" like outputs

[still rose]

Wait this is pretty cool how did u do that

[spark finch]

now since MJ went the route it did, even with a still simplfiied prompt, to ideogram, which will use its prompt magic llm to negotiate a better prompt from "two women beautiful women in brown bikinis eating mcdonalds icecream in a dim seaside bar with a stained wood aesthetic and antique glass lighting, sitting very close and grinning seductively "

now this prompt came out with prompt magic to be : A captivating and nostalgic photograph featuring two stunning women wearing brown bikinis, enjoying McDonald's ice cream cones in a dimly lit seaside bar. The warm, antique ambiance is accentuated by the stained wood and vintage antique glass lighting fixtures. The women, seated to the right of the bar, face each other with a sense of camaraderie, their smiles inviting the viewer into their moment of happiness and enjoyment. The bar, occupying the left half of the image, frames them off-center, creating a balanced and visually appealing composition.

[deft bison]

as you know and in general that among the rules of a site is not to mention the names of other sites, it is considered an act of advertising, it is for this reason that I responded to you with a general way...

[crisp stream]

😄

[jovial tiger]

it's ok, stability is a non-profit.

i keeeeeeed

[hazy warren]

generating babes I see

[jovial tiger]

lol

[hazy warren]

I don't understand what your saying, but im just going to nod my head

[spark finch]

I was responding to the guy earlier who asked about MJ and how could SD compete with , and I said that the prompt they used is not the "prompt" they used

[spark finch]

what? that's ridiculous and gating as hell, of course you can mention other sites lol

[jovial tiger]

it's frustrating that they don't let you see the expanded prompt. doesn't truly let you evolve the prompt to work towards what you want.

[spark finch]

yeah that's why I mentioned ideogram, I use it sometimes FOR magic prompt, and then take to SD to use that across other models to see how I can get it more in line with mj / ideo style

[jovial tiger]

yeah.... although ideogram has almost no style.

it's the opposite of MJ which goes too far.

[hazy warren]

MJ = Mostly Jam

[jovial tiger]

haha

pixart, MJ, sd3

[hazy warren]

batman!

[jovial tiger]

i feel like mj's prompt understanding has been getting better very recently.

i think they see others are chasing them and they had to get better quickly.

[hazy warren]

Batman eating ice-cream

[jovial tiger]

definitely a difference compared to when v6 launched.

what else is he doing? who is he with?

is the ice cream alive?

[deft bison]

I thank you for this information which enlightened me... and it's really good to evolve in a vast space in terms of expressions of words! ...

[hazy warren]

https://tenor.com/view/what-batman-superheros-nge-eh-gif-6068419068561014816

[jovial tiger]

sd3 works sometimes. "when your bat suit is at the dry cleaners and all you had was this black cellophane"

[wispy nest]

Does someone use Stable diffusion Api endpoint

[jovial tiger]

yep

[wispy nest]

Can you show me?

[jovial tiger]

what do you want to know?

[wispy nest]

Your endpoint

[jovial tiger]

it's not MY endpoint, it's THEIR endpoint.

https://platform.stability.ai/docs/api-reference#tag/Generate/paths/~1v2beta~1stable-image~1generate~1sd3/post

[wispy nest]

I used many endpoints, none of them working

I mean

[jovial tiger]

  -H "authorization: Bearer sk-MYAPIKEY" \
  -H "accept: image/*" \
  -F prompt="Lighthouse on a cliff overlooking the ocean" \
  -F output_format="jpeg" \
  -o "./lighthouse.jpeg"```

that's the simplest implementation just on the command line. see if you can get that working with your api key

[wispy nest]

Also, there are 2 platform for Keys

One stability.ai and stable diffusion

Which one is choose

[jovial tiger]

i use the one mentioned in the url above

it has the url in there.

https://api.stability.ai/v2beta/stable-image/generate/sd3

[wispy nest]

From where you get the key

[jovial tiger]

https://platform.stability.ai/account/keys

[wispy nest]

Is it working?

[jovial tiger]

sign up for an account and then go to the page I just linked. you can create a key there.

I think you get 100 free credits, and then you have to put in money

[wispy nest]

https://stablediffusionapi.com/settings/api

@jovial tiger are you in pc right now?

[jovial tiger]

why do you ask?

[wispy nest]

      const body = {
        key: process.env.NEXT_PUBLIC_STABLE_DIFFUSION_API_KEY,
        prompt: newPrompt,
        width: 512,
        height: 512,
        samples: 1,
        num_inference_steps: 21,
        enhance_prompt: "yes"
      }
      const imageResponse = await fetch("https://stablediffusionapi.com/api/v3/text2img", {
        method: "POST",
        headers: {
          "Content-Type": "application/json",

[wispy nest]

Can you check console logs?

vega.js.org

[deft bison]

a thousand apologies to Mr. vortex for the inconvenience. I sent you a message by mistake and which was intended for another person and I thank Mr @Lorentz a second time who enlightened me on a question that I was unaware of... 🤝

[jovial tiger]

you'll probably want to go here https://discord.com/channels/1002292111942635562/1002602742667280404

[hazy warren]

😮

[jovial tiger]

that's probably actual stablity people

[jovial tiger]

you'll probably want to go here https://discord.com/channels/1002292111942635562/1002602742667280404

[wise field]

Holy Hell who has 1000 dolalrs to spend on that shit a month

[wispy nest]

😐

https://cdn.discordapp.com/attachments/1230799113197457461/1230821630331060284/image.png?ex=66258db8&is=66243c38&hm=21b95c2c3f29099e0a920925f2d03c367f75299c26920931852e5de11a05288a&

https://cdn.discordapp.com/attachments/1230799113197457461/1230805361909305344/image.png?ex=6634a7d1&is=662232d1&hm=379802da1e523c875304cffb974eee69ea69d9e6f17e2c33727a694568c22ed7&

[deft bison]

Thank you for the information... 👍

[jovial tiger]

It's stupid fast. 6 images in 10-15 seconds most of the time, at sd3 quality which will only increase over time. it's a pretty great way to supplement local rendering.

[nimble mason]

they don't exactly make it sound like a great deal by just saying 4090... is it a bank of them?

otherwise it's just like... shit, buy 6 of them and you'll break even in a year

[jovial tiger]

lykon has mentioned that his workflow is a lot slower than what's on the api... so i'm not expecting sd3 to be fast even with a 4090

[nimble mason]

guess it's good i'm getting myself acclimated to 400-500 step workflows lol

[jovial tiger]

lots of green border statuses dancing around the screen

[hazy warren]

Fallout!

[crisp stream]

though I at least assume I know what you mean, dunno 😀

[hazy warren]

Needs to be more Fallout

[crisp stream]

Funk Fallout 😀

[deft bison]

It's powerful when you think about the damage that these giant and deadly waves will cause! ...I think it was a great success! ... 👍

[hazy warren]

just a scratch

[deft bison]

😀

[wary adder]

Hello spirits

[velvet briar]

Sup yall, i just recently made this art piece using juggernaut xl model. I call it :

THE SPHERE

What i did was, basically generate a bunch of images using that model of an assortment of household objects. I then print it on some thick paper, cut each individual object, and then glue them together on an A3 paper to create the sphere

here's some proof

lemme know what you guys think

[hazy warren]

time for some pixel art

[languid pebble]

Pretty detailed ...

The bad Wizard ...

[hazy warren]

https://media.discordapp.net/attachments/194262303183601673/1231529407458578513/00032-3482591988.png?ex=66374a23&is=6624d523&hm=648c3846b2ca18a7096aa49e72df5c69354700944197f01865f0a73f4b1c05a7&

[sterile kiln]

a magician never reveals his tricks 🙂

[primal ermine]

child green dress

#🌠｜show-and-tell girl

[jovial tiger]

https://x.com/tedtalks/status/1781351036877156452?s=61

TED Talks (@TEDTalks) on X

What will TED look like in 40 years? For #TED2024, we worked with artist @PaulTrillo and @OpenAI to create this exclusive video using Sora, their unreleased text-to-video model. Stay tuned for more groundbreaking AI — coming soon to https://t.co/YLcO5Ju923!

Twitter · Apr 19, 2024, 3:55 PM

[quartz ravine]

Very cool, I like to see it

[jovial tiger]

I don't think I've ever seen waves crashing down on something... SD or some other service?

[uneven cypress]

how can we create these images?

[languid pebble]

2001 Space Odyssey: Good morning coffee

2001 Space Odyssey: Trouble starts ...

2001 Space Odyssey: A romantic story coming?

2001 Space Odyssey: But we need an end fight first ...

Happy end ...

[junior robin]

This is insane

[weary urchin]

@late sorrel @dense nova

[dense nova]

LOL you already have A1111! You don't even need anything. 🙂

[weary urchin]

BUT HOW DO I DO THE STUFF

i type stuff in and it turns out garbage

[dense nova]

Let's see...

[weary urchin]

ok lemme give u an example

im gonna copy a promt thing now

and show u what i get

[late sorrel]

ok im happy we are here, so we can skip some instructions lol

[dense nova]

Try this prompt: "desktop cat". That's the whole prompt.

[weary urchin]

ok

its loading

[late sorrel]

he is using magicmix, isnt that one of those anime ones?

[weary urchin]

i thought it was a realistic one for asians

i got these

[late sorrel]

so you got a lot of stuff haha

[late sorrel]

nice

[weary urchin]

it all turned out suckish so i wanted a good one

[dense nova]

Your setup is fine.

You just need a better model.

[late sorrel]

yea better model and settings

[weary urchin]

like what model

[dense nova]

Settings are fine.

[weary urchin]

and what do i to the settings

oh

[late sorrel]

can you run sdxl? cause that will give you better quality stuff

[dense nova]

If you're sticking to SD1.5, RevAnimated and RealisticVision are the best ones I know of, but I've been using SDXL for ages now.

[weary urchin]

probably if i knew what that is

[dense nova]

What are your PC's specs?

How much VRAM? Assuming an NVidia GPU.

[weary urchin]

how do i check again

[late sorrel]

well where did you get the models from? use civitai website, you can even filter by model type, lora, etc

[weary urchin]

thats the gpu

memory

cpu

[dense nova]

You just cropped it off of the pic you posted. It's right under the render info.

(On mine says A:10.43GB, R:14.15GB, Sys...)

[weary urchin]

where is it

[late sorrel]

you dont always need a high cfg scale btw, it tends to cook pics, i usually like to start with 5

[weary urchin]

oh

found it

[dense nova]

16GB VRAM. You can run SDXL no problem.

[weary urchin]

ah

[weary urchin]

how i do that

[dense nova]

Just download a model from Civitai like you did for the other models you have.
https://civitai.com/models/363210/crystal-clear-xl-prime

[weary urchin]

kk

[dense nova]

Not sure what version all the cool kids are using though...

[late sorrel]

get a general xl model like dreamshaper, or juggernaut, etc

[weary urchin]

yeah i got both already

[late sorrel]

nice

so you are set man

[dense nova]

No, he's talking about SD1.5

[weary urchin]

these are the ones i got

i just want to be able to generate like actually realistic looking images

[late sorrel]

well he has juggernaut xl i see

and dreamshaper

[dense nova]

See how the one on the left says "XL" in the title? The one on the right is bad quality SD1.5.

[weary urchin]

oo

why

[dense nova]

Because it's older.

[late sorrel]

no but he has xl version already

[weary urchin]

ic

[weary urchin]

i can just downlaod the normal one

[late sorrel]

so you want lower quality?

[dense nova]

No, he's right. You have DreamshaperXL in your list.

[weary urchin]

oh

yeah i got that

[late sorrel]

you literally already have both dreamshaper and juggernaut xl

you can use them

[weary urchin]

yeah but the pics turn out garbage

[late sorrel]

ok but let's go step by step and try to fix that

[weary urchin]

lemme generate one

[dense nova]

Yeah, now your settings are the problem! Easy fix.

[weary urchin]

ill show u

[weary urchin]

yay

[late sorrel]

there is also a possibility your VAE is bad

[weary urchin]

whats a vae

ok like here

[dense nova]

VAE should be automatic. But try the settings I just posted, and share the results.

[dense nova]

Wrong resolution. Try the settings I dropped.

[weary urchin]

what should i prompt

[late sorrel]

you are using low resolution

[weary urchin]

kk

[dense nova]

"Desktop cat". Or anything quick.

[late sorrel]

he wants some beautiful asian woman ok? :3

[dense nova]

"beautiful asian woman" then

[weary urchin]

damn right

[dense nova]

How'd it turn out?

[weary urchin]

uhh

its stuck on 0%

[dense nova]

Please don't crop. What's the error???

[dense nova]

The console error you cropped off.

[dense nova]

Hmm.

[late sorrel]

out of memory

[dense nova]

Hang on. I know the problem. Just looking for instructions to give you.

[wild sorrel]

well,
There is not enough GPU video memory available!

[weary urchin]

i thought i had enough lol

[late sorrel]

you prob need to unload the model before using another one

[dense nova]

Just need a low-vram launch flag. Takes two seconds. But where do you put it again...

[late sorrel]

but he has 16gb vram no?

[weary urchin]

oh ik

nvm

i frogot

[dense nova]

Me too. 😦

[wild sorrel]

Which gpu are you using?

[weary urchin]

i remember setting it to midvram

[weary urchin]

6950 XT

[late sorrel]

im not familiar with amd ones

[dense nova]

Oh LOL I'm in the wrong folder. Hang on.

[weary urchin]

xD

[dense nova]

Edit "webui-user.bat"

[wild sorrel]

he shouldn't need to change to low\mid vram if he has 16gb

[late sorrel]

but wait why you telling him to change to low vram

[weary urchin]

yh here

found it

[weary urchin]

yeah xD

[late sorrel]

this is simply a problem i think because he didnt unload the model before using another one

[dense nova]

Okay... Well, let's try anyway.
Flag: --lowvram

[weary urchin]

how i do that

[dense nova]

Doesn't it default to just 1 model in VRAM?

[weary urchin]

in front of the direct ml thing?

[late sorrel]

well im not sure how automatic1111 handles that

[weary urchin]

like that?

[dense nova]

Yep. Now exit the console window, and open it again by double-clicking webui-user.bat.

[weary urchin]

kk

[wild sorrel]

I don't think that's good idea...issue somewhere else, 16gb should be enough for most models, including sdxl-based

[late sorrel]

yea there is no need for lowvram i think

[dense nova]

We'll check settings in a second. Already found it:

First, let's just prove we can get an image onscreen.

[wild sorrel]

either memory leak (just reload console and webui) or settings or install issue

[weary urchin]

generating

r those prompt things fine?

[dense nova]

Wrong model LOL.

[weary urchin]

oh which one should i use

[dense nova]

Anything with "XL" in the name.

Like DreamshaperXL I think was in your list.

[weary urchin]

kk

[dense nova]

Console?

[wild sorrel]

reload webui

close the console, launch the thing again

[weary urchin]

ok

just staying on 0% again

WAIT

LOADING

ok nvm

it died

[dense nova]

Okay. What's the directml flag? Do you have an AMD GPU?

[weary urchin]

yes its amd

[weary urchin]

how do i check

[dense nova]

Hang on.

Keyboard Ctrl+Shift+Esc.

[weary urchin]

yep

[dense nova]

Click "Performance" in the column on the left.

[weary urchin]

yep

[dense nova]

Click "GPU 0" in the new column that shows up.

[weary urchin]

yup

[dense nova]

AMD Radeon RX 6950 XT...

[weary urchin]

yeah

[dense nova]

I've never dealt with it before but I know it's a pain...

[weary urchin]

o ok

[junior sky]

15.8gb used? A1111 is such a piece of crap 🤣

[weary urchin]

dosent work

git config --global --add safe.directory C:/stable-diffusion/stable-diffusion-webui-directml
Python 3.10.6 (tags/v3.10.6:9c7b4bd, Aug 1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]
Version: 1.9.0
Commit hash: <none>
Traceback (most recent call last):
File "C:\stable-diffusion\stable-diffusion-webui-directml\launch.py", line 48, in <module>
main()
File "C:\stable-diffusion\stable-diffusion-webui-directml\launch.py", line 39, in main
prepare_environment()
File "C:\stable-diffusion\stable-diffusion-webui-directml\modules\launch_utils.py", line 593, in prepare_environment
raise RuntimeError(
RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check
Press any key to continue . . .

[wild sorrel]

are you sure you installed automatic1111 specifically for amd GPU's?

[weary urchin]

no

yeah i thought i needed it to bypass the gpu thing

cuz i dont have cuda cores

[wild sorrel]

hmmm...

[long folio]

Is there a way to take a style from an image (drawn) and apply it onto a photo (real)? I have the A1111 webui and ComfyUI webui, controlnet

[wild sorrel]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs

[dense nova]

I think this is what he has and it's right. https://github.com/lshqqytiger/stable-diffusion-webui-directml/

[weary urchin]