🆕｜sd3

[low stone]

I have automation going to pass it images and have it describe it for image prompting. I did gpt4 vision compared to gpt4o and while vision was already really good, this new gpt4o is seriously good. Able to describe things in as short a sentence as possible which is obviously one of the most important factors.

[stoic turtle]

Automation the car company tycoon game?

pog

[severe phoenix]

yeah i'm using it that way too, its really better than gpt4 in terms of descriptions

[low stone]

Retro urban photography, vibrant neon sign reading 'Torcello's Art!' glowing in red and yellow against a clear blue sky, classic motel buildings with red roofs in the background, palm trees providing contrast, sharp shadows indicating midday sun, desolate street emphasizing a nostalgic feel, 70s color grading for a vintage aesthetic.

So it was able to read the text of the larger part, although not the clip drop piece.

[severe phoenix]

do you have paid gpt4o?

[low stone]

Yeah I'm using it through the paid api endpoint.

[raven fern]

kek

[teal fossil]

Depends on how trainable it'll be. If it can pick up concepts way better than XL etc.

[dull star]

I have different intentions

[dreamy sundial]

better than sd3 is debatable, i've ran some tests and it seems to be really similar, finetuned sd3 will be at the same level or even better

[queen spindle]

how do i create image'

[stoic turtle]

pencil

paper

draw

[queen spindle]

??

broo

ai image

how is that done

[low stone]

Click this https://discord.com/channels/1002292111942635562/1237461679286128730

[little quarry]

2 weeks

[lofty laurel]

I prefer watching videos, but now I'm tired of watching pictures.

Good things are also used.You won't see bad things anymore.Poor things can no longer be seen, who told them to send a video and send 1080P P.Such a high-definition video is still a long video, which makes me feel bad when I see the pictures now.

Are you all Asians? It is reasonable to say that new york time is only now.It's 4:00 AM

PM

[hallow lion]

We are all Asians.

[civic egret]

All of us.

[severe phoenix]

have you tried the img2img thingy where you can turn imgs to carcatures and upload a character thatcan be used to tell consistent stories

[low stone]

i think you have to use the full app for that. i don't have plus, i just have the api keys so i probably can't use that.

[severe phoenix]

ohh ok, mahn i wanna try that so bad

[low stone]

ella / pixart / sd3

In a hyper-realistic, cinematic style with a shallow depth of field and warm, golden lighting, a massively oversized CPU and motherboard sprawl before us, their intricate circuits and components gleaming amidst a sea of tiny, furry creatures - rabbits, mice, and squirrels - that swim through a flood of glowing, iridescent "memory leaks" like a digital tidal wave, as shocked and disbelieving onlookers, their faces illuminated only by the soft glow of nearby diagnostics screens, stumble backward in awe and horror.

ella vs sd3

so this is with a much shortened prompt. maybe that's why sd3 did so much better.

Gigantic cybernetic squirrel with glowing eyes emerging from a mechanical gateway flanked by screens displaying underwater scenes, with a crowd of people observing, amidst intertwining wires, in soft ambient light.

[cunning lintel]

SD3 definitely prefer the prompts short, the longer the prompt the loser the interpretation it seems

[low stone]

i had gpt4o rewrite the prompt and now sd3 is way better.

[bitter hearth]

Yes

Sd3, doesn't, need, this, too

And every single type of camera settings in the prompt x3

Makes me go insane

[cunning lintel]

i used to be telling llama "compress" but "bring down to its essence" works better,, now i need to figure out how to make my system prompt do that by default without filling 1k tokens with examples 🙂

[low stone]

I've been having pretty good look with brief and terse and limit to 5 or 10 descriptive elements. It's hit or miss.

[lucid swift]

https://fxtwitter.com/gdb/status/1790869434174746805 do you think that gpt4o generates vae(image) , audio, and text tokes with the llm? all in one thing?

[odd basalt]

Festival man here’s ur prompt in chatgpt4o

[upper snow]

the sd3 one is much more atmospheric, this squirrel doesn't look very cybernetic and the brightness is very offputting

[restive halo]

allegedly it still asks dall-e for it and they haven't enabled the direct gpt-4o images in the UI yet

same for voice and whisper

[vapid radish]

I see what you mean!

[vapid radish]

I have found GTP 4o is a great prompt interrogator, if you have an image you want to reproduce and don't have a prompt for it , it gets so close when you re-generate it with Dall.e 3 somehow.

[restive halo]

I think 4o might have the image to tokens part in already, just not the generate image from it part in the UI yet

but not sure, they've been very unclear about it

[low stone]

Gpt4o really is incredible as a vision model. I put in this stuffed animal and dog image from YouTube, got this prompt, and made the image on the right with sd3. Amazing stuff. The prompt: A vibrant, illustrative style with detailed textures and warm lighting shows a golden retriever with soft, expressive eyes and velvety fur, being petted while a fluffy stuffed tiger with bold black stripes is held nearby, office setting with people at desks, modern chairs, computer screens, cables under desks, carpeted floor, distant shelving with various objects, colorful post-it notes on partition walls, and soft ambient indoor lighting enhancing the scene's warmth.

[vapid radish]

Yeah I was trying to recreate this without a prompt and got the 2nd image, which is amazing for something so abstract!

[low stone]

I would even say the new one is better than the original

[restive halo]

is it actually better at giving you a similar prompt to generate a similar image than say using an ipadapter instead

[vapid radish]

another example, but I changed the prompt to add the arrows.

[low stone]

Hah I was about to say and then you can change it up easily but you're way ahead of me.

[severe phoenix]

i'm about to test this using movie screenshots to see really how good it is at description

[low stone]

Hyper-realistic, vibrant, wide-angle shot of a massive, chaotic desert dystopian vehicle convoy with a fire-spewing guitarist atop a towering speaker-laden truck, surrounded by armored cars and motorcycles kicking up clouds of sand, under the harsh midday sun casting sharp shadows and flaring highlights.

One on right is Ella. Sd3 didn't do well with it but that just needs fine tuning.

[vapid radish]

It seems to be trained on movies as well.

[severe phoenix]

crazzy, can i see sd3 version

[severe phoenix]

ohh wow i never thought of doing this too. incredible

[low stone]

Sdv made a neat animation of it though.

[sullen moss]

https://the-decoder.com/stability-ai-reportedly-runs-out-of-money-and-in-negotiations-to-sell/

[oak pier]

release sd3 before you go out of business please

also music ai xD

[restive halo]

theyve said multiple times they will

[idle current]

As expected

[idle current]

Or they save it for a potential buyer that will own kt

It

And then offer the product with a price tag

[oak pier]

yeah sd not worth a price tag

[severe phoenix]

oh yeah, sd3 is not that great here

[low stone]

Exactly this. There's no way they're going to release it when a potential buyer will probably want to monetize it. Sigh.

[oak pier]

only real selling point of sd is being able to train anything on it xD

[severe phoenix]

its still confusing to me how stability didnt even try to have a leading ai gen site like leonardo nd playground. its mind boogling to me. they should be leading in that space right now.

[oak pier]

if monetize sd just become a worse midjourney xD

[sullen moss]

I have exactly the same suspicions

[low stone]

The problem is that as awesome as sd3 is for us, it's far inferior to almost every other big commercial image creator out there. With these new announcements by meta, gpt4o, imagen from Google, they're already better than sd3.

[restive halo]

not only at all, the tooling around it as well

[low stone]

So where's the value in monetizing it? Sigh getting bigger.

[oak pier]

tools disapper if it become a paid platform

people stop investing in to it xD

[restive halo]

literally none of them are better if you want to do complex stuff

for pure text to image sure

[low stone]

Dude, did you see the new demos for gpt4o? It's an another planet compared to sd3

[severe phoenix]

meta got an nnouncement too? gaddamn when?

[restive halo]

sure, and you cant generate say a gif with my face in it

[oak pier]

i did not i wll have look at that later

time for work for me 🌊

[low stone]

When they released llama3, they also released their image creator. Does lightning fast amazingly prompt following images and even does short animations of them. Only problem is that it's highly censored.

[ancient cape]

dalle3 actually gets the job done 80% of the time, on the first generation attempt.
be it something for work, something for marketing, something for dnd, for a one-time-use meme, or background assets when i design printable sheet

[oak pier]

llama3 is great i enjoy it

[low stone]

Gpt4o can absolutely do that.

[ancient cape]

sdxl + loras usually take me like 10~30min, to get the same value or a bit better than dalle3 in seconds

[severe phoenix]

very true, imagetoimge nd model finetuning is sd's biggest leg up. i cant imagine what this img gen space would've been without them honestly

[ancient cape]

tried sd3 via the api... and uff... I wept

[severe phoenix]

ohh yeah i saw that, i thought you meant they announced sthng like yesterday or sthng.

[restive halo]

what demo have they shown that can do that? or most stuff you use say controlnets for

[ancient cape]

I really really hope the final model is leagues better

same for a bunch of other finetuners I know. they also all tried sd3 by buying like 10$ of credits... and our experiences were sadly all the same 😦

[low stone]

Have to get to work, but have a look at the demos they put out in the last day or so. Full image upload with manipulation

[restive halo]

what I really hope is that the tooling, especially controlnets are better for sd3, it's so sad that sdxl has worse tooling than 1.5

[noble coyote]

<<<Splutter!!!>>>

[restive halo]

'image' yes, which is why I said gif

[low stone]

Googles new stuff can do sora like videos. You think it can't make gifs? A lot has changed in the last 2 days.

[restive halo]

I think it can make gifs... but not with my face on them

which is why I said 'say a gif with my face in it'

none of those so far look like they come with the kind of control and tooling we have with SD for complex workflows

also google's video model looked kind of bad? the videos weren't that great

not that I could signup anyway, Europe isn't allowed to

[sullen moss]

Unfortunately, for now, open source is lagging behind, and at the moment, there are no circumstances that could change that.

[noble coyote]

Cash will win this race!

I've been with software companies whose original business model was "free updates for life!" And it was great - while it lasted. Topazlabs were then hit by the development of AI - so had to renege of the F4L business model - lost a lot of goodwill, and customers; but they had to do it! They're still trading!

FLStudio also does free-upgrades-for-life - but their code is so old! They do not have the income resources to afford a complete rewrite; but I am grateful for over 21 years of free upgrades!

I feel the Stability AI Community should Crowdfund SAI: we'd all be stakeholders, ensuring a continuously affordable product ...

[low stone]

Now that I think about your request, you're asking for it to make deepfakes which they'll never do.

You could do that face swap locally though and then have them animate it

[idle current]

More than that

Although Midjourney is not in the best spot either

But they are keeping themselves above the water for now

But they also got millions of people and some companies paying subscription to them

And they dont have to deal with other expenses in comparison to SAI

[idle current]

Unrealistic

And also not enough

If you are at a dept of 100 million dollars, have a current year quarter bill of 30 million while having a revenue of 5 million...that aint gonna work out

And the reaction in the community says it all as well

[low stone]

Yeah I'm seeing more articles and reactions today on Twitter. It's really bad.

[idle current]

its almost like it was predestined to end this way

basically the game was rigged from the start

it was borrowed time

and the lawsuits are not to be ignored either

i mean i know a lot of people here laughed about the lawsuits against Stability AI, Midjourney and DeviantArt, but those can have big consequences even if the lawsuits dont end up very satisfying for the people that did sue

[low stone]

Do I believe stability would win the lawsuits? Most likely. Do I believe they have the money to fight it properly, doubtful.

[idle current]

open ended book as of now

but the lawsuits might have had consequences already

think of the customers and especially companies that refuse to become customers because of the risks

"who cares about companies, we care about our OS community here!" thats what some ppl think and this is such a stupid one as well

[desert garnet]

like what companies

[idle current]

Coca Cola for example

i mean they are using Adobe one but just an example

[desert garnet]

right i remember when nvidia and intel invested in SAI

they must be quackin on their boots rn

[idle current]

Nvidia and Intel arent its customers tho, how much and what did they invest

[desert garnet]

intel like 20million

[idle current]

ah okay, that was the one that was mostly computing power they sent

forgot about Nvidia in detail

but thats one time investment simply isnt enough. You need actual long term partners and customers

[desert garnet]

yes and SAI couldnt get them because they released the weights for free + their models are not very user friendly, midjourney and openai prob have some sort of software to help big companies generate stuff easily

[idle current]

well yeah they want to get investment back

Midjourney doesnt offer more than their Discord stuff + recently also their web UI. OpenAI on the other hand does more, yeah

OpenAI, Google and Adobe

and Microsoft

[bitter hearth]

steven talking

[desert garnet]

https://tenor.com/view/yapping-yapper-yapping-man-old-guy-yapping-keep-yapping-gif-16586719177455162521

[cunning lintel]

Not really. I don't see why open-weights with hosted services and commercial restrictions don't work. But seeing what SAI has produced, the llms, the sd3d thingy, the audio, it never caught on, and competitors that did specialize outperformed each. Can't help but feel SAI just squandered their resources, look what pixart put out with minimal resources but by focusing on one thing only, or the recent chinese PoCs. Apart from that all recent news screams mismanagement and frivolity, it might have been the attitude that got us stable diffusion, but now it's what makes investors walk away. Over promise/hype under deliver might get you attention, but you do antagonize your investors really really badly when you can't deliver.

[low stone]

It didn't do what I wanted, but it made some neat pictures.

[idle current]

yeah but just look at the reaction of people whenever pricing is brought up

a bunch of people would rather watch SAI die as long as SD3 gets released than donating money or paying for service to keep the development going etc.

although its unrealistic anyway in this case

[low stone]

That's a lot of Reddit too. I've raced to do the sd3 api etc because I want to support them where possible. Even people like datavoid on Twitter who heavily uses stability products hasn't hesitated to say #notsd3 and down with stability when they mentioned that sd3 would be downloadable with just an email sign up. That was a step too far apparently. Sigh.

[cunning lintel]

That's just shouting on social media, another thing altogether.

[low stone]

Well, you find out who your friends are when you're not in good shape. If the people who are only where they are today because of stability immediately go to abandon them....

[cunning lintel]

Though I myself would never just donate (in the sense of gift) money to SAI as is either. If it goes that way it should be a proper non profit with a solid governance structure.

[idle current]

selfishness

people get one finger, want the whole hand but refuse to give anything back

profit amongst all

and then they shit around against corporations how greedy those are lol

the irony

i said often enough, "the people" arent better than corporations (besides that its not all black and white anyway)

[cunning lintel]

Yeah true, that's madness. I just meant the "popular opinion" of random reddit users

[noble coyote]

SD3@ClipDrop

SD3@ClipDrop

SD3@ClipDrop

SD3@ClipDrop

[cunning lintel]

It's ironic, the main use for SD3s text capabilities is memes about SD3/SAI 🤡

[gusty trail]

Gpt4o text capabilities is way more advance than others. I think it might used some text planing pipeline to allocate the text area.

[keen falcon]

GPT-4o vision is from other planet

[gusty trail]

Agree, like an alien tech

[cunning lintel]

"One model to rule them all" great, but how can opensource ever catch up to that, how can you even run inference such a model locally. Yet it seems the way forward.

[idle current]

unless the production cost gets lower i dont see that happening

[bitter hearth]

no, shrek is the main use for sd3

[bitter hearth]

What they showed is indeed from another level xD

the open ai logo with text in it, I'd believe it was really generating text in image and not placing with some technique

This example I mean

[gusty trail]

The closest thing I know is this. https://huggingface.co/spaces/modelscope/AnyText

It is like a controlnet to guide the generation

[dull star]

well thankfully SD3 will be either released or leaked anyhow

only ethical concern is that many people will lose their jobs

[bitter hearth]

well

it is what it is

[robust junco]

Where can a donation be made?

[sterile pendant]

By simply paying for a membership, even if you don't actually need it. I'm sure if you contact them, they likely have a means for just donating to them as well.

[robust junco]

contact them here on the server?

[dusky thistle]

i wouldn't count on a leak

[dull star]

I would

[dusky thistle]

that's their most valuable IP (as far as we know) and any buyer would want that carefully guarded

there's a reason we don't have weights for dalle or know what the MJ pipeline is

[hallow lion]

Nosumers unite.

[cosmic pelican]

Sell it to X lol

So far, do you guys like SD3 or MJ v6?

[tropic aspen]

But then again, those companies didn't outsource their models to other companies

And are not on the verge of bankruptcy

I do agree though that it still may not be leaked due to legal issues

[idle current]

Dont use either of them anymore but if i had to choose it would be MJ i guess

[dusky thistle]

most ppl aren't going to be willing to risk their careers for an AI image model leak

[cosmic pelican]

Why you don’t use any of them anymore? 🤔

[tropic aspen]

I feel like the only reason ppl believe that is because of NovelAI, but the only reason that leaked was because 4chan managed to crack their GitHub accounts (aka, figured out that the password was "password")

[idle current]

One doesnt fit my workflow and current situation amongst all, the other one i dont want to pay for

[little quarry]

2 more weeks?

[cosmic pelican]

oh ok

[tropic aspen]

Ig also some of the devs said that as well, but again that's prob basing it off NovelAI's leak

[idle current]

I would have been genuinely curios tho how SD3 compares to others and to what i use

[cosmic pelican]

SD3 is claimed to be better than Midjourney v6, but I haven’t seen much comparison done.

[idle current]

In terms of quality and composition?

Basic model?

[cosmic pelican]

Quality and composition

The claim is on their site I think.

[idle current]

Would have to see that and more importantly comparison to DALL E

[cosmic pelican]

I’m assuming you use DALL E?

[idle current]

People complained that in reality it didnt turn out that well

[idle current]

To supplement Adobe Firefly/Photoshop

Or vice versa depending on how you look at it and usecase

[cosmic pelican]

Oof

[idle current]

And no i dont pay for D3 ^^

[cosmic pelican]

That’s a good idea

[vapid radish]

SDXL 0.9 also leaked early.

[idle current]

Yeah and the same approach i have with creative software and workflow i have in doing art and gamedev

[dull star]

SD3 can be aesthetic, but nowhere near the quality of midjourney

prompt adherence though

[low stone]

MJ is really good. It can't do what sd3 can do, but what it does do, it does very well. I use all of it. 🙂 I usually tend towards fhe stuff that doesn't censor what I ask for, which isn't that edgy, just that many of the services are overly tight

[low stone]

I feel like anything sd3 can't do is just because it needs community finetunes, not that the engine can't do it

[cosmic pelican]

🤔

[bitter hearth]

you guys see what 4o image gen can do? it is wild. can even make its own fonts. not release just yet though

[idle current]

Well in art professionals generally have a whole pipeline and ecosystem of tools they use, rarely do you find people thst work on one single package

[dull star]

SD3 is still the best in text, knowledge and prompt adherence out of all OPEN models

[idle current]

This can apply to genAI too

[dull star]

yeah SD3 finetunes would be SUPER sick

[bitter hearth]

dalle will be over, 4o has a built in image maker

[dull star]

but even as a base it can do decent aesthetic images

I just hate bokeh 24/7

[idle current]

Wdym

[dull star]

gpt4o's image generator is really OP

from what I've seen

(like 3-4 images lmao)

[cunning lintel]

MJ seems the lowest bar, MJ is just aesthetics. It's Dalle-3 and Ideogram (and probably the new gtpt4-40, google and meta image gens) that are the real benchmark

[low stone]

I also saw that sd3 is better than all of these. Oh it's not released yet, so I can't show it off. 🙂

[dull star]

I hope finetuned SD3 will become Ideogram quality

ideogram is the perfect meme maker in my experience

SD3 doesn't perform as well

[bitter hearth]

no chance

[dull star]

but for open standards, SD3 is still the best

[bitter hearth]

gpt4o is some multi modal omni model and built ground up with image, text, video creation things. they show off an example that was wild

[low stone]

I took ideogram and dalle images and made a Lora of robots punching their fist through buildings. Now sdxl can punch through buildings. Same will be with sd3.

[bitter hearth]

but right now it does not have the image mker turn on yet

[dull star]

I hope, with a much larger scale (so like 3038712847192874 more actions and examples)

[idle current]

you need to consider the usecase tho

[low stone]

I was joking about boasting of models that haven't been released. 🙂

[bitter hearth]

i have use sd3 at my cousins place, he buy some credits. it is just another stable meh release i think. impressive tech but it really lack too, like all they put out xl was over hype so will sd3 imo

[idle current]

the biggest advantage Stable Diffusion has over its competitors is the customization

[dull star]

SD3 will probably do the types of memes ideogram can if I would train a lora on it

its simple really

a synthetic dataset could be easily made even

[cunning lintel]

Still can't help but wonder what a fully trained SD3 looks like, i'd expect that does away with halfway ending limbs, the horrid hands and such

[bitter hearth]

here is the one image they tease with 4o, which will also be able to create you own font. look at that image and also the text unreal imo: https://twitter.com/gdb/status/1790869434174746805?s=46

Greg Brockman (@gdb) on X

A GPT-4o generated image — so much to explore with GPT-4o's image generation capabilities alone. Team is working hard to bring those to the world.

Twitter · May 15, 2024, 10:18 PM

[dull star]

I would simply expect slightly (yes, only slightly) less hallucinations

[bitter hearth]

4o going to leave the rest in the dust at release and it should be in a few week time

[idle current]

i guess someone will bring SD3 to Photoshop via that A111 plugin as well

[cosmic pelican]

i’m banned from gpt, so im downvoting it 🤣

[bitter hearth]

haha

[dull star]

I'm gonna make colour blobs and use them in SD3 img2img

kek

[cunning lintel]

Well, maybe just that, but there's also talk about looking at the text encoder, at the least, maybe longer prompts, at best, better understanding. But that's not so much aesthetics.

Time will tell, or not

[bitter hearth]

i really did not see any difference between sd3 and xl when it was hype. just my opinion of course and still impressive

[dull star]

if they somehow uncap the prompt length (if they have the non truncated prompts) and train further with soejmthing like longclip

I wonder if that would make a change

[cunning lintel]

Also curious what our ruskie friends are brewing, kandinsky was always "just not there" but an sd3 clone just not there would be pretty good :p

[dull star]

we need longclip-g as well

[low stone]

Dalle is one of the most censored models out there. If gpt4o is too, it's not of much use to me.

[bitter hearth]

dalle provide what i need, i am not after some gore or sex stuffs

but yes i am sure 4o will not be open like sd3 could be

[dull star]

I'm after gore if I'm trying to do zombies or war depictions or artistic violence, so SD3 will do just fine

[low stone]

When I was grabbing the images for the Lora, and I wanted a soldier carrying an ar-15, it denied the majority of images. I kept hammering away and got it to give me enough of them. That's not gore or sex.

[cunning lintel]

Right now SD3 is more problematic than Dalle to me 😂 I just gave up prompting females, it always ends in blurrs, though i get the point, sd3 without that filter is less restrictive than dalle 🙂

[idle current]

im hoping for a soonish release of some of the SD available features to Adobe Firefly and the ecosystem

[bitter hearth]

well sam altman say they are going to unleash more of that kind of thing too. who knows. they have to have guard rail of course because they are the first mover in all this, to most people who know ai, they of chatgpt, so they will also get all the bad press too if people overeact

[idle current]

where do you use SD? Clipdrop?

[mortal mesa]

i do not still wear the first pair of socks ive bought

[bitter hearth]

look at the global panic when someone make some taylor swifts haha. they cant have that kind of thing

[cunning lintel]

started with the api, then clipdrop, now mostly pixeldojo (though they have fake delays, it is advertised as unlimited)

[bitter hearth]

any of you guys ever use meta image maker? i am curious how that is

[low stone]

It's really good. Couldn't do text but it's great for other stuff. Only does square images though which kinda blows

[cosmic pelican]

Do you guys think weighted models will be outphased by cloud based?

[idle current]

oooh okay

[bitter hearth]

square image is good some time too. i got a co pilot pro subscription once and it only would make widescreen dalles haha. nice but some time i want the option of square too

[idle current]

what do you mean?

[bitter hearth]

depend on the tech. we probably could not run dalle3 or 4o model or whatever on a home computer

[idle current]

i dont have issues with those usually, if any i can outpaint them with ease

[bitter hearth]

in time i suppose a home computer tech will be cheap enough and a model advance enough to run on a phone. who would think 20 years ago we could have the compute power we do in our cell phones

[cosmic pelican]

I mean instead of installing a webui, and using it locally on a device, everything will be moved to a cloud server where it will be updated routinely

[idle current]

oh yeah i suppose so

but that also means people gotta pay

for the service

[bitter hearth]

has to be some profit motive for a company to make a local model of course. stable is/was rare but they still were getting funds not do it for charity

[idle current]

which is where SAI failed at

[bitter hearth]

a lot of people seem to think everything should be free, but why would a company just give away a free model you know

[idle current]

people are very selfish

[mortal mesa]

to get buisness

[cosmic pelican]

welcome to subscription era lol

[idle current]

it pays out xD

[cosmic pelican]

this

[idle current]

if i actually had to buy my whole software pipeline at once i couldnt afford it at all

i would have to sell my family xD

companies know this

and they offer subscriptions

[mortal mesa]

or free models

[idle current]

rarely

[mortal mesa]

to get corporate buisness

[idle current]

SAI is exception

but SAI also has interests

[cunning lintel]

There really aren't that many people able to use those big models at home. Just need to limit commercial use (which SAI now does) so a part of that profit flows back to you. (then you have the best of both worlds, generate income, and "free" research using your models)

[idle current]

Stability AI isnt a charity organization

[cosmic pelican]

tbh i prefer cloud based where it’s $1 a month lol

[bitter hearth]

and i do not think SAI is at top of the game either, im not sure how well their approach work for them

[idle current]

never was

[mortal mesa]

realistically $$$ wins, look at Grok, OpenAI pissed off Elon and Poof Grok was made

[idle current]

at the top of the game are the corporations which large ressources and influence

Google, Meta, Adobe

not even OpenAI is on their level

oh and Microsoft

[mortal mesa]

alibaba

[bitter hearth]

nvidia could release some big thing too if they want.

[idle current]

and then there is Nvidia

Nvidia is the hidden master lol

[mortal mesa]

there are players for sure

[idle current]

they rather put their technoly underneath the products of the big boys

technology*

[cosmic pelican]

A while back ago, Elon Musk was in discussion of a partnership with MJ. I wonder if that’s still ongoing, or if Stable Diffusion is the new canidate, or even canidate.

[bitter hearth]

yes and seem to pay off big time for them for now

[idle current]

exactly and even those boys are compromising on their products

to ot overspend

not

[idle current]

Midjourney or OpenAI?

[bitter hearth]

look how far we have come in just... idk was dalle 2 or stable 1.4 even two years ago? in another year maybe even we will probably be close to 1 shot image creation where it is perfected. then it will just be a matter of the person pick which company they like best as the image creators will all be near perfect

[dull star]

yeah the api censorship is super aggressive

[idle current]

i dont believe in that tbh

[mortal mesa]

no one will be talking about image generation cuz all the models will be able to, will be talking about its other multimodel capabilities

[bitter hearth]

yes kagi. we will be onto video haha

[mortal mesa]

pieces of the puzzle

[hallow lion]

Anything online is utter trash compared to SD. I don;t even know why ayone would even use them other than being uninformed or lazy.

[bitter hearth]

i prefer dalle for my needs i think it much superior model

[idle current]

bullsh*t Dodge

[hallow lion]

0 control and trahs quality + sencorship

why anyone even woudl consider using dalle or MJ is beyond me

[idle current]

you are talking purely from ideology

[bitter hearth]

trash quality i would say is more sd even sd3

[hallow lion]

peoplea r eclueless

[idle current]

that doesnt make it true

i think you are clueless one here

[hallow lion]

MY and dalle are junk compared to SD

prove me wrong

[cosmic pelican]

I just want a model where I don’t need to keep installing 😂

[idle current]

you have no idea

[cunning lintel]

Does anyone really think this looks good Yes text is impressive, the board itself and hands though, awful. If this is wat photo-realistic looks like, it's not that great. I feel often we're wowed by certain aspects of imagen, while other aspects degrade. (And that's one thing SD3 does seemingly pretty well, their model in no way feels like a sidestep, but is fully better than SDXL)

[bitter hearth]

well some of us just look for best option, of course we have to get into the fan wars like any corps haha

[idle current]

i dont even know why i talk to clueless trashtalkers like you lol

[bitter hearth]

idk it would have fooled me if i did not know it wasnt real. i would love to see sd3 try to duplicate it

[hallow lion]

you have no control, no ccustom trained checkpoints and cencorship.

[idle current]

comes into the talk and just starts trashtalking

[hallow lion]

the onyl reason they seu MJ and dalle is coz either they dont know about SD or ar elazy to install it

[tropic aspen]

Honestly, idk why ppl care about text so badly. If you wanted to put text in an image you'd... put text on an image

[idle current]

have you considered that people use multiple products?

[tropic aspen]

Why have the AI do it?

[hallow lion]

who cares about multiple procudts

[bitter hearth]

well if you really want an image you would paint it or draw it. why have the AI do it?

[idle current]

tbh true

[hallow lion]

you have so many custom chekpoints on civitai you can do whatever you want

[cosmic pelican]

doesn’t sd3 have censorship cause you can’t makws nsfw?

[idle current]

thats how to know you are clueless

[hallow lion]

noway you can do more variety on MJ or dalle

pluhleaZe

[tropic aspen]

Because that's requires more effort than using the text tool in Paint.net?

[hallow lion]

yes SD3 has cencorhsip

[bitter hearth]

and adding natural chalk text to a guy at a blackboard in some other method does not?

[hallow lion]

but sdxl doesnt

cascade doesnt

sd15 doesnt

and SD3 wont have any cencosrship too ass son as the weights drop

[idle current]

if we are to get to your level, Why would i use Stable Diffusion if i have MJ or DALL-E + Adobe Photoshop/Firefly or for vector Illustrator? Nobody needs VAE control and similar

[hallow lion]

so whats your point?

[tropic aspen]

That can prob easily be done with the correct font and a bit of img2img/Cnet

[hallow lion]

dalle looks fake and ceoncored

MJ looks arts fartsy and cencored

[bitter hearth]

lets see you try it then, post your image here

[hallow lion]

i just dont get it?

[bitter hearth]

should be able to make it in a few minute right? it so easy to do

[idle current]

i think we have to deal with a fanatic here lmao

[hallow lion]

why would anyone sue them indtead of free and far superior SD?

hmm?

[idle current]

dude casually joins the channel and starts trashtalking bs

[cosmic pelican]

sd being sued?

[bitter hearth]

i feel like i am on some comic discord where someone is upset we say iron man cooler than batman

[idle current]

Stability AI is sued, yes

[cosmic pelican]

for what lol

[idle current]

for real

[idle current]

copyright infringement

together with Midjourney and DeviantArt

[cosmic pelican]

haha sued by who

oof

[idle current]

Karla Ortiz and someone else

forgot the names

[cosmic pelican]

dunno them

[idle current]

the case just got forwards since few days

OpenAI was sued as well but separately

[bitter hearth]

but i guess i apologize some too. this is a sd3 area and we are talking about ai in general. i suppose 4o and other thing are ot haha

[idle current]

true

if you guys want we can talk on another channel

[cosmic pelican]

🤔

[bitter hearth]

but to me, i am no fan of any, i use what is best and use them all, not plant my flag like some keyboard warrior for any corp only one digital landscape

[hallow lion]

SD for the win

[bitter hearth]

sure

[hallow lion]

Nothing comes close

[cosmic pelican]

Your computer must be strong af lol

[idle current]

yeah thats why i cant take the guy previously seriously. Professionals dont use exclusively one tool

[bitter hearth]

haha i wish. i can not even run sd really.

[hallow lion]

anyone who uses PC seriously has 6-8gigs VRAM these days

[idle current]

my pipeline exist out of like 10 software

that i use

at the same time sometimes

well not literally all open

[bitter hearth]

i guess i am not serious enough pc user 😂

[idle current]

but switching in between

when i build my PC next year i can open em all then lol

[bitter hearth]

do you even p.c., bro?

[idle current]

or you simply dont need that much xD

[hallow lion]

you cna fooooocuse with 4gigs please

[idle current]

also my software pipeline is extremely expensive

[cosmic pelican]

SD staff reading these comments 🫣

[idle current]

you dont need that i suppose

[idle current]

and eating popcorn

maybe Dodge is their bot

[hallow lion]

well

[mortal mesa]

someone that works with OpenAI said youll never need more than 16KB memory

[idle current]

to put in some salt

[hallow lion]

if you are reading please SD3 weights now

you are the heroes please dont leave us

[idle current]

wasnt it supposed to be released already?

[bitter hearth]

they wont until the money dry up

[hallow lion]

I am not their bot

[idle current]

maybe i am

O_O

[bitter hearth]

a bot would have more nuance i think

[idle current]

Infidelis.bot

[hallow lion]

I had to actually upgrade my PC just to use SD but I know its the future and a good oivnestment

MJ and dalle are hot garbage

[idle current]

bruh come on xD

[hallow lion]

cencored and bad quality and no variety

theres one lora that emulates midjourney for instance

one single lora

reporduces midjourney style for 200MB

for free

[cosmic pelican]

is that midjourney v6 asthetics?

[hallow lion]

If people ar etoo intimidated to install all these scary apps like A1111 and comfy well its their loss

[idle current]

people dont use SD for quality or prompt adherence, they use SD because of local usage and customization possibilities

[hallow lion]

it took me months to also get used to it and learn it but thats what ti is

[idle current]

and those customizations are absolute overkill and unnecessary for a lot of people

for me for example

[hallow lion]

nooooo

[cosmic pelican]

Btw is it just me or does A1111 run up to only Python 10

[hallow lion]

we want perfect awesome and consistent images

this is the future

not just a hobby

[idle current]

the future is at corporations

just saying

[bitter hearth]

4o is going to have consistent character, another big feature for it

[idle current]

i just got access to it

i mean to the model

[bitter hearth]

not 4o i think, it is not turn on yet inside of 4o

[hallow lion]

Thats why Sd is so greta

[bitter hearth]

i mean the image creation

[hallow lion]

so the future is not at corporations

[bitter hearth]

it still run on dalle3 for now

[idle current]

yeah its not there yet

[bitter hearth]

oh ok 4o is awesome

[idle current]

also i wonder how many messages i can send before it gets set to 3.5 again

because its limited

[bitter hearth]

if you are free i think not many

[idle current]

dont tell me like 5

[hallow lion]

lol "death is typing...

[bitter hearth]

idk i am a plus user haha. but so many people are try to make use of it it may be 5 haha

[idle current]

i deleted my original OpenAI account, before that i got rid of MJ sub

[cosmic pelican]

by customization possibility do you mean pkgs, loras, etc?

[idle current]

well i got rid of MJ much earlier

[hallow lion]

MJ is artys fartsy

[cosmic pelican]

lol

[idle current]

yes, controlnet too

[hallow lion]

too high contrast and dramatic and too much color

[cosmic pelican]

im signing off

[hallow lion]

I think Sd was always the perect ballance between grey boring flatness of dalle and artsy fartys wildness of MJ

[idle current]

the death is signing off

i can feel it

[cosmic pelican]

🤙🏻

[idle current]

have a nice day/night 😄

[hallow lion]

🙂

[bitter hearth]

say hi to Dream for me

[idle current]

La what are you using again? SDXL?

we might have talked before about that

[bitter hearth]

i dont use any sd at the moment

just curious to see sd3 when it release

[idle current]

aaah okay

id test some stuff as well but i wont pay for Clipdrop and similar

[bitter hearth]

i test sd3 a few week ago, but on my cousins pc since he pay for some credit haha

[idle current]

oh

[bitter hearth]

but i will wait to see it on here if they offer free test. i think they did for xl

it was not very good with hands still i know

[hallow lion]

hands will still suck at SD3 i feel

[idle current]

if any i would gladly replace D3 with SD3 if SD was better and preferably for free but without local usage lol

[bitter hearth]

we had much better success with the image 2 image which is of course always a nice feature

[idle current]

on the other hand Firefly/Photoshop remains untouched for me

[bitter hearth]

i hear firefly 3.0 is release. good quality?

[primal summit]

It may be released in August

[idle current]

if you ask me why not local, because of my specs

and yes its more demanding than my usage of Unreal Engine 5

lmao

take that

[hallow lion]

😮

[primal summit]

Are you a game developer?

[bitter hearth]

how do you even use firefly? in photoshop only?

[idle current]

text2image is good, but in general (worse in some cases, better in other) MJ and D3 beat it in that regard, in editing, in further features it blows them out of the water

[bitter hearth]

yes i am just trying to give some compliment to sd too haha

[idle current]

at this point almost exclusively there and rarely via web UI, they fully integrated it into Photoshop Beta and will so with the major one probably in october

then there is the model for vector and soon for video

but thats not Firefly 3.0

[idle current]

yes

[bitter hearth]

what are the big player then: dalle, mj, firefly, meta, sd... anything else?

[idle current]

no, im not well known lol

famous

[bitter hearth]

cool job

[idle current]

its a side hustle currently

until get eventually freelancer/self-employed or end up in the industry

i*

[bitter hearth]

starting in the alley, one day on the boulevard haha

[primal summit]

Me too but I use Unity. Have you mastered the engine?

[idle current]

if i was employed there i wouldnt even have to pay 250-300€/month for creative software

lol

the company would pay for me

[idle current]

nah, sadly not. I dont even work with C++ yet, i play around with blueprints

well C++ is still in you know what i mean but i work with visual script simply

how is it going with Unity in your experience? ^^

[primal summit]

I suggest you don't waste your time with it. It is very difficult and the demands are great. Unity is much simpler and powerful at the same time, especially as a visual programming language. You will not even need to read the documentation if you understand its basics.

[primal summit]

I mastered it almost completely in about one year. I am currently working on a big game that I will finish soon

[idle current]

for my case its not a waste at all, its exactly what im looking for tbh. Well i dont look for C++ ofcxD

but i have to learn it at some point

otherwise i would have taken Unity indeed

[idle current]

from absolute beginner with no coding experience?

[primal summit]

I still don't know things related to online games and also the C# language, I only have the basics

[primal summit]

I learned basics but have no experience

[idle current]

oooh okay, we can go to off topic if you want 😄

i think slowly someone might punish me here for being OT all the time lol

[primal summit]

Oh, I see. Are you planning to work as a C++ programmer?

[idle current]

not really but since im in Unreal Engine i have to learn it anyway

sooner or later

[primal summit]

Lol, I remember that I tried to learn it for 6 months, but I did not understand anything

[hallow lion]

Don;t fail us Emad. Pleas eremember what this is about.

Money is dead.

Even if you make it taxes and inflation will take it.

Don't deviate.

Stay on target.

[primal summit]

By the way, is there a way to train stable-cascade online using Colab or Kagle?

[idle current]

possible lol. Some people get along with it, others not so much. This happened to me with 3D software Blender

[primal summit]

I tried a lot but the bugs never end

[bitter hearth]

didnt emad resign?

and i wish someone tell my bank money is dead haha

[primal summit]

ho

[tropic aspen]

Yes

[hallow lion]

Money being dead wont be announced on the news I am afraid just like the pandemic... It wa snever annoucned officially beign over.

Also now we are in WW3 but you will not hear it on news

spo same for money bign dead

its not something media will proclaim

even if its true

You have to kind of use your brain to see it

[idle current]

it can be frustrating ^^

[hallow lion]

sure is

[primal summit]

Maybe if they take the Blender Foundation's business model, it would be great for them and for us

[idle current]

what should i say with me juggling with prototypes currently lol

[idle current]

mhm doesnt work out tbh

Blender is a completelly different case

[hallow lion]

XD

Why doesnt Emad just put google Ads in his sweet lovely arxe models?

[idle current]

Emad is gone

[hallow lion]

idk lol just some ads for every gneration

and u cna use SD adblocker to avoid them

[bitter hearth]

i will tell the bank next time they want a house payment, money is dead, you guys just need to see it haha

[idle current]

as well as some of the best researches from Stability AI

[hallow lion]

allegedly

[idle current]

they also left the company

researchers*

[primal summit]

What I mean for Stability Company, if they work like the Blender Foundation, is to focus on only important models while accepting volunteer developers and scientists

[hallow lion]

you see 1 google ad for every SD3 generation...

And then u can use SD3 adblocker of course eventually but dont tell anyone

[bitter hearth]

stable probably will get bought by someone else if they even worth the effort at this point idk. hard to see a future beyond sd3 at this rate

[idle current]

oooh so narrow down to for example only image models and there focus on one

the rest should be handled if at all by community and co.

[hallow lion]

Well money almst died three times in 20tyh century so...

2008 2016 and the stock market crash

[idle current]

Breaking News we are happy to announce that Adobe acquired us today

[hallow lion]

I use photoshop from 20 years ago

[idle current]

(they wont do that)

[idle current]

uffff

[hallow lion]

lol

i use some 20 year old programs in my pipeline

some programs even deunct

f

[idle current]

my man what else do you have xD

"ey, dont touch that software. Its from 1976"

[hallow lion]

because i just cant stand the montly whatever subs model they sue now

its laughale

what if i dotn have internet?

or a phone

u can tlive without your phine now?>

are we serious?

its absolutely bizare and crazy that peopkle assume everyone has phine now

is a phoen an extension of your body now

[idle current]

i used to torrent them longer time ago, but im a full fledged customer since meanwhile like 2 years

[idle current]

to a certain point, yes

[hallow lion]

well not for me

i dt eve use my phone to sign into anything

[bitter hearth]

i mean you are on the internet right now, hardly some luddite

[hallow lion]

still

this is a PC

[bitter hearth]

which is just a fatter phone that need a cord to be in the wall socket

[hallow lion]

the idea that i need a phone for everythign is not acceptable

what if i dont have phone?

[bitter hearth]

then you wont have that advantage. idk i dont care if you have a phone or not

[hallow lion]

yeah

my passwords are so long adn weir di will neve rtype them in on the phone

i juts use my phone to chekl them time and read stuff on the toilet

Poll: how many people read stuff on the toilet on their phones?

[idle current]

i even do draw and paint on my tablet while being on the toilet

soon i will also sculpt in 3D on iPad when Zbrush gets released for it

[hallow lion]

😄

[primal summit]

[little quarry]

2 weeks

[trail frost]

Hey

Can someone please recreate this and share the image and the prompt please

[cyan blaze]

good

[primal summit]

for what

[little quarry]

[low stone]

you'll have a better time over on ideogram.

[crude yarrow]

Until the open source Freesound audio model drops. Yep. That's probably what he's taking about

[noble coyote]

Very steady tracking!

[low stone]

yeah i figured out so far the best couple of settings for sdv. it now gives good output more than it doesn't.

[noble coyote]

I must try SDV one day ...

[bitter hearth]

godamn

You unlocked the chef class

[frigid saffron]

whers is sd3?

[low stone]

click here: https://discord.com/channels/1002292111942635562/1237461679286128730

[noble coyote]

Clipdrop.co

[stoic thistle]

？

[icy drift]

Guys we will get SD3. It has limited useful lifespan, and SAI values community engagement. HanyuanDiT already publicly released with similar performance (and basically same tech). In a year, SD3 will be outdated, and in ten years, it will cost about $100 to train a similar model.

[cobalt moon]

SD1.5:

HanyuanDiT are still below the performance of fullbaked SD3

as well it is just a forecast of SD3's new captioning

[desert garnet]

ok where is the fullbaked sd3 to test

[cobalt moon]

You can safely assume Artisan SD3 is like 75-85% baked

[desert garnet]

we cant really assume anything since we dont have the weights

[cobalt moon]

We do have another SD3 clone apart that Chinese model. That one project carried by Simo Ryu in Twitter

If I won't wrong he literally train a model from scartch with those paper.

[abstract plank]

If the company is facing a sale, the development team is lost, and the training equipment is an external cloud platform, then the existing big model results are the only thing that attracts buyers. Until the funds are resolved, it is highly likely that they will not be visible. For buyers who spend money to purchase large model assets, with a previous operating loss of 20 million and a debt of over 100 million US dollars from the supplier's cloud platform, how to exchange for profits above the acquisition costs is also a constraint. Ideals are good, but difficulties are great.

[desert garnet]

do not worry,in 10 years we will be living in underground bunkers so we wont have to worry about ai models

[cobalt moon]

https://twitter.com/cloneofsimo/status/1791135683362631698

Simo Ryu (@cloneofsimo) on X

moon is high, model is 44k steps in, I stopped the run to check on everything and use multi-nodes, didn't expect anything at all.... However, safe to say, i've trained my FIRST ever 5.6B Text2Image MMDiT from scratch!!!

Twitter · May 16, 2024, 3:56 PM

pretty sure you guy know how much is 44k parameter

[cosmic pelican]

So what are you actually implying?

[noble coyote]

Mebbe that SAI is joining MJ, OAI, Clipdrop etc in a monthly subscription business model?!

[restive halo]

https://x.com/EMostaque/status/1790451196915831018?t=YJuHqJJ-YCivInuOrZ2_Lw&s=33

Emad (@EMostaque) on X

@morew4rd @GoogleDeepMind Sd3 is due to drop now I don’t think folk will need that much more tbh with right pipelines

Twitter · May 14, 2024, 6:36 PM

[idle current]

I mean they pretty much are in but its not working well when it comes to revenue

[cobalt moon]

seem pretty funny thinking that it is 2 days already and you are the first one posting it here

[verbal epoch]

Weights when?

[raven fern]

[icy drift]

Just tried hanyuanDiT (got it running on my PC), and it's just bad at prompt following. 😕
A red-haired woman wearing a blue hoodie is standing on the right. A white-haired man wearing a green tuxedo is standing on the left. In the background is an abstract watercolor design featuring arabesque patterns.

Lemme test form consistency.

Oh sweet mercy what am I even looking at right now??!!

This is unbelievable. 😱

I mean, it got the neck a little wobbly, but I bet picking a different sampler would help with that. It has the right number of fingers on both hands, 6 tuner pegs on the head, and 6 strings over the hole (sorry IDK guitar terminology).

Will test more have to go to work.

[little quarry]

Pick that pillow up off the floor!

[icy drift]

It's for his cat.

[little quarry]

[icy drift]

HunyuanDIT is a fully-censored model! 💯
I've never seen that for a chinese model before, but mostly I'm just amazed at how well they managed it.

[paper frigate]

how to use mid journey

why can i join this server

cant

[icy drift]

As a large language model trained on Stack Overflow, this question is off-topic and will be closed.

[paper frigate]

so you are ai

[desert garnet]

i cant join this server either,its broken

[icy drift]

Hmn.

[icy drift]

It can't do text. Wow this is the worst chinese innovation so far. Lightning and Hyper were legit.
A cute kitten holding up a sign that says "SD3".

[paper frigate]

Is there any other way

[lament summit]

3

[paper frigate]

sd3

[lament summit]

the photo just says 3

[paper frigate]

3

[little quarry]

Two weeks SD3

[jolly abyss]

Wait a second. I thought Temu is the worst.

[paper frigate]

I want to create some original pictures with any convenient AI

[icy drift]

Isn't that a shopping app? IDK. Just mean bytedance and tencent have done really awesome stuff with SD, but HunyuanDiT is terrible.

[jolly abyss]

Yeah, not the best innovation. I am Team Pixart. No text at all, but nice prompt following. They just need a model with more parameter.

[jolly abyss]

How? Are you a magician?

[low stone]

I ran your comment as a prompt and paid money for this. I call it, octopus hands.

[dull star]

so does artisan use a different version of SD3 compared to the API?

[cobalt moon]

hard to say

probably same

[dull star]

either way, you think SD3 is 75-85% baked?

hope end of may is the release of SD3

with controlnets and finetuning and stuff

I wanna see how highresfix works with it

cause pixart had a lot of issues

[jolly abyss]

Me too, without the paying part. I call it: Basket Hand

[cobalt moon]

nvm

[low stone]

On a slight rant, the fact that sd3 is still just as terrible at hands as it always has been, makes it worthless for so many generations. I want to be able to make a joke picture based on something that happened at work and send it to my coworkers, but I can't do that if there's something blatantly obvious that is majorly distracting from the actual content of the picture. Pixart and Ella are amazing, but they generate octopus hands at random too. Knowing that it's not going to get any better with sd3 is disheartening.

The other image services seemed to have figured it out but not sd3.

I've tried using hand detailer but the original image's hands are so bad, it can't even fix it

[cobalt moon]

DALL E 3 cant even do it well I mean

or in fact all of the AI art generation model

[low stone]

I get great results from dalle all the time.

I generated 30-40 images of men carrying rifles from dalle for a Lora training set and 100% of the images had perfect fingers and perfect hand position on the rifle.

[cobalt moon]

because under machine learning you can't "exclusively" training hand than other part of the body which are much simpler in anatomy movement.

and plus there are like 100 different hand position with 100 different perspective

in 100 image dataset

[low stone]

I just did 2 images from dalle. Lots of hands. All perfect.

I just paid for 5 images from sd3. Every single hand is a mangled mess.

[cobalt moon]

Well, while training hands is damn hard. it is entirely possible that you could lessened the chance of malformed hands. I actually don't quite know how they can able to achieve it without slight "deformation"

increasing parameters and bombard training may worked but eh...

just like how people try to do their job weakening the power of funny hands with finetunes training and stuff.

but anyway thx for sharing

https://www.youtube.com/watch?v=24yjRbBah3w

Vox did a great job explaining it

[cunning lintel]

If only it was just hands (after a request of "how to create this image earlier in this topic, tried to create it with sd3) . It's also terrible at limbs in general. Those feet don't look healthy. And I think everyone knows how a lying person looks like

What's even going on here?

[cobalt moon]

I think that's because the complexity of number of characters

SD can't count how much limbs are there

[cunning lintel]

I hope this is part of what "sd3 not ready yet" means. i said it before, it reminds me of the early sd1.x leak in many ways, it garbles so many things

[trail frost]

Thanks

[cunning lintel]

But really, use ideogram if the intention is to get as close as possible with a promptL https://ideogram.ai/g/s_dmCBfhTPGFjFiQi3kPoQ/2
4 large men in the background wearing sunglasses, pointing their finger at a tiny man, as big as the laptop in front of him. The big men seem to have a newspaper texture all over them,, as if they're made from it, and the small man wears a yellow jersey, the small man sits on what appears wooden boarding, face looking at the laptop on the boarding, legs crossed. his back to the viewer so the laptop screen is visible. Only the jersey and the laptop screen have color, the laptop screen shows some graphic design, the rest of the image is back and white

[trail frost]

This is good

[dull star]

yeah ideogram is ridiculously good

If only SD3 could perform as good

maybe with finetunes one day... 😔

it actually got the newspaper part a bit better, but the size difference is non-existent

not that bad tbh

especially for a base model

and yes, I did cherry pick this lol

[rain current]

ideogram works very well, I like it a lot. Sometimes it produces real garbage, other times it creates wonders. The main problem is that the results are very similar, even if you change or add prompts

[cosmic pelican]

Well, with MJ subscription, we get unlimited generations… So SAI loses here with it’s credit system, no?

[hallow lion]

Cage: Not the beeeez! not my eyes! not the beeeeeeez!

[noble coyote]

Agreed - I have a one year subscription at MJ - and don't really use it!!! LOL

[abstract nymph]

:D

[dull star]

fr

cant wait to see how it performs with larger step counts, highresfix, etc

how much difference there is between the model sizes

what if I only use clip-G+T5

etc

[hallow lion]

are SD3 checkpoints going to be more than 6 gigs in size?

[desert garnet]

lemme check the ones i downloaded from hf

o wait

[restive halo]

sdxl is more than 6gb and sd3 is bigger so yeh of course

[hallow lion]

hmm

[restive halo]

or I guess the smallest sd3 will be under, the biggest will be over

[dull star]

4B 8B absolutely

don't exactly know what MMDiT weighs like

but fp16/bf16 weights will probably match expectations

SDXL is 3.5, so between 2B and 4B of course

you also have to take into account the T5 model, which is an extra few gigs

https://huggingface.co/city96/t5-v1_1-xxl-encoder-bf16/tree/main this one is 10GB and is recommended over the useless fp32 one which consumes much more RAM in comparison and loads like a hog

you can run this one (or the fp16 one) on CPU RAM just fine and it loads in just a minute or two on SSD

T5 on cpu usually takes like 5-10 seconds to create the encoding, on GPU it's near instant

(talking from Pixart-Sigma experience, which also uses the t5-xxl encoder)

[wide pagoda]

T5 can be turned off entirely (and probably should be when you're not trying to generate text)

[cosmic pelican]

If you don't use it, can you give it to me?

[hallow lion]

👍

[dull star]

since I make memes I will have it on, and even then I'll have it on just in case

I hope its mostly just text, cause I don't want to have T5 a massive advantage, cause that would turn users away from SD3

like "oh I have to download +10GB and consume more VRAM/RAM just to make it better than SDXL? Why even have SD3"

here's a sneak peak at using only each or special combos of the clip/T5 models

by mcmonkey/Alex

sadly prompts are cut out so its annoying to compare

[wide pagoda]

I was about to post that lol

[dull star]

T5 likes to make photos/photoreal images

[cosmic pelican]

@hallow lion What are your personal thoughts on SD3 moving to cloud-based (subscription based) if the company were to be bought by X (Elon Musk), and migrated to their website X with the full functionality (idk if this is possible, but let's assume it was; my knowledge in the tech field is limited) of what is similar to Automatic1111?

[hallow lion]

It woudl be better than MJ and dalle as there would be more features obviously... but still nothing beats local... I would hope Musk would take that opportunity to really free the weights and alllow open AI sicne his attempt with öpenäi backfired and they went the corp direction

I really hope SD3 gets out and we can refine it and it should be enough for a few years as far as image generations go really...

[cosmic pelican]

Ok thanks for your thoughts on that subject! lol

[hallow lion]

The next thing we should focus on is the tools

consistency is the next big thing

And not you know 90% and ipadapter and hacks and tricks

but real consistency

down to a tea

[cosmic pelican]

Yeah that would be a good approach.

[tranquil owl]

你好

[primal summit]

He is the worst person to get the company because he will make it pay in the end

[teal fossil]

If Huggingface would buy SAI (like they teased a while ago), that'd be amazing.

[hallow lion]

yes facehuggers would be epic choice

[versed elk]

Increase resolution and create a HDR photograph

[bitter hearth]

Enhance

[knotty meadow]

Will be sd3 available on replicate?

[idle current]

Impossible

Huggingface cant even afford to acquire Stability AI

They would go bancrupt (almost) instantly

Ups i meant @teal fossil

[jolly abyss]

🤢 That hand and the bump on the arm. Disgusting

[low stone]

Are you guys here to wait for SD3 like me?

[idle current]

Curious about it

Not excited tho

[low stone]

I think it's gonna be better than sloth disco 2

[idle current]

XD

[abstract nymph]

[removed image] oops that was the same one as last time

there we go :)

[raven fern]

nice

[raven fern]

now try to lift two cats, you can't... :3

[hallow lion]

You can if they only carry 4GB of ram.

[raven fern]

true

[dull star]

oh so SD3 will come in december

[bitter hearth]

2 weeks in december

trust

[dull star]

only 2 weeks after decemeber? 🙏

[bitter hearth]

some preety cool armor

is it supposed to be anything in particular

[low stone]

I'm taking the Guardian website photos of the day, putting them through gpt4o describe, then changing up details.

I have to give it credit, pixart/ella couldn't do the powerwash sprayer.

sd3 could

although it only got it right 1 out of 12 images, which makes me feel like it was luck.

[hallow lion]

Wow SD3 is better than ella sigma?

[low stone]

it IS better than those 2, but only where the finetune trainining happens to be better than my sdxl refiner model. that doesn't happen often, but like in this robot one, it was.

[hallow lion]

Refined SD3 will rule the universe.

[low stone]

it's why it would be tragic if sd3 didn't get released. pixart/ella is awesome, but only because it's a great stopgap until sd3. not instead of.

SD3 ^^

pixart/ella ^^^

In a surrealist style, vibrant colors, and exaggerated forms, Will Smith confidently strides through a bustling, modern airport terminal under dramatic, high-contrast lighting, aggressively pursued by animated, anthropomorphic spaghetti paparazzi with flashing cameras, surrounded by travelers, sleek architecture, and large windows with incoming sunlight.

neither one could quite nail it

[violet escarp]

@dull star they are experimenting with t5 only.
https://old.reddit.com/r/StableDiffusion/comments/1cgr74j/april_30th/l2bxv66/
https://old.reddit.com/r/StableDiffusion/comments/1ciyzn5/sd3_weights_are_never_going_to_be_released_are/l2dhd6q/

They originally trained it with t5 limited to 77 tokens to match clip. Training t5 only allows them to bump that back up to 512.

[bitter hearth]

challenge: this chat 1 day withtout doom and gloom

[violet escarp]

if they decide to limit it to 77 tokens then t5 is basically irrelevant

[dull star]

interesting?

[low stone]

pretty much

[dull star]

yeah

I didn't get this initially

then again, we have longclip-L, now we need longclip-G as well

[cunning lintel]

While SD3 is 100% the better/more complete model, often when i have a prompt that looks really good in SD3, i try it in pixart, and to my surprise, it looks really good there as well.

[dull star]

to make them ~240 token length globally

[bitter hearth]

(I have nothing to say)

[low stone]

clownshark and I have been messing around with gpt4o, and the prompts it make are massively more following for sd3 if they're under 77 tokens. it's kinda frustrating because pixart/ella are both 300+ tokens

[violet escarp]

pixart sigma trains on 300 token length

[violet escarp]

I think sigma is underbaked and doesn't have the best dataset

but the arch is still pretty good

[low stone]

pixart is awesome and frustrating at the same time. I'm ready to start training it but I spent hours last night trying to get it working and couldn't. too many wacky python dependencies.

[bitter hearth]

[cunning lintel]

kinda sucks right. I'll take the challenge!

[jolly abyss]

I like the subtle SD3 tattoo.

[low stone]

hah yeah i try to sneak it in there.

[cunning lintel]

Textured tempera painting, billowing digital anime by Pascal Blanche, Ross Tran, and Glen Keane, big tareme eyes, -- dramatic, low-angle shot of the girly explorer, standing at the edge of a misty, abandoned funhouse, a crumbling, creepy clown face looming in the background

[low stone]

very cool

[bitter hearth]

reminded me of tomb raider

[cunning lintel]

haha, now that you mention it, she's a girly explorer too 😉

[bitter hearth]

https://tenor.com/view/lara-croft-tomb-raider-angry-gif-13282398

xD

[hallow lion]

There. Gonna be released. Got that! No more fear mongering. XD

[hallow lion]

Is that Bexos?

[bitter hearth]

lmao

[rich iron]

If his eyes were any further apart, they would be in his ears.

[bitter hearth]

That's the punishment for becoming one with the spaghetti

[low stone]

never go full noodle

[hallow lion]

lol

[raven fern]

go full nude instead

[raven fern]

mhm

[low stone]

that didn't do the animation. :\

[raven fern]

lol

[violet escarp]

I think it also implies that it will take longer to release compared to what they said earlier

[little quarry]

SD3 two more weeks!

[bitter hearth]

sounds like stable is going to be sold

not sure that sd3 is coming from things you read online

[stark cave]

If it's musk related then you can pretty much ignore it.
If they did sell to musk then StabilityAI would kind of be backpedaling away from their whole vision of safety.
It is just taking a bit of time for them to finalize the safety DPO training they are currently doing. That is why they released over API first. In order to gather enough data on model behavior to ensure the Safety training worked before full release

[bitter hearth]

well emad tweet about it, it seems more a breaking news kind of thing. i guess we'll see

[low stone]

he tweeted what what's breaking news?

[bitter hearth]

i was just doing some catch up on reddit ai news and in stable they talking about a sale that is pending, emad tweeted about it too. look over maybe first 4 or 5 top stories.

[dusky thistle]

where?

just post a link if you have one

[bitter hearth]

stable reddit

https://www.reddit.com/r/StableDiffusion/comments/1cu1c4d/well/?rdt=35057

[dusky thistle]

can't find that tweet

wish ppl would actually provide their sources

real? probably? maybe?

[bitter hearth]

well its posted on they reddit i assume they would not leave it up

this was the other thread i was reading

https://www.reddit.com/r/StableDiffusion/comments/1cub74a/so_sad/

[dusky thistle]

oh wait now i found it

https://twitter.com/EMostaque/status/1791148539013411217

Emad (@EMostaque) on X

@literallydenis Apparently its been about to be sold and run out of money for a couple of years 🤷

Also apparently it spends more money than it makes unlike other AI companies, crazy.

Twitter · May 16, 2024, 4:47 PM

don't know why ctl f didn't find that

yeah either way the only buyer i could see them finding is someone who has the means to monetize models via online generation, a plan to do so, or some that is a near trillionaire who just throws away tens of billions for the f of it

[low stone]

look if we all just blow into the cash worm, money comes out the top. problem solved.

[dusky thistle]

just ask good ol' Polyjaws for help

[low stone]

this is what happens when you train a lora on dall-e output. I can just say yeah, dall-e made that.

it's uncensored for me, i don't know what your problem is.

[dusky thistle]

hahaha

great plan

[low stone]

I feel like that could be an impulse buy on etsy

[dusky thistle]

i would buy it that's for sure

[mystic flare]

i would buy it that's for sure

[wooden hare]

gm

[wraith ferry]

help

cat

/help

[noble coyote]

"Lego Lara" 🙂

SD3@ClipDrop 🥳

SD3@ClipDrop - the red knight drinks beer and eats pancakes with the white witch!

SD3@ClipDrop

[cunning lintel]

In a misty dawn forest, a majestic Cucumber Creature poses, its slender, elongated body adorned with intricate, swirling patterns resembling tiny seeds. Delicate, leaf-like fins decorate its arms, and its gentle, smiling face features expressive, sparkling green eyes. Soft, pale green skin glows with a soft, luminescent light, set against a whimsical, romantic backdrop. with enchanting, dreamy colors and intricate details.
Negative: boring, threatening, desaturated, comic

[cinder junco]

Emad was being sarcastic. Like “they were saying we were running out of money and would be sold imminently two years ago, too. And Stability is the ONLY AI company not turning a profit right now, oh noes!!!” That is, it wasn’t and it’s not.

[low stone]

Yeah he was deflecting.

[wooden temple]

presentation background, about it and business, keep it simple

[hallow lion]

Lol how come this isn't blurred? 😄

[dull star]

its not about content

I bet the nsfw detector was only trained on "too much exposed skin = bad!!!1!"

its unreliable

[cosmic pelican]

Thanks for this!

[sullen moss]

https://twitter.com/Lykon4072/status/1791820252243800253?t=KD7CZRhCRARO3CdbIM-BuQ&s=19

Lykon (@Lykon4072) on X

Also, as far as I know, we're still planning to release #SD3 as Christian said. Hopefully the wait is almost over.

Twitter · May 18, 2024, 1:16 PM

[dull star]

https://tenor.com/view/office-michael-scott-thank-you-gif-5278681

[dull star]

Christian is the CTO right?

whaat???

I thought it's upsacled with dreamshaperxl or somethign

[bitter hearth]

hes right

stop asking when kek

[dull star]

some people are still in denial

they still believe that SD3 won't release