💬｜general-chat

[soft iron]

"good that there are people with the entire brain left in their skull nowadays
guess 5%
from this joke planet"

any year is good and it just one of the turns of human development

That's what people have always said

but that's just senile conservatism

[sudden ruin]

[bright trench]

bonk

[sudden ruin]

No space for hate

[soft iron]

why

he didn't hate anyone

[west mirage]

Eshku, you are a great asset for the globohomo agenda

[hasty hornet]

I just read through what he said before...there's some "interesting" stuff.

[vast ingot]

A discord server about AI drawings is no place to start political discussion like that.
Just a reminder : to call for mods, don't ping every one of them, but use Right click => Application => Report to staff
or react to the message using ⚠️
both of those ways go right into the moderation queue

[sudden ruin]

Calling people sick in their head for their sexual orientation

[hasty hornet]

yea I'm sorry for a ping, but I didn't want it report to be limited to one message, there's more from him in his older messages

[soft iron]

but this is often what happens...

Due to psychological trauma, people go into such a defensive reaction

[vast ingot]

reporting one message links to the message reported, and moderators will read the context to act. it's the best way to contact all of them

[soft iron]

this in no way cancels out people from another "sexual orientation"

but it exists

[sudden ruin]

As already stated above, this isnt the right place to discuss this, please stop it here and now or actions will follow

[hasty hornet]

it's not just about context of this conversation, it's about his older messages too (which I think was partially cleaned already either by him or mod)

I wish there was a way to add message or image to this "report to staff" thingy for additional info

[vast ingot]

I started reporting the first message that got on the trolling path personally. For having moded this discord for some time (quite long ago now), we did analyze a user's post history in cases like those anyway, so it would have been caught. but anyway, thanks a lot for reporting it like you did too

messages reported get copied in there too, so if the message is altered, it's still visible to mods

[hasty hornet]

oh that's interesting

[upper panther]

Of course

[bleak matrix]

Hello everyone! 🙂

It's a really nice day outside today; what's everyone up to?

[quick hemlock]

Hi! Did someone try the new model yet?

Trying to sign up keep getting error messages. Is it worth it if I already have Midjourney pro?

[fervent thunder]

guys SD available on m1 max studio?

i m thinking of purchasing it but i wont purchase it if it doesnt have SD

[surreal jay]

Hi Sunny, stuck at work 😭 you?

[weak scroll]

SD3 kind of is not so great! SDXL produces better results. What's the updated in SD3?

[forest trout]

Would love to train models rn but stuck at work too 🎵

[weak scroll]

What's new in SD3?

[forest trout]

At least Harmonai office hours are (likely) happening today so that's a nice perk.

[bleak matrix]

Oh, just studying and doing research! And trying to plan things! You know a regular day, haha!

It's a fairly nice day, though!

[weak scroll]

I got an exam tomorrow, CASP+ certification

Cybersecurity stuff

[bleak matrix]

I hope it goes well for you!

[surreal jay]

Research? On what? If you don't mind me asking

[bleak matrix]

Etymology, and, also, sign language! Genetic research is also part of the program, along with historical research.

[surreal jay]

That's a lot!

[mortal delta]

What are the best stable diffusion interfaces that work with amd and windows, if any at all?

[warm junco]

Auto1111 would be the best. But comfyui works too.
For AMD guides checkout my site:
https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides

If your GPU has more than 4gb vram and is relative new then go for the Zluda guide

[mortal delta]

thanks ive tried here and there but i possibly did something wrong so i will try again.

[warm junco]

No problem, if you run into any issue or have a question, just ask me in #🤝｜tech-support

[bleak matrix]

It all blends together with my historical research, to be perfectly honest! History is such a fascinating subject--it relates a lot to art, and I love being artistic, especially with AI. So I get to use all of the things I learn about here. It's really fun!

[mortal delta]

ill try on my own first but i will be there if anything happens.

[weak scroll]

Does A1111 have SD3 support yet?

[warm junco]

Not yet

[mortal delta]

i know comfyui does...

[weak scroll]

I've been using Swarm

[warm junco]

Yea for now comfyui and stableswarm

[weak scroll]

https://www.reddit.com/r/StableDiffusion/comments/1de65iz/how_to_run_sd3medium_locally_right_now/

[warm junco]

@mortal delta BTW what's your GPU?

[weak scroll]

I have a 3060 12GB

I need more VRAM so I can run LLM + SD same time

Right now I can only one one at a time.

[surreal jay]

Indeed history is fascinating, I never liked it much in high school, but I didn't know anything back then! Are you doing it for University? Or is it a personal project?

[mortal delta]

oh its an amd 580 with 8vram.

[warm junco]

Ah okay then you can try the zluda guide. It should work with that GPU. (Needs RocmLibs) from the Guide.
If not you can only use the directml version.

[fervent thunder]

if you freaky then own it

[verbal osprey]

Comfyui is going to be way too overwhelming for a newcomer. Get automatic1111 and learn the basics there.

[mortal delta]

oh also i would like to try sd 1.5 again are there any models any of you would like to suggest by chance?

[mortal delta]

ive tried comfyui which was honestly a mistake it was just too hard.

[bleak matrix]

I really loved history back then, but it's been so long. I've forgotten what that's like! Hahaha! It is for a personal project, yeah! I suppose I'm quite the nerd for everything, but, I guess we all are, bhahahahahaha!

Creative nerds, I summarize!

[surreal jay]

Lol, so long ago that it has become history itself. Yeah I bet us even being here is pretty nerdy on its own. Enjoy your project!

[warm junco]

For realistic stuff try Juggernaut reborn

[mortal delta]

what about like anime or somethign i dont feel ready for realsitc models yet, im not even sure what to do with stable diffusion but im sure ill figure it out.

[bleak matrix]

Fff, yeah, pretty much! Feels like an age ago, so it might as well be! I certainly well! Hahaha! You as well, enjoy your day!

[warm junco]

Divinie animemix and sudachi are nice

[mortal delta]

thank you for the suggestions.

[warm junco]

Np when I'm back home I can give you more

[mortal delta]

you dont have to i think ill be fine with what you suggested.

[warm junco]

Dreamshaper v8 is also an good Allrounder for anything creative

[flint raptor]

is there a self hosted sd3-large ?

[weak scroll]

Waiting for SD3 fine tunes on CivitAI

[flint raptor]

any eta ?

[soft iron]

never

💀

[flint raptor]

and another question:
anyone knows the speed difference between a100 and 4090? currently running on a100 but wondering if 4090 will have same performance

[vast ingot]

well, the code for it got released less than 24h ago, and people need to try it a little, so give it some days to start being good, but you should see the first SD3 finetunes in not that long. anyone can train one now

(and by code for it, I mean the diffusers library got updated with training code for SD3 architecture)

[frail path]

Check vram.

[vast ingot]

24GB Vram on 4090

[flint raptor]

i see my a100 using up to 18gb

so 4090 should be ok but what about the speed

[vast ingot]

well, hard to say. just from the cuda cores, RTX4090 is a little faster than A100, but A100 has so much more VRAM available, you can do lots more things in your workflow before hitting your VRAM limit (and slowing down your process a lot).
So on a pure technical "speed" comparison, 4090 would be the best.
But if you already have an A100 up and running at home, stay on it, you can make up for that lower speed by using the overhead VRAM in so many ways (bigger pictures, controlnet, more models loaded at once, ...)

[desert dagger]

https://opendata.blender.org/ benchmarks. yeah, this is a blender site but you'll still find it useful

[vast ingot]

https://technical.city/en/video/A100-PCIe-vs-GeForce-RTX-4090
this is the once I was checking

[flint raptor]

actually i have it on azure, still have credits but soon will need to buy my own machine

and i dont need bigger pics than 1080p

[vast ingot]

what I was going to answer is outdated it seems (I wondered the same 2 years ago, and bought the 3090TI, but the A100 was almost 10k€ at that time)
Not sure right now. graphics card market is moving so fast

[flint raptor]

yeah a100 is very expensive.... so if the model can fit the vram of 4090 and its faster, i rather buy many 4090s....

btw, its for commercial use (just bought the commercial liscence today)

[somber orchid]

What is stopping the coding from using both ram and the CPU and the GPU to make images?

[fervent thunder]

Need a #suggestions channel for those god awful black nicknames that no one can see if they are using the default theme

white is always an option

[cedar salmon]

yes white background for you would fix your issue

[vast ingot]

the graphics card is what has those "CUDA cores", the ones that are required for diffusing pictures with stable diffusion. You can use the RAM of the computer instead of the vram, most software let you do that if you really want to, but this means that the model is loaded/unloaded multiple times during a single picture, since it needs to be on the VRAM during processing
All in all, if you at least have the 8GB VRAM on the graphics card, you can run most models but at a very slow speed. very very slow. Like 15 minutes per pictures instead of a few seconds in the worse case scenarios

[vast ingot]

yeah sorry, black nickname comes from that Stable Society role, and is supposed to transport that "secret society feel" and be hard to see. I suggested it to be changed back in the days I got it to be honnest, but we don't seem to have a suggestion channel anymore right now
(and people did get it by winning community contests that were run every week => #🏆｜winner-gallery , got mine with this submission #🏆｜winner-gallery message)

[desert dagger]

nothing. it's just a lot slower if you pull the CPU into it

[fervent thunder]

jazzy coffee, nioce nioce

I am just a young sprout what do I know!

For real feel like I am in FF

[jovial ridge]

Knok knok

[forest trout]

Here's the submission I got the role with: #🏆｜winner-gallery message

[gray fern]

about 4 seconds standard, 20 with hires fix

@28 steps

[flint raptor]

What res?

[fervent thunder]

that first one is very interesting, but I am not sure what I am looking at

[gray fern]

1024x1024 for “standard and 2048x2048 hires

i was interested in a100 speeds too

[flint raptor]

Sounds about same as a100, thats great! Thanx

[gray fern]

damn that’s crazy

[fervent thunder]

The winner's gallery is a bit better than I will probably ever do

[flint raptor]

Actually a little faster than a100

[gray fern]

how bout H100 🤔

[flint raptor]

Maybe after going public on nasdaq 🤣

[gray fern]

i think runpod has some comfyui templates maybe i’ll rent one later

[tulip yarrow]

Is there SD3 Lora Trainer yet

[fervent thunder]

guys does SD work well on mac as it works on PC?

[fervent thunder]

I've open sourced my SD3 Discord Bot if anyone wants to use it.

https://github.com/RocketGod-git/stable-diffusion-3-discord-bot

[desert dagger]

Stable isn't machine dependant. but it should run better on the mac, really, than on the PC.

[red otter]

Does sai keep working on svd2 ?

[pure owl]

Hi

[fervent thunder]

Hi

[nova zodiac]

Well no - most stable diffusion codebases are optimized for cuda, which has to be emulated on the mac. Yes there are mac specific torch implementations but they arent as good yet

[hallow nacelle]

@visual lagoon Have you trained any SDXL loras before?

[visual lagoon]

yeah

[hallow nacelle]

I'm thinking of training an animated character for xl which can be used for realistic and anime gens but mainly realism

Last time I used about 60 images with different angles of the character for 1.5 and turned out great

So I did the same for xl but the results turn out way too cartoonish/animated (resembling the original style drawing)

Idk if it has to do with sdxl models or my training

I used 768x768 for sdxl training if I'm correct

Can't remember it was like a year ago

Have you trained anything like this before? @visual lagoon

[visual lagoon]

nah i've not, but i have used loras of animated chars that can look realistic if the base model is meant for it

768x768 training might also be too low for sdxl, standard is 1024x1024

[hallow nacelle]

Hmm, I'ma try 1024x1024 next

[visual lagoon]

maybe find some loras that work

from civitai

https://civitai.com/articles/3168/lora-inspector

then use this tool

lora inspector

apparently shows you the settings they used to make their lora

[hallow nacelle]

Yeah the model is crucial for the lora as well, just remembered SD1.5 Realistic Vision will never generate my chracter lora at all

[hallow nacelle]

Nice, thanks for that

[lime gust]

https://www.instagram.com/reel/C8K2yVXChbB/?utm_source=ig_web_copy_link&igsh=MzRlODBiNWFlZA==

[hallow nacelle]

Looks like a lot of furries, not into that 😂

[visual lagoon]

maybe it was made for that idk but doesn't seem so atm, seems mostly best for anime style

and semi-realistic anime style

and it has a lot of built-in anime series knowledge already

so you might not even need lora for the character you're interested in

[hallow nacelle]

I'm gonna build it anyway

Just found out I used 32490 Max Train Steps for 1.5 lora and 2025 for XL version lol

And precision bf16 for SDXL lora idk if it matters

[fervent thunder]

[mortal delta]

you guys think it spossible to run stable diffusion and other ai on cosnoles like xbox or playstation? just curiouse because this is a random thought i just had.

[warm junco]

If you jailbreak an Xbox I see it could working. As consoles are using AMD CPU/GPU it should work with directml

[wet grotto]

why someone will traine controlnet? It's something else to be trained there?

to same model I don't understand why... to create new controlnets yes

[sudden cobalt]

is clip-l completely broken in sd3 or?

maybe skill issue lol

imma just replace it, seems to have fixed whichever problem it was lol

[dawn mulch]

anyone messed with AnyNode yet?

[vestal dew]

Ive seen articles, youtube videos stating SD3 isn't good with human figures, hands. SF claims it fixes hand issues. Which is it?

[desert dagger]

you've seen a lot of articles, and not one of them have also stated that the version of SD 3 that was released is the low end 2B model - and yeah, it does have issues with hands and feet IF you haven't worked with it enough to know how to tweak the settings correctly

[vestal dew]

I see.

[willow crescent]

talking about tweaking, how come it looks like SD3 adds a preset-style to every image (even if that is not what is technically happening)?

[desert dagger]

that's probably the effect of your scheduler and sampler, but there are other settings that can affect that. Are you using SD 3 inside comfy? or something else?

[willow crescent]

yes comfy and I'm using the default settings

[desert dagger]

okay, but for which workflow?

[willow crescent]

the example workflow from huggingface and I've already checked on many of the samplers and schedulers and so far found the default one dpmpp 2m with the sigma scheduler the best working, yet there is simply that sort of preset style inhabited by the images

[desert dagger]

okay, here are the settings i have for this specific image. this is the basic SD3 workflow from huggingface. i never use negative prompts so mine consists of just . and is out of my way

[gilded wolf]

Give me a prompt for SD3

[desert dagger]

cfg: 4.0, sampler: ddim, scheduler: normal, ModelSamplingSD3: 0.70,

all of those settings can drastically affect the look of your image. I think the defaults were chosen just to try to make sure the images would look okay

[willow crescent]

gonna check, ty

[desert dagger]

an apply spinning in space

[civic gorge]

hey folks

[gilded wolf]

#🏞｜general-with-images message

[civic gorge]

so, question. anyone know what it takes to get finetuning on SD3? like...what specifically we should be watching for to drop?

[trail lion]

the commits in the kohya_ss github is what I'll be looking at

[willow crescent]

checked yet doesn´t change the issue I was referring to. Here is the example I tested it with (with the default settings) yet with the same effect: #🏞｜general-with-images message

Also when using flat-painted input images for example SD3 adds some sort of glossy 3D finish

[civic gorge]

Thanks. notice any movement or still just up in the air? I know there was something about diffusers dropped...which was a big thing...

[desert dagger]

i don't think anything is missing. i think the problem most people are running into is that they are trying to use SD 3 as if it was the previous versions of stable diffusion. but the network is totally different. you cant' just use it like that. you have to learn it all over again

[willow crescent]

yes, that´s what they keep saying, yet I got the feeling like it´s not only that

[trail lion]

seeing a couple loras on civit already, so I was way wrong on that bit...the irony, I dont think there's an easy way to load them, unless some people have a way in comfy

[willow crescent]

And what do you mean by "i don´t think anything is missing."?

[gilded wolf]

#🏞｜general-with-images message

[desert dagger]

download them, add a lora loader node, load them?

[trail lion]

that worked for you?

[desert dagger]

i usualy don't use loras, but the node isn't lora specific, just make sure you use them with the right base model

[willow crescent]

that´s all nice and stuff, yet can you produce anything flat painted looking like this? #🏞｜general-with-images message or you @desert dagger?

[desert dagger]

give me a bit and I'll see. that's not a style i like so not one i've tried yet

[willow crescent]

👍

[desert dagger]

#🏞｜general-with-images message a few different flat painting images

[nova zodiac]

What? If you cant prompt a person lying down in any way shape or form, that isnt a fault of the prompter!

[desert dagger]

have you read nothing that's been discussed about that, yet?

[nova zodiac]

Ive been reading most of the stuff, the latest theory being that someone applied a leco to the model to obliterate certain concepts, which holds up given that it can do lots of other things quite well. What did I miss?

[desert dagger]

the theory only looks good on paper. please explain to me exactly how it is possible to rmeove anything from the weights once they are done - you can't very well jsut edit the files

and it's not the humans that decides how the data is stored, it's the AI

the humans still dont' know HOW it does what it does

[nova zodiac]

Its a tensor - if you know where you are aiming, you can overwrite areas by rewriting the data easily enough

[desert dagger]

nope. sorry. it's 100% someone's assumption becasue they havne't worked with it enough to learn how to use it. everyone that's saying this nonsense is also trying to use it like it was previous versions of stable. it's an entirely new thign, with a totaly new type of network.

[dawn mulch]

@gilded wolfA cow sitting on a ladybug.

[desert dagger]

black cat in a coal mine

[nova zodiac]

How that affects the end result, we dont know, but you can definitely zero out weights on the first layer if you know what the encoder returns. And its not a new network. Its new for image gen, but its still the same as other multimodal models like llava

[gilded wolf]

#🏞｜general-with-images message

[dawn mulch]

oh thats fast

but uh, ill have to rate it.

[desert dagger]

it's a totally new network for 99% of the people that are having issues using SD 3, and who are assuming that because THEY can't get it to work, then stability did 'something" to it - all the way from taking stuff out (nonsense) to deliberatley making it worse (stability.AI isnt' openAI)

[nova zodiac]

Well they had to do something negative to it, given that same prompt and seed comparisons with the “same” medium model on the API give different results. If its workflow then why would they post an inferior workflow with the model card??

If its not workflow, then the other theories may hold credence. Either way, theres a lot of speculation and a lot of memes of lying in grass at SAIs expense

[desert dagger]

except they didn't. it's a low end model. the API is using a higher end model

[nova zodiac]

The api has medium, large and turbo no?

It does in the api reference…

[broken smelt]

didnt an sai dev confirm that the sd3 medium model was literally a newly trained model different from the one on the api

[nova zodiac]

Mcmonkey4eva on reddit I believe now you mention it

[desert dagger]

not sure but i think that's all it has. i don't think it has ultra yet

[desert dagger]

that would be @finite cloak and i would listen to him

[nova zodiac]

Ultra, Core, and the 3 flavours of sd3 in the post endpoint as at this moment in time

[desert dagger]

okay then if you're using the API you can run the same prompt with each and make comparisons fairly easily

[nova zodiac]

Im not on the api, but have seen reddit threads from others that have, some results are similar but even a picture of a single flower was different between the two models

[desert dagger]

shoot, you don't need different models for that, you just need to change one setting on the current model

[nova zodiac]

Yeah but for identical settings they should match

[desert dagger]

no they shouldn't. why would they match?

you're only going to get that if you are using not only the same settings, but the same model.

[nova zodiac]

My Point exactly - the medium model in the api SHOULD be the same as the one released but it isnt

[desert dagger]

it is. what isn't the same is what else they are running along with it for the API - and there's no guarentee they aren't running prompt enhancers or filters (they are defintley running filters), and other things

and they are almost certainly running prompt enhancers

[nova zodiac]

Im a quit pissing into the wind here and go back to work. Something fishy is definitely afoot, and it will take some time to fix it. In the meantime we sit and wait. Have a great arvo/morning/rest of your day!

[late surge]

hi

[broken tiger]

Can i use this discord to generate images privatly from another image, or chat with a bot?

[desert dagger]

privately? everything you prompt on here is public

[verbal osprey]

Privacy? Imagine you're using an Adobe product. There is zero privacy now.

[stray wolf]

i need to get more deditated wams

SD3 is gonna give my SSD cancer

[desert dagger]

love the people that are so positive that SD 3 medium is lousy that they have to find something tiny in an image to nitpick

[stray wolf]

i think sd3 is insane

especially for a foundation model, finetunes will probably make it even crazier

i think i just barely managed to fit the whole thing without having to swap

now i can spam images

[desert dagger]

@finite cloak do you know of any technical documentation that talks about how the ModelSamplingSD3 node in the ComfyUI workflows is actually affecting the network and what it's doing under the covers?

[finite cloak]

Not "new" but yes it's a different model.
SD3 at time of the initial research paper was 4 models: 800M, 2B, 4B, 8B
The API model is an early preview of the 8B
The released SD3-Medium is the first completed training of the 2B.

Note that the larger models tend to have better structure and understanding, but need much more training to have clear quality.
Medium (2B) has weaker understanding but better quality.
Note by "quality" I mean eg if you zoom in close and look at fine details, 2B does so much better than the 8B preview on API does. (vs the posts people on reddit have of the corrupted humanoid blobs are structural/understanding issues that the 8B is better on).

[desert dagger]

it needs come conceptual training on humanoid creatures laying on flat surfaces. other than that, it's really working well

[finite cloak]

explanation #🆕｜sd3 message

demonstrative graph #🆕｜sd3 message

(full disclosure offhand i'm not sure if i inversed which way is which or not but the concept's all explained well there, and you can play with the value eg in a Grid to see the real world affects)

[finite cloak]

funny thing is if you just prompt it differently it doesn't explode anymore. Like, a woman lying in grass breaks but (a woman lying in grass:0.5) doesn't, or a painting of a woman lying grass doesn't either. There's something weirdly specific causing the fleshpile explosion to be so bad

(generally just needs investigation still)

[desert dagger]

i've been playing with it, i've spent since the release just changing one decimal point at a time 🙂 I'm wanting the deep dive into the actual funcitons. i'll look at the links

[finite cloak]

you can make a grid of it easily in swarm like so:

[desert dagger]

oh that's interesting

[finite cloak]

(i love making grids to understand what parameters do)

might add an axis for CFG since that parameter has tied in effects

[desert dagger]

i don't use swarm, just comfy. I'm planning on doing an document though with a bunch of example gens in it. so far, almost ever 1.5 prompt i have is returning expected, but better results

[sage reef]

@finite cloak hey, just something i want to ask you. according to you, what would be the main reason as to why the released 2B model generates some weird human anatomy? im not talking about the hands, i am more talking about some of the generations where the human looks like spaghetti or just lot of missing or disconnected body parts.

[desert dagger]

yeah. in order of critical: ModelSamplingSD3, then cfg, then sampler/scheduler

and at this point i can not get eulerA to return anything but noise and blobs of color. Euler works however

[desert dagger]

just tried this in comfy, and it warped her all over the place

[opal hedge]

Not an SAI employee, but it's probably due to a hasty effort to censor the model

[desert dagger]

yeah... no

[broken smelt]

interesting, i had just now gone over the technical paper of sd3 and this part stuck out to me

We also observe that some variants perform well in some
settings but worse in others, e.g. rf/lognorm(0.50,
0.60) is the best-performing variant with 50 sampling
steps but much worse (average rank 8.5) with 5 sampling
steps.

[opal hedge]

Uh... yes.

[broken smelt]

had just seen comfy talk about timesteps too

[sage reef]

hmm... idk... i hope this gets fixed one way or another :3

[broken smelt]

sounds like more fun stuff to experiment with

[desert dagger]

nope. i can make highly NSFW images with SD 3 just fine. it's got nothing to do with that

[stray wolf]

are we sure there was censorship, or was it just a lack of NSFW in the training data

[opal hedge]

Then why are landscapes perfect, but only human anatomy grokked to hell

[stray wolf]

besides, im sure finetunes will give us all the booba we need

doesnt seem like a good move business-wise for them to train it on NSFW images

[desert dagger]

i'm gonna let you do two things - 1. decide to stop assuming stuff, and then posting it as possible facts when you have no information and 2. search this discord and read

[desert dagger]

oh you should see the images i got by accident today that i can NOT post on this discord. it's not the least bit shy about making any of that sort of content

[opal hedge]

OK I'll read SAI employees saying the meatpiles being a prompting skill issue mb

[sage reef]

but i mean are you basically saying the following: they removed so much nsfw or suggestive content that the model basically forgot what a human looks like, even in a simplistic scenario like someone lying on the ground or grass, etc?

[desert dagger]

tell you what, why don't you go install it on your machine, and start experimenting to see what you can get it to do instead?

[stray wolf]

sd3 can booba?!

[desert dagger]

not only can it, it can be hard to keep it from doing so

[stray wolf]

oh my

for me it fails to do so

[broken smelt]

it really is a skill issue tho, there are so many levers to mess with with sd3

[stray wolf]

maybe im bad at proompting

[stray wolf]

all these new nodes have me confused, i have no idea what shift is supposed to be

[desert dagger]

depends on your settings: sampler, scheduler, what you have other things set to. and if you're using it local or you are using it somewhere that has filters running

[opal hedge]

And none of them give good anatomy with any precision

[desert dagger]

you ahve absolutly no idea what you are talking about.

[opal hedge]

You can cherry pick 1 out of 100 but I haven't seen it yet

[broken smelt]

comfy explains that here #🧣｜comfy-ui message

[desert dagger]

go to #🏞｜general-with-images and scroll through, you'll see LOTS

[opal hedge]

Alright, post your perfect anatomy workflow in sd3 images and I'll try out the prompt

[desert dagger]

nope. not giving it to you, you're too much of a troll

[opal hedge]

Kek

[finite cloak]

https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Why Use Swarm.md

[desert dagger]

thank you 🙂

[finite cloak]

#💬｜general-chat message

[sage reef]

@finite cloak huh interesting... il try that tomorrow

[desert dagger]

and you could have simply said that the ModelSamplingSD3 node is moving the timesteps

[broken smelt]

the technical paper says that . 5/.6 gave the best results when used with 50 steps but gave awful results when used with 5 steps. changing the timesteps would be changing when to stop structuring the image as a whole and when to focus on detailing. explains the skill issue and wide variance of outputs we're seeing.

[finite cloak]

Yes "Euler A" the "A" is "Ancestral", which is similar to "SDE" samplers the "S" is "Stochastic", both of these words effectively mean "noise is randomly altered between steps". These samplers are incompatible as SD3 uses Rectified Flow sampling, ie it's designed to linearly design one set of noise directly to an image instead of individual steps separately moving random from noise towards an image

[broken smelt]

man people need to start reading more technical papers, it goes over rectified flow in the first paragraph lmao

[desert dagger]

yeah. but dpm_2_ancestral works. something's odd with EulerA

[desert dagger]

how many of the people in this discord have you seen posting since the SD 3 release who seem even slightly technical?

[broken smelt]

man i wish i could read

[opal hedge]

It's not even just the meatpiles that are bad, the hands are somehow a step back from sdxl. I'm getting up to 10 fingers per hand

[sage reef]

does this mean that we need a new sampler that is very tuned to rectified flows?

[finite cloak]

If you come here asking for help to improve your gens, I'm sure many people will love to help you. If you come here wanting to cause conflict, all you're gonna get is conflict. And if all you want is conflict... please take that somewhere else.

[desert dagger]

the denoisers seem to work fine

[finite cloak]

Imagine the utopian society we could have if everybody was eager to read technical/scientific papers

[opal hedge]

I don't want conflict, I was posting my opinion before crystalwizard yelled wrong at me. So he should leave too I assume

[desert dagger]

we'll need somewhere other than arXiv to put those papers first

[sage reef]

i personally found tons of cool papers on arxiv when im working on some projects

[desert dagger]

sure, but they also have stuff like the one i saw today where the authors basically refuse to provide any proof to what they are saying, and offer to let people contact them if they want it

[sage reef]

can you link that? im curious

[desert dagger]

if i can find it again

[sage reef]

because depending on the publication, some papers have a page limit, and maybe the proof part cannot fit in the paper

[nova zodiac]

Im happy to read technical papers, but id still like to know why a headshot photo of a human works great and yet (an elegant woman lying in the grass in a serene and peaceful manner:0.5) gives me eldritch horror (or woman lying on a car bonnet etc)

[desert dagger]

switch over to #🏞｜general-with-images and read through the chat that was just taking place with alex

[sage reef]

i actually dont really use any video models myself, so sadly no idea. but the license information should be available in the model card or readme, etc, there are a lot of projects these days, like open sora, svd, etc, but i have no clue about the licenses sadly, sorry :3

[desert dagger]

pika has commercial licenses

[olive cradle]

hi im new. using img2img, how do I get a filter/rotoscope?

[desert dagger]

no, but pika has commercial terms.

[pearl ocean]

guys for ComfyUI I downloaded "sd3_medium_incl_clips.safetensors" and "sd3_medium.safetensors", and put them in the models checkpoints folder.

I also downloaded "stableDiffusion3SD3_textEncoderT5E4m3fn.safetensors"
And put that in the models clip folder.

Wheres is the node thing for me to add the SD3 Text Encoder?
have I done the right stuff?

[sage reef]

im assuming you skipped this: https://comfyanonymous.github.io/ComfyUI_examples/sd3/

[late surge]

hi

[pearl ocean]

Weird, it runs off the file called Clips, but doesn't fro the medium

[sage reef]

well again, maybe you misunderstood how this all works. the one that is called medium strictly is only working with the other components separately, while the incl versions have everything included. so you have to choose what pipeline you want, either just components or mixed, you cant mix match

[pearl ocean]

😮

oh cool, it actually does words well

[sage reef]

that is why that page has two workflows, mixed and components

[mossy river]

hey yall, is SDXL seed space discrete or continuous?

[narrow kernel]

Because those kinds of poses are also very heavily associated with concepts pertaining to NSFW content. This model was meant to be as safe as possible and the vast majority of users that will use sd3 will never prompt for people in "exotic" poses. It's not impossible to get decent laying photos, but you'll have to word things better aka not use prompts like 1grill, splaying in grass, bobs, vegine, waifu, 4k, 8k, 16k, 32k, 64k, ♾️k, best quality, breathtaking

[tight mason]

are there any good podcasts regarding stable diff?

[sage reef]

hey guys welcome to my podcast... so today... i generated a lot of booba... 🤣

[pearl ocean]

Sometimes I wonder if SD3 is actually more realistic.
It definitely is with text.

[nova zodiac]

As a base model doing a head and shoulder portrait, it is definitely more realistic than sdxl or cascade

[pearl ocean]

oh yeah, I forgot they ahven't released the full model yet

[fresh barn]

Hello, I have a question. Recently, I was using SD Automatic1111 and installed the CivitAI Browser+ extension. When I restarted the Docker, the new extension (Colab Command) started, and my RTX began mining due to a script called "miner." Has anyone else experienced this problem? I have now deleted that extension and removed the files from my Docker, but I’m not sure what happened.
Imagen

https://media.discordapp.net/attachments/1002602742667280404/1250889607533301922/Captura_de_pantalla_2024-06-13_a_las_20.35.52.png?ex=666d3dfa&is=666bec7a&hm=d6f67f3fe4adc18a3133aac4a87aa001976a1b7e2fb66066d6550ab2cb416c33&=&format=webp&quality=lossless

[charred thistle]

seems to be a lot of updates since I checked last

[pulsar venture]

hi, in the end how is good SD3?

a slight step forward or a bit step forward?

[fervent thunder]

I have not used it, but I understand it's a huge leap forward in censorship.

And isn't that what we were all hoping for!

[pulsar venture]

Explain please?

[vast ingot]

This is really subjective @pulsar venture . People don't like the anatomy right now it seems, I think this is what Midare is talking about ?
For me right now, it feels like a big step forward on some front, and a step backwards on others, especially compared to finetunes of SDXL.
The thing is, we need to adapt our prompting style to that new model for a start : there are 3 clip models now, so 3 prompts to be used. So it's hard to compare directly to XL
But also, there are great leap forwards technicaly, some features that just didn't exist, like text drawing, or being able to specify in long prompts multiple subjects

[wet grotto]

hi

its possible install forge on colab?

I'm using fast stable diffusion notebook

[warm junco]

Forge is not recommended for normal users anymore. With the next update it will break everything

[wet grotto]

@warm junco but its possible? I can experiment

[pulsar venture]

Usually i had used some lora from anime (Chizuru Mizuhara from rent a girlfriend) but the character didn't like as i want, so the lora are now more precise?

[warm junco]

I don't know but its possible to get it working on docker or other cloud services

[wet grotto]

I'm using colab and now Kaggle

kaggle disponibilize two T4, 32vram o/

[vast ingot]

LoRA are trainings done on top of a model. LoRAs for SDXL or for SD1.5 don't work directly with SD3.0 for now, from what I see currently at least.
New LoRA trainings for 3.0 should come out at some point, or ways to use previous LoRAs with this new structure of model, but as the architecture of the checkpoint has changed, I'm really not sure this will be possible. We'll need to wait a little to know more

[low moon]

Then never update.

[warm junco]

Yea but then it won't get SD3 support

[pulsar venture]

So the loras didn't get betters?

And isn't possible to train a lora with SD3?

[warm junco]

It will be possible in the future

[pulsar venture]

Eta is not available right?

[wet grotto]

anyone use this?
https://github.com/BlafKing/sd-civitai-browser-plus

civit ai browser plus

its good? how it works? don't have to download the model from civitai?

[warm junco]

Some people use it.
With that you can search and download any model or lora etc from Civitai.com directly in the webui interface

[vast ingot]

LoRAs don't "get better", never. They are trained like a model is. A lora is like an add-on, a mod on a game. it changes the model to fit more your given concept, the one it was trained on. But it is made for a single version of a model. It does have some cross use when the base model you put it on is of the same form/architecture, but other than that, it's about as usefull as a brick if you change model. A new LoRA will need to be trained for that new SD3.0 model, yes.
ETA is not out yet, but 2 days ago SD3.0 got released, and the code for training models/LoRAs got released the same day. Finetuners are individuals, the LoRA you got is not produced by Stability, so the question would be : when and if that individual will update their lora, do a new training from scratch

[wet grotto]

sounds good, must keep the information of models, can help me

[warm junco]

Yea it also saves the infos, trigger words and you can let it download the preview images too

Very handy for webuis in the cloud or docker

So you don't have to move stuff around manually

[wet grotto]

I don't undestand this last part

I'm using Kaggle for example

you are sayng that it will download the model in the correctly folder, is it? so I don't have to move around manually

[warm junco]

I mean that the extension downloads the model directly into the correct folder. So you don't have to download the model manually and move it around on the file system level

[pulsar venture]

Ok, so if I understand right, i need to wait for new version of the lora trained on an SD3 model, and if it is done, the lora "get better", right?

[wet grotto]

nice

my kaggle dont open the link to a1111, give me a error msgm

anyone know how can I corret it?

I put the port on firewall but yet keeps blocking =/

[vast ingot]

that new LoRA would be usable with that 3.0 model. Your current LoRA could also be trained better on the model it's comming from (not sure but I imagine it's for SDXL?).
Best thing you can do right now if your lora doesn't fit your need is one of the following :

• find another lora trained better. that would most likely be a lora for SDXL currently
• wait for a new lora for that character to be trained, for SDXL or SD3.0.
• train yourself that new LoRA
  LoRAs are "fan made" mostly, and some are professionally made but rarely for free, and their quality depends on how much work and knowledge is behing them, how much finetuning, and how much good dataset is available to train them.
  LoRAs for 3.0 will also be like that : their quality will only depend on how they were trained. The quality ceiling will get a little higher, but it's already quite high with SDXL

[pulsar venture]

thank you

[tropic frost]

Anyone here know the AI artist Yu Arisato?

[ionic crown]

Hey, how can I generate images? And do I need to pay for them now?

[tropic frost]

most image generation platforms charge.
The most popular one seems to NovelAI.
The most popular free one is stable diffussion, but you gotta get it up and running on your PC, if its powerful enough to work.

[restive robin]

————————
A note to stability AI.

First, I appreciate the effort that went into creating SD3, it excels at creating landscapes, text, objects, rooms etc etc and has phenomenal prompt adherence. I can make ridiculous things like a pixelated volcano encased in a bubble with the bubble itself containing an icy city and it comes out perfectly.

I can make silly things like cats with multicolored bows on their heads or neck jumping around, digging in the beach, or flying planes.

I can make text made out of ice that is on fire without any confusion on what’s supposed to be ice and what’s supposed to be fire.

And when promoting only a single spacebar and hitting enter I can get anything from vivid city landscapes and oceans, to artistic portraits of people with perfect anatomy mind you, etc etc so I know the model for sure has really good data.

HOWEVER, and it boggles the mind. It’s as if the ability to pose humans for anything other than a headshot portrait or standing straight like a mannequin has been surgically removed.

I’ve tested with posing various mammals, birds, fish, even chimps and they had far less of an issue with anatomy.

Keep in mind that I know the model has the capacity for humans as the blank prompt produced realistic men and women.

So, please explain what has occurred for SD3’s 2B public release model to have worse anatomical problems in specifically humans than the very first model you’ve produced.

In addition, I implore you to make the same model as the API, be it the 4B or 8B available. Without whatever restrictions were placed on 2B to specifically mess with human anatomy prompting.

I truly believe that SD3 can be a fantastic model from what I’ve seen in my own testing and the amazing things people have managed to make (that are not human)

So, let’s have an open and honest dialogue while some people still give you the benefit of a doubt.

It’s not producing the anatomy that was promised, and as of this morning with people trying to fine tune it, it’s not easy to train like it was promised.

So, what happened, and what’s the plan going forward to address the community’s concerns other than one person refusing to acknowledge the problem, insulting fine tuners, and effectively giving us the finger saying to “get good”?
————————

Because let’s be honest, if you want to have a good reputation which will attract investors.

Being known as the company that intentionally mangles its product is not a good look.

It’s a much better look to say “here’s the issue, and here’s how we’ll move forward, we have a plan of action”

————————
That’s all I will say for now. I’m giving the largest benefit of a doubt I ever have here. Let’s be open and have an actually meaningful discussion.

[restive robin]

(Shortcut to the top)

[vale steeple]

Dont you think it is better to post that in Reddit than here

[restive robin]

Maybe, but let’s be real here. They’re not looking at Reddit. So I’m coming to their front door so-to-speak.

[trail lion]

even better yet there's a contact link on their website

so it'll go right to the staff

[vale steeple]

You realized the only real devs that could see this are comfy, mcmonkey, and Lykon

The rest is just there not touching Discord

Lol

[trail lion]

I dont care about the post, personally, but it's unlikely to be read by their staff here being realistic

[vale steeple]

They are not the main dev of SD3 either

One is UI dev, one is postprocessing, and one is also another UI dev

[restive robin]

True, but there’s also the very likely outcome of… it being ignored.

Here there is at least a somewhat public record that the question was asked.

I shall follow your advice to contact directly. But I shall also leave this here as a show of goodwill/that there is some level of a benefit of a doubt if there is actual discussion.

[restive robin]

Thank you for the information

[trail lion]

fair enough, maybe do both

[vale steeple]

But they definitely got pressured after the dissapointment they getting with human anatomy

Especially the majority who dont actually have in depth technical knowledge

And saying "You know what, this is what you get by removing NSFW"

[restive robin]

True, just figured I’d actually try to throw them a bone to make a semi-public statement that they recognize the issue and have a plan of action.

Which would be a MASSIVE improvement over insults and intentional silence, and would actually salvage their reputation as of now.

[vale steeple]

And "OpenAI release their paper about using NSFW to train anatomy"

[restive robin]

(The bone)

[proper scroll]

do I have to train loras with sd3 to make my loras work with sd3?

[still glacier]

yes and not possible yet

[proper scroll]

I don't understand nwo

now*

I can't train loras with sd3?

[trim nymph]

it would have been wiser to not release anything, hope was already lost anyways

[still glacier]

Apart from this method https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_sd3.md I dont think any tool has been updated for sd3 training yet.

[vale steeple]

I mean you can but it is tricky

We got Diffusers training

And SimpleTrain?

Simple something?

[still glacier]

diffuser yup

[restive robin]

Who knows? Maybe they’d actually be willing

[restive robin]

To have an honest discussion

[proper scroll]

what is a diffuser training?

[vale steeple]

Uhhhh...

Simply saying it is a package made by Hugging Face to run and generate stuff on code

Library for pretrained diffusion model a little bit advanced

[vast ingot]

Diffusers is the underlying code library powering stable diffusion (thus the name). Diffusers training is also known as fine-tuning : it's training a base model even more using a restricted dataset usually. It's how most models shared on civitAI or hugging face are made

[still glacier]

gui-less training, no visual interface

[vast ingot]

Yeah, this is the inner "brick", this is the cement of model making, and not the easiest to use. All training method use it, but usually it's hidden deep under the UI you use

[bleak matrix]

Goor mogning, everyone! How are you all doing?

[vast ingot]

Good morning Sunny ! doing HOT over here ! temperatures are climbing like crazy over here ^^ and you ?

[nova kindle]

Hey are there any sections on this discord documenting current optimisations?

[vast ingot]

not sure we do have those. #🤝｜tech-support may be of help if you do have some specific things you need to optimize, but general help/documentation on this, not on this discord, since we don't promote a specific tool to use for Stable diffusion. Each and every tool has its own github that does have a wiki usually, and some optimization parameters are listed there. Usually tech support can point you to the ones you would need

[nova kindle]

https://tenor.com/view/my-man-my-man-hd-my-man4k-gif-23532578

thanks!

[past frigate]

Holy shit is it too much to ask that if you are being puritan and censhoring shit despite NSFW community being pretty helpful to you,and insulting said community, THAT YOU ATLEAST DONT MAKE IT IMPOSIBLE TO GENERATE HUMANS ? You went so far you created Ctulhu Abominations. Wtf ?

I tried SD 3 and jesus christ You guys overfocused safety like some white neolib Karen.

[proper scroll]

dude its their model

they are free to make it the way they want

[sly eagle]

Are Loras working gor you in automatic1111?

[steep arrow]

And we are free to remind them that being Hr representatives will lose them support

[proper scroll]

Hr representatives?

human resources?

[desert dagger]

you're free to say anything you like, but if you want people to pay attention, there's a good way and a bad way to say it

[restive robin]

Well hopefully this is a good way then.

Every hour I will make a shortcut back to this note for anyone at SAI to see. And the public as well to note if they remain silent.

If they wish to have an honest and open discussion, I am holding the door open for however long they need to think.

[past frigate]

Thank god for Muslim Migration to Europe lol, if you only want the part of Liberalism that benefits Neoliberal Feminism, then why should Liberalism be defended ? apparently OF is okay because it empowers women but you know Gen AI Nsfw is BAAAD because its competetion and god forbid Feministic Karen dont get special treatment.

A "THE MESSAGE" Yeah you are just Neolib

And only the Progresive "Messages" are good right. Jesus fucking Christ

[steep arrow]

Absolutely, but I have been following stability since sd1.5 and they have the same problems with communication and model training everytime. Then we have a confusing mess with a bunch of people saying "get gud" and "its a skill issue" with a model that has a fundamentally different model structure, and presumable issues in associating input text to the latents.

Stability did nothing to explain how to prompt, or what to expect. The composed elements are disappointing, and claiming it is the users that are the problem is only going to make this worse.

Why can't stability be honest?

Did they censor the model to make it safe and thusly crush the models capacity for anatomy? Is it something else with the architecture?

They did something to lessen the capacity of the model, and I would like to understand it instead of getting gaslit. But knowing the community, I expect I will just get more gaslighting.

[pale latch]

just get gud. ez. gg

how fragile are you when "get gud" destroys your whole world?

[hazy gulch]

#✍🏼｜rules-and-tos

[steep arrow]

No, it doesn’t.

Its just fruitless banter that shows your immaturity.

The sd3 chat is full of “what this doesn’t have problems, with my undisclosed prompts and process I can do it no problem”

This is a toxic way to behave, and causes more confusion.

I have had plenty of success, and none of it comes from the kind of snark so many want to deliver.

[pale latch]

skill issue

confused? just get gud

[steep arrow]

Send an Application into stability, you have a perfect mindset to work with them.

I am going to make images and ignore this nonsense, because nothing ever comes of it, and you can flex your “skill at deciphering poorly implemented software” and I hope that makes you feel good.

[past frigate]

Check general chat with images to see what i mean when i talk about them censoring them model, its not possible to have males be normal with normal artifacts and having women have ctulhu legfoots without them trying to censor NSFW and fucking it up.

i get the feeling that male anatomy is better and that points to this.

[pale latch]

if you're already gud i dont know why your ego would be so destroyed by "get gud" then

[steep arrow]

My ego is fine kid.

You are unproductive.

[pale latch]

The funniest thing is that "get gud" has been around for decades and no group has ever vehomently declared it's offensive until now

[proper scroll]

err kind people, I was seeing the examples of SD3

and it asks for clips that I do not have?

where could I possibly get this clip files?

thanks in advance

[pale latch]

they're in the civit ai page too i'l get a link

[proper scroll]

thank you very much

[pale latch]

https://civitai.com/models/497255 yu'll see at the top. various text encoders and clip models to select. you'll want to refer to a starter guide to know which ones you''ll want and where they go

https://education.civitai.com/quickstart-guide-to-stable-diffusion-3/

[proper scroll]

I see, thanks a bunch!

another question, in which folder do clips belong in comfy ui? xD

ah models/clip

[pale latch]

you got it

[hardy sedge]

What are the best settings to train an SDXL character model with kohya at the moment? All the videos on YouTube are pretty old. For example, how many steps are necessary? What other settings should be made? Thank you ...

[restive robin]

Another hour, another shortcut to the note if SAI actually wants to talk to the community

[soft iron]

They just lie endlessly, do shit, and don't try to save their situation

They have problems with the community

They have financial problems

They have problems with the quality of models

And they don't want to solve any of these problems

[restive robin]

Oh I’m well aware. I’m giving the opportunity to come clean and set out a reasonable plan of action.

That’s the only thing that could possibly salvage their reputation

This is not for talking points

[soft iron]

I don't think anyone will answer you

[desert dagger]

it's a low end model. they didn't do anything to it after it was trained. anything that would have been done is during training. Alex also said, several times yesterday that it isn't finshed, and he bluntly said they rushed it out because the community was impatient. The biggest issue is that the images people have seen are from 8b, not 2b - and Stable medium is 2b. It's got better prompt coherance than SDXL or other versions, it does a better job with a lot of things - but at the end of the day, i'ts just one tool. If it won't do what someone wants, and they don't want to use the beta test version of 8b that's available in various places - then they should use a different tool.

[soft iron]

model 1.5 has an excellent training base and knows more than other models

subsequent models were completely castrated

Problem 2b is castration, not that it is small

[mint fiber]

Why is the smallest one called "medium"?

[desert dagger]

and is unique in other ways. all the models have their own pros and cons - they're all different tools in the toolbox. it's sort of like cooking in the kitchen. if you're a chef, you have a lot of spices and tools, and you use the right ones for the desired out put. you don't try to make just one thing - salt or paper - or a fork - do everything

[desert dagger]

that's how it's called in the API

[soft iron]

no, they just cut down the dataset due to problems with copyright and other nonsense

[subtle badge]

It's (ostensibly) not the smallest one

[soft iron]

the basic model is basic because it should know the maximum

[desert dagger]

not sure why anyone is calling it the smallest - it's the low end model. Use it for what you need it for, use something else for other things

[subtle badge]

I mean it is the smallest (and largest) so far with weights publicly available

But there's an 800m version 'in the works', right?

[desert dagger]

the API is publically available. we're using SD 3 ultra on one of the discords i'm on.

[desert dagger]

you can play with that one in the artisan channels here if you don't want to grab the API and impliment it somewhere

[subtle badge]

Oh really
Neat, thanks

[desert dagger]

after you put in your prompt, click the +5 more at the end of it, and then for model, choose ultra

[subtle badge]

Oh I thought you meant the 800m is available
As a 4gb VRAM peasant I'm really curious about its performance

[vocal leaf]

random question but how do i make my own GPT assistant without limitations?

[subtle badge]

Probably a llama derivative would be your best bet

(Not technically GPT though)

[vocal leaf]

does it take coding experience to use?

[subtle badge]

Not to use per se, but probably a bit of poweruser skills for the setup I'd guess, and more if you wanted to train it

[vocal leaf]

ill look up tutorials online, is there no other better ones or is that the best one?

I know someone mentioned AgentGPT

by langchain i think

[subtle badge]

Honestly I don't know much about the current state of LLMs
I know llama made waves by not pretending to have a doomsday machine like the clowns at openai

[vocal leaf]

thanks, also what you mean by doomsday machine?

they fucked up something?

[subtle badge]

Nah, openAI likes to act like if GPT got out into the public it would be horrible

[verbal osprey]

People having access to information is a positive, not a negative.

[vocal leaf]

ah, anyways what discord server should I join to discuss about LLMs n stuff

[restive robin]

Another hour, another shortcut to the note if SAI wants to address concerns and have a semblance of their goodwill back from the community.

(This shall continue until open dialogue is had. And during the wait it shall measure the silence)

[desert dagger]

it's available in the API, it's available in the #1237459938901491852 channels here

[desert dagger]

something tells me that if you really want their attention, you should email them at their corporate email

[subtle badge]

Issue with this is that there is no PR cost to just ignoring such an email

[desert dagger]

issue with this entire thing is that it's just discord noise. you want them to take you seriously, use the correct communication channels

i.e don't post on twitter or discord, email them at their official email address

[restive robin]

I have already been advised as such, but let’s be honest with ourselves. If they actually responded to these we’d already have the full story as… you bet they got a LOT of emails Thursday.

Therefore I have decided that the correct avenue is to put it in the public eye. Meaning that there is now a measurable amount of silence from their official server if they choose to remain quiet.

[restive robin]

I have also been careful to give them the highest benefit of the doubt. The model CAN be good, but it’s not with humans specifically.

I believe it’s not too much to ask why

[desert dagger]

you're on discord - a fairly toxic at the best of times social media server that is mostly focused on games and gaming. you want to talk to the people in charge of the company, this is NOT the place to do it

i highly doubt they read anything posted here. they're also not likely to be reading posts at the exact moment you post yours. and if they aren't, they're never going to see it in the post spam that happens in these channels

[restive robin]

To each their own. I don’t disagree with your points.

However there’s not many options left when nothing is being said anywhere period

You are correct in what you are saying, just as it is correct that they are not communicating in their contact emails.

[desert dagger]

except a LOT has been said, by the devs and by stability itself. what IS not happening though is that people are NOT accepting what is being said. the internet is currently full of people with no real knowledge making assumptions, screaming, hollaring, and spouting off things as fact that are made up entirely in their own imaginations.

and others are picking that up and regurgitating it

[solemn sky]

blblblblblblblblblblblbl

[desert dagger]

and trolls. let's not forget the large number of trolls everywhere too

[restive robin]

Please explain then.

I’d be more than happy to learn. (This sounds sarcastic but I genuinely mean it.)

[solemn sky]

trollolololololololololol

[restive robin]

It’s entirely possible that I am misinformed

[desert dagger]

start by searching this discord for posts from @finite cloak and read what he said, fairly blunty, yesterday evening

he's a dev. he rather knows what he is talking about

[solemn sky]

i want the small model plz ! 😄

[restive robin]

To summarize. 2B was intended to be a beta but is was not labeled as such… because of unspecified reasons.

It needs to “cook more”, if I’m to borrow some of the wording used, to understand humans.

This means that at least the issue is acknowledged on their end, and it seems that they are aware of the awkward prompting.

However there is no timetable for fixes or updates as of now. (Other than vague promises of 8B and that they’ll circle back to 2B after that is released…. With no timetable for that either.)

Do I understand correctly?

[desert dagger]

sure. Alex said that clearly yesterday. he specifically agreed it wasn't finished, and bluntly said it was rushed to release because the community was impatient. so y'all got an unfinished model. you wanted it, you got it. now play around with it and let them work on fixing stuff and finishing it

[subtle badge]

Thanks for this

[restive robin]

Thank you for the information.

I personally don’t think it makes sense in the slightest to not call a beta a beta however. It would have been the easiest excuse in the world to have for complaints. Again, giving the highest benefit of the doubt here: It makes them sound like they just said it to retroactively cover their backs even if it’s genuine because of the current optics.

BUT! I appreciate that you’ve pointed me to this as up until now I was under the impression that they were simply not saying anything other than “get good”

Turns out there’s at least an acknowledgement of the situation, just no clear plan of action communicated yet.

I shall limit to asking for a realistic timeline then as that is still undefined.

[soft iron]

https://civitai.com/models/497255/stable-diffusion-3-sd3?modelVersionId=552771

[desert dagger]

doesn't do any good to call anything a beta. and they didn't say it was the final product, they just said they were releasing the weights. however people have been told over and over and over it's not finished and they, one and all, ignore that and continue to rant

yeah, or put Not Applicable

[restive robin]

I understand that from your standpoint you have definitely seen that they may have said it’s unfinished.

However from my standpoint up until ten minutes ago and the majority of everyone who was just casually looking… released means released as in full. (Especially if they’re promoting it on their website with no clear indication it’s still under development, only that it’s the most sophisticated and that it’s out.)

This is actually exactly what I hoped to achieve. Open dialogue so that people like me can be informed of what’s actually happening.

And now I know.

If anything though I’d suggest making an official announcement instead of relying on the grapevine to communicate this.

#📣｜announcements message

This one here for example says nothing about it being a beta or unfinished as of now.

If that would have been CLEARLY communicated everywhere, I bet there would have been much less vitriol

[desert dagger]

i definately was sitting here talking to Alex yesterday, and talking to the other user he was talking to when he specifically said it was unfinished and agreed there were issues. i was talking to him about technical stuff.

[desert dagger]

i don't think that's necessary. why? no one would listen. they don't listen when the devs tell them to their face, they aren't going to listen if it's posted in annoucements. all they do is post about how people are lying and go on with wild assumptions.

[restive robin]

Not everyone is up late at night on discord to catch things in the moment.

[restive robin]

Hence it would be good to have it somewhere where it won’t get lost.

And it should have been done here where it won’t get lost from the beginning

Of course now people wont believe it. They’ve been going on the information where it wasn’t said that it was in beta at the time.

I completely agree that people (myself included) are quick to jump to conclusions. But see how fast I turned around after understanding what happened? You didn’t even have to tell me exactly what, you just pointed to the information and I went out to see it and learn what’s actually happening.

Imagine how much headache they could save by making an official announcement —which is what I and others have been looking for— instead of just saying it in a fast moving chat and hoping enough people saw it at the exact moment.

———
That is all I’ve ever been asking for. An official announcement where they say that they understand what’s happening and “XYZ” is the plan of action moving forward.

[soft iron]

nothing will happen, they have been burying themselves since SD 2.0

I'm surprised they're still alive

[restive robin]

Well that’s true if they don’t say anything from an official capacity. Nothing will happen if nothing is said

It’s up to them to sort themselves, I’m advocating that they do so.

[restive robin]

Anyway, here is the note from earlier.

I understand that the model is in beta.

Just post a realistic plan of action anywhere that won’t get lost and people will be much more understanding than with nothing (or relying on the grapevine where things will inevitably be misconstrued)

Have a good evening

[soft iron]

all I saw was that they said that 2b is ready, and 8b is in beta

if after the failure they started saying that this is beta, then this is just damage control

[restive robin]

I will say, that is pretty much what I believe too. I want to give them the benefit of a doubt, I really do, but it makes no sense if it only comes out now. Hence why communication is vital.

It’s more than a little strange that they only call it a beta now.

They need to have

A: a realistic explanation with actual examples other than “he said this”

B: a realistic plan of action to fix things.

In an official announcement

[mortal delta]

hello again, i want to do something with ai tools but im not sure what? im kinda looking for suggestions or exmaples, ive thought about making a comic or a video game but im pretty clueless.

[warm junco]

Hey, did the setup worked yesterday?

[soft iron]

use the engine for visual novels

and generate visuals using pony

even a child can handle it

[desert dagger]

the information they've been going on is all the assumptions users on reddit have posted, or assumptions they've made themselves. and i've watched them on here totally ignore any information they've been given that doesn't match their assumptions. The reddit users being just as bad, i'm sure that's what they are doing there, as well

[mortal delta]

oh sadley i coudent get it to work i even tryed sd next but still got nothing but errors. i swtich to comfyui + diretml and now i get images at the moment. still cant run sdxl due to hardware sadley.

[mortal delta]

like renpy visaul novel maker?

[soft iron]

you need to program in renpy, if you are not a programmer it will be very difficult for you, take something simpler

[restive robin]

Can you, point to any website or announcement that says that this was a beta or under development in an official capacity during release and that I was not a full release?

Again, not everyone is on discord to see things live, and even so, seeing it live only after the fact will be a lot harder to believe.

Of course people will be skeptical now, the burden of proof is a lot higher when you’re retroactively trying to say something existed when it wasn’t announced in any official capacity.

[warm junco]

Oh if you want to try fixing it again. Let me know in #🤝｜tech-support

[soft iron]

I have P106-100. It cost me $30 and when I didn’t have a video card I easily used it for SD.

[mortal delta]

i think ill stick with directml for now, i guess.

ill post an image or 2 in the image chat if your curiouse what i created.

[river tusk]

So 8B when released is going to be what the API uses too, right

[restive robin]

This is exactly the kind of question I’m asking for them to answer and communicate clearly

Plan. Of. Action. What is it?

[pale latch]

its what it uses right now. they might introduce other payment tiers. up to the ceo to decide

[river tusk]

Well yeah, they're going to be fine tuning it then releeasing it, the API is going to be using that fine tune too right?

[regal rain]

has anyone found a fix for one button prompt and infinate zoom yet?

[desert dagger]

to try to finish training it while the discord community apparently rushes ahead with burning down the discord - no, i'm not going to do your work for you and go find you links to annoucements that you are perfectly capable of finding yourself

[regal rain]

https://www.reddit.com/r/StableDiffusion/comments/133z9kl/i_mashed_infinite_zoom_and_one_button_prompt/

[river tusk]

You know this doesn't actually accomplish anything but instead cause more confusion and problems right? Clearly he can't find it, if its so easy just point to it, maybe try putting out fires instead of lighting them bruh

[desert dagger]

sorry, i don't agree with you. the people, like him, that i see demanding information instead of going and getting it don't actually want it. i've provided it ,and been ignored. i've also been told that the information is lies. so i feel no obligation to do what he can easily do if he really does want the information

[river tusk]

okay yeah no thats aa lot to read through im not even going to contest if you're right or wrong about answering it, if you answered my question tho lemme know, legally blind here and its a PITA to go through lots of text

[restive robin]

First, I wasn’t asking you to magically state what their plan is. Obviously you are a third party and I thank you for what you did point me to.

Second, I’m not sure where the animosity is coming from, I’ve been nothing but cordial with you. If you don’t wish to point to anything that’s fine, I was simply asking if there were any as when I looked at this server’s announcements and the website I could not find any.

Third, when you have pointed me to Alex’s comments I thanked you and even confirmed if I understood correctly and merely stated that I would encourage SAI to make that information more widely spread in an official capacity instead of relying on word of mouth as that could lead to poor optics or it being misconstrued.

I see however that you are frustrated with people assuming things. So I’ll pretend that you’re totally right and I didn’t acknowledge and support more people learning what you’ve shown me (because I wouldn’t have found it in other means than Reddit where I’ve just seen people posted those comments made my Alex) so you can vent I guess??

[desert dagger]

sorry for the way it came across, i have no animosity toward you. and yes, i'm extremely frustrated with quite a bit of what's going on, both on this discord and on reddit. I do appologise for it coming across wrong to you. not my intention

[restive robin]

It happens. Don’t stress about it.

[desert dagger]

part of what stability said was at the conference apple just had where the CEO annouced that sd 3 medium weights would be open sourced. it's on youtube, i don't have the link.

[restive robin]

Good to know, I will look it up when I have time. (Literally just got back from mowing the lawn in 95F weather so I gotta take a sit down and I happened to check if there were updates.)

[desert dagger]

and appologize again, i really didn't mean for it to come across as it did. you didn't deserve that.

[fervent thunder]

[restive robin]

We’re all a bit wired lately.

I’ll admit, I may be a little over eager for SAI to make an official acknowledgement for people to read, along with a roadmap where nobody has to go hunting down chat logs or search through hours of conference video for a few specific lines.

On that end, I am a little overzealous.

But anyway, have a good evening.

I’ll be off for the day since if something was going to happen, it would have by now.

[desert dagger]

you have a good evening too, and don't spend too much time out in the heat tomorrow

[dim forge]

anyone got lora training scripts working on multi gpus with flash attention? I have heard performance issues occur making it not worth doing

[pliant reef]

I'm trying to make a model on my art but I'm getting errors

When at the download step

[desert dagger]

post the errors?

[regal rain]

no reply about one button prompt and infinate zoom yet? 😦

[desert dagger]

what's a one button prompt?

[regal rain]

https://www.reddit.com/r/StableDiffusion/comments/133z9kl/i_mashed_infinite_zoom_and_one_button_prompt/

they changed the py code it dont work lol

looks cool dont it? 😄

[desert dagger]

i'm not sure what i'd use it for, actually

[regal rain]

i installed the extension in stable diffusion web ui extensions it makes really cool pictures i can select from a dropdown of what to make

i cant use infinate zoom anymore because they changed the code

https://imgur.com/mUflon8

its a shame that infinate zoom video with one button prompt looks cool

[astral goblet]

extension that generates prompts

[regal rain]

https://github.com/AIrjen/OneButtonPrompt yes it only makes pictures now though

[regal rain]

im also looking for the controlnet discord to make https://www.youtube.com/watch?v=wmaL-fhJ8Pw

[naive rose]

which sampler you use?

[regal rain]

i found the video on facebook i wish i made it lol

you havnt heard of draw things?

https://imgur.com/gallery/draw-things-complicated-fMAP1Gx it has allot of things to choose from a dropdown like one button prompt

it has an age slider 😄

[lucid bane]

How would you guys run SD3 with option for mobile interface (via lan)? (asking cuz a1111 still doesn't support sd3)

[regal rain]

i installed stable diffusion web ui in sanoma after 24 hours of following chatgbt I posted it in the apple fb group and I have sd3 😄

[soft iron]

0 finetunes of sd3 is death

💀

[regal rain]

https://imgur.com/gallery/pinokio-sd-forge-sd3-lFpfX9c

i wish that one button prompt guy had a discord I messaged him on github

[regal rain]

is there an infinate zoom discord?

[astral goblet]

it'll be fine

[weak scroll]

yes

7tg4ZTq7 here is the invite code

[regal rain]

thanks

[hazy gulch]

show can I use this channel

[desert dagger]

type in it.

[desert dagger]

@finite cloak

[finite cloak]

AMD not apple

[finite cloak]

do dis button

(im not a mod im not supposed to touch things unless mods are asleep and there's a big issue)

[desert dagger]

sorry. yes. amd - and thank you

[fervent thunder]

guys wait for 5090 for SD or go with 4090 now?

[potent star]

If you make money doing SD, or have too much, go for it. If not wait I would say.

[sudden ruin]

It doesnt seem to have much more VRAM and its probably gonna be pricy too sooo...

[potent star]

Just checked how many images I have made in a year just with a1111. It's 105k images, 370gb. Hmmm

[fervent thunder]

go for 4090 or wait for 5090 and yea i would probably make money out of SD for yotube thumbnails ,and my intro converting myself into anime charcter

i m getting impatient man some people are saying 5090 wont release until Q2 of next year

[potent star]

Who knows, what are you using now?

[fervent thunder]

dont ask man trash stuff

laptop with igpu 11the gen intel i5

[fervent thunder]

is 4090 good fr sd or is it terrible like its terrible for ray tracing kinda terrible

[sudden ruin]

Dont you delete anything?

[potent star]

It's good for both.

[fervent thunder]

no tons of youtbers are saying how trahs 4090 is for fps in ray tracing

15-20 fps lol in ray tracing at 4k

[potent star]

Working with clients so I keep everything until I am 100% sure we won't need anything. I still have to go through old generations sometimes.

[sudden ruin]

Thats bs. I have a 4090

[fervent thunder]

how many fps u get

[sudden ruin]

Why would anyone listen to youtubers

[sudden ruin]

That totally depends on the game lol

[potent star]

How is it shit when it's the strongest card for raytracing in games atm 😄

[sudden ruin]

Most games these days arent very well optimized on launch

[potent star]

I play at 1440p and never had an issue with max settings in any game.

[fervent thunder]

for SD?

[potent star]

Fastest consumer card for AI, 2x the speed of my 3090.

[fervent thunder]

coz 5090 advertised lora and lots of AI AI stuff which 4090 doesnt seem to have

lora speed also increased in 5090 i guess

[potent star]

Man, you have no idea what you are talking about huh? 😄

[fervent thunder]

yeah i know lora is important for sd

[sudden ruin]

I also play on 1440p but 5120 wide

[fervent thunder]

and 5090 advertised it

for videos how is 4090

in SD

[potent star]

Thats not how any of it works, educate yourself first.

[fervent thunder]

converting urself into anime charcter 30 sec video

or even 8 sec video

30 sec wont need for intro but 8 sec needed for intro and rest is for generating youtube thumbanils

[potent star]

I can't run a ultrawide screen, I am a monitor hoarder, run 5 atm.

[sudden ruin]

The best you can buy and put into a normal PC

[sudden ruin]

I was used to 2 screen as well, but ultra wide is so good

[fervent thunder]

people said this when 3090 realsed and look how much difference in 3090 amnd 4090

day and night difference

i will buy 4090 if 5090 wont have bigger difference in AI

[sudden ruin]

Youll get around 30 FPS in the stock Cyberpunk 2077 with Pathtracing, not raytracing.
If you get the pathtracing mod you can improve on that and upscaling tech will obviously also improve results

Thats for 4K

[fervent thunder]

by AI i mean SDXL STORY DIFFUSION ETC

how many fps have u got highest in ray tracing 4k

@sudden ruin

[sudden ruin]

What game my dude

[fervent thunder]

any game name will do which u played

[sudden ruin]

Games with lighter implementation you can easily run 100+ FPS on my almost 4k screen

[potent star]

It's newer, of course its going to be better. But anything you can do with a 3090 you can do with a 4090.

So if it's important for you that an image is made in 1 sec instead of 2 seconds then go for it. And when the 5090 comes out maybe 0.5 second.

I personaly can't tell a big difference between 3090 and 4090 in real world use, as its useless to create 500 images rather than 250 in the same time, you still have to look through those images to find one you like, so then it comes down what you do with it.

I only upgraded because I teach custom models and there is a difference if it takes 2 days or only 1 day.

[sudden ruin]

For example spiderman remastered has RT and runs at least with 100 FPS on 4k

[fervent thunder]

cynberpunk

[sudden ruin]

While the 3090 barely does 60

[bleak matrix]

Good morning, everyone! How are we all today?

[queen whale]

i saw that some checkpoints have as options to download things like: Pruned Model fp16, Pruned Model fp32, Full Model fp32

what is that?

[broken smelt]

essentially means a good chunk of usually unnecessary stuff is cut out to save space (sometimes lots). usually doesnt result in loss of image quality. use them if youre not planning to train the model

from what i recall, there is minimal image quality loss going from fp32 to fp16 but idr too much on it

[queen whale]

so for image generation prune is enough

[stuck summit]

hi, i need some help with a model and a LORA downloaded from civitai. It doesn't show them on the UI

[queen whale]

did you restart webui after putting them on the folder?

[broken smelt]

there should be a refresh button next to the model loader that you can click, saves ya waiting for everything else to also refresh

[lone hawk]

SD3 is sooo good, am i right guys?

[kindred quartz]

how to use this tool？

[stuck summit]

yes, refreshed and restarted. So would you like to show you what i've downloaded so you can say me where should i put them?

https://civitai.com/models/377060/retro-fantasy-style

and the relative anythingelseV4_v45

[wet grotto]

my jesus

[low moon]

civitai has that effect on some people

[strange ruin]

Hey! Can anyone please give me advice of any AI desktop offline tool that can clone voice#1 and change voice#2 at video to voice#1

I mean i have just voice of a person #1.
I want to change some random video's voice to voice of a person #1.

Thank in advance

[marsh steeple]

Why SD3 ripped off in every single generation the body horror of Cronenberg? I thought SAI wanted to remove copyrighted stuff from the datasets?

[desert dagger]

dont' even go there please. you're gonna have every member of this discord flooding the channels with that sort of stuff

[marsh steeple]

Yeah, having copyrighted images generated here would force the discord to be took down by irs

[desert dagger]

no, you're gonna have them flooding the channels with body horror of cronenberg

[subtle badge]

Bark and Tortoise are both good.
Iirc tortoise is probably closer to what you're looking for

As far as voice replacement goes I don't really have any suggestions

[strange ruin]

thank you for help, i'll check it

[opal hedge]

All my homies waiting for a useful sd3 finetune

[waxen summit]

I'm batman

[floral umbra]

will most A.I stuffs suffice with python 310? Or is it recommended to go 3.12+ these days?

[bright wren]

Is Instant id for img to img for a1111 released?

[desert dagger]

it does take time to create such

[fervent thunder]

[marsh steeple]

Is SAI still interested in making the 800m SD3 available? It can be trained more easily even with this "license" so we can delete the copyrighted cronenberg body horror from its generations way faster

[desert dagger]

they are going to, yes. it's a long way from done.

[wanton lotus]

question for forge users: are you gonna keep using it

[hoary jasper]

I have some questions can anybody join the call

I am beginner and want to get started

zero zero size zero

[unborn lion]

yo

[vast ingot]

Is dreamstudio being abandoned ? I can't find mentions of it on https://stability.ai/, and it doesn't seem to propose the latests 3.0 models, nor the SDXL turbo model, still having a "New" tag alongside "SDXL 1.0".
I know it's not perfect, but I still liked that platform a lot, for making pictures when not at home for example

[fervent thunder]

That's a good question

I like it also

[clear oyster]

sd3 out??

[vast ingot]

yep (also here #🆕｜sd3 )

[desert dagger]

also expect to have to learn it from the ground up, it doesn't work like other versions of stable do

[sage reef]

i mean it depends what you want to generate tho, some stuff you can do with just tags

[azure nebula]

anything worth using SD3?

[sage reef]

i personally would recommend you wait for finetunes

or until they fix it

it is however very good at some stuff, for example, i did some pixel art and it does it nicely

[steep arrow]

It can write things out about 60-80% of the time. Thats a pretty decent trick.

[sage reef]

but it seems to be struggling to generate isometric views

[pseudo coyote]

hello

anyone knows how to enable xformers on any repo

just like the command used in SD?

[rancid fjord]

I noticed after I started using stable diffusion our electric bill went up 😭

I don't think that's normal

[wheat isle]

Hello! I wrote an article about using the paid version of the image generation API with my friends. Please read it if you are interested.

https://x.com/xqdior/status/1801995625745252765

If you want to know the best performance of stable diffusion 3, this article will be useful. Please like and repost it.

[desert dagger]

running your gpus a lot?

[rancid fjord]

yeah

it draws about 300w at full power

either 300 or 240 I forgot

[desert dagger]

then that's normal. that's why it costs so much to rent gpus

[rancid fjord]

my 1200w power supply probably doesn't help

[forest apex]

you can power limit the GPU with nvidia-smi, depending on what model you have

[desert dagger]

yeah, that probably isn't helping

[rancid fjord]

I have an amd

[wicked vine]

hi, is there a technique or methodology to transform the left white blouse to the right blouse (white background and standard view) ?

ops, I cant attach the screenshot

[golden gale]

Hi, does anyone knows alternatives free image generator on discord like stable diffusion? Thanks

[desert dagger]

use #🏞｜general-with-images

[full lark]

Ye U.U

[wary belfry]

With the SD3 license agreement: Is a LORA a derived work? Since a LORA is entirely new data that it exactly NOT what the core model is, it can't be a derived work.... at least that's my understanding

[desert dagger]

go ask the crowd in #🆕｜sd3 who are currently arguing about the license

[wary belfry]

ah thanks!

[warm garnet]

I'm trying to run SD locally but I have an AMD, is there a way to achieve compatibility?

[desert dagger]

are you running it in comfy? auto1111? some other interface?

[warm garnet]

Try auto1111 and Zluda
I got it to create images with the CPU but it never recognized the GPU

[desert dagger]

i don't know if there's a solution then. try comfyUI

[warm garnet]

I'll try! tanks!

[muted bloom]

the reason for sd3 nuking nsfw is the same reason they did it for sd2, cheese pizza. they got rid of nsfw so people couldn't generate cheese pizza, it was either get rid of nsfw or get rid of kids so they chose nsfw

that's the reason they got threatened with a lawsuit btw, they got threatened to be sued for the model sd1 being able to generate cheese pizza

it really sucks but that's their reasoning for it

and i imagine it hasn't changed from sd2 to sd3 so that's why i'm mentioning it

[warm garnet]

Has anyone tried Fooocus with AMD?

[oak lantern]

Of course getting rid of nsfw means getting rid of the ability to handle humans without turning them into mutants as well or so it seems.

[steep arrow]

It means stability doesn't know how to properly make their model SFW without making it a mess.

Sd2 never got fixed because of this strategy.

[full lark]

Idk if any other AI companies could do that, they block the nsfw pictures using the API but not at model level, dalle3 for example can do nsfw, mjv6 was able too but of course you got blocked

[hasty hornet]

yea other models can do nsfw, they just censor it

[full lark]

On api level*

[hybrid jacinth]

• images with *Ultra, using prompt *photorealistic medium portrait photograph of a psychedelic beautiful goddess with flowing robes hologram in space galaxy, neon teal and neon peach color palette, wide angle cinematic masterpiece, symmetrical, centred, cinematic masterpiece, high detail, sharp focus (and negative prompt *drawn, illustrated, sketch, cgi, asymmetrical.), seed **1298201248, aspect *4:5! 🛫

[hasty hornet]

sometimes mj generate occasional booba too , fails to recongnise it

[oak lantern]

Which is impossible with a locally run model.

[hasty hornet]

idk why it's a problem if it's something local tbh

[wet grotto]

annotator in SD is what?

[full lark]

Because companies lawyers and journalists only throw 💩 at open source models, they want to know everything you generate, they dont want you to run something local, only on their servers, in their perfect world they would have cameras inside everybody´s house just to make sure "they don´t do something unsafe/illegal"

[steep arrow]

I would suspect if you actually removed nsfw from the training data and then did some very intentional caption organization, you could guide the model to be ignorant of NSFW content without creating a model with incompetent latent connections.

Whatever Stability are doing, its hamfisted and doesn't work

[full lark]

Idk, google tried doing the same thing with "racism" and end up generating black popes and indian Cristobal Colon and stuff like that

[steep arrow]

Which is what stability wants with NSFW

Ask for nudity, get a person in clothing. Google did create an offensive model, but it wasn't botched.

[full lark]

Google tried creating a non offensive and it did inverse racism, stability tried creating non nsfw and did inverse anatomy?¿ .
In my opinion journalists and everyone that try to sue open source models should go f*ck themselves.
Maybe they did some mistakes in the training like you said, but that doesn´t change the fact that when companies try to censor things in AI they break it

[nova zodiac]

Probably too far. Deepfakes and CSAM can do a lot of harm…

[steep arrow]

You are right, but there are different kinds of breaking.

A black george washington is coherent, a boneless spaghetti person is not.

[nova zodiac]

Ideogram has managed censorship well so it is possible

[steep arrow]

Exactly, and stability has already "learned this lesson" once, and then gave us SDXL, and now here we are again in SD2 land.

[full lark]

A black george washington is coherent ... if you ask for a black george washington

[full lark]

But they didn´t ban deepfacelab did they?

[steep arrow]

"coherency" in a technical sense is about the sensible organization of the elements in the contextual space, not the intention.

[nova zodiac]

The question with ideogram though is “Is that through the base model or is that through prompt sanitation and post-creation image checking”?

[mortal delta]

What are the best simple comfyui workflows and addons for making comics and getting the same chacraters? or are there andy tips and tricks or should i not even bother?

[weak scroll]

https://www.youtube.com/watch?v=HmxbsQDRTSs

[dense shoal]

Here is another video about this weird Licence worts then Adobe. https://youtu.be/r8oA42saTNI?si=g0aIpBHftWleWpaW

[desert dagger]

@finite cloak the license trolls moved over to this channe 😦

[dense shoal]

You have to be open to critics. Criticism, like rain, should be gentle enough to nourish a man's growth without destroying his roots-Frank A. Clark

[broken smelt]

criticism? sure. disingenuous trolls who wont change their mind despite any evidence contrary to what theyre posting?? lmaooo 👋

[dense shoal]

I did not use any bad or rude words. I will take that. and finish this coversation.

[pearl ocean]

https://www.youtube.com/watch?v=HmxbsQDRTSs&t=2s&ab_channel=OlivioSarikas

😮

[finite cloak]

#🆕｜sd3 message

[pearl ocean]

I just got modded

[full lark]

You already posted the video bro

[pearl ocean]

#💬｜general-chat message

[hasty hornet]

despite any evidence contrary - I don't think there was any until now, so...

Alex's version makes more sense

[tropic sequoia]

https://www.youtube.com/watch?v=HmxbsQDRTSs

[hasty hornet]

let's not spam same thing

[thorny relic]

Hey

[finite cloak]

#🆕｜sd3 message

[tacit oyster]

https://youtu.be/HmxbsQDRTSs fk em this is nonsense

[finite cloak]

#🆕｜sd3 message

[desert dagger]

you realize those are probably bots? or alternate accounts for one person?

[finite cloak]

i'm assuming it's just dumb people that don't read

like the last person to post it has a discord account from 2017 with linked twitch, youtube, etc. accounts

[desert dagger]

too consistently the same. I tend to be really skeptical. no dialog, just a blurb and the same link as everyone else. unless there's a campaign going somewhere and people are being told to do this

[wary ember]

I think the best explanation right now has been made by Invoke AI https://www.youtube.com/watch?v=r8oA42saTNI&t=603s , he analyses every section of the terms and also provides insight into the "grey areas" while also providing his take on why they wouldn´t use/deploy it... maybe not as viral but I think most complete atm... I'm honestly very concerned about why SAI is making these decisions and deployment of terms and careless release of censored models without previous review of performance and metrics. It's almost like they're working against themselves, and I don't think their engineers are stupid at all, the 8B model is very close to MJ level of quality, so I don't know if there's a weird agenda within the Executive department or if plain delirium on the exec side due to sheer financial desperation or greed is undermining the brilliancy of the developers.

I really hope this gets reviewed and modified, not to mention finding some new community figures that have a minimum of communication skills, it'd be sad to see the most widely used Open Source model developer company go in a fire like that... shitty or not lately, all the open-source community ecosystem that has derived from SD is thanks to them. I'd in fact be happy if they get a stable revenue from deployments of their model in a commercial license (like what krea.ai is doing with SD1.5 and steerablemotion, but you know... actually paying something for enriching themselves based on open-source community work), imagine the amazing stuff that can be developed with the new text clip architecture, making a more coherent time-aware text interpretation of instructions to represent motion accurately and with a lot more control... IF SAI is able to find the middle ground to profit from widely used deployments that make revenue for third parties, while not absolutely obliterating the Open-source community developers, that make their models actually usable on production level work...one can always dream I guess.

[wary ember]

And yes, I read what you posted, I appreciate you taking the time to share your thoughts on this, I also think someone within the legal team should in fact handle clarifying this info so that nobody complains about who's saying what and who's really right, but hopefully what Alex is saying is true, and it'd be a great start... that however doesn't solve in the slightest the 6k limit for derivative work and its implication on training and finetuning, take into account that atm there's a kind of limbo where someone who doesn't exceed the creators 3 point criteria of Million users etc, but also is a community developer of tools, can't use the creators license for training, which is nuts, there are lots of developers who can't work with either license, not to mention the issues with the response time for enterprise license requests.

This to say, that I'd LOVE to be corrected, and be wrong, so that LOTS of developers, who would also LOVE to be corrected, have the accurate facts, and can decide whether to invest in finetuning and developing once again a great tool (i.e what the community has done with SD1.5 and SDXL, and Comfy, and AnimateDiff, SUPIR, etc etc), that hopefully this time can be monetized for SAI, and their work can go on. We love the tools you have brought to the community, and the community is ready to put hours and resources to improve on it, if you just embrace and empower the amazing wave of community development you have spawned!

[desert dagger]

i keep hoping for some community members that want to do something other than complain, but apparently that's a lost hope

[finite cloak]

the 6k limit is for commercial usages of images, this doesn't impact most finetuners. The pony dev is a unique case because they have a whole complex generation service dealio, which does get affected by the 6k limit. Almost everyone else that's a finetuner or developer or researcher or etc. and monetizing it, is fine with the regular Creator license. Anyone in that group not actively monetizing is of course completely free under noncommercial license.

[wary ember]

Look, I agree, the shitstorm was brutal, and I won't lie, I'm upset because I have high hopes, I see the huge potential it has, what I'm saying is we should try to work together, or well, the development community and SAI as a whole, to change the narrative and get from this toxic thing, to something that moves forward the community like it has with previous models... I'm not hating, but I want this to just move forward

[desert dagger]

it's a beta release. you do understand that, correct? and it is moving forward. but continuing to just go on and on about it isn't helpful. what have you found that needs fixed?

[wary ember]

Yeah, I understand, and that's cool and all, but what happens when it does indeed get monetized, Pony is not the only example, what about krea, or Invoke, or Civitai, or any service that deploys your model (Or ANY model that is finetuned to SD3 OR SD3 Outputs) but DOESN'T exceed the 3 point million threshold, 6k is just impossible for anyone other than a single individual, and even then it's quite restrictive for a single individual looking to use it for commercial purposes for an ad campaign or work in general... @desert dagger oh, sorry, tagged you here, this is something I've found, to answer your question, again, great to have a conversation

[broken smelt]

this all ignores the point that sai has literally not mentioned the legality of anything -- it was just random idiots assuming what little they could glean from reading what's probably their first legal document and spreading that misinformation around lmao

[desert dagger]

is it saturday where you live? because it is where the rest of the world lives.

[spare frigate]

Section 4 a. "The Stability Technology, any pricing information, and the terms and conditions of this Agreement and any Orders will be deemed Stability Confidential Information"
Section 9 c. "Also upon termination or expiration, each party will return or destroy (or in the case of electronic information, render practically inaccessible) the Confidential Information of the other, including Stability’s Software Products and any Derivative Works. Additionally, each party will cease using the other party’s Marks"

[finite cloak]

If you're doing an ad campaign, you are not commercially using 6k images in a single month.
You're generating 6k and then commercially using maybe 5 images. You'll never hit 6k commercial usages of an image if you're not selling an image generation service

[broken smelt]

sunday as of 11 mins ago

[tacit oyster]

i love what did they delivered to community, but this like they forgot what community delivered for them. 6k limit is bullshit, like take a look for example on img2img 60fps video, it is 100sec now, mayber 4 YT shorts in total, wow...

[spare frigate]

Was that in the creator license agreement? Because I cannot find the 6000 images limit there

[broken smelt]

someone promote this mans to be the legal hr guy

[desert dagger]

i seriously doubt that any one at SAI that could make official annoucements is at work

[spare frigate]

Oh the limit is only in the FAQ. No need to focus on that. Stability needs to specifically add that to licence because a note in a FAQ is not binding.

[desert dagger]

it's specifically on the proper page. and has been screen shot, highlighted, and posted. multiple times so far

[tacit oyster]

this ain't cool at any way, even my avatar profile is gen image by myself

[spare frigate]

It is not on the license page https://stability.ai/creator-license-agreement

[desert dagger]

then don't act like a bot

[tacit oyster]

lol

[wary ember]

I understand, yes, this is a good thing, but there should be a clarification on WHAT output constitutes it to be "commercially used" and what isn't. Let me give you an example, a real life example. Let's say I work on an ad campaign for a product or service company as an individual, and I want to generate images for commercial use, but IT IS THE CASE that for the client, the coherence and consistency of THEIR product needs to be 100% accurate. Then you may have to train a LORA or Embedding, so that almost every time you generate an image, you get what you're looking for (so for every image you use in the end, you first used a handful of images i.e regularization images, iterative training etc. used to derive into a LORA or Embedding). You may say to just use photoshop and then do some compositing, and some editing tricks, all cool and dandy. But then what if what someone wants to do is Video instead, or, something that is more complex than just an image, and you need coherence across the WHOLE sequence, i.e training, or using another community developers tools (let's say a ControlNet, or a temporal model, you get the idea), then what happens there? is that training/finetuning that is ultimately used for commercial use, considered commercial use or not? either/both for the developer/s and the user who makes use of said tool/s.

I hope I'm getting what I'm trying to convey in a clear way. This is for example just MY take, based on MY experience using Open Models and Open Tools derived from them. I'm sure MANY community developers have similar and different concerns as well...

[tacit oyster]

i pay PS i use it for free or commercial, because it is my bossiness how i use it. I download free Davinci same way. Why it can't be same way here?

[wary ember]

I mean yeah the community ecosystem that has derived is amazing and almost a dream, the idea is to continue fostering it! We can all benefit from this, including SAI achieving profitability with a good, accurate, and clear framework

[desert dagger]

different companies

[wary ember]

I mean they still have to be profitable, I understand a commercial license is necessary to guarantee a retribution for using the model commercially, I don't mind that, but you know, that while also being coherent and not punitive for the Developer Community ecosystem, or the slightly more advanced commercial individual

The reason why it's called a base model is because... well, it's a base, a foundation to build on, you can't make the foundation and then tell the construction workers the only 2 options is to only place 6 bricks per month (the base will crack due to weather and rain if not covered haha!), or be a big developer, which AGAIN I'm not saying are the facts, but that's where a proper official statement is wanted

[tacit oyster]

why is tastes so good and this so bad? because it is what u guess? This is stupid one, there is free usage and paid licensee in for example game engines, where u can do what you for free and realize game for money, but this one is very dif. For example for one good image you need something like maybe 20 tries and then like many images as you want from it, right? So, one image which you can sell costs you like 0.067$. For one 15 sec short you need like 900 images just to get it done when you are already chose style, parameters, etc and it's already 3$ without time on render you need on this. And why do u get so different spendings on products which will give+- identically profit. the hack is this? And what if i want perfect one so i'll inpaint all night long, again, different story. The more i skilled or lucky i pay less, eh?

[desert dagger]

every company does things different. as do all people. your house, your rules. my house, my rules. your company, your rules. my company, my rules. that's just sort of how it is.

[finite cloak]

(also saying adobe photoshop licensing/pricing "tastes so good" is ... extremely weird phrasing, but also just an unusual take considered how unpopular adobe's financial model usually is considered)

Stability definitely can't ever do a photoshop-like license setup: Photoshop costs $10/month to use. We'd have a riot instantly if somebody said you had to pay $10/month just to use SD at all

[sage reef]

maybe models could be like triple A games, you buy them for like $60 :3

[desert dagger]

and if we want to talk about filters... you think the filters on SD3 are tough? you should try using firefly in photoshop on your own machine. you even think about doing stuff that isn't G-rated, you get told it's against their community standards

[feral pike]

can someone please explain to me what Hyper models are ?

like SD 1.5 -Hyper

[sage reef]

just less step models

hey ❤️

[feral pike]

isnt that same as lightening and turbo models?

[sage reef]

kinda

[wary ember]

And hey, @finite cloak, it is needless to say that I appreciate you taking the time to read all my rant and probably lots of other people's rants, just understand that there is a community that is already present and was super hyped, and very good teams that build on top of your models... well anyhow, maybe relay that info to the team, or see if it's possible to have a conversation with the community developers, actual developers, not me just a random who knows very little and just enough for his job, and is probably not as smart as either those developers, or the engineers behind SD

[feral pike]

what is the diff

[sage reef]

i mean the diff is more technical, for you, lets just say its about the same haha

[feral pike]

oh

i would like to know the technical difference too 🙂

[sage reef]

well then you have to do some reading :3

[feral pike]

where

[sage reef]

all those models usually are on huggingface with a model card that has information for you to read, especially a research paper attached

[sage reef]

yea im busy with all sorts of projects these days

[feral pike]

so turbo , lightening and hyper are all stepped down faster models?

are there also , say , bigger slower models?

like opposite of hyper

🧠

good idea

[sage reef]

you changed your name right?

haha

[feral pike]

or are they just called XL models?

[sage reef]

no i remember the profile pic

[feral pike]

so for this model , they only released Hyper versions? https://civitai.com/models/4201/realistic-vision-v60-b1

[sage reef]

i mean you can technically merge models yourself and make your own hyper lighting or turbo versions

or use the loras to "transform" them into faster versions

like lcm, pcm, etc

[tacit oyster]

i dont say adobe's financial model is anywhere fair, it is totally not. But bieng even worse than this is something like "masterpiece" token

[desert dagger]

adobe's model is far worse than this one. now stop

[sage reef]

Narrator: he did not stop

[feral pike]

Ooo interesting

[tacit oyster]

worse in which scenario? This shit and adobe's garbage runs on my machine, right? It's costs roughly same 20$. But PS+LR gives you month of use, while with SD3 u can reach limit even in 2-5 projects, omegalul

[desert dagger]

worse in all scenarios. i want photoshop and inDesign. i have to pay for the ENTIRE suite in order to get just those two at a decent price

and forget using photoshop's gen fill to do anything except squeaky clean, g-rated images

i can't even prompt for a blurry background of someone sleeping in bed. that's "against our community standards" - on my OWN machine

[crude notch]

"u can reach limit even in 2-5 projects,"
?!

dont tell me its about the 6k images thing

[desert dagger]

she can't - she's ignoring everythign alex has said

[tacit oyster]

but this is runs not on your machine. Anyway being even slightly better than adobe is shitty achievment

[tacit oyster]

img2img video for example

[desert dagger]

we're discussing adobe here. photoshop runs on my machine, firefly runs on my machine, i still can't prompt for anything that isn't 100% squeaky g-rated clean. and i have to pay through the nose per month for that privlidge

[tacit oyster]

firefly doesn't runs on your machine

[desert dagger]

i just created an almost 2 hour movie of 3 second video img to video clips and didn't hit 6000 gens

[desert dagger]

yes. it. does.

[crude notch]

the 6k gens thing is for monetizing, in which case you still needed a license, just a cheaper one if under 6k gens sold

[desert dagger]

you're wasiting your time - it's been explained over and over and over. she, and the rest of them, are ignoring what's been said. on purpose.

[crude notch]

possibly

[broken smelt]

what if firefly actually doesnt run but instead flies, ever thought of that? checkmate in the argument tbh

[crude notch]

holy hell

[hasty hornet]

2 hours = 7200 seconds, is it <1 frame per second?

[crude notch]

using svd-xt to extend i assume

[desert dagger]

24 fps

2 months of work generating and editing

[desert dagger]

all on leonardo, actually

[crude notch]

neat

[desert dagger]

i made the mistake of generating music on udio, and then had to make a video for it

[hasty hornet]

I tried it once and got dissapointed after suno lol

[desert dagger]

it's on my youtube channel under the title of Strange Realities if you want to watch it

[tacit oyster]

we have tested few times you disables internet and it says you no connection to abobe servers, also no gpu load, etc.

[desert dagger]

don't really care what you have tested. it runs on your machine. part of it is in the cloud, but it's on your machine. it is NOT on their public website where filters matter. what is running in the cloud is not seen by any member of the public, and can't be.

and the entire point is that if you want to use photoshop, do you honestly think adobe is going to care if you don't like their terms?

[tacit oyster]

what kind of bullshit are talking? how it can be same speed on rtx4070 and intell uhd 600 graphics?

[tacit oyster]

i'll go with affinity or other stuff or even torrent, what's the deal?

[desert dagger]

i think you must have dropped into an alternate dimension as you no longer appear to be discussing anything even remotely close to the previous conversation

[tacit oyster]

aBobe is just big company, when SD is SAI's and community project, it is different story

[regal rain]

did anyone figure out one button prompt and infinate zoom as an animation yet? the guys coding dont work with the new shorter code lol 😄

[tacit oyster]

ok, enough, idk why r u so stressfull about big compnies money, just really tied of u

[desert dagger]

no, it's not. and it doesn't much matter. you didn't read what was explained, you wouldn't accept it anyway.

[tacit oyster]

i saw this pretty clear, that u can generate billion images and sell 1 and goes with img2img video. U sad u got another type of work, which isn't img2img and u r ok with, i'm glad with u, but this is not my case.

and this is just one simple example, ok? What if i make 10 images and stuck it in 1? is 10 images i sell, or 1 image, or even 0, because i made finale image by myself?

have you heard of card pack which you can buy in any market, full of AI image? is this also 1 image of billion? i can go with this all day

[tacit oyster]

i mean i'll make 200 image for printing and we goes with 50. Note, that all of 200 was for commercial use and i can say i got paid for making 200 image or 50, even if only 50 will be printed later. So i can fool u, or fool myself or what? this is inpractical shit

[desert dagger]

go re-read everything alex said and that you ahve ignored

[broken smelt]

man i wish i could read

[tiny wagon]

New to all this AI stuff—I've just installed ComfyUI and downloaded SD3 medium. I'm a bit unsure where exactly to properly place all the files to use them in Comfy

[hasty hornet]

licenses are confusing

6k thing is fine for most cases, deleting confidential info is fine too (if it's limited to it)
But I feel like being forced to pay to (use any fine-tuned model based onSD3 ___ and ___ train fine tuned models) isn't gonna do good for community, which was the main driving force of SD models

[desert dagger]

https://www.youtube.com/watch?v=Di1KqPXxx2Y&t=30s watch this install tutorial

[tiny wagon]

thanks!

[hasty hornet]

I see why that is, gonna have money to compete, but that's just sad

[tacit oyster]

ah ofk i cant upload image. so when he says most finetuners that means NOT ALL. When he says unique case this means is FOR NOW. I kinda hate this slather like eula

[desert dagger]

use #🏞｜general-with-images

[hasty hornet]

if bigger model will be significantly better than current one I feel like - making it $20 and lame one free and open would be better approach, but eh, what do I know

[sharp fulcrum]

Is it possible/effective to use SD3 as a refiner for SDXL?

[tacit oyster]

I disagree with this one. What if prompting and LORAs a part of commercial secret? Still image for home use, i guess? Yes u r not selling, but whole workflow is part of commercial use, or SAI is dumb at money charging, idk

[desert dagger]

probably.

[hasty hornet]

I'm wondering how people gonna keep track of 6k images per month thingy

[tacit oyster]

it is part of EULA being suspicious and unclear, okay? If big company can charge 1 cent, then 100$ ins't deal at all.

[desert dagger]

i suppose it's too much to ask that people be honest?

[sharp fulcrum]

Is 6000 a month for the free license or the $20 one?

[hasty hornet]

<6k commercially used images for 20$ (but I'm not sure)

[desert dagger]

free license is unlimited but non-commercial

[tiny wagon]

It seems I'm missing the manager and share buttons on my ComfyUI interface. Did I do something wrong during install?, or are these optional plugins?

[sharp fulcrum]

Using this stuff for commercial purposes seems kind of dubious to me anyways considering a lot of the training data was gathered without explicit consent
Idk if there have been any landmark cases to decide if training AI off of copyrighted works is considered fair use or not

I enjoy using AI for fun but I hate seeing it on places like pixiv and patreon

[desert dagger]

you ahve to install the manager.

[hasty hornet]

I think it's fine as long as people don't say "I made this thing" or "it's a real photo" or do missleading \ fake news \ propaganda nonsense

[tiny wagon]

I see, thx

[sharp fulcrum]

I'm just hoping that eventually patreon will ban using primarily AI content as the focus of the page (using AI for things like backgrounds would still be acceptable use) like pixiv fanbox did
What I see happen a lot is AI posters get popular really quickly on pixiv thanks to the generous algorithm, and they redirect people to their patreon

[desert dagger]

they won't

what you'll see is that fewer and fewer sites will care

[hasty hornet]

I see nothing wrong with AI art, some of it is really good, some artists modify them in editor tools too, AI is a part of the world \ jobs now, whether you like it or not...its just need to be clear where it's made with \ by AI and where isn't

[sharp fulcrum]

I never thought I'd be on this side of the argument

[hasty hornet]

at some point it probably won't matter at all once it will get better...

[sharp fulcrum]

But like it's incredibly annoying browsing boorus and such and seeing a constant flood of AI stuff even when I have it blacklisted

There's literally aibooru but people feel the need to post it elsewhere
At least danbooru just simply doesn't let the posts get through approval

[desert dagger]

what you will also see is that human made will become more expensive, as it will be more rare

and instead of paying 10 bucks for somethign, it'll be 100 bucks or more

[sharp fulcrum]

Even if you can't tell something is AI just from looking at an individual image, it'll be easy to tell if there's a ton of variations that look like completely different pieces posted close together

[desert dagger]

not really.

[hasty hornet]

that depends...bad \ low skill artists might be completely replaced, but good ones with specific style yea
But I don't think that's just about "pure" art, artists who can incorporate AI in their workflow will be more efficient.

[desert dagger]

no. that's what always happens. books used to be bound by hand, machines came along. now handbound books are thousands of dollars. clothes that are sewn by hand are far more expensive than in the store, etc

[hasty hornet]

well, if we're talking about real world art - yea, that's gonna be more expensive, digital is different tho...