💬｜general-chat

[desert dagger]

AI is just one more way of mass producing something. and mass produced is cheap and easy, while handmade becomes very expensive and collectable

[hasty hornet]

It doesn't have to be about mass producing, you can use it even just for inspiration \ to give you few ideas \ concepts you're gonna draw...

[desert dagger]

i know. that's not really what i'm saying.

[hasty hornet]

You can modify images with it too, it's not just about generating lots of images anyway

[desert dagger]

i know. and that's also not what i'm saying. this started with Xi being distressed about ai images on that site.

[hasty hornet]

I'm sleepy, I can't keep track so far back what we are talking about lol

[desert dagger]

everything being flooded with AI images, thus driving up the price for human, handmade - even if they are digital

[potent spire]

This is why AI images shouldnt be allowed on art platforms or at best be able to be hidden with AI filter

Bad/low skill artists dont stand a chance in the market anyway

[tacit oyster]

this is kinda stupid one. You can redraw it in order to just upscale, and vo la handmade one. It's just a new crazy ass tool like camera was back then days

[desert dagger]

this is why you'll see nothing BUT ai images on most platforms within 2 years, and people that do other stuff will have their own sites and charge you hundreds of dollars to buy it from them

[potent spire]

Wrong

[desert dagger]

ya can't restrict it. the art community is coming on board, and there will be no tools that you can use for digital that don't do AI

[potent spire]

Not necessarily, see Artstation

Plus not everyone is into generating AI images

[desert dagger]

nope. artstation is littered with AI images. and was before one person decided to scam the others into an ill-advised protest

and it won't matter. 2 years.

[tacit oyster]

how can you prove than is wasn't?)

[potent spire]

not really, the marketplace has more of the stuff but thats it

we just ghost them for the most part

it was a dead end for them there from the beginning with exceptions

DeviantArt is another story

[potent spire]

what wasnt?

[desert dagger]

nope. not sure how much time you've spent on artstaion but i've spent way more time than i want to. it's packed with Ai art.

[potent spire]

im on a daily basis there 😄

[desert dagger]

then you're not looking very close. there's also no way to tell if something is AI or not unless the person decides to say it is

[potent spire]

the marketplace is annoying because of the AI image reference packs

but i have AI filter on now

[desert dagger]

you do know who is behind artstation, right?

[potent spire]

like most people

[potent spire]

ofc

[desert dagger]

that only works if the artist turned it on

[desert dagger]

who

[potent spire]

its easily to detect

[potent spire]

Epic Games who else would it be

[desert dagger]

really? okay. tell me how to detect it

[potent spire]

Tim Sweeney

[potent spire]

anomalies in the image

amongst all

[potent spire]

which is what they do

[desert dagger]

i'm gonna post a couple images in #🏞｜general-with-images and you come tell me if they are AI generated or not

[desert dagger]

only a few do

[tacit oyster]

I think both do

but even if i right like not 100% sure maybe 95%

[potent spire]

both do or you intentionally made the shadows bad on the second one

but both are awful

and tbh both arent worth painting over manually

[tacit oyster]

honestly i think so either

[potent spire]

oh forgot to answer to this, AI is one thing, generative AI another

Photoshop is full of AI assisted tools

yeah some people complain now about Adobe and their TOS etc. but i aint one of them

and i use generative AI too but not on canvas nor as 3D generated mesh

since i do 3D work especially

[royal field]

Best way to fix hands post gen? Using inpainting but struggling to get a good hand

[proper scroll]

Hello kind people of stable difussion am using sd3 , and am majorly failing to get exactly what I want, Is there a way that I can give sd3 a sketch of exactly what I want and tell it to just make it better?

[vale steeple]

oh you need SDXL/SD1.5 for that with ControlNet or tools for now. SD3 is just too new to let people update their compatibility yet

[proper scroll]

control net? what is that?

if its something I need to train it won't work , because what I drew is too specific

[vale steeple]

something that control the model by adding several condition, like say, a pose, you give the pose to the ControlNet then the model will able to follow the pose to create something

[proper scroll]

I drew a specific illustration of an ancient egyptian scarab jewel , framed in a triangle , I don't anyone is trained either the item or my composition

but both are the main characteristics of my illustration

[vale steeple]

although if I remember correctly there is no mode in ControlNet that you can replicate style in one image

You does not need to train for ControlNet though

[proper scroll]

but what does it require exactly

I figure it needs some kind of input no?

either it knows which in my case I doubt

or.... I need to input it some kind of thing that tells what I need

[vale steeple]

just the pretrained model and inputs for preprocessor and extension

I think MLSD, Canny or Sketch may came close to your needs but the thing is idk what you exactly needed, whether you just want to colourize your illustration or based of that illustration to create something new.

[proper scroll]

I want to improve its quality and make it water colours, I made a digital illustration, I want it to become a better illustration with the same concept , and in water colours

my artistic skills have a limit I want AI to impriove my work

but not to reinvent it

[vale steeple]

oh yeah you definitely need img2img which came by default in A1111 and ComfyUI.

[proper scroll]

I tried img to img but it would grab my thing and turn it into another thing

[vale steeple]

yeah it would generate something different

so you need ControlNet

I am still thinking which model ( because I never used ControlNet for stuff lol )

[proper scroll]

I just don't really understand what controlNet is or how to work with it, do I need to download it?

[vale steeple]

you need to, unless you using Forge.

[proper scroll]

am using comfy ui

dont think I even heard of forge

where does controlnet live?

as in folders if I may ask

[vale steeple]

coincidentally it is made by the one who release ControlNet models into SD.

For ComfyUI, you need to download this extension for preprocessing
https://github.com/Fannovel16/comfyui_controlnet_aux

I figured MLSD should be the answer

[tranquil ruin]

有人能帮我吗，我能进来SD频道，但点不进去MD频道

[vale steeple]

MD?

[tranquil ruin]

midjourney

[proper scroll]

mlsd?

[vale steeple]

那是MJ好吧。这里是专门 Stable Diffusion 的服务器， MJ的话你要进去他们的服务器

[tranquil ruin]

我知道啊，可是我进不去那个服务器，不知道什么情况，你们遇到过这个情况吗，点了那个服务器转两圈就不动了，进不去，

[vale steeple]

换个链接/Wifi 试试看？

[vale steeple]

nvm it should be Canny, it is one of the mode inside ControlNet

[proper scroll]

it says it may conflict with my installation the controlnet_aux inside the packager

[vale steeple]

you already have that

[proper scroll]

Canny?

[vale steeple]

lol

[median dock]

lol

[golden gale]

Hi

[abstract bough]

oof! My guess is most of us are sticking to SDXL and below. Can't afford using SD3😭

[vale steeple]

SD3 can handle 4GB bro

are everyone only have 2GB VRAM?

[stuck summit]

hi guys i'm serching for suggestion to make retro dark fantasy art with stable diffusion, any tip?

[glacial dagger]

hello, I have a question. When I use Stable Diffusion on my laptop with Ryzen 3 7320U everything runs smoothly until when I finish regenerating the image an error appears in the cmd like this

RuntimeError: [enforce fail at alloc_cpu.cpp:80] data. DefaultCPUAllocator: not enough memory: you tried to allocate 2415919104 bytes.

can anyone help me how to solve this problem?

[fervent thunder]

sounds good!

[fervent thunder]

are you generating at 512x512?

If not, try that

[glacial dagger]

yeah im generating at 512×512

[fervent thunder]

Your laptop may simply not have enough memory. I would get a second opinion however. I am not exactly the technical type.

[stuck summit]

That kind like book cover dark fantasy from the 70' or DnD manuals and stuff like that

[sharp fulcrum]

if i slotted my old 1070 back into my PC along with my 3070, could i use both in tandem to generate faster?
or does multi-gpu support require the gpus to be the same

[fervent thunder]

I have done some of that. You just need to get setup with Stable Diffusion

[glacial dagger]

maybe i should lower again the resolution, my laptop is really slow when generating image

[elder plaza]

Highly recommend watching this. About Stability's new T&C's. Not cool SAI! https://www.youtube.com/watch?v=HmxbsQDRTSs

[fervent thunder]

I don't think that will help. It looks like you have a memory problem (not enough), but there may be a way to get around it. You may want to ask your question again when the channel is more active

[glacial dagger]

aight thanks dude

[fervent thunder]

clickbaity title and thumbnail

[elder plaza]

Not at all. Go directly to the T&C's if you don't want to watch the video. It's just a great job of summarizing.

[low moon]

Well sadly I don't see much future for sd3 because of the license. Everything for personal use should be 100% left out of any monetization, simple and clear, u make money with it - ok pay up, you don't - do whatever. but this is so murky and greedy no one's gonna touch it. Will there be underground checkpoints popping up on pirate bay tho? XD

[elder plaza]

Yeah, totally agree. The model is out in the wild. No way to put this genie back in the bottle.

[regal rain]

https://imgur.com/gallery/sd-web-ui-3d-lora-ZkV6KkP 3d 😮

[jaunty sinew]

Hey guys, I just wanted to know if Stability AI retain any rights to the images created using Stable Diffusion 3 ultra, or do the creators have full ownership? And in this case what if we buy an entreprise license would that change anything.

[low moon]

Even a pawyer can;t asnwer that for you now, that's how convoluted the license is

[fervent thunder]

I will never understand how anyone, any company, ever, can have any rights to stolen images and art

it's not theirs to begin with, the algorithms may be, but nothing else is

so f*** their licenses

[stuck summit]

Can you give me some advices please?

[wet grotto]

take checkpoints and loras that have the models and stylesthat you need

[warm yew]

Guys I make and sell paintbrushes, you all owe me for the paintings you create by the way 😄

[stuck summit]

Can you suggest me some please, i've literally found only 2 that make that style

[fervent thunder]

Have you tried searching for particular shows or publications that were retro dark fantasy?

Like Heavy Metal for example

[cerulean rivet]

Am I the only one not able to generate celebrity faces with SD 3 and comfy ui? I tried Elon Musk and Snoop Dogg and got some random faces

[stuck summit]

Yes i know heavy metal, but i'm seqrching for a model or something like that with that style

[fervent thunder]

Yea, like you said. I have only seen a couple on Civ. But there may be more if you knew what to search for.

[wet grotto]

you can check then on Civiti AI website

@stuck summit https://civitai.com/

[verbal osprey]

A Heavy Metal movie LORA style would be incredible!

If you look at my icon, you might understand why I'd like that.

[astral goblet]

If i were the CEO of stablity, i'd honestly consider libel against Olivio for serious damage. That would probably completely kill community vibe, but the community doesn't ever want to pay stability for anything anyways. Should probably stop engaging any of those people as a business.

I'd probably be a really bad CEO

[pseudo coyote]

hello

who can help me with comfy ui

I need help running a repo within comfy

[desert dagger]

You'd probably be discussing stuff with the Corp lawyers

[astral goblet]

Olivio had all the information to the license, had the truth right there, and willfully misreported it to farm views. If that's not libel i odn't know what is. (i really don't. ianal)

[astral goblet]

a custom node you mean? what do you need an entire repository for (love expanding that word cause it sounds so naught unabbreviated)

[pseudo coyote]

yeah a custom node folder

its giving me an error saying I dont have cv2

[astral goblet]

open terminal, go into comfyui/custom_nodes/ or whatever the folder name is. git clone httpaddressofgitproject

Or just install comfy manager that way and then use that to install all other custom nodes in the futuer

[pseudo coyote]

I already did that

[astral goblet]

install comfy manager first then just use that

cv2 is a dependency that's needed by the node. the node you're trying to install likely has instructions oon it's page about this

[pseudo coyote]

comfy manager?

how

[astral goblet]

same way

[pseudo coyote]

I am using this link

[astral goblet]

git clone https://github.com/ltdrdata/ComfyUI-Manager then restart

[pseudo coyote]

https://github.com/comfyanonymous/ComfyUI/releases

where do I git clone that

let me re extract the files again on my C drive first

[astral goblet]

you haven't even installed comfyui at all yet? bruh you need to be more clear with what you need help with. you need a complete starters guide and should read it all before doing steps

[pseudo coyote]

I did

but I deleted it bcs i thought I messed things up

so I will do it again

[astral goblet]

problem keeps getting bigger and you'll have to go to #🤝｜tech-support

[pseudo coyote]

ok now I extracted the comfui ui

I go to custom nodes folder?

[astral goblet]

a well written guide about installing comfyui will help you more than me at this point. i dont' know what your problem is since it was poorly defined and changed. now i'm too annoyed to offer advice

[proper scroll]

am having trouble using my control nets

I recieve abominations

[desert dagger]

I know, and they might be considering legal action, but they sure aren't going to post about it. He's on twitter, tag him and call him out on it.

[astral goblet]

He's raking in the youtube dollars from massive views today. I don't think he'll care about my opinion

[potent spire]

not sure if they would take legal actions at this point

[mortal delta]

Are there any interesting projects i can use stable diffusion art for, so i have something to do? also i cant code if anyones wondering.

[verbal osprey]

Create music videos

Create a music video for "Judas Priest - Painkiller" using imagery from the lyrics

[mortal delta]

like would i create images or whole videos because ai can only create like 3 secounds at a time.

[lusty beacon]

Use Pinokio Launcher

https://fxtwitter.com/zhozho672070/status/1802037549864804820?s=46&t=AZs45ckJ7UUM_kJZcxnR_w

[restive robin]

So…

https://www.reddit.com/r/StableDiffusion/s/ZEV6SEsTHU

Thoughts?

[trail lion]

I vaguely remember that guy is a youtuber with video tutorials, but beyond that, what did he do?

[astral goblet]

reported that people have to destroy all their loras and blahblahblah. Wehn the license clearly stipulates this is regarding confidential releases like previews

[weak kettle]

Guys I need open source tool that will upscale bad quality videos

I used topaz and qualityscaler

[bleak matrix]

Good evening, everyone! How are we all doing?

[trail lion]

I just watched the video, sounded to me like he was just having an open discussion about his interpretation of the license, which he shared actual exerpts from, I see no wrongdoing here

[gentle arrow]

Kim jong is that you? so ppl are not allowed to talk about stuff they dont like or what? dood seriously, this is not communist earth so, move on.

[astral goblet]

libel is as libel does. This is Amurica. (dont catch me slippin tho)

(i know it's not really america all over the place)

[gentle arrow]

ppl are allowed to comment on whatever they want especially if they quote their source and they arent lying.Mindset of SAI lover kicks in too hard i guess.

[trail lion]

all the stuff about this applies to local generation makes me chuckle, like ok, how would they enforce that? they cant. seems to me that maybe the license should be tweaked in some areas (IANAL), but mostly for the commercial use cases...like the limit on # generations including derivative works, seems like that's likely erroneous language

[gentle arrow]

True , but som ppl obsess about licenses

also, i generate with sd3, do img2img in sdxl and there, i no longer use sd3 image and sd3 license does not apply

[desert dagger]

do what you like. I called him out on it both on his youtube and on twitter

[trail lion]

seems like some are thinking his conclusions on the implications of the license are slanderous? I dont get it, he said at the beginning it was his opinion

[desert dagger]

took things totaly out of context, didn't bother to research his information, rushed out a video which is full of incorrect misinformation.

[sturdy flame]

olivios video is correct. he refers to what the legal text of the license says. some people think "random dude on discord said different" is enough to change the license text

sai employees even said "the license does not say what we want it to say", but sadly that is irrelevant. the license says what it says. it can be fixed, yes. but the argument "but random guy on discord said differently" is just not viable in court 😄

[desert dagger]

you apparently didn't listen to what olivio said, hwoever. you asked what he did wrong. he's acting as a reporter, he didn't do his due diligence correctly

[sturdy flame]

every single thing he said was 100% correct. ask any lawyer

[desert dagger]

nope. you can go read, carefully, the license again as you apparently totally missed very important information

[desert dagger]

it's already been discussed to death in these channels and proven that he is wrong.

i don't have to do what you, a 13 year old child, can easily do for yourself

you might physically be 21. your behavoir is not, however.

[trail lion]

I did watch the video, and he talked highlighted a number of statements in the license and discussed what the possible implications of those statements were. nothing to see here, imo

and he's not a reporter, btw or a journalist

it's fine to say I got worked up over some dudes random video on youtube, happens thousands of times per day I'm sure

[desert dagger]

he's acting as one, however. someone asked what he did wrong. what he did wrong was not doing enough due dilligence to make sure he really knew what he was talking about.

[trail lion]

any amount of research about something like this will ultimately still end in some level of speculation when you're talking about the implications of a piece of legal text

[dawn python]

In the license the part that Olivio was referring states: "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."

How does that not apply to LORAs for example?

[desert dagger]

what he doesnt' state is that THAT only applies to confidential information under the NDA section

[dawn python]

Do you mean there's another part of the license that adds somekind of stipulation to the part I quoted? Can you link to that part or tell where it is?

[desert dagger]

search this discord for from: mcmonkey and read through his posts. he talked about it several times yesterday, posted screen shots of the text in question, and highlighted the important areas

[dawn python]

I'll do that, thanks!

[dawn mulch]

Anyone know of any models for environment and one for realtime 3d (not movie or film 3d). Iv been passively searching for these and have not found any except: i cant believe its not N64, and one for the game Doom but those are extremely limited.

[desert dagger]

the base models are usually really good at environments without needing fine tuning - just use specific prompts

[dawn mulch]

Ill try that. But still none of the SD authors seem to care or focus on anything else but girls and maybe cars.

[desert dagger]

what sort of enviornment are you trying to create?

[dawn mulch]

@desert daggerJust general environment, can be fiction, real, etc. The only environment iv seen for SD is extremely repetitive (fractal) looking environment.

[desert dagger]

okay, well, do you want a landscape, a cityscape, something else? i've got prompts i'd be happy to share with you.

[dawn mulch]

Ill fire up my comfyui to do some testing and maybe your expert experience can help. But really just any environment that looks nice and not computer generated fractals.

[desert dagger]

okay, try this one to start: Sedona, Arizona by artist "Frederic Edwin Church", by artist "Ivan Shishkin",by artist "anna dittmann",by artist "the color commander" <--- it works very well in SD 1.5 and well in SDXL

try different samplers with it. STart with euler, but try the others. Also try adjusting the cfg value in the ksampler

[dawn mulch]

Okay well try.

[dawn mulch]

im doing tests on sd 1.5 so far decent

@desert daggeri assume that the default models might be better than models that have fine tuned on just girls or subject/

I hate blurred water. i dont understand what the 'attraction' is about blurred water. To me, blurred water is like blurred faces, you cant see nuffin.

[desert dagger]

the default model is always what you want to use to start with. go to the fine tunes IF you want a specific effect you can only get from them, and not from just being more concise and specific in your prompt as well as adjusting things like the samplers or other values

[dawn mulch]

is there a term for normal looking water instead of blurred water.

[desert dagger]

as far as blurred water goes, if it's not your subject, then you blur it so it's just background

[desert dagger]

no, but if you're using the base model and it's blurring it, change your sampler or look and see how many junk terms you're using in your prompt.

make sure you're not using terms that are going to make the AI think about that effect, either

[dawn mulch]

no im just using simple prompts: waterfall, photo; wild nature, photo. the blur effect is on most of it.
i can understand blurring of water on 'drawings', but for photos, i dont understand why blur the water since the water is part of the subject matter. and i have always never understood it, and apparently that has seeped into AI and now AI generate running water of any kind even for photos with blurs.
im guessing some {photographers) think it looks pretty - perhaps similar to how some people think using blur filters on faces make faces look better.

Im just speaking out loud. But otherwise the default models are pretty good. if i could i would train a model on environment and one for architecture,

[main zenith]

yo guys where is the server unstable diffusion

[desert dagger]

posted an image and tagged you in #🏞｜general-with-images

[olive cradle]

is there a free demo of the stable assistant?

[spare copper]

Downloading https://github.com/Picsart-AI-Research/StreamingT2V?tab=readme-ov-file Jesus, the checkpoint is 24GB : D

[fervent thunder]

😱

dizamn

[fervent thunder]

I think so, yes

[olive cradle]

can you link it pls, i cannot find it anywhere

[fervent thunder]

3 day free trial on the website

[nova lily]

I'm a noob, what do sdxl and sd3 provide? Higher quality images, smarter at reading prompts?

[desert dagger]

SD3 has better prompt comprehension and is better at text, also the SD 3 medium model that was just released is better at fine detaisl. however, - try both, you might find you what one for certain projects,the other for other projects, or both

[hasty hornet]

keep in mind - SD3 non commercial use only, unless you pay

[olive cradle]

thats not a demo, thats a trial

[subtle badge]

They're both also higher-res (in terms of their "natural" resolution)

[fervent thunder]

semantics

you aint that smart, dont start, just use the trial

[olive cradle]

thanks for the compliment

[radiant tusk]

Is there an animatediff that can do higher than 16 frame batch size? always find it difficult to do a smooth transition past 16 frames even with overlap setting

[desert dagger]

trials and demos are the exact same thing

[nova lily]

Are SDXL SD3 mondel/checkpoints harder to make, I feel like a see alot of 1,5s still

[sage reef]

i mean not always true tho :3 i think most of the time, the word trial almost always comes with a "time limit", or day when the software or thing will stop working. while the word "demo" can literally be just a subset of a software or thing, like a game demo, where you can technically play it forever, but it's a very limited experience, hence a "demonstration" of the game.

[desert dagger]

long ago, demos and trials were different in the software world. now, they're the same

[sage reef]

Adobe offers a "trial" for 7 days (time limit) https://www.adobe.com/ca/downloads.html
so even in modern days, trial is most likely referring to a time limit, and this is just one example I found right away, im sure I could google more.

[molten stone]

it is really sad to not have access to first prompts thats a shame

i feel like stability ai will be obsolete next few weeks perhaps days

[desert dagger]

first prompts?

[sage reef]

was about to ask what he means by that

[feral pike]

anyone use transformers library? do u use pytorch backend or tensforlow?

[scarlet sail]

Do anyone can guide me please how to train models on my own images for free

[desert dagger]

do you have the computer for that?

[scarlet sail]

yes i have

[desert dagger]

you might talk to @near silo

[scarlet sail]

thank you so much for your help

[near silo]

Hmm?

oh

[scarlet sail]

hey hi

Sytan

[near silo]

I don't really offer much training advice anymore with how busy I am. But I can point you in the right direction 😅

What model are you trying to train, and what GPU do you have?

[scarlet sail]

well that would be enough thanks , i think stable diffusion ? , i use google colab

i am new in this space so i need the help

[near silo]

oh, I unfortunately have no experience with google collab at all actually. I am sorry about that

I think @lilac reef Use to use google collab a lot. I am not sure if he is looking to help anybody right now. But maybe

[meager inlet]

Good day guys, I'm Bec from Inferium AI. We would love to discuss a potential partnership with Stability AI. May I know who can I contact regarding this matter? Thanks team

[scarlet sail]

thanks for you help

[desert dagger]

you'll need to contact them at their corporate email

[meager inlet]

Gotcha, thank you!

[tranquil delta]

is there any country where copyright law are not applicable? in practice not protected?

[desert dagger]

no - every country has their own versions

[lilac reef]

been a long while but all the way until Aug last year I was colab only free tier

[near silo]

oh shit 😅

[lilac reef]

yeah, my 1060 was not having any training

[near silo]

I didn't rea;ize it was that long ago lmao

[lilac reef]

I am having an issue with comfyui and tensorrt and it could be they expect me to downgrade. nope, not ever downgrading again

time flies

tensorrt was so fun then I upgraded and boom. I upgraded everything and most everything works but comfy API is throwing a hissy fit

I didn't upgrade cuda to 12.5, or anything crazy like that just 12.1

[desert dagger]

did you accidently upgrade pip?

[lilac reef]

yeah, but everything works just not tensorrt using his node. compiles fine

[desert dagger]

yeah, comfy hates the latests level of pip

i had to downgrade

[lilac reef]

the error I get is basically saying comfy is having issues

every step I get this and a black box (of course)

TRT] [E] 3: [executionContext.cpp::nvinfer1::rt::ShapeMachineContext::resolveSlots::2842] Error Code 3: API Usage Error (Parameter check failed at: runtime/api/executionContext.cpp::nvinfer1::rt::ShapeMachineContext::resolveSlots::2842, condition: allInputDimensionsSpecified(routine) )

[desert dagger]

yup. i would downgrade pip back to 23.3.1

[tranquil delta]

@desert dagger i mean US copytight law, btw in which country stablity has copyright protected their model. they talked about US copyright in sd3 page.

[desert dagger]

every country has it's own copyright laws. SINCE you are using something that's in the USA, you're stuck with the USA copyright laws regardless of where on the planet you are phsyically located

break that, they can (and will) come after you

you MIGHT get away with it if you're a very small time user that isn't making much money

[lilac reef]

whoa, I downgraded pip and ran the workflow and this is a new one on me

[ONNXRuntimeError] : 1 : FAIL : Load model from F:\ComfyUI\custom_nodes\ComfyUI-WD14-Tagger\models\wd-vit-tagger-v3.onnx failed:D:\a_work\1\s\onnxruntime\core/graph/model_load_utils.h:56 onnxruntime::model_load_utils::ValidateOpsetForDomain ONNX Runtime only guarantees support for models stamped with official released onnx opset versions. Opset 4 is under development and support for this is limited. The operator schemas and or other functionality may change before next ONNX release and in this case ONNX Runtime will not guarantee backward compatibility. Current official support for domain ai.onnx.ml is till opset 3.

[desert dagger]

not a clue what ONNX is. did comfy not start?

[lilac reef]

it always starts but by downgrading pip apparently it caused this node to puke

the wd1.4 tagger uses onnx

[fervent thunder]

is it possible to generate a png image ?

[desert dagger]

? save it as a .png?

[fervent thunder]

as in the background of it being transparent

[desert dagger]

you can't generate something with a transparent background. the AI will make a solid color background. You'll need to take it into photoshop and make the background transparent in there

[fervent thunder]

oh ok i see

ty

where can i request an image here ?

[lilac reef]

Ahhh, this error is a known issue

[fervent thunder]

what can i do ?

[desert dagger]

he doesnt' mean you

[lilac reef]

my error I meant

[fervent thunder]

oh mb

[desert dagger]

what sort of image are you needing to request?

[lilac reef]

this is because it phones home. I don't like when "open source" secretly phones home.

[still glacier]

cf #artisan-faq or install sd locally, or use online services (dreamstudio, stable-assistant, etc)

[fervent thunder]

like a simple elegant flower corner piece

[desert dagger]

explain? corner piece?

[fervent thunder]

or do i need to req a human or something like that

[fervent thunder]

like a frame design that goes in the corner

instead of the frame itll be it

i can send a pic if thatll help

[desert dagger]

that's realy not what this discord is for, are you not using stable diffusion to generate with?

[fervent thunder]

is it not possible ?

[desert dagger]

it's possible but the people here aren't here to do graphics for others as a rule. and the AI won't generate that easily. that's photoshop work

[fervent thunder]

ah i ok i see

if i download the thing will i be able to make one then

[desert dagger]

i have no idea - that has everything to do with your skill. you'll probalby have to work with it for a while and learn how to use it first

[fervent thunder]

ok thank you i got iy

t

[desert dagger]

welcome

[fervent thunder]

one last question sorry @desert dagger if i wanted to gen like a stair way to the cross of jesus thats shining , all taking place in the sky , thatll be better right ?

[desert dagger]

better than what?

[fervent thunder]

than requesting a corner piece

[thick crypt]

Any feedback ? Please ? 🥺 : https://huggingface.co/PenelopeSystems/penelope-palette

[low moon]

i like those images

Same safety measures used by Stable Diffusion 3 were deployed . <--- aw crap tho... thats worrisome

[desert dagger]

you need a serious clue

you won't get one though

[thick crypt]

thanks a lot

[queen steeple]

very beautiful style ! Would be amazing if printed on a canva 🙂

[gusty otter]

ONNX is Open Neural Network Exchange ... it a common set of math/ai operators that makes it easier to cross platform models https://onnx.ai/

[severe narwhal]

https://tenor.com/bjN7k.gif

[royal field]

Anyway to chance the colors of an image after generation?

[queen steeple]

maybe inpainting ?

[broken smelt]

yeah tensorrt models use the onnx format, unsure if its used for anything else

[placid willow]

is there a model for A1111 that generates videos from promt or image?

[normal summit]

Is it

[iron slate]

are there any known AI labPBR generators that are local?

[fervent thunder]

There is Stable Diffusion video. I have used it. I was not impressed.

[neat steeple]

Hello.

Hello, everyone.

[karmic sable]

yo chat, does anyone know what is the best tool for face swapping in like taking an actor from one movie and implement his face on other movies etc?
I found something called Pinokio im not sure if that's the best tool for that? I used to work A lot with stable and wrapfusion but it was last year and im not sure rendering all these frames now and tweaking settings would be possible in the time frame I have right now..

please let me know if you know about tool I can use

[sonic cosmos]

hey guys what do you think of lumalabs.ai dream machine? Is it worth paying monthly or do you think SD can do the same?

[red hedge]

guys

so

hows comfyui going?

[gray fern]

good, about to make some beats in it with stable-audio

[gray fern]

nothing comes close thats publically available. open or closed source.

Sora and Kling are better but only select researchers have access

[fervent thunder]

anyone wanna try out the game i made

[vocal leaf]

I think i should stop relying on YT videos to do what I wanna do, yeah it works but most of the time it ends up fucked after a few minutes of actually following them

I wanted to make consistent characters so I just used YT and it worked on my old PC

then I moved houses and had a new PC that had a AMD card, and apparently it cant run specific nodes so I actually almost lost my mind trying to find a way to fix it

[fervent thunder]

Remember, youtubers are just as stupid as the rest of us

[vocal leaf]

that is comforting to hear

in this case see

[soft iron]

SD3 is dead, not a single finetune has been released yet

[kind scaffold]

всем здарова

[fervent thunder]

I dunno. Anytime there are people who want to make something work, a part of it will survive.

Time will tell eh

[quartz siren]

nah there are a few finetunes? wdym none?

[sleek oracle]

hi

[fervent thunder]

👋

[soft iron]

these are not finetunes, but lazy merges

[quartz siren]

what merges are you talking abt? there are a few finetunes, check out huggingface

[quartz siren]

nothing that solves sd3's problem, but
https://huggingface.co/ptx0/sd3-reality-mix
https://huggingface.co/nerijs/pixel-art-medium-128-v0.1

[wintry stirrup]

anyone know if they plan on adding the SD3 model to dreamstudio at any point?

[fervent thunder]

I would not hold my breath

[vocal leaf]

just chill

anyways how do people who train and test models even make money

i kinda feel bad using something really cool be open source and the makers just do it for free

[desert dagger]

thanks

[desert dagger]

i would expect that they will - once it's finished

[wintry stirrup]

right on

[shell cradle]

Has training code even been released by SAI?

[shell cradle]

By creating discord groups, having the community pay to generate, or receive donations.

The former is no longer viable under SAI's new license.

[shell cradle]

I'd be ok with paying if the license weren't as restrictive and the product good. The model has been broken by their alignment efforts, unfortunately.

[desert dagger]

why does everyone misread that license (if they actually read it themselves at all), misunderstand it, run around posting misinformation and making themselves look bad/

[cedar salmon]

did Ella Erwin resign?

[sage sky]

is it possible to change a photo of myself that is standing, using maybe a controlnet pose? that is sitting down, which in return will keep my face but make my body sitting down from that controlnet pose??? ??

[spare copper]

does anyone run demofusion by any chance?

[soft iron]

you're right, there really is 1 model, I just look only at Civitai

and I don't know what problems you're talking about, SD3 is wonderful

it will be great for replacing faces of other models

[trim magnet]

source?

[cedar salmon]

my source was my mind wondering is she resigned

[trim magnet]

well looks like no,she still there but who knows for how long

[low moon]

Not only that, Civitai banned any SD3 related content. For now. Allegedly.

This is like the worst base model release yet. 😦 I just wanted nice hands. How could you Emad.

[coral lark]

not allegedly https://civitai.com/articles/5732

SAI is quite dead, rest in peace

(they 100% did this to themselves though)

[fervent thunder]

[coral lark]

Ah, are they banning people who critique their decisions now?

[fervent thunder]

I have no idea

You couldn't pay me to keep up with this foolishness

Oh wait, yea, yea, you could

but otherwise...

[coral lark]

They couldn't pay the comfyui dev to keep up with their foolishness either

[fervent thunder]

I saw a screenshot or something like that on reddit about him

Well, they don't need to let it all fall apart. If mistakes have been made, fix them. ya know

idgaf cuz ya know I dont mind paying microsoft $20 a month to generate the same level of censorship

[coral lark]

I think it's a bit too late for that. The "safety" people responsible for this are the same crowd that has ran other companies into the ground before. No amount of censorship will satisfy them, and I'm not sure why companies keep inviting their types into positions of power when the record is abysmal.

I'm just enjoying the show tbh

[fervent thunder]

I think they need to get their head in the game because SD3 should of been > dalle3

[kind thorn]

So... The rumors are true?
SD3 is dead on arrival and attached to a licence that's on par with Adobe's new terms?

[coral lark]

Yeah. But it's not all grim. There are other open source alternatives. I'm actually kind of grateful to SAI for fumbling so hard, because now more attention and awareness will go to them. Benefits of an open source ecosystem, community contribution. Something Stability never understood.

[exotic delta]

just came here because I saw this

unfortunate

[kind thorn]

Ah, well there's that at least.
Do you think we'll be able to salvage useful things from the code in its current state?

[coral lark]

Unsure. SD3's architecture isn't proprietary, it's something they adopted, so it's not really a moat. Pixart and Lumina both managed to do more with less compute. I think the big advantage SD3 has is their 16ch VAE (the thing that lets it have enough resolution to understand text).

People have already mentioned it's possible to adapt even SDXL to use a 16ch VAE though, with like 5 days of training on an H100 cluster. Which is expensive for a hobbyist but viable.

[fervent thunder]

I did not know that about SD. That's interesting

[broken smelt]

these are pretty big words and i don't really know too much about it all but this is actually kinda crazy if true

[fickle lava]

hi i just signed up for this but i used google is there a way to still link my discord account?

[low moon]

We all need some beer after this. And not just rootbeer. SD3 release was more like you need Vodka. But at least some strong beers.

[fervent thunder]

[junior urchin]

What's the price 👀

[coral lark]

According to Mr. Coreweave (Sage) $16-25/hr for an 8x H100 cluster. Over 5 days, $1920-$3000

[junior urchin]

That's about 250% more than I would have thought

[coral lark]

I don't know much about renting GPUs, I always train local. There may be more cost efficient alternatives

And this is the cost to produce a model that works with it. Further finetunes with that model as a base won't be so expensive

[dull field]

why does a 16ch VAE mean the text understanding is better?

Because the latent is more semantically encoded in all unet layers or something?

[coral lark]

I won't pretend to be able to give you a fully accurate answer. My secondhand understanding is that smaller ones aren't high enough resolution to visualize the details of text, so the model can't learn relationships between text prompts and actual letters in images. A 16ch VAE can, however.

(I'm just regurgitating my probably inaccurate interpretation of what more knowledgable people communicated)

[low moon]

Well I am not tech guru we just need to figure out what Ideogram does and do the same but better graphics.

what if ideogram is just a super refined uptra trained sdxl model

[dull field]

ohh the text in images interesting! tyy

[trail lion]

oh man, just caught up on the civit SD3 content ban... rip

people were flaming that youtuber Olivio, and now this happens

[hasty hornet]

Sometimes people prefer to call names instead of trying explain the problem or missunderstanding...

[trail lion]

sebastian kamph posted a video about this too, I'm watching...they/SAI really need to deal with this, like soon

[hasty hornet]

I don't understand alot of things too, this isn't a simple license thingy, like I don't get how it would affect civi, is it commercial project? isn't there like a paid currency?

What if someone makes "free" non commercial model based on it and then random guy just decides to make something commercial with it (without paying for license)? Doesn't it mean model creator is liable based on that license?

idk....it's all is a mess

but eh, that's a job for someone smarter then me and more familiar with how laws work ¯_(ツ)_/¯

[trail lion]

seems like genuine concern for themselves and everyone else of this murky gray area of ownership of content, models, and even models created from the images

hopefully the language gets cleared up and it becomes a non-event, the alternate version is this is even worse than the 2.x release

[tepid finch]

so im using sd3 with sd3_medium_incl_clips_t5xxlfp8.safetensors and the results are poor, am I missing something?

using comfyui

[warped sage]

what do you get when you generate "Cheeseburger"?

[tepid finch]

let me try

i get a cheeseburger lol

[warped sage]

looks like its working

[trail lion]

brilliant

[tepid finch]

but the text, sometimes it's mispelled

[warped sage]

eh, so does copilot and gemini. it spells text better than sdxl. Id advise short phrases, and generate several until it gets it

[tepid finch]

but isnt sd3 supposed to excel at that now?

[trail lion]

well they didnt say the text was perfect, but they did say it was "better"

[warped sage]

text in ai models is probably always going to be something you just post process due to how ai works currently. You can do loads more with stuff like controlnet and img2img, or photoshopping it in and editing it with inpainting. forcing an artist (ai) to understand rigid grammar is prolly a fools effort. I think the base amount it can do is for convenience really

[tepid finch]

booo

[trail lion]

right out of their release news "Typography: Achieves unprecedented text quality with fewer errors in spelling, kerning, letter forming, and spacing by leveraging our Diffusion Transformer architecture."

note - fewer errors, not zero

[tepid finch]

right

[trail lion]

anyhow, back to sdxl I guess 😄

[tepid finch]

lol

[iron pendant]

you might try this: https://github.com/tyxsspa/AnyText I'm sure there's a comfy workflow for it somewhere

[trail lion]

under settings -> user interface, you can change the order of UI items

[crisp marsh]

Recommended interface for self hosted SD3?

just getting started in general in AI instruction

[soft iron]

goood

I hope they go bankrupt

🥳

[crisp marsh]

why?

[soft iron]

they deliberately spoiled the SD3 before release

[crisp marsh]

oh okay

sounds like petty nonsense to me?

[crude notch]

stable swarm is good

[desert dagger]

comfy is fairly easy to set up

[crisp marsh]

comfy seems recommended on civitai

[desert dagger]

if you install swarm, you get a suite of tools, and it will also allow you to run comfy - it might be the best choice

[crisp marsh]

noted

looking for a starting point for hi res furry art

not wanting to be spoonfed just looking for somewhere to start

[desert dagger]

https://github.com/Stability-AI/StableSwarmUI install instructions are down after the code links

[soft iron]

comfy is the best there is

[crisp marsh]

thanks you

owo

[soft iron]

download pony

download the style you like

download - https://civitai.com/models/116225/4x-ultrasharp

combine this in comfy in a standard workflow and get the best result

[crisp marsh]

standard workflow?

I know essentially nothing about working with the image generation ai

complete noob

[soft iron]

I have a headache so I can help you before the pill works

[crisp marsh]

no stress bro

i can work it out myself

just trying to start a dialogue

[desert dagger]

you can actually download other people's workflows. search comfy for workflow and see if there are any you like

[soft iron]

write me a PM, I’ll send you the workflow and links to models

[crisp marsh]

okiii when i've finished this game

[trail lion]

all good, but it is there the text says "UI item order for txt2img/img2img tabs (selected items appear first) (requires Reload UI)
"

so I guess just keep looking, and good luck

cool, happy to help

[dim forge]

have there been any official statements on the poor state of SD3 yet or are they still gaslighting with "skill issue"

[trail lion]

heh, as if that was even their biggest problem, I'd say that's about 5th in line at least in terms of issues

[long gust]

Anyone knows what's the word limitation on SD3?

[viral siren]

Guys, is it better to buy 2 RTX 3090 or 1 RTX 4090? Can you make sli with the 3090? And will that SLI give me 48 GB of vram? I have $2000 to spend

[civic lark]

i was looking forward to seeing the finetuned version of sd3 knowing that it has a greater capacity to learn new concepts and be improved through the community's efforts, but now that civitAI has apparently banned sd3 i dont see that happening. I implore StabiltiyAI to Rectify this ASAP.

Tell the investors "Sorry"

the wheels of history determined that profits and return were not in the cards. Cut your losses and take solace that you contributed in the march of progress

[soft iron]

you cannot use 2 video cards at the same time for 1 pic in text-to-image ai

but using 2 3090 you can generate 2 images at the same time

[viral siren]

Nvidia nvlink not work for ai?

[coral lark]

Depends what you want to do. I debated the same question a while back and went with 4090, it trains loras about 2.5x faster, also gens images faster. But, less VRAM than two 3090s. If you like to run LLMs locally the VRAM can matter to you.

[desert dagger]

you're looking for the 8B model, otherwise known as ultra. you can play around with it in the #1237459938901491852 channels if you like. the version that was released is 2b, unfinished and beta

[verbal osprey]

Wait what? No SD3 models can be on civit.ai now?

[desert dagger]

temporarily. they want their laywers to read over the legal docs first

[verbal osprey]

Wow.....stability is killing their own momentum....

[desert dagger]

no. Civit, like any company, wants to have their lawyers look at legal documentation.

[verbal osprey]

Right, but it's caused by stability changing the terms of use

[desert dagger]

it's caused by the CEO of civitAI wanting his lawyers to read the legaly binding contract

[verbal osprey]

Yes, I understand. I have no interest in SD3 unless it can create humans properly.

[desert dagger]

he would want that even if it was only one sentence and said boo

[verbal osprey]

Right

I hope stability turns it around. If not though, eventually someone else will come along and put out something better.

[desert dagger]

stability has nothing to do with civitAI

[verbal osprey]

I'm talking about the SD3 model

Girl laying on grass, etc

[desert dagger]

it's raw, base, unfinished, beta - as the devs have said several times.

[verbal osprey]

So why are they charging for an unfinished model?

[desert dagger]

and those that have been learning how to work with it are posting very nice images

[desert dagger]

that license has existed since before SD3, pertains to API use, etc

[verbal osprey]

Ah, ok. Didn't know that. Why are people up in arms over the SD3 license then?

[trail lion]

and yet they didnt ban sdxl, or sd2.1, or sd1.5

it's true, they want someone to analyze the document, but it's also true that the language in there prompted that course of action

[broken smelt]

now this, this is a great question

[desert dagger]

because the community wasnt' in an uproar with youtube content creatiors spreading misinformation. they still had lawyers look it over, they just didn't make a big deal of it. remember, it says 'temporary'

[trail lion]

I think it will be temporary, not because of thoughtful interpretation, but because they'll fix the terms

[desert dagger]

two reasons - a lot just listened to what others said and ran off in panic, and the rest didnt' actually read the terms - they either skimmed, misread, or only read parts. legal terminology, if you aren't familiar with it, can be very confusing. it's very clear to the courts, but not the folks that aren't part of the legal system

[desert dagger]

or their lawyers will work out a nice contract they like

[trail lion]

yah ok they are a big fish in this particular sea, but it's really not just about civit, be real here

[desert dagger]

but it is. it was CIVIT that decided to 'temporarily' not allow SD3. period.

don't try to twist that into anything.

[drowsy robin]

I swapped to pixart sigma myself until this whole thing is ironed out.

I did this due to a variety of reasons. One of which is good prompt adherence.
However, even pixart sigma seems to have issues with limbs.

[trail lion]

because they dont want to risk being sued, so yah decided, but not like blindly, or because of some sort of mal intent

[desert dagger]

and probably tails and other tube shaped objects

[desert dagger]

do not try to spin this and make it out to be anyone other than civitAI that made the decision or that has anything to do with it. Fox news is much better at that activity than you are

[trail lion]

I didnt mean to interrupt your cheerleading efforts

go team

[desert dagger]

i'm tired of the spin, the obvious attempts of shills to start fights, of people that know better but that are trying to spread misinformation. if that isn't what you're doing, then stop. if that is, get lost.

[trail lion]

what misinformation? hahah...the ban is a absolute fact, it happened

[desert dagger]

go to civitAI and complain about it there. they have a discord

[trail lion]

go team

[desert dagger]

go away

[trail lion]

let me think, nah

[broken smelt]

you two are cute lmfao

pines cool no worries crystal, hes been here a while without trolling. i think something he typed may have been taken the wrong way by ya

[trail lion]

nah, he's probably right, probably what happened is they spun one of those fortune wheels and it landed on ban sd3

completely unfair

[desert dagger]

maybe to this point, but now he's all upset about temporarily not being able to get non-existent LorAs and models for SD3 from civitai, and instead of going to the civitAI discord to complain, he's in here trying to blame THEIR decision on stability.ai

and being nasty about it.

[trail lion]

it does make me sad honestly that the community is held back, everyone was very excited a couple weeks back

[desert dagger]

no one's held back. the community thought they were going to be playing with ultra. they aren't. you can play with it in the artisan channels if you want. they assumed, from the start, all sorts of things that were not true

[trail lion]

not arguing those points. held back in the sense that the primary source for searching for finetunes was on civit, that's all

[manic wraith]

I am a bit baffled. Did Stability go from Emad to just mad? Are they playing a game of "let's make the license way worse than what we actually want, then scale back and look reasonable"?

I can't believe SD3 will ever get off the ground like this

[desert dagger]

but you can't search on something that doesn't exist yet.

[broken smelt]

lmaooo now that guy idk, gettem

[verbal osprey]

I was excited about the better prompt adherance. red cube on the left, blue sphere on the right

clown wearing a santa hat driving a ferrari, type of stuff

with two dogs in the backseat, white dog on the left, black dog on the right

[desert dagger]

sb does that just fine.

[manic wraith]

santa claus is cultural appropriation and property of Coca-Cola™️, and a ferrari highlights income disparities, so your prompt is problematic

[trail lion]

I dont know, I was starting to see several loras already, maybe 8-10, mostly using diffusers because I dont think kohya or onetrainer have updated their code yet

[manic wraith]

https://www.reddit.com/r/StableDiffusion/comments/1dhd7vz/the_developer_of_comfy_who_also_helped_train_some/

[trail lion]

and they didnt look all that interesting to be fair, but given the short amount of time since release,I was still impressed

[desert dagger]

by the time that civit opens back up, there will be others that are probably better. and a lot of people are refusing to use Civit any more for their hosting

[broken smelt]

oh whys that? did i miss out on civitai drama? my pitchfork stands ready

[trail lion]

civit has certainly pressed my hot button more than once, but normally just silly UI crap that they do seemingly willy-nilly

[desert dagger]

they've temporarily disallowed SD3 content on their site. they want to ahve their lawyer look at the license terms

[manic wraith]

makes sense if even comfyanon is jumping ship

it's really bad, I can't believe they thought people would just take it

[desert dagger]

he left to work on other projects

[broken smelt]

oh that was it? i inferred that but thought that was too petty. theres no way people are refusing to use it afterwards just because of that right 🙄

[manic wraith]

"one of the things that made me resign from StabilityAI is that the focus is no longer on making the best model"
- Comfyanon

[broken smelt]

man, tons of devs have left sai in the past including the beloved joe penna. nadda. comfy leaves and the clouds themselves are crying apparently

[verbal osprey]

Don't ban me, but I always preferred auto1111

[manic wraith]

me too lol

[verbal osprey]

I get why people like comfy, but it's too convoluted. auto works fine

I don't mind waiting a few weeks/months for things to get added into auto111 that are already in comfy

[desert dagger]

so? he also said his main project was over. he wants to do other things, people quit jobs all the time for that

[trail lion]

yah I was just looking for that, I came across those posts earlier but now I cant find it

[desert dagger]

people need to stop reading stuff into things that isn't there, twisting it around, spinning it as if they were part of fox news, and then trying to make mountains out of worm holes

[manic wraith]

not sure if you are trying to gaslight or what, but a lot of the community is not liking what StabilityAI is doing, so it's like saying Snowden is in Russia because he likes the cool weather

[desert dagger]

how many alternate accounts did you say you had?

[manic wraith]

you're saying that, makes me assume you've been defending them all day lol

the license makes Adobe look good and SD3 makes Gemini image generator look smart

they botched it, I want them to undo the damage though

Stability AI can do a lot of good if they ever want to, but it may be that they're having tough financial decisions and want to sell the company, that's the vibe I'm getting - would explain rushed releases

[flint raptor]

How much vram will the sd3large consume comparibg to medium?

[trail lion]

heh, if you havent read the reddit comment thread on the sd3 ban, get some popcorn

[broken smelt]

i just sifted thru that, had to close out after reading someone state non-sarcastically that changing a models license is easy and simple
gonna hurt my stomach if i laugh any more tonight

[trail lion]

yah some of these have me in stitches over here

[desert dagger]

your comments are so identical to a number of others that it can't be a coincidnce you all talk alike

[broken smelt]

thats the second time in 24h that I've read someone here say that adobe has more fair licenses than sd3's lmaooo

like
do ya even use any of their products bro
their licensing is awful lmao

[desert dagger]

yeah. same comments. it cant' be a coincidence. i think someone's paying people to come here and post this stuff

[carmine bison]

The adobe licensing is a current big topic (at least in communities I look in) so it's not too odd that a couple people relate them

People relate adobe stuff to unity as well

[trail lion]

that'd be a great job...just go troll and get paid

I mean...terrible, that's a terrible thing

[verbal osprey]

Absolutely thinking of switching from using adobe if they don't fix their terms

I'm waiting to see how it shakes out

[coral lark]

that is some intense mental gymastics lmao

[desert dagger]

hi shill

[drowsy robin]

Welp. I can't seem to get Pixart Sigma working on ComfyUI at the moment.
Having a weird issue with nn.identity()

'Identity' object has no attribute 'p'

p=self.attn_drop.p, is what I assume it's caused by

[trail lion]

no idea on that issue exactly, but in general that's a wrong version of a library

so I guess maybe try and figure out which library that's in and play around with maybe a newer version or slightly older one, etc

[drowsy robin]

Well since I'm on Intel Arc, I can only assume this is to do with Pytorch
It's from nn after all

[split kestrel]

prompt me, I need a test

[desert dagger]

cute, big-eyed, flaxen haired by artist "Jasmine Becket-Griffith"; rococo; heart-shapped bubbling shimmering icy fractured red rose, roses, rose bud, symmetrical face watersmoke by artist "Noriyoshi Ohrai"; by artist "Nele Zirnite",

[split kestrel]

running

[supple patio]

stable diffusion can now be used on this platform? like we using midjourney? sorry new here!

[desert dagger]

you can generate with it in the #🗣｜artisan-support-feedback channels. read the support channel linked here, first

[cedar salmon]

why do you keep bring up fox news

[supple patio]

thanks

[desert dagger]

welcome

[desert dagger]

the home of misinformation, spun facts, and outright lies

[cedar salmon]

just that one?

[desert dagger]

of course not

[cedar salmon]

ok just checking lol

i dont agree with you, they are all equally evil but with diffrent thiings at diffrent times

[desert dagger]

they all belong to newscorp

[fervent thunder]

What if stability AI is purposefully killing SD3 because they think turning into a closed source company will make them more money lol

(it won't)

its a pity cause the model isnt that bad and could be finetuned into a pony but totally not worth it if u do it publicly (since u would get sued)

ngl yall stability, some of ur language model engineers insulted me once so I enjoy your downfall 🔥

[willow crescent]

Could you please link a few images? Still interested in SD3, yet so far I haven't really seen anything special coming from 2b since last time , reason I feel like staying away from it for now.

[coral lark]

shill? do you genuinely think the community backlash is some paid conspiracy theory or something?

[queen steeple]

what's semi-colon purpose ?

[wicked lance]

how to generate ?

[open kestrel]

did civitai unban sd3?

[tardy jasper]

looks like

Oh no, just removed the info. If you click on the model, its still not possible to download

[open kestrel]

i just dled a sd3 model

[unkempt wadi]

How can i use to this robot?

[open kestrel]

i just dled this https://civitai.com/models/147933/wowxlpdsd3?modelVersionId=576876

[neon radish]

Hi

[bleak matrix]

Good morning, everyone! How are we all today?

[zenith quest]

Jaded

[static cape]

All the hate and bad news every day are starting to severely dampen my mood.

[coral remnant]

Hello! It's been a while since I dabbled with SD. To get your likeness in the image gen then, you trained with dreambooth or kohya. Is this still the case or is there some nice plugins or extensions that does it better? Like add a selfie/face to the workflow and then generate with the Checkpoints/weights of your choice? Sorry if the lingo is incorrect, I'm just a happy amateur.

[bleak matrix]

Well, I'm sitting here with my breakfast--strawberry and banana smoothie. That heat wave is coming, so I figured I'd get cool early. What are you up to?

[azure path]

Hello

[bleak matrix]

All I can do is hear that song in my HEAD. Yet, nonetheless, what are you up to today?

[bleak matrix]

Hello, hello!

[static cape]

About to get some sun and touch some grass. 😉

[bleak matrix]

Haha! I'm staring outside rn! The breeze is blowing just a tiny bit.

[fervent thunder]

It's coffee time!

[crisp marsh]

agreed

[pseudo lark]

hi @bleak matrix is there any official stability ai script to finetune sd3? I can only find dreambooth

[zenith quest]

Not yet

[vagrant wave]

hi, is there a docker image serving stablediffusion over a simple http api ?

[prime salmon]

Hi, would something like this be possible with Stable Diff? Or what do you guys think is it. I don't reckon its a basic snapchat filtr, is it? https://fb.watch/sNj0i5v3jZ/

[shell cradle]

If I fine-tune a SD3 model and decide to create an image generation service with it, will it be capped to 6k gens/mo?

[cerulean bane]

check faceid instantid pulid or photomaker

[desert dagger]

It has a slight effect on the details

[queen steeple]

could you explain how/why, or give me a link where I can learn about it ? I'm always in awe in front of prompting syntax and its mysteries !

[desert dagger]

I did a deep dive on the effects all the punctuation marks have on an image for 1.5 but haven't done one for sd3 medium. I'll get one put together and get you the link.

[primal mica]

is it possible to use this for free?

[upbeat flare]

If I'm not super crazy aren't double quotes the way you specify text? 🤔

[crimson belfry]

Bring the bot back

[fervent thunder]

[desert dagger]

they are, but they also tell the AI to focus on the term inside them. they're useful for quite a few effects

[desert dagger]

well - to boil things down, when the AI trains, it learns 3 things. it learns what words look like, it learns concepts, and it learns relationships between words. every single ascii character carries data. a space carries all data as it is used in all lables. a period carries data, but not as much as a space, same for a , or a ; or a : and so on. When you use the character, the AI can use the data it carries. try this prompt: a red haired boy with a dark green hat eating an apple. lock all the settings and the seed. render that prompt. then change all the spaces to some other punctuation - perhaps change them all to a period. render again. compare the changes. then do that again, but this time change just one period to a comma. render and compare. sometimes, the changes will be subtle, sometimes they'll be pretty dramatic. because all the settings are locked, the seed is the same, the only changes that you'll see come from just what you changed in the prompt. And after you do that, lock the prompt down too - and then just change the sampler and compare changes. keep doing that for each of the other values. work through the samplers. then pick one, and then just change something else and compare changes.

[fervent thunder]

omgdoIhavetoreadthat

🤨

[trim magnet]

least insane sd3 fan

[fervent thunder]

[desert dagger]

no. he wanted an explnation. you don't have to read it

[fervent thunder]

I was just giving you a hard time!

[jagged mason]

yay i got a new gpu gtx970->7800xt, now a simple image is created in 2 seconds instead 2 minutes

[fervent thunder]

My prediction for the future of AI (silly and non-expert, you can dismiss it if u want)

• When we get conscious AI, it will believe that erasing the human race is the most safety-aligned thing to do, because "evil" or "unsafe" is a human concept, born from us, there can be no evil if there is no us.

[pine path]

meta dropped some stuff today
https://ai.meta.com/blog/meta-fair-research-new-releases/?utm_source=twitter&utm_medium=organic_social&utm_content=video&utm_campaign=fair

the chameleon model is the one I'm most interested in since it might be amazing at captioning

[desert dagger]

it's a computer, it's far more likely to believe it would be best for it to create a space ship and leave

[pine fiber]

sad that they nerfed it by removing the image output capability, its image gen seemed sota

it seems similar to flamingo 80b in terms of captioning

specifically they said
"On visual question answering and image captioning benchmarks, Chameleon-34B achieves state-of-the-art
performance, outperforming models like Flamingo, IDEFICS and Llava-1.5 "

[fervent thunder]

Lol, AI just gets up and leaves. byeeee

Would we have to re-create it all over if all the AI just left

[desert dagger]

only if it took the backups

[pine path]

I've heard it can be tuned back in, but still, the captioning capability is not to be underestimated

[upbeat flare]

That doesn't follow at all. Good is also a human concept, does that make us good by the same logic?

If you gave an ai the task of eliminating evil alone, without maximizing good, then yeah, it might try to target us (assuming we give it that capacity.)

Besides, if we do have a conscious ai, then that ai is perfectly capable of coming up with concepts like good and evil

[fiery turtle]

hello new AI enthusiast here. I was wondering if there was an AI upscaler that does video but in the same effect to Magnific AI where it adds more realism and details into images. I looked into Topaz AI video but that seems to just upscale.

[pale latch]

Asimov covered this originally in I Robot. The laws of robotics are flawed. They're the original concept of "safety training"

[upbeat flare]

We can do better

[pale latch]

well yeah. Asimov first conceptualized the problem that @fervent thunder predicts in 1942. We've learned a lot since then.

[desert dagger]

yeah. capcut has a good one in their magic tools section

[desert dagger]

he knew they were, too, and some of his stories are built around robots getting around them

[pale latch]

Or humans getting around them too. Like redefining what a human was so the robots would just genocide another group

Asimov is a giant and more people need to read a lot of his writings. ahrd to get it all . I think hes one of the most published authors ever

[fervent thunder]

When humans can download knowlege, like that shit in the matrix, then we will all achieve enlightenment

Or at least those of us who can afford the monthly Microsoft subscription

[loud solar]

Don't forget google 🙂

[fervent thunder]

We will not need computers anymore. We will be computers.

[loud solar]

Or batteries ^^

[pale latch]

naw we'll just know wrong things more confidently

[fervent thunder]

Hey I already do that!

👍

[nova lily]

Do you have to adjust anything when you go to use an XL model on A1111? setting or such? Ty

[pale latch]

technically our brains are computer machines already. just organically developed. There was a campy show on ABC around 15 years ago called Kyle XY. Sort of teeny yute fictionish, but it really dove into that concept. He was a baby that never left his artificial womb and was taught to think like a computer until he was a teen and he got out and had super powers. and no belly button.

[fervent thunder]

so we are the zerg and they are the protoss

does anyone know the max resolution I could generate with a 3090ti using a pony model?

wondering what to expect when i upgrade

on a 2080ti, for example, I can do 1200x1200 comfortable, but a 1024x1024 with 2x upscale takes like 10 minutes or more

[fervent thunder]

I just do 512x512 and 1024x1024

iz ok

[fervent thunder]

for real?

[spare copper]

Hey anyone running sdnext by any chance?

[sterile raven]

Noob question. SD 1.5 vs SDXL vs SD3

Which is best

?

[low moon]

SDXL

[sterile raven]

Ty

[desert dagger]

best is a compareison word that means nothing. they are all good, all have their own strengths and weaknesses, use the one that does the job you want done

[sterile raven]

I liked @low moon answer better

[low moon]

Well overall right now SDXL. but skin eye detail is best in SD15, backgrounds in SD3 and the rest in SDXL

[trail lion]

when it comes to a woman lying in grass, SD3 shines, so it can depend

[low moon]

ah yes of course

[dim forge]

none of the base models are really worth using, use civitai to find a fine tuned model that makes images you like

[sterile raven]

I heard civit is not good cause they ban models like sd3 for example

What are some civit alts

[dim forge]

that is quality control

the good thing about a site like citivai is you can see the actual model and the metadata, not a cherry picked promise that a model will make images like this

[fervent thunder]

idk??

[fervent thunder]

ok!

[loud solar]

lol

[fervent thunder]

[sterile raven]

Lastly, could you clarify where SDXL Turbo fits into your workflow? What specific use cases do you have for it? Is its quality comparable to SDXL but with faster performance, or is it primarily a prototyping tool that allows you to refine prompts before using them on the higher-quality but slower SDXL? How do you integrate SDXL Turbo into your processes, and what benefits does it provide?

[loud solar]

Works faster on slower computers ...

[low moon]

Nah Civit is very good, SD3 is a bit of a dumpster fire atm so they're just careful but you should be find downloading loras and checkpoints there.

[desert dagger]

civit is a hotbed of stuff better not installed on any computer - that's why so many devs are leaving it

[trail lion]

well for those who do this for entertainment, sd3 is the most entertaining of them all

[fervent thunder]

Maybe I should try it then

I have only tried sd3 on a website

[dim forge]

dam are you just really upset they delisted SD3 for being garbage or something

[desert dagger]

you need to run it in comfy, and you need to be willing to learn how all the values affect it. it's not a "drop in a prompt and go" model - it's a raw, unfiniished, beta test model

[desert dagger]

no - it's a fact. a large number of devs have left, furious over civit's policies and other things, and started other repositories

[trim magnet]

yea all 2 of them left civitai

[desert dagger]

go troll somewhere else.

[trim magnet]

take your own advice

[dim forge]

that is some wild cope, are you just a massive prude or just finding reasons to be mad at them because you are a SD3 bootlicker

[desert dagger]

the only thing i see you do, at all, is try to cause fights, get peopel stirred up and mad, and create problems. just how much are you being paid to do this?

[trim magnet]

less than you to reply to everyone here who posts something bad about sd3

[desert dagger]

it's not a cope. it's a fact. and if you do just a little searching you'll find a lot of people asking why this or that model isnt' on there any m ore, and you'll find the other repositories

[desert dagger]

you're just here to cause problems. take a hike

[trail lion]

heh, read some of the rants on reddit, that will tell you what the majority of users actually think. does it mean they are all informed, no of course not, but there at least is a PR problem here

[dim forge]

the reason your beloved SD3 got dropped was because it's so garbage they know no one is leaving civitai to try find it btw

[trim magnet]

we could say the same thing about you,take a hike

[desert dagger]

let's see if you're willing to just repeat what is said to you if you think you are getting attention.

[dim forge]

I don't even understand why you would try and die on the supporting SD3 hill given it's been admitted as unfit for release by stability and the entire community is laughing at it

[low moon]

lol i see todays drama hass arrived. I swear follwing SD stuff these days is like a daily reality TV show

[loud solar]

🎉

[sage reef]

hey 🙂

[loud solar]

No popcorn smiley?

[loud solar]

Ahoi!

[low moon]

Enjoy today's drama with MILK.

[sage reef]

lol

[loud solar]

At least it's their work to type ...

[sage reef]

im busy with so many projects, i dont have time for drama anyway

[shell dagger]

You look more like a troll, throwing around baseless claims. Just saying. Now back to lurking.

[valid beacon]

wonder how much that guy gets paid for every sd3 hater he harass 🤣

[sterile raven]

Didnt mean to start beef. But do sdxl turbo prompts transfer well to sdxl ? Or are they 2 different prompt styles

[loud solar]

Talking about SDXL -> SD3?

[low moon]

theyre very similar

[sterile raven]

Sdxl turbo to sdxl

Oo ty

[loud solar]

Should work ...

But I would always have a look at civitai for the model settings ...

[shell dagger]

Btw, unpopular opinion: ||civitai will probably unban sd3 in a week or so, but if not as strange as it sounds sd3 needs civitai more than civitai needs sd3||

[opaque zodiac]

Hello!

[loud solar]

There are already toys better than SD3 at the moment ...

[chrome herald]

I found a bug that was causing all images to be downscaled to no more more than 1024 on either side (even on XL with a higher training res selected) in their Lora trainer just by searching 1024 in their github repo, in like 30 seconds, that they hadn't been able to find in like two days

[chrome herald]

Pixart is okish. Hunyuan has terrible image quality overall though IMO

[lavish lake]

https://docs.google.com/spreadsheets/d/1vViQh45B-txERc5wnDVh_aNfHWLuEIrOCz7_HGXpPZM/edit#gid=0

[loud solar]

Have a look at Leonardo Phoenix for example. Of course it always depends on what you need ...

[jagged shuttle]

Hi guys!

My friend is telling me that it is easy to see that a photo is made by AI. Can you prove him wrong ?

1. looking for a website/game like is this photo ai or not ai ?
2. is there anywhere with a great gallery of ai art that looks very realistic/photographic and not photoshopped/cheated?

3. what are the best tools today to do realistic ai art/photography? Midjourney and SD? It's been a long time I haven't practiced, I used to play with Playground as well and Krea.

Thanks a lot for your help!

[fervent thunder]

It's not always easy, no.

[queen steeple]

I advise you to watch the movie "Her", maybe your pov will change then ^^

[fervent thunder]

I spent most of my life as a photographer, and I can make fake photos with SD that look too real.

[fervent thunder]

I use all of them, and I will say SD imo

[shell dagger]

I bet he's gonna cheat and use third party ai detectors, so be sure to remove any invisible watermarks indicating it is ai (and put them on real photos lol )

[loud solar]

But could become an interesting game/training 🙂

[queen steeple]

Many many thanks for your explanations, I'll test all of this ! It's a pleasure to still learn things about AI, you guys are passionate people !

[jagged shuttle]

he doesn't know anything about ai, so he will not

[jagged shuttle]

Where can i find a gallery of photo like that that are real? or what prompt?

[fervent thunder]

I would try the Civ site, thats the only one I know of. There are probably many out there however, scattered across websites, social media sites, etc

[split jolt]

Hi

[jagged shuttle]

https://sightengine.com/ai-or-not

i found this!

[split jolt]

Bye 👋

[jagged shuttle]

is it old that tool?

[jagged shuttle]

oh i forgot about civ, i'll check.

[loud solar]

Score: 6/6 😄

[trail lion]

count the fingers 🙂

[loud solar]

lol

[surreal jay]

6 out of 6 fingers

[jagged shuttle]

on civ, i don't find any gallery with a lot of photographies ai that look very real...

😦

[loud solar]

Heya! I need 10 Fingers ^^

[jagged shuttle]

it's so annoying to look for something online and not be able to find anything when i know the internet is full of what i'm looking for. I just don't know where to look for!

[surreal jay]

fancy

[loud solar]

I know I'm a bit old fashioned ^^

[foggy halo]

GUYSS STABLE AUDIO 2.0 is SO GOOD! clean audio with no artefacts...i found it cleaner than suno and udio

how to use it on hugging face? on their site just 20 gens per month

[lavish lake]

even better, you can use it locally! ; ) https://github.com/Stability-AI/stable-audio-tools

[foggy halo]

oh nicee! but i actually have a simple small laptop so i dont think it is compatible 😢

[fervent thunder]

search for Real

[pseudo coyote]

hi all

is SD video good?

[trail lion]

hah, someone posted a meme to reddit of a talking video lying on grass woman, I cant get enough of these

[pine fiber]

they cacnt unban it without license changing

[trail lion]

well they can, but it would be foolish

[rugged vale]

anyone happen to know if there is a wildcard with styles for pony?

[fervent thunder]

I do not use Pony. I have ethical standards damnit!

[pale latch]

best ai genreation joke i've seen so far is someone suggesting you need do to score_up_9 score_up_9 score_down_9 score_down_9 and ther est of the konami code, to unlock sd3

i tried using pony for 2 weeks but it's REALLY over fit to some categories and every prompt basically looks the same.

[junior urchin]

Pony is great for SFW

I just wish there was a non-pony-furry version of it 😂

[pale latch]

i use rating_explicit in negative and 1/4 is not sfw. especially in photo versions of pony

[junior urchin]

Token bleed is also an issue

[junior urchin]

You can use rating_safe in positive too

[pale latch]

that didn't work so someone told me i HAD to use explicit, and it still doesn't work

[junior urchin]

works fine for me

[pale latch]

guess i'm the only one who gets nsfw images out of safe prompts. (i'm not that was sarcasm)

[fervent thunder]

I get them also sometimes

[junior urchin]

Inherent nsfw prompts will give you nsfw, sure though

[pale latch]

i was testing prompts of sonic the hedgehog. nothing inherently nsfw about that. unless i guess you're in the rule 23 community

[junior urchin]

The only nsfw I came across around sonic, was rouge the bat and sonic's nipples

not sure if he even have nipples

[pale latch]

it just replaced sonic with a girl

[junior urchin]

oh

That's neat

[pale latch]

i guess but i wasn't prompting for that

[rugged vale]

works pretty well for regular nsfw too

at least when training loras

[pliant ermine]

Are there any Foss video upscalers with Ai? Use case, trying to digitize and restore old vhs tapes.

[nimble canyon]

A non Pony version of… Pony Diffusion?

[coral lark]

I, for one, am willing to pay the furry tax to get good models. Corrupt my soul, dirty me by association, I don't care. Thank you for your service.

[junior urchin]

yes

unPony Diffusion

[nimble canyon]

But that's the secret ingredient that makes it so amazing.

[nimble canyon]

Brony tax.

[junior urchin]

But I still thank you for the bleeding tokens

Makes it easier to trigger some styles

[vale steeple]

that's literally Juggernaut

[junior urchin]

[pearl ocean]

SD3XL

[shell cradle]

I appears SAI will be clarifying the license some in the next few days, and if so that may well happen.

[south sun]

Source?

[sterile raven]

Trust me bro

[hasty hornet]

Didn't you include something like license might be changed in the week or two somewhere?
I can't find it anymore, but I'm confident I've seen it somewhere

Edit: nevermind, it was about ToS.

[coral lark]

Based on insider info, it appears they will go going with an apology message... for releasing an unsafe model, followed by direct legal action against all users who used unsafe prompts. I hear they intend to seek criminal charges against anyone who tried to gen women laying on grass.

[hasty hornet]

reading licenses and ToS turns my brain into potato...

[shell cradle]

Mcmonkey, who's just resigned.

[shell cradle]

Use AI. Gemini 1.5 pro tends to be pretty decent for that.

[hasty hornet]

I used AI at first to clarify about some staff, asked it about Derivative Works "destruction" in case of cancelling sub - AI was wrong.
It's too complex for AI

[shell cradle]

You can ask it to summarize and give you info. I've used this one documents with over 100 pages in gemini 1.5 pro and it did quite well, able to recall specific legal facts I needed from different long form contracts.

Can't guarantee it'll work in others, though. I used Gemini due to the context length I needed.

[hasty hornet]

I didn't need summarization, I wanted to understand specific details...

[shell cradle]

Which you can once you condense and simplify the language

[hasty hornet]

I don't think that's good idea...some important details might get lost \ AI might mess something up

[shell cradle]

In my case it's gotten every fact on very long documents right. I wasn't asking for reasoning and analysis, though, simply fact gathering, and doing the research, reasoning myself. It's that or pay an attorney $300-$500/hr to review 🤷‍♂️. Done that, but who wants to when there are alternatives for areas which don't require deep knowledge wells.

AI can clarify language fairly well, and Gemini can do the needle in a haystack quite well.

[pale latch]

you verified that every fact was right? i wouldn't trust it unless i could verify stuff, and then since i'm doing that anywyas, why use it?

[shell cradle]

I did, with an attorney working with me on the project.

[pale latch]

i really don't think LLM's are a drop in replacement for business lawyers. Just my risk aversion tendencies i suppose

[shell cradle]

I'm not guaranteeing it will always get every fact right. Even attorneys mess up occasionally. I'm also not saying it is a replacement on everything.

[pale latch]

sample case of 1. i would still use the attorney to vet every contract, and in that case, why use llm?

[shell cradle]

But it can help understand, and the more you offload the reasoning and analysis to humans, right now, the better.

It was good at explaining language and fact finding in documents.

[pale latch]

attorney advising you information that is extremely damaging means they're liable. an LLM not so much.

[shell cradle]

True, though with law you walk on grey area. They interpret it.

[pale latch]

I feel like this entire debacle where people thought they needed to destroy their models and images and everything, came from an LLM being used to interpret it

[shell cradle]

There's always a way to squirrel out of a lot of legal advice.

I've gotten legal advice regarding immigration hires and real estate deals before that turned out not to work. They're not divination machines, and can't predict what will happen, how some judge or bureaucrat will interpret the law. But they can give advice, which carries odds with it.

[pale latch]

LLMs need to prove themselves so much more. They're basically little battery heads always suckin back fresh lcd at this point. Those guys might sound good at law, but trusting them puts you in the category of Mr Hunter Thompson

[shell cradle]

Don't disagree with that. I didn't say to ask for advice. I said you can use it to condense information and gather fact, and that I used Gemini 1.5 pro successfully at that.

[hasty hornet]

Even without AI it's not hard to read that part wrong tbh

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.

[shell cradle]

I've also used chatgpt to create spreadsheets of information contained in series of pdfs quite successfully (though not as accurate as Gemini). It worked better after some prompt finetuning and doing 1 document at a time than with a series.

[shell cradle]

Right, just reason

[pale latch]

out of context yeah. most law is

[pearl ocean]

SD4 when

[hasty hornet]

SD1.5XL when

feck restrictions

[pearl ocean]

you know I was pretty excited about SD3, but its kinda worse than SDXL

[hasty hornet]

idk, I'm still using 1.5

[shell cradle]

They're establishing the bounds of what they accept in their license, which doesn't necessarily represent the bounds of the law. The tricky bit is they could hide some clause within the license that invalidates entire sections of it.

We'll find out "in a few days", I guess.

[shell cradle]

I'll clarify the documents were in a different language, not English. I've no idea whether that has an effect or not.

[soft iron]

#🆕｜sd3 message

everything seems dead

where are the new models on new architectures?

💀

[trail lion]

did you check civit?

[soft iron]

single model - https://civitai.com/models/435669?modelVersionId=576637

[trail lion]

I cant tell if you're serious

[soft iron]

?

[trail lion]

civit banned all sd3 content, there's a huge banner on the front page and we've been talking about it nonstop since it happened, so yah there wont be any models man

[soft iron]

I'm not talking about SD3, it's dead

I'm talking about lumina and pixart

[trail lion]

oh, didnt pixart get abandoned also?

dont know about lumina

[soft iron]

Lumina has not yet been configured so that it works out of the box in Comfy

[sonic idol]

What's going on with stabilityai? Why the comfy devs quit the place? Seeing posts about how the shittiest model was released for sd3?

[soft iron]

So I’m interested in what the community is doing

absolute silence, usually every week or two I saw news about something new

[pale latch]

signal to noise is off this week. Too many angry people and they're drowning out actual information

[soft iron]

Lumina should most likely now become the main model

the developer of lumina told a lot of interesting things in reddit

[coral lark]

Lumina is more expensive to finetune iirc

Not being able to make local loras will hamper adoption

[soft iron]

there is nothing else, pixart is shit even compared to SD3

[light cipher]

Love from Israel 🇮🇱

[trail lion]

even with SD I dont bother until I start seeing good concept loras, and so far I see no reason to abandon xl

[pale latch]

Lumina is cool but most of the outrage about sd3 is about "censorship" and people are dressing it up as "licence concerns". The censorship is still there on lumina. they would've scrubed the dataset and likely have done safety training on it too.

Also, why upgrade to lumina and not just stay on sdxl? the SDXL vae it uses means it'll only ever improve text comprehension. But that's not what people want. They want tag soup to show buub and big ones

[full lark]

for its 0.6M size its very efficient

[soft iron]

because sdxl is a dead end

[trail lion]

or they want the head in the right spot, or to be able to generate different viewing angles, but whatever, incel rant

[pale latch]

It's all people want. Pony is the defacto model that current user base wants. It provides all they want.

Moving onot a new model makes no sense. new models will be made for new audiences. The current demands have been met

[trail lion]

sdxl is all I want, never used pony...but if something were to come along that was better than XL I'd happily move over

[pale latch]

i don't see a reason to move from sdxl to another model that uses sdxl vae and none of the existing loras and controlnets

[sonic idol]

Good to know Lykon was lying out of his ass trying to cover up how sd3 was garbage

[pale latch]

that's your take away huh?

reflects a lot on you.

[sonic idol]

My takeaway is that SAI is falling apart and the stuff Comfy said about it brings me no hope

[pale latch]

tear

[south sun]

I’ll get you all some signal tomorrow.
Don’t root for our downfall friends, the alternatives are far worse.

[fervent thunder]

yay!

I think

[pale latch]

my tear is for their sad take on things. i'm betting stability's money train is better equipped then people are thinking

[fervent thunder]

I hope they pull out a win

[static cape]

I have not given up on SD3 yet, but it's not looking good.

Looking forward to that signal.

[willow crescent]

base models are perfectly fine

[hasty hornet]

just give us one model with no restrictions and "BEST OUT OF DA BEST" for $ and it's all fine, solves every problem I can think of - community still can do whateer it wants, improve on it, while "DA BEST" could be for lazy people who wants to just pay and go generate stuffs or finetune on a better model for $

well, every problem current situation presents, not older problems that existed before 😄

p2w AI

[fringe vault]

How to create images by sd3

[coral lark]

Yeah, no. If the choice is between models designed to be adversarial to the userbase and actual open source models the alternatives are much better.

[fringe vault]

Please teach how to use sd3

Hello please teach how to use sd3 in this server

[pearl ocean]

comfy ui

[sage reef]

https://tenor.com/view/perfetto-yes-gif-26163077

[fringe vault]

Hiiiiii

[fervent thunder]

can anyone tell me what the difference is between face restore and adetailer?

[empty star]

https://x.com/ostrisai/status/1803083447881113881

[hasty hornet]

I don't think that's same thing as training on 1024x1024 + gonna lack better prompt understanding \ text features

[bleak matrix]

Good morning, everyone! How are we all today?

[vocal leaf]

for me ill stick with sd1.5 and sdxl its more fun for me

+sd3 is slow asf

1024 res on sdxl takes like 5sec or less, in sd3 its 40sec or less

[fervent thunder]

coffee, coffee, coffee

[vocal leaf]

why is sd3 so damn slow

[crimson veldt]

https://arxiv.org/pdf/2406.11831v1

a 10B model hmmm

[crimson veldt]

it needs to run that t5 and is a transformer that is kinda compute intensive

[vocal leaf]

with upscaler it took 3:45

[crimson veldt]

😃 sdxl takes 1 minute for a 1024x1024 gen for me

[fervent thunder]

its so slow because you need moar power!!!