💬｜general-chat

[trim magnet]

1.5 is the old arch of models,most of them are good enough,sdxl are more coherent than 1.5 but they require at least 8gb of vram and good ammount of ram, then theres sd cascade and upcoming SD3 but they are new models so havent been finetuned yet

[true locust]

Are you sure you want a laptop pc ?

[warm junco]

if you want to get sd to work on your 1050, checkout the SD install guide in #🤝｜tech-support

[normal steppe]

hey

i'm new here

[warm junco]

hey

[wispy oasis]

best model available for stable diffusion i can download is "sdxl"? and will a 4060 mobile iwth 8gb ram cut it. with 16 or 32gb system ram

[wispy oasis]

i need it to be fast, i have to make thousands of images

[wispy oasis]

yes

[trim magnet]

u can run it on a 1050 low vram mode,it will run but it wont be fast

[true locust]

The issue with laptops is that they are slower than desktop computers with the same components, and it's really more expensive

[trim magnet]

32gb of ram would be ideal

[carmine herald]

is A1111 still the best tool to use now, or has something new gotten better?

[normal plover]

ComfyUI hands down!

Any hints when V3 will start being available?

[wispy oasis]

can someone link me the cheapest laptop to get for stable diffusion's best model to run reasonably fast

[split osprey]

Is there a way to download the image or video generator for free so i dont have to run it in my browser?

[glossy willow]

comfyui, forge, a1111 but itll be locally in your browser

[noble star]

it is cheaper to run a desktop then remote into it

I run 2 servers and control them from my laptop

[wispy oasis]

what do you mean

[noble star]

you can run a desktop then use remote desktop connection over your network

I wouldnt want to use a laptop running stable diffusion it would be 50000 degrees

[wispy oasis]

i want to use stable but i dont have a good GPU. is it possible and feasible to rent a server that has gpu and use stable? how can I interface with it to generate images from text prompts in bulk? is there an api

[noble star]

there are many online services.. You interface with it in a web browser

[wispy oasis]

can you give example? I also can not be inputting prompts manually in to my browser. i need an api that will let me generate thousands of prompts and return a link/image

[noble star]

I haven't played with any of the api's yet

[wispy oasis]

dalle 3 cost $800 for 20,000 images

i need WAY cheaper

[noble star]

https://platform.stability.ai/pricing

their api is 10$ for 5000 images

[fervent thunder]

what is the estimated time for SD3 public realease date?

[wispy oasis]

thanks

[fervent thunder]

those of us that find just making images boring end up gravitating towards making animations (videos) where we do have to use SD 1.5 with ipadapter and SM, SDXL isn't fast enough for us to get there yet -- best we have is LCM but the amount of time we can render out of those is minimal so we make endless loops or we glue it all together with something like premiere pro / davinci resolve

[winter pike]

id recommend you waiting for stable diffusion 3 to come out if you want prompt adherence

[wispy oasis]

what is prompt adherence

[fervent thunder]

less quality in fp16? let me doubt about that

what does it sound like

if you adhere to a prompt, what are you doing?

[winter pike]

basically its that the ai will follow your prompt accurately, like dalle 3 does most of the time

[fervent thunder]

people used to have to walk hours in the rain/snow/whatever weather to the library to find information, you have google, gpt, gemini in front of you, ???

[wispy oasis]

I am also in a chat titled "General chat about all things Stable!". i know what the two words are and i still don't get why its an issue because why would the image not be what i tell it to be if thats the point

[fervent thunder]

because it's not 2 words the people from stable diffusion made up, they are using clear english

i just used gemini to look up prompt adherence even though i know exactly what it is, it's 100% accurate

no drive to figure it out yourself

[wispy oasis]

you live in your basement and hide from the sun. im not afraid of human interaciton. im in a general chatof humans and you telling me to converse with ai instead get a life

you dleeted your rude messages i see

[fervent thunder]

i do work

my own work.

[wintry chasm]

Prompt adherence is how well an ai adhere to the prompt given for an image.

[wispy oasis]

talking with people instead of ai is lazy?

youre cracked

[fervent thunder]

no you're begging for help

you're not talking, you're asking a question that google can answer you

it's your ignorance / lack of drive

[wispy oasis]

touch grass

im not even armerican. you gotta be pretty stupid to make that assumption

[fervent thunder]

you are definitely not a baller if you cant add 1+1

[wispy oasis]

i got an accountant for that

you're mad sus for your response to asking a question here. take your personal problems out elsewhere weirdo

[fervent thunder]

but that's discord i.g.

[winter pike]

its just a question and its not that serious, i dont get the drama

[fervent thunder]

i hate people with 0 drive and no ambition / full on ignorance

lol

[wispy oasis]

dude talking with other humans is not free work. i dont use chat gpt or gemini. im not asking how to use those this is for stable diffusion

if its general chat about all things stable why are you telling me to go talk to chatgpt? this isnt event chatgpt server

[fervent thunder]

is this guy still confusing talking to someone on discord as talking to humans when you're really just begging for help online

[wispy oasis]

confusing talking to someone on discord as talking to humans

touch grass

im not begging for help its a conversation

[crimson belfry]

What do you need help with

[fervent thunder]

"what is prompt adherence"

Do you guys see this as conversating, or begging discord for help that google can fix

[wispy oasis]

thanks for speaking for me

you are mad sus, i can only imagine you irl. stop getting worked up over normal conversation

[crimson belfry]

Explain what you need help with and I will guide you in the right direction

[wispy oasis]

I need to know why someone is telling me to wait to use prompt adherence on a new version of stable model instead of using stable currently. is it insufficient?

[crimson belfry]

No

The upgrade into SD3 will be compared to Dall-E, Prompt adherence will go beyond anything we have seen from stability

It would be recommended to wait for the update

This is my own observation

[winter pike]

sd3 is better at understanding your prompt than sdxl

[crimson belfry]

I have stability on my X and emad and the nine, I'm keeping up with what's being said.

[winter pike]

but sdxl is still good

[wispy oasis]

this guy keeps deleting his messages like a weirdo

good anyway, useless messages

[crimson belfry]

Of course but sdxl does not do text well

SD3 does words and letters and text

That alone makes sdxl obsolete

[wispy oasis]

heres my idea thetruth

[fervent thunder]

just not gonna chat in here anymore, below average IQ

not 4 me, peace

[wispy oasis]

bye you said bye like 4 times just leave

[crimson belfry]

Bye fam

[wispy oasis]

no one cares

thetruth i want to generate thousands of images based on a text prompt using sdxl. the images should be very similar to eachotehr with slight difference based on the text prompt

[crimson belfry]

All in all SD3 will be better for you to use and it will combine with stable cascade

[wispy oasis]

SD3 is coming out when

[crimson belfry]

https://twitter.com/StabilityAI/status/1760656767237656820?t=itr_r_DSoZXxc_ZHX-XQRQ&s=19

[wispy oasis]

cool

[winter pike]

this is the reason i recommend to wait for SD3, its just too good understanding instructions: https://twitter.com/EMostaque/status/1760668434772156552

[wispy oasis]

nice

[cobalt loom]

hello, can i get some help with installing control net models. i have the download but i don't know what to do after that.

[jovial wraith]

In your extensions folder, there's a controlnet model which should contain a models folder. You place them there

[cobalt loom]

thank you!

[jovial wraith]

You're welcome! :)

[crimson belfry]

Which model is best for fantasy realism

[grand elbow]

Is there a Russian chat room?

[crimson belfry]

I want to try something new but there are so many models lol

I'm not sure @grand elbow

With the power of AI language barriers are a thing of the past

[dense tulip]

is there a function on SD and comfy, that have the midjourney "blend" function?
ipadadapter can use only one photo as reference, but the blend in MJ can use up to 5 pics
so i was wondering is that the same tech behind it.

[grand elbow]

where can I find out here what tags to put when creating a picture?

[burnt kite]

Can SDXL run on AMD by now?

[still hare]

I trained a lora of my face with google colab, it works fantastic. My problem is that most of the shots had down to around my shoulders. What is happening is automatic1111 refuses to generate any clothing other than one that was in some of the photos.

Do I need to retake photos and make sure no clothes are in them? Or maybe it's the text files like "man in grey hoodie taking a selfie".

Or I missed some way of telling the Lora that this is "only the face please"

[cedar salmon]

ide think its more native like SD 2.1 Unclip and Stable Cascade where you can do that out of the box without big comfy spaghetti

[hushed quarry]

Someone could take the current ip-adapter and write some inference code to mix the cross attention activations and you should get that, could try mixing the clip embeds as well pre-unet cross attention. Likely wouldn't get as raw or defined a mix though, more conceptual vs literal

[tranquil elk]

I wonder if SD3 is able to generate celebrities

Cascade is a downgrade with this stuff. 😦 SDXL is still capable of generating some celebs. Rly disappointed with the cascade. Hope SD3 is not excluding celebrities from it's dataset

[trim nymph]

thats not important: even if not, one could easily finetune or insert faces into generations.

[pale latch]

all these companies are taking the taylor swift event that happened seriously. worse than "The Fappening" in many people's perspectives. There is a tsunami of consequences thats going to hit shore soon .

dont expect celebrity capabilities in future generative models from million dollar funds

[pulsar lintel]

I agree that was a huge disappointment

[vestal dew]

anyone have any experience with InstantID? It does a good job on face, but if I add controlnet2 openpose, it mostly ignores the pose. If I try it with inpaint, the controlnets are ignored. It is good for a face, and pretty much only a face/head shot, as far as I can tell.

If I turn the Control Weight way down, say around .15, the result is following the openpose, but the head is tiny.

[vestal dew]

I guess it's just loras or something else trained to get a certain face

[foggy halo]

any idea when sd3 will be out to public?

[foggy halo]

like seriously the blending feature in MJ is phenomenal!! that's what makes it unique and i wish they still offered free trials 🥲

[vestal dew]

it's not unique - it's hosted. Someone else is in control.

MJ can change parameters, settings, and they can change terms/price whenever they want.

[hollow dust]

Is there a way to stop Animatediff from generating looped videos? (I'm using comfyUI)

[tranquil elk]

You can still use midjoirney for free in their niji app

[wispy oasis]

can Intel Core i5-12600KF 3.7 GHz 10-Core Processor run stbale diffusion with a 4070 or should i get Intel Core i7-12700K?

[feral stone]

What happened here? Is it no longer possible to create images?

[foggy halo]

OMG thankyou so so much!

i thought niji was paid as well

[pseudo bough]

anyone got a workflow in comfy or a method in forge to do what SD is now calling "creative upscale" maybe with controlnets or something idk

[jovial wraith]

the bot is down, presumably to prepare ahead of the release of SD3

[wispy cloak]

when release of SD3?

[jovial wraith]

we don't know yet. i'm speculating in around 3 weeks

all we know is that it's coming soon, we've seen some images and we're able to sign up for the waitlist, although noone has gotten access yet afaik

[bronze needle]

I ran out of hours for midjourney does anyone have any suggestions on similar sites or how i can still use mid journey for free

[foggy halo]

GUYS!! does anyone know a free alternative to the "blend" feature of Midjourney? that is as accurate as MJ

[wispy oasis]

is it reasonable to expect to generate 10,000 images in 6 hours using a 4070 and stable diffusion?

[surreal kite]

I don't care much about live celebrities but I've been having fun with cartoon characters in Bing. and I don't want to have to install one of the weird furry fan art trained models just to generate stuff like "Donald Duck in a data center".

[vestal dew]

I say Face ID is superior to Instant ID. Face ID works well with Inpaint, while Instant ID doesn't

[opal hedge]

If you're using XL lightning then this should be very possible

But it also depends on the size of the images and the level of post-processing you have in your workflow

[broken otter]

is There are any difference between the lighting ksampler and normal

[opal hedge]

The ksampler is the same but you have to change cfg and the number of steps

For more information I suggest watching a vid or two on the tube

[broken otter]

👌

thx

[jade mason]

When SD3?

[broken otter]

depends sora i guess

[opal hedge]

Allegedly within two weeks from now but no one knows for sure

[sage ledge]

ya no sirve para generar imagenes?

[paper spire]

great

[rose meteor]

hi

[mild oasis]

How does this app wor 🙂

Kinda new to this

I found clipdrop? Is this using the newest stable disfusion 3?

[grim aspen]

do regularization images make sense for a pixel art style? If yes what kind of images should I use?

[wet kestrel]

hi

[nova zodiac]

Not really - ask in #🔧｜finetune for more responses

[nova zodiac]

Id seen 2-4 weeks on reddit from on of the Stability staffers

[shrewd linden]

I'm pretty new to stable diffusion, been experimenting with WebUI forge and some models, currently mainly using "AldeboBaseXL" but I keep getting bad hands. Is there a guide somewhere to fix hands/get better hand generation with AldeboBaseXL (I suppose this is based on SDXL). I pretty much know how all this works by now, but the hands are still hard and some LORA's I checked seem to only work with 1.5

[nova zodiac]

6 hours is 360 minutes or 21600 seconds which is 2.16s/img. A 4070 should be able to get 15it/s at 512x512 euler a, so if doing 30 step images with no controlnet and no hires fix then yes its possible…

Might create a lot of junk tho

[nova zodiac]

Try a different model if looking for better hands.. i dont rate albedobase at all, there are far better options (juggernaut, dreamshaper lightning, blank canvas xl, aesthetic mix)

[shrewd linden]

Ok I'll try these. Pretty hard to find a decent model for a mix. JuggernautXL I tried already. Others I don't think I have them yet.

[nova zodiac]

Pony diffusion xl also good (requires a slightly different prompt but results are good - you just have to get past the name)

[shrewd linden]

Also, is there a way to 'save' settings when using webui forge? For example, the sampling steps/method etc are alway reset when starting again

[shrewd linden]

Is this the right dreamshaper one? dreamshaperXL_v21TurboDPMSDE

[nova zodiac]

Any of the last couple of dreamshaper ones are fantastic

[nova zodiac]

You can use the png info tab - drag the original image into the tab and the generation metadata will appear

It doesnt always populate the extensions (freeu etc)

[shrewd linden]

ok thx, I find it a bit of a missing feature, would be nice to be able to save like some presets

Learning prompts too. For example, I'm trying juggernautxl to create an image of trump sitting on a large bomb, but I never seem to get what I want. Tips. This is the prompt I use : "caricature of Donald Trump with a giant nose and a smile from ear to ear, windy, sitting on a large bomb, one arm in the air holding a USA flag lora:Caricatures_V2:0.8"

[nova zodiac]

You might be best with some photobashing there. Create seperate images of the components, throw em together in the right order in gimp/krita/photoshop/paint, then save it and do img2img on that

[signal dome]

who all have access to sd3 ?

[pure crystal]

Bros, I am wondering if I can use img2img to perform a eye-contact correction on a video (not a anime)? Many thanks

[tawny scroll]

hi guys!

[deep spade]

sup guys!

[pearl ocean]

SD makes really good images of people, other than that, it’s hard to make to do cool animals or objects, or like sonic or Mario.

Maybe I need a different model

[jovial wraith]

it sounds like what you need is a lora!

[pearl ocean]

😮

[jovial wraith]

would you like a brief explanation?

[pearl ocean]

Sure

[jovial wraith]

lora stands for "low rank adaptation". think of it as a smaller model that you put into your prompt to get specific results. think art styles, characters, that kind of stuff.

[clear dust]

Did someone will suit SD3 like https://github.com/AUTOMATIC1111/stable-diffusion-webui this?

[pastel lynx]

Do we know when SD3 is coming ?

[jovial wraith]

nope

[torn bramble]

have anyone got the access to SD3?

[jovial wraith]

not afaik

[frail wind]

did anyone try sd-forge?? did it eat a lot of ram??? cause mine it will lag after i use a couple of hours

[grizzled palm]

Someone in a YT comment a few hours ago said he got access to SD3. I asked if it was an API or weights, but no response.

[pine fiber]

and u believed them?

[tacit oyster]

guys, can someone help with problem. i'm using SD XL with control net inpaint model and it doesnt work. Canny and other processors are fine, only inpaint fks my brain hard * Error running process: D:\Stable Diffusion\extensions\sd-webui-controlnet\scripts\controlnet.py
Traceback (most recent call last):
File "D:\Stable Diffusion\modules\scripts.py", line 718, in process
script.process(p,script_args)
File "D:\Stable Diffusion\extensions\sd-webui-controlnet\scripts\controlnet.py", line 1143, in process
self.controlnet_hack(p)
File "D:\Stable Diffusion\extensions\sd-webui-controlnet\scripts\controlnet.py", line 1128, in controlnet_hack
self.controlnet_main_entry(p)
File "D:\Stable Diffusion\extensions\sd-webui-controlnet\scripts\controlnet.py", line 894, in controlnet_main_entry
model_net, control_model_type = Script.load_control_model(p, unet, unit.model)
File "D:\Stable Diffusion\extensions\sd-webui-controlnet\scripts\controlnet.py", line 367, in load_control_model
control_model = Script.build_control_model(p, unet, model)
File "D:\Stable Diffusion\extensions\sd-webui-controlnet\scripts\controlnet.py", line 396, in build_control_model
control_model = build_model_by_guess(state_dict, unet, model_path)
File "D:\Stable Diffusion\extensions\sd-webui-controlnet\scripts\controlnet_model_guess.py", line 221, in build_model_by_guess
ksize = int(state_dict['body.0.block2.weight'].shape[2])
KeyError: 'body.0.block2.weight'

i have 0 knowlege about python, 10 years ago i made hello world and dropped 4ever

do u have some ideas plzzzzzz

searched everywhere i could

[pine fiber]

change yaml

[tacit oyster]

what is this and where xD

[pine fiber]

https://github.com/Mikubill/sd-webui-controlnet/issues/672

[trim magnet]

inpaint is buggy when u use it with a1111 webui

better switch to focus or the forge webui

[tacit oyster]

mmm. i see

[tacit oyster]

i saw this one but couldn't find this parametr. (can't stick scrsht)

[glossy drum]

guys please someone help me out to get this type of results
https://www.instagram.com/p/C3cuSRpvGOX/

[tacit oyster]

omg, what i have realised that i haven't conda, i have installed it and miniconda and idk... and still cmd doesnt recognise conda package

[hollow summit]

Is there any controlnet model to use my exact logo (with colors and shape) and generate different kinds of background, themes?

[wispy oasis]

on stability ai API what is differnece between positinve and negative prompt

[nova zodiac]

canny is probably a good bet

[hollow summit]

basically write what you wanted to generate in positive section and write what you don't wanted to see on generated image in negative section

[wispy oasis]

well that doesnt make sense idd have to list 100000000 negative things... is it only for remaking photos that already have something in it and you dont want it anymore

[hollow summit]

Nope, canny just picks the shape not the colors :/

[nova zodiac]

combine it with ipadapter then

or just use krita in post

[hollow summit]

Can I just use ipadapter?

[wispy oasis]

is it recommended to leave negative prompt blank

[nova zodiac]

cant hurt to try

[nova zodiac]

for sdxl yes, for sd1.5 - start with "worst quality, jpeg artefacts, blur, noise, watermark, logo" and build up from there

also, for the model you are using, find the model card on civitai and see what prompts people are using there

[hollow summit]

you can download negative embeddings and use it, some of the SDXL models is better without negatives

[wispy oasis]

thanks folks

[fervent thunder]

when is it comming out? Pls ping me when replyign I don't check it too often

[wispy oasis]

is "steps" more the better for sdxl?

what is recommended steps for sdxl

[hollow summit]

it's up to model you use look at the description of the model

[wispy oasis]

ok and another thing

i generated say 20 images

about 75% of them look the same style while about 25% look different styles

75% look how they should

how do i make sure it keep the same style and my character standing in the same position

[trail lion]

controlnet depth map

for the position, for style, you'll probably need to keep the seed constant

[wispy oasis]

do i just put a random number as the seed and keep it the same each imgae?

[trail lion]

probably better to use one from a generated image that did what you wanted

[wispy oasis]

ok thats weird

i used a seed from a perfect output

and it is giving me images that look nothing like it

same prompt set

the image that i took the seed form generates the same

others dont look like it

[wicked gust]

any luck with the logo regeneration?

I have searching for that for so long

[hollow summit]

Stable diffusion is basiclly try and see different results. There is no exact optimal parameters, of course there is things that will improve quality of your outputs but you have to try different things and make some research then determine your own workflow or use someones workflow. You can try to use img2img section to improve your images, just upscale it to 1.5x or 2x with high denoising (0.70 or something)

[hollow summit]

I tried to use both of the models but it just burnt the logo, I am still searching it.

[wicked gust]

the closer I got I think was with inpaint, and it would update the logo to match the surroundings. Ping me and I can share the research and tests I did so you can save some time

[twin skiff]

Is anyone really good with generating Realistic Ai models?
would pay someone decent to create one for me plz

[wispy oasis]

im having trobule getting stability api to generate images that are the same looking anmimal and portrait shot and style. if i include eyewear and headwear in the prompt description it always generatees perfectly with the matching style, but if the prompt i send doesnt have eyewear or headwear on the animal it generates a way different picture with the wrong pose and colors/style

i dont want all my animals to have hats and glasses just some

[jovial wraith]

models or images?

[twin skiff]

both plz

[winter pike]

it doesnt have a certain release date but i heard "2 weeks" which may change, we're gonna be pinged in #📣｜announcements

[fervent thunder]

thx mate

[wispy oasis]

what channel is best for this

[wispy oasis]

^

[normal ermine]

Why we can't generate images here in stable diffusion now? Any one can explain?

[jovial wraith]

the bot is down, presumably to prepare for the release of SD3. we don't know when it will be back up.

[solid rock]

Any forge users here ? Is controlnet working fully for you guys ?

[jovial wraith]

yup!

it's even preinstalled!

[solid rock]

For some reason, mlsd and depth is not working for me. Sdxl I mean

Could you try on urs ?

I was using the controllllite model

It was working in a1111

But on forge it just shows 'nonetype' attribute error

[jovial wraith]

SDXL isn't really an option on my hardware i'm afraid. it takes like 10 minutes to generate a single image, and even longer with controlnet

[solid rock]

Oh no. Even for forge ?

[jovial wraith]

well on forge it's like 7 and a half minutes

but that's not forge's fault, i just have crappy hardware

[warm junco]

Hey what it/s do you get on forge compared to auto?

[jovial wraith]

me? or the other guy?

[warm junco]

Anyone who compared them ^^

Didn't saw actual comparisons, just on github it claims massive improvement

[jovial wraith]

the improvement is definitely noticeable, especially on lower end hardware. i'll do a comparison rq

[cloud robin]

What's the best way to upscale generated images, to a resolution your computer can't handle generating directly? (Or maybe even generating in parts works as well.)

[jovial wraith]

https://www.youtube.com/watch?v=EmA0RwWv-os&t=195s this still works pretty well i think

[cloud robin]

Thank you. I'll have to try and see.

[jovial wraith]

A1111 gives me roughly 3s/it, forge gives roughly 1.8s/it

on SD1.5, that is

[solid rock]

i think it's similar on my side. but im on 4090 so it's understandable that the gain is minimal. however. loading checkpoints is way faster , and it has many features built in that made me switch over.

Also im able to upscale images for like 3x now, on sdxl. which is impossible on a1111

if anyone is using SDXL on Forge, please try controlnet MLSD and let me know if it works

https://huggingface.co/bdsqlsz/qinglong_controlnet-lllite/tree/main you can download the controllllite model here if you havent already

[warm junco]

Thanks for the quick check!
Yea I assumed its more noticeable on lower end hw

[wispy oasis]

can someone talk about controlnet

how it works, can I use it in a script to enerate many images that are similar to eachother?

[trim nymph]

anyone already invited into sd 3 cooking?

[tidal thorn]

Any one know how to fix " Error occurred when executing KSampler:

Given groups=1, weight of size [2048, 16, 1, 1], expected input[2, 4, 24, 24] to have 16 channels, but got 4 channels instead
"

[shrewd linden]

can anyone explain how I can install "photomaker" on webui forge?

can't seem to find info how to do this? it's "included" in forge, but I think I need to download some extra files to use it

[pine fiber]

I dont believe there was an agreement like that on the application

at the very least people will be able to show that theyve been invited

[crimson belfry]

I need SD3 in my life

[humble onyx]

how is this still down lmao

[worn trail]

will you be able to run SD3 locally though

[ornate iris]

Hey guys, quick question. I finally want to upgrade my GPU and can't really decide between the rtx 4070 ti super and the 4080 super. Is the extra 200€ you pay for the 4080 worth it?

[wispy oasis]

im not sure to asnwer your question but im about to buy 4070

i think its good enough for sd

[ornate iris]

Yeah, probably, but I think its smart to future proof your setup a bit so you dont have to upgrade in the near future.

[wispy oasis]

i hope the new sd will run on a 4070

i will be so disappointed

[crimson belfry]

You will need a quantum computer 🖥

[summer tinsel]

The 4070 ti super I use is pretty good

[pine fiber]

they are presumably releasing multiple models with different sizes

[static cape]

I don't understand why SDXL never got the Tile-Controlnet. It's like the most powerful one... I hope we'll get something similar for SD3.

[pine fiber]

"The Stable Diffusion 3 suite of models currently ranges from 800M to 8B parameters. This approach aims to align with our core values and democratize access, providing users with a variety of options for scalability and quality to best meet their creative needs."

[fervent thunder]

#💬｜general-chat why cant i generate arts now. where is the bots

[proper karma]

SD3 might include some more advanced tools such as what stable cascade does with image variation and the like

which is basically controlnet

[pine fiber]

emad mentioned they should work with sd3 (planned)

[static cape]

But did people get the Cascade features to actually run yet? When I looked the last time we were happy we got Cascade working in Comfy at all.

[fervent thunder]

#💬｜general-chat can someone tell me how to generate arts like before

[wispy oasis]

#1047610792226340935

[sharp spoke]

Maybe someone knows more better bot for image generation?

[fervent thunder]

#💬｜general-chat @wispy oasisno i m not able to .show me how you are doin

[winter pike]

the bot is turned off for now

[fervent thunder]

a quantum computer would be entirely useless

they are entirely useless

[late garnet]

Any idea how i can fix this warning?

warning control could not be applied torch.Size([2, 1280, 10, 7]) torch.Size([2, 1280, 19, 13])

[proper karma]

ah, you are talking about specific implementations of third party projects that have nothing to do with stability ai :v

yes, stable cascade is working. The code is open-source, and the weights are released.

[wispy oasis]

how fast can a 4070 generate 1 image or how many images per minute

on sdxl 1024x1024

[undone garden]

Is the 29th the release date of SD3?

[static cape]

I am talking about all the Cascade "features" they were talking about when they released it.

[static cape]

Where have you read that? I've only heard talk of "a few days up to 2 weeks" so far. But that's probably for the testing Phase, not the full Open Source release.

[grizzled harness]

Is it normal for Comfy to have worse performance than Auto on 1.5 models at higher res?
I'm seeing quite a big difference at 1536x1536 generation.

Comfy 1024x1024:
1.5 - 1.35it/s
XL - 1.49it/s

Auto 1024x1024:
1.5 - 1.54it/s
XL - 1.32it/s

Comfy 1536x1536:
1.5 - 3.17s/it
XL - 1.72s/it

Auto 1536x1536:
1.5 - 2.13s/it
XL - 1.83s/it

[undone garden]

https://twitter.com/EMostaque/status/1762187618513457628

[static cape]

Interesting.

[grizzled palm]

The robot does say 29-02-(20)24. Could be SD3, could be something else. Could be full launch, could be some NDA lift, could be waitlist invites.

[undone garden]

It's even more obvious in the replies. The image he quoted said 22/02 which was the announcement of SD3 so I'd say it's a safe bet the 29th will be the release

[winter pike]

i hope so

[proper karma]

release, or maybe it will be the time the beta starts and they start contacting waitlist

[pine fiber]

"safe bet" they havent even done testing yet

its obviously the waitlist embargo

[topaz drift]

I can't wait to try SD3

[shrewd linden]

Any idea why? I loaded a picture of me and my friends with snowboards, did a pre-process, I see the black/White outline but when trying to generate an image this error appears

same problem when using "lineart", photomaker works fine

Probably stupid question, but I tried using canny and got the following error : TypeError: 'NoneType' object is not iterable

[pine fiber]

post full error

[shrewd linden]

Traceback (most recent call last):
File "E:\webui_forge_cu121_torch21\webui\modules\call_queue.py", line 57, in f
res = list(func(*args, **kwargs))
TypeError: 'NoneType' object is not iterable

(is this what you want?)

so I select "canny" in controltype, preprocessor "canny" and model "control_v11p_sd15_canny"

the checkpoint I'm using is aamXLAnimeeMix_V10

nevermind, found it. was using a 1.5 controlnet model, needed XL version

hm nope, worked a few times, not getting the error again 😦

well, it gives me an error when using "kohya_controllllite_xl_canny_anime.safetensors"

[bright marsh]

Hey guys I have a question, I keep seeing the exact same face all over SD posts of girls/anime, who is she (you know the one, I hope) and where does she come from?

it's getting old, I want to avoid it at all costs in my images

[warm junco]

She is 1girl 😁

No, Seriously she is the "girl, woman" that SD combines of all trained images

[bright marsh]

ooh I see now

[warm junco]

Extension like adetailer that fix faces will also make these face

[bright marsh]

im new so im learning all this, its a bit over my head

[warm junco]

No problem at all ^^

You can avoid such face with a trick. Like giving your subject a name. Like Zoe or something

[bright marsh]

oo ill try that, thanks

never thought to use names

[warm junco]

Yea SD knows celebreties too, for each name it will generate something different, we dont know what SD thinks when generating a Zoe but it will make it look not to generic

Face fixes can still morph it

Upscaling with lower denois also helps preserving the face

And if your in SD a bit, you should take a look at various extensions which you can use to "faceswap"

[bright marsh]

ok, I haven't gotten that far yet. Just learning the basics/interface right now and optimizing. idk if this is the place to ask but is 6it/s bad for a 4090 on sdxl? I'm doing 1024 x 576 then hires fix to 1440 (landscapes) seems to work well but it seems slow? just want to know if 6it/s means I need to optimize

[warm junco]

Hmm I can't say that based on sdxl. Would need the its for 1.5
But check If your using xformers.
Take a look at the bottom of the webui "versions"
If that is xformers N/A your missing out performance

[bright marsh]

ok, I was getting around 25it/s on 1.5 iirc

512x512

[warm junco]

Okay that sounds good. Still check for xformers

[bright marsh]

alr, gotta head out. thx for the info

[warm junco]

Np cya

[buoyant current]

Hi,
I very new to SD and started toying with control net. A few questions about the models: Are t2i adapters better then the v11_sd15 ones?

[lofty garnet]

will you be able to make loras on stable diffusion 3?

[weak cosmos]

Well duh yeah

supposed to

can't wait to get access. This shit will go so hard

[hallow sable]

what happend to SD2?

or 1

[weak cosmos]

Compared to diffusion 3

Eh in quality and composition

They're gonna be classified as the "old" models now tbh

[charred mesa]

There's 3 factors that will dictate if the current models will get replaced by SD3.

• Can there be corn (even if finetuned eventually)?
• Will it run well on most systems?
• Will it run well on +8 GB systems?

if the first two are true, then it will replace SD1.5

[weak cosmos]

Gyat

[charred mesa]

if its only the last, then it will replace SDXL

yes, gyatt

skibidi ohio rizz found in Stable diffusion 5

with fanum tax

[weak cosmos]

My man fr fr

[charred mesa]

on god

[weak cosmos]

No cap

Can't wait to generate toy chica doing the griddy in stable diffusion 3

[charred mesa]

holy shit yes

I forgot about more meme-y stuff

I need to add that to my prompt list

[weak cosmos]

I PRAY it has a wide knowledge base just like google Imagen. Google imagen ai knows like EVERYONE

[charred mesa]

I have a list full of stuff I want to try when I get into the server to try SD3

[weak cosmos]

Please man PLEASE. If not I'm gonna cry

[winter pike]

so emad in reddit said that we are in the tuning phase of SD3 and the "Discord beta/preview access to improve it is imminent", so i asume 29/02 MAY BE the day the bot is gonna be available again in the server

[charred mesa]

I bet it still won't know Tekken Characters or it won't replicate Source FPS games that well lol

okay mabye FPS games are gonna look more convincing

[weak cosmos]

If so I won't use it

I need the knowledge base shit. that's what made Imagen so great (minus everything wrong with it like horrible composition, prompt adherency etc)

[charred mesa]

but finetuning might be viable for 800M

[opal hedge]

SD3 when. I have booba to gen

[charred mesa]

but yeah I want knowledge too

[charred mesa]

if not naked then probably will be possible

if it's only censored as much as SDXL then we're golden

[opal hedge]

I hope not, then we might as well wait a few days for midjourney 7

[charred mesa]

midjourney will do NSWF??

[weak cosmos]

Just to be clear I don't have an expensive ass computer

I have a laptop from 2015

[opal hedge]

No, but if SD3 doesn't do nudity then SD3 is pointless

[weak cosmos]

So there's no tinkering nothing

[weak cosmos]

You're saying stable diffusion 3 needs porn to be useful 💀

Never let bro speak again

[charred mesa]

I have 12GB of vram and I heard fp32 8B model will be 24GB, so idk if fp16 + xformers will be enough 😬😬😬😬😬

I want to have the best DALLE3 quality shit with the MOST knowledge

[weak cosmos]

I don't know what the hell you just said

[charred mesa]

SD3 will have multiple sizes

800Million to 8Billion

[winter pike]

prompt adherence is def not pointless

[charred mesa]

I don't know about the inbetweens

[weak cosmos]

Yeah yeah prompt adherency is cool and all but can it generate sonic and tails doing the griddy in Fortnite with Mario chug juggin with babyshark

If not then nah screw that

[opal hedge]

What else would you use it for then that Midjourney won't be able to do? Midjourney is getting actual features beyond a prompt. SD's only real advantage is a lack of censorship once trained

[charred mesa]

bro people still keep using SD1.5 just because of corn (not me lol)

[weak cosmos]

Actually

Yeah fair point

[charred mesa]

we'll get to know when we get the waitlist and stuff

I also want the paper already!

:(

[weak cosmos]

It's not good in quality with its only upside being uncensorism and customizability

[radiant tusk]

Is there a tutorial that can lip sync and match for animatediff?

[weak cosmos]

Imagine we get disappointed ☹️

[charred mesa]

nah I think it'll be fine, it'll def know much more than SD1.5 and SDXL, it's going to be 8B!

[winter pike]

sd1.5 is also low vram friendly for most user including me, thats another reason

[charred mesa]

800M though...

[charred mesa]

exactly

[weak cosmos]

Hope it's as skibidi as your rizz

[opal hedge]

SDXL is already mindblowing in my opinion. SD3 will be insane

[charred mesa]

I am really scared that 800M is going to be marginally worse

[weak cosmos]

Yeah....

No

Sdxl, no

Sdxl turbo, yeah

[charred mesa]

if you aren't afraid of gore (zombie image), I posted some Super highres pics in #✨｜sdxl that you can zoom into, its really nice for being able to run it offline and all

using juggernautXL 8

[opal hedge]

Why turbo? It has worse quality than sdxl. At least say SDXL lightning

[weak cosmos]

Midjourney try not to suck out people's wallets challenge

[charred mesa]

I did not like turbo

[weak cosmos]

No.

It's on par with Dall e 3

[opal hedge]

That's just not true

[weak cosmos]

And Dall e 3 lighting is extremely cringe

[weak cosmos]

Blud tripping in Ohio rn

Is bro an ohian

[opal hedge]

If you're talking prompt comprehension, image quality, and overall stylistic integrity Dalle 3 is several marathons ahead of sdxl, and even further than sdxl turbo

[weak cosmos]

Blud didn't do his fantum tax to be spouting such TOMFOOLERY

[charred mesa]

idk if we'll get to run SD3 8B with the same amount of VRAM as SDXL then we're golden, if censorship is ONLY as bad as SDXL (not 2.X level lobotomy), we're golden, we are eating really fricking good then

I can generate gore-y zombies just fine with these SDXL models and possibly corn with the other ones, I just haven't tried cause I'm not into AI corn loll

[weak cosmos]

Only thing sdxl turbo impressed me with extremely well is the type shit where your prompt changes as you type

[opal hedge]

Yeah, that was pretty nuts

Did you see the lightning one?

[weak cosmos]

I do not want to be model hunting. If you're telling me to generate a style which means I STILL need to download some goofy model from the internet, I'm done

[opal hedge]

https://fastsdxl.ai/

[weak cosmos]

Back to my sad dall e 3

[charred mesa]

also, it's going to possibly divide users when it comes to training,
imagine going on CivitAI and seeing multiple base models like (as a guess, idk the model sizes): SD3 [800M], SD3 [3B], SD3 [8B]

[weak cosmos]

Blud is high

[charred mesa]

and they are not compatible with eachother

like loras and controlnets

we'd have to train them separately

[opal hedge]

Yeah, it'll be a rough transition

[charred mesa]

hopefully controlnets will be ""cross platform"" if you know what I mean

so it works on all of tem

[opal hedge]

There'll be no point in investing further into SD1.5 with 800m parameter SD3, and SDXL will quickly be overcome by 8b paramter sd3

[opal hedge]

SD3 uses a new type of architecture so sadly I don't think it'll be possible

But I can hope

[weak cosmos]

Stable diffusion 4: 64 trillion parameter

No need for external character models

I would be so happy

[charred mesa]

yeah SDXL will be replaced probably, since people use it cause it's "bleeding edge" and stuff, higher base res, slightly better anatomy as a base model, better prompt comprehension

so SD3 will replace SDXL guaranteed

[opal hedge]

Agreed

[weak cosmos]

Fo surr

[opal hedge]

SD1.5 will take a bit of time because of how much was developed for it

[charred mesa]

they'll get the vram usage down to 8-12GB on the LARGEST SD3 model

[weak cosmos]

Lmao I just realized, Clint hardwood

[charred mesa]

even if it takes fp8

[weak cosmos]

Lmao the name

[opal hedge]

That's why booba is important. I'd be Clint Softwood with Dalle 3

[weak cosmos]

AYYYY

[charred mesa]

I want dalle3 fun without as much censorship, using copyrighted characters and stuff!

It's 1000% be less censored or content limited than DALLE3, so there's that

[weak cosmos]

Dall e 3 when it first came out was the best days of my life, I'm dead serious

It was amazing

[charred mesa]

yessss

[opal hedge]

Dalle 3 would be so good without censorship, but they're so mainstream that AI regulation would drop on them like a hammer

[charred mesa]

I was so f*cking impressed

[charred mesa]

at this point I wonder if they mention safety just to satisfy the public and the investors and it might just be lightly censored loll

I sound like a broken record, but I just hope it's going to be as censored as SDXL

so far it looks good for a ""censored"" model

[winter pike]

100%

[opal hedge]

I think so, it's a really smart way to do it. That way SAI isn't liable for nudes produced of the unsavory and often illegal type

[charred mesa]

FineTuning will take care of the rest

[opal hedge]

It'll be the fault of the person who 'mispurposed' and trained SD3

[charred mesa]

hopefully SD3 will come out around April or May with controlnet and stuff at launch with Comfyui and Forge support

[trail lion]

fav chat of the day 😄

[charred mesa]

if it's March I'd be the happiest man on earth

[opal hedge]

Allegedly SD3 is coming out sometime in March

[charred mesa]

Oh my god

[scarlet moss]

Are you saying it is worse now? I am new to this field

[charred mesa]

already in march

That is amazing

even if at the end of march or even delayed in early April, that' really good news

[opal hedge]

I've heard SAI employees say anywhere from two to four weeks

[charred mesa]

god....

its closer than I thought man

[opal hedge]

So hopefully sooner than later so that people smarter than myself can prepare the GUI for it

[charred mesa]

no we'll get comfyui support on almost day 1

[trail lion]

I've never used dalle or midjourney, still dont care

[charred mesa]

comfy is a stability dev

that's how we got SDXL support before even release

[scarlet moss]

How would you use SD3 when it 'comes out'? I only have experience with the Automatic1111 webui

[winter pike]

emad said SD3 wont take as long as SDXL did, i bet the release date is next month

[opal hedge]

Gigachad Comfy users vs A1111 waiters

[charred mesa]

now its forge lol

cause its kinda objectively better

OHHHHH

and Fooocus will thrive

that gui adds stuff to your prompt, so it will yield much more aesthetic results

SD3 takes your prompts seriously

[opal hedge]

I don't like this, but I can imagine why some people would

[hallow sable]

nodes > text line

[charred mesa]

Dalle3 prompt: a catgirl, anime artstyle
Dalle3 Revised Prompt: Generate an image of a feline-human hybrid, a girl with characteristics of a domestic cat, using the aesthetics commonly found in Japanese animation (anime). She should have kitten-like attributes such as a fluffy tail and soft ears on her head. Her mannerisms should be playful and curious, similar to a young feline. Dress her in casual attire, typical among anime characters, which could include a comfortable t-shirt and a skirt. Her eyes should be large and expressive, another common trait in anime characters, and her hair can be shoulder-length and styled in a cute manner. Remember to portray her as a lively and engaging character.

Example

[weak cosmos]

Ayy thanks man

[charred mesa]

DALLE3 does this

fooocus won't be this advanced maybe, but still

[opal hedge]

Wait, is this real?

[weak cosmos]

Yeah it's not great compared to back then

[charred mesa]

I don't like this either, but if people want ease of use then this is going to be their GUI of choice

[hallow sable]

no this is patric

[opal hedge]

Maybe we'll have LLava running in comfy to do that for us as well

[charred mesa]

yes somehow you can get the revised prompt from the API

heh, the smallest model quantized is already a heavy task

if it uses gguf (cpu with gpu offloading) then sure

[weak cosmos]

Never let bro near a kindergarten

[opal hedge]

Sigh, every day an H100 looks more and more tempting

[charred mesa]

lmao

[weak cosmos]

*Men

[charred mesa]

honestly I was thinking of getting a 3090 sooner or later

24GB would make me happy for the rest of my life lmao

(1-2 years at best)

[opal hedge]

Lol I was gonna say

[charred mesa]

unless AIs are gonna get way more efficient in the future

[trail lion]

we'll all be doing cloud compute at some point, it'll be too big to do local

[opal hedge]

I heard Sora AI uses several hundred H100s to generate even half-decent results

[charred mesa]

then 24GB might stay the gold standard for evne the heaviest tasks

[weak cosmos]

Open ai says a lot of shit

[charred mesa]

I think SD3 8B will be doable with 8-12GB vram soon enough when it comes out

[weak cosmos]

They said the world would end when gpt 69 comes out

[charred mesa]

and for lower vram, there's gonna be the smaller variants of course

in the end, if it all goes south, they will still have their corn generators (SD1.5 finetunes)

[weak cosmos]

Crige

[charred mesa]

so they don't really lose anything lmao

[opal hedge]

They said the higher parameter SD3 will take about the same amount of resources as SDXL

Well, by 'they' I don't mean SAI but people who work with SD to a higher degree of capability

And generally that's the entire appeal of the model so it makes sense

[charred mesa]

*M. Bison gif of him saying yes*

I will get to have all the fun

I hope that the smaller models won't suffer that much though

they'll be finetuned for sure

[opal hedge]

I guess we'll see how much of a difference parameters really make

Logically the difference between 4b and 8b shouldn't be too great

But I have no idea about the finer workings of AI so I could be entirely wrong

[trail lion]

I'd say the difference would 4

[charred mesa]

fr

yeah we'll have to see the paper, they might tell us more about it there

[opal hedge]

In the future we'll also have chips dedicated for AI usage like GROQ

[charred mesa]

except you won't have to buy 300x of them, each costing 20K bucks

[opal hedge]

Apparently each GROQ chip is 100x as efficienct as an H100 for LLMs so I think you're right

[charred mesa]

yeah but it has pitiful amount of RAM in it, but yes

[opal hedge]

Yep

Must be nice to be a bitcoin millionaire right now, casually sitting on an 8x H100 server

[charred mesa]

I would have personally bought one of those 48GB gpus from nvidia

[hallow sable]

I would just wait a few years when you can pick a better card for 500

let everyone bash their head experimenting and post videos and documents on how to do stuff

[charred mesa]

if AMD gets more traction with the ROCM and such, their cards would be more viable thanks to their vram

[opal hedge]

This is the hope

[hallow sable]

what about the open sauce cuda?

[opal hedge]

I haven't heard about that

[charred mesa]

NVIDIA RTX A6000

[hallow sable]

yeah its a thing I saw gamers nexus talk about.

[charred mesa]

ZLUDA too

[hallow sable]

also nivida junk actually combing all their programs together into one single program...

like how AMD is

[charred mesa]

I guess its kinda like Wine, but instead of Linux running Windows apps, its AMD running Nvidia CUDA stuff

[bright marsh]

I figured out why my renders were slow, 25-28it/s, I had freaking windows HAGS enabled this whole time.. getting 35it/s now on the bench. I feel like an idiot lmao.

so has anyone got news if we'll be able to run SD3 on consumer hardware?

[pine fiber]

why are people asking this

read the official announcement

clearly states there are multiple models 800-8b params

[hallow sable]

what is SD 2? and SD 1?

im guessing we are on SD 2???

[pliant jasper]

we are on srtable cascade/sdxl

[fervent thunder]

who the hell is we

i would never install cascade lmao

[pliant jasper]

the entire community?

[fervent thunder]

comfyUI or gtfo

[hallow sable]

768 SD 2.1 sucks... cant get dreambooth to output a working checkpoint

[pliant jasper]

?

cascade is comfy ui

[fervent thunder]

oh

wait what

[pliant jasper]

bro

[fervent thunder]

no its not

[pliant jasper]

im literally using cascade

in comfyui rn

[hallow sable]

might be a automagic 11111 thing tho

[fervent thunder]

thats not comfyUI

[hallow sable]

I have not used comfy ui

[fervent thunder]

u dunno what ur talking about, cascade != comfyUI

cascade was said to actually have ripped comfy off

they went through the code and found thousands of lines ripped

[pliant jasper]

what r u talking about

[fervent thunder]

the cascade developer used to say he used comfy's code

[pliant jasper]

stable cascade???

the AI model???

made by stability???

[fervent thunder]

... no

[hallow sable]

yes?

[wispy oasis]

😂

[pine fiber]

maybe?

[pine fiber]

?????

#1207078178510872636

[wispy oasis]

unglued said yesterday the average IQ of this place was too low for him to chat so he was leaving... he never left

[fervent thunder]

ye i heard the person who made cascade ripped like 1000s lines of code from comfy is what i heard

[pliant jasper]

STABLILITY AI MADE CASCADE

WE ARE NOT TALKING ABOUT YOUR CASCADE

this one: https://huggingface.co/stabilityai/stable-cascade

[fervent thunder]

im not talking about my cascade either, i dont use cascade

[pine fiber]

https://tenor.com/view/peter-griffin-meg-griffin-i-just-wanna-talk-to-him-put-that-gun-away-family-guy-gif-14677957

[pliant jasper]

what is happening rn

[pine fiber]

dude wtf are you TALKING ABOUT!!!!

[pliant jasper]

u can use stable cascade with comfy

[fervent thunder]

yea cause thousands of lines of code was stolen from Comfy

theres a huge fight with the SD guys rn

[pliant jasper]

...

what was stolen?

could i see some links

[fervent thunder]

thousands of lines of ComfyUI code

[pliant jasper]

... for what?

[fervent thunder]

look it up yourself, i dont wanna do your work for you

the SD guys are fighting amongst each other rn tho

comfy is fighting the cascade dev

cascade dev originally said they did use comfy's code

then they turned around and said they didnt

so everyone went and looked, found thousands of identical lines

cascade (do i have the right name?) dev continues to deny

even after we've seen the proof

[pliant jasper]

wym cascade dev

cascade is a model

not a GUI

all i could ifnd was this: https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/169#discussioncomment-8428689

[fervent thunder]

sec, lemme get you the whole scoop

[hallow sable]

imagine using a mac :S

[pearl ocean]

Imagine using a phone

[hallow sable]

yeah, like macs they have a ton of bad games and poor hardware

[pliant jasper]

?

m3 max is probably the best consumer device for LLM inference

[hallow sable]

I really dont care about the hardware of a 3000 USD laptop with horrible OS made for children.

using ARM cpu

[trail lion]

OS shaming ftw

[pearl ocean]

TempleOS is all you really need

[wispy oasis]

lol

[fervent thunder]

I tried a bunch of Stable Diffusion UIs but Automatic1111 is the only one that was both feature-rich and user-friendly

[pliant jasper]

...

m3 max at max specs has 128gb ram.

i'll let you consider that.

[graceful peak]

how do i use hi res fix with img to img

[hallow sable]

ive heard talks about desktop arm cpus coming out and also a 4080 is just gonna blow a mac m3 or whatever out of the water for the same price

ok ddr5 128 gb is like $200 for 6000mhz

[pearl ocean]

I use my 4080 for D2, seems pretty fast.

[pliant jasper]

128gb vram

at 400gb/sec?

[hallow sable]

its not vram... its shared ram...

[pliant jasper]

same thing

also 1/3th the power draw

[hallow sable]

not even close

[pearl ocean]

Is 16GB of VRAM fine?

[pliant jasper]

what u talking about bro

[hallow sable]

yes i have a 4060 ti and its pretty good

[pliant jasper]

the me bandwith is very good

thats higher than your card

obviously it costs more

but you can't get a device w/ 128gb of shared ram that can inference an llm at half the speed of 4090 anywhere else

[trail lion]

in img2img it's just sd_ upscale under scripts, accomplishes more or less the same thing

[pearl ocean]

text2game when

[graceful peak]

its making a bunch of images and merging them as a final result

[fervent thunder]

mk the boys that talked about this sunday at the meet aren't on rn, ill update with the exact info later for ya N8

[pliant jasper]

got it thanks

[fervent thunder]

I just know the SD guys are bickering amongst each other atm, and someone stole thousands of lines of code

i'll point names, and name fingers later

[pliant jasper]

got it

good to know

[warm ibex]

Hi All, I followed the git (https://github.com/camenduru/stable-diffusion-webui-colab/tree/drive?tab=readme-ov-file) that's tried install stable diffusion on the google drive. Did i miss something?

it's no work

There is no gen the Public WebUI URL.

[opal hedge]

Are you talking about the forge drama?

[fervent thunder]

ahh that might be it

i think cascade was said to be decent at one thing but really crap at another, then they rolled into the comfy / some code thief's drama

[pliant jasper]

ah

yes

cascade

isn't the drama

i dont think

[fervent thunder]

ye i dont think it is either, now that im recalling it, i believe they just said cascade is a bit overhyped, but still does some things really well -- it was forge

[opal hedge]

The forge drama was weeks ago and it was a minor disagreement on github. Not the drama you think it is

[fervent thunder]

well to me it's real drama because if you originally say you used comfy's code, then you retract it and say it's all yours, people scrape you code and find thousands of lines in it, then you continue to deny that you use comfy's code

this is the drama i think it is

this is as bad as it gets

funny part is, doubt comfy would care that he used his code if he admitted and gave credit

[pseudo bough]

suno ai released a new version for paid customers v3 alpha its really good now, made this christian pop country song with it https://soundcloud.com/artofficial-753519323/he-was-there

if you have a suno ai sub check it out toggle the version thing to v3 it can handle a much higher lyrics input vs v2

[opal hedge]

I tried suno and it kept adding lyrics to all of my music

I told it specifically 'no vocals' and 'instrumental only'

[pseudo bough]

do the custom lyrics thing leave the lyrics blank and put "style" instrumental in the style

orchestral instrumental, hillbilly instrumental, hip hop instrumental, etc

i never prompt it to write the song i love writing my own music

[restive skiff]

guys

do you think T4 GPU can run SD 3.0?

[shell tendon]

man, now i'm getting tempted to upgrade my PC again... it's looking like the 4090 is botlenecked by a f'n 5950x when training sdxl >_>

if any of you are on a 7950x and a 4090 lemme know...

curious if you get more than 50% utilization of the 4090 which is where i'm hitting the wall (1.5 it/s with 1024x1024 unet only... i know that's fast, but <50% utilization has me feeling greedy)

[hallow sable]

how would a 5950x bottle neck anything

[trim magnet]

didnt know that training used cpu at all

you are the first one here

[hallow sable]

you use that stupid bottleneck website that says a 1st gen epyc cpu is better for a 4060 than a 7700x? because more core = more bebberz

[trim magnet]

or maybe he bought one of those cheap pcie. 2.0 motherboards

[shell tendon]

yup it's the first time i've ever had that issue

inference etc isn't any issue whtasoever

x570e wifi ii mobo

i'm wondering if this is why a lot of people are saying they're not getting the huge gains they expected jumping from a 3090 to a 4090

i'm still getting better than just about anything i've seen anyone report getting (not by a mile or anything - most are saying they're happy getting to 1 it/s and i'm getting 1.5-1.6)

[hallow sable]

it highly depends on the program. always look at single core performance, cache and the number of cores the programs you use need.

[shell tendon]

kohya and onetrainer both appear to pile all the cpu work onto one or two cores

max use for a core during inference is like... 30-40% maybe

with the gpu going 100%

[hallow sable]

AM4 has 25% better single core performance than the AM3 cpus

[shell tendon]

yeah

just wondering if this is actually what's going on or not, hence why i'm wondering if anyone has a 4090 with a 7950x or something with major single core performance gains over the 5950x

if it really is a cpu bottleneck and i can buy my way outta it i'll do it to cut training time in half

[hallow sable]

7600x would be more than enough with 6 cores than am3 with 16, 32 cores

[trim magnet]

enough for what? for training with cpu? lol

[shell tendon]

i've thrown so much at my tech shit in the last year that another $300 for a top of the line cpu is nada

[hallow sable]

I dont see the point in going with more than 8 cores. maybe if you do a lot of FEM in cad or something you can benefit from those cpus

[shell tendon]

i do a lot of stuff that requires very high multicore performance (computational chemistry)

which is the reason i'm not considering swapping in a 5800xd3

[trim magnet]

well if u do that id go for a threadripper instead

[hallow sable]

threadripper is honstly too expensive

and the singlecore preformance is not that good

[shell tendon]

yeah

other thing is i got an oled tv in the basement that's lonely without a gaming system and a recently displaced 3080 12gb card on my shelf

so i have all the excuses i need to buy a couple more parts and have my PC undergo mitosis lol

but yeah, just wanna have some kind of confirmation that an upgrade would actually unlock the rest of the power of the 4090, and that it's not some kind of nvidia driver BS or bug in the trainers that will scale with the cpu

[trim magnet]

no

u prob training the text encoder or have added way too much regularization imgs

[shell tendon]

neither, no reg images, no text encoder

issue is present in two different trainers

[trim magnet]

try to disable or enable cache latents

[shell tendon]

already done, along with many other parameters

again, i'm already getting speeds as good or better than any i've heard of

[hallow sable]

thats crazy that people buy such expensive gpus

1k, 2k

[shell tendon]

if you think about what we spend on dumb bs on a daily basis and how it adds up, not so bad

[trim magnet]

try to ask general awareness he too trained a lot of loras on a 4090 maybe he can give u more optimized settings

[shell tendon]

sounds good thanks for the tip

[pearl ocean]

is 4080 good enouph?

[shell tendon]

you can def do it on a 4080 but you're gonna have to do gradient checkpointing and other stuff if it's sdxl that'll slow you down a bit

i wwas just BARELY able to get a few loras trained on a 3080 12gb

[trim magnet]

yea even a 3060 can train its just slower

[pearl ocean]

how do you know if you got sdxl or just sd?

[shell tendon]

two different main popular versions

check out the model section on civitai

[pearl ocean]

ok

I think I might have the standard

[shell tendon]

what's the model you're using and what's the software

[trim magnet]

if the chekpoint is 2gb its a 1.5 model if its 6gb its sdxl

[shell tendon]

mannn the combination of a lora with a lycoris-ia3 with mixed weighting, trained on the same set, is just amazing

[pearl ocean]

this video
https://youtu.be/onmqbI5XPH8?si=smJ7ZS1NiUok-4dn

[pearl ocean]

oh, also whats the difference?

[trim magnet]

xl is more coherent but is heavier so u need a better gpu to use it

[shell tendon]

if you have a 4080 already you'll have no issues

if you're shopping for a gpu i'd either go for the 4090 or a 4070 super i guess

[pearl ocean]

I got a 4080

does coherent mean better results?

[shell tendon]

not necessarily

but usually

sdxl lacks the power of SD15 controlnet though

[trim magnet]

if u only generate girl portraits u wont notice any difference

its just more accurate but with controlnet u can create anything u want on 1.5

[shell tendon]

well, sdxl is def better with hands and fingers, but it's a problem in both for sure

[pearl ocean]

like a pokemon?

[trim magnet]

yes either a pokemon lora,or controlnet ipadapter,canny,reference etc

[pearl ocean]

😮

[shell tendon]

hop into general chat with images for a really simple example

[rocky flame]

/Tesla

[shell tendon]

Here is the image you requested. #🏞｜general-with-images message

[pearl ocean]

Model F

[stuck sun]

so looks like Supir is a way better upscaler than Magnific

so that'll save a few bucks ha

[shell tendon]

this? https://github.com/Fanghua-Yu/SUPIR

[stuck sun]

yep

[pearl ocean]

they say looking at pretty people can increase life expectancy by 10%

[stuck sun]

https://www.youtube.com/watch?v=PqREA6-bC3w

just redid all the images he upscaled in magnific to see how it compared and in every case Supir absolutely smokes Magnific

[shell tendon]

sweet

i don't watch videos so i'll take your word for it

[stuck sun]

:/

[nova zodiac]

Interesting 🙂

[stuck sun]

cant post images here, so uhh, "trust me" lol

[shell tendon]

post in general-with-images

[stuck sun]

done

[fallow veldt]

I wondered that aswell, but nowadays I'm using SD Ultimate Upscaler and I think it is just that, like CN tiles. I did an upscale with high denoising and it generate just like with the loaded SD XL model, not just the pure upscaler (upscaler generation it is kind of lame by itself, cannot understand really what is it doing, I test it a lot with Topaz upscaler).

I haven't use CN tiles for a while, but I think they work in a pretty similar way, or do similar results.

I use 1024 pixel tiles and it works really good. Only problem is to upscale x4 images with large pure background, the seams were kind of noticeable in some gens. Solution may be upscale by 2 two times.

[digital moon]

[shell tendon]

i'm starting to think it's just not going to happen

[pearl ocean]

Me when I make very realistic cat images

[hallow sable]

what does the control net do?

[lone hawk]

gives you control, duh

[hallow sable]

like birth control? what it do

nvm my bro bing chad just told me

[pearl ocean]

I’m watching the first airbender!

[static cape]

And that's just absolutely sad.

[static cape]

I just don't like the ultimate upscaler. I never get "great" results.

[nova zodiac]

Got released a couple of days ago

Its on huggingface

https://www.reddit.com/r/StableDiffusion/comments/1aypn18/controlnet_lllite_for_sdxl_tile_realistic_come_out/

[static cape]

Oooh... but never got LLlite working. Shouldn't it be plug and play like other CN's?

[fervent thunder]

if you click on the link provided in reddit, download one of the png's and drop the png onto comfy, do you still have issues?

[static cape]

I just saw that the LLlite use a custom Node to be loaded instead of a normal CN Loader Node. I'll definitely check that later!

[trim magnet]

yea it sucks tested already

[zinc imp]

Does anyone here know of a recent and up to date Dreambooth tutorial? The ones I find online seem slightly out of date and I end up meeting not getting what i want from it.

[lusty beacon]

yo my 4090 is fire in stable cascade

[charred mesa]

cool

it works fine with 12GB too

[jovial wraith]

it does not work with 3GB :(

[trail lion]

nice

[bleak matrix]

Good morning, everyone!

How are you all doing?

[trail lion]

not sure

[jovial wraith]

i'm doing fine, what about you?

[gusty oriole]

https://tenor.com/view/monkey-eating-chewing-chew-good-morning-gif-23364266

[bleak matrix]

Just dandy!

What's everyone up to?

[gusty oriole]

Computer is in shop.
No local img gen
Sadge

[bleak matrix]

rip

[jovial wraith]

i'm trying out realism for a change. i got bored of stylized images and pixel art

[trail lion]

checking out that controlnet tile posted above, we'll see

[wispy oasis]

i am making thousands of pictures but i dont have stable diffusion and stability api is too slow so im buying a new pc on the weekend and then i will make thousands of pictures

[jovial wraith]

nice! got any specs in mind?

[fervent thunder]

do some kaggles

[wispy oasis]

4070 (12gb), i7 12600k, 32gb ram

[jovial wraith]

not bad!

[wispy oasis]

pc cost so much money

[fervent thunder]

thats y ur getting really old parts?

2 processor gen's behind

going back in time for the upgrade, dope

[trim magnet]

you should help him out then

[fervent thunder]

i read my own docs and figure out my own stuff

id request everyone do the same

[trim magnet]

whats the matter dont have 1k$ to spare to him?

[jovial wraith]

everyone's acting like 2 generations of cpu is such a big deal. it really isn't imo

[fervent thunder]

i do but i don't have 1k$ i want to spare him

[trim magnet]

suboptimal

[fervent thunder]

do i look like a charity case lol

im not the do it all for you dolly

ima go the same way i came in, alone without help, and without helping anyone else

[trim magnet]

are u even anything here

[fervent thunder]

yea i dont make boring ass images like most of you guys

im into video gens

[trim magnet]

damn u create videos,such a revolutionary creator no one can do that here

[fervent thunder]

im sure a few others here do it, but most of yall are img herders

i'd never scrape the bottom of the barrel like that

haha

who would want images when you can do video, seriously?

make some sense to me and ill listen

[trim magnet]

indeed,it takes a very high iq to generate a video

[alpine wren]

yo guys, how can i generate images using this ai?

[fervent thunder]

to do it on the first day you install comfyUI and ever used SD, you do actually have to be semi smart

some of us program in 15+ langs

people like you, well, probably do ???

[trim magnet]

15 langs? thats insane,he really is the stable diffusioner

[fervent thunder]

no im only 1 month into SD

but this stuff is so easy lmao

i used to write markov chains

in 2010, yall are just feeding off our old code, improving it

[jovial wraith]

wow that's insane. would you like a medal?

[fervent thunder]

bodyweight in gold plz

[wispy oasis]

I want images instead of video so I can make a new cutting edge smart contarct that is a hybrid of erc20 and erc721, cobining tokens and NFTs and creating fractionalized NFTs

[trim magnet]

how much money have u made with NFT's?

[jovial wraith]

NFT's are still a thing?

[wispy oasis]

no

but this will be

its a hybrid protocol

[wispy oasis]

0

i have made small amounts on erc20 but not erc721

[fervent thunder]

i already made a little bit of money from my videos

lol

2 customers irl and ive only done this 1 month

[jovial wraith]

who, your mom and sibling?

[trim magnet]

🗿

[fair pecan]

can you send one here

if they’re public

or smth

[fervent thunder]

its all on my drives, i can share in general-with-images what i make

[fair pecan]

I’d like to see it

[trim magnet]

we are not worthy

to witness such genius

[bleak matrix]

I wish you joy in your journey, and tons of fun in the process!

I made a bunch of watercolor cats this morning!

[wispy oasis]

thanks i already have the prompts pre generated. they each have 5 traits that will be injected in to the prompt like {eyecolor}, {expression}, {background color}, etc

i am not using controlnet or lora i am going to first attempt to generate images that are similar to eachother without

i have crafted a prompt format and using the same seed, it works most of the time on stability api

[bleak matrix]

I love that kind of stuff! It's always fun to play around with these kinds of settings. I really love doing fine details--my kind of jam.

Do you have any particular themes, etc, that you're going for?

[wispy oasis]

All images are supposed to be a portrait shot of the same species of animal, done in a detailed cartoon style. Each image should have random eye colors, faciail expressions, hats, glasses, backgroung colors injected in to the prompt to make each image unique. For my test samples the animal I chose was a Beaver but i dont know if thats what i am going to use yet. i might pick another animal

[bleak matrix]

I approve this message, to be perfectly honest! Animals in hats, are, for certain, the way to go! I love the random eye colors--it's always fun to do that. I've had some ideas in the past about doing pupil shapes and things like that, but I never quite got around to doing that whole thing.

Beavers are pretty cool!

[sonic cosmos]

Hi guys i dont know if i'm correct here but i downloaded controlnet but it cant be seen here

on my page

[meager kayak]

You're not supposed to download it. If you're using Auto1111, you'll need to add it to the extension tab. You can add it to the extension folder manually, but there's no point in doing that.

[sonic cosmos]

but in the extention tab there is no ControlNet only "Controlnet-fastload" Thats what I found so I thought thats the wrong one and downloaded it via url

[meager kayak]

its called sd-webui-controlnet

[sonic cosmos]

whats controlnet fastload? Do i need that?

[meager kayak]

no, you don't

[sonic cosmos]

ahh thank you so much

[blissful furnace]

Hey guys and girls, i'm super really ultra new to stable diffusion and i'm a little lost. I'm watching tons of videos on where to start, sd 101, automatic1111, comfyui, fooocus, and well, did i mentioned i'm new to this hehehe, so, could anyone maybe give some directions on where i could start, which platform to use. My main goal is to learn how and to create an virtual influencer and give life to it. In advance, tyvm for any help. Cheers!

whta?

kkkkkkkkkkk

not that kinda virtual influencer ahhaha

[wispy oasis]

@sudden ruin @bleak matrix look above please

[jovial wraith]

Your username is that and you expect no action be taken against you? Not to mention you're being very unhelpful qnd disrespectfup towards that dude that asked for help

That's exactly what i mean. Disrespectful. Let me tell you, you won't be making friends acting like that

[blissful furnace]

maybe i asked for help on the wrong channel?

[modest star]

But he/she/it is just doing it for the attention LOL

Awe, did it already leave, I didn't get a chance to give my condolences

[modest star]

Have you started with any of the YT videos that are out there for this subject? What you are wanting to do, isn't going to be done with just an install of SD/SDXL etc, you are going to use several different tools to get to your end result.

[full lark]

Nah, this people are just trolls/bots

[blissful furnace]

yeah i've been watching a lot of videos on lora, checkpoints, but still i'm in doubt of what and where to start with?

seriously?

[full lark]

Bruh, where are the mods?

[jovial wraith]

Start by installing Automatic1111 if you're a bit tech savvy. If you're not, install Fooocus

[blissful furnace]

i'm not gonna lie, i tryed to install automatic but failed like a pro hahahaha

[full lark]

Yeah, I would recomend you start with a1111 or with sd forge, I think forge is a bit faster @blissful furnace

[hasty hornet]

have you tried official github instructions?

[blissful furnace]

yup

[hasty hornet]

hm, how did it fail?

[blissful furnace]

yup, installed everything like it said over there and puff there it went my try

got a lot of error messages, so i unistalled everything and got comfyui working

[bleak matrix]

There ya go

[hasty hornet]

welp, good enough

[modest star]

You are getting a lot of responses right now. You can also install a product called Pinokio, which will help you install many of the mentioned products in one click, it will keep you from haivng to do a lot stuff on your own, like python installs, the cuda stuff, etc. I would start there, and practice making images, then move onto vidoe/3d etc.

[hasty hornet]

I reported him to discord too, but it's easy to make new acc anyway

[modest star]

Thank you!!!

[wispy oasis]

thanks sorry for ping

[blissful furnace]

saw something about pinokio yesterday

will try it 🙂

[modest star]

I reported him as well, but his IP needs to be reported too, but a Mod should do that.

[blissful furnace]

yup, same here

[bleak matrix]

Ya can ping me anytime!

I might not be around, but please always feel free to do so!

[gusty oriole]

https://tenor.com/view/kung-fu-panda-gif-24482528

[lusty beacon]

why my images looks bad in sdxl turbo

[jovial wraith]

what's your CFG scale?

[arctic ginkgo]

Has anyone been working with Intel ARC A750?

[lusty beacon]

using this turbo gradio app from pinokio

[gray wolf]

Stable diff 3 wen

I mean is there an ETA or Smth,1 week, a month,etc

[trim nymph]

lets hope the starting shot is the 29.2, for gathering user input

[lone hawk]

hype for Stable Diffusion 3 is fading...

[winter pike]

i bet 29.2 is the day the discord bot is coming back

[trim nymph]

i dont think so. we're about to get an open source model that can compete with midjourney. think about how awesome this is!

[lone hawk]

midjourney is far behind... it's main competition is Dall-E

and Sora

[turbid bay]

hi guys 🙂

I've been experimenting with Infinite Zoom in A1111 and I wish there were an option to infinitely zoom into the initial image. Is there a plugin like that?

[gray wolf]

no ETA or semi public access...

[turbid bay]

It does have the "Zoom-in" mode, but that only reverses the generated video.

[gentle plume]

Hello, for several weeks it is no longer possible to access stable streaming bots via Discord to create images or videos for free. Does anyone know if it's possible to access stable diffusion for free and how? Thank you very much:)

[cerulean robin]

hey guys

I need help with something. which channel is best suited for the same?