💬｜general-chat

[loud solar]

Idea is ok for all those who don't want to be independent ...

[opal hedge]

Actually it's the middlemen

Buy gold ore from miners, sell in bulk to refiners/jewelers

[teal pagoda]

I kinda hate this idea.

An open-source software is 100% better

or even like Topaz AI

[trail lion]

agree but can SD become like a blender or a linux, with so much front end capital needed for training

maybe that's why emad wants to do decentralized, swarm fashion

[teal pagoda]

decentralized is a very good idea

this way, they can't monopolize it, but it's already too late

companies like OpenAI are already doing that

[trail lion]

is openai opensourced? I thought it wasnt

ok, forge was stupid easy to setup, took like 5 git commands, not sure why it took me this long to try it

[fleet portal]

Hi new here

[teal pagoda]

no, I meant they monopolized

[trail lion]

well, there's already enough out there for someone to pick up where sdxl left off, if we dont see more from SAI

doesnt have to be the leader, because like you said, there will always be those who will just gravitate toward the open solution

[teal pagoda]

that's right, but who would take this to the next level remaining open-source at the same time? :))

[loud solar]

https://sakana.ai/evolutionary-model-merge/

[lavish lake]

I believe they are committed to releasing SD3 as open model, including weights. it was stated recently by the new interim CEO, after that we don't know... https://x.com/chrlaf/status/1772226646365397311?s=20

[whole orchid]

who's this?

who's the owner of sakana ai?

[loud solar]

Japanese newcomer ...

But I like the idea ...

[whole orchid]

yeah its more inspired by nature

its google form says applicant need to be based in japan

[loud solar]

An A.I. model chooser could be fine, too. You enter a prompt and it loads the best model for it. AFAIK there's already something like that out there ...

[teal pagoda]

After that it's "good luck have fun, you're all on your own!" :))

[charred mesa]

I mean I'd be okay with that lol

the sketch controlnet looks good

so I suppose the other controlnets are gonna be fire as well

[narrow kernel]

Yes and no. For now, they need to lock down on these newer models until countries can pass laws and regulations on what they can and can't be used for. There's a lot of awesome stuff they can be used for, but the amount of potentially bad stuff almost outweighs it. Right now, AI in general is almost as big of a wild wild west was when the internet started taking off.

Another example like this would be drones. It didn't take long for laws to be enacted that require you to register them and if you're caught operating one without it, you can be fined

(of the ones that go beyond a certain height or w/e it is)

[sick garden]

Any new rumors on the topic of potential release date of SD3?

[trail lion]

I'm anti that kind of regulation. Using your own example, that would have been like restricting internet use until controls could be put in place. The way it played out was best. Pass the laws about what users can and cant do, but dont restrict first. Let the technology blossom.

[amber bloom]

The internet was much better back in the wild west days than it is nowadays.

[hasty badge]

Who is this Thibaud dude that he can commit to "training complementary models, or developing additional tools like ControlNet, IPAdapter, AnimateDiff.". Always thought it were different groups doing that, seems there's a single entity doing more for opensource imagegen than SAI, they surely could just train SD3 themselves! (yeah, that's sarcastic, what a load of fluff is that tweet)

[gleaming topaz]

Does anyone know who to contact at stability for buying credits at scale?

[honest spear]

not much, but we have a window of release between 4 to 6 weeks from the interim CEO of Stability

[pseudo nacelle]

https://civitai.com/models/339604/how-to-generate-multiple-different-characters-mix-characters-andor-minimize-color-contamination-or-regional-prompt-adetailer-and-inpaint-or-my-workflow

have anyone tried this? :v

[mint token]

source?, he said it will open but never gave a window as far as i've read

[honest spear]

https://twitter.com/chrlaf/status/1772228848387522728?t=vqMeYQQCg0uTTpVw-ZfLpA&s=19

[sterile gust]

I have made peace with the fact that the search for an "ethical AI" has condemned SD3 and as a result it might never come out.

[narrow kernel]

No, it really wasn't best and those that romanticize it are usually anti-authority types that don't want to be told what they can or can't do with something. We get it, you want to have your cake and eat it... You want to have the latest and greatest model so you can keep secretly making your even more realistic waifus with your favorite actress' faces on them lol

[twilit solar]

Is SD done for?

[trail lion]

fair enough, we can disagree on the politics. if I have the chance to vote my conscience, it will always be the least amount of regulation

[static cape]

So May if we are lucky... but if we get API access before that (Discord Bot) - that would already be something and I'd make sure I give my input for better hands etc.

We should try to make the SD3 Base Model the best it can be.

[sterile gust]

Probably

[charred mesa]

[trim magnet]

why is blud so pressed about ppl maing waifus

[trail lion]

who?

ppl should be able to generate what they want. my only personal gripe is really just that portals like civit should do a better job with categorization, so that content can be separate when you're looking for a model. same with nsfw, their tagging is better, but it's still lacking

[sick garden]

That's not so bad. I can wait :)) Thanks.

[stone latch]

If they're talking about the same person, someone essentially said that research is not driven by people generating waifus

And I fully agree - just like how people can draw whatever they want, they can generate whatever they want.

To play the devil's advocate it can be exhausting to constantly see suggestive images of women fill 99% of civitai, but for different reasons than "waifus bad"

[calm leaf]

Hello everyone 👋

I would like to create this kind of video but I don't know how to go about it.

https://www.youtube.com/watch?v=1sN6U5dV1Os

How can I achieve this, how did he manage to create the background animation?

I know that you need 2 controlnets including openpose for subject pose, and a depth sensor for subject depth.

But for the rest, I'm lost.

Do you have any idea how to achieve this?

Thank you in advance.

[storm shard]

https://huggingface.co/thibaud

[hasty badge]

ok, i stand corrected, they really are a pretty big contributor 🙂

[narrow kernel]

99% of people using SD to make their waifus don't contribute a single penny toward the actual research and development of better models... Actually, I'd estimate it to be more like 1:10,000 that actually contribute monetarily

[trail lion]

yah there should be some kind of reward program for donors

encourage contributions

granted that ratio will always be that, it's the nature of the beast

but imagine if SD3 was early release for donors

[teal pagoda]

Yea, I hate the government regulations. Most of the time they introduce these regulations only for their benefit anyway. The only things which should be banned are the images with celebrities (NSFW) and the images with kids (for pedophil3s).

[crude notch]

current sd3 testing is for researchers as of now

[teal pagoda]

The biggest fault's here if the one of Civitai for real.

[stone latch]

I was referring to heteronormativity more so than Civitai being the problem, but that's a can of worms.

[stone latch]

But this means everyone who generates images (irrespective of their nature) must feel guilty and/or useless if they don't pledge their support for the sake of aiding research and development.

[trail lion]

guilt or lack thereof is between everyone and their conscience, but at the moment, there's really not a strong message to donate, so hardly can blame the users can we?

and certainly only a small fraction are developers, and only a fraction of those understand AI programming

[fast warren]

guys i dont remember, i used to have an extension or whatever that when you were typing it was displaying the correct tags that corresponds but since i reinstalled sd i cant find the name

[true canopy]

what is it that drags a bunch of weirdos in here ....

[ivory merlin]

Hi everyone, I'm running an experiment for school and would appreciate if you could fill this form out (completely anonymous if you want it to be)
https://forms.gle/DqMih6Z9wCmTRvZN8

[wary hazel]

Does text2video even work at all

[astral goblet]

animatediff extension in a1111. lots of nodes for it in comfyui

[wary hazel]

Thanks for replying @astral goblet

[astral goblet]

#🏞｜general-with-images message here's a weird one i baked off yesterday. took 6 min

[wary hazel]

Looks great man I wish I knew more about open sora ai @astral goblet

[astral goblet]

lol "open". its not open

they're going to keep that one locked away

[wary hazel]

Damn any way to get close to that @astral goblet

[true canopy]

close to what? open ai sora?

[wary hazel]

Yeah lol

[true canopy]

its not open to the public, and it wont be for quite awhile

[astral goblet]

creating pre animations in a 3d program. a lot of experience and skill for editing.

[true canopy]

actually im not sure its ever going to be open in any public way?, maybe someone correct me on that tho

[wary hazel]

I saw that today imA look into it @astral goblet

[trim nymph]

100% will be

[astral goblet]

sora likely needs 100s of gb of vram to execute. I wouldn't set hopes on something like that running locally

[true canopy]

im not sure 100s will even suffice

[trim nymph]

just wait till open source alternatives catch up

[astral goblet]

and each render takes a long time. 1min video renders for an hour i heard. even with their massive compute

[true canopy]

if thats true its really far away from normal people

[astral goblet]

They'll sell it to studios with hollywood budgets

[true canopy]

yeah in a few years prob

[astral goblet]

not the public api

[true canopy]

im sure they could do some kinda public api in a few years, but i wonder what that would cost, just the electricity to support that service

[wary hazel]

I’m currently looking up how to work animatediff @astral goblet

[true canopy]

somsone on reddit is doing prompts on sd3 btw, if anyone wanna try

[wary hazel]

Interested

[true canopy]

https://www.reddit.com/r/StableDiffusion/comments/1bnjm3i/stable_diffusion_3/

theres a few images as well

[molten sage]

I got 5 answers "right" 🙌

[split kestrel]

I was up way too late learning

[true canopy]

seems like more people have gotten access to sd3

[astral goblet]

none for me. i dont expect one thoughh. signed up for the list but i'm not a popular model maker

[split kestrel]

is SD3 supposed to be a vast improvement?

[ivory merlin]

Yay! You are currently the record holder for the most!

[charred mesa]

In prompt adherence, yes, a lot

DALLE3 level of prompt adherence, quality wise, for 2B and 800M model probably not much

[true canopy]

and it rly shows, especially from that guy on reddit, hes posting images with prompts now

[charred mesa]

yes exactly

I asked one myself, it looks amazing

[true canopy]

it rly does, for being this early

[charred mesa]

we need more people like this from the beta

when I get the beta I'll do a post similar to this

if the stability team allows it

I know how everyone wants to try it just as I would

[split kestrel]

I assume it requires a new model set/

?

[charred mesa]

if you mean loras and controlnets? yes

Controlnets will come officially, by default, so you won't have to wait for those

[split kestrel]

Im still noob, so my bad if some of my questions are misunderstandings

[charred mesa]

nah its okay

[true canopy]

peopel are spamming him a bit now haha

[charred mesa]

I've been using stable diffusion since 2022 august and there's still some crap I don't understand

[charred mesa]

exactly lol, this is what I would expect if I were to do one

but I would be happy to help

gotta give something back to the community

[split kestrel]

Almost all of my work and knowledge is less than 4 days old lol

[pallid otter]

Heya, any idea when bots will be working again? ty

[split kestrel]

I spent like… 24 hours learning what steps do

And even now… it still throws me through a loop

[astral goblet]

the t5 is what helps the prompt adhearance a lot. and now that we have lavi-bridge, which works, we got t5 on sd15 models!

I tested the library myself, but i have no idea how to transplant it as a node or extension. i talked to a couple of the devs on projects that could implement it easily with diffusers, but they argued with me about its value

[tawdry plaza]

just finsihed a little desktop application for the new(ish) CORE API 😄

https://imgur.com/H4PmCXo

[astral goblet]

everyone sleeping on lavi-bridge

[charred mesa]

cause im waiting for comfy implementation + SDXL support

[split kestrel]

I assume it wont be long to see a SD4 or a set that runs on SD3 that is technically a version of sorta

Sora

If prompt adherence is the focus…
Is there a leap in graphical performance?

[astral goblet]

it should be easy to implement into comfy. the code is straight forware and from what i can tell, not doing anything so out of spec that it wouldn't work.

it's just diffusers code, so it won't copy paste. but the projects that do use diffusers, not even those devs will look at it

[charred mesa]

we'd have a LOOOT of fun with T5 + SDXL

[astral goblet]

t5 + sd15 then refine with sdxl

all the parts are there.

damnit. i know the pieces fit cause i watched them fall away

[split kestrel]

What is T5?

[astral goblet]

text encoder , like clip, but a lot more parameters

[split kestrel]

Ok

[astral goblet]

its what allows sd3 to have huge contexts

[split kestrel]

And SDXL is better than 1.5 ?

Or just… different tool different job?

[astral goblet]

more parameters and higher base resolution

[true canopy]

what is the base res on sdxl? 1024?

[split kestrel]

That part I know

Yes

1024x1024

[astral goblet]

yeah, but also the base model was trained with bucketing. so the base resolution is like a list of megapixel resolutions

[split kestrel]

(I think)

[true canopy]

alright, i see

[astral goblet]

https://www.reddit.com/r/StableDiffusion/comments/15c3rf6/sdxl_resolution_cheat_sheet/ all these aspect ratios got juice

[split kestrel]

I need to start running an Xcel sheet for this stuff

also, thanks for that

[astral goblet]

notes are good to take

[split kestrel]

I think I understand the checkpoint and CFG / STEPS a bit more.

[astral goblet]

i'm not spreadsheet level organized. more sticky notes saved in various places. it helps though

[split kestrel]

I got basics of upscaling and output

[astral goblet]

once in a while i circle back to this video, to ground myself and see if i understand anymore of his jargon. https://www.youtube.com/watch?v=1CIpzeNxIhU

[split kestrel]

Ill check that out 🙂

There are a few points in the rats nest I still need to learn.

All the things in the sampelr

[astral goblet]

https://stable-diffusion-art.com/glossary/ there are a few glossaries out there. i like this site's. they actually have a bunch of solid guides

[split kestrel]

Ooooo

Animate diff does not use sdxl?

Or it does… and it’s bad?

[astral goblet]

you need a motion module for sdxl to use it. there is one and its very beta

[split kestrel]

I was trying when I didn’t know what I was doing and got garbage

[astral goblet]

a lot more parameters means training is a lot more difficult

same reason why some controlnets don't work as well with sdxl. like openpose or tile

[split kestrel]

But animate is better for animations due to the controlled environment, just forced to used 1.5?

Until that levels up?

[astral goblet]

researchers are working on it!

I like to do experiments with sdxl animatediff using input guidance frames.

higher resolutions need more vram though of course

[split kestrel]

I really wish you worked in comfy lol

I would love to see some of these flows to better understand the process

But I’m getting there

[astral goblet]

i like consistent uis ❤️

[split kestrel]

I hear you with that!!!

I just like to understand everything under the hood before I let stuff take control

[astral goblet]

i've learned comfy and load it up now and then to try new stuff with it. like say , stable cascade. the most cutting edge image model we've got access to

no support in auto1111 last i checked.

[charred mesa]

what about forge

[split kestrel]

Stable cascade?

You have a link to learn ?

[charred mesa]

also stable cascade is kinda mid no matter what I try, I hope to see someone tell me what I should try

[astral goblet]

forge i don't think supports it yet either. After comfydev took a dump on illy i've seen a lot less updates on forge and foooocus

[split kestrel]

Stable swarm is new… looks interesting

Buggy from what folks have said

[astral goblet]

stable swarm been around for a while. it's a front end for comfy ui

[split kestrel]

Yep

Didn’t dive in too deep yet

So many ways to get distracted

[astral goblet]

i'd use it more but i'm always debugging comfyui node graphsh in it anyways

[split kestrel]

Is there a chosen sampler setting of champions? lol if that makes sense?

Scratch - reading it now

[astral goblet]

the sampler debate. ahh yeah.

not really. i stopped careing so much about samplers. more about avoiding certain ones.

find one you like and stick to it. they're not that big of a deal

[split kestrel]

Got it!

I’m sure some have some methods that are better than others

[astral goblet]

at some point in the future, if you're upscaling img2img stuff and playing with fine grain photo details, experiment and worry about samplers then, you know?

for most image generations the difference is negligeable.

[nova zodiac]

You got a link to that??

[astral goblet]

Turbo models and LCM need specific samplers

[split kestrel]

Sooooooo much to learn

[astral goblet]

to what? stable cascade?

https://comfyanonymous.github.io/ComfyUI_examples/stable_cascade/

[charred mesa]

no I think the dumping part lol

https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/169#discussioncomment-8428689

[astral goblet]

oh the drama lol. i'm not going to dig that up. it was in illy's repo discussion

you got it

[split kestrel]

That’s for the link to cascade. I’ll look at that next

It’s just a ui?

[astral goblet]

no it's a base model

#1207078178510872636 is where people post stuff they generate with it

[split kestrel]

Ohhh ok

[astral goblet]

its the newest image model they've released. comes from the wurstchen model guys

i think it's basically their version 3, just renamed

[split kestrel]

Goes in the checkpoint?

[astral goblet]

i got it in /cascade/ inside my checkpoints folder. just cause its a couple models that act as one

[split kestrel]

Interesting

So, can you reduce the noise level to .5 and the step count to 60 and get a cleaner image than 1 / 30. ?

[astral goblet]

i think prompting would lead to a cleaner image than renoising tricks like that tbh, but some people have found gains with that kind of work

[split kestrel]

Just made me wonder if you are more likely to follow prompt by increasing the steps and reducing the amount of stuff between each step

I’m a nerd… so I deep dive into this crap

[astral goblet]

prompt comprehension is really limited by the text encoder layers unfortunately

[split kestrel]

Got it

So I’m reading too deep into this

[astral goblet]

which is where things like t5 come in. sd3 will use it but optionally

[astral goblet]

take everything with a grain of salt. a lot of people are just bullshit weaving. i find myself doing it sometimes too.

[opal gorge]

trying to generate an ai image from depth map in txt2image section + controlnet but after inserting my image as reference then generating it gives me a whole another image and nothing from reference why is that

[split kestrel]

CFG?

[astral goblet]

make sure you've got the controlnet model selected too

sdxl models need sdxl depth control and sd15 need sd15 depth control

[opal gorge]

u right i dont have depth model in controlnet but which one should i download from

[astral goblet]

https://civitai.com/models/38784/controlnet-11-models these are all the sd15 models

https://civitai.com/models/136070?modelVersionId=267507 these are all the sdxl ones

[split kestrel]

Karas seems to be the best on paper in that zone

[opal gorge]

how to import these into controlnet?

[astral goblet]

inside the controlent extension folder, there's /models folder. put it in there and then in the UI, click the little refresh button

[opal gorge]

yeah i already tried that but the options are not shown its only 'controlnet11Models_tileE [e47b23a8]' shown

[astral goblet]

i can't really troubleshoot your folders from here sorry. it'll work if you do what i say though.

[opal gorge]

holdon u mean the controlnet folder from my pc or on site?

[astral goblet]

i've never said anything about using a site? so i don't know. i assume you're using a111 webui on a local machine

this is why i dont like to hold people's hands. i offer insight instead. i knew you didn't have the controlnet model downloaded yet because it's so common of a problem. some popular guide out there is glossing over that part or something.

theres always something they're not telling you

[opal gorge]

you not forced to help me im just asking there in case somebody wants to guide me in the tutorial am watching nobody said anything about downloading the model

[astral goblet]

i bet the youtuber left those details out specifically and said they're in another video, causing you to go watch another one of their videos for more monetization traffic

i'm really bitter about youtubers

[split kestrel]

As an I

YouTubers are not honest about a workflow and often do tricks to ensure they don’t fully work

Drives me up the wall.

Comedy is that you need to have an understanding of what they are doing almost better than what they are showing to get around what they are not showing

[astral goblet]

when the 3d comfy workflows came out, i saw people straight up faking results. they'd sketch the shittiest car, and show a video of comfyui generating out a professionally made model of a sports car.

of course they'd have a patreon too

stay frosty out there

[split kestrel]

lol

I’ve seen a few of that stuff with animate

Maddening

[astral goblet]

enshitification isn't limited to the corporate domains

[split kestrel]

Part of the solution is listening for an Ai voice over

If they don’t build it from scratch, and run it… I don’t watch it

DPM adaptive looks fun

Sucks, cuz I can’t run these tests at work…

[astral goblet]

i think dpm adaptive is one to avoid.

[split kestrel]

It looks like it’s deceiving charts

[astral goblet]

on automatic1111 you can set up an xy grid. different prompts over a series of settings. helps to see sampler differences.

[split kestrel]

I like uni-pc

[astral goblet]

its a stock script in a1. i think you need a few custom nodes and specific workflows for it in comfyui

unipc is nice. sometimes doesn't play well with extras. i found compatibility issues in the past

[split kestrel]

I run into stuff now and again, but it’s “latest and greatest” lol

Euler looks to be just fine as well

[astral goblet]

eular a is my go to for animatediff'ing. dpm++ sde 2m karass, whatever its called. i typed that from memory holy shit. did i get it right?

[split kestrel]

Sure did

Need

[astral goblet]

hurrray

[split kestrel]

Nerd

lol

[astral goblet]

yeah but as we've all seen the movies, the nerd eventually gets to grape the girl. i mean get the girl.

[split kestrel]

Grape all you want

[loud solar]

To much blabla ... learn! 😛

[split kestrel]

I am son!

[astral goblet]

one of hollywoods greatest speeches https://www.youtube.com/watch?v=5mA_tzJije0

[split kestrel]

You calm your ti*s

I might have to get a 50 series

Provided invidia decides to make connectors that can stand more than being disconnected 4x

[nova zodiac]

Ah right, yeah that drama

[astral goblet]

the power connector issue was over blown by youtuber hype artists

[nova zodiac]

Wont rehash that

[astral goblet]

it happened like 1/1000 units

[junior pulsar]

$20 if someone can guide me through the process of downloading / using stable diffusion with amd gpu

[astral goblet]

a lot of them were from people just yanking on them by the wire and damaging the connectors

[split kestrel]

It happened a lot. I have personal insite on that

[loud solar]

Look at youtube ...

[junior pulsar]

i have

[nova zodiac]

Go to #🤝｜tech-support and look in the pinned posts for a guide

[junior pulsar]

theres always some other error they dont explain

[astral goblet]

its an avoidable issue. a lot of people are stupid af though. no offense intended to stupid people. they can't help it

[split kestrel]

Most… MOST do not unplug the gpu a lot

[junior pulsar]

ive checked a bunch of tutorials and theres always some error and im not a coder

im not a computer geek i dont know much about coding

[split kestrel]

NOCODE*

lol

[junior pulsar]

yea but theres some sort of error everytime

[loud solar]

Awww... sorry for you!

[junior pulsar]

even with the "no code"

[nova zodiac]

Its not coding, but coders are used to seeing that sort of error

[junior pulsar]

well i dont need you to feel sorry for me thats why im offering $20

[astral goblet]

when you're in a hospital and the nurses are like "HES CODING!" and all the doctors rush in to stop it. thats how you know coding is bad

[split kestrel]

It’s open source knowledge

[loud solar]

Good idea ...

[junior pulsar]

yea youre so cool and smart and im dumb, okay, still offering the same thing

[nova zodiac]

Post your error in tech support and ping @warm junco

[split kestrel]

No on is ragging on ya friend.

[astral goblet]

i'm just a picture that i'm holding of someone who is cool. pretty dumb too usually

[split kestrel]

This isn’t easy…

[loud solar]

We all have those problems sometimes ... and it's always hard work to make it run again 🙂

[nova zodiac]

Sd isnt easy, and amd doubly so

[split kestrel]

I’m at baby step level. So….

You might be smarter than me.

[junior pulsar]

lol everyone starts feeling bad for me lmaooo

[astral goblet]

naw we relating

been there

[split kestrel]

Well, stop sucking and get awesome!

lol

[astral goblet]

git gud nub

[junior pulsar]

you guys know the absolute best way to face swap videos

[split kestrel]

As we say in skydiving…. Two things will happen. You will leave the plane, you will land.

[junior pulsar]

akool is expensive and im sure there are better ways

so im trying roop that will be ran through sd

[astral goblet]

honestly, best advice i can offer for people catching python errors that have no ptyhon idea or care to learn about python code...

use a package manager. There is a great one called "Stability Matrix" and it'll install all these different ui's and manage their python environments for you

[split kestrel]

If python is broken, scrub it… find a different thing

[astral goblet]

roop is old news. that's just pasting the transformed face on top of the finished pic. there are IP-adapters now (image prompt) that swaps the face in at each step, so that the finished product is more naturalized to the rest of the pic

[split kestrel]

I am under no circumstance gonna fight python code lol

[astral goblet]

heheuhehe scrub the python. /butthead

[slate dust]

Visions of Chaos ftw

[loud solar]

Don't use the latest python version ....

[junior pulsar]

i want to offer you money to help me solve my problems, but not sure how much is the right amount because it may be harder or easier than i think. and also dont know how rich you are lmao

[loud solar]

3.10.10 should be better

[astral goblet]

honestly, i'm pretty aprehensive about taking money online. you'll just charge it back (is what i figure)

[split kestrel]

Could be like me… and read 🙂

[warm junco]

Come to #🤝｜tech-support and tell me what's your GPU

[junior pulsar]

what is the best way to pay online for me not to scam you

we can facetime

we can literally do aynthing lmao

[astral goblet]

https://github.com/LykosAI/StabilityMatrix/releases/download/v2.9.2/StabilityMatrix-win-x64.zip here's the stability matrix download

[split kestrel]

Ai, it’s a trap!

[astral goblet]

https://github.com/LykosAI/StabilityMatrix/blob/main/README.md here's the readme

i'm a big believer in allowing people to fail their way to success 😉

[split kestrel]

Heun looks good

[junior pulsar]

here's basically what i need done from anyone, but been trying to figure this out for a while and figured i should join one of these groups and ask some smart people; i need a consistent and efficient way to face swap loads of content every day, and also have the most realistic results possible. at very minimum i will need about 35 videos and 35 photos a day. and i want it to be as easy / automated as possible. preferably would like it where our model can just drop the content in somewhere, and it automatically face swaps for us. i know this may be hard, but not sure, because some things I think would be hard to code / automate are apparently easy, and vise versa.

[astral goblet]

ooof. goals.

[split kestrel]

Sounds like you’ve got your work cut out

We all have our missions 😉

[junior pulsar]

i have a way to finnesse free trials from akool, and also have a way to get rid of the watermarks. but its time consumiing, probably not sustainable, and there may be a way to get better results out there

[split kestrel]

Deep fake

[junior pulsar]

yea. well basically what im setting up is modeling services for new OFM agencies. take one model and essentially turn her into 10 with 10 different faces. I have spent the past six months developing a website (hired a developer), and i have a whole plan on this already. It's not a new idea to me. But recerntly I wanted to see if there are places for better results than Akool.

[split kestrel]

I’m not your guy. I’m literally scrubbing through sampler settings. And going ground up.

[junior pulsar]

where you think i should go

where's the nerdiest of geeks that live and die this stuff

[split kestrel]

I think you should learn what your doing 😉

Here…

[junior pulsar]

thats not the most efficient way

tho

[shell tendon]

easy with comfyui

[junior pulsar]

wym

[shell tendon]

got a custom method i made

[split kestrel]

If you don’t wanna learn what you are doing you are failing forward 😉

Cuz you’ll never be able to troubleshoot

[junior pulsar]

if i hire someone who can, i can always troubleshoot

[warm junco]

For install instructions for SD on AMD, visit #🤝｜tech-support and check the Pinned Messages there.
For Faceswap you need to install Reactor Extension.
For your website you can integrate the API of the webui.

[junior pulsar]

why waste time learning something i dont need to learn

[split kestrel]

Heh… that’s like web developer speak

Our company pays people 1k-5k to edit a document each time we make a change

[shell tendon]

#🏞｜general-with-images message

[split kestrel]

So… that 20.00 venture sounds like capital gains now.

Make it 1,000.00 and you’ll have someone’s attention.

[junior pulsar]

mhm, but i was offering $20 just to set up with AMD driver. i assumed it wouldnt take longer than an hour

obv i understand that my whole plan is worth way more than $20 to set up

[shell tendon]

$20 🤣

[split kestrel]

20.00 an hour is free if you learn.

[shell tendon]

i wont even bother walking across a parking lot for a five

[astral goblet]

should know what you're selling ebfore you're selling it tbh.

[split kestrel]

Bingo

And you should read your license agreements before you sell someone’s open source project

[loud solar]

At least he is not one begging us to do everything for free ....

[astral goblet]

insight face is for non commercial use only

[split kestrel]

I’ll take a workflow if we are handing out free stuff 😉

[astral goblet]

big drawback of it

it's what all the face swaps are based on

[junior pulsar]

1. me finding someone would take less time, and time is money so no it wouldnt just be free for me to do it. 2. I do know what I'm offering, I need to setup stable diffusion with amd

[split kestrel]

The time you’ve taken to argue your point… you could have read your resolution. But… good luck!

[junior pulsar]

and also, you can negotiate of course

[split kestrel]

The room bid starts at 1k

And I’m out

[junior pulsar]

$1k for one hour of work

max

sounds fair

i would go like $50 maybe

[split kestrel]

Technically… it is

[astral goblet]

its not the hour of work you're paying for. its all the time it took that person to learn this shit in the first place

i hear the same arguments about contractors from my dad when i tell him to modernize his home

[shell tendon]

here's basically what i need done from anyone, but been trying to figure this out for a while and figured i should join one of these groups and ask some smart people; i need a consistent and efficient way to face swap loads of content every day, and also have the most realistic results possible. at very minimum i will need about 35 videos and 35 photos a day. and i want it to be as easy / automated as possible. preferably would like it where our model can just drop the content in somewhere, and it automatically face swaps for us. i know this may be hard, but not sure, because some things I think would be hard to code / automate are apparently easy, and vise versa. <---- i can do thousands a day on my 4090 with better than reactor quality

[astral goblet]

$1000 is a lowball bid . anyone looking to boot strap your commercial endeavor for that is leaving money on the table

[junior pulsar]

that argument makes no sense. no one gets paid for the time they learned something. you get paid for the value you bring. at this point im just here to prove my point, this is wasting my time i know. but i can easily find someone else that would help me for $50 right now

[astral goblet]

not to mention, insightface isn't licensed for commercial use

[split kestrel]

Ps… a band charges 120,000.00 for 2 hours. For a medium touring act

So…

[junior pulsar]

are you guys like multi millionares or something?

like who tf u think u are

[astral goblet]

it doens't make sense if your head is up your ass about expertise. no offfense

[split kestrel]

If you have to ask

You can’t afford it

[junior pulsar]

you cant copyright anything ai

[astral goblet]

go find someone to help you set up your specific needs for $50. they'll check out before you're finished i'm 100% sure

[split kestrel]

Not true

[shell tendon]

@heavy lark is setting up a bank of 4x4090s just to mmake silly pictures so maybe we are

[split kestrel]

But go on

[junior pulsar]

thats at least what google says

[shell tendon]

yep gotta pay too for the time listening

[split kestrel]

I’m moving on. Time is money and you are taking away from real learning. Good luck! It can be done. Do your homework

[heavy lark]

FedEx is holding my stuff hostage currently. 🙂 hopefully I get it before sd3

[astral goblet]

i'm chewing thorugh machine learning street talk interview right now. this is a thick channel

https://youtu.be/sJXn4Cl4oww

[shell tendon]

oh god not fedex

[teal pagoda]

That's really not possible atm. And yea, I guessed what you want to do - automating OF accounts.

[astral goblet]

another thing i've found about lowball offers for expertise. these people think they own you after that. like if you can't respond in 1min, or help them a week later, they'll flip the fuck out

[shell tendon]

i can and have done it

[astral goblet]

or the contract will renegotiate suddenly. they'll pay you with "exposure" which is worth $50.... 🙄

[split kestrel]

I trade knowledge in the industry

All you have is your name

[teal pagoda]

with what? ReActor is a joke, facefusion looks 100% off and other tools are private and really expensive

🙂

[teal pagoda]

In what industry?

[split kestrel]

Most

Actually

[astral goblet]

reactor is bullshit yeah lol. it's just pasting the low resolution face on top of the final product, and codeforming it

[shell tendon]

yuuup

[shell tendon]

my own comyfui workflow

[astral goblet]

ip-adapter face swap is the best way to use insightface so far.

[teal pagoda]

:)) exactly, man

[astral goblet]

instant-id too but its sdxl exclusive

[teal pagoda]

then why are you still asking us here then? It makes no sense

[astral goblet]

holy shit. satisfactory is almost 1.0!

/tangent

[shell tendon]

oops i copied their question

[teal pagoda]

ID is the best atm, but it's more resources hungry than anything else for now

because it uses 2 CNets at the same time

[astral goblet]

resourcess schemozorces. just buy a fatter card

[trail lion]

omg just got here that conversation, lol

[shell tendon]

instantid struggles with anything beyond the original pose

[teal pagoda]

and believe me, for video faceswap nothing free is good atm. I tried almost all of them

[astral goblet]

with forge, you dont have to use both networks. 1 will do and it'll just freeball the new face position. start it at 0.1 steps or later so that the prompt can guide the positioning more than the network

[trail lion]

I wouldnt get out of bed for $50

[astral goblet]

not in my experience

[shell tendon]

faceid portrait is the best zero shot imo

[split kestrel]

But it’s worth 20.00 to someone who plans on making more extorting people for money.

[teal pagoda]

that's a good idea

[split kestrel]

Till they get shut down and run off

[shell tendon]

maybe its improved since i tried it, its been a while

[astral goblet]

i'm gonna post some quick tests in images in a min. i'm curious now

[shell tendon]

def curious too

[split kestrel]

Euler is just fine…. As I’m reading

[shell tendon]

are you using it in comfyui or elsewhere

[teal pagoda]

Man, 99.99% of the people who ask for videos faceswaps will use them for p*rn/OF. Kinda disgusting and anyway, they'll be banned in the near future. Just don't think that only by replacing a face, your content is "so" original.

[shell tendon]

mhm

[split kestrel]

I have some animations I want to run. But I’m trying to rig creatures

Face swap might make sense for legal morality…

[astral goblet]

i use the ui's mostly. comfyui i foun has the worst instant-id nodes. they break in fp8 too

but i haven't tested there for a bit

[split kestrel]

But uh… p*rn…. Sigh

[astral goblet]

not just banned. laws are coming down the pipeline and enforcement will hit hard

[split kestrel]

As it should

[astral goblet]

in canada i expect retroactive justice and i've been warning people for over a year

[split kestrel]

Wait till ai is just used for cats and objects lol

[teal pagoda]

I think that the banning of faceswaps (especially for p*rn/OF/destroying one person's reputation) is really worth

it can be even identity theft

[split kestrel]

Can be loads of unethical stuff

Just be ethical.

[shell tendon]

a1111/forge?

[teal pagoda]

all these faceswaps/deepfakes jokes from nowadays...

[shell tendon]

i havent used cubiqs

making it criminal will change very little

[teal pagoda]

Man, the only ethical use for faceswaps (not even faceswaps, but image generation with someone's face) is if that person is asking you to put her in a different environment/in different clothes and so on. But with her agreement and without using others' bodies.

[split kestrel]

For me it’s a likeness on a creature that I want to transform to identify

Zombie

Whatever

Permission to use****

And yes what you said

lol

[teal pagoda]

yea, this works too. But nowadays the Instagram is filled up with OF models/AI influencers which use faceswaps for useless dances and half naked (or even 80% naked).

:))

[sly pawn]

does anyone here use stabilitymatrix?

[real burrow]

Hi! The server is down ?

[astral goblet]

yeah thats right. something's changed though. instant-id isn't freeforming the faces as well, which defeats the whole purpose of my testing. i could use 2nd keypoints netowrk still though

[astral goblet]

consent. simple.

[teal pagoda]

from both parties - the one with the face and the one with the body

[split kestrel]

Exactly

[teal pagoda]

but 99% of the people won't even want to hear that

[sly pawn]

are there any really handsome really intelligent cool people here willing to help me with an error 🥺

[split kestrel]

Anyone know if you can batch render svd?

[astral goblet]

face swapping celebrities jjust for playing around, nothing sexy, i do that all the time. working with tobey maguire right now

[shell tendon]

yup its shouting into the wind

[astral goblet]

99% of a subreddit niche community. a minority of people to begin with

[teal pagoda]

The dangerous part is when someone make n*des with celebrities and try to sell them saying that they are 100% real - big problems incoming

[astral goblet]

1% of which are posting content at all

[shell tendon]

The day is coming where an AI virus finds photos of us all and swaps our faces and bodies and voices into acts with farm animals

It's gonna be chaos

Nothing we can do to stop it

[split kestrel]

When a computer understands itself, it destroys itself

[astral goblet]

there was that teenage boy who was extorted with deep fakes. they sent them to his family and he , well, you know. 💀

[shell tendon]

The sooner ppl realize nothing is real anymore the better

[teal pagoda]

I meant that no one will willingly agree to use her face/body for faceswaps

[split kestrel]

The control will just be a merge of data for face swapping of models

[teal pagoda]

You really have some fantasies :)))

[split kestrel]

Birds arnt real

[teal pagoda]

Nah, because you can't trust anyone then.

that's not a solution

[split kestrel]

Ps… this guide is perfection!!!

Thank you for this reading

Not really a guide… but you know

[shell tendon]

But that's where we are headed and there is nothing that can stop it short of nuclear war, asteroid impact, etc

[split kestrel]

Animatediff is fascinating stuff!!!

[astral goblet]

yeah instant-id even on forge now is busted. he changed something in the last few updates i guess.

[teal pagoda]

sadly, yea

[astral goblet]

it requires the keypoints

[teal pagoda]

told you, man

I don't even remember if it ever worked without keypoints

[astral goblet]

regardless, the model does work on it's own without keypoints, given the code implements it properly

[shell tendon]

Which is why ppl need to stop fighting it and accept nothing they see or hear is real just because it appears to be

[astral goblet]

we figured it out before the internet existed. we'll figure it out again

[teal pagoda]

Only the real face-to-face talk (until there will be clone robots) :))

[junior pulsar]

You have absolutely no idea what I plan on doing. You just assume that I am a scummy scammer who doesn't give a fuck about people simply because I am in the OFM space.

[split kestrel]

Sorry, I moved on. Next.

[shell tendon]

Or verified by trusted sources etc

It's going to get really hard to know whats true

[junior pulsar]

Next to where? Where are you going in life? From your pfp you're like 30s-40s

on discord

[split kestrel]

(I also don’t care that you are in OF space)

[junior pulsar]

yes u do, u assume im extorting people. you assume that i am scamming

[split kestrel]

I’m 43. Turning 44 in a few weeks if you are trying to stab at my age.

[teal pagoda]

Man, you just shame yourself. Imagine selling illusions to braindead lonely men who really think that the OF girl really thinks about them while she doesn't care about nothing but about what purse to buy next - with some indians as chatters (imagine chatting with p3rverts as a male chatter - you'll get braindead at one point or another).

[teal pagoda]

Yep

[shell tendon]

I have a chemistry PhD and here's my loser ass on discord

[junior pulsar]

im just providing models to agencies if you want to know exactly what my plan is

agencies have a hard time finding models to mangage their accounts

[trail lion]

it has other uses, like character consistency, but who knows, your 99% could be right

[teal pagoda]

That's legit what you're doing, but you don't call it "scamming", but selling "illusions"

[astral goblet]

#🏞｜general-with-images message

[junior pulsar]

so i am providing them real content, but to turn 1 model to 10 i have to use face swaps

[junior pulsar]

is gambling scamming?

[split kestrel]

Then learn it

And do it

[astral goblet]

aaaaahahaha you're so old. i'm only 41 oldieboyyy

[junior pulsar]

every relationship online is fake

[split kestrel]

lol 🙂

Ancient aliens

[junior pulsar]

so every OF girl is messedup?

even if they are the ones chatting?

[split kestrel]

No one cares… that you do OF

[teal pagoda]

it's the usage of the weak apes which are easily addicted

[junior pulsar]

these people choose to spend their money because it makes them feel good

[split kestrel]

Its work

Glad you are making money

[junior pulsar]

i would argue that selling alcohol is way worse

its way more damaging

[split kestrel]

However… you are not wanting to do the work.

[junior pulsar]

fast food is for weak apes

[teal pagoda]

I really researched this and yea, sadly, 99% of them are doing that.

[junior pulsar]

literally junk food, all ice cream, is feeding off "weak apes"

[split kestrel]

So do the work, and ask questions

[astral goblet]

what i find contentious about faking an of girl, is you as the operator still have to play the role of the girl and seduce the audience

[junior pulsar]

yes, they are paying for a fantasy

if they wanted to actually talk to a girl, they would go talk to someone in real life

[teal pagoda]

:))))

so true

[junior pulsar]

these people arent EVER actually meeting up with these of girls ever

so whats the difference between a fake girl and real

its all fake

[astral goblet]

the alcohol example. while it's legal, you can't give it to minors, drive while drunk, poison people without their knowledge, etc....

[junior pulsar]

okay, no one under 18 is allowed to use the site

mf

[split kestrel]

Well, good luck with your project.

[junior pulsar]

and how am i taking these peoples money without their knowledge

they totally understand how much they are paying

everytime

[split kestrel]

Move on time

Time is money

[astral goblet]

i think since you're asking that, you have issues with it of your own

[junior pulsar]

plunder fest we are having an argument and it is interesting to me to see why people think this is a bad thing so please stfu

[astral goblet]

but i think when people pay for a donut, they dont wanna get a hotdog

[shell tendon]

Can we take this to General with images so I can respond with clown Sharks riding cyborg centipedes into battle

[split kestrel]

My bad, go on then. It’s your name.

[red lynx]

🍿

[charred mesa]

🍿

[trail lion]

tbf move on time came and went so long ago, this is more like that slow train wreck that you cant look away from

[teal pagoda]

Just don't destroy the world more than it is right now. You know, OF destroyed this whole generation of girls/women as they think they only need a good body to get paid - so no real work.

[astral goblet]

part of the fantasy of OF patronage is a feeling of connection to the performer. People dont' want THAT much fake. They're consenting to a relationship with a performer who is misrpresenting? kind of breaks the rules of engagement regarding consent then don't it?

if you're up front and say "i'm a man producing this content" i bet you'd get a lot less customers

[split kestrel]

https://tenor.com/view/jason-momoa-chair-interested-gif-9751403

Really… a link?

[junior pulsar]

the value of what we bring to the customers is them feeling connected, and for them to be able to talk to someone, and also they are horny so yes, nudes. so you are right. but regardless if the model is real or not, they beleive the model is real, which gives them the same outcome. The outcome of them feeling good. Feeling connected. The relationships are never going to be in person anyways so what is the difference.

[brave gull]

Hello, does anyone know what's going on that the images are always blurry (at least 1/10 images are fine) with any prompt I'm sending, they appear the same whether in "SDXL 1.0" or "SD 1.6" ?

[junior pulsar]

Your brain feels the same stimuli when something is real and when something feels real.

[karmic brook]

feel free to take this convo to DMs, but this server isn't the place for this chat

[teal pagoda]

You and your braind3ad ("intelligent" business men wannabes) comrades are contributing to the destruction of moral girls/women by showing them that they can easily get money by selling themselves. Not even you will date an OF girl, let alone marry her.

[astral goblet]

disclosure is ethical and required by EU law now

[teal pagoda]

yea, right. Wasted my time with an "oxyg3n w@ster"

[junior pulsar]

having a phd does not make you a valuable person. congrats, but that doesnt make you valuabvle

[trail lion]

you have enough steps? sometimes the sampler you are using needs more or less....like euler a likes probably 35-45, but others like dpm2a you can get away with 20 - 25

[split kestrel]

Depends. He might be a heart surgeon.

[astral goblet]

i don't think he was puffing himself up there

[junior pulsar]

so you're saying that its bad that girls want to get into OF? And they shouldn't?

[brave gull]

Of course, I have tried many eprompt configurations but the result is still the same.

[teal pagoda]

only the SDXL 1.0 has a more accentuate problem with blur

[astral goblet]

kind of weird that your response was to cut him down after he called himself a "loser ass" lol

[split kestrel]

NO ONE is bringing up OF except the person doing the OF

It’s like vegans…

[astral goblet]

||insecurities much?||

[karmic brook]

moving on

[astral goblet]

or engineers

[split kestrel]

lol

I spent my morning making pictures of plants eating vegans

[teal pagoda]

:)))))) legit delusional

[brave gull]

This is my prompt, among other variations: Beautiful girl, silver hair, full body, Los Angeles beach in USA, 4k, style of Salvador Dali

[astral goblet]

hey speaking of tobey maguire! I heard spiderman 4 is in the works to be made!!

SQUEEEEEEEE

[trail lion]

ok, and you've tried different prompt variations? blurry image output is not normal...do you have a vae loaded?

[red lynx]

What is your workflow? Maybe you haven't set resolution to 1024x1024 if you are using sdxl

[teal pagoda]

If you're asking this, your values/ethics/morality are at a below 0 level

[astral goblet]

ethics are highly subjective

[brave gull]

im using sd1.6 and 512x512 y sdxl 1.0 and 1024x1024

[teal pagoda]

blurry background or the entire image?

[astral goblet]

lol i see so much hate for bokeh

[split kestrel]

Are those the standard resolutions?

[brave gull]

I have been sending prompts from the api with python for many days and everything has always been fine but today the problems started.

[brave gull]

complete image

[split kestrel]

512 and 1024

[teal pagoda]

then that's not normal

[trail lion]

too few steps or vae issue are the only things I can think of

[brave gull]

if i use sd1.6 -> 512x512 if i use SDXL1.0 -> 1024x1024 .... all prompts are blurred

[split kestrel]

Interesting

Oh oh oh oh

[brave gull]

of course, it's not normal, I've been using the api for about 1 month and no problem but today almost 100% give me problems.

[split kestrel]

I learned that a code was updated 2 days ago!!!

Uhhhh. I forget the workflow that was broken

[astral goblet]

i can assure you that sdxl is not a model that just produces blurry results. something is wrong in the api calls.

[split kestrel]

^

[trail lion]

random question, there's no sv3d extension yet right?

[split kestrel]

It exists

It does not produce anything g good

[shell tendon]

its on comfy

[teal pagoda]

Do you guys think that Stability AI will continue to be open-sourced after releasing SD3? I mean Emad and the best people left already...

[shell tendon]

at this point yeah

its their big advantage

[split kestrel]

It’s just a matter of time before someone cracks the code so to speak and it drifts out to everything else

[shell tendon]

and the statement recognized that

[split kestrel]

Always a smarter fish

Does automatic111 have an option to input a video source for animate ? Video - video ?

[trail lion]

there's one for gif to gif, many of those are 1.5 I think, same with animatediff

[split kestrel]

Hmmm I keep wanting to try it, and then feel like comfy is just “more”

lol

[trail lion]

wouldnt know, never tried comfy, no desire

[split kestrel]

What’s your reasoning?

Curious

[trail lion]

umm, I have blender experience, and when I think of the shader which is node based, I cringe

[split kestrel]

Yup

[honest spear]

stable diffusion before 1.5 was already open source and wasn't even released by Stability but Runway. I have no evidence of them changing policies, in the letter they relased a couple days ago a the end they said they continue support the community and open source. But things can always steer nothing is certain

[split kestrel]

I would agree

[trail lion]

then I see people post their node arrangements, and I'm like...yup, that's what I thought

[split kestrel]

lol

Some of it… is insanity

[teal pagoda]

For AnimateDiff, yea. But it's resources hungry. Without atleast 24 GB VRAM, you can't really do much.

[split kestrel]

I have the room

I think…

[teal pagoda]

:)))

So only the AnimateDiff for 576x1024 video is using around 15-16 GB of VRAM. Imagine that by using the vid2vid option, you have to enable at least one ControlNet. And the results without hires. fix look so bad... And you'll want to use a SDXL model. So you'll have an usage of around 20-24 GB of VRAM (minimum).

[trail lion]

funniest part is when people are requesting 'workflows', because it eventually becomes more effort to manage the app than the actual thing you're trying to generate

having said that, if I was trying to automate, I'd use it, but I dont need that

[teal pagoda]

:))) kinda right, that's why I dislike comfy

its interface reminds me of the Unreal Engine program

[trail lion]

it's a thoughtful powerful design, and wouldnt take that away from it...but hard to beat pasting a prompt and hitting generate

[split kestrel]

I saw a flow on comfy that I might chase when I’m back to the zone

Used a video to extract the motion and … outlines? Then converted that to allow the model to follow that motion

Interesting stuff!

[teal pagoda]

I think I saw something similar

[split kestrel]

Now, if I can take a video of a human model doing what I need them to do for exact motion… convert that, then animate THAT motion with a created model, then batch that model for extended frames…

That might be game set match for what I want to build. Provided the graphics card doesn’t die or overload / crash lol

[teal pagoda]

:))) try until you get one crash and never after

[split kestrel]

lol

I’m back to learning animated diff now

Which brings me full circle 🙂

Is there a frame rate / fps where animate breaks apart?

If 32 frames at 8fps is normal… could you crank to 120 frames at 30fps / 60fps?

[astral goblet]

the frame rate setting just defines the final file. nothing about the process. the context size is what breaks it. 16 context batch size is my main sale. 24 blows it out. 20 is sketchy. 18 sometimes works.

the model is trained at 8fps, so if you set the final file at 60fps, it'll seem really hasty . thats where interpolation comes in

[split kestrel]

The model being the motion model?

[astral goblet]

yes

[split kestrel]

Understanding 🙂

I need to understand the batch stuff

[astral goblet]

i often crank the frame rates because smooth gifs are hot

[split kestrel]

Right!

In video.. I usually prefer to edit at 60fps

[astral goblet]

think of the generated frames as keyframes

interpolation works REALLY well in these domains

[split kestrel]

Batch size… I imagine… is simply the amount of pictures generated to follow said motion?

I spoke too soon. I’m now at that reading part lol

Needs to be in a scale of 8

Wierd

So 40fps is max

Same rule of context batch

8-16-32

Also says the quality degrades sub or plus 16

[forest trout]

Batch size tells you how many images will be produced simultaneously. It's resource intensive and uses up vram.

So in the case of video generation, I'm assuming it's generating x images simultaneously.

[split kestrel]

Makes more sense why 8’s now

[forest trout]

I mean it's going to be faster, but at the cost of vram.

[astral goblet]

the context batch is how many frames it's thinking about in memory at a time. when i generate 100s of frames, it'll act on 16 at a time, with a sliiding window. my taskman shows a ton of small copy operations as it caches memory back and forth

[split kestrel]

16 seems to be the match

I don’t know if I have the same controls in comfy but it helps me to understand the reasoning

[astral goblet]

instead of doing one 16 batch all at once though, it'll do 1 step, then move to the next frame window and do 1 step there, going through it all before it starts on the second step. it requires more system memory in that case

[split kestrel]

And animate says it breaks

[astral goblet]

i think in comfy it doesn't come stock. you got to wire it up to work that way

[split kestrel]

Says everything beyond 16 is not a good idea

Right 🙂

[astral goblet]

yeah i've seen that too. i pretty much keep it on 16 frames only

[split kestrel]

It’s helping me learn all the steps in place

So I can ask more thoughtful questions lol

[astral goblet]

its trained on 2 second clips cut down to 8fps. if you're ever thinking of making your own motion module or motion lora

[split kestrel]

I was. At some point

Open pose?

[astral goblet]

you can use open pose across frames yeahh

[split kestrel]

I’m gonna get home in 4 hours and build stuff lol, then get real mad when it breaks and I don’t know why 😉

[astral goblet]

moms spaghetti

[split kestrel]

Does animate require 1024/512 or it breaks?

[novel narwhal]

btw does thiis make generation slower?

[astral goblet]

uses sd15 so you're going to want to use 512 ish resolutions. i play pretty loose here though. often 640x640

[warm junco]

No, but if you store models on a HDD, it can take longer to load them

Generation speed is not affected

[novel narwhal]

ah ok

idk why but after resetting my pc and reinstalling SD it feels slower

[warm junco]

Maybe your missing the performance arguments in your webui-user.bat

[novel narwhal]

o probably

[low moon]

Could a mega well trained lora fix hands for good?

[warm junco]

What's your GPU?

[novel narwhal]

3060

[warm junco]

With 12gb vram?

[novel narwhal]

16

[warm junco]

There is no 16gb 3060

[astral goblet]

lol. adetailer is sufficient.

[novel narwhal]

wait might be looking at wrong one

[astral goblet]

or a detailer node

[novel narwhal]

its 8

ill just send this

oh cant send pictures

[warm junco]

So its a 3060ti?

[novel narwhal]

yh

[warm junco]

Okay then you need --xformers --medvram-sdxl --no-half-vae
As COMMANDLINE_ARGS in your webui-user.bat

[warm junco]

There are some hand loras but I dont think they can fix everything

[split kestrel]

SDXL for animate doesn’t look all that great at the moment

[novel narwhal]

so like thhis?: set COMMANDLINE_ARGS= --xformers --medvram-sdxl --no-half-vae

tyty

[astral goblet]

kind of curious how people get such clean open pose frames. i run openpose on this video for a batch operation, and 30% of the frames are blank, missing limbs, broken all over the place, whatever. is it because it's a battletoad? https://i.kym-cdn.com/photos/images/newsfeed/001/459/662/f79.gif

[warm junco]

Np, should be much faster now.

[split kestrel]

I just ran over that

Open pose has a limit

[warm junco]

Dw-pose could help a bit

[astral goblet]

i thought so too. same problem. i'll try it at various resolutinos and see what comes out

[warm junco]

And enable pixelperfect

[astral goblet]

been using pixel perfect which sets the resolution to the image's resolution. thats why i'm gonna try doing like 400 or 600 resolutions and i'll see what happens

i think its cause its a dancing toad

[warm junco]

Haha xD what's the initial resolution?

[split kestrel]

I think I read that it needs a specific input

[astral goblet]

300x300

[warm junco]

Maybe thats to small for a proper recognition

[west geyser]

Is there a bot to watch for SD3 Beta output ? i saw some posts on Reddit but wasn't sure if they were legit.

[astral goblet]

#🏞｜general-with-images message the poses coming out from dw_pose at 640 resolution aren't so bad. i'll run with this and see what happens

[split kestrel]

I imagine you will have to upscale output after and pray?

[astral goblet]

they're just controlnet inputs frames. it should scale to my render resolution well enough

[split kestrel]

I was questioning if I film a thing, and make it so the human was small… run it though op, then rig it to a cat for proportions

Instead of stretching the cat, stretch the pose.

[astral goblet]

open pose resolutions are quite decoupled from the final resolution. it determines the thickness of the pose lines more than anything. and how many pixels it's trying to base the pose off of

the controlnet openpose model is pretty forgiving

posted the final openpose bake into #🏞｜general-with-images

[split kestrel]

On average… how many steps do you run on your base image?

[astral goblet]

30 is where i like it but i feel thats often not necessary. depends if i'm using lcm or not

[split kestrel]

I was playing around with 64 last night

[astral goblet]

its so easy to fiddle with these settings in something like automatic1111 so i often twiddle and tweak all the nobbies

[split kestrel]

But I had no idea what I was doing

[astral goblet]

higher step counts got some nice results too.

[split kestrel]

42, had some real nice elements

But it also has me in believe that it’s based off 8’s

Instead of pose, would a control net with edge detection work better?

[astral goblet]

alrrright. i'm hucking a sonic prompt at this dancing battletoad skeleton. lets see what happens.

[split kestrel]

Gooooo Gettum!

[astral goblet]

next thing i want to figure out how to accomplish is regions and prompt travelling

[split kestrel]

I was juuuust learning about that too

Seems it’s not great

Small motion is better

[astral goblet]

this may take 30min or so. will be back once the smoke has cleared

[foggy halo]

sighs SD3 release when!?!?!

[split kestrel]

Right now, you missed it tho…

So now you wait

[astral goblet]

@split kestrel if you're interested #🎥｜animation message its cursed though

[split kestrel]

lol

[astral goblet]

the dw-openpose brings the face markers in, and i don't think i want those for sonic

[split kestrel]

I like you people

I’ll get there

[astral goblet]

theres a few moments its working almost. you see how the clothes shift ever couple seconds? thats the context window

[karmic cedar]

If you’ve never tried feeding the script for the turbo encabulator video on YouTube to Suno, well

[split kestrel]

Does animate require a specific sampler / scheduler?

lol

[astral goblet]

https://civitai.com/models/120149/controlnet-for-densepose i just found this model. going to try it instead

densepose was waaaay faster and more accurate

[split kestrel]

Im now itchy to play again

[astral goblet]

i've not experimented much there

[split kestrel]

Wondered if that sampler is different from the k

[astral goblet]

which?

[split kestrel]

Animate sampler

Same as ksampler = good to know!

[astral goblet]

in auto land, we dont have seperate samplers / schedulers. well, 1.9 will

[split kestrel]

Well if you ever need that info. It’s the same

lol

[astral goblet]

i didn't know there was a densepose controlnet till just now. cool beans

[karmic cedar]

i’m still waiting for a weird random controlnet to drop called nerfherder or flargenstomp

[split kestrel]

Can a parent image be a created image?

Sorry

Latent

Like a reference image?

I’m starting to see the variations in comfy / auto 🙂

[astral goblet]

after hucking sonic at this densepose sheet, i'm gonna double up the controlnets and try a depth map with it too. so the leg positions stay consistent

[nova zodiac]

Nerfherder lol

[split kestrel]

Ok, I think I have a final question before I get home and play lol

VAE

My understanding is that it’s baked in on some of the data sets?

Checkpoints

Gonna read this real quick before I ask a dumb question

Answer my question 🙂

Answered **

[fervent thunder]

👍

[crystal fable]

hi

[crystal edge]

how can i use an image as the style for all other images that i want to generate

[crude notch]

server has 100ish people now

[willow crescent]

IPAdapter+

[crystal edge]

ok i will try it rn

[crystal edge]

i downloaded it where can i find it in gui

[willow crescent]

in Comfy?

[crystal edge]

a1111

[willow crescent]

don´t know 😄

[crystal edge]

xd

[pearl ocean]

GTA7

[regal snow]

Ip adapters are used via controlnet extension

Put ip adapter models into models/Controlnet/

Then it’s a matter of using the correct preprocessor… refer to controlnet github. May need to check Discussions tab

[astral goblet]

unless you're in comfyui land then theres a specific node for them

[regal glacier]

yoooo hook me up ..
my first installation was succeed to open SD on my browser (http://127.0.0.1:7860/)
but it's out of memory

now my SD literally do nothing

[karmic cedar]

https://tenor.com/view/you-go-girl-gif-12815320275574392740

[trim magnet]

get a better gpu

[opal hedge]

4-6 weeks

Copium

[astral goblet]

https://civitai.com/models/232042/loosecontrol-use-the-box-depth-map-to-control-the-protagonist-position ahah! i knew i wasn't hallucinating this. might try this depth model with animatediff

https://shariqfarooq123.github.io/loose-control/

[remote badge]

anyone use forge webui and know what all the integrated extensions do?

ive been trying to find documentation for things like what hypertile "scale depth" and "depth size" should be

but find nothing

[regal glacier]

I thought there is some script to support low end gpu ...

[gaunt pulsar]

I want to believe

[trim magnet]

whats your gpu?

[opal hedge]

That came from the stand-in CEO of SAI, but they said release in 2 weeks a month ago, so who knows tbh

[clever mountain]

Hey folks - Can someone explain to me the difference between say SD-1.5 and SDXL. At first I assumed they were completely different platforms so to speak, but now as I'm familiarizing myself with SD more, they just seem to be differently collections of differently tagged and weighted images. Similar to one of the dozens of checkpoints I can find on civitai. Is each version simply a new set of images, specially weighted and tagged? If this is the case why do so many factors change, like render time, resolution, controlnet support etc. Juggernaught XL is a checkpoint based on SDXL... right? Is it just the SDXL model collection with additional tagged images and weights? Is it completely separate? Why more excitement around these models landing as opposed to literally any other checkpoint? Is it just the volume of reference content?

[opal hedge]

They're different base models of stable diffusion. SDXL is newer and better, but SD1.5 has more stuff for it

[clever mountain]

I appreciate the prompt response, but doesn't really make the distinction im after. Is Juggernaught XL not a base model? Is that an iteration on SDXL? Is SDXL Lightning a completely different base model?

[opal hedge]

If a model has XL in the name, it's a finetune of SDXL, meaning the weights were changed to give specific and/or better results. SDXL Lightning is a different base model from SDXL

[pseudo kettle]

Como gerar imagens

[clever mountain]

Alright, so when they are talking about SD3 coming out, it's just going to be another collection of differently weighted and tagged images (likely a larger collection i assume)? Why then is it that something like controlnet will work well with 1.5 but not as well with sdxl? Why are Lora's SD version specific... is it because LORA resolution might not align with the foundational model?

[karmic cedar]

Yes

Essentially

[opal hedge]

sd1.5, sdxl, sdxl lightning, sd3 are all different base models meaning they have different AI architecture. What works for one will not always work for the other, and occasionally a finetune of a base model will not work with tools meant for that base model. Welcome to cutting edge tech

[clever mountain]

Ah got it, so models aren't just a collection of photos and defined metadata. Theres an AI element integrated there. So something like PONY also has their own proprietary AI architecture for interpretting prompts etc

[opal hedge]

PonyXL is a fine-tune of SDXL. The architecture is the same, but PonyXL was fine-tuned using a very cut-throat method meaning not all the tools that work for SDXL will work for pony

[karmic cedar]

Look up convolutional neural networks

[clever mountain]

Interesting, so the framework for interpretting these checkpoints/base models isn't contained with the SD framework (which i can leverage using a variety of UI solutions), but rather the models themselves?

[clever mountain]

will scope it, thanks for the suggestion

[karmic cedar]

Essentially these models are giant rubber band balls.

[clever mountain]

this sounds like the start to an amazing allegory -- is there a part two to this haha

[karmic cedar]

Sort of! The rubber bands aren’t really closed loops—it’s just a superhighway of superhighways contained in a larger superhighway, and there are control mechanisms placed at different points to direct traffic to various processing “stages”

And some models have eight lane highways, others have two

Some models have two parking garages, some have one

Some have a parking garage attached to a mall (controlnets)

[pearl ocean]

Question, is the apple M3 Pro any good SDXL image creation, in terms of speed?

[karmic cedar]

I don’t know I stopped paying attention after I learned that one image generation these days is equivalent to a full smartphone charge

[trim magnet]

fake information

[karmic cedar]

I sure hope so Steven

https://youtu.be/0ZraZPFVr-U?si=JwzM4Gu17VyuHxtv

[floral nimbus]

anyone familiar with StableSwarm and if so is it possible to hook all my models from A1111 to StableSwarm without duplicated files

[iron pendant]

symlinks

[floral nimbus]

elaborte

[iron pendant]

just google "symlinks" and your operating system

basically you put all your model files in one folder, just pick one, and symlink the model folders for other applications

[green mauve]

is stable diffusion paid only ?

[floral nimbus]

ig if you count the VRAM aspect

[floral nimbus]

couldn't of been said better, amazing tool that should be used and easier to do

[opal hedge]

It's free if you run it on your computer

Though you do need at least a 'decent' computer

[vast berry]

me with my 8 year old laptop that may as well be made of cardboard with parts i've had to replace at least 3 times each

[pearl ocean]

Your laptop is probably in year 2 at school

[plush sand]

Hey guys, could you please tell me where can I find the bot1-10？

[warm junco]

Hey, the bots are currently down

[plush sand]

Got it, thanks buddy

[wet dawn]

gm

[woven grove]

Is there a way to view checkpoint configuration parameters?

[still glacier]

No because they re not stored anywhere. So read the readme/description and store the (recommended) settings yourself if needed

[woven grove]

There is no way to view information about a checkpoint xl or sd1.5

[bleak vale]

Does anyone know the download address of chillgenV2 checkpoint?

[still glacier]

you can check if a checkpoint is sd1.5, sd2, sdxl but that s about it. If there s any trigger words, preffered samplers, etc that s not gonna be indicated anywhere else than in the readme/description

[loud solar]

Have a look at civit.ai ...

[woeful forum]

Hi everyone, I'm a complete beginner to AI image generation and I'm currently using a fresh installation of AUTOMATIC1111.

In an article I came across, the author suggested using a specific prompt along with a negative prompt called "FastNegativeEmbedding." It sounded a bit unusual to me, so I queried ChatGPT about it and they confirmed that it's indeed something I should use.

However, while searching online, I found this on Civit AI: https://civitai.com/models/71961/fast-negative-embedding-fastnegativev2. It seems like something I can download, and it appears to achieve a similar outcome to what the article's author was aiming for.

So, my question is: does this "FastNegativeEmbedding" negative prompt work right out of the box, or is there more to it?

[loud solar]

For SD 1.5 only ...

[nova zodiac]

yep, just add it to the Embeddings folder, then use FastNegative in your negative prompt

assuming you are using a model based on stable diffusion 1.5

[woeful forum]

So, if I won't add it to this 'embeddings' folder, this will just not work, even when I work on SD 1.5?
Plus, one more question - what does it actually do? I mean, I can guess, but is there any way to find out specifically? Am I right that somehow this 'something' adds lots of words/phrases to the negative prompt? If so, then what are those words?

[loud solar]

And you have to write the magic words in the negative prompt

[woeful forum]

Right 🙂 BTW, I'm wondering what does it do exactly so I can test if this actually work.

[loud solar]

I only have the idea it's a collection of negative prompts ...

[loud solar]

Try to create the same picture with and without it using the same seed ...

[woeful forum]

Did that, but... ah! Using the same seed! I have read about the seed yesterday and forgot 🙂 Information overload 🙂 Thanks!

[grand relic]

what’s the difference between SDXL and Cascade ?

[flint fox]

Cascade is based on the Wurstchen architecture

and SDXL is just a bigger SD1.5 with a base and a refiner

[still glacier]

SDXL has a different architecture than 1.5

[flint fox]

isn't SDXL just using a 3 times larger UNet and a second text encoder? idk because I only watched the abstract of the paper

[still glacier]

changing the text encoder is already a big enough change in my book to call it different

[flint fox]

SD1.5 and SDXL both uses CLIP iirc?

[still glacier]

different clip if I remember correctly, also sdxl has 2 text encoder, 2 tokenizer, etc there is the refiner models (most people don t use it), etc

[unkempt hatch]

so I wonder how well SDXL understands relativity. Say I'm training an IRL Homer, so I'm finding various men who look like Homer might if he were real.

In one photo, he looks like homer, but not as fat. Should I tag it "skinny" even though he's not really skinny? But he's skinny compared to Homer

[still glacier]

fit homer maybe

#🔧｜finetune is probably the best channel to discuss that kind of things

[unkempt hatch]

Ya, I'm wondering about this specific example, but also wondering about the SD "logic"

Because I might do the opposite, like someone who's normally very skinny, and in the example where he's average, tag "fat". Or someone who's muscular, but in the image that he's not "scrawny"?

[green mauve]

where can i find instructions to install?

[opal hedge]

youtube a1111 stable diffusion good luck

or if you're feeling particularly adventurous youtube comfyui starter guide

[zinc egret]

civitai is one guy ... he train an ai ... sorry say so but get a girl friend lol

sd can do so much more

[trim magnet]

sorry i prefer a man friend

[zinc egret]

try this prompt , : intergalactic nebula

NY city in 3033 lol

SD is great for creativity

@trim magnet that s not the point ,,, i use it for backgrounds and it s great

😉

by the way you should have a section for a gallery

so we can show them the power of SD

[fervent thunder]

hello

is sd 3.0 available yet?

[true canopy]

no

[topaz parcel]

So I am a Mac user (M2 Pro Max) and getting so frustrated at the inability to do anything with img2img and ControlNet with any XL model any more. It's basically broken on a Mac. And the solution to run -- no-half is useless as it makes any generation take way too long - 30 minutes plus. SO. Any solutions does anyone know? Or if it means the truth is I need to ditch the Macbook for a PC, then what equivalent PC Laptop off the shelf does anyone recommend that will match the spec of my Macbook M2 Max, but also be able to run A1111 and similar with no issues.

[karmic cedar]

Your thinking is correct—these models are extremely demanding and resource-intense.

You can use cloud infrastructure or target a PC computer with a graphics card that supports the tensor architectures you’re working with.

[topaz parcel]

ahh - ok - so rent like a cloud rig you mean

[karmic cedar]

Indeed. You can pay as little as $0.30 USD an hour for some

[topaz parcel]

ooh - thats a great idea

[karmic cedar]

However, keep in mind that one image generation costs about the energy equivalent of a full smartphone charge

🙂

That’s actually just me projecting, sorry, watched a video and I’m still like…uhhh

[topaz parcel]

Really. Did not know that.... ouch. So the cloud rigs, do most of them allow yoiu to upload your own LoRAS, etc

[karmic cedar]

Yes, as long as you have a decent upstream you can use them much like your local rig

[true canopy]

but, can u even use comfy and stuff wth cloud gpus?

[karmic cedar]

of course

you can even run a virtual Linux desktop and play with Blender if you want.

[topaz parcel]

def gonna try that.. perfect idea. i'd forgot about this option, as My mac was coping fine, until I started to experiment more with XL models.. then it kind of crumbled on me

[karmic cedar]

Give it a try. I use a phone and an iPad, and that’s all I’ve got.

You at least have some understanding of the basics, which is more than a lot of folks have when they start using cloud services.

[true canopy]

can u rly get a 4090 for that 0.3/hour tho?

ive never looked into it

[karmic cedar]

nah, those are more like $0.75-$1 ish

[topaz parcel]

I'll have a look - and see wahts on offer

[karmic cedar]

but you can 🙂

[true canopy]

i was wondering how they made money with that 0.3

but im sure its a decent offer for some people

[karmic cedar]

If you know what you want to create, and you know how to get the distributions together and line up your workflow, you can start a cloud session and get a lot done in very little time. You just have to pre-plan your work a little to make sure you don’t waste.

[true canopy]

it would be a sort of stress for me to have some kinda limit

and if the gpu is going to go warm, might as well warm up mu room ><

[topaz parcel]

yep - i'm pretty okay at the set up and settings and the use of UIs like A1111 and comfy, etc. so should be easy enough to get going. and def cheaper than going out and buy a $3000 laptop just to do this one task my mac can't

[karmic cedar]

For me, it’s like…a roll of film. When you’re shooting medium format, you only have 12 shots per roll of film. That puts more pressure on you to compose, etc.

for sure