💬｜general-chat

[vague pond]

people feel like they're too good to use a jupyter notebook as their SD frontend, smdh

[rugged mirage]

I'd prefer a Jupyter over comfy tbf, I just use whatever has the most tooling support

[steep arrow]

Comfy lets me think in a spacial way about what the code is doing, really wish it was a little more stable in between nodes, but it is "freeing" compared to other software.

[vague pond]

i would prefer just doing the vector calculations by hand over using comfy

[hot kettle]

I'd only use python/jupyter but for some reason comfy and automatic1111 works faster and with lower vram

[rugged mirage]

I actually did use only Jupyter (and colab at first) until I wanted to try some workflow that used a111 extensions

[rugged mirage]

probably just comes to settings and tricks they use under the hood

[vague pond]

it has been 5 weeks since stable-diffusion-webui-forge's last commit

[hot kettle]

I tried everything I've found and still not quite as efficient

[rugged mirage]

but like I'd really rather feed functions into functions than connect fucking nodes

[hot kettle]

Especially if you don't see what's exactly under each node

[vague pond]

spaghetti doesn't belong in the frontend where users see it, it belongs in the backend where it frustrates developers

[rugged mirage]

yup, I wish there was a right click > convert to code option so I can easily edit it or look at it at least to figure it out

[hot kettle]

And i don't wanna open github every time I need to check something

[rugged mirage]

not just an opaque node name

[hot kettle]

I just need to find every optimization they did and copy it into python

But I'm lazy lol

[vague pond]

the node system doesn't even make sense. Like the "model" type which simultaneously covers model weights and merges, unet hijacks (like FreeU), and sampling modifications (like CFG Rescale/ZSNR). It allows you to make complete nonsense configurations that then behave unpredictably.

[steep arrow]

You just have to feel it out, like wandering through a cave in the dark lol

[hot kettle]

Also don't like the way nodes load inputs into memory

Sometimes it goes in the opposite direction it should

[vague pond]

What does it mean if you merge a model with ModelSamplingDiscrete set to V-prediction and zero terminal SNR with a model set to regular epsilon prediction (which for the uninformed are very much mutually exclusive)? Who knows! But ComfyUI will not stop you from setting that up.

[charred mesa]

^

[hot kettle]

Will that work in any way?

[sage reef]

i dont think it's nonsense, it patches part of the model unet in a certain way, freeu targets specific parts in the model that for example a lora or controlnet doesnt, etc, so i wouldnt say its nonsense, it makes sense, but it's up to you to adjust then the parameters to fit your needs

[steep arrow]

I honestly like having the ability to break things in simple ways.

[eternal ledge]

But diffusers practically requires you to use huggingface and oftentimes breaks if they have downtime (and they only have two 9s of uptime last I looked for a period of just a few months)

[hot kettle]

Either way I was lost when I needed to use some funny nodes for playground 2.5 to work

[vague pond]

I'm saying that it makes no sense to have them before a merge node, that would imply that one model in the merge has it and the other does not, when really you either have it or don't

[hot kettle]

Does it break even if I've already downloaded the models?

[eternal ledge]

If you've already downloaded the models, you should be fine, but if you're not using a persistent machine, you've rarely downloaded the models

[hot kettle]

Well good for me since I've got them all on my ssd

Ready to be used by a 6gb GPU...

[vague pond]

One or the other will win, and it'll just sample from one of the model's prediction mode

[eternal ledge]

Also if you're not using the exact right flavor of a diffusers pipeline, you can't use safetensors files, which is pretty annoying

[vague pond]

But you can point diffusers to a file path just fine?

[eternal ledge]

Last time I tried to use SVD, it was hf repo only, but maybe they updated it

[vague pond]

I mean you can still just git lfs clone it and have the files locally and outside of HF's cache management system

[hot kettle]

I could load some diffusion models no problem from a file path

[vague pond]

you can even load regular model checkpoints, StableDiffusionPipeline.from_single_file(path)

[eternal ledge]

Some, and with the right pipelines. I haven't figured out the full rhyme or reason to whether a pipeline has .from_single_file, if you're doing just the basicmost types of generation, Diffusers are convenient, but it seems easy to accidentally end up in hf-dependent land

[hot kettle]

Yeah haven't ventured too far from diffusers

[eternal ledge]

It's still within diffusers that some of this is annoying: https://huggingface.co/docs/diffusers/main/en/using-diffusers/svd - if I have an SVD finetune safetensors file (e.g. SV3D), I can't just use diffusers to run it. I have to get it into a huggingface repo with the right shape and so on

Because StableVideoDiffusionPipeline doesn't have the .from_single_file(path) method

[hot kettle]

I think at this point it's just better to download model architectures from github

[steep arrow]

The world of image gen outside of comfy looks incredibly intimidating lol

[worn aspen]

What do you mean by this? I feel a local install is better than a cloud instance.

[eternal ledge]

It's easier just to use comfy 😂 the stability-ai gh repo for svd uses an inordinate amount of VRAM for some reason

[hot kettle]

Yeah probably

[cerulean kraken]

I got it working and have generated quite a few images... but my credits aren't going down I am confusion.

[sage reef]

it do go down

[woven panther]

😅 don't say that out aloud

[cerulean kraken]

😬

[sage reef]

you have glitched the system

[woven panther]

Probably just a delay

[sage reef]

the system is overwhelmed with requests

[junior urchin]

While the model is available via API today, we are continuously working to improve the model in advance of its open release. In keeping with our commitment to open generative AI, we aim to make the model weights available for self-hosting with a Stability AI Membership soon.

[rugged mirage]

the license is almost definitely the same as what it is now https://stability.ai/membership

[bright dragon]

How to use sd3 guys?

[karmic cedar]

With butte

plz try

[pine path]

I know one person planning a 4-5m finetune for cascade

[karmic cedar]

I hope they call it the Big Sausage edition

[shell tendon]

well, they just deleted the #stable-cascade channel.

sadly

[pine path]

they're not in this discord so it doesn't really matter to them

[steep arrow]

Very cool, I am working on a stable cascade inpainting detailer to correct generated text.

It is a nightmare to scale everything to a padded 1024 square in comfy, but it does work.

[pine path]

it's furry+anime btw

[steep arrow]

Do you think it would have improved text gen at all?

Are they training with better text labeling in anyway?

[pine path]

I don't think so. They're not training on text caption stuff, just booru tags

[fierce swift]

Hi,
I'd like to use Google Colab (pro) to run SD, then use Dreambooth to train it on my own drawings and create a style and then also use contronet or similar plugin to designate areas for different things, like 'trees', 'creek', 'rocks' etc (to make a forest scenes).

Is it possible to install SD with Colab and then save out the configuration and then move on to adding Dreambooth, and then move on to add other plugins, such as the controlnet type functionality?

I've found some setups but most of them seem old and throw errors - if the above is possible, can someone point me to up-to-date resources for making it all happen? - I'm fairly low experience with git etc.

[steep arrow]

RIP me.

I have been trying to duplicate the features of SD3 with Cascade and SDXL, text gen is possible but its so finicky.

[pine path]

have you looked into the anytext model? I haven't used it, but it might be good for an artificial dataset.

[shell tendon]

i'd love to see one that actually cleans up the leftover latent noise/oversampling issue with stage B

[neon oriole]

comfyui always had the correct implementation of the refiner, compare it with the official example, A1111 has the incorrect one

[neon oriole]

you can merge vpred and regular models they are not mutually exclusive, I even have an example: https://comfyanonymous.github.io/ComfyUI_examples/model_merging/#advanced-merging

[steep arrow]

Looks incredibly useful. Thanks.

[vague pond]

I know you can merge them and get a working model, I've seen people do it, but when you are sampling, you are going to be sampling the model output as either epsilon or V and not some interpolation of the two, and I have seen some people confused by this because intuitively you might think it's appropriate to put the ModelSamplingDiscrete node between any v-prediction model and the merge node, and for everything to work flawlessly after that. This won't turn out as expected, though.

[neon oriole]

yeah because it requires a minimum knowledge of how things work

but that's something people can learn by experimenting with stuff in the ui

[vague pond]

That's fair I suppose, but my experience with seeing people fail to use included config files or the appropriate node does tend to bias me in the direction of "make things completely idiot proof".

[vague pond]

if this is samplers.py line 709 to 733, then it doesn't look correct. It looks like it is doing the same thing A1111 was doing (wrongly) -- it calculates the sigmas used for sampling, and "denoise" determines which sampling step it cuts off at, instead of which model timestep the switch should happen. If you use the default scheduler, this will work fine, but if you use the Karras scheduler it will be off. If you're using ZSNR, it will be very visibly off

If it was doing it correctly, then doing 0.8 denoise on a base of 50 steps would perform something other than 40 steps on the Karras scheduler (it would go either way depending on whether it's the zsnr schedule or not, the ZSNR karras schedule is very extreme and probably actually shouldn't be higher than 160 sigma max or so)

[neon oriole]

just look at the official comfyui workflow for SDXL

[sage reef]

is sd3 confirmed day 1 on comfy? ❤️

[shell tendon]

it's already been used in comfy afaik

[neon oriole]

once the weights are released yes

[sage reef]

nice

[neon oriole]

it's releasing with ComfyUI and StableSwarm support on release day

StableSwarm is getting a simple UI mode

[sage reef]

im assuming they will give us t2i and i2i first i guess and then we wait for controlnet, etc right?

[shell tendon]

excellent. are there going to be any significant changes to comfyui with this release, or can we more or less plan on being able to adapt current workflows as support for controlnet, ipadapter, etc gradually makes its way in?

[neon oriole]

the only big change is the latents are going to be 16 channels instead

[sage reef]

ah yea

[neon oriole]

so any nodes that treat them as 4 channels are going to have to be updated

but all core nodes should work

[shell tendon]

great, good to know! thanks

[sage reef]

i wonder how modular the t5 part will be, hopefully we can plug and play any t5 variant

[shell tendon]

are there going to be early preview weights kinda like with the API, or are we going to be waiting for the final ones? if you can say, of course 🙂

[neon oriole]

only the t5 version the model was trained on will work

[sage reef]

hmm maybe at the start until people find ways i guess

[shell tendon]

definitely very excited after some pretty extensive testing today - even though it's an old version, SD3 is blowing away ideogram and dalle3 and MJ6 on many prompts

[sage reef]

yea some awesome results so far

[shell tendon]

so great job over there at SAI

[sage reef]

[shell tendon]

there's some challenging concepts SD3 has pulled off today that dalle3, MJ6, ideogram, and SDXL all failed horribly at

[sage reef]

i just wish they released stable audio too :3

[shell tendon]

i would love that...

[sage reef]

and then we have the full set, video, animation, image, audio and text

@neon oriole what is PerpNeg based on? is that also a variant of like cfg rescaler? is there a paper on it?

[neon oriole]

yes there's a paper somewhere on it

[vague pond]

perpneg is more of a different approach to negative prompt than something like cfg rescale, CFG rescale is more about un-breaking CFG overexposure

[sage reef]

yea i mean it does have neg in the name haha

[vague pond]

https://arxiv.org/pdf/2304.04968.pdf this is the paper

[sage reef]

thx

[sterile gust]

So what's this core thing people have been talking about?

Is it a new model like cascade?

[ornate flame]

I made some improved FreeU values in forge based on FreeU v2: B1 1.05 B2 1.06

[sage reef]

freeu gives me cooked pics most of the time idk...

[lusty beacon]

https://x.com/StabilityAI/status/1780599024707596508

[ornate flame]

Yeah we know, it's been in #📣｜announcements for a while now

[woven kindle]

Hi

anyone know how to keep consistence in body?

for face I understand but for body is more tricky

[ornate flame]

a lora trained on a specific body type should work

[tropic frost]

small curiosity when it comes to lora training.

if you train a lora for a character and most of the image used have a certain object or piece of clothing in them, it tends to cause them to appear in all images you try to generate with the lora

right?

[ornate flame]

yeah like Ryuko Matoi from Kill la Kill often has her red sword in gens of her

Because the red sword was in a lot of the training data

[tropic frost]

hhhmm

wouldn't it be a solution to add more neutral images of character?

Like if you can't find any, if you cut off the character's head and paste it over a body without additional artifacts reduce this result?

[sage reef]

cut off the head? so violent! :3

[woven kindle]

you mean I must train a lora myself?

[ornate flame]

yeah probably, body type loras are very rare

[woven kindle]

ok thanks you , I will try to progress on that way

do you know a tuto about lora training ? never did before

[ornate flame]

If you search koyha ss tutorial on yt, that will probably help

[sage reef]

onetrainer seems good too

[woven kindle]

txs

[ornate flame]

or if you don't mind paying a bit for an easier process, civitai has a lora trainer

[woven kindle]

no I prefer to do it in local

[steep arrow]

Just use IpAdapter Style/Composition and an input image of what you want the clothes to be.

Works like a dynamic lora loader, having a lora for everything is too space consuming.

[sage reef]

yet another cloth model https://github.com/ShineChen1024/MagicClothing

or maybe it's a bit old idk, they say it's a branch

[rich kestrel]

I luuuuuv that my GPU gets so steamy hot and heats up my legs when I'm training a model

if it wasnt for the fact that its 2024 and 9/12 months are summer now

[lusty beacon]

Check out my simple web ui for SD3 https://github.com/PierrunoYT/stable-diffusion-3-web-ui

[karmic cedar]

sd4 when

when will we be able to communicate with donkeys

[pine path]

I don't know if there will be anything beyond sd3

[karmic cedar]

(there won’t)

it’s the last T2I according to Emad himself

[arctic sedge]

So the open weights of SD3. I'm a little bit confused. I have been hearing people saying we need a SAI subscription to access the weights. Does that mean you need a SAI subscription to get the weights to train the model? Where would all the custom fine tuned models go? To Civit AI?
Sorry if these sound like really stupid questions.

[karmic cedar]

weights will go up in a couple of weeks on huggingface. to use, you’ll have to do the typical license agreement thing

[arctic sedge]

I thought they meant through a subscription of SAI? Idk. I guess i just got the whole thing messed up. xD

[karmic cedar]

you can have a free subscription

you just have to be tied to the model usage somehow, esssentially

[arctic sedge]

But arn't they refering to their own service? When they say that. Not HuggingFace.

[karmic cedar]

you can access SD3 via the developer API right now, yes.

and to use the dev API, well

subscription

[arctic sedge]

Yeah. I just thought we needed to sign up to SAI's services to get the weights, when they come out. But if they end up on HuggingFace, I guess that is like normal.

[foggy halo]

SD3 RELEASED OMG!!!???? YAYAYYYYYYY 🎉🎉🎉🗣️🗣️🥳🥳🥳🥳

FINALLYYYY

I am a simple person...just want to generate pics using SD3 to access the state of the art features and coherency of the model

so how to access it for free?

[karmic cedar]

hi

[arctic sedge]

The weights are not available yet. We need to wait a bit longer for those.

[tight robin]

Hi everyone, I am Hans Paul cofounder at speedlegal

[foggy halo]

what is the meaning of weights?

[karmic cedar]

what is magnets?

real talk, this is why prices will be going up a lot soon: https://arstechnica.com/ai/2024/04/power-hungry-ai-is-putting-the-hurt-on-global-electricity-supply/

[foggy halo]

guys I am new to these terms...what is weights not really sure

[karmic cedar]

i recommend watching some introductory videos on youtube

[foggy halo]

i just use ai for simple ai imge gen

sure

[karmic cedar]

https://youtu.be/ZqRMHhkeylw?si=nD3GXOvA89PJ7hpp

looks like a legit one 🙂

much much easier to digest than explaining via chat.

[arctic sedge]

Sorry. Weights are the main files that contain SD3. Basically.

[stuck sun]

is there somewhere that breaks down the relative use of electricity across home/industry/manufracturing/AI/other uses worldwide

and military

[karmic cedar]

good question—but i wouldn’t be surprised if someone leverages an agentic AI to build a website tracking those particular stats…

would be a good idea for many different cases

[foggy halo]

Thanks a lot for the link!! 👍

[karmic cedar]

sure, hopefully it was a good help

[foggy halo]

oh i get it..thanks a lot! 👍

[karmic cedar]

hopefully by the end of the month we’ll have SD3 proper

whatever that ends up looking like 🙂

[foggy halo]

yes...even i need the full release
have seen some images not look that impressive th

i think that is an older version maybe??

[eternal ledge]

Playing with the API, I've definitely had some better non-text outputs with SDXL finetunes

[karmic cedar]

not sure, i haven’t really looked very closely at the images but so far the text placement has been very jarring

[foggy halo]

yea even i observed that

[foggy halo]

fr tho...wish they improve it on mass release

[karmic cedar]

i seem to remember the first images they demonstrated as being much better looking.

[foggy halo]

yes those images were what got me so hyped up.

[karmic cedar]

that’s why i suspect they’re…doing things to the model to check off more…mystery boxes related to social politics etc.

[foggy halo]

oh yes maybe

[alpine hull]

I'm disappointed

Especially at Stable Assistant

I have "access", as in access to PAY FOR IT

[karmic cedar]

🙂

it’s a business, at the end of the day.

[alpine hull]

It's a downgrade, that's what

a downgrade from Stable Chat

[karmic cedar]

oh, i feel ya.

I sense the doors closing on this company, doors that used to be much easier to open

[foggy halo]

guys is api key safe to enter in a huggingface space?

[alpine hull]

I legitimately don't even care about the open source part, but WHY would this be paid???

[foggy halo]

Seriously!! same thoughts

[karmic cedar]

because this is a company that needs money very badly

[alpine hull]

This is actually a cash grab

[alpine hull]

freemium

give, idk, a few credits a day

make the user pay more for more stuff

but no

[karmic cedar]

🙂

[alpine hull]

you have to pay $9 a month MINIMUM for a 12B LLM and SD3 access

[karmic cedar]

there is a cadence forming out of all these open source startups

[foggy halo]

🔥 SD3 is out now on ImagineArt with unlimited free generations" is this true!?!?

[karmic cedar]

where they develop and develop, and then….smack

compute costs skyrocket

investors question the future

etc.

[alpine hull]

they couldn't even BOTHER with using SD3 TURBO in it

and SD3 Turbo isn't even a downgrade

it's just cheaper

right?

[sage reef]

idk that website is fishy to me

[steep arrow]

It is probably a downgrade in quality (might not be that noticeable), but not necessarily diversity or composition coherency.

I don’t think we have anything to go off of specifically yet.

[alpine hull]

bruh at least give us free access to stable assistant for a little while :/

doesn't make sense

hell, im not sure if it's even a limited beta

[sage reef]

sad kirby

[alpine hull]

afaik, everyone here can access it, right?

[steep arrow]

You will not catch me defending SAIs deployment strategy lol

[foggy halo]

omg how do you think it is? better to be safe

[sage reef]

this is why Emad left

[alpine hull]

there are less stupid ways to launch smth like this

[karmic cedar]

businesses need profit

emad is building the future

[alpine hull]

this isn't the way to make a profit

[sage reef]

i mean im not using that website, so no idea, it's just fishy whenever you see some new stuff and then combined with the word "unlimited"

[alpine hull]

I'm not paying $9 for a 12B MODEL

[karmic cedar]

no smart person is

[foggy halo]

ah yes correct

[karmic cedar]

unlimited is starting to become a vibe word

[sage reef]

the best I can do... is $2 (Pawn Stars moment)

[karmic cedar]

lol

[alpine hull]

I'd probably accept that

[sage reef]

deal

[alpine hull]

I'm not one to complain about stuff like this

[low moon]

Holy shit...

[sage reef]

the shit is holy

[low moon]

so checkpoints when.

[sage reef]

may 10

[low moon]

oh

[alpine hull]

usually i can just

accept it

but this is like openai replacing chatgpt with this paid only downgrade for 3.5 only

[steep arrow]

I might pay for midjourney for a month 🤔

Zero chance I pay to help test.

[sage reef]

these days they are called closedai 🙂

[low moon]

Its not the 9 bucks fr the weights its the hours and gpus for trainign i dont have

[alpine hull]

instability ai

[sage reef]

it was unstable when Emad left

[pearl ocean]

how do you feel to pretty much be the next evalution of man?

[low moon]

Anyway its too cutting edge for me to bother with it right at this moment but it's nice to finally see the SD3 channel there 😄 I knew they day would come.

[sage reef]

too cutting edge for you?

[cedar salmon]

the sharp credit card edge

[low moon]

But where is the stable cascade shtroom :<

[low moon]

3 months after the community had time to refine it- implementation into fooocus/forge/A111/comy/controlnets. Then it won't cut me anymore probably. It's too hot now, sizzling hot. It needs to cool, settle into shape.

probably some optimization too - not sure 12gigs is enough for this

[karmic cedar]

different architecture made by a dev team that has left Stability

it’s the odd cousin of stability models now

[low moon]

SDXL wasn;t exactly welcomed with open arms too

took like 6 months to take off

[karmic cedar]

but when it did drop, it dropped quickly and without pomp

there’s been a lot of pomp and little circumstance this time around

[low moon]

yeah

subvert your expectations

[steep arrow]

SDXL didn’t start getting good in a unique way until some more recent nodes dropped, its just now hitting it’s stride.

[low moon]

exactly

[sage reef]

yea sdxl getting real good these days

[karmic cedar]

the latest text examples from SD3 are either not being handpicked carefully or the quality of the model has somehow changed versus what was posted with the paper

[low moon]

yeah im using it almost 100% now it just beats 1.5 now in all aspects

[karmic cedar]

1.5 with ELLA is pretty nice though

[sage reef]

not sure about all aspects

[low moon]

whats ella

[karmic cedar]

it just doesn’t work always with other finetunes

[steep arrow]

1.5 with Ella is nice, but that should only be temporary that we don’t have an SDXL version.

[sage reef]

i mean if they dont release sdxl ella, maybe some other team will release something similar for sdxl

[karmic cedar]

https://github.com/TencentQQGYLab/ELLA

[steep arrow]

The Proteus model guy is working on SDXL ella

[karmic cedar]

but if you can have better semantic control with ELLA, there’s really nothing wrong with taking your output to SUPIR for a quality upscale.

it’s just extra computation, at the end of the day

[steep arrow]

Ella still looks really really SD1.5ish.

And Ella has a lot missing from Clip.

[karmic cedar]

i think that’s because it hasn’t been demonstrated with anything beefy, just the base model

if it were fine tuned to a more advanced / refined checkpoint, perhaps

[steep arrow]

Would have to be retuned in a major way, ella doesnt have the Artist/actor labeling that we have in clip.

[karmic cedar]

this is true

[pearl ocean]

the question is, will we ever get SD3

[karmic cedar]

there are limits to the model itself

i think we’re going to get something more like vanilla cascade

and then we’re going to find out that extending it is much more of a bitch than what we’re used to now, etc.

[steep arrow]

Ella first pass, Clip finish pass 👍

[sage reef]

may 10

[low moon]

I thought SD3 is sposed to be easier to train for

[karmic cedar]

i hope so!

[low moon]

like cascade

[karmic cedar]

i hope i’m wrong

[low moon]

the sausuge wurst tech

[karmic cedar]

lol

i like it too.

just that, as far as diffusion models go it’s more like your distant uncle you don’t talk to very often

he’s like, cool and stuff, but you don’t really like, have stuff to talk about

[sage reef]

cascade is tiny sausage, sd3 is big sausage

[karmic cedar]

indeed

[pearl ocean]

😮

[karmic cedar]

we’ll call it the Emaddening

[pearl ocean]

meant to have more detailed and do more of what you say

[steep arrow]

If the small sd3 can give good starting composition and fix text with inpainting/detailing, I will be a very happy camper.

[karmic cedar]

as long as it doesn’t look like someone’s amateur distortion tool work in photoshop, which is what a lot of it has looked like so far

and that is jarring to me

being used to text looking more baked in

[steep arrow]

Lmao, I think that is probably down to prompting.

I have seen the “copy paste” looking text, but it also does styled text so…. At worst this will be addressed with finetunes.

[karmic cedar]

yeah, i’m taking a closer look at the #sd3 channel feed and i’m sure you’re right

[crude notch]

i dont have any numbers, but according to them, it would be easier to train than expected

(e.g. compared to an 8b model)

[steep arrow]

Isn’t training supposed to be applicable for all of the model sizes?

Or am I hallucinating that?

[crude notch]

eh, sort of

flow-rectified is not diffusion

[sonic shard]

How do I use the bot again to generate?

or is there a bot I can add to my server if anyone knows?

[red lynx]

It is offline

[livid blaze]

Hey, all. Complete moron here. Is there a specific discord for the automatic1111 / stable diffusion webui client, or is it worth poking around here?

[sage reef]

i mean folks here use all sorts of things

we try to help each other 🙂

[sonic shard]

know of any other bots anywhere?

[shell tendon]

go ahead and send me a prompt in #🏞｜general-with-images

sometimes ppl get lucky and are happy with my creations

[verbal osprey]

Not to my knowledge....but if you find it let me know. I understand why people like comfyUI for the flexibility but I just prefer auto1111

There was supposedly a fork that includes the front end look of auto111 but the backend flexibility of comfyui. I bookmarked it somewhere but I forget the name

[livid blaze]

Roger that. I might end up swapping, I'm sure I've screwed up something with the install.

[shell tendon]

forge

[verbal osprey]

My auto1111 install was messed up for a long time. The lora tab was just gone. I think it was because of the extensions I had installed. I did a fresh install and it all works now

[mortal delta]

I dont know if this is off topic or not,
but how can i save up money for a better pc/gpu to use with ai and other things if im not allowed to have a job yet due to highschool?

[livid blaze]

When I was a kid, I did Swagbucks stuff. Kind of a shitty predatory service, don't know if it's still in business, but you could automate it to get a trickle of amazon gift cards. Otherwise, I bet Amazon Mechanical Turk is still around.

[mortal delta]

Ive actually looked into swagbucks and such sites but it seems supper grindy theses days compared to back in the day also i dont have time to grind due to school and life. also what is amazon mechanical turk?

[knotty rivet]

how do i save my settings/make tempalte in a1111.
like my tiled diffussion and tiled vae. settings
i tried the png info but it only save my promps and basic stuff

[tepid finch]

Anybody making any money with your creations? Just curious how many ways you can make a living off Ai creations.

[wise bear]

Advertise locally and find a group that does wall vinyls printouts, the cost of the vinyl would be the only expenditure, which can significantly reduce cost, or compete with designs on the market.

[mortal delta]

I wish but it seems oversaturated and youtube dumb gurus are not helping.

[tepid finch]

Keep em coming!

[mortal delta]

I have an idea, what about concept art for others projects or freelancing?

Are you thinking of trying to make money with ai art, and what model do you plan on using?

[tepid finch]

I have a 48in plotter at work

[tepid finch]

Yes i tried selling Ai created tshirts, canvas prints, etc on shopify but Noone was buying. So I shut it down

[mortal delta]

Shopify is pricey, ive also been thinking about making money with ai but im scared to start and i have no plan on what to do.
also try to stay away from youtube gurus....

[opal hedge]

sd3 without fine-tuning feels no better than sdxl on release

Can't wait until it's usable

[tepid finch]

Don't be scared just dive in. That's what I did.

[mortal delta]

But just like you i need a ai sidehustle idea, possibly.

[tepid finch]

U can sit around all day wanting to do something but nothing will happen unless you just dive in. Learn as you go. Just do it

[mortal delta]

Ive thought about making comics but that would be un-consistent.

[tepid finch]

Y inconsistent ?

[mortal delta]

I dont know how to get the same person with stable diffusion, ive asked a few times, and i was told loras and seeds is the way to go for some but things might have changes sense then?

[tepid finch]

Yah keep the seed the same. I'm coming from midjourney so I'm relearning

U Def can get consistentancy in sd

[mortal delta]

I guess ill try seeds again, ive tried them with online services but not locally. also im not sure how to market (my bloodline is bad at marketing).

[tepid finch]

Some guy was paid like $230k for like 2hrs of Ai work. Hired to do game cards like pokémon

[mortal delta]

thats crazy.

anyways what method are you planning on using to make cash?

[tepid finch]

Sorry 90k https://www.pcgamer.com/games/card-games/champions-tcg-ai-artist/

[mortal delta]

and failing is part of the journey.

[tepid finch]

Don't know yet hopefully I get the idea in here lol

[mortal delta]

We all want that at some point in terms of ideas from this server, i might make a comic and try to sell it locally.

[sterile gust]

Well SD3 might be as bad as SD2 but the question is, can it be fixed?

[tepid finch]

First comic made entirely by ai

It's all in the sales pitch

I could have made some sales with my site but I didn't push ads as much as I should have and I was dissatisfied with the quality of print on demand for the shirts.

[shell tendon]

it's way more powerful

[tepid finch]

I was thinking of making album covers for bands .

Children's books

Short youtube anime series. There is so much u can do

[mortal delta]

Might work but as i said the market is oversaturated with ai products.

[tepid finch]

Redisigning company logos is another

$500 a pop

[rose sedge]

is stable diffusion 3 not going to be released open source?

[autumn vector]

I hope this message finds you well. I wanted to share some thoughts regarding the pricing structure for API credits on Stable Diffusion 3 (SD3) in comparison to other platforms, specifically Ideogram.

While I have no intention of disparaging SD3, I believe that the cost of API credits for generating images is relatively high. In contrast, Ideogram offers a more cost-effective solution, providing 1,000 images for $20.00, along with text generation capabilities. Using the API system on SD3, 1,000 credits would only yield approximately 154 images, which I find to be a significant difference.

As a supporter of Stable Diffusion and its open-source philosophy, I understand the importance of sustaining the platform. However, given that SD has traditionally been open-source and free, I feel that the current pricing structure is quite steep.

Please note that this is merely my personal opinion, and I wanted to share my perspective with you. I would be interested in hearing your thoughts on this matter and discussing potential solutions that could benefit both the users and the platform.

Thank you for your time and consideration.

[tepid finch]

What is everyone's reason for creating images with Ai? Mine is it's intriguing and want to profit from it.

[narrow kernel]

At generating what? Lemme guess... Anime and realistic anime waifus? That's like 90% of the complaining I've been seeing and hearing lol

[narrow kernel]

Base sdxl is still phenomenal, you're just biased by the overtuned models that make close up portraits with the same 5 face shapes

[twilit blade]

https://kailashsiri.medium.com/the-impact-of-over-reliance-on-ai-balancing-technology-and-critical-thinking-88f72f6f8298

[tepid finch]

OK I'm jumping on xbox for some mw3. Catch yall later and brainstorm some more

[sage reef]

ah so I just saw the reddit post by a SAI staff member, Core is SDXL Turbo finetuned and it's basically their best model at any given time or better than any new base model, which makes sense why it gives higher quality pics than sd3, it's probably finetuned like crazy

so it wasnt sd4 alpha... sigh... :3

[vague pond]

but what if I don't want to think about whatever matplotlib plot I'm trying to generate

[pearl ocean]

SD5 when

[sage reef]

[trim magnet]

cant wait for SD69

[rotund remnant]

Can anyone successfully use the SD3 API? I got 500 response code

[karmic cedar]

uhhh well basically i’m using SD3 to create a ridiculous AI themed game show with the same spirit as Takeshi’s Castle

actually that’s just idea #7218256

[peak orchid]

I've heard that SDXL 0.9 is the best for finetuning, because of some the tricks they used to make the SDXL model outputs look better. 0.9 is what model trainers have been using

[peak orchid]

The new head of the US regulatory agency for regulating AI would say no open SD4 and above should be allowed: https://arstechnica.com/tech-policy/2024/04/feds-appoint-ai-doomer-to-run-us-ai-safety-institute/

[trim magnet]

he cant stop me though

[sage reef]

omg i remember watching that, but it was called MxC or something right? takeshi castle im assuming was the original name but translated to english

[peak orchid]

In the past they have a period of exclusivity on their services before releasing the full model, and its probably going to be the same here as they are at risk of going under

[peak orchid]

he can certainly hurt you though

[karmic cedar]

yeah, they actually released a remake of the original last year and it’s still really entertaining

[trim magnet]

he cant,hes like 90 and can barely walk

[sage reef]

nice

[karmic cedar]

https://youtu.be/kzL7vjAwg5M?si=aTwtx3vcSYP3sdkt

[peak orchid]

if you use American platforms, services, software, etc... then you will be impacted by his actions

[trim magnet]

not really ,only use Saudi Arabian Cloud services

[karmic cedar]

locking down technology on account of fear scenarios is a practice of convenience for the powerful

then it becomes about the having and the eating

the cake becomes a lie to us, and two truths to them

it is a time of great innovation

[onyx notch]

Hello! I had a quick question. Involving SD and Animatediff I guess? I'm only familiar with generating static images using Automatic1111, it's an awesome tool. I found a tiktok video on my fyp and honestly just want to know how this account makes these videos, the tools they use. I can't post a link but can send it in a DM. I'm just looking for some guidance tbh

[karmic cedar]

you can’t share the link publicly?

[sterile gust]

I have seen people trying it out and comparing output results to DALL-E 3 and so far i can see that SD3 has a lack of understanding of styles, composition might be there and better but the variety of styles has definitely been lost in comparison to SD 1.5 which is why I wonder if it can be fixed via LoRAs and finetunes?

The compassions i saw were low poly/N64 style and Rembrandt painting, both of which SD3 seemed lacking.

Personally i don't like anime so i try to avoid generating anime as much as possible

[onyx notch]

Rule #5 says no external links?

[karmic cedar]

oh—i guess i interpreted that as being about self-promotion, but you’re technically right.

[onyx notch]

no I have no affiliation with the poster of the tiktok in reference

[karmic cedar]

as far as tiktok vids…i feel like the majority of them have been some kind of animatediff workflow

but anyone feel free to correct me

https://comfyworkflows.com/videos anything on that page remind you of it?

[onyx notch]

Not exactly. Some buddies and I play DnD and the results from the tiktok are so awesome that I find it hard to believe there isn't any post processing on them. I'd be happy to send you the link if you'd want. I guess I just want an expert to be like "yeah this is edited" or "nah this isn't possible with SD"

[acoustic pelican]

where can find all prompt var or attributes？

[onyx notch]

(I'd love to make something similar and implement it into our campaign basically)

[trail lion]

there is sure to be a contingent that's super critical out of the gate, I'll personally reserve judgement, but some of the images I've seen posted today in the images room does not disappoint

[sterile gust]

I have been hearing about core not sure what that is

[trim magnet]

send one here

[onyx notch]

In this chat?

[vague pond]

there will be people saying for years after the release that the model is complete unusable trash because of "censorship" (the model cannot produce anime titties on day one)

[trim magnet]

yes

[onyx notch]

https://www.tiktok.com/@voidstomper.ai/video/7355215076783131935
This is the only video from the creator that I care about really, the others are not as impactful.

[trim magnet]

text2vid with animatediff and some postprocessing

[onyx notch]

Thank you so much, I love you. I'll look into those.

[sterile gust]

Wouldn't that just show that the model lacks understanding of anatomy both photorealistic and drawn?

Maybe just drawn

But i have seen sdxl having touble with photorealistic anatomy

[karmic cedar]

models are going to lack that kind of contextual understanding anyway. they generate from vector alignments, paths followed.

[sterile gust]

Then how come models be fine-tuned towards generating better anatomy? Does that have to do with the dataset?

[karmic cedar]

I think they’re exposed to more body pose configurations through sheer repetition

via the fine tuning

but that’s not providing them any anatomical context, just more pose context

or rather positional context, if you want to simplify it more

[sterile gust]

Is that why models can't do hands?

[heavy lark]

They can. Just not this one

Hah. Yet.

[sterile gust]

So you are positive it can be "fixed"

[karmic cedar]

with enough developer attention, it could eventually be yes

the sky’s the limit. people just gotta figure out where they’re looking when they look up

[opal hedge]

I wonder if SAI had released SD3 a month ago if the fine-tunes of today would've been 10x better than what we get with the API

[heavy lark]

If you have a look at lykons twitter feed, you'll see that even the latest internal version is much better than what they're serving in the api

[opal hedge]

Yeah, it's odd that they released the outdated version

[fervent thunder]

hi

is there a way to do constant objects or a lora of an object?

[karmic cedar]

how to date LORA?????????????????

[opal hedge]

Send her witty text messages

[karmic cedar]

😦

… f o r f r e e ? ? ?

[slow tundra]

hi there

is there a channel for job opportunity

[sage reef]

what kind of job?

[slow tundra]

prompt editing lora creating

[slow tundra]

are you free to work?

[trim magnet]

no,i am free to sleep

[sage reef]

oh you mean you are looking for someone to make you a lora?

[slow tundra]

not only lora

will work on a website to edit prompts to make user experience better

long time job per month salary

[honest mica]

What user experience? (And no, i have a job).

[slow tundra]

you will create prompts

by using english language

[terse dust]

Where to apply?

[slow tundra]

we have a panel for moderation

you will check update create new daily

there is a monitor page where we take users requests where a moderator manually checks updates prompts

[slow tundra]

you can text me in dm

[wet dawn]

hi

[karmic cedar]

I know you

[trim magnet]

miss

[pearl ocean]

what about me

[stuck sun]

has anyone been getting a sense of how easy/hard SD3 will be to train?

[sage reef]

let me activate my 6th sense

[pearl ocean]

7th sense when

[sage reef]

may 10, the same day sd3 weights release

[vocal leaf]

Whats the app again with models

I mean site

[sage reef]

huggingface.co

[pearl ocean]

imagine hugging a face

[sage reef]

i will hug you instead

[placid oak]

Hi

[karmic cedar]

I know you

[next dawn]

Virtual assistant available here ✌🏻

[wide tendon]

maybe it turns out Cascade was a better architecture than SD3?

because Cascade on release was better than SDXL on release

how trainable is SD3? That's the most important thing. SD1.5 still beats SDXL sometimes because of it's community of folks with average GPUs

I hope it's not like "SD3 heavy and SD3 light... because I want the poors to fine tune SD3 heavy. Cascade was perfect for that, it had a tiny middle-block model that was easily tunable, and then the larger upscale/pseudo-VAE that would need more VRAM to finetune, but always worked together

[fresh vigil]

sd is not going to be open source?

[karmic cedar]

Weights going up on HF in a few weeks

[fresh vigil]

Ok, thank you

[wide tendon]

so hard to wait 3 weeks

[pearl ocean]

go into a coma

[wide tendon]

OTOH SDXL + ELLA + SelfAttentionGuidance + FreeU + PonyFineTunes +IPadapter... maybe SD3 will take months to catch up

[karmic cedar]

Watch fallout

[hot kettle]

Otoh?

[wide tendon]

no, I'm a s addict. Either I'm training or out IRL getting a boyfriend.

[hot kettle]

Since cascade was already mentioned, does anyone else get blurry outputs?

[karmic cedar]

Blurry, no

Polished to the point of being vapid, yes

[obsidian roost]

When I download models from civitai, how do I know which folder to put them in?

[unkempt hatch]

I found the skin-tunes had a blurred look. I think it just needs more FineTuning, and some heroes would tune the model B as well.

OnTheOtherHand that probably won't happen because of SD3

[karmic cedar]

It depends on what your setup is

[hot kettle]

I was using fp16 with comfyui and couldn't get anything not blurry

[karmic cedar]

Is BP16 an option?

Errr

Bf16

[unkempt hatch]

I wish the MixOfExperts node had been maintained, and we could use SD15, SDXL, Pony, Cascade, SD3 all at once, working together in a genius model

[hot kettle]

My bad it was bf16

No idea why I said fp16

[pearl ocean]

BF2042

[karmic cedar]

That’s a lot of heavy lifting for a node

[karmic cedar]

Put up some examples if you have any

[unkempt hatch]

that's what it was designed to do though, right? Not saying it would be easy, but amazing

though not sure how that'd work with e.g. a controlnet that's meant for SD15 being used while all the other models are working.

but

[karmic cedar]

Deconstructing the model architectures into a generalized latent space just seems really taxing, resource wise

[unkempt hatch]

I never really understood how MixofExperts worked at all.... like you got a model that does better hands, how does the MoE node know that it's the best hand model? And how does it make that hand consistent with the rest of the image drawn by other models

[unkempt hatch]

Ah, and MoE was doing more multiple finetunes of the same architecture, rather than multiple architectures

still, I just imagine how awesome it'd be

[hot kettle]

I'll make some now

[karmic cedar]

I’d like to see more cross-configurations and especially alternatives to the CLIP encoder

Like what gets used in ELLA

If you could swap out CLIP with an LLM like Llama-34B, you’d be able to prompt a short story into an accurate image 😂

[small frost]

i need some help with comfyui

[past charm]

[sweet chasm]

[keen kettle]

SD3可以下载吗？

Can SD3 be downloaded?

[sudden ruin]

No it was an API only release so far

[pearl ocean]

🥱

[small frost]

can you pls help me out with comfyui

[sudden ruin]

#🧣｜comfy-ui feel free to ask there

[sudden ruin]

Trust me I want the weights as much as anyone else

[small frost]

i have asked over there

[pearl ocean]

I like local, it just makes images faster, and I don’t have to worry about tokens either 🙂

[sudden ruin]

Same

[vale steeple]

eh I mean it is always like that

SDXL also release its API first

before 0.9 release

[sudden ruin]

Yep nothing unexpected

Also a good thing, people with the API get to test an early version of the model, give feedback and help improve the final model

[vale steeple]

I didnt expected them to actually release API for SD3 Turbo in the same time

But I guess that make sense

[bleak matrix]

Good morning, everyone! I hope you are all having a lovely day!

[uneven obsidian]

Good morning to you 😊

[cinder salmon]

Hello everyone I have a question can I create here some ai images? Or it's possible only on website.

[uneven obsidian]

Maybe it's not supported here

[cinder salmon]

ok thank you

[uneven obsidian]

Please let me inform of u find any such source, cause i am looking for the same

[fervent thunder]

Hey guys, i want a partner for my project ~

[bleak matrix]

Good morning, good morning! What are you up to today?

[fervent thunder]

I'm not getting the main couse of not generating the image

[bleak matrix]

The bot is not currently active! You can see the status on #1047610792226340935 for more information on that. In the meantime, feel free to check out https://beta.dreamstudio.ai/generate or https://clipdrop.co/

[fervent thunder]

Link is safe ~

[cinder salmon]

Thank u for informing!

[bleak matrix]

Not a problem!

[fervent thunder]

@bleak matrix can you pls look at the repo, if you are free

https://github.com/Priyansxu/vega

[trim nymph]

anyone an idea what causes this inside auto111: RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 1152 but got size 992 for tensor number 4 in the list.

[twin skiff]

does anyone know any professional Ai designers?... im looking to hire one i would pay $300 to who ever can put me in contact with someone and they get the role please 🙏

[frozen helm]

Overpaying

[fervent thunder]

Nvm

?

[sudden ruin]

Please use #1092446741984444416 and stop advertising your site here

[fervent thunder]

Okay sorry, although the site is not working so why would i 😅

[fervent thunder]

I want support btw

[sudden ruin]

This discord isnt for support by SAI, you can get help from the community here for pretty much everything. If you need help with the API please contact their support directly

[fervent thunder]

Server link?

[sudden ruin]

https://stability.ai/contact

[brisk tiger]

Can anybody help me what exactly Discord about?

Is it more like a chat where everyone has a discussion about particular topic?

[frozen helm]

discord is discord

[trail fox]

AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

[vale steeple]

@sudden ruin @bleak matrix

are you racist to the chat?

[low moon]

hello mods?

[cerulean kraken]

TY

[vale steeple]

thank you mod!

[low moon]

Why did you axe the Cascade channel 😦

[cerulean kraken]

That was a hell of a chat to log onto 🤣

[late holly]

can anyone tell me why do i keep getting bad generations with newer models?

I've recently picked up stable diffusion again but after downloading a few newer models, it seems like i cannot get good generations as easy as it used to be back in the day

body proportions comes out wonky, the interpretation of prompts are really bad

[rich kestrel]

bc newer models suck at anatomy...

stick to 1.5 custom

[rich kestrel]

bc cascade became irrelevant 2 weeks into its release l0l...

[cerulean kraken]

I need some help.. has anyone used Ollama with ComfyUI?

[sage reef]

using game of thrones terminology, cascade was a bastard 🙂

just like sd2.0

but i guess research is still research.. so they are going "towards" greatness eventually

[blissful ore]

is it possible to improve images when inferencing a adapter+base model?

[karmic cedar]

Cascade is like a Volvo while SD is like a Honda Civic lol

[peak orchid]

The layoffs are brutal, like why get rid of the most competent researchers still working at the company? Its a complete bloodbath

[rose sedge]

We can help you. We work with AI and are video editors and graphical designers.

[green prawn]

Hey, what is the difference between buying membership and buying credits seperately?

[narrow kernel]

Yeah I've used it a ton and even made some custom wrapper nodes. What's up?

[eternal ledge]

Membership is for self-hosting the models in a commercial setting iirc

Credits let you use the APIs/tools that StabilityAI publishes

[rich trench]

Hey everyone my company has one open position to work with AI.
https://apply.workable.com/palazzo-inc-1/j/877AE4A35A/

[rich kestrel]

5yrs python experience lol...

only thing I hate more than jobs is job hunting and thx for reminding me why

[rich trench]

@rich kestrel hahaha, you don't need to have it, it's because our workers are write using python.

[cerulean kraken]

nevermind I figured it out, but thanks!

[eternal ledge]

That stuff is written by HR at most places anyway, not the people who actually understand the position. I've seen more than a few that asked for more years of experience in a technology than that technology had existed (e.g. "3 years experience in using SDXL" - not that I've seen this one, but it illustrates what I've seen)

[sage reef]

the 5 years of experience is like their ideal candidate, they usually accept below that

but then again, depends what they mean by "experience", cause if experience is strictly in some professional setting, then maybe out of luck, but if they see you made some python projects over the course of couple years, like on github, i guess that counts too. idk.. i hate job searching too, thats why im working on my own business so i dont have to work for someone ugh :3

[tall mantle]

I've created a cool tool using Wordware that runs Stable Diffusion 3 with both an input prompt and a refined prompt. I'm really excited to hear what you think! Give it a try and let me know if you find it useful—I'd love your feedback!

https://app.wordware.ai/r/b137b2f5-a971-420c-a594-4f6350c24fa5

[rich trench]

I understand I think the part of SD is more valuable.

[gusty oriole]

[eternal ledge]

This, I also tend to personally think of "experience" as something a little more precise and valuable than "tenure", but job reqs always are really stating tenure when they say experience. Most people do not turn every year of tenure into a year of experience, which is part of why I generally recommend one apply for a job if they meet even half of the requirements.

I'm really enjoying working at the startup I'm at right now, but if I had to be searching in the current economy, there's a pretty high chance I'd be looking at starting my own thing.

[sage reef]

i wish there was a startup i could join, i love joining and building a company from scratch with someone, but no luck so far 😦

[eternal ledge]

Yeah, I wish we were in a spot to be hiring 😦 seems like even when anyplace hits Series A now, they're not being encouraged to increase headcount so fast as they once were

[sage reef]

yea

[teal pagoda]

Testing atm

[honest musk]

Quick merge of the Forge and UI-UX forks.

All credit goes to lllyasviel and anapnoe.

https://github.com/codaloc/sdwebui-ux-forge-fusion

[teal pagoda]

The problem is that the image is blurred.

[pine cedar]

What would be a decent it/s for a hires fix 512x768 x2 upscale 20 steps 0.25 denoise?

[teal pagoda]

like maximum 1 min for that

[pine cedar]

I have a 4090, Ryzen 7950X3D etc

I get like 2.5

Feels slow

[teal pagoda]

what?

[teal pagoda]

for real

Automatic1111?

[pine cedar]

Forge

But same on "vanilla"

Not 2.5 min btw

2.5 it/d

s

[loud aspen]

Does switching from Automatic1111 to ComfyUI make any difference in the quality of the images generated?

Like does it make it better?

[tall mantle]

hmm.. good feedback, let me think it through

[eternal ledge]

It can, if you use a better workflow for your particular use case. But just switching won't automatically make things better

[normal rock]

hello everyone, can anyone direct me to how to learn and use RAGs in my LLMs. I am new to machine learning so simplicity will be appreciated. Thank you

[fallow garden]

Where can i find sdxl temporalnet controlnet?

[sage reef]

im personally using the combo lm studio with anythingllm, lm studio gives you text+vision and anythingllm adds the document support, so basically all is there

[normal rock]

thank you , and I also wanna use IP adapter for my project, where I wanna mix facial features (which I can choose myself) of two persons. Can you direct me to some useful resources as I tried using styleGAN3 but didnt worked.

[foggy halo]

guys how to use sd3 for free now?

[eternal ledge]

It's only provided via API atm, for about $0.065/img, so, if anyone's offering it for free, it's for a very small number of gens to try to capture you as a long term or subscribing user

[foggy halo]

oh thanks for the info...i wanted unlimited free generations as in like huggingface where no complex code is required

[eternal ledge]

Ah, unlimited free generations probably won't exist until they release the model weights "in the near future" per #📣｜announcements message

[sterile gust]

Is there a way to try SD3? Like online? Been looking for a HF space but nothing comes up

[rich kestrel]

ya its on imagine.art

[sterile gust]

Does the phone app also have SD3?

[rich kestrel]

havent tried the app sowwey

[orchid quiver]

Hola!

I keep having this error with Animatediff
AttributeError: 'NoneType' object has no attribute 'batch_size'

Any help? thanks in advance

[trail lion]

that's usually the version of a package is not new enough, from that we wouldnt be able to do a deeper dive, but essentially you might be able to determine which software package is complaining and just update that to a newer version and it'll probably go away

[slim scarab]

when u see a sick image being generated but u put denoising strength to 0,8

[eternal ledge]

Doesn't look like it on Android

[opal hedge]

Which paramter size of SD3 is available via API?

It surely isn't 8B right?

[pulsar venture]

Hi, can someone explain me what is the schedule type?

[teal pagoda]

Nah

[unkempt depot]

Hi, I want to create AI images and videos that feature celebrities and celebrity faces. (Links to example videos of what I want to achieve: https://www.youtube.com/watch?v=XfkgiXaaCY4&list=LL&index=2 , https://www.youtube.com/shorts/ASkd9Oxk1Eo ).
-1st question: Is there any Stable Diffusion Model or generally any AI that is able to create images of celebrity faces? (especially like shown in the first linked video)
-I think I could recreate the dancing animation in the second linked video with AnimateDiff from Stable Diffusion, but how would I integrate celebrity faces onto my animated characters? Therefore my 2nd question is: How could I create videos like shown in the second linked video?
-3rd question: I want to create AI images of family guy cartoon characters, ss there any Stable Diffusion Model or generally any AI software that could do this? Btw, before someone mentions it, for other softwares: Microsoft copilot is able to generate some characters like Peter Griffin, but for example isn't able to generate other less famous characters like Cleveland or Quagmire.)

If anyone knows even only a single answer to one of my questions, this would help dramatically, thus help is very much appreciated 🙂

[rich kestrel]

just use a lora

for characters

[warm junco]

Hey, you can use extensions and tools for face swapping to get that

For specific characters look for Lora's of them

[trail lion]

context? are you referring to the learning rate scheduler for training? how can you ask a qusetion like that with absolutely no context about what program, use case, context

[warm junco]

I guess he means the Schedule type menu in auto1111 since its appearance in version 1.9.0

[trail lion]

auto1111 would have been a decent hint I guess

reminds me of that hobbit scene when bilbo asks what's in his pocket

[unkempt depot]

Oh ok, that would mean that I would have to download a lora for every celebrity I want to create and edit, right? There basically is no way to have access to how celebrities or cartoon characters look within Stable Diffusion without downloading the Lora's of the celebrities and cartoon characters, correct?

[warm junco]

No not every character needs a lora.
The more famous the character the higher is the chance that a model can recognise and create it.
Lora's help to get exactly the character you want.
There are a lot of character Lora's available already.
If you need a specific character then you can also train a lora on that yourself.

Checkout civitai.com for models and Lora's

[unkempt depot]

Oh ok, will do that, thanks for the help. 🙂

[stuck sun]

whats the latest on SD3? Any idea when model is being released for local use? Or how hard it will be to train

[sage reef]

may 10

[stuck sun]

but thats too late, i have a bomb in my chest, with a timer set to EXPLODE on may 9th 😦

[sage reef]

welp

[karmic cedar]

Okay, Game of Thrones AI challenge: which big AI companies are which factions

OpenAI = House Tyrell for sure

Stability = House Stark, since, y’know, winter is apparently coming for them

[final schooner]

Google = House Baratheon

[sage reef]

lol

[low moon]

AI will kill the ecomony and money is dead so soon you won't have to work.

[karmic cedar]

Intel = Lannister

[sage reef]

Hotel = Trivago
oops... wrong game

[red lynx]

Meta = House Martell

[pearl ocean]

SD3 = Soon

[gusty oriole]

[boreal mauve]

omg ... I'm testing SD3 on clipdrop... that's barely an improvent from XL. Any secret how to prompt?

[fervent thunder]

what are you弄啥嘞

[final heron]

When changing just the face, the result often come out head too small and neck too long, you guys have any tips for it?

[stuck sun]

huh base version of SDXL did stuff like this

[heavy lark]

If you think it's barely better than sdxl, you're giving it really boring requests. Ask for more stuff

[shell tendon]

"Disclaimer: I have no idea what I am doing" great... 🤣

[gleaming pilot]

So i saw this video about making an AI influencer and it made me wonder how easy it might be. how much would i have to study to make something consistently and believeable?

[heavy lark]

why would anyone want to make an ai influencer except a creepy person?

seriously the stuff that gets posted by bots on here constantly

[gleaming pilot]

whatever dude, dont judge

[heavy lark]

why not? it's either that or the constant requests for making images of soft porn constanlty.

[gleaming pilot]

everyone has reasons, i dont owe an explanation

[trim magnet]

for money, what else do u think,use your 🧠

[heavy lark]

why would anyone want to watch a fake person with a fake voice say fake stuff on the internet?

it's all just cringe

[gleaming pilot]

maybe i just wanna make pretty images of what i think my ugly ass looks like, asshole

[trim magnet]

thats your opinion

[trim magnet]

ppl watch clowns on reality tv like the kardashians all the time

[gleaming pilot]

like i said, dont judge

[trim magnet]

or your politicians

[heavy lark]

it just strikes me as yet more tiktok style thirst videos.

[gleaming pilot]

I probably shouldnt have even asked for help in a public forum

[trim magnet]

do not care the reasons only thing ppl care is that tiktok/instagram/twitter pays u if u have lots of followers

[pearl ocean]

S D 3

[strange ridge]

大家好呀

[sage reef]

nani

[strange ridge]

discribe

[boreal mauve]

I figured in the meantime... I just got weird results with hands and awkward deformations

[heavy lark]

no excuse for that one, it resets the clock on some of the stuff that sdxl finetunes have fixed.

people hold things better, but if you ask to show hands, they'll still have 3 fingers.

[sage reef]

show feet

[heavy lark]

i've seen less doubling of fingers though.. it's just less, but those less fingers look "good"

[boreal mauve]

True, hands are bad. But it's very variational. Huge plus. A d very artsy

[shell tendon]

my q there is whether that's due to it being a months old model that's badly undertrained

[zinc bane]

Hey guys, I need the opinion of the veterans here.

Right now I have an 1070Ti and on sdxl which I mainly like to use I get like 1,5 minutes per image on fooocus with a simple generation.( Only adding upscale)

I would like to have a faster workflow so I can try more things.

I was thinking of waiting to buy a 5080 or used 4090 when the 50xx cards drop later this year and the prices will hopefully be more manageable.(1200€ range)

Right now I have the opportunity to buy a 3090 for 700€ or I can also go for a 4080 Super. I am not really into training models but I would like the flexibility to run multiple Loras, controlnet and such. (My PSU is 750w gold + btw)

What would be the best move right now in your opinion? I am a hobbyist atm.

[sage reef]

im personally saving money and waiting for 5090 to drop and then il upgrade my whole pc

[trim magnet]

if u wanna train stuff 3090 is better than super

specially if u buy 2 because it supports nvlink

[final heron]

Im using sd 1.5

[zinc bane]

Nice, I believe that 5090 will be extremely overpriced so I was thinking of buying 4090 used when priced drops. But it may go a long time until then.

[zinc bane]

I want it purely for inference so speed is a must. Would a 4080 be a better option for that or still the 3090 is king?

[trim magnet]

for training 3090 is better

[heavy lark]

In the limited time I've been looking at cards, buying last generation cards are difficult. I wouldn't wait too long after the 5090 comes out before grabbing that 4090, because they stop making them pretty quickly when the new one is out

I tried to get 3090's and other than the used market, I couldn't.

[stuck sun]

4080 is gonna be faster than 3090 but not if you are generating multiple images at once or high rez image that need more vram

[zinc bane]

I was thinking of grabbing a used one as I don't believe the retail price will be too significant of a difference.

[zinc bane]

Oh ok, so 4080 seems everyone agrees it is not optimal. Is 3090 then the better choice right now or I should wait for 4090 used price when the new gen comes out later this year

[trim magnet]

for training the speed diff between 4090 and 3090 is not that big plus u cant nvlink with a 4090

now if u only generate imgs the 4090 is faster

[stuck sun]

depends what you are using it for! If its pure speed, and you dont need super high res stuff, 4080 will be faster. Btw nvlink isn't going to be much use to you, and you wont be able to power 2 cards with that PSU anyway.

the purpose of NVLINK is to pool the ram, which some professionally 3d animation apps take advantage of for rendering, stuff like that, and even then its not all that great, i have 4 quadros with nvlink and i turn it off most of the time, its buggy as hell

[zinc bane]

What super high res corresponds to? Scale, generation steps?

[stuck sun]

image resolution

[zinc bane]

Realistically what can I achieve with a 4080? Iam only thinking about that as a) it will have warranty, b)it will be a tiny bit faster and not used. On the other hand though the 3090 suprim x goes for 700 lightly used( as said by seller) which is a good price for something like this. I could also wait and give money to something like a 5080 or as I mentioned a used 4090 . I can wait but the thing is I don't know if it will be worth it. ( Maybe price won't drop for 4090 or 5080 will be too expensive.)

[stuck sun]

how much is 4090

[zinc bane]

Right now at 1700€.

New

[stuck sun]

lol fuck that

get the 3090

[zinc bane]

😅

[trim magnet]

yea more vram = better for ai

[stuck sun]

4080 is supposedly 29% faster ish than a 3090, so like, 700 vs 1700, for less vram, this is a no brainer

not even a question worth considering

if it was 800 bucks yeah, 850 maybe

[zinc bane]

4090 is 1700, 1100 for 4080

For new

[stuck sun]

ah shit sorry i had meant to type 4080

[zinc bane]

Ah ok

[stuck sun]

anyway 700 vs 1100, same deal, not worth 400 bucks more for less vram and a marginal speed boost

you can prob get a 3090 a bit cheaper, friend just sold one for 675

have one sitting in a box gathering dust i might sell

[zinc bane]

Well 25 more is not that much when paying something like this I guess 🤔

[stuck sun]

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

game frame rate comparison

[pearl ocean]

I got this

[rich kestrel]

UBI

[uncut surge]

is there some service that helps me figuring out the right models/checkpoints and prompts to generate a specific artstyle which i then can use myself to create hundreds of different images in that style? if so, under which name or where can i find that? 🙂

[final heron]

How to reduce glow or just remove it from face? People said turn doen cfg scale to 4-5, but the face still glowing too much, especially on the nose part

[uneven obsidian]

Hello 👋

[pearl ocean]

Maybe a 4080 is all you really need

[fervent thunder]

Yo

Anyone wanna collaborate to my Stable diffusion text-to-image Project?

[sudden ruin]

Ill let you know one last time to post it into #1092446741984444416

[fervent thunder]

Uh okay

[uneven obsidian]

Sure thing

[fervent thunder]

DM pls

[inner wharf]

help me plz

[sudden ruin]

How

[inner wharf]

check mess plz thank youy

[sudden ruin]

Ask your questions here, I dont accept random DMs

[inner wharf]

ok men

but not image

[sudden ruin]

#🏞｜general-with-images

[inner wharf]

thank you

[unkempt hatch]

will Stability use people's prompts in the API to 'refine' or better censor the model?

like, suppose I figure out some useful prompt to get images that Stability doesn't like; will it be watching and figuring out how to better neuter it, so that the released weights will be worse than the API?

[vale steeple]

you know, people always have trust issue with corporations

But what can I say really, with the look of Internet not trusting Emad say "Stability is in safe hands"

sure it did mention safety in the thread that announces API release. And some people apparently don't appreciate that

[nova zodiac]

the prompts themselves wont censor the model itself..

they may use the prompts to improve a prompt filter before they send it to the model for processing, but the resulting image wont be improved over time

[viral atlas]

Hi

or is it censored

like if I pay on the site rather than use a local github version

i rlly wanna do it with sd3

[rugged mirage]

you cant run sd3 locally yet, and you cant really generate porn from the APIs

[bleak matrix]

Good morning, everyone! How are you all today?

[bleak matrix]

Please see the #✍🏼｜rules-and-tos regarding NSFW; thank you!

[viral atlas]

those are the server rules though I meant the api

I'm not trynna discuss or post nsfw here

[viral atlas]

Ahhhh so it's filtered?

[hot kettle]

Can I encode 2 prompts and sum their features for generation?

Comfy, automatic1111 or hf

[sweet flint]

ok. so ive heard that forge is faster than automatic 1111. how much of that is actually true. Ive got a 3080ti for reference. will I notice any difference?

[wheat hazel]

maybe this is stupid because it's either not really working or it's probably already done but... Has anyone tried to get a low rank QR decomposition of the weights of the UNet layers?

[rugged mirage]

just try

tho forge is less updated, a111 had an update last week, forge hasnt had one in a month

[shell tendon]

Yes, it's several times faster

[still glacier]

depends of your hardware. I did not notice any significant speed improvement on my setup (2070S) but there has been a slight memory usage improvement.

so yup, try it and see for yourself.

[sudden ruin]

Please open a ticket and get in contact with either outerside or fruit

[shell tendon]

Interesting, it'd be nice to have actual data on this

It made a huge difference on my 3080 12gb and my 4090

At least 2-3x faster

[still glacier]

from my last test a month or two ago

Auto1111, xformers, fp8 enabled, JuggernauteXL, 1024x1024, 30 steps Euler A => 1.2~1.4 it/s (avg 1.4) + Reserved memory 8.6gb
Auto1111, xformers, fp8 disabled, JuggernauteXL, 1024x1024, 30 steps Euler A => 1.0~1.6 it/s (avg 1.5) + Reserved memory 11.15 gb
Forge, --cuda-stream --cuda-malloc --pin-shared-memoy, fp8 enabled, JuggernauteXL, 1024x1024, 30 steps Euler A => 1.5~1.85 it/s (avg 1.6) + Reserved memory 7.4 gb

and when switching to SD1.5 512x512 speed goes from 9.1it/s to 9.8it/s when switching to forge

[unkempt hatch]

like, baking is done, right? They're not gonna do more fine-tunes because I figured out how to make someone too handsome for their dogma, right? That's what you're saying?

[still glacier]

So nowhere near what they estimated on their readme but many people did notice significant improvement on their setup sooooo \(〇_ｏ)/

[shell tendon]

I presume the forge test is with xformers?

I don't have the data handy right now but on my 4090 it goes from something like 1.5-2.5 it/s to 4-6.5it/s depending on the sampler

What's weird is even though it's the comfyui backend I get significantly higher speeds with forge vs comfyui

So either some optimization was made in forge, or something is wrong with my comfyui venv

[still glacier]

It didn't change much to turn it on. In fact iirc. the dev discouraged people from using it.

[shell tendon]

Gotcha

Yeah it didn't have any real effect for me either

[rich kestrel]

Bro, have you seen how brain dead This generation is?

Tiktok has microwaved the youth's brain

15 seconds of going in there and the amount of brain rot is unbelievable

And you wonder why these people are willing to pay for AI influencers? You could make money out of shorts of wet farts if you are lucky enough these days. There are no standards anymore. That also means the future is fucked…

[foggy halo]

guys any idea when will SD3 release on HuggingFace?!

[frozen helm]

This generation? Which generation?

[foggy halo]

i need it for unlimited sd3 generations

[rich kestrel]

Zoomers and alphas

[frozen helm]

People my age still use TikTok? wtf?

That shit died a long time ago

[rich kestrel]

Who is the target demographic?

[frozen helm]

like you’ll get laughed at over here

[frozen helm]

I guess middle schoolers now.

[rich kestrel]

Same goes for youtube. Kids make these websites a lot of money it seems.

[low moon]

-80 IQ folk

[rich kestrel]

Rofl. Which these days make up 75% of the population.

[trim magnet]

all ppl in every age say the same thing about the young generation

[low moon]

i think it is objectively determined empirically and scientifically and stuff that everythign declined in the past 20 years

regardles sof the usual old geezer grumblings

across the board, economy culture etc everything in the toilet

hell wer ebasically in WW3

[heavy lark]

probably inside a month.

[hushed aspen]

hello 🙂 can someone point me in the right direction?

I want to take an existing image (a simple diagram) and use stable diffusion to do the styling without losing the text or (to a lesser degree) overall shape of the diagram

any pipelines you can recommend for a workflow like this?

[trim magnet]

indeed and is scientifically and factually known that most of the politicians and billionares that run the world are 70+ year old geezers

[rugged mirage]

I mean it's not, the average age of billionaires is 60s, and top politicians is more like 50s

Trump and Biden are the exception, look at the ages of other and previous presidents

[trim magnet]

yea who can forget the younglings at the congress

like my man schrumer such a youngling at 73 or mitch mconell at 82 or mancin at 76

[rich kestrel]

Yes, and the good news is that those clowns aren't going to live forever

No more than 20 more years before we get a complete change in management after those mfs bite the dust

[trim magnet]

thats prob why they keep pushin for the big ww3 they know its their last chance in power

[rich kestrel]

Hot take

But I kinda want ww3. Extra nuclear please

At this point a Fallout-style postapocalyptic world seems better than this shit

Working ur ass off all day just to own nothing. Nah. I say bring on the nukes and lets shake things up a bit

[karmic cedar]

welcome to the reason escapism is a genre

[low moon]

work all day to own nothing sums up very well everyone's situation now

[karmic cedar]

‘wage slaves’ is an evocative term

[sage reef]

yea man, i dont own a 4090... it sucks ass

[low moon]

my parents in communist eastern europe had an apartment paid for and enough saved for a car in that terrible system in the 80s

[karmic cedar]

but we do have cloud

[low moon]

now you cant do anything,buying a home is like a pipe dream

work pay your taxes and die idiots

[karmic cedar]

the folks running the world think mathematically and building mathematical models, but they themselves are sloppy, instinct-prone apes who can’t think beyond their own clever heuristics to have the imagination to change it all

[low moon]

they take even the little money we manage to make

put their little hand sin our pockets

every goverment is the same

[karmic cedar]

🙂

it’s all a football game to them

it’s all “if this than this”

it’s time to make your own meaning in this world, no more looking for it elsewhere

[low moon]

ive lost all faith on cops goverments culture and education and healthcare

[karmic cedar]

everyone is programmed

🙂

[low moon]

cops only good at harassing regular citizens, illegal immigrants murderers and burglars are ok

spineless swine just get regular people

jawalkers

[karmic cedar]

i’m sure the majority of those folks deep down never thought they’d become dangerous people one day

[low moon]

and bus tickets

[karmic cedar]

that’s naivety

[boreal dew]

what are the spec requirements for sv3d? I can't seem to find it.

[verbal osprey]

Guys, are we live?!

I'm talking about SD3

[karmic cedar]

dev API version is live

based on an older build of SD3

[verbal osprey]

So this is like a beta version and the "full" version is coming later?

[karmic cedar]

something like that. probably May 10th

[verbal osprey]

awesome thanks

[quiet forum]

#💬｜general-chat how do I generate a picture

[sage reef]

probably :3

[viral atlas]

wait so how filtered is sd3?

let's say no nsfw but just like with DALLE3 sometimes it refuses to generate certain prompts for whatever reason

I wish they would specify this stuff on the website

before you deposit any money

[karmic cedar]

Hard to specify the effects of concentration gradients

[heavy lark]

it will generate anything, but currently on the api it will send back a blurred image if it feels it's edgy. that won't be the cast in the final and i've already gotten lots of stuff that would definitely have been censored by dall-e.

anything else will be added by finetunes.

[mortal delta]

how can i afford better hardware for ai if i cant get a job because of life? it just feels like tech keeps going up in price.

[low moon]

Rank abunch of debt and buy whatever

with the end of the world coming might as well enjoy what time we have left

lol

remember no one owns anythign anywya people who seem to have things are in debt for decades most of the time

its all a charade

[red raptor]

hello guys, what should i type in prompt to get a half face image please ? i tried (half face) (left side) (partial face)... none worked. is there a way to do that ? Thanks!

[low moon]

uhm make the image and crop it in photoshop?

[red raptor]

what about the background ?

if anyone got an idea, please assist!

[low moon]

just type "neutral grey background"

or black or dark

plain

[red raptor]

this isn't my question. i asked if there is a part of the prompt i'm missing that could give me the result i need. thanks for the ideas

[granite topaz]

Guys
I just made a video with warp fusion.
But, I don't have davinci studio. How do I deflicker this?

[severe otter]

Can someone guide me to where I could post a Stable Diffusion (Comfy) tutorial link?

[still glacier]

depends what it's about.

[severe otter]

Perturbed, Advanced Masking, and a Fashion Surprise - https://youtu.be/j3xHNmEWWCI

[still glacier]

Then probably #🧣｜comfy-ui . Just don't start spamming auto promotion stuff and it should be fine.

[severe otter]

I appreciate that, thank you!

[sage reef]

lol

[fervent thunder]

whats wrong with https://stablediffusionapi.com/api/v3/text2img

[fervent thunder]

guys, what do i need to install to use pony xl ?

Im getting really weird images, i assume im doing something wrong. I did install the loras, vae, and clip is set to 2 :/

[graceful anchor]

Is there any information about there about how they plan to enforce licenses of models and such without using the metadata itself?

[vernal rock]

everytime i generate a img2img i get IndexError: list index out of range
i dont know what to do before it worked fine
i didnt use any extension i think

[narrow kernel]

Hidden watermarks in the image. There are some good ones out there that can survive a lot of cropping and jpg compression. Kind of like how most modern QR codes can still be read even if they're missing a bunch of the code

So I imagine they will use some kind of variant of it in the latent space

And by watermark, I don't mean something you can read. It's a formula that subtly changes pixels across the whole image to where you can't even tell they are changed. Like going from a red value of 197 to 198

In some color of some specific pixel

But do that for thousands of pixels and have an equation that can decode them

[deep horizon]

What is all this pony ai/diffusion stuff I am seeing on civit?

what is pony

[rich kestrel]

a breed of small horse

[jade mason]

When SD3 weights?

[boreal dew]

its a furry model using sdxl i think

[deep horizon]

ah... name checks out

[sage reef]

may 10

[pearl ocean]

When SD4 weights?

[sage reef]

one thing at a time

[dusk tiger]

so umm what was the point of joining the waitlist for SD3 a couple months ago? 😅

[pearl ocean]

Probably to get how many ppl are interested in it

[sage reef]

which is at least 5 people

[pine path]

I remember something like this for XL. The "hidden" watermarks were in fact pretty visible. Also, since the watermarks were handled by code you could just go in and remove the part responsible.

[graceful anchor]

thank you

[narrow kernel]

They can also be baked into the inner workings of the model, which I presume they will do. Sdxl's 1.0 vae I think tried it, but it backfired because it would give every image that chromatic aberration color fringe look. Though it likely wasn't encoding anything other than just being a tell that the image was made with sdxl

[opal hedge]

I'm on a no boob generating strike until sd3 comes out

[rich kestrel]

and on a no-boob strike you shall remain

not sure what youre hyping for

[opal hedge]

sdxl did too once upon a time

[rich kestrel]

tru. so 6 more months without b00bies?

l0l gl

[opal hedge]

The community and our collective autism will fine-tune sd3 for boobs within a week

[rich kestrel]

I just want realistic b( . )Y( . )b textures dammit

I wanna see those goosebumps on those melons

🤤

[opal hedge]

You can already do that with some models of sdxl

[rich kestrel]

share pix?

[opal hedge]

go on civit ai for a few minutes and roam the images

You'll find all sorts of booba

[viral atlas]

lol why are all of us here mentally unstable

[viral atlas]

idk but this is only a good idea for me who has dual Russian citizenship bcus then I can flee America and the banks won't get me there, they've sanctioned themselves off:D

[viral atlas]

lol you can say that about anything "how can i afford starbucks if i cant get a job because of life?" lmao

Also you should rent gpu hardware for ai unless its for gaming or cheap otherwise it becomes obsolete super quick

[pearl ocean]

Should be first item

[viral atlas]

I just won a lawsuit though so i'm gonna buy or build an obese computer:D not for ai though just for gaming n compiling stuff so I'm focusing on the cpu being good as well. Maybe a threadripper pro 7900 series 😍

[pearl ocean]

[viral atlas]

theres uncensored cloud sites you can use sdxl and sd2 and foooocus to make porn

im not gonna advertise but just look for them

a lot of them even give you like $5-$10 in free credits to start out

[low moon]

Coz the world is nuts and wer eon the brink of WW# and no one can afford anything lol

[viral atlas]

it depends on where you live though like where living costs for food and housing / rent are higher the wages are higher

but tech seems to cost the same everywhere

[opal hedge]

The world is fine

[viral atlas]

Yeah I think it's fine

idt ww3 gonna happen

[rich kestrel]

ur one to talk bruh especially since ur country and ur soy president wont leave ukraine alone

bc of yall we keep sending billions overseas too

[viral atlas]

Russias gonna win against Ukraine the West is alrdy giving up, and America might destroy Iran or something but so be it

[rich kestrel]

russia will win when we stop sending ukraine money.