💬｜general-chat

[charred mesa]

okay damn damn damn

[near silo]

exactly, and thats the only quote I will state for obvious reasons

Thats why we have had multiple companies reach out wanting to train their own models, cause thats make your own model money

Pixart was made for like $26k

[charred mesa]

yeah

though I hated the huggingface space images I generated with pixart lol

no matter the CFG

but yeah you could literally make a model for cheaper than $350K lmao

[near silo]

but yeah, I work with a few different companies on contract now, so I am not too worried about singling anybody out, letalone any specific quotes

But yes, I hope that puts it into perspective a little bit more 😅

[charred mesa]

If stability wants that sweet sweet license cash they need to decrease it or something

[full lark]

True

[near silo]

Granted, said company could have way higher income than we know of. I wouldn't be surprised if income is a factor

[charred mesa]

maybe they need that cash to make an offline Sora model lol

nah but yeah

[near silo]

But still, just wanted to put that out there as an actual factual quote give to one of thje companies potentially looking to work with us

[pine path]

they need cash period

[near silo]

^

As much as Emad said a while ago that they would likely be bigger than open AI and have more resources (yes, he actually said that), it has certainly not happened 😅

[charred mesa]

at least these won't affect us for now 😔

im waiting for the 800M, 2B and 8B models...

[pine path]

there might be less support

[near silo]

I am quite curious what the param counts will be

[charred mesa]

yeah that's why for now... 😬
I hope they won't be on the brink of bankruptcy again or whatever

I don't remember what happened to that

[pine path]

I meant from the community

[charred mesa]

ohhh

[near silo]

My biggest hope and hold out (arguably the thing i think will make SD3 either a success or a flop) is if LoRA's will be compatible across all of the different param count models

if they can do that, I think SD3 will be a massive success off of that alone

[charred mesa]

the 2B "edit" might become the most popular 🤷

perfect balance between quality and VRAM usage

[charred mesa]

that's exactly what I've been thinkin about too, ""cross-platform"" LoRAs loll

[trim magnet]

yea gotta wait and see how good it is

[charred mesa]

there are a few examples on the paper

it's not bad thanks to DPO!

[near silo]

I was already right when I assumed that T5 would be a chocie you could opt in or our of, so I am seriously hopeful I am right about SD3 having inter-model LoRA support

[charred mesa]

same architecture, difference model size, same clip and T5 though... could it work? 🤔

[near silo]

SAI has already proven they could do inter-model LoRA's. As SDXL LoRA's work natively on SDXL Turbo, so fingers crossed

[charred mesa]

yeahhh hmm

[near silo]

I seriously think if they got that working, it will MASSIVELY improve the success of SD3

being able to allow people to train LoRA's on their 8GB cards again would massively speed up the quality improvements of SD3 in a way that SDXL was cut off from

[charred mesa]

also fingers crossed for runnin T5 at int8 + xformers helping to get the 8B model working on 12GB

I still kinda want that boost in prompt adherence

even though the benchmark seemed to show that EVEN WITHOUT T5 it was pretty coherent alongside DALLE3 and Ideogram

🔥

[near silo]

also, I can already see it coming if they don't support it

800M will be reserved for people withw eak GPU's, 2B will the go to, and 8B will be the best of the best on launch

Rapidly, 2B will approach and surpass 8B in terms fo quality because of the massive ammount more of community support and fixes, and 8B will be a waste of time

thats my thought on it as of now

[near silo]

I really do not have much hope for 8B running below 16GB, but people might be able to distill it down quite nicely

[charred mesa]

yeah that sounds like the most likely outcome

800M could probably run (very slowly) on androids too lmao

that would be fun I guess

[near silo]

you'd be surprised. It doesn't have to be slow anymore

Huawei's new phone can run SDXL turbo in real time

[charred mesa]

if 2B w/ T5 performs than 8B w/o T5 (in prompt adherence) then we know the superior choice 👍

[trim magnet]

yea huawei has the new npus on their phones

[near silo]

also, android phones have enough RAM on the high end to likely run all of these models

[charred mesa]

yeah

[near silo]

my phone has enough to run SDXL if I wanted to, no problem

[severe vine]

Hello guys, do you know about any 2d sprite animation game asset generators? I need sort of consistent character running, jumping attacking etc. I think it's more easy to make that model than realistic video, but for some reason i can't find any of that.

[jovial wraith]

your best bet would be pixellab, but it's paid

[severe vine]

Yeah, i saw that. As far as i remember there was some other project that animates existing images but i don't remember name. Anyways it's really strange that noone did this already, considering existing SORA, atleast at pika labs or runway level but special for 2d games

[jovial wraith]

i'm currently working on a workflow that does this, albeit not very actively. i can generate still images but consistent animation is an issue right now. here's my workflow if you're interested: https://discord.com/channels/1002292111942635562/1212330984192614440

[fervent thunder]

is there an alternative to runpod?

[buoyant moss]

if you can get an account, coreweave

But unlikey you can get an account

together.ai also has cheapish dedi GPUs

Note, 7 seconds is without xformers but fp16. I suspect that SD3 was also using fp16, but we have no proof of course

So I think 5x slower than SDXL is probably a reasonable ballpark

[pine fiber]

thats kinda bad

would definitely need to get faster

[fervent thunder]

my issue with runpod is that my gpu is always unavailable when i want to use it

[buoyant moss]

together.ai has GPUs, but their sales didn't contact me again. I guess they have a ton of requests lol

Coreweave, fat luck getting an account

[fervent thunder]

thanks for helping me

[near silo]

I don't see a model with more than 3x the param sbeing only 2x as slow, personally

[pine fiber]

a full size sdxl image takes me like 40 minutes on my machine LOL

but that was last year

[fervent thunder]

what did u change

[pine fiber]

idk I havent tried since

I have an amd card on windows

until rocm is properly implemeneted in pytorch

i cant do anything locally

[still glacier]

you should take a look at zluda

[fervent thunder]

what effects me if i use a cloud gpu far away from where I live?

[broken cave]

do you have the resources or motivation to create such a model?

[broken cave]

the benchmark belies that most people put in vague prompts

[broken cave]

the differently sized model checkpoints will not have compatible loras.

[near silo]

How come the certainty?

[broken cave]

because that's how the math works

[broken cave]

it is frustrating but the alternative vendors are too good to be true

[fervent thunder]

ah damn lol

[broken cave]

the alternative cloud vendors are like the fuckup kid from your high school whose grandparents left him a trust, and instead of doing volunteering or philanthropy or something positive, he jacks off and plays dark souls all day, until one day the ethereum he mined because he read about it on a white supremacy forum is worth millions of dollars

and now he has all these GPUs he doesn't know what to do with, all this capital, and no knowledge

[versed yoke]

But you could try to make q x adapter model to make them work together. ( ther was a paper about it)

[near silo]

Is there a specific reason why they were able to do it with SDXL and SDXL turbo even though they use different sampling methods, and are trained at drastically different resolutions with different weights?

Surely those are not any more compatible than the SD3 models will be

[fervent thunder]

lmao

[broken cave]

in my experience, a fine tuning of sdxl used on the sdxl turbo checkpoint does not work very well

it performs similarly to combining multiple sdxl LoRA fine tunings, which is to say, it does not work very well

[fervent thunder]

tbh i find it easier to obtain my desired results using 1.5 compared to sdxl, does that mean im just a noob?

[near silo]

In my experience and the experience of many others, it works fine. I am not sure why that would work so well and SD3 would not be able to do the same with them all built from the same arch

[broken cave]

the differently sized model checkpoints are essentially checkpoints for different architectures

[broken cave]

well, the math is different, i guess is the answer.

[near silo]

Are you able to say why with any less vague un-certainty?

[versed yoke]

With x adapter u can use sd 1.5 loras wizh sdxl. I am sure u could make a xadapter version for cascade and sd3 https://github.com/showlab/X-Adapter

[near silo]

ah man, i always forget about this thing

[versed yoke]

The code is finally out : D

[charred mesa]

yeah they'd have to update it to be compatible

[broken cave]

yes. all diffusion models are functions that turn noisy images from timestep t to slightly less noisy images at timestep t+1. sdxl turbo and sdxl are the same architecture, but sdxl turbo has a postprocessing to do more denoising per time step. a "lora" is like another piece of postprocessing that, because the numbers on "the left hand side" of the "lora" between sdxl and sdxl turbo "Are the same," the "loras" are "compatible"

[jade mason]

Do we have a timeline for weight drops?

[charred mesa]

but if that would work well I'd be heavily surprised

[charred mesa]

don't remember any being shown or announced

hopefully this Month (kinda unlikely, don't know how much it has been trained so far) or April, in the worst case probably May

[pine fiber]

emad said same timeframe as previous models

[charred mesa]

ohhh

[pine fiber]

which was like 3 months

[charred mesa]

so I guess May 😔

[pine fiber]

but stuff he says is rarely accurate lol

[charred mesa]

we are literally gonna forget about SD3 existence for like a whole month or two and then get surprised when it finally comes out lmao

at least there's gonna be that super secret server

[shell tendon]

that would be incrdeible

[charred mesa]

I just don't know how long we would be there to make images

[broken cave]

@near silo i guess there isn't a succinct way to explain any of this because it's really in the weeds

you have to know about everything all at once

[jade mason]

Yeah, he should drop it before the hype dies down.

[charred mesa]

ehhh

unless its like some SD 2.9 that gets leaked then it wont happen

SDXL moment

[near silo]

I know how Ai models work, and your explaination makes sense. I didn't realize SDXl and SDXL were the same architecture. Does that suggest they are the exact same amount of parameters? Cause if not, then that explaination doesn't make any sense

if they are then yeah, i get you then

[broken cave]

i haven't read a paper specifically about sdxl turbo, i only know about distillation generally. when creating sdxl turbo, stability did ad-hoc experimentation to find which weights in sdxl should be frozen / "slowed" such that, because LoRAs on e.g. attention layers seem to perform best, those LoRA weights in SDXL turbo would produce aesthetically similar results as sdxl. they may have even systemized some part of the distillation process to make that happen, maybe it wasn't ad hoc. i don't know.

[near silo]

Alright, I get you in that case. That explaination makes more sense then

Thanks for the clarification

[broken cave]

yes sdxl turbo and sdxl are the same architecture. https://huggingface.co/stabilityai/sdxl-turbo/blob/main/unet/config.json versus https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/unet/config.json, they differ only in "sample size" which i think corresponds to the 512x512 versus 1024x1024 image target

but i'm, not sure

this is the most succinct "documentation" for the model

[near silo]

That makes sense in that case

[broken cave]

afaik if i looked at the actual thing as declare in pytorch it would be the same

[pine fiber]

what res does sd3 target

1mp?

[charred mesa]

1024x1024 iirc

[buoyant moss]

I wonder if SD3 is using v-prediction + zero SNR

Haven't had the chance to read the paper yet

SDXL 1.0 didn't use the fixed zero SNR, although I am told that they have internal zero SNR models + v-prediction on SDXL fine tunes

[radiant tusk]

how many steps do you guys use for aniamtediff?

[broken cave]

the strength of stable cascade suggests it is their dataset that has improved enough to benefit from larger models.

[buoyant moss]

Not sure how that relates to zero SNR + v-prediction, but still a good thing to hear

I think the zero SNR fix can be applied without too much breakage, but v-prediction would almost certainly break all the LoRAs

[pine fiber]

log-snr and v prediction

[buoyant moss]

Ah v-prediction is nice

If they were going to break the LoRAs, doing v-prediction with a new architecture (not SDXL/SD1.x) would be the time to do it

Since old LoRAs wouldn't work anyways

Fine tuning scripts would have to be changed too

Not sure if the log-snr thing is relevant, the none zero-snr was a bug in their code. I'll take a look at a white paper!

[mellow meteor]

[buoyant moss]

Guess this will cause some breakage in all the webuis & plugins until they sort it out

[gaunt pulsar]

So if the paper released today, when do you think they will start calling people from the wait-list to test it?

[charred mesa]

according to a tweet reply from emad, today or tomorrow (but it was a reply to a random twitter user instead of an announcement so I wonder if only they gonna get access lmao)

[gaunt pulsar]

Nice! I can't wait to test it, the hype is too high xD

[bright hedge]

Does TripoSR export models with textures? So far I'm only getting obj without texture.

[static cape]

Thanks for the research paper. I'd like my invite now (or the open source release). 👼

[charred mesa]

the model's still far from finished

lool

but the early test invite would be great

[brittle maple]

hello

[shell tendon]

i'm willing to run naked through the streets screaming if that's what it takes to get access to the weights today

[vagrant fox]

any way to fix black screen withouth using -no half command on broken models?

[hasty hornet]

doubt that's something devs interested in

[brittle maple]

4060 good for xl model?

[brittle maple]

i have same error

i change broken model and fix it

[vagrant fox]

its mostly with models downloaded from civica

[brittle maple]

you need redownload

it help me

[vagrant fox]

redownload certain model, correct?

[brittle maple]

yes

[vagrant fox]

because no half slowed the shit out of my generation like so much i was like its worse than my 3080 that i get rid of and i was pissed for half night lol

[brittle maple]

--xformers --medvram --no-half-vae my argument i have 1050 ti

[vagrant fox]

i have modef from 10 to 16 gb so medvram is not for me also no half slows your generations so i wont be using it

no reasong buying 4080 super and then limiting speed to less than my old card had, i will try redownloading models

[brittle maple]

how big your model?

[vagrant fox]

6.46 gb

[brittle maple]

for 3080 its ok i think

[vagrant fox]

im on 4080 S now, 16 gb vram, should do just fine

but with command no half it generates slower than 3080 on no half with small models somehow

so i will try getting this model again

i know is known issue with models downloaded from civit ai

[brittle maple]

when i try download from civ its stop every time

[vagrant fox]

are u logged in?

[brittle maple]

yes

[astral goblet]

is amiliafox spamming anyone else here? they've never said anything in the server and messaged me

[vagrant fox]

what did he msg? watch out for any links in DMs

[astral goblet]

just a friend req and a sticker wave. cliche intro for bots. but the lack of any messages on any servers sketches me out

ameliafox1 is actual username

she apparantly

[vagrant fox]

yeah sounds like bot name fishing out for weak ass simps xD

[astral goblet]

yeah

[brittle maple]

faces in the distance are greatly distorted

[vagrant fox]

tell her to suck a pretzel and block

[astral goblet]

think thats a limit of the vae architecture. sd3 incoming

[vagrant fox]

people who dont go straight to point why they msg are most of the times bots/scammers

[near silo]

@astral gobletAmmelia fox is a personf rom the Pixja server. Its a crappy "AI Art" server

[astral goblet]

already did hahaha. was wondering if it was a wide net cast or targeted at me

[brittle maple]

Midjorney do this good

[near silo]

I got messages from her as well as other people from that server. Joined it for a bit, but its a bit of a cesspool

Shes not a bot, just fishing for people to join their server, which is still annoying

[vagrant fox]

yeah but u cand run MJ on your machine thats biggest issue, i rather own tools i am using

than subscribe to them

main reason im using davinci resolve instead of adobe

[brittle maple]

yes sd my best ai because freedom

[near silo]

I personally recommend against anybody joining Pixja, the server seemed kinda toxic to me. Lots of gamer words in their chats

[vagrant fox]

also mj is build on sd anyway they just do their model training

[astral goblet]

i went to the newbs room in midjourney and spotted 8 pics with distant distorted faces instantly. you're leaning into confirmation bias. MJ uses vnet too

[charred mesa]

ah thank you I ignored their messages

[vagrant fox]

never heard of it anyway but i was out of loop for over year

[astral goblet]

should get banned from here then. never participates and just does unsolicited messaging

[near silo]

The people in the research group I am in and I have been looking at MJV6, and we have no idea what they have done to it that has degraded its quality so much. I understood it being wonky when it first started training, but it looks worse now than it did before

lots of deformations and oddities that we do not understand

[astral goblet]

cant understand them. v6 could be 8 different model weights

can't know. its proprietary

[vagrant fox]

would not be surprised if their training data was corrupted by ai generated img's

[brittle maple]

but when I tried the faces were always good

[near silo]

I honestly have no idea, all I know is that V6 feels like its worse and worse every day

[charred mesa]

😔

is this the end for midjourney?

[near silo]

damn it, this is the non image chat

[charred mesa]

has it ever been this over for midjourney?

[vagrant fox]

nothing to be sad about

[near silo]

I am gonna send some examples real quick in #🏞｜general-with-images if anybody wants to peep

[charred mesa]

yeah I really care about SD3 anyway lol

[astral goblet]

go use mj then i don't know what the problem is

[brittle maple]

i use mj for desktop picture)

[trim nymph]

calm..lets first see how good sd3 really is. especially with diverse styles+textures of each-my hopes are high!

[brittle maple]

i want it on computer

and free

[vagrant fox]

you could of saved all that subscription to get used gpu from 3000 series and run sd on your machine over time

[charred mesa]

I wonder how good the Pony model will be, people say that PonyXL is very good 🤔

[astral goblet]

distant faces are a limit of the architecture. its a vae problem.

Use extensions like adetailer

[charred mesa]

I am not a brony myself but I am interseting in how good these models can get for OTHER artstyles

[vagrant fox]

3080s are cheap af currently

[brittle maple]

and without censorship

[shell tendon]

interesting. i've never used MJ tbh, just seen other generations

[astral goblet]

its a porn meme model. it actually isn't very great. just hyped.

destroys lora compatibility and if you don't heavily prompt against it, throws anthro cocks on everything

[vagrant fox]

MJ had some adventages at start when tere were not many trained models for SD and not many tools, now u can do all stuff like tilling, face swap etc on sd

[astral goblet]

Bronies over blow anything that is MLP related

[brittle maple]

i want free sora)

[near silo]

I sent the images I was referring to about MJV6 in #🏞｜general-with-images

[brittle maple]

yes extention on sd big plus

[vagrant fox]

shit looks like first generation with 1.5 sd model xD

[near silo]

Yeah, its all types of messed up, and its getting worse by the day

[vagrant fox]

great to see greedy guys overcharging for tool they build on open source stuff getting slowly rekt

[pine path]

they will use offset noise most likely

there was acutally a zsnr vpred XL model

[vagrant fox]

like i said if u don;t have decent gpu just hold your sub and get secondhand 3080 for 400 usd or so

[charred mesa]

ah what the hell 💀

nvm then?

[pine path]

they had a vote between regular, offset, and zsnr vpred. Offset won

I doubt they're going to bother with zsnr now

[pine path]

it's one of the best anime models

which is why it's so popular

[vagrant fox]

hentai more likely Kappa

[pine path]

public, at least

[charred mesa]

lmao

damn

[vagrant fox]

there are shit ton of anime models u dont need to be stuck with porn garbage

anime models are one of easiest to achieve decend quality since it's easy style to replicate in my opinion

[pine path]

haha

[astral goblet]

soryr i'm not supposed to talk about this stuff here. i forgot

[vagrant fox]

you think its harder than photorealism?

[near silo]

I see somebodys messages getting nuked

ag, I see now why

[astral goblet]

self nuked. anytime i criticize this topic, mods catch me later with a warning

[near silo]

As somebody who has spent months trying to fix the realism issues with SDXL, I am extremely hopeful that SD3 won't be as damaged for realism. I have had to undo sooo many issues with SDXL for months and months, and am still fighting with its issues every day

[near silo]

I feel that, very strong censoring when even fair crticism it brought upon anything they do. Can't say I haven't been the victim of it as well

[pine path]

I think you're overestimating the quality of the anime models. If you think I'm wrong then please name some good ones.

[near silo]

I would agree

[pine path]

1.5 is plagued by NAI garbage and barely anyone has enough compute for XL

[vagrant fox]

it's simple, how many anime models and how many upscalers that at first worked fine just with anime were intorduced? that's clear indicator that's easiest style to replicate, little cartoonish heroes with cute eyes, wow, super hard bro

[astral goblet]

barely anyone might be projection. a ton of people have more than 8gb gpus

8gb gpus are the new 4gb gpus

[broken cave]

the problem with removing human anatomy from their dataset is that it makes many creative tasks harder. to a certain extent i wish they would just make a license to ban pornography, so that they can put it back in their dataset and run takedowns against whatever they want, so that the model works better for real creative tasks

[pine path]

I am talking about finetuning

pony was trained on 3x a100 80gbs over 3 months

[broken cave]

they could make whatever license they want, and maybe an ordinary person wouldn't have to follow it, but civitai would

[astral goblet]

rented. ez. not hard to budget for

[near silo]

I am not even so much talking about fixing the anatomy, that's even more sadistic. I am meaning even just basic lighting, focal planing, background coherency, skin rendering

[broken cave]

hmmm...

[vagrant fox]

1.60 usd per hour per gpu on nvidia website

[pine path]

please name a quality model then

[astral goblet]

all i notice nudes in the data set causing is belly buttons over all outfits. and nipple ornaments.

base xl doesn't do it. all popular refines do

[charred mesa]

I hope Bokeh is not going to be a huge problem

[vagrant fox]

and most of it's training time was eaten by making sure sex poses and furrys are replicated correctly, same thing with any other porn model

[charred mesa]

hopefully negative prompts are gonna be much more efficient

cause it'll know exactly what you want removed

[broken cave]

the steam hardware survey is pretty representative of layperson VRAM. 74% of people have 8GB or less of VRAM.

[brittle maple]

i have 4 gb)

[vagrant fox]

and whats procentage of people on steam using ai tools. 0.5%?

[broken cave]

the steam hardware survey is so large, and video gamers so disproportionately the owners of video cards, that this is almost certainly a representative sample of GPU owners among laypeople

[brittle maple]

dont know why need more 6 gb if you not use ai

[vagrant fox]

well if u play 1080p in 2024 then i guess u dont need lol

[broken cave]

that said, they are releasing larger models

i agree there's no point in making things work for video cards with less VRAM

[brittle maple]

you have 4 k?)

[vagrant fox]

it's hard to expect cutting endge tech working on budged card from 8 years ago

[astral goblet]

you read that wrong

[vagrant fox]

i got main monitor 2k res 240 hz and 2x full hd as side monitors

[brittle maple]

and 6 gb not good for it?

[near silo]

hell, even then a lot of modern games use more than 8GB VRAM for 1080p ultra now

[near silo]

absolutely not

[astral goblet]

100 - 34 is 66. and 12gb has 17% and growing

[vagrant fox]

now new ports from consoles would need 12 or 16g to run at max i seen some article

[near silo]

10GB should be the minimum for 1080P ultra, 16GB the minimum for 1440P ultra, and 20/24 for 4k ultra, based off of what the highes AAA games are pushing now

[broken cave]

i want you to think very deeply about what i wrote

[astral goblet]

8gb cards are an age of done. they barely run the new generation of unreal engine 5 games with nanite and lumen. they're late last gen now.

[brittle maple]

yes ports very bad for vram

[vagrant fox]

it's really not that expensive to get used card from 3000 series those days, people rather use money on subscription ai tools and stuff that dont pay off

Or generate some income slowly with your creations, music videos etc (upscale your frames for higher res) and get gpu

[astral goblet]

hah i get cha now. my points still stand. 8gb is old news. get with the time pleb

[vagrant fox]

i used to make 100-200 euro per day of work creating music animations spaming on fb groups with musicians some time ago when i needed some cash

[brittle maple]

i have 1024 [768 monitor lol

[vagrant fox]

be creative and u will upgrade your hardware

[vagrant fox]

i feel your pain

[astral goblet]

unreal engine is a pc engine first. It kind of killed the age of bad console ports. master chief collection is probably the last relic of such an age

maybe final fantasy 16 when it comes out

[vagrant fox]

most entry level laptops have now better resolution, i would focus on upgrading your setup / generating some cashflow with ai and saving up for used parts this will get you to generate better stuff in long run

[astral goblet]

i brought up UE, because it's the bar.

[near silo]

I got a 3090 since I have been in this scene, and it has EASILY paid for itself

and I enjoy it day to day as well

[astral goblet]

yup. no reason to hang out on 8 if you're doing gneerative ai

self owning might be a good reason if you wanted to do that

[near silo]

I started with a 3060ti, upgraded to a 3080 for no money, then got a 3090 for $250 more

[astral goblet]

guess it depends on what your goals are

[vagrant fox]

i got 4080s instead od 90 just because i were out of loop and were not interested in buying up beaten 3090 with no warrianty, might upgrade with time if my needs bump up and my backlog of orders

[near silo]

I have all 3 cards still, and I use my 3060ti with my 3090 to run bigger LLM's. Gonna be selling the 3060ti for a 3060 so I can have 4GB more VRAM

[astral goblet]

can your 3080 and 3090 nvlink? there's multi gpu rendering now

[near silo]

I bought a second hand 3090 with a warranty. I can't blame your aprehension to getting a used 3090. My first one was a scam that took 3 MONTHS to settle and dispute, and stressed the fuck out of me

[vagrant fox]

don;t you need same chip gpu to link?

[astral goblet]

thats how it was with sli

i'm not sure though.

[near silo]

I have no room for the 3080, only the 3060ti and the 3090 as of now. I only use the second GPU for more VRAM for LLM's

I never need more VRAM or more speed for image gen honestly

I got the 3090 for training anyways, which is something not parallelized for SD just yet

[astral goblet]

riser cable 😉 just hang it off the side of your case

its alive! ITS ALIVE

[astral goblet]

i think koyha has it

[vagrant fox]

case also would be cheaper to upgrade than gpu itself

[near silo]

I would need to get a new mobo in order to properly use the 3080. Thats the biggest issue I have right now. I can onyl use my 3060ti for VRAM overflow for the 3090 cause of the severely bandwidth limited PCIE slot it is in. I wasn't expecting to run a multi GPU system when Ibought my new mobo

[brittle maple]

I have strange artifacts in games, smoke with squares and transparent textures with a spruce noticeable mesh, but in general the picture is not critical, is it a lack of memory?

[near silo]

That does not sound good-

what GPU?

if its an intel Arc, maybe thats normal 😅

[brittle maple]

1050 ti

[vagrant fox]

you fried it with ai

[near silo]

hmm, that shouldn't have any issues like that

[vagrant fox]

put it to owen for some time and bake it might help but not permanently

[astral goblet]

seems like amd texture filtering problem.

[near silo]

They have a 1050ti, but that does sound like an AMD issue

[vagrant fox]

nah he just got some artifacts due to overstressing card had simillar stuff with 3080 after running it 24/7 for few months

[brittle maple]

lol its before ai

[near silo]

I actually had a friend fix their GPU by using that trick haha

[vagrant fox]

yes its old trick and works

[brittle maple]

i want show you

[astral goblet]

the oven trickk lol. you have to be very informed to pull that off without fluking it

[near silo]

Just be careful what plastic the card is made from

[vagrant fox]

he can take off cooling system and bake just board

[astral goblet]

reseating the solder typically isn't needed anymore, but i could see a 1050ti needing it potentially. wouldn't cause texture artifacts though

shitty chinese sub par solder is discovered by the market now and it's self corrected. same with shitty capacitors from 2005 era

[near silo]

related note, ended up fidning out the plastic tray my mom got from the dollar store many years ago is made fo rextremely high tem resistant plastic haha

Was cooking a meal and she didn't tell me she recently put our serving trays in our oven. It was at 400F and it ended up being fine haha

[brittle maple]

can i put images in this chat?

[near silo]

no, you need #🏞｜general-with-images

[vagrant fox]

can't

[near silo]

this is general without images

[astral goblet]

/spartan

[brittle maple]

i can put image game?

[vagrant fox]

https://tenor.com/view/nike-just-do-it-gif-26636514

[brittle maple]

with afrefacts

[near silo]

you should be able to, its a general chat and we are helping you

just send it in #🏞｜general-with-images

[brittle maple]

ok

i need take sreen

[high falcon]

Civitai in non active tab consumes all my cpu endlessly -_- Why so?

[winter pike]

its "a bug" they were suppose to fix months ago, im grateful for their website but damn

[high falcon]

And I was wondering why the fans were humming so loudly... It's better to not forget about this tab.

[shell tendon]

i've noticed the same bs

[astral goblet]

all animated images just keep animating in the background forever

[high falcon]

Only one animated image - user avatar. So, this is unlikely.

[astral goblet]

its a webapp and caches in the background

[clear oyster]

im not having that issue apparently

[harsh parrot]

Що робити? В мене пропали боти для генерації зображень

[fervent thunder]

Why are you typing in Ukrainian

[velvet nexus]

Sory guys I willl use it first time, which channel I should write prompt and which command I should use

[fervent thunder]

"What shall I do? I've lost bots for generating images"

[cloud plume]

there's currently no bot/command i.e. see #1047610792226340935

[sudden marsh]

where can i use SD 3?

[vagrant fox]

not yet released to public afaik

[sudden marsh]

ok.

bye guys lol

[charred mesa]

based

[sudden marsh]

even if i'm a famous company guy needed this tech

[vagrant fox]

???

[sudden marsh]

no prb sorry for the inconvenience

[prisma horizon]

lol

[sudden marsh]

I will wait

and this is fair

now i wait.

ok when is it released?

[vagrant fox]

ofc u wait tool is free don't expect them to run around you xD

[prisma horizon]

🤣

[sudden marsh]

even if i subdue them?

🤣

[prisma horizon]

they should run, but jump, and i they can't jusmp, crawl XD

[sudden marsh]

i understand so SD 3 is 100% free?

or is it going pay for use?

like GPT 4

[prisma horizon]

usually, sd is free for private use, in case of a business use, the question is different

[sudden marsh]

business use

[vagrant fox]

send email to stability ai

we donb't know future

[prisma horizon]

but this is in case of every sd model

[sudden marsh]

you're not stability ai here?

ok gotta check the email

[vagrant fox]

I am, give me your bank account and left arm

[sudden marsh]

[e-mail]. Kidding.

[prisma horizon]

we are just users,

[vagrant fox]

+1

[sudden marsh]

and the pen? no pen = no contract

[prisma horizon]

but if youwant to send us money to buy new hardware, why not? XD

i'm so needy for new vram XD

[sudden marsh]

but you're not Mr Stability AI, you are Shaikorth

or Mss

[vagrant fox]

what u on rn Shai?

[sudden marsh]

I wanna talk to the boss

[vagrant fox]

boss dont have time for small fish like u

[prisma horizon]

i only have 4gvram

do you know how hard is genning and testing on a 4gvram? lol

[sudden marsh]

how do you know if im a small fish?

[vagrant fox]

well get used 3000 series, 3080s are like 400 usd now or so

[prisma horizon]

the other day i attempted to train a lora, without results

[vagrant fox]

if you were not you would know how to use email and communicate like proper company

[sudden marsh]

well for your info, i communicate how i want

[vagrant fox]

cool, don;t get mad small fishy fishy

[sudden marsh]

i guess your boss is a small fish too then

[prisma horizon]

zie, don't pull his leg, let him play his game how he want lol

[vagrant fox]

if i ever had a boss he would know how to ask company based questions and use email

[brittle maple]

i have 4 gb you can ask me)

[prisma horizon]

respect his freedom

[sudden marsh]

lol if im a small fish then you are a small fish too lol

[prisma horizon]

from now on, we are bro in misery -,-'

[sudden marsh]

LOL

[brittle maple]

small bro)

[sudden marsh]

As an AI, I have no feelings, will, and I respect every fish in the planet

[brittle maple]

big fish was small fish

[prisma horizon]

never trained a style lora? it looks so funny to do >.<

i want to add details and shadows in some precise points, but it looks so hard

[brittle maple]

you know ?small steps and big things

[brittle maple]

no

[prisma horizon]

gens fatally lacks shadows, it's my mission to solve this peril >.<

[sudden marsh]

I hope your boss will be a big fish one day lol.

[prisma horizon]

too bad bro XD

[vagrant fox]

why not became big fish yourself?

[sudden marsh]

who said I am not?

[brittle maple]

i no have money for big fish)

[prisma horizon]

big fishes get fished and eaten

[vagrant fox]

where is Nemo

[sudden marsh]

i have food for big fish lol

[vagrant fox]

mid rare steak?

[sudden marsh]

ok i dont wanna talk about your boss but those who say people are small fish are really wrong

(the only truth is : you don't know) bye and good luck to respectful people here

[vagrant fox]

we respect fish every size

https://tenor.com/view/fish-wink-gold-fish-swim-cute-gif-17467988

[sudden marsh]

🙂

[brittle maple]

chubby fish

[sudden marsh]

And please forget what i said about me. I am just like all other users. Nothing to take seriously...Unless....kidding

[brittle maple]

want pizza with fish?

[sudden marsh]

sure

[brittle maple]

🍕 🍕 🍕 🍕 🍕 🍕 🍕 🍕 🍕 🍕

[sudden marsh]

🙂 ok guys and don't forget to do pressure on the boss to have equal shares ok? Not kidding, yu guys gonna be rich. Hopefully

(stability ai team i mean)

[brittle maple]

[sudden marsh]

gota watch time to be alive now. I mean Two Minutes Paperz (I gotta)

[boreal canopy]

Damn, trying to read any PC How to tutorial, in this case the install tutorial help for AMD, i feel illiterate XD

[nova zodiac]

hold on to your papers!

[tropic frost]

out of curiosity, did any of you guys modify your gpu settings to use SD better?

[mellow meteor]

🧇

[still glacier]

check pinned messages of #🤝｜tech-support

[static cape]

Go and take a closer look at the paper, everyone - there are several more SD3 example prompts at the end: https://stabilityai-public-packages.s3.us-west-2.amazonaws.com/Stable+Diffusion+3+Paper.pdf

[lusty beacon]

SD3 Research Paper https://stability.ai/news/stable-diffusion-3-research-paper

When can we expect SD3?

[static cape]

TIL Karras is an actual researcher called "Tero Karras".

[nova zodiac]

SoonTM

[lusty beacon]

Okay waiting for it badly 🙂

[nova zodiac]

We all are chief

[fervent thunder]

Have you tried waiting harder?

[lusty beacon]

ehm

[fervent thunder]

Like, wait faster.

[lusty beacon]

https://tenor.com/view/confused-wth-wtf-really-meme-gif-17151805

[opal hedge]

SAI be teasing us with SD3 😦

These announcements aren't even for us, they're for investors

[vagrant fox]

they don;t need to do public annoucements for investors tho

it can all happen behind closed doors

[nova zodiac]

hype train doesn't build itself.. just look at how many people have posted "wen sd3?" in here/reddit/other discords since the announcement

[dim minnow]

These mfs can write a 26 page research paper on stable diffusion but can't fix the discord bot 😂 😂 😂 😂 😂

[vagrant fox]

and what's so broken here?

run your sd locally

or you expect them to build tool for you for free and also always provide resources? :v

[fervent thunder]

in the sd3 paper they state t5 encoder needs 19.05 gb of memory

[vagrant fox]

i have sent you screen on #🏞｜general-with-images

[arctic sedge]

Where does it say that in the paper?
I saw that a unoptimized version of the full 8B model + T5 fits onto 24Gb, but i didn't see anything referring to the T5 model requiring 19Gb.

[agile umbra]

when will sd3 be open-sourcely released?

[fervent thunder]

posted in #🏞｜general-with-images

[fervent thunder]

You are among the, like, 0.0001% of sensible people who are capable of putting themselves in the shoes of someone else (an overworked SD employee).

[crimson belfry]

@dim minnow The bot is not broken

Once we get access everyone will be happy im sure, The quality is amazing and tbh I am hyped up as well.

[radiant tusk]

How do you fix overlap setting giving blank images or just have to turn to -1

[pearl ocean]

imagine running A.I models on the M3 Pro MacBook Pro 16"

[pale latch]

well i mean yeah it's sort of what they've been doing this whole time. not weird to expect more of the same. if they change course, i would be like "woah a surprise!"

[fervent thunder]

it’s more about the tone that some ppl have, like it’s their god given right to be able to generate anime waifus on a discord bot

[pale latch]

i getcha

[vagrant fox]

yeah entitled brats thinking someone will provide all resources for them, no electricity needed, no hardware needed

if they want 99.99% uptime, go buy MJ subscription

[fervent thunder]

Yep. That's why I take slight issue with DarkLaw's "everyone will be happy" comment. Someone will always be like:

"What do you mean, you won't provide ten warehouses full of RTX 4090s and pay a megawatt power bill so that I can send 69,420 queries per millisecond for all eternity?"

[spring oar]

Hello someone one

[foggy halo]

guys any idea when SD3 release!!??!

[vagrant fox]

this year but could be wrong

[foggy halo]

rough idea or what do you guys think

[fervent thunder]

Hello, your profile says 'new to Discord' and you seem to be in this "server" for no reason, so I assume you're a scammer (crypto-scammer), but if you are actually new to Discord, I would warn you, there are lots of scammers and degenerates here. Be careful.

[fervent thunder]

Likely within a month. 2MinutePapers did a cool video on it. It will initially be remote-only (so you can't generate anything cool).

[vagrant fox]

he might just be exploring, he's on MJ's server too

[foggy halo]

oh like what is remote only?

[fervent thunder]

Who is that?

[vagrant fox]

i have no clue but I can see mutual servers

[trim magnet]

be careful theres a guy with a pepe profile picture trying to scare new people

[fervent thunder]

You won't be able to download the model (initially); you'll just be able to send prompts to it and get results back (like MidJourney), and SD will scan your prompts for naughty words, etc.

[fervent thunder]

I am discheesed with that comment. New people SHOULD be scared.

[vagrant fox]

let's hope model won't be gatekeeped for long

[foggy halo]

Ohh now i understand...thanks for letting me know! will it be free as in on huggingface? just like Stable Cascade

[spring oar]

I want free ai tools please help to get

[fervent thunder]

To be fair, I am new to Discord too (don't like it much, really I only signed up to get on the SD3 waiting list).

guessing the invites happen this week and model release in 3 weeks

[fervent thunder]

We are not Encarta.

[vagrant fox]

use google notebook or run sd on your local machine, there is plenty of tutorials on yt

[foggy halo]

i didnt understand the SD3 model comparison graph it is kinda confusing

[fervent thunder]

depends they are doing a/b testing or fine-tuning runs

[foggy halo]

also what is our w/o t5 (or smth like that)

[fervent thunder]

What are you talking about? nobody promopts less than full DD cups.

[fervent thunder]

lmaoo, i was trying to guess when sd3 will have full release

[fervent thunder]

Ahhh, got it. My guess would be Q1/2 2024.

[fervent thunder]

I don't like unrealisticaly huge bewbs

[full lark]

That bot is expensive to run 😁

[vagrant fox]

let him root in lone silence

[dim minnow]

you're right... at the end of the day, running the bot is a hard thing to do... sorry devs...

[vagrant fox]

get your own hardware and electricity and stop crying like lil boy

[trim magnet]

whats the matter,no local? 🗿

[dim minnow]

[vagrant fox]

i suppose daddy didn't bought him new pc since 2016 and bro didn't started to make his own money yet

[dim minnow]

[shell tendon]

jeez, they already give out the source code and the weights and the user interface for free lol

[trim magnet]

but not the gpu

[shell tendon]

that's the real crime

[vagrant fox]

yeah, gpus should be included in every pack of chips, how unfair life is!

[dim minnow]

but how... how does one access it...

[shell tendon]

you have to book a flight to pyongyang and complete a vision quest at a mountaintop shrine to totalitarianism

then, and only then, will the answer reveal itself to you

[dim minnow]

oh, ok

✈️

[pearl ocean]

so yeah, lol, now on chome, I have Pinned tabs for, Gemini, ChatGPT, CoPilot, Claude, and Stable diffusion

[fervent thunder]

y'all should just wait 50 years, you'll be able to run sd on a phone at 200it/s

[lone hawk]

in 50 years this will be primitive ancient tech

[pearl ocean]

we even going to be around in 50yrs?

[short crater]

hello!
Stble Diffution Rest Apis is not working now

Anyone having the same problem?

[shell tendon]

no shit lol

[waxen path]

Here D:

[winged dove]

Yeah same. Can't check my credits either on the site. Seems like the backend is busted.

[sterile gust]

Wish i could say the same but i have made a life here sadly, knowing SD3 is going to get as heavy censorship as SD2 and SDXL i don't have my hopes up.
But i got nothing to lose so signed up for SD3 beta

[rich kestrel]

Same. I am ready to be yet again thoroughly disappointed.

[pearl ocean]

is SD3 model on local being restircted?

[desert spoke]

how to use ?

[pale latch]

funny and so narrow minded. when all you want is nsfw porn, you only need a 300m parameter model. stick with sd 15 because you've got what you want already and you sure don't need more parameters for one concept

[wet dust]

Is there a way to specify a subject from a concept? For example, I want to generate Doomsday, from the Superman comic books. But just prompting "Doomsday" doesnt really give me that character, it gives me generic doomy things

[pale latch]

sd3 is going to cover every other possible concept, or as much as possible

[rich kestrel]

l0l

seems like someone else is narrow minded if you think porn has a limit

[pale latch]

sure you can go all you want in that narrow scope but it's a narrow scope

300m parameters is all you need. run wild with it

[rich kestrel]

its not. more detail is more detail

[sterile gust]

It can't be which is why the training data will most likely be free of anything that can be considered inmoral or NSFW.
Which just cripples quality and creativity as we have seen with other cloud restricted models.

[rich kestrel]

We can still hope it will be easily trainable

[pale latch]

sd2's big failing wasn't nsfw . it was because it was hard to refine and make loras for.

[rich kestrel]

in which case just give the community 3 days to have it learn female anatomy

[pale latch]

which meant you couldn't train it to do porn

[rich kestrel]

yah. 1.5 revamped is what we need

is that what we're getting? I doubt it

but we'll see

[pale latch]

if you think xl is incapable, load a few models. try pony. hah

[nova zodiac]

Nope… wayy different architecture

[rich kestrel]

pony is relatively new

[pale latch]

pony is xl refined by the community

[trim magnet]

yea pony the guy who used 2 a100's and 2million imgs for a couple of months

[nova zodiac]

Yep - it shows you can take a censored model and train it to your whims if you have the time and budget

[rich kestrel]

yea, but relatively new

[rich kestrel]

exactly this

[pale latch]

there's 100 of those guys out there. there'll be more in the future

[rich kestrel]

think of it like modding

a game like fallout NV has a shitload of good mods bc it's worth investing time into

starfield is not

same goes for models like these

1.5 was worth investing time training, 2.1 l0l hell no

[pale latch]

lots of good mods for starfield. i shoudl update mine good call

starfield just caught the usual hate train. every bethesda game ever gets it

[rich kestrel]

compared to oblivion, skyrim and NV? no

nope. skyrim was revered

the top modders for those game said it themselves

[nova zodiac]

It depends if the architecture allows for good training

[pale latch]

modders for previous bethesda games rarely crossed over in the past

they're used to what they know

entrenched

[rich kestrel]

another wrong statement

[nova zodiac]

Modders can only mod if the game is pliable

[pale latch]

not a lot of skyrim modders jumped onto fallout 4

[rich kestrel]

my dude here spittin 💦 fax 📠

[pale latch]

or fallout 3 to skyrim

or oblivion to fallout 3

or morrowind to oblivion

[trim magnet]

just buy a couple of A100's easy game right

[pale latch]

or daggerfall to morrowind

or arena to daggerfal

[nova zodiac]

Lol with enough compute most models become pliable lol

[sterile gust]

Everyday i feel like SD3 will become Stability's Starfield.

[pale latch]

the youtubers will 100% milk this. you'll get hate baited

[nova zodiac]

I reckon sd3 is gonna be close to sdxl in terms of trainability

[pale latch]

my guess is aientrepreneur is going to kick it all off and start the hate cycle

usual youtuber stuff

[rich kestrel]

aitrepreneur has been right in the past

I don't think his channel is still relevant for SD

[pale latch]

eeeeeeeeeehhhh.. stopped clock

[rich kestrel]

I unsubbed after he went full on for LLMs

but for early SD he was the ultimate guider

[pale latch]

not really. he relayed tons of misinformation

[nova zodiac]

I had Nerdy Rodent as #1 source for how to guides but he went quiet for a bit and then never really went into comfy

[rich kestrel]

olivio did. if ur not following him

all he does is comfy which is koo. I no longer hate on comfy users

to each their own. I used it for cascade and like it

[nova zodiac]

He got too political with the whole clubs civitai debacle

And sells out to hype up models

[fervent thunder]

you don't like comfy?

[rich kestrel]

I like it, it's just hella overrated

[pale latch]

he's a hype man. he hypes topics for views.

[rain sierra]

give me one reason not to leave this server

[rich kestrel]

I'd love even more development for A111

[pale latch]

just enough research to fill 15min

[rich kestrel]

but I get why some ppl like comfy

[rain sierra]

whats so great about stable diffusion

sell it to me

[nova zodiac]

Sd3 bot coming here first - likelyhood 80%+

[trim magnet]

joe biden

[rain sierra]

whats sd3 bot

[pale latch]

text 2 images

[rain sierra]

heh

is it

[pearl ocean]

😮

[pale latch]

The UNDERTAKER

[rain sierra]

going to get better later

or is stable diffusion just is what it is

[pale latch]

its ai. it will grow and evolve until it is our god incarnate

[nova zodiac]

It gets better daily as people add their new resources

[rain sierra]

I for one welcome our new AI overlords

ok

you convinced me to stay

[pearl ocean]

hey I said that yesterday

[rain sierra]

its the line from timmys dad that lives rent free in my head

from fairly odd parents

[pale latch]

#🏞｜general-with-images message

[rain sierra]

lol

[pale latch]

triggered

it is NOT the line from fairly odd parents!

[rain sierra]

wheres it from

[pale latch]

(╯°□°)╯︵ ┻━┻

[rain sierra]

lel

[full lark]

┳━┳ ノ( ゜-゜ノ)

[pearl ocean]

(╯°□°)╯︵ ┻━┻

[pearl ocean]

imagine SD3 being more restrictive

[pale latch]

skill issue

[trim magnet]

dont need to imagine,already happened here

[pale latch]

unskilled generators

[pearl ocean]

🙏

[nova zodiac]

I doubt it would be more restrictive than the sd2.0 release

that thing was a eunuch compared to sdxl

[full lark]

Why they just dont launch iiiiiit? 😭

[nova zodiac]

cause they are still finetuning it

[full lark]

Aaa

[nova zodiac]

they have decided on the architecture and have shown us things from well along the training, but there are still more epochs happening which may give us a better result yet

I'm guessing we will get a similar release to sdxl where they will have a couple of checkpoints they can decide between internally and shove both on the bot for some RLHF a/b testing and there will be a final release later in april/may

[pine fiber]

most likely

they are censoring violent and nsfw prompts fyi

[round rapids]

does anyone know how to update Stable Diffusion from being from 22.2.1 -> 24.0 using this command \Stable Diffusion\stable-diffusion-webui\venv\Scripts\python.exe -m pip install --upgrade pip

[forest osprey]

So you mean the bots are back in April or may, am I right?

[static cape]

He probably does not know and can't know.

Soon TM.

[pale latch]

how? you can't censor models that way. you can block prompts, but if we can run the weights, we control the prompts

[shell tendon]

man, differential diffusion in comfy blows away soft inpainting

[pale latch]

thye could not have any violence in the training data, but i mean, thats not a thing

[pine fiber]

for the api idk they didn’t say exactly how

[pale latch]

even 2.1, as censored as it is, still has nudes in the training data as photography

[pine fiber]

I hope they didn’t prune the training data like that

[pale latch]

dataset needs to be pruned. you can't not cull it

3 has an all new dataset. its not laion 5b

anything through their service will likely be censored for liability reasons

[nova zodiac]

Estimate, take with a dump truck full of salt

[tropic mason]

how do i use a checkpoint model from civitai.com and setup the APIs to build an application?

[nova zodiac]

Whats your use case for the API?

[wise bear]

I hope they don't lobotomize it ;_;

I hope people get a reason to step away from SD1.5

[pale latch]

thats not how censoring models work. they don't gut any weights post training. they only train

[wise bear]

Talking to me?

[pale latch]

2.1 wasn'te ven as censored as people think. it was a whole different clip language plus vprediction was hard to train

things will be fine.

[nova zodiac]

All depends on the dataset they train with. Given this is a public model, I would not be expect large amounts of nsfw to be present in the dataset

[wise bear]

That's neat

[nova zodiac]

2.1 wasnt, but 2.0 was

[wise bear]

I then just need to hope that it can be trained easily

[nova zodiac]

Dont know until we try it

[wise bear]

I want to spend some good time on the paper for SD3 :3 see if there's any further info or hints.

[pale latch]

even then it wasn't as bad. with nsfw. it was just really haggard and they didn't have experience of what to look for during training. 2.0 was stability's own team's first model. runway ml did 1.5

you prompt nude photographer names and it'd do boobies

[nova zodiac]

True about runway - it wasnt terrible but they did admit they had messed up the threshold at which they would remove the image from the training data

[pearl ocean]

Hopefully it adds imperfections

[tropic frost]

say, usually how many images you generate at once when you for a silghtly higher res?

[jovial wraith]

what do you mean exactly? i usually generate until i'm happy with an output, which i then either upscale or hires fix

[pale latch]

setting the cut threshold to .9 instead of .99

thats a lot of difference on 2 billion images haha

[tropic frost]

basically my method is, you generate images with low resolution, say 512x512 until you get something close to what you are looking for. Then you take it to image to image, and set resolution to 1024x1024 so you get a clearer image.

[jovial wraith]

ah, like that. well, i just keep going until i found a prompt that works well

[tropic frost]

not what im asking about, how many images do you generate at once?

[jovial wraith]

just one. my pc can't handle more than that

[tropic frost]

for some reason, when i was generating images at a higher res a few days ago, i could got 4 at once.
not it won't allow it, and can do 2 at once max

[astral goblet]

i do prompt craft at low resolutions with lcm loras. then once i'm happy with prompting, i crank on quality. sometimes i rely on large batches, generally not needed. i have a very low gear brute force. i like to rock crawl the latent space

[jovial wraith]

don't click

masked link

@bleak matrix

[warm junco]

I timed him out

Thanks

Also Users can report any Message by racting with ⚠️

[jovial wraith]

oh, i didn't know that. will do that in the future! :D

[shell tendon]

thanks for the info, didn't know either.

[thorny sedge]

ok

[crisp zinc]

I think I figured out why Midjourney looks so aesthetically pleasing.

It's well known that they have extensively utilised user rankings and are collecting them, even giving people GPU hours to do it.

Their most aesthetical model, V3, is able to reproduce niche and extensive features from variety of concepts without overfitting or underfitting, and without breaking the aesthetic.

I think they trained a aesthetic predictor AI on user votes and have used it to create a text encoder token to represent the aesthetic of a given image. They have done this for every single image in their dataset. This allows the AI to learn features from 'uglier' images (like maybe some shitty mobile screenshot) while being able to apply these features in a beautiful way, without breaking down the style.

And when a user generates an image, they just inject that token with maximum beauty. Hell! They even have a parameter called --stylize that do exactly that and they don't let you put it lower than a given value (at least for V3 and I think for newer models they just normalised it so that 0 is still giving it some aesthetic)

This would also explain why their models find it difficult to generate 'ugly' people.

[desert nebula]

I had a dream about an IP-adapter error

[nova zodiac]

Ugly people being hard isn't just unique to Midjourney - it's to do with how humans find symmetrical faces aesthetically pleasing, and the fact that an AI model when learning is taking an average of all of the faces with that token in the dataset, and when you average out a lot of faces you end up with something very symmetrical

[crisp zinc]

SD has no issue generating 'ugly' people, MJV3 can't even generate old people.

I think my hypothesis is strong, what do you think?

I'm actively training models based on midjourney rn.

[nova zodiac]

The aesthetic predictor hypothesis.. plausible and probable

stability did the same for sdxl

my question is why focus on mjv3 when they are up to v6?

[crisp zinc]

Oh? Can you prompt it?

[crisp zinc]

V3 still looks more beautiful than XL, Cascade, V4, V5 and V6.

you can be extremely lazy with prompting and still get a masterpiece

[nova zodiac]

no.. they did a whole heap of A/B testing which would have created an aesthetic predictor under the hood for the final training epochs of sdxl 1.0

[crisp zinc]

I'm training a Lora on V3 style and its giving me masterpieces with one word prompts or other lazy stuff, and the images aren't even incoherent like V3.

[crisp zinc]

But that can lead to catastrophic forgetting and other undesirable stuff. It would be better if they labelled the entire dataset and told the model.

Because that would allow the model to generalise and understand aesthetic. Like if the final epochs only include 'aesthetic' images, the model has no incentive to preserve any prior features learned from unaesthetic images.

This can be aided partially by creating a LoRA but it won't be as good because a LoRA can only hold so much, especially one that is probably trained with under 10,000 images of the concept, so it will likely be less flexible.

Of course I'm totally grateful for the existance of Stable Diffusion. However improvement necessitates criticism.

[frank kernel]

gotha thanks!

[hexed chasm]

Is it possible to run Stable Video Diffusion on the cloud?
I have 8 GB of vram on my video card, and am getting lots of memory overflow errors ("OutOfMemory")

[warm junco]

Hey what's your GPU?

You may need just some performance settings

And yes SD can run in the cloud too

Some services offer virtual GPU power

[bleak matrix]

Good morning, everyone! How are we all today? 🌥️

[brave crest]

Is there anyone who has experience with generating images that look the same?
To be more precise, I have a customer that want to update their fitness catalog of their app and they want to be able to show how the movement is done and which muscles are involved. They wanted to know if it was possible to do with AI - my only concern is that with 100+ exercises, the images would look different to eachother

[bleak matrix]

I would suggest that you look into training for this. You can find more information in #1080946152318443610

[brave crest]

Alright cool. Do you think it would be feasible with a limited set of images for it to resemblence? I would like to provide enough training data for it to have a baseline for what the image should look like

Right now they have just videotaped people doing the exercises, so there isnt much data to be provided in the new context

[bleak matrix]

I would look into LoRAs for that.

[chilly hinge]

Hy! Where is the stable diffusion bot?

[nova zodiac]

#1047610792226340935

[tropic mason]

to create an img generation app

[nova zodiac]

for yourself, or for wider consumption?

[wary sparrow]

Joined before sd 1 released any way I could get early access?

[mortal sphinx]

I have a blurry picture, can I use Stable Diffusion to fill in the gaps and upscale it?

[brittle maple]

[sudden ruin]

Join the wait list and sit tight

[brittle maple]

flowwolf DDU not work

[mellow zinc]

hey peeps, what model you think is best for generating non human images? like scenery and things like that

[brittle maple]

i found best for human dont know )

[worthy bone]

@sudden ruin Like many others, I have been on the waitlist for a few weeks, will it take much longer? 🙂

[zealous pine]

How to use the prompt feature to generate images

[sudden ruin]

I have no insights, I know as much as you do pretty much. Just wait and relax

[sudden ruin]

#1047610792226340935

[worthy bone]

Ok thanks 🙂

[rich kestrel]

that depends on several factors

if its just a random image ur better off upscaling with topaz

if its a known subject or a trainable subject, you can upscale with cnet + img2img

SD is so underrated for image restoration

[mortal sphinx]

What's the best upscaler with some sort of AI correction then?

Free of course.

I don't know how to use cnet + img2img yet

[warm junco]

Upscayl is nice:
https://github.com/upscayl/upscayl

or you can also use the sd upscale script in img2img or hires fix in txt2img

controlnet is not needed for upscaling

[broken cave]

this has a lot of product development for something that is basically a 7 line python script...

[rich kestrel]

I use topaz and didn't pay for it wink wink nudge nudge

it's just a great general use upscaler

but if you want something basic then yeah, the Extras tab upscaler also works

img2img would only work if the AI knows who the subject is, aka loras or dreambooth

[broken cave]

if only these developers would harness their manic energy all into one goal

[vagrant fox]

i got topaz ai and tbh isn't anything special

auto detection ai feature sucks kind of

[rich kestrel]

which one is better

I also use it for video upscaling but always looking for the next best thing

[vagrant fox]

video upscaling isnt issue since u can cut video to frames anyway

and use any upscaler

[warm imp]

Is there somewhere here I can go for help with img2img upscaling on AI Horde?

[gloomy bison]

Where do I install the original august 2022 stable diffusion

[broken cave]

why are you using ai horde?

[broken cave]

why?

[gloomy bison]

stable diffusion came out august 22 2022 i thought

[warm imp]

Bcuz my PC not good enough for local + don't want to use Google shit

[gloomy bison]

just for fun

[vagrant fox]

get yourself 1.5 model then i suppose

[gloomy bison]

just to see how far ai image gen has come

[broken cave]

https://github.com/CompVis/stable-diffusion

there is sd-v1.1

[vagrant fox]

must be painful to use it

[broken cave]

when you say google shit, do you mean collab?

[vagrant fox]

most likely

gemini is shit also

[gloomy bison]

is therea a ckpt or safetensors

[warm imp]

Yes

[broken cave]

i think you can handle it from here

[broken cave]

but why?

[brittle maple]

zielone_ciastko hello your problem fix with redowndload model?_)

[vagrant fox]

it did fix with this command line --no-half-vae

[brittle maple]

like in my bat file

[vagrant fox]

yes

[brittle maple]

i help i think?)

[vagrant fox]

?

[brittle maple]

you use my setting or experiment?

[gloomy bison]

no i can't

[vagrant fox]

i have been helped by @warm junco

[brittle maple]

he help me too)

i use DDU today

its not work

[vagrant fox]

well sucks, think of saving up for new pc setup

[brittle maple]

hardware bug

[sour mica]

Has Stable Diffusion completely become pay to use? It's been a while since I used it and nor the website or the discord server seems to work anymore, is there any other way?

[shell tendon]

just did some tests with a fully updated comfy and a fully updated a1111

with sdxl 1024x1024 and dpmpp_a/karras (50 steps) i'm getting 1.75 it/s on a1111 and 2.76 it/s on comfy

so even with those recent updates... even with running the same version of python and torch... the speed gap remains present

[still glacier]

local installs work just fine, bot channels are offline for now #1047610792226340935

[vagrant fox]

piss off

[still glacier]

what ?

[vagrant fox]

it was to msg that was deleted someone posted some porn dc inv

[shell tendon]

screencap em next time, paste the screencap with the exclamation emoji and it'll attract the mods

[rich kestrel]

I actually joined it before they removed it

obvious scam

to steal your login

[static cape]

I got some "I do cheap commissions" Spam yesterday.

[astral goblet]

reply to that spam with "i can out perform you"

[teal osprey]

do we have the technology to generate prompts from images? to recreate or remix similar stuff. the image is a screenshot, not an ai generated that has gen. data

[rich kestrel]

reference function on controlnet

or maybe you mean captioning which is also a thing

[teal osprey]

but i want to see and learn the logic behind the generation through the prompts

this should even be a contest or something where contestants try to recreate the closest image to he original one by guessing the prompts or creating their own

[sour mica]

Ohh alrightt Is there any guide for local install?

[astral goblet]

image prompt adapters are good too

[hallow sable]

I hate how everyone just makes people with SD lmao

[shell tendon]

gets boring

[true canopy]

im trying to make the ultimate cute frog right now tho !

[grizzled harness]

yeah there's like 4 pictures people keep remaking 😛

the waist-up-walking-toward-me digital person, probably of Asian descent ...
the cyberpunk whatever...
the political figure doing something stupid...
the origami version of whatever...

[hallow sable]

the sat map model someone made is pretty cool but somewhat useless, ive made a composite image.

https://cdn.discordapp.com/attachments/1004159122335354970/1215009351283118121/image.png?ex=65fb30a3&is=65e8bba3&hm=ff43dec2993a69d9a8dcc9f6abbf45af47a40d7fa14cafe6b78d8e930d738810&

but it wont replace gaia or blender

Gaia can be done fully procedurally

sadly the sat map model doesnt work with img 2 img 😦

[static cape]

I feel like the answer is just a simple goggle search away...

[sour mica]

Yeah found it

[static cape]

Interesting. I think it's strange how currently the trend goes towards "Let's just prompt for something ridiculous and ignore the details".

[fading gyro]

yo yo. just starting with stable diffusion! my first question (and thanks in advance for the guidance) is how do I choose checkpoints/LORAs? do I choose the checkpoint first? do I need a checkpoint?

my goal is to be able to create flat/drawn cartoonish characters of a consistent style

[gaunt pulsar]

Might be better to just wait for SD 3 imo

[fading gyro]

wdym?

[gaunt pulsar]

There is a really good stable diffusion model that is probably entering early preview very soon.

It's so much better than current models that anything you make today will likely become outdated as soon as it's released

[gloomy bison]

Why is stable cascade and sd 3 being developed around the same time

if sd 3 is better than stable cascade why not focus all development on sd 3

[brittle maple]

lora its like minimodel help get best result

[fading gyro]

lora = fine tuning, checkpoint = ?

[brittle maple]

checkpoint is base

[fading gyro]

i know civitai is a good place for LORAs, where can I find good checkpoints?

[brittle maple]

yes you can get lora and model

[hallow sable]

dont know bro, watch some videos instead of asking everyone to hold your hand on the most basic stuff

[fading gyro]

kick rocks buddy

[gloomy bison]

kick rocks buddy

[brittle maple]

i start use lora maybe week ago

[hallow sable]

what is this czechpoint <:^) and how me useing

[vagrant fox]

maybe creating one gives them experience and testing grounds to push another one further

[hallow sable]

is cascade made by the same devs? ive never tried it or looked much into it

[gilded lantern]

hey eveyone, i made a multi function AI bot if anyone wants to try it out.. text, image, voice, music

[fading gyro]

noob question... whats the difference between SDXL vs Stable Cascade? more so what are the use cases

[gilded lantern]

i also have video, avatar, and more programmed in but limited on GPU

[brittle maple]

how work your bot ? local in pc?

[vagrant fox]

have you got any videos showcasing tool already

?

whats voice tools? can you teach it certain voice and it will replicate from text on what it has been learning?

I am happy to cover new tools on my youtube but I need to know this works and its not just some virus garbage

i am currently checking your website

[still glacier]

Check #🤝｜tech-support 's pinned messages

[gilded lantern]

yes! all open source local, all using transformers library, not even langchain

[brittle maple]

what is it exe file?

[vagrant fox]

I don't see any voice and sound examples, just videos/animations

[gilded lantern]

hey cool! I am still doing lots of coding and will be active for a long time. I do have an entire website, AI platform and multi-platform bots

[vagrant fox]

i mean i see 3 sec sample

[gilded lantern]

on my YT i only have the videos, I have a discord server (link in my profile) where i have the multi AI bot

[gilded lantern]

where at? all the times can be set, and still need to add the controls for that

[vagrant fox]

well you got only one voice post on your website

and it's that one

[gilded lantern]

all python

[vagrant fox]

what are your features as voice and sound generators since there is not much info about it

[gilded lantern]

oh okay yep! I havent added many of them yet but they do work and will add the time controls.. i've been working on the slideshow, and the web app connect to YT which is nice, got that working

I've been nearing a release, it's still not realesed yet.. i can send access codes if you're interested and would be able to check it out..

all in the same AI thread, you can AI chat, voice, image, slideshow, storyteller video

[vagrant fox]

I am mainly focused on voice generating/replicating and music creation features, could have look into video generation

what is the size of it with all models required to run?

[gilded lantern]

and more.. theres a whole AI blog maker, coder (the code also posts to gitlab)

[brittle maple]

sometimes it seems to me that SD itself learns based on my queries and later produces similar results when I do not specify details

without train

[vagrant fox]

we live in matrix, matrix listens to us, lol

[gilded lantern]

I think it's around 2GB quantized to 4bit

[brittle maple]

sometimes is crippy)

[gilded lantern]

ya all this is lately

[crimson kernel]

One of the stable devs (specifically the one responsible for getting data for training) brought down the midjourney servers over the weekend for 24hrs trying to scrape prompts and images, and as a result the CEO is banning all Stability employees from the service indefinitely.

[gilded lantern]

the music is only mediocre at the moment, I'm still looking for a better one but the one i'm using is fully open source

[crimson kernel]

News just broke from the MJ CEO on his office hours

[vagrant fox]

source, links

[gilded lantern]

wow lol

[crimson kernel]

Office hours on the MJ server right now

[vagrant fox]

they release recordings of those voice meetings when they end right?

would be nice stuff to cover and make video about if its true, sounds crazy

[crimson kernel]

No 🙅‍♂️ recording is bannable

[vagrant fox]

yikes, then why they make it, should be like twitter spaces, you can listen to it when u miss opportunity to be live at event

[crimson kernel]

There are people who take notes and journalists who show up weekly

Only way to hear it is to be there

[gloomy night]

So what is this thing that allegedly SD tried to scrape the MJ site Satuday?

David Holz Just said so in Office hours

[crimson kernel]

Prompts and images

Really shady and not cool of them

[gloomy night]

Is that actual real?

So far it's only an allegation from DH

[vagrant fox]

why they didnt think of making scape bot that just read their channels on dc or their website instead of breaking into database

[gilded lantern]

ya that's crazy

[crimson kernel]

It's real

[gloomy night]

Evidence?

So far it's only DH word

[vagrant fox]

get different model thats not fucked up

who else got issue with outdated weird model?

https://www.reddit.com/r/StableDiffusion/comments/yyv487/anyone_else_has_a_modelcpkt_with_hash_a2a802b2_is/

seems like you are not only one looking for it

there is link in comments but site does not work

[brittle maple]

i eat many chilly pappers mouth burned

[vagrant fox]

sounds like one of prompts people on fb groups do xD

why do you empty your trash bin without checkin contents then? it didn't delete itself i believe, what was so special about this model?

i believe it's improved and contains more parameters

For text-to-image, pass a text prompt. By default, SDXL generates a 1024x1024 image for the best results. You can try setting the height and width parameters to 768x768 or 512x512, but anything below 512x512 is not likely to work.

[gilded lantern]

@vagrant fox have you seen suno bark for voice? I mainly use that because it's smaller size to run. there's also metavoice https://github.com/metavoiceio/metavoice-src and its nice they have a ready-to-go docker but theyre model uses maybe 6GB or so

but does have a better voice

[vagrant fox]

i will check this out

[gilded lantern]

yes bark is slow, i need to find better.. coqui is good too but cant find a good commercial open model besides metavoice(i think they use coqui?)

ya that coqui does run nice

[vagrant fox]

i will look today into that since i am using browser based service for tts and would be nice to have something locally with good quality, have you got some selection of voices?

can you teach it your voice?

[gilded lantern]

coqui and metvoice you can

not the best time for traditional content creators lol

[vagrant fox]

i need to find working workflow, some yt tutorial then test it myself and release vid about it on my channel 😛

"Efficient, flexible, lightweight but feature complete Trainer API" so i suppose u can teach coqui voice

[gilded lantern]

i am working on tutorial videos too, that will be really valuable, 1 click text to tutorial video

[vagrant fox]

why not? you can still create content, and people still preffer to see real person than no person or some vtuber

[gilded lantern]

metavoice is the best i found for that, and if you install docker, you can start it up in 1 command line

[gilded lantern]

ya for sure. I'm sure it will eventually all fall in its place

which model?

[crimson kernel]

they have 2 credit cards that match a name on LinkedIn for that particular employee there

[gilded lantern]

damn im not sure im using 1111

[crimson kernel]

oh also the SD employee responsible waited until midnight California time to start the scrape, wanted to make sure the devs were in bed. it was clearly a calculated scrape/takedown

[vagrant fox]

you could train your own model with paitings tho

#🏞｜general-with-images

i am hoping someone will ask for more info about that sd drama on office hours, staying there on listen

[gilded lantern]

@vagrant fox what'd you think about the website overall? i suppose i do still need to add more audio samples!

[vagrant fox]

yes, you have shit ton of video examples, but only one 3 sec voice example and 0 for music generatior, also would be nice to have video showcasing your's tool ui

[gilded lantern]

yep! the music will post but yes i will have to make demonstrations

[fathom osprey]

Not sure if I am asking in right place but ...
Hello, I'm new to stable-diffusion and seeking guidance on generating visually appealing anime character images in low resolution, around 150x300 pixels. I'm looking for vibrant, colorful images that stand out. Could someone recommend the best model, such as VAE, or suggest any extensions that could assist me in achieving this? I'm using Automatic1111. Any insights or advice would be greatly appreciated. Thank you!

[gilded lantern]

nice! hope so. i was wondering if it would be good to start keeping track of all those offline models for that reason. i hope you find it!

[gilded lantern]

i dont make much of those but the aniverse, animerge models on civitai would probably work good

i like dreamshaper8 in auto1111 and right in python

that turbo seems to have the tradeoff of using a little more VRAM.. the dreamshaper seems to run really good for speed and minimal vram

cool

ya i noticed that.. havent really checked it out much

[hallow sable]

you have any programming videos?